574 files changed, 104532 insertions, 0 deletions

diff --git a/pkg/sentry/BUILD b/pkg/sentry/BUILD
new file mode 100644
index 000000000..d18cf3555
--- /dev/null
+++ b/pkg/sentry/BUILD
@@ -0,0 +1,12 @@
+# This BUILD file defines a package_group that allows for interdependencies for
+# sentry-internal packages.
+
+package(licenses = ["notice"])  # Apache 2.0
+
+package_group(
+    name = "internal",
+    packages = [
+        "//pkg/sentry/...",
+        "//runsc/...",
+    ],
+)
diff --git a/pkg/sentry/arch/BUILD b/pkg/sentry/arch/BUILD
new file mode 100644
index 000000000..a88f57ac7
--- /dev/null
+++ b/pkg/sentry/arch/BUILD
@@ -0,0 +1,66 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "arch_state",
+    srcs = [
+        "arch.go",
+        "arch_amd64.go",
+        "arch_state_x86.go",
+        "arch_x86.go",
+        "auxv.go",
+        "signal_amd64.go",
+    ],
+    out = "arch_state.go",
+    package = "arch",
+)
+
+go_library(
+    name = "arch",
+    srcs = [
+        "aligned.go",
+        "arch.go",
+        "arch_amd64.go",
+        "arch_amd64.s",
+        "arch_state.go",
+        "arch_state_x86.go",
+        "arch_x86.go",
+        "auxv.go",
+        "signal_act.go",
+        "signal_amd64.go",
+        "signal_info.go",
+        "signal_stack.go",
+        "stack.go",
+        "syscalls_amd64.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/arch",
+    visibility = ["//:sandbox"],
+    deps = [
+        ":registers_go_proto",
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/cpuid",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
+
+proto_library(
+    name = "registers_proto",
+    srcs = ["registers.proto"],
+    visibility = ["//visibility:public"],
+)
+
+go_proto_library(
+    name = "registers_go_proto",
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/arch/registers_go_proto",
+    proto = ":registers_proto",
+    visibility = ["//visibility:public"],
+)
diff --git a/pkg/sentry/arch/aligned.go b/pkg/sentry/arch/aligned.go
new file mode 100644
index 000000000..193232e27
--- /dev/null
+++ b/pkg/sentry/arch/aligned.go
@@ -0,0 +1,31 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+import (
+	"reflect"
+)
+
+// alignedBytes returns a slice of size bytes, aligned in memory to the given
+// alignment. This is used because we require certain structures to be aligned
+// in a specific way (for example, the X86 floating point data).
+func alignedBytes(size, alignment uint) []byte {
+	data := make([]byte, size+alignment-1)
+	offset := uint(reflect.ValueOf(data).Index(0).Addr().Pointer() % uintptr(alignment))
+	if offset == 0 {
+		return data[:size:size]
+	}
+	return data[alignment-offset:][:size:size]
+}
diff --git a/pkg/sentry/arch/arch.go b/pkg/sentry/arch/arch.go
new file mode 100644
index 000000000..021789e4b
--- /dev/null
+++ b/pkg/sentry/arch/arch.go
@@ -0,0 +1,351 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package arch provides abstractions around architecture-dependent details,
+// such as syscall calling conventions, native types, etc.
+package arch
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Arch describes an architecture.
+type Arch int
+
+const (
+	// AMD64 is the x86-64 architecture.
+	AMD64 Arch = iota
+)
+
+// String implements fmt.Stringer.
+func (a Arch) String() string {
+	switch a {
+	case AMD64:
+		return "amd64"
+	default:
+		return fmt.Sprintf("Arch(%d)", a)
+	}
+}
+
+// FloatingPointData is a generic type, and will always be passed as a pointer.
+// We rely on the individual arch implementations to meet all the necessary
+// requirements. For example, on x86 the region must be 16-byte aligned and 512
+// bytes in size.
+type FloatingPointData byte
+
+// Context provides architecture-dependent information for a specific thread.
+//
+// NOTE: Currently we use uintptr here to refer to a generic native
+// register value. While this will work for the foreseeable future, it isn't
+// strictly correct. We may want to create some abstraction that makes this
+// more clear or enables us to store values of arbitrary widths. This is
+// particularly true for RegisterMap().
+type Context interface {
+	// Arch returns the architecture for this Context.
+	Arch() Arch
+
+	// Native converts a generic type to a native value.
+	//
+	// Because the architecture is not specified here, we may be dealing
+	// with return values of varying sizes (for example ARCH_GETFS). This
+	// is a simple utility function to convert to the native size in these
+	// cases, and then we can CopyOut.
+	Native(val uintptr) interface{}
+
+	// Value converts a native type back to a generic value.
+	// Once a value has been converted to native via the above call -- it
+	// can be converted back here.
+	Value(val interface{}) uintptr
+
+	// Width returns the number of bytes for a native value.
+	Width() uint
+
+	// Fork creates a clone of the context.
+	Fork() Context
+
+	// SyscallNo returns the syscall number.
+	SyscallNo() uintptr
+
+	// SyscallArgs returns the syscall arguments in an array.
+	SyscallArgs() SyscallArguments
+
+	// Return returns the return value for a system call.
+	Return() uintptr
+
+	// SetReturn sets the return value for a system call.
+	SetReturn(value uintptr)
+
+	// RestartSyscall reverses over the current syscall instruction, such that
+	// when the application resumes execution the syscall will be re-attempted.
+	RestartSyscall()
+
+	// RestartSyscallWithRestartBlock reverses over the current syscall
+	// instraction and overwrites the current syscall number with that of
+	// restart_syscall(2). This causes the application to restart the current
+	// syscall with a custom function when execution resumes.
+	RestartSyscallWithRestartBlock()
+
+	// IP returns the current instruction pointer.
+	IP() uintptr
+
+	// SetIP sets the current instruction pointer.
+	SetIP(value uintptr)
+
+	// Stack returns the current stack pointer.
+	Stack() uintptr
+
+	// SetStack sets the current stack pointer.
+	SetStack(value uintptr)
+
+	// SetRSEQInterruptedIP sets the register that contains the old IP when a
+	// restartable sequence is interrupted.
+	SetRSEQInterruptedIP(value uintptr)
+
+	// StateData returns a pointer to underlying architecture state.
+	StateData() *State
+
+	// RegisterMap returns a map of all registers.
+	RegisterMap() (map[string]uintptr, error)
+
+	// NewSignalAct returns a new object that is equivalent to struct sigaction
+	// in the guest architecture.
+	NewSignalAct() NativeSignalAct
+
+	// NewSignalStack returns a new object that is equivalent to stack_t in the
+	// guest architecture.
+	NewSignalStack() NativeSignalStack
+
+	// SignalSetup modifies the context in preparation for handling the
+	// given signal.
+	//
+	// st is the stack where the signal handler frame should be
+	// constructed.
+	//
+	// act is the SignalAct that specifies how this signal is being
+	// handled.
+	//
+	// info is the SignalInfo of the signal being delivered.
+	//
+	// alt is the alternate signal stack (even if the alternate signal
+	// stack is not going to be used).
+	//
+	// sigset is the signal mask before entering the signal handler.
+	SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt *SignalStack, sigset linux.SignalSet) error
+
+	// SignalRestore restores context after returning from a signal
+	// handler.
+	//
+	// st is the current thread stack.
+	//
+	// rt is true if SignalRestore is being entered from rt_sigreturn and
+	// false if SignalRestore is being entered from sigreturn.
+	// SignalRestore returns the thread's new signal mask.
+	SignalRestore(st *Stack, rt bool) (linux.SignalSet, error)
+
+	// CPUIDEmulate emulates a CPUID instruction according to current register state.
+	CPUIDEmulate(l log.Logger)
+
+	// SingleStep returns true if single stepping is enabled.
+	SingleStep() bool
+
+	// SetSingleStep enables single stepping.
+	SetSingleStep()
+
+	// ClearSingleStep disables single stepping.
+	ClearSingleStep()
+
+	// FloatingPointData will be passed to underlying save routines.
+	FloatingPointData() *FloatingPointData
+
+	// NewMmapLayout returns a layout for a new MM, where MinAddr for the
+	// returned layout must be no lower than min, and MaxAddr for the returned
+	// layout must be no higher than max. Repeated calls to NewMmapLayout may
+	// return different layouts.
+	NewMmapLayout(min, max usermem.Addr, limits *limits.LimitSet) (MmapLayout, error)
+
+	// PIELoadAddress returns a preferred load address for a
+	// position-independent executable within l.
+	PIELoadAddress(l MmapLayout) usermem.Addr
+
+	// FeatureSet returns the FeatureSet in use in this context.
+	FeatureSet() *cpuid.FeatureSet
+
+	// Hack around our package dependences being too broken to support the
+	// equivalent of arch_ptrace():
+
+	// PtracePeekUser implements ptrace(PTRACE_PEEKUSR).
+	PtracePeekUser(addr uintptr) (interface{}, error)
+
+	// PtracePokeUser implements ptrace(PTRACE_POKEUSR).
+	PtracePokeUser(addr, data uintptr) error
+
+	// PtraceGetRegs implements ptrace(PTRACE_GETREGS) by writing the
+	// general-purpose registers represented by this Context to dst and
+	// returning the number of bytes written.
+	PtraceGetRegs(dst io.Writer) (int, error)
+
+	// PtraceSetRegs implements ptrace(PTRACE_SETREGS) by reading
+	// general-purpose registers from src into this Context and returning the
+	// number of bytes read.
+	PtraceSetRegs(src io.Reader) (int, error)
+
+	// PtraceGetFPRegs implements ptrace(PTRACE_GETFPREGS) by writing the
+	// floating-point registers represented by this Context to addr in dst and
+	// returning the number of bytes written.
+	PtraceGetFPRegs(dst io.Writer) (int, error)
+
+	// PtraceSetFPRegs implements ptrace(PTRACE_SETFPREGS) by reading
+	// floating-point registers from src into this Context and returning the
+	// number of bytes read.
+	PtraceSetFPRegs(src io.Reader) (int, error)
+
+	// PtraceGetRegSet implements ptrace(PTRACE_GETREGSET) by writing the
+	// register set given by architecture-defined value regset from this
+	// Context to dst and returning the number of bytes written, which must be
+	// less than or equal to maxlen.
+	PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, error)
+
+	// PtraceSetRegSet implements ptrace(PTRACE_SETREGSET) by reading the
+	// register set given by architecture-defined value regset from src and
+	// returning the number of bytes read, which must be less than or equal to
+	// maxlen.
+	PtraceSetRegSet(regset uintptr, src io.Reader, maxlen int) (int, error)
+
+	// FullRestore returns 'true' if all CPU registers must be restored
+	// when switching to the untrusted application. Typically a task enters
+	// and leaves the kernel via a system call. Platform.Switch() may
+	// optimize for this by not saving/restoring all registers if allowed
+	// by the ABI. For e.g. the amd64 ABI specifies that syscall clobbers
+	// %rcx and %r11. If FullRestore returns true then these optimizations
+	// must be disabled and all registers restored.
+	FullRestore() bool
+}
+
+// MmapDirection is a search direction for mmaps.
+type MmapDirection int
+
+const (
+	// MmapBottomUp instructs mmap to prefer lower addresses.
+	MmapBottomUp MmapDirection = iota
+
+	// MmapTopDown instructs mmap to prefer higher addresses.
+	MmapTopDown
+)
+
+// MmapLayout defines the layout of the user address space for a particular
+// MemoryManager.
+//
+// Note that "highest address" below is always exclusive.
+type MmapLayout struct {
+	// MinAddr is the lowest mappable address.
+	MinAddr usermem.Addr
+
+	// MaxAddr is the highest mappable address.
+	MaxAddr usermem.Addr
+
+	// BottomUpBase is the lowest address that may be returned for a
+	// MmapBottomUp mmap.
+	BottomUpBase usermem.Addr
+
+	// TopDownBase is the highest address that may be returned for a
+	// MmapTopDown mmap.
+	TopDownBase usermem.Addr
+
+	// DefaultDirection is the direction for most non-fixed mmaps in this
+	// layout.
+	DefaultDirection MmapDirection
+
+	// MaxStackRand is the maximum randomization to apply to stack
+	// allocations to maintain a proper gap between the stack and
+	// TopDownBase.
+	MaxStackRand uint64
+}
+
+// Valid returns true if this layout is valid.
+func (m *MmapLayout) Valid() bool {
+	if m.MinAddr > m.MaxAddr {
+		return false
+	}
+	if m.BottomUpBase < m.MinAddr {
+		return false
+	}
+	if m.BottomUpBase > m.MaxAddr {
+		return false
+	}
+	if m.TopDownBase < m.MinAddr {
+		return false
+	}
+	if m.TopDownBase > m.MaxAddr {
+		return false
+	}
+	return true
+}
+
+// SyscallArgument is an argument supplied to a syscall implementation. The
+// methods used to access the arguments are named after the ***C type name*** and
+// they convert to the closest Go type available. For example, Int() refers to a
+// 32-bit signed integer argument represented in Go as an int32.
+//
+// Using the accessor methods guarantees that the conversion between types is
+// correct, taking into account size and signedness (i.e., zero-extension vs
+// signed-extension).
+type SyscallArgument struct {
+	// Prefer to use accessor methods instead of 'Value' directly.
+	Value uintptr
+}
+
+// SyscallArguments represents the set of arguments passed to a syscall.
+type SyscallArguments [6]SyscallArgument
+
+// Pointer returns the usermem.Addr representation of a pointer argument.
+func (a SyscallArgument) Pointer() usermem.Addr {
+	return usermem.Addr(a.Value)
+}
+
+// Int returns the int32 representation of a 32-bit signed integer argument.
+func (a SyscallArgument) Int() int32 {
+	return int32(a.Value)
+}
+
+// Uint returns the uint32 representation of a 32-bit unsigned integer argument.
+func (a SyscallArgument) Uint() uint32 {
+	return uint32(a.Value)
+}
+
+// Int64 returns the int64 representation of a 64-bit signed integer argument.
+func (a SyscallArgument) Int64() int64 {
+	return int64(a.Value)
+}
+
+// Uint64 returns the uint64 representation of a 64-bit unsigned integer argument.
+func (a SyscallArgument) Uint64() uint64 {
+	return uint64(a.Value)
+}
+
+// SizeT returns the uint representation of a size_t argument.
+func (a SyscallArgument) SizeT() uint {
+	return uint(a.Value)
+}
+
+// ModeT returns the int representation of a mode_t argument.
+func (a SyscallArgument) ModeT() uint {
+	return uint(uint16(a.Value))
+}
diff --git a/pkg/sentry/arch/arch_amd64.go b/pkg/sentry/arch/arch_amd64.go
new file mode 100644
index 000000000..23526fe8e
--- /dev/null
+++ b/pkg/sentry/arch/arch_amd64.go
@@ -0,0 +1,302 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+import (
+	"bytes"
+	"fmt"
+	"math/rand"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// These constants come directly from Linux.
+const (
+	// maxAddr64 is the maximum userspace address. It is TASK_SIZE in Linux
+	// for a 64-bit process.
+	maxAddr64 usermem.Addr = (1 << 47) - usermem.PageSize
+
+	// maxStackRand64 is the maximum randomization to apply to the stack.
+	// It is defined by arch/x86/mm/mmap.c:stack_maxrandom_size in Linux.
+	maxStackRand64 = 16 << 30 // 16 GB
+
+	// maxMmapRand64 is the maximum randomization to apply to the mmap
+	// layout. It is defined by arch/x86/mm/mmap.c:arch_mmap_rnd in Linux.
+	maxMmapRand64 = (1 << 28) * usermem.PageSize
+
+	// minGap64 is the minimum gap to leave at the top of the address space
+	// for the stack. It is defined by arch/x86/mm/mmap.c:MIN_GAP in Linux.
+	minGap64 = (128 << 20) + maxStackRand64
+
+	// preferredPIELoadAddr is the standard Linux position-independent
+	// executable base load address. It is ELF_ET_DYN_BASE in Linux.
+	//
+	// The Platform {Min,Max}UserAddress() may preclude loading at this
+	// address. See other preferredFoo comments below.
+	preferredPIELoadAddr usermem.Addr = maxAddr64 / 3 * 2
+)
+
+// These constants are selected as heuristics to help make the Platform's
+// potentially limited address space conform as closely to Linux as possible.
+const (
+	// Select a preferred minimum TopDownBase address.
+	//
+	// Some applications (TSAN and other *SANs) are very particular about
+	// the way the Linux mmap allocator layouts out the address space.
+	//
+	// TSAN in particular expects top down allocations to be made in the
+	// range [0x7e8000000000, 0x800000000000).
+	//
+	// The minimum TopDownBase on Linux would be:
+	// 0x800000000000 - minGap64 - maxMmapRand64 = 0x7efbf8000000.
+	//
+	// (minGap64 because TSAN uses a small RLIMIT_STACK.)
+	//
+	// 0x7e8000000000 is selected arbitrarily by TSAN to leave room for
+	// allocations below TopDownBase.
+	//
+	// N.B. ASAN and MSAN are more forgiving; ASAN allows allocations all
+	// the way down to 0x10007fff8000, and MSAN down to 0x700000000000.
+	//
+	// Of course, there is no hard minimum to allocation; an allocator can
+	// search all the way from TopDownBase to Min. However, TSAN declared
+	// their range "good enough".
+	//
+	// We would like to pick a TopDownBase such that it is unlikely that an
+	// allocator will select an address below TSAN's minimum. We achieve
+	// this by trying to leave a sizable gap below TopDownBase.
+	//
+	// This is all "preferred" because the layout min/max address may not
+	// allow us to select such a TopDownBase, in which case we have to fall
+	// back to a layout that TSAN may not be happy with.
+	preferredTopDownAllocMin usermem.Addr = 0x7e8000000000
+	preferredAllocationGap                = 128 << 30 // 128 GB
+	preferredTopDownBaseMin               = preferredTopDownAllocMin + preferredAllocationGap
+
+	// minMmapRand64 is the smallest we are willing to make the
+	// randomization to stay above preferredTopDownBaseMin.
+	minMmapRand64 = (1 << 26) * usermem.PageSize
+)
+
+// context64 represents an AMD64 context.
+type context64 struct {
+	State
+	sigFPState []x86FPState // fpstate to be restored on sigreturn.
+}
+
+// Arch implements Context.Arch.
+func (c *context64) Arch() Arch {
+	return AMD64
+}
+
+func (c *context64) copySigFPState() []x86FPState {
+	var sigfps []x86FPState
+	for _, s := range c.sigFPState {
+		sigfps = append(sigfps, s.fork())
+	}
+	return sigfps
+}
+
+// Fork returns an exact copy of this context.
+func (c *context64) Fork() Context {
+	return &context64{
+		State:      c.State.Fork(),
+		sigFPState: c.copySigFPState(),
+	}
+}
+
+// Return returns the current syscall return value.
+func (c *context64) Return() uintptr {
+	return uintptr(c.Regs.Rax)
+}
+
+// SetReturn sets the syscall return value.
+func (c *context64) SetReturn(value uintptr) {
+	c.Regs.Rax = uint64(value)
+}
+
+// IP returns the current instruction pointer.
+func (c *context64) IP() uintptr {
+	return uintptr(c.Regs.Rip)
+}
+
+// SetIP sets the current instruction pointer.
+func (c *context64) SetIP(value uintptr) {
+	c.Regs.Rip = uint64(value)
+}
+
+// Stack returns the current stack pointer.
+func (c *context64) Stack() uintptr {
+	return uintptr(c.Regs.Rsp)
+}
+
+// SetStack sets the current stack pointer.
+func (c *context64) SetStack(value uintptr) {
+	c.Regs.Rsp = uint64(value)
+}
+
+// SetRSEQInterruptedIP implements Context.SetRSEQInterruptedIP.
+func (c *context64) SetRSEQInterruptedIP(value uintptr) {
+	c.Regs.R10 = uint64(value)
+}
+
+// Native returns the native type for the given val.
+func (c *context64) Native(val uintptr) interface{} {
+	v := uint64(val)
+	return &v
+}
+
+// Value returns the generic val for the given native type.
+func (c *context64) Value(val interface{}) uintptr {
+	return uintptr(*val.(*uint64))
+}
+
+// Width returns the byte width of this architecture.
+func (c *context64) Width() uint {
+	return 8
+}
+
+// FeatureSet returns the FeatureSet in use.
+func (c *context64) FeatureSet() *cpuid.FeatureSet {
+	return c.State.FeatureSet
+}
+
+// mmapRand returns a random adjustment for randomizing an mmap layout.
+func mmapRand(max uint64) usermem.Addr {
+	return usermem.Addr(rand.Int63n(int64(max))).RoundDown()
+}
+
+// NewMmapLayout implements Context.NewMmapLayout consistently with Linux.
+func (c *context64) NewMmapLayout(min, max usermem.Addr, r *limits.LimitSet) (MmapLayout, error) {
+	min, ok := min.RoundUp()
+	if !ok {
+		return MmapLayout{}, syscall.EINVAL
+	}
+	if max > maxAddr64 {
+		max = maxAddr64
+	}
+	max = max.RoundDown()
+
+	if min > max {
+		return MmapLayout{}, syscall.EINVAL
+	}
+
+	stackSize := r.Get(limits.Stack)
+
+	// MAX_GAP in Linux.
+	maxGap := (max / 6) * 5
+	gap := usermem.Addr(stackSize.Cur)
+	if gap < minGap64 {
+		gap = minGap64
+	}
+	if gap > maxGap {
+		gap = maxGap
+	}
+	defaultDir := MmapTopDown
+	if stackSize.Cur == limits.Infinity {
+		defaultDir = MmapBottomUp
+	}
+
+	topDownMin := max - gap - maxMmapRand64
+	maxRand := usermem.Addr(maxMmapRand64)
+	if topDownMin < preferredTopDownBaseMin {
+		// Try to keep TopDownBase above preferredTopDownBaseMin by
+		// shrinking maxRand.
+		maxAdjust := maxRand - minMmapRand64
+		needAdjust := preferredTopDownBaseMin - topDownMin
+		if needAdjust <= maxAdjust {
+			maxRand -= needAdjust
+		}
+	}
+
+	rnd := mmapRand(uint64(maxRand))
+	l := MmapLayout{
+		MinAddr: min,
+		MaxAddr: max,
+		// TASK_UNMAPPED_BASE in Linux.
+		BottomUpBase:     (max/3 + rnd).RoundDown(),
+		TopDownBase:      (max - gap - rnd).RoundDown(),
+		DefaultDirection: defaultDir,
+		// We may have reduced the maximum randomization to keep
+		// TopDownBase above preferredTopDownBaseMin while maintaining
+		// our stack gap. Stack allocations must use that max
+		// randomization to avoiding eating into the gap.
+		MaxStackRand: uint64(maxRand),
+	}
+
+	// Final sanity check on the layout.
+	if !l.Valid() {
+		panic(fmt.Sprintf("Invalid MmapLayout: %+v", l))
+	}
+
+	return l, nil
+}
+
+// PIELoadAddress implements Context.PIELoadAddress.
+func (c *context64) PIELoadAddress(l MmapLayout) usermem.Addr {
+	base := preferredPIELoadAddr
+	max, ok := base.AddLength(maxMmapRand64)
+	if !ok {
+		panic(fmt.Sprintf("preferredPIELoadAddr %#x too large", base))
+	}
+
+	if max > l.MaxAddr {
+		// preferredPIELoadAddr won't fit; fall back to the standard
+		// Linux behavior of 2/3 of TopDownBase. TSAN won't like this.
+		//
+		// Don't bother trying to shrink the randomization for now.
+		base = l.TopDownBase / 3 * 2
+	}
+
+	return base + mmapRand(maxMmapRand64)
+}
+
+// userStructSize is the size in bytes of Linux's struct user on amd64.
+const userStructSize = 928
+
+// PtracePeekUser implements Context.PtracePeekUser.
+func (c *context64) PtracePeekUser(addr uintptr) (interface{}, error) {
+	if addr&7 != 0 || addr >= userStructSize {
+		return nil, syscall.EIO
+	}
+	// PTRACE_PEEKUSER and PTRACE_POKEUSER are only effective on regs and
+	// u_debugreg, returning 0 or silently no-oping for other fields
+	// respectively.
+	if addr < uintptr(ptraceRegsSize) {
+		buf := binary.Marshal(nil, usermem.ByteOrder, c.ptraceGetRegs())
+		return c.Native(uintptr(usermem.ByteOrder.Uint64(buf[addr:]))), nil
+	}
+	// TODO: debug registers
+	return c.Native(0), nil
+}
+
+// PtracePokeUser implements Context.PtracePokeUser.
+func (c *context64) PtracePokeUser(addr, data uintptr) error {
+	if addr&7 != 0 || addr >= userStructSize {
+		return syscall.EIO
+	}
+	if addr < uintptr(ptraceRegsSize) {
+		buf := binary.Marshal(nil, usermem.ByteOrder, c.ptraceGetRegs())
+		usermem.ByteOrder.PutUint64(buf[addr:], uint64(data))
+		_, err := c.PtraceSetRegs(bytes.NewBuffer(buf))
+		return err
+	}
+	// TODO: debug registers
+	return nil
+}
diff --git a/pkg/sentry/arch/arch_amd64.s b/pkg/sentry/arch/arch_amd64.s
new file mode 100644
index 000000000..10d621b6d
--- /dev/null
+++ b/pkg/sentry/arch/arch_amd64.s
@@ -0,0 +1,135 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// MXCSR_DEFAULT is the reset value of MXCSR (Intel SDM Vol. 2, Ch. 3.2
+// "LDMXCSR")
+#define MXCSR_DEFAULT	0x1f80
+
+// MXCSR_OFFSET is the offset in bytes of the MXCSR field from the start of the
+// FXSAVE/XSAVE area. (Intel SDM Vol. 1, Table 10-2 "Format of an FXSAVE Area")
+#define MXCSR_OFFSET	24
+
+// initX86FPState initializes floating point state.
+//
+// func initX86FPState(data *FloatingPointData, useXsave bool)
+//
+// We need to clear out and initialize an empty fp state area since the sentry
+// may have left sensitive information in the floating point registers.
+//
+// Preconditions: data is zeroed
+TEXT ·initX86FPState(SB), $24-16
+	// Save MXCSR (callee-save)
+	STMXCSR	mxcsr-8(SP)
+
+	// Save x87 CW (callee-save)
+	FSTCW	cw-16(SP)
+
+	MOVQ	fpState+0(FP), DI
+
+	// Do we use xsave?
+	MOVBQZX	useXsave+8(FP), AX
+	TESTQ	AX, AX
+	JZ	no_xsave
+
+	// Use XRSTOR to clear all FP state to an initial state.
+	//
+	// The fpState XSAVE area is zeroed on function entry, meaning
+	// XSTATE_BV is zero.
+	//
+	// "If RFBM[i] = 1 and bit i is clear in the XSTATE_BV field in the
+	// XSAVE header, XRSTOR initializes state component i."
+	//
+	// Initialization is defined in SDM Vol 1, Chapter 13.3. It puts all
+	// the registers in a reasonable initial state, except MXCSR:
+	//
+	// "The MXCSR register is part of state component 1, SSE state (see
+	// Section 13.5.2). However, the standard form of XRSTOR loads the
+	// MXCSR register from memory whenever the RFBM[1] (SSE) or RFBM[2]
+	// (AVX) is set, regardless of the values of XSTATE_BV[1] and
+	// XSTATE_BV[2]."
+
+	// Set MXCSR to the default value.
+	MOVL	$MXCSR_DEFAULT, MXCSR_OFFSET(DI)
+
+	// Initialize registers with XRSTOR.
+	MOVL	$0xffffffff, AX
+	MOVL	$0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x2f // XRSTOR64 0(DI)
+
+	// Now that all the state has been reset, write it back out to the
+	// XSAVE area.
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x27 // XSAVE64 0(DI)
+
+	JMP	out
+
+no_xsave:
+	// Clear out existing X values.
+	PXOR	X0, X0
+	MOVO	X0, X1
+	MOVO	X0, X2
+	MOVO	X0, X3
+	MOVO	X0, X4
+	MOVO	X0, X5
+	MOVO	X0, X6
+	MOVO	X0, X7
+	MOVO	X0, X8
+	MOVO	X0, X9
+	MOVO	X0, X10
+	MOVO	X0, X11
+	MOVO	X0, X12
+	MOVO	X0, X13
+	MOVO	X0, X14
+	MOVO	X0, X15
+
+	// Zero out %rax and store into MMX registers. MMX registers are
+	// an alias of 8x64 bits of the 8x80 bits used for the original
+	// x87 registers. Storing zero into them will reset the FPU registers
+	// to bits [63:0] = 0, [79:64] = 1. But the contents aren't too
+	// important, just the fact that we have reset them to a known value.
+	XORQ	AX, AX
+	MOVQ	AX, M0
+	MOVQ	AX, M1
+	MOVQ	AX, M2
+	MOVQ	AX, M3
+	MOVQ	AX, M4
+	MOVQ	AX, M5
+	MOVQ	AX, M6
+	MOVQ	AX, M7
+
+	// The Go assembler doesn't support FNINIT, so we use BYTE.
+	// This will:
+	//  - Reset FPU control word to 0x037f
+	//  - Clear FPU status word
+	//  - Reset FPU tag word to 0xffff
+	//  - Clear FPU data pointer
+	//  - Clear FPU instruction pointer
+	BYTE $0xDB; BYTE $0xE3; // FNINIT
+
+	// Reset MXCSR.
+	MOVL	$MXCSR_DEFAULT, tmpmxcsr-24(SP)
+	LDMXCSR	tmpmxcsr-24(SP)
+
+	// Save the floating point state with fxsave.
+	FXSAVE64	0(DI)
+
+out:
+	// Restore MXCSR.
+	LDMXCSR	mxcsr-8(SP)
+
+	// Restore x87 CW.
+	FLDCW	cw-16(SP)
+
+	RET
diff --git a/pkg/sentry/arch/arch_state_x86.go b/pkg/sentry/arch/arch_state_x86.go
new file mode 100644
index 000000000..cb38d098a
--- /dev/null
+++ b/pkg/sentry/arch/arch_state_x86.go
@@ -0,0 +1,97 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+import (
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+)
+
+// warnOnce is used to warn about truncated state only once.
+var warnOnce sync.Once
+
+// afterLoad is invoked by stateify.
+func (s *State) afterLoad() {
+	old := s.x86FPState
+
+	// Recreate the slice. This is done to ensure that it is aligned
+	// appropriately in memory, and large enough to accommodate any new
+	// state that may be saved by the new CPU. Even if extraneous new state
+	// is saved, the state we care about is guaranteed to be a subset of
+	// new state. Later optimizations can use less space when using a
+	// smaller state component bitmap. Intel SDM section 13 has more info.
+	s.x86FPState = newX86FPState()
+
+	// x86FPState always contains all the FP state supported by the host.
+	// We may have come from a newer machine that supports additional state
+	// which we cannot restore.
+	//
+	// The x86 FP state areas are backwards compatible, so we can simply
+	// truncate the additional floating point state. Applications should
+	// not depend on the truncated state because it should relate only to
+	// features that were not exposed in the app FeatureSet.
+	if len(s.x86FPState) < len(old) {
+		warnOnce.Do(func() {
+			// This will occur on every instance of state, don't
+			// bother warning more than once.
+			log.Infof("dropping %d bytes of floating point state; the application should not depend on this state", len(old)-len(s.x86FPState))
+		})
+	}
+
+	// Copy to the new, aligned location.
+	copy(s.x86FPState, old)
+}
+
+type syscallPtraceRegs struct {
+	R15      uint64
+	R14      uint64
+	R13      uint64
+	R12      uint64
+	Rbp      uint64
+	Rbx      uint64
+	R11      uint64
+	R10      uint64
+	R9       uint64
+	R8       uint64
+	Rax      uint64
+	Rcx      uint64
+	Rdx      uint64
+	Rsi      uint64
+	Rdi      uint64
+	Orig_rax uint64
+	Rip      uint64
+	Cs       uint64
+	Eflags   uint64
+	Rsp      uint64
+	Ss       uint64
+	Fs_base  uint64
+	Gs_base  uint64
+	Ds       uint64
+	Es       uint64
+	Fs       uint64
+	Gs       uint64
+}
+
+// saveRegs is invoked by stateify.
+func (s *State) saveRegs() syscallPtraceRegs {
+	return syscallPtraceRegs(s.Regs)
+}
+
+// loadRegs is invoked by stateify.
+func (s *State) loadRegs(r syscallPtraceRegs) {
+	s.Regs = syscall.PtraceRegs(r)
+}
diff --git a/pkg/sentry/arch/arch_x86.go b/pkg/sentry/arch/arch_x86.go
new file mode 100644
index 000000000..5cc4f8377
--- /dev/null
+++ b/pkg/sentry/arch/arch_x86.go
@@ -0,0 +1,613 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64 i386
+
+package arch
+
+import (
+	"fmt"
+	"io"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	rpb "gvisor.googlesource.com/gvisor/pkg/sentry/arch/registers_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// System-related constants for x86.
+const (
+	// SyscallWidth is the width of syscall, sysenter, and int 80 insturctions.
+	SyscallWidth = 2
+)
+
+// EFLAGS register bits.
+const (
+	// eflagsCF is the mask for the carry flag.
+	eflagsCF = uint64(1) << 0
+	// eflagsPF is the mask for the parity flag.
+	eflagsPF = uint64(1) << 2
+	// eflagsAF is the mask for the auxiliary carry flag.
+	eflagsAF = uint64(1) << 4
+	// eflagsZF is the mask for the zero flag.
+	eflagsZF = uint64(1) << 6
+	// eflagsSF is the mask for the sign flag.
+	eflagsSF = uint64(1) << 7
+	// eflagsTF is the mask for the trap flag.
+	eflagsTF = uint64(1) << 8
+	// eflagsIF is the mask for the interrupt flag.
+	eflagsIF = uint64(1) << 9
+	// eflagsDF is the mask for the direction flag.
+	eflagsDF = uint64(1) << 10
+	// eflagsOF is the mask for the overflow flag.
+	eflagsOF = uint64(1) << 11
+	// eflagsIOPL is the mask for the I/O privilege level.
+	eflagsIOPL = uint64(3) << 12
+	// eflagsNT is the mask for the nested task bit.
+	eflagsNT = uint64(1) << 14
+	// eflagsRF is the mask for the resume flag.
+	eflagsRF = uint64(1) << 16
+	// eflagsVM is the mask for the virtual mode bit.
+	eflagsVM = uint64(1) << 17
+	// eflagsAC is the mask for the alignment check / access control bit.
+	eflagsAC = uint64(1) << 18
+	// eflagsVIF is the mask for the virtual interrupt flag.
+	eflagsVIF = uint64(1) << 19
+	// eflagsVIP is the mask for the virtual interrupt pending bit.
+	eflagsVIP = uint64(1) << 20
+	// eflagsID is the mask for the CPUID detection bit.
+	eflagsID = uint64(1) << 21
+
+	// eflagsPtraceMutable is the mask for the set of EFLAGS that may be
+	// changed by ptrace(PTRACE_SETREGS). eflagsPtraceMutable is analogous to
+	// Linux's FLAG_MASK.
+	eflagsPtraceMutable = eflagsCF | eflagsPF | eflagsAF | eflagsZF | eflagsSF | eflagsTF | eflagsDF | eflagsOF | eflagsRF | eflagsAC | eflagsNT
+
+	// eflagsRestorable is the mask for the set of EFLAGS that may be changed by
+	// SignalReturn. eflagsRestorable is analogous to Linux's FIX_EFLAGS.
+	eflagsRestorable = eflagsAC | eflagsOF | eflagsDF | eflagsTF | eflagsSF | eflagsZF | eflagsAF | eflagsPF | eflagsCF | eflagsRF
+)
+
+// Segment selectors. See arch/x86/include/asm/segment.h.
+const (
+	userCS   = 0x33 // guest ring 3 code selector
+	user32CS = 0x23 // guest ring 3 32 bit code selector
+	userDS   = 0x2b // guest ring 3 data selector
+
+	_FS_TLS_SEL = 0x63 // Linux FS thread-local storage selector
+	_GS_TLS_SEL = 0x6b // Linux GS thread-local storage selector
+)
+
+var (
+	// TrapInstruction is the x86 trap instruction.
+	TrapInstruction = [1]byte{0xcc}
+
+	// CPUIDInstruction is the x86 CPUID instruction.
+	CPUIDInstruction = [2]byte{0xf, 0xa2}
+
+	// X86TrapFlag is an exported const for use by other packages.
+	X86TrapFlag uint64 = (1 << 8)
+)
+
+// x86FPState is x86 floating point state.
+type x86FPState []byte
+
+// initX86FPState (defined in asm files) sets up initial state.
+func initX86FPState(data *FloatingPointData, useXsave bool)
+
+func newX86FPStateSlice() []byte {
+	size, align := cpuid.HostFeatureSet().ExtendedStateSize()
+	capacity := size
+	// Always use at least 4096 bytes.
+	if capacity < 4096 {
+		capacity = 4096
+	}
+	return alignedBytes(capacity, align)[:size]
+}
+
+// newX86FPState returns an initialized floating point state.
+//
+// The returned state is large enough to store all floating point state
+// supported by host, even if the app won't use much of it due to a restricted
+// FeatureSet. Since they may still be able to see state not advertised by
+// CPUID we must ensure it does not contain any sentry state.
+func newX86FPState() x86FPState {
+	f := x86FPState(newX86FPStateSlice())
+	initX86FPState(f.FloatingPointData(), cpuid.HostFeatureSet().UseXsave())
+	return f
+}
+
+// fork creates and returns an identical copy of the x86 floating point state.
+func (f x86FPState) fork() x86FPState {
+	n := x86FPState(newX86FPStateSlice())
+	copy(n, f)
+	return n
+}
+
+// FloatingPointData returns the raw data pointer.
+func (f x86FPState) FloatingPointData() *FloatingPointData {
+	return (*FloatingPointData)(&f[0])
+}
+
+// NewFloatingPointData returns a new floating point data blob.
+//
+// This is primarily for use in tests.
+func NewFloatingPointData() *FloatingPointData {
+	return (*FloatingPointData)(&(newX86FPState()[0]))
+}
+
+// State contains the common architecture bits for X86 (the build tag of this
+// file ensures it's only built on x86).
+type State struct {
+	// The system registers.
+	Regs syscall.PtraceRegs `state:".(syscallPtraceRegs)"`
+
+	// Our floating point state.
+	x86FPState `state:"wait"`
+
+	// FeatureSet is a pointer to the currently active feature set.
+	FeatureSet *cpuid.FeatureSet
+}
+
+// Proto returns a protobuf representation of the system registers in State.
+func (s State) Proto() *rpb.Registers {
+	regs := &rpb.AMD64Registers{
+		Rax:     s.Regs.Rax,
+		Rbx:     s.Regs.Rbx,
+		Rcx:     s.Regs.Rcx,
+		Rdx:     s.Regs.Rdx,
+		Rsi:     s.Regs.Rsi,
+		Rdi:     s.Regs.Rdi,
+		Rsp:     s.Regs.Rsp,
+		Rbp:     s.Regs.Rbp,
+		R8:      s.Regs.R8,
+		R9:      s.Regs.R9,
+		R10:     s.Regs.R10,
+		R11:     s.Regs.R11,
+		R12:     s.Regs.R12,
+		R13:     s.Regs.R13,
+		R14:     s.Regs.R14,
+		R15:     s.Regs.R15,
+		Rip:     s.Regs.Rip,
+		Rflags:  s.Regs.Eflags,
+		OrigRax: s.Regs.Orig_rax,
+		Cs:      s.Regs.Cs,
+		Ds:      s.Regs.Ds,
+		Es:      s.Regs.Es,
+		Fs:      s.Regs.Fs,
+		Gs:      s.Regs.Gs,
+		Ss:      s.Regs.Ss,
+		FsBase:  s.Regs.Fs_base,
+		GsBase:  s.Regs.Gs_base,
+	}
+	return &rpb.Registers{Arch: &rpb.Registers_Amd64{Amd64: regs}}
+}
+
+// Fork creates and returns an identical copy of the state.
+func (s *State) Fork() State {
+	return State{
+		Regs:       s.Regs,
+		x86FPState: s.x86FPState.fork(),
+		FeatureSet: s.FeatureSet,
+	}
+}
+
+// StateData implements Context.StateData.
+func (s *State) StateData() *State {
+	return s
+}
+
+// CPUIDEmulate emulates a cpuid instruction.
+func (s *State) CPUIDEmulate(l log.Logger) {
+	argax := uint32(s.Regs.Rax)
+	argcx := uint32(s.Regs.Rcx)
+	ax, bx, cx, dx := s.FeatureSet.EmulateID(argax, argcx)
+	s.Regs.Rax = uint64(ax)
+	s.Regs.Rbx = uint64(bx)
+	s.Regs.Rcx = uint64(cx)
+	s.Regs.Rdx = uint64(dx)
+	l.Debugf("CPUID(%x,%x): %x %x %x %x", argax, argcx, ax, bx, cx, dx)
+}
+
+// SingleStep implements Context.SingleStep.
+func (s *State) SingleStep() bool {
+	return s.Regs.Eflags&X86TrapFlag != 0
+}
+
+// SetSingleStep enables single stepping.
+func (s *State) SetSingleStep() {
+	// Set the trap flag.
+	s.Regs.Eflags |= X86TrapFlag
+}
+
+// ClearSingleStep enables single stepping.
+func (s *State) ClearSingleStep() {
+	// Clear the trap flag.
+	s.Regs.Eflags &= ^X86TrapFlag
+}
+
+// RegisterMap returns a map of all registers.
+func (s *State) RegisterMap() (map[string]uintptr, error) {
+	return map[string]uintptr{
+		"R15":      uintptr(s.Regs.R15),
+		"R14":      uintptr(s.Regs.R14),
+		"R13":      uintptr(s.Regs.R13),
+		"R12":      uintptr(s.Regs.R12),
+		"Rbp":      uintptr(s.Regs.Rbp),
+		"Rbx":      uintptr(s.Regs.Rbx),
+		"R11":      uintptr(s.Regs.R11),
+		"R10":      uintptr(s.Regs.R10),
+		"R9":       uintptr(s.Regs.R9),
+		"R8":       uintptr(s.Regs.R8),
+		"Rax":      uintptr(s.Regs.Rax),
+		"Rcx":      uintptr(s.Regs.Rcx),
+		"Rdx":      uintptr(s.Regs.Rdx),
+		"Rsi":      uintptr(s.Regs.Rsi),
+		"Rdi":      uintptr(s.Regs.Rdi),
+		"Orig_rax": uintptr(s.Regs.Orig_rax),
+		"Rip":      uintptr(s.Regs.Rip),
+		"Cs":       uintptr(s.Regs.Cs),
+		"Eflags":   uintptr(s.Regs.Eflags),
+		"Rsp":      uintptr(s.Regs.Rsp),
+		"Ss":       uintptr(s.Regs.Ss),
+		"Fs_base":  uintptr(s.Regs.Fs_base),
+		"Gs_base":  uintptr(s.Regs.Gs_base),
+		"Ds":       uintptr(s.Regs.Ds),
+		"Es":       uintptr(s.Regs.Es),
+		"Fs":       uintptr(s.Regs.Fs),
+		"Gs":       uintptr(s.Regs.Gs),
+	}, nil
+}
+
+// PtraceGetRegs implements Context.PtraceGetRegs.
+func (s *State) PtraceGetRegs(dst io.Writer) (int, error) {
+	return dst.Write(binary.Marshal(nil, usermem.ByteOrder, s.ptraceGetRegs()))
+}
+
+func (s *State) ptraceGetRegs() syscall.PtraceRegs {
+	regs := s.Regs
+	// These may not be initialized.
+	if regs.Cs == 0 || regs.Ss == 0 || regs.Eflags == 0 {
+		regs.Eflags = eflagsIF
+		regs.Cs = userCS
+		regs.Ss = userDS
+	}
+	// As an optimization, Linux <4.7 implements 32-bit fs_base/gs_base
+	// addresses using reserved descriptors in the GDT instead of the MSRs,
+	// with selector values FS_TLS_SEL and GS_TLS_SEL respectively. These
+	// values are actually visible in struct user_regs_struct::fs/gs;
+	// arch/x86/kernel/ptrace.c:getreg() doesn't attempt to sanitize struct
+	// thread_struct::fsindex/gsindex.
+	//
+	// We always use fs == gs == 0 when fs_base/gs_base is in use, for
+	// simplicity.
+	//
+	// Luckily, Linux <4.7 silently ignores setting fs/gs to 0 via
+	// arch/x86/kernel/ptrace.c:set_segment_reg() when fs_base/gs_base is a
+	// 32-bit value and fsindex/gsindex indicates that this optimization is
+	// in use, as well as the reverse case of setting fs/gs to
+	// FS/GS_TLS_SEL when fs_base/gs_base is a 64-bit value. (We do the
+	// same in PtraceSetRegs.)
+	//
+	// TODO: Remove this fixup since newer Linux doesn't have
+	// this behavior anymore.
+	if regs.Fs == 0 && regs.Fs_base <= 0xffffffff {
+		regs.Fs = _FS_TLS_SEL
+	}
+	if regs.Gs == 0 && regs.Gs_base <= 0xffffffff {
+		regs.Gs = _GS_TLS_SEL
+	}
+	return regs
+}
+
+var ptraceRegsSize = int(binary.Size(syscall.PtraceRegs{}))
+
+// PtraceSetRegs implements Context.PtraceSetRegs.
+func (s *State) PtraceSetRegs(src io.Reader) (int, error) {
+	var regs syscall.PtraceRegs
+	buf := make([]byte, ptraceRegsSize)
+	if _, err := io.ReadFull(src, buf); err != nil {
+		return 0, err
+	}
+	binary.Unmarshal(buf, usermem.ByteOrder, &regs)
+	// Truncate segment registers to 16 bits.
+	regs.Cs = uint64(uint16(regs.Cs))
+	regs.Ds = uint64(uint16(regs.Ds))
+	regs.Es = uint64(uint16(regs.Es))
+	regs.Fs = uint64(uint16(regs.Fs))
+	regs.Gs = uint64(uint16(regs.Gs))
+	regs.Ss = uint64(uint16(regs.Ss))
+	// In Linux this validation is via arch/x86/kernel/ptrace.c:putreg().
+	if !isUserSegmentSelector(regs.Cs) {
+		return 0, syscall.EIO
+	}
+	if regs.Ds != 0 && !isUserSegmentSelector(regs.Ds) {
+		return 0, syscall.EIO
+	}
+	if regs.Es != 0 && !isUserSegmentSelector(regs.Es) {
+		return 0, syscall.EIO
+	}
+	if regs.Fs != 0 && !isUserSegmentSelector(regs.Fs) {
+		return 0, syscall.EIO
+	}
+	if regs.Gs != 0 && !isUserSegmentSelector(regs.Gs) {
+		return 0, syscall.EIO
+	}
+	if !isUserSegmentSelector(regs.Ss) {
+		return 0, syscall.EIO
+	}
+	if regs.Fs_base >= uint64(maxAddr64) {
+		return 0, syscall.EIO
+	}
+	if regs.Gs_base >= uint64(maxAddr64) {
+		return 0, syscall.EIO
+	}
+	// CS and SS are validated, but changes to them are otherwise silently
+	// ignored on amd64.
+	regs.Cs = s.Regs.Cs
+	regs.Ss = s.Regs.Ss
+	// fs_base/gs_base changes reset fs/gs via do_arch_prctl() on Linux.
+	if regs.Fs_base != s.Regs.Fs_base {
+		regs.Fs = 0
+	}
+	if regs.Gs_base != s.Regs.Gs_base {
+		regs.Gs = 0
+	}
+	// Ignore "stale" TLS segment selectors for FS and GS. See comment in
+	// ptraceGetRegs.
+	if regs.Fs == _FS_TLS_SEL && regs.Fs_base != 0 {
+		regs.Fs = 0
+	}
+	if regs.Gs == _GS_TLS_SEL && regs.Gs_base != 0 {
+		regs.Gs = 0
+	}
+	regs.Eflags = (s.Regs.Eflags &^ eflagsPtraceMutable) | (regs.Eflags & eflagsPtraceMutable)
+	s.Regs = regs
+	return ptraceRegsSize, nil
+}
+
+// isUserSegmentSelector returns true if the given segment selector specifies a
+// privilege level of 3 (USER_RPL).
+func isUserSegmentSelector(reg uint64) bool {
+	return reg&3 == 3
+}
+
+// ptraceFPRegsSize is the size in bytes of Linux's user_i387_struct, the type
+// manipulated by PTRACE_GETFPREGS and PTRACE_SETFPREGS on x86. Equivalently,
+// ptraceFPRegsSize is the size in bytes of the x86 FXSAVE area.
+const ptraceFPRegsSize = 512
+
+// PtraceGetFPRegs implements Context.PtraceGetFPRegs.
+func (s *State) PtraceGetFPRegs(dst io.Writer) (int, error) {
+	return dst.Write(s.x86FPState[:ptraceFPRegsSize])
+}
+
+// PtraceSetFPRegs implements Context.PtraceSetFPRegs.
+func (s *State) PtraceSetFPRegs(src io.Reader) (int, error) {
+	var f [ptraceFPRegsSize]byte
+	n, err := io.ReadFull(src, f[:])
+	if err != nil {
+		return 0, err
+	}
+	// Force reserved bits in MXCSR to 0. This is consistent with Linux.
+	sanitizeMXCSR(x86FPState(f[:]))
+	// N.B. this only copies the beginning of the FP state, which
+	// corresponds to the FXSAVE area.
+	copy(s.x86FPState, f[:])
+	return n, nil
+}
+
+const (
+	// mxcsrOffset is the offset in bytes of the MXCSR field from the start of
+	// the FXSAVE area. (Intel SDM Vol. 1, Table 10-2 "Format of an FXSAVE
+	// Area")
+	mxcsrOffset = 24
+
+	// mxcsrMaskOffset is the offset in bytes of the MXCSR_MASK field from the
+	// start of the FXSAVE area.
+	mxcsrMaskOffset = 28
+)
+
+var (
+	mxcsrMask     uint32
+	initMXCSRMask sync.Once
+)
+
+// sanitizeMXCSR coerces reserved bits in the MXCSR field of f to 0. ("FXRSTOR
+// generates a general-protection fault (#GP) in response to an attempt to set
+// any of the reserved bits of the MXCSR register." - Intel SDM Vol. 1, Section
+// 10.5.1.2 "SSE State")
+func sanitizeMXCSR(f x86FPState) {
+	mxcsr := usermem.ByteOrder.Uint32(f[mxcsrOffset:])
+	initMXCSRMask.Do(func() {
+		temp := x86FPState(alignedBytes(uint(ptraceFPRegsSize), 16))
+		initX86FPState(temp.FloatingPointData(), false /* useXsave */)
+		mxcsrMask = usermem.ByteOrder.Uint32(temp[mxcsrMaskOffset:])
+		if mxcsrMask == 0 {
+			// "If the value of the MXCSR_MASK field is 00000000H, then the
+			// MXCSR_MASK value is the default value of 0000FFBFH." - Intel SDM
+			// Vol. 1, Section 11.6.6 "Guidelines for Writing to the MXCSR
+			// Register"
+			mxcsrMask = 0xffbf
+		}
+	})
+	mxcsr &= mxcsrMask
+	usermem.ByteOrder.PutUint32(f[mxcsrOffset:], mxcsr)
+}
+
+const (
+	// minXstateBytes is the minimum size in bytes of an x86 XSAVE area, equal
+	// to the size of the XSAVE legacy area (512 bytes) plus the size of the
+	// XSAVE header (64 bytes). Equivalently, minXstateBytes is GDB's
+	// X86_XSTATE_SSE_SIZE.
+	minXstateBytes = 512 + 64
+
+	// userXstateXCR0Offset is the offset in bytes of the USER_XSTATE_XCR0_WORD
+	// field in Linux's struct user_xstateregs, which is the type manipulated
+	// by ptrace(PTRACE_GET/SETREGSET, NT_X86_XSTATE). Equivalently,
+	// userXstateXCR0Offset is GDB's I386_LINUX_XSAVE_XCR0_OFFSET.
+	userXstateXCR0Offset = 464
+
+	// xstateBVOffset is the offset in bytes of the XSTATE_BV field in an x86
+	// XSAVE area.
+	xstateBVOffset = 512
+
+	// xsaveHeaderZeroedOffset and xsaveHeaderZeroedBytes indicate parts of the
+	// XSAVE header that we coerce to zero: "Bytes 15:8 of the XSAVE header is
+	// a state-component bitmap called XCOMP_BV. ... Bytes 63:16 of the XSAVE
+	// header are reserved." - Intel SDM Vol. 1, Section 13.4.2 "XSAVE Header".
+	// Linux ignores XCOMP_BV, but it's able to recover from XRSTOR #GP
+	// exceptions resulting from invalid values; we aren't. Linux also never
+	// uses the compacted format when doing XSAVE and doesn't even define the
+	// compaction extensions to XSAVE as a CPU feature, so for simplicity we
+	// assume no one is using them.
+	xsaveHeaderZeroedOffset = 512 + 8
+	xsaveHeaderZeroedBytes  = 64 - 8
+)
+
+func (s *State) ptraceGetXstateRegs(dst io.Writer, maxlen int) (int, error) {
+	// N.B. s.x86FPState may contain more state than the application
+	// expects. We only copy the subset that would be in their XSAVE area.
+	ess, _ := s.FeatureSet.ExtendedStateSize()
+	f := make([]byte, ess)
+	copy(f, s.x86FPState)
+	// "The XSAVE feature set does not use bytes 511:416; bytes 463:416 are
+	// reserved." - Intel SDM Vol 1., Section 13.4.1 "Legacy Region of an XSAVE
+	// Area". Linux uses the first 8 bytes of this area to store the OS XSTATE
+	// mask. GDB relies on this: see
+	// gdb/x86-linux-nat.c:x86_linux_read_description().
+	usermem.ByteOrder.PutUint64(f[userXstateXCR0Offset:], s.FeatureSet.ValidXCR0Mask())
+	if len(f) > maxlen {
+		f = f[:maxlen]
+	}
+	return dst.Write(f)
+}
+
+func (s *State) ptraceSetXstateRegs(src io.Reader, maxlen int) (int, error) {
+	// Allow users to pass an xstate register set smaller than ours (they can
+	// mask bits out of XSTATE_BV), as long as it's at least minXstateBytes.
+	// Also allow users to pass a register set larger than ours; anything after
+	// their ExtendedStateSize will be ignored. (I think Linux technically
+	// permits setting a register set smaller than minXstateBytes, but it has
+	// the same silent truncation behavior in kernel/ptrace.c:ptrace_regset().)
+	if maxlen < minXstateBytes {
+		return 0, syscall.EFAULT
+	}
+	ess, _ := s.FeatureSet.ExtendedStateSize()
+	if maxlen > int(ess) {
+		maxlen = int(ess)
+	}
+	f := make([]byte, maxlen)
+	if _, err := io.ReadFull(src, f); err != nil {
+		return 0, err
+	}
+	// Force reserved bits in MXCSR to 0. This is consistent with Linux.
+	sanitizeMXCSR(x86FPState(f))
+	// Users can't enable *more* XCR0 bits than what we, and the CPU, support.
+	xstateBV := usermem.ByteOrder.Uint64(f[xstateBVOffset:])
+	xstateBV &= s.FeatureSet.ValidXCR0Mask()
+	usermem.ByteOrder.PutUint64(f[xstateBVOffset:], xstateBV)
+	// Force XCOMP_BV and reserved bytes in the XSAVE header to 0.
+	reserved := f[xsaveHeaderZeroedOffset : xsaveHeaderZeroedOffset+xsaveHeaderZeroedBytes]
+	for i := range reserved {
+		reserved[i] = 0
+	}
+	return copy(s.x86FPState, f), nil
+}
+
+// Register sets defined in include/uapi/linux/elf.h.
+const (
+	_NT_PRSTATUS   = 1
+	_NT_PRFPREG    = 2
+	_NT_X86_XSTATE = 0x202
+)
+
+// PtraceGetRegSet implements Context.PtraceGetRegSet.
+func (s *State) PtraceGetRegSet(regset uintptr, dst io.Writer, maxlen int) (int, error) {
+	switch regset {
+	case _NT_PRSTATUS:
+		if maxlen < ptraceRegsSize {
+			return 0, syserror.EFAULT
+		}
+		return s.PtraceGetRegs(dst)
+	case _NT_PRFPREG:
+		if maxlen < ptraceFPRegsSize {
+			return 0, syserror.EFAULT
+		}
+		return s.PtraceGetFPRegs(dst)
+	case _NT_X86_XSTATE:
+		return s.ptraceGetXstateRegs(dst, maxlen)
+	default:
+		return 0, syserror.EINVAL
+	}
+}
+
+// PtraceSetRegSet implements Context.PtraceSetRegSet.
+func (s *State) PtraceSetRegSet(regset uintptr, src io.Reader, maxlen int) (int, error) {
+	switch regset {
+	case _NT_PRSTATUS:
+		if maxlen < ptraceRegsSize {
+			return 0, syserror.EFAULT
+		}
+		return s.PtraceSetRegs(src)
+	case _NT_PRFPREG:
+		if maxlen < ptraceFPRegsSize {
+			return 0, syserror.EFAULT
+		}
+		return s.PtraceSetFPRegs(src)
+	case _NT_X86_XSTATE:
+		return s.ptraceSetXstateRegs(src, maxlen)
+	default:
+		return 0, syserror.EINVAL
+	}
+}
+
+// FullRestore indicates whether a full restore is required.
+func (s *State) FullRestore() bool {
+	// A fast system call return is possible only if
+	//
+	// * RCX matches the instruction pointer.
+	// * R11 matches our flags value.
+	// * Usermode does not expect to set either the resume flag or the
+	//   virtual mode flags (unlikely.)
+	// * CS and SS are set to the standard selectors.
+	//
+	// That is, SYSRET results in the correct final state.
+	fastRestore := s.Regs.Rcx == s.Regs.Rip &&
+		s.Regs.Eflags == s.Regs.R11 &&
+		(s.Regs.Eflags&eflagsRF == 0) &&
+		(s.Regs.Eflags&eflagsVM == 0) &&
+		s.Regs.Cs == userCS &&
+		s.Regs.Ss == userDS
+	return !fastRestore
+}
+
+// New returns a new architecture context.
+func New(arch Arch, fs *cpuid.FeatureSet) Context {
+	switch arch {
+	case AMD64:
+		return &context64{
+			State{
+				x86FPState: newX86FPState(),
+				FeatureSet: fs,
+			},
+			[]x86FPState(nil),
+		}
+	}
+	panic(fmt.Sprintf("unknown architecture %v", arch))
+}
diff --git a/pkg/sentry/arch/auxv.go b/pkg/sentry/arch/auxv.go
new file mode 100644
index 000000000..70e0e35b7
--- /dev/null
+++ b/pkg/sentry/arch/auxv.go
@@ -0,0 +1,28 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// An AuxEntry represents an entry in an ELF auxiliary vector.
+type AuxEntry struct {
+	Key   uint64
+	Value usermem.Addr
+}
+
+// An Auxv represents an ELF auxiliary vector.
+type Auxv []AuxEntry
diff --git a/pkg/sentry/arch/registers.proto b/pkg/sentry/arch/registers.proto
new file mode 100644
index 000000000..437ff44ca
--- /dev/null
+++ b/pkg/sentry/arch/registers.proto
@@ -0,0 +1,55 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package gvisor;
+
+message AMD64Registers {
+  uint64 rax = 1;
+  uint64 rbx = 2;
+  uint64 rcx = 3;
+  uint64 rdx = 4;
+  uint64 rsi = 5;
+  uint64 rdi = 6;
+  uint64 rsp = 7;
+  uint64 rbp = 8;
+
+  uint64 r8 = 9;
+  uint64 r9 = 10;
+  uint64 r10 = 11;
+  uint64 r11 = 12;
+  uint64 r12 = 13;
+  uint64 r13 = 14;
+  uint64 r14 = 15;
+  uint64 r15 = 16;
+
+  uint64 rip = 17;
+  uint64 rflags = 18;
+  uint64 orig_rax = 19;
+  uint64 cs = 20;
+  uint64 ds = 21;
+  uint64 es = 22;
+  uint64 fs = 23;
+  uint64 gs = 24;
+  uint64 ss = 25;
+  uint64 fs_base = 26;
+  uint64 gs_base = 27;
+}
+
+message Registers {
+  oneof arch {
+    AMD64Registers amd64 = 1;
+  }
+}
diff --git a/pkg/sentry/arch/signal_act.go b/pkg/sentry/arch/signal_act.go
new file mode 100644
index 000000000..36437b965
--- /dev/null
+++ b/pkg/sentry/arch/signal_act.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+// Special values for SignalAct.Handler.
+const (
+	// SignalActDefault is SIG_DFL and specifies that the default behavior for
+	// a signal should be taken.
+	SignalActDefault = 0
+
+	// SignalActIgnore is SIG_IGN and specifies that a signal should be
+	// ignored.
+	SignalActIgnore = 1
+)
+
+// Available signal flags.
+const (
+	SignalFlagNoCldStop    = 0x00000001
+	SignalFlagNoCldWait    = 0x00000002
+	SignalFlagSigInfo      = 0x00000004
+	SignalFlagRestorer     = 0x04000000
+	SignalFlagOnStack      = 0x08000000
+	SignalFlagRestart      = 0x10000000
+	SignalFlagInterrupt    = 0x20000000
+	SignalFlagNoDefer      = 0x40000000
+	SignalFlagResetHandler = 0x80000000
+)
+
+// IsSigInfo returns true iff this handle expects siginfo.
+func (s SignalAct) IsSigInfo() bool {
+	return s.Flags&SignalFlagSigInfo != 0
+}
+
+// IsNoDefer returns true iff this SignalAct has the NoDefer flag set.
+func (s SignalAct) IsNoDefer() bool {
+	return s.Flags&SignalFlagNoDefer != 0
+}
+
+// IsRestart returns true iff this SignalAct has the Restart flag set.
+func (s SignalAct) IsRestart() bool {
+	return s.Flags&SignalFlagRestart != 0
+}
+
+// IsResetHandler returns true iff this SignalAct has the ResetHandler flag set.
+func (s SignalAct) IsResetHandler() bool {
+	return s.Flags&SignalFlagResetHandler != 0
+}
+
+// IsOnStack returns true iff this SignalAct has the OnStack flag set.
+func (s SignalAct) IsOnStack() bool {
+	return s.Flags&SignalFlagOnStack != 0
+}
+
+// HasRestorer returns true iff this SignalAct has the Restorer flag set.
+func (s SignalAct) HasRestorer() bool {
+	return s.Flags&SignalFlagRestorer != 0
+}
+
+// NativeSignalAct is a type that is equivalent to struct sigaction in the
+// guest architecture.
+type NativeSignalAct interface {
+	// SerializeFrom copies the data in the host SignalAct s into this object.
+	SerializeFrom(s *SignalAct)
+
+	// DeserializeTo copies the data in this object into the host SignalAct s.
+	DeserializeTo(s *SignalAct)
+}
diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go
new file mode 100644
index 000000000..4040b530f
--- /dev/null
+++ b/pkg/sentry/arch/signal_amd64.go
@@ -0,0 +1,476 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package arch
+
+import (
+	"encoding/binary"
+	"math"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// SignalAct represents the action that should be taken when a signal is
+// delivered, and is equivalent to struct sigaction on 64-bit x86.
+type SignalAct struct {
+	Handler  uint64
+	Flags    uint64
+	Restorer uint64
+	Mask     linux.SignalSet
+}
+
+// SerializeFrom implements NativeSignalAct.SerializeFrom.
+func (s *SignalAct) SerializeFrom(other *SignalAct) {
+	*s = *other
+}
+
+// DeserializeTo implements NativeSignalAct.DeserializeTo.
+func (s *SignalAct) DeserializeTo(other *SignalAct) {
+	*other = *s
+}
+
+// SignalStack represents information about a user stack, and is equivalent to
+// stack_t on 64-bit x86.
+type SignalStack struct {
+	Addr  uint64
+	Flags uint32
+	_     uint32
+	Size  uint64
+}
+
+// SerializeFrom implements NativeSignalStack.SerializeFrom.
+func (s *SignalStack) SerializeFrom(other *SignalStack) {
+	*s = *other
+}
+
+// DeserializeTo implements NativeSignalStack.DeserializeTo.
+func (s *SignalStack) DeserializeTo(other *SignalStack) {
+	*other = *s
+}
+
+// SignalInfo represents information about a signal being delivered, and is
+// equivalent to struct siginfo on 64-bit x86.
+type SignalInfo struct {
+	Signo int32 // Signal number
+	Errno int32 // Errno value
+	Code  int32 // Signal code
+	_     uint32
+
+	// struct siginfo::_sifields is a union. In SignalInfo, fields in the union
+	// are accessed through methods.
+	//
+	// For reference, here is the definition of _sifields: (_sigfault._trapno,
+	// which does not exist on x86, omitted for clarity)
+	//
+	// union {
+	// 	int _pad[SI_PAD_SIZE];
+	//
+	// 	/* kill() */
+	// 	struct {
+	// 		__kernel_pid_t _pid;	/* sender's pid */
+	// 		__ARCH_SI_UID_T _uid;	/* sender's uid */
+	// 	} _kill;
+	//
+	// 	/* POSIX.1b timers */
+	// 	struct {
+	// 		__kernel_timer_t _tid;	/* timer id */
+	// 		int _overrun;		/* overrun count */
+	// 		char _pad[sizeof( __ARCH_SI_UID_T) - sizeof(int)];
+	// 		sigval_t _sigval;	/* same as below */
+	// 		int _sys_private;       /* not to be passed to user */
+	// 	} _timer;
+	//
+	// 	/* POSIX.1b signals */
+	// 	struct {
+	// 		__kernel_pid_t _pid;	/* sender's pid */
+	// 		__ARCH_SI_UID_T _uid;	/* sender's uid */
+	// 		sigval_t _sigval;
+	// 	} _rt;
+	//
+	// 	/* SIGCHLD */
+	// 	struct {
+	// 		__kernel_pid_t _pid;	/* which child */
+	// 		__ARCH_SI_UID_T _uid;	/* sender's uid */
+	// 		int _status;		/* exit code */
+	// 		__ARCH_SI_CLOCK_T _utime;
+	// 		__ARCH_SI_CLOCK_T _stime;
+	// 	} _sigchld;
+	//
+	// 	/* SIGILL, SIGFPE, SIGSEGV, SIGBUS */
+	// 	struct {
+	// 		void *_addr; /* faulting insn/memory ref. */
+	// 		short _addr_lsb; /* LSB of the reported address */
+	// 	} _sigfault;
+	//
+	// 	/* SIGPOLL */
+	// 	struct {
+	// 		__ARCH_SI_BAND_T _band;	/* POLL_IN, POLL_OUT, POLL_MSG */
+	// 		int _fd;
+	// 	} _sigpoll;
+	//
+	// 	/* SIGSYS */
+	// 	struct {
+	// 		void *_call_addr; /* calling user insn */
+	// 		int _syscall;	/* triggering system call number */
+	// 		unsigned int _arch;	/* AUDIT_ARCH_* of syscall */
+	// 	} _sigsys;
+	// } _sifields;
+	//
+	// _sifields is padded so that the size of siginfo is SI_MAX_SIZE = 128
+	// bytes.
+	Fields [128 - 16]byte
+}
+
+// FixSignalCodeForUser fixes up si_code.
+//
+// The si_code we get from Linux may contain the kernel-specific code in the
+// top 16 bits if it's positive (e.g., from ptrace). Linux's
+// copy_siginfo_to_user does
+//     err |= __put_user((short)from->si_code, &to->si_code);
+// to mask out those bits and we need to do the same.
+func (s *SignalInfo) FixSignalCodeForUser() {
+	if s.Code > 0 {
+		s.Code &= 0x0000ffff
+	}
+}
+
+// Pid returns the si_pid field.
+func (s *SignalInfo) Pid() int32 {
+	return int32(usermem.ByteOrder.Uint32(s.Fields[0:4]))
+}
+
+// SetPid mutates the si_pid field.
+func (s *SignalInfo) SetPid(val int32) {
+	usermem.ByteOrder.PutUint32(s.Fields[0:4], uint32(val))
+}
+
+// Uid returns the si_uid field.
+func (s *SignalInfo) Uid() int32 {
+	return int32(usermem.ByteOrder.Uint32(s.Fields[4:8]))
+}
+
+// SetUid mutates the si_uid field.
+func (s *SignalInfo) SetUid(val int32) {
+	usermem.ByteOrder.PutUint32(s.Fields[4:8], uint32(val))
+}
+
+// Addr returns the si_addr field.
+func (s *SignalInfo) Addr() uint64 {
+	return usermem.ByteOrder.Uint64(s.Fields[0:8])
+}
+
+// SetAddr sets the si_addr field.
+func (s *SignalInfo) SetAddr(val uint64) {
+	usermem.ByteOrder.PutUint64(s.Fields[0:8], val)
+}
+
+// Status returns the si_status field.
+func (s *SignalInfo) Status() int32 {
+	return int32(usermem.ByteOrder.Uint32(s.Fields[8:12]))
+}
+
+// SetStatus mutates the si_status field.
+func (s *SignalInfo) SetStatus(val int32) {
+	usermem.ByteOrder.PutUint32(s.Fields[8:12], uint32(val))
+}
+
+// CallAddr returns the si_call_addr field.
+func (s *SignalInfo) CallAddr() uint64 {
+	return usermem.ByteOrder.Uint64(s.Fields[0:8])
+}
+
+// SetCallAddr mutates the si_call_addr field.
+func (s *SignalInfo) SetCallAddr(val uint64) {
+	usermem.ByteOrder.PutUint64(s.Fields[0:8], val)
+}
+
+// Syscall returns the si_syscall field.
+func (s *SignalInfo) Syscall() int32 {
+	return int32(usermem.ByteOrder.Uint32(s.Fields[8:12]))
+}
+
+// SetSyscall mutates the si_syscall field.
+func (s *SignalInfo) SetSyscall(val int32) {
+	usermem.ByteOrder.PutUint32(s.Fields[8:12], uint32(val))
+}
+
+// Arch returns the si_arch field.
+func (s *SignalInfo) Arch() uint32 {
+	return usermem.ByteOrder.Uint32(s.Fields[12:16])
+}
+
+// SetArch mutates the si_arch field.
+func (s *SignalInfo) SetArch(val uint32) {
+	usermem.ByteOrder.PutUint32(s.Fields[12:16], val)
+}
+
+// SignalContext64 is equivalent to struct sigcontext, the type passed as the
+// second argument to signal handlers set by signal(2).
+type SignalContext64 struct {
+	R8      uint64
+	R9      uint64
+	R10     uint64
+	R11     uint64
+	R12     uint64
+	R13     uint64
+	R14     uint64
+	R15     uint64
+	Rdi     uint64
+	Rsi     uint64
+	Rbp     uint64
+	Rbx     uint64
+	Rdx     uint64
+	Rax     uint64
+	Rcx     uint64
+	Rsp     uint64
+	Rip     uint64
+	Eflags  uint64
+	Cs      uint16
+	Gs      uint16 // always 0 on amd64.
+	Fs      uint16 // always 0 on amd64.
+	Ss      uint16 // only restored if _UC_STRICT_RESTORE_SS (unsupported).
+	Err     uint64
+	Trapno  uint64
+	Oldmask linux.SignalSet
+	Cr2     uint64
+	// Pointer to a struct _fpstate.
+	Fpstate  uint64
+	Reserved [8]uint64
+}
+
+// Flags for UContext64.Flags.
+const (
+	_UC_FP_XSTATE         = 1
+	_UC_SIGCONTEXT_SS     = 2
+	_UC_STRICT_RESTORE_SS = 4
+)
+
+// UContext64 is equivalent to ucontext_t on 64-bit x86.
+type UContext64 struct {
+	Flags    uint64
+	Link     uint64
+	Stack    SignalStack
+	MContext SignalContext64
+	Sigset   linux.SignalSet
+}
+
+// NewSignalAct implements Context.NewSignalAct.
+func (c *context64) NewSignalAct() NativeSignalAct {
+	return &SignalAct{}
+}
+
+// NewSignalStack implements Context.NewSignalStack.
+func (c *context64) NewSignalStack() NativeSignalStack {
+	return &SignalStack{}
+}
+
+// From Linux 'arch/x86/include/uapi/asm/sigcontext.h' the following is the
+// size of the magic cookie at the end of the xsave frame.
+//
+// NOTE: Currently we don't actually populate the fpstate
+// on the signal stack.
+const _FP_XSTATE_MAGIC2_SIZE = 4
+
+func (c *context64) fpuFrameSize() (size int, useXsave bool) {
+	size = len(c.x86FPState)
+	if size > 512 {
+		// Make room for the magic cookie at the end of the xsave frame.
+		size += _FP_XSTATE_MAGIC2_SIZE
+		useXsave = true
+	}
+	return size, useXsave
+}
+
+// SignalSetup implements Context.SignalSetup. (Compare to Linux's
+// arch/x86/kernel/signal.c:__setup_rt_frame().)
+func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt *SignalStack, sigset linux.SignalSet) error {
+	sp := st.Bottom
+
+	// "The 128-byte area beyond the location pointed to by %rsp is considered
+	// to be reserved and shall not be modified by signal or interrupt
+	// handlers. ... leaf functions may use this area for their entire stack
+	// frame, rather than adjusting the stack pointer in the prologue and
+	// epilogue." - AMD64 ABI
+	//
+	// (But this doesn't apply if we're starting at the top of the signal
+	// stack, in which case there is no following stack frame.)
+	if !(alt.IsEnabled() && sp == alt.Top()) {
+		sp -= 128
+	}
+
+	// Allocate space for floating point state on the stack.
+	//
+	// This isn't strictly necessary because we don't actually populate
+	// the fpstate. However we do store the floating point state of the
+	// interrupted thread inside the sentry. Simply accounting for this
+	// space on the user stack naturally caps the amount of memory the
+	// sentry will allocate for this purpose.
+	fpSize, _ := c.fpuFrameSize()
+	sp = (sp - usermem.Addr(fpSize)) & ^usermem.Addr(63)
+
+	// Construct the UContext64 now since we need its size.
+	uc := &UContext64{
+		// No _UC_FP_XSTATE: see Fpstate above.
+		// No _UC_STRICT_RESTORE_SS: we don't allow SS changes.
+		Flags: _UC_SIGCONTEXT_SS,
+		Stack: *alt,
+		MContext: SignalContext64{
+			R8:      c.Regs.R8,
+			R9:      c.Regs.R9,
+			R10:     c.Regs.R10,
+			R11:     c.Regs.R11,
+			R12:     c.Regs.R12,
+			R13:     c.Regs.R13,
+			R14:     c.Regs.R14,
+			R15:     c.Regs.R15,
+			Rdi:     c.Regs.Rdi,
+			Rsi:     c.Regs.Rsi,
+			Rbp:     c.Regs.Rbp,
+			Rbx:     c.Regs.Rbx,
+			Rdx:     c.Regs.Rdx,
+			Rax:     c.Regs.Rax,
+			Rcx:     c.Regs.Rcx,
+			Rsp:     c.Regs.Rsp,
+			Rip:     c.Regs.Rip,
+			Eflags:  c.Regs.Eflags,
+			Cs:      uint16(c.Regs.Cs),
+			Ss:      uint16(c.Regs.Ss),
+			Oldmask: sigset,
+		},
+		Sigset: sigset,
+	}
+
+	// TODO: Set SignalContext64.Err, Trapno, and Cr2 based on
+	// the fault that caused the signal. For now, leave Err and Trapno
+	// unset and assume CR2 == info.Addr() for SIGSEGVs and SIGBUSes.
+	if linux.Signal(info.Signo) == linux.SIGSEGV || linux.Signal(info.Signo) == linux.SIGBUS {
+		uc.MContext.Cr2 = info.Addr()
+	}
+
+	// "... the value (%rsp+8) is always a multiple of 16 (...) when control is
+	// transferred to the function entry point." - AMD64 ABI
+	ucSize := binary.Size(uc)
+	if ucSize < 0 {
+		// This can only happen if we've screwed up the definition of
+		// UContext64.
+		panic("can't get size of UContext64")
+	}
+	// st.Arch.Width() is for the restorer address. sizeof(siginfo) == 128.
+	frameSize := int(st.Arch.Width()) + ucSize + 128
+	frameBottom := (sp-usermem.Addr(frameSize)) & ^usermem.Addr(15) - 8
+	sp = frameBottom + usermem.Addr(frameSize)
+	st.Bottom = sp
+
+	info.FixSignalCodeForUser()
+
+	// Set up the stack frame.
+	infoAddr, err := st.Push(info)
+	if err != nil {
+		return err
+	}
+	ucAddr, err := st.Push(uc)
+	if err != nil {
+		return err
+	}
+	if act.HasRestorer() {
+		// Push the restorer return address.
+		// Note that this doesn't need to be popped.
+		if _, err := st.Push(usermem.Addr(act.Restorer)); err != nil {
+			return err
+		}
+	} else {
+		// amd64 requires a restorer.
+		return syscall.EFAULT
+	}
+
+	// Set up registers.
+	c.Regs.Rip = act.Handler
+	c.Regs.Rsp = uint64(st.Bottom)
+	c.Regs.Rdi = uint64(info.Signo)
+	c.Regs.Rsi = uint64(infoAddr)
+	c.Regs.Rdx = uint64(ucAddr)
+	c.Regs.Rax = 0
+	c.Regs.Ds = userDS
+	c.Regs.Es = userDS
+	c.Regs.Cs = userCS
+	c.Regs.Ss = userDS
+
+	// Save the thread's floating point state.
+	c.sigFPState = append(c.sigFPState, c.x86FPState)
+
+	// Signal handler gets a clean floating point state.
+	c.x86FPState = newX86FPState()
+
+	return nil
+}
+
+// SignalRestore implements Context.SignalRestore. (Compare to Linux's
+// arch/x86/kernel/signal.c:sys_rt_sigreturn().)
+func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, error) {
+	// Copy out the stack frame.
+	var uc UContext64
+	if _, err := st.Pop(&uc); err != nil {
+		return 0, err
+	}
+	var info SignalInfo
+	if _, err := st.Pop(&info); err != nil {
+		return 0, err
+	}
+
+	// Restore registers.
+	c.Regs.R8 = uc.MContext.R8
+	c.Regs.R9 = uc.MContext.R9
+	c.Regs.R10 = uc.MContext.R10
+	c.Regs.R11 = uc.MContext.R11
+	c.Regs.R12 = uc.MContext.R12
+	c.Regs.R13 = uc.MContext.R13
+	c.Regs.R14 = uc.MContext.R14
+	c.Regs.R15 = uc.MContext.R15
+	c.Regs.Rdi = uc.MContext.Rdi
+	c.Regs.Rsi = uc.MContext.Rsi
+	c.Regs.Rbp = uc.MContext.Rbp
+	c.Regs.Rbx = uc.MContext.Rbx
+	c.Regs.Rdx = uc.MContext.Rdx
+	c.Regs.Rax = uc.MContext.Rax
+	c.Regs.Rcx = uc.MContext.Rcx
+	c.Regs.Rsp = uc.MContext.Rsp
+	c.Regs.Rip = uc.MContext.Rip
+	c.Regs.Eflags = (c.Regs.Eflags & ^eflagsRestorable) | (uc.MContext.Eflags & eflagsRestorable)
+	c.Regs.Cs = uint64(uc.MContext.Cs) | 3
+	// N.B. _UC_STRICT_RESTORE_SS not supported.
+	c.Regs.Orig_rax = math.MaxUint64
+
+	// Restore floating point state.
+	l := len(c.sigFPState)
+	if l > 0 {
+		c.x86FPState = c.sigFPState[l-1]
+		// NOTE: State save requires that any slice
+		// elements from '[len:cap]' to be zero value.
+		c.sigFPState[l-1] = nil
+		c.sigFPState = c.sigFPState[0 : l-1]
+	} else {
+		// This might happen if sigreturn(2) calls are unbalanced with
+		// respect to signal handler entries. This is not expected so
+		// don't bother to do anything fancy with the floating point
+		// state.
+		log.Infof("sigreturn unable to restore application fpstate")
+	}
+
+	return uc.Sigset, nil
+}
diff --git a/pkg/sentry/arch/signal_info.go b/pkg/sentry/arch/signal_info.go
new file mode 100644
index 000000000..ec004ae75
--- /dev/null
+++ b/pkg/sentry/arch/signal_info.go
@@ -0,0 +1,66 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+// Possible values for SignalInfo.Code. These values originate from the Linux
+// kernel's include/uapi/asm-generic/siginfo.h.
+const (
+	// SignalInfoUser (properly SI_USER) indicates that a signal was sent from
+	// a kill() or raise() syscall.
+	SignalInfoUser = 0
+
+	// SignalInfoKernel (properly SI_KERNEL) indicates that the signal was sent
+	// by the kernel.
+	SignalInfoKernel = 0x80
+
+	// SignalInfoTimer (properly SI_TIMER) indicates that the signal was sent
+	// by an expired timer.
+	SignalInfoTimer = -2
+
+	// SignalInfoTkill (properly SI_TKILL) indicates that the signal was sent
+	// from a tkill() or tgkill() syscall.
+	SignalInfoTkill = -6
+
+	// CLD_* codes are only meaningful for SIGCHLD.
+
+	// CLD_EXITED indicates that a task exited.
+	CLD_EXITED = 1
+
+	// CLD_KILLED indicates that a task was killed by a signal.
+	CLD_KILLED = 2
+
+	// CLD_DUMPED indicates that a task was killed by a signal and then dumped
+	// core.
+	CLD_DUMPED = 3
+
+	// CLD_TRAPPED indicates that a task was stopped by ptrace.
+	CLD_TRAPPED = 4
+
+	// CLD_STOPPED indicates that a thread group completed a group stop.
+	CLD_STOPPED = 5
+
+	// CLD_CONTINUED indicates that a group-stopped thread group was continued.
+	CLD_CONTINUED = 6
+
+	// SYS_* codes are only meaningful for SIGSYS.
+
+	// SYS_SECCOMP indicates that a signal originates from seccomp.
+	SYS_SECCOMP = 1
+
+	// TRAP_* codes are only meaningful for SIGTRAP.
+
+	// TRAP_BRKPT indicates a breakpoint trap.
+	TRAP_BRKPT = 1
+)
diff --git a/pkg/sentry/arch/signal_stack.go b/pkg/sentry/arch/signal_stack.go
new file mode 100644
index 000000000..7c6531d79
--- /dev/null
+++ b/pkg/sentry/arch/signal_stack.go
@@ -0,0 +1,58 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package arch
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// SignalStackFlagOnStack is possible set on return from getaltstack,
+	// in order to indicate that the thread is currently on the alt stack.
+	SignalStackFlagOnStack = 1
+
+	// SignalStackFlagDisable is a flag to indicate the stack is disabled.
+	SignalStackFlagDisable = 2
+)
+
+// IsEnabled returns true iff this signal stack is marked as enabled.
+func (s SignalStack) IsEnabled() bool {
+	return s.Flags&SignalStackFlagDisable == 0
+}
+
+// Top returns the stack's top address.
+func (s SignalStack) Top() usermem.Addr {
+	return usermem.Addr(s.Addr + s.Size)
+}
+
+// SetOnStack marks this signal stack as in use. (This is only called on copies
+// sent to user applications, so there's no corresponding ClearOnStack.)
+func (s *SignalStack) SetOnStack() {
+	s.Flags |= SignalStackFlagOnStack
+}
+
+// NativeSignalStack is a type that is equivalent to stack_t in the guest
+// architecture.
+type NativeSignalStack interface {
+	// SerializeFrom copies the data in the host SignalStack s into this
+	// object.
+	SerializeFrom(s *SignalStack)
+
+	// DeserializeTo copies the data in this object into the host SignalStack
+	// s.
+	DeserializeTo(s *SignalStack)
+}
diff --git a/pkg/sentry/arch/stack.go b/pkg/sentry/arch/stack.go
new file mode 100644
index 000000000..6c1b9be82
--- /dev/null
+++ b/pkg/sentry/arch/stack.go
@@ -0,0 +1,246 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package arch
+
+import (
+	"encoding/binary"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Stack is a simple wrapper around a usermem.IO and an address.
+type Stack struct {
+	// Our arch info.
+	// We use this for automatic Native conversion of usermem.Addrs during
+	// Push() and Pop().
+	Arch Context
+
+	// The interface used to actually copy user memory.
+	IO usermem.IO
+
+	// Our current stack bottom.
+	Bottom usermem.Addr
+}
+
+// Push pushes the given values on to the stack.
+//
+// (This method supports Addrs and treats them as native types.)
+func (s *Stack) Push(vals ...interface{}) (usermem.Addr, error) {
+	for _, v := range vals {
+
+		// We convert some types to well-known serializable quanities.
+		var norm interface{}
+
+		// For array types, we will automatically add an appropriate
+		// terminal value. This is done simply to make the interface
+		// easier to use.
+		var term interface{}
+
+		switch v.(type) {
+		case string:
+			norm = []byte(v.(string))
+			term = byte(0)
+		case []int8, []uint8:
+			norm = v
+			term = byte(0)
+		case []int16, []uint16:
+			norm = v
+			term = uint16(0)
+		case []int32, []uint32:
+			norm = v
+			term = uint32(0)
+		case []int64, []uint64:
+			norm = v
+			term = uint64(0)
+		case []usermem.Addr:
+			// Special case: simply push recursively.
+			_, err := s.Push(s.Arch.Native(uintptr(0)))
+			if err != nil {
+				return 0, err
+			}
+			varr := v.([]usermem.Addr)
+			for i := len(varr) - 1; i >= 0; i-- {
+				_, err := s.Push(varr[i])
+				if err != nil {
+					return 0, err
+				}
+			}
+			continue
+		case usermem.Addr:
+			norm = s.Arch.Native(uintptr(v.(usermem.Addr)))
+		default:
+			norm = v
+		}
+
+		if term != nil {
+			_, err := s.Push(term)
+			if err != nil {
+				return 0, err
+			}
+		}
+
+		c := binary.Size(norm)
+		if c < 0 {
+			return 0, fmt.Errorf("bad binary.Size for %T", v)
+		}
+		// TODO: Use a real context.Context.
+		n, err := usermem.CopyObjectOut(context.Background(), s.IO, s.Bottom-usermem.Addr(c), norm, usermem.IOOpts{})
+		if err != nil || c != n {
+			return 0, err
+		}
+
+		s.Bottom -= usermem.Addr(n)
+	}
+
+	return s.Bottom, nil
+}
+
+// Pop pops the given values off the stack.
+//
+// (This method supports Addrs and treats them as native types.)
+func (s *Stack) Pop(vals ...interface{}) (usermem.Addr, error) {
+	for _, v := range vals {
+
+		vaddr, isVaddr := v.(*usermem.Addr)
+
+		var n int
+		var err error
+		if isVaddr {
+			value := s.Arch.Native(uintptr(0))
+			// TODO: Use a real context.Context.
+			n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, value, usermem.IOOpts{})
+			*vaddr = usermem.Addr(s.Arch.Value(value))
+		} else {
+			// TODO: Use a real context.Context.
+			n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, v, usermem.IOOpts{})
+		}
+		if err != nil {
+			return 0, err
+		}
+
+		s.Bottom += usermem.Addr(n)
+	}
+
+	return s.Bottom, nil
+}
+
+// Align aligns the stack to the given offset.
+func (s *Stack) Align(offset int) {
+	if s.Bottom%usermem.Addr(offset) != 0 {
+		s.Bottom -= (s.Bottom % usermem.Addr(offset))
+	}
+}
+
+// StackLayout describes the location of the arguments and environment on the
+// stack.
+type StackLayout struct {
+	// ArgvStart is the beginning of the argument vector.
+	ArgvStart usermem.Addr
+
+	// ArgvEnd is the end of the argument vector.
+	ArgvEnd usermem.Addr
+
+	// EnvvStart is the beginning of the environment vector.
+	EnvvStart usermem.Addr
+
+	// EnvvEnd is the end of the environment vector.
+	EnvvEnd usermem.Addr
+}
+
+// Load pushes the given args, env and aux vector to the stack using the
+// well-known format for a new executable. It returns the start and end
+// of the argument and environment vectors.
+func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) {
+	l := StackLayout{}
+
+	// Make sure we start with a 16-byte alignment.
+	s.Align(16)
+
+	// Push our strings.
+	l.ArgvEnd = s.Bottom
+	argAddrs := make([]usermem.Addr, len(args))
+	for i := len(args) - 1; i >= 0; i-- {
+		addr, err := s.Push(args[i])
+		if err != nil {
+			return StackLayout{}, err
+		}
+		argAddrs[i] = addr
+	}
+	l.ArgvStart = s.Bottom
+
+	// Push our environment.
+	l.EnvvEnd = s.Bottom
+	envAddrs := make([]usermem.Addr, len(env))
+	for i := len(env) - 1; i >= 0; i-- {
+		addr, err := s.Push(env[i])
+		if err != nil {
+			return StackLayout{}, err
+		}
+		envAddrs[i] = addr
+	}
+	l.EnvvStart = s.Bottom
+
+	// We need to align the arguments appropriately.
+	//
+	// We must finish on a 16-byte alignment, but we'll play it
+	// conservatively and finish at 32-bytes. It would be nice to be able
+	// to call Align here, but unfortunately we need to align the stack
+	// with all the variable sized arrays pushed. So we just need to do
+	// some calculations.
+	argvSize := s.Arch.Width() * uint(len(args)+1)
+	envvSize := s.Arch.Width() * uint(len(env)+1)
+	auxvSize := s.Arch.Width() * 2 * uint(len(aux)+1)
+	total := usermem.Addr(argvSize) + usermem.Addr(envvSize) + usermem.Addr(auxvSize) + usermem.Addr(s.Arch.Width())
+	expectedBottom := s.Bottom - total
+	if expectedBottom%32 != 0 {
+		s.Bottom -= expectedBottom % 32
+	}
+
+	// Push our auxvec.
+	// NOTE: We need an extra zero here per spec.
+	// The Push function will automatically terminate
+	// strings and arrays with a single null value.
+	auxv := make([]usermem.Addr, 0, len(aux))
+	for _, a := range aux {
+		auxv = append(auxv, usermem.Addr(a.Key), a.Value)
+	}
+	auxv = append(auxv, usermem.Addr(0))
+	_, err := s.Push(auxv)
+	if err != nil {
+		return StackLayout{}, err
+	}
+
+	// Push environment.
+	_, err = s.Push(envAddrs)
+	if err != nil {
+		return StackLayout{}, err
+	}
+
+	// Push args.
+	_, err = s.Push(argAddrs)
+	if err != nil {
+		return StackLayout{}, err
+	}
+
+	// Push arg count.
+	_, err = s.Push(usermem.Addr(len(args)))
+	if err != nil {
+		return StackLayout{}, err
+	}
+
+	return l, nil
+}
diff --git a/pkg/sentry/arch/syscalls_amd64.go b/pkg/sentry/arch/syscalls_amd64.go
new file mode 100644
index 000000000..41d8ba0d1
--- /dev/null
+++ b/pkg/sentry/arch/syscalls_amd64.go
@@ -0,0 +1,52 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package arch
+
+const restartSyscallNr = uintptr(219)
+
+// SyscallNo returns the syscall number according to the 64-bit convention.
+func (c *context64) SyscallNo() uintptr {
+	return uintptr(c.Regs.Orig_rax)
+}
+
+// SyscallArgs provides syscall arguments according to the 64-bit convention.
+//
+// Due to the way addresses are mapped for the sentry this binary *must* be
+// built in 64-bit mode. So we can just assume the syscall numbers that come
+// back match the expected host system call numbers.
+func (c *context64) SyscallArgs() SyscallArguments {
+	return SyscallArguments{
+		SyscallArgument{Value: uintptr(c.Regs.Rdi)},
+		SyscallArgument{Value: uintptr(c.Regs.Rsi)},
+		SyscallArgument{Value: uintptr(c.Regs.Rdx)},
+		SyscallArgument{Value: uintptr(c.Regs.R10)},
+		SyscallArgument{Value: uintptr(c.Regs.R8)},
+		SyscallArgument{Value: uintptr(c.Regs.R9)},
+	}
+}
+
+// RestartSyscall implements Context.RestartSyscall.
+func (c *context64) RestartSyscall() {
+	c.Regs.Rip -= SyscallWidth
+	c.Regs.Rax = c.Regs.Orig_rax
+}
+
+// RestartSyscallWithRestartBlock implements Context.RestartSyscallWithRestartBlock.
+func (c *context64) RestartSyscallWithRestartBlock() {
+	c.Regs.Rip -= SyscallWidth
+	c.Regs.Rax = uint64(restartSyscallNr)
+}
diff --git a/pkg/sentry/context/BUILD b/pkg/sentry/context/BUILD
new file mode 100644
index 000000000..ff39f94ba
--- /dev/null
+++ b/pkg/sentry/context/BUILD
@@ -0,0 +1,14 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "context",
+    srcs = ["context.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/context",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/amutex",
+        "//pkg/log",
+    ],
+)
diff --git a/pkg/sentry/context/context.go b/pkg/sentry/context/context.go
new file mode 100644
index 000000000..e0dffafba
--- /dev/null
+++ b/pkg/sentry/context/context.go
@@ -0,0 +1,103 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package context defines the sentry's Context type.
+package context
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/amutex"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+)
+
+// A Context represents a thread of execution (hereafter "goroutine" to reflect
+// Go idiosyncrasy). It carries state associated with the goroutine across API
+// boundaries.
+//
+// While Context exists for essentially the same reasons as Go's standard
+// context.Context, the standard type represents the state of an operation
+// rather than that of a goroutine. This is a critical distinction:
+//
+// - Unlike context.Context, which "may be passed to functions running in
+// different goroutines", it is *not safe* to use the same Context in multiple
+// concurrent goroutines.
+//
+// - It is *not safe* to retain a Context passed to a function beyond the scope
+// of that function call.
+//
+// In both cases, values extracted from the Context should be used instead.
+type Context interface {
+	log.Logger
+	amutex.Sleeper
+
+	// UninterruptibleSleepStart indicates the beginning of an uninterruptible
+	// sleep state (equivalent to Linux's TASK_UNINTERRUPTIBLE). If deactivate
+	// is true and the Context represents a Task, the Task's AddressSpace is
+	// deactivated.
+	UninterruptibleSleepStart(deactivate bool)
+
+	// UninterruptibleSleepFinish indicates the end of an uninterruptible sleep
+	// state that was begun by a previous call to UninterruptibleSleepStart. If
+	// activate is true and the Context represents a Task, the Task's
+	// AddressSpace is activated. Normally activate is the same value as the
+	// deactivate parameter passed to UninterruptibleSleepStart.
+	UninterruptibleSleepFinish(activate bool)
+
+	// Value returns the value associated with this Context for key, or nil if
+	// no value is associated with key. Successive calls to Value with the same
+	// key returns the same result.
+	//
+	// A key identifies a specific value in a Context. Functions that wish to
+	// retrieve values from Context typically allocate a key in a global
+	// variable then use that key as the argument to Context.Value. A key can
+	// be any type that supports equality; packages should define keys as an
+	// unexported type to avoid collisions.
+	Value(key interface{}) interface{}
+}
+
+type logContext struct {
+	log.Logger
+	NoopSleeper
+}
+
+// Value implements Context.Value.
+func (logContext) Value(key interface{}) interface{} {
+	return nil
+}
+
+// NoopSleeper is a noop implementation of amutex.Sleeper and
+// Context.UninterruptibleSleep* methods for anonymous embedding in other types
+// that do not want to notify kernel.Task about sleeps.
+type NoopSleeper struct {
+	amutex.NoopSleeper
+}
+
+// UninterruptibleSleepStart does nothing.
+func (NoopSleeper) UninterruptibleSleepStart(bool) {}
+
+// UninterruptibleSleepFinish does nothing.
+func (NoopSleeper) UninterruptibleSleepFinish(bool) {}
+
+// Background returns an empty context using the default logger.
+//
+// Users should be wary of using a Background context. Please tag any use with
+// FIXME and a note to remove this use.
+//
+// Generally, one should use the Task as their context when available, or avoid
+// having to use a context in places where a Task is unavailable.
+//
+// Using a Background context for tests is fine, as long as no values are
+// needed from the context in the tested code paths.
+func Background() Context {
+	return logContext{Logger: log.Log()}
+}
diff --git a/pkg/sentry/context/contexttest/BUILD b/pkg/sentry/context/contexttest/BUILD
new file mode 100644
index 000000000..5977344de
--- /dev/null
+++ b/pkg/sentry/context/contexttest/BUILD
@@ -0,0 +1,34 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "contexttest_state",
+    srcs = [
+        "contexttest.go",
+    ],
+    out = "contexttest_state.go",
+    package = "contexttest",
+)
+
+go_library(
+    name = "contexttest",
+    testonly = 1,
+    srcs = [
+        "contexttest.go",
+        "contexttest_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/platform/ptrace",
+        "//pkg/sentry/uniqueid",
+        "//pkg/state",
+    ],
+)
diff --git a/pkg/sentry/context/contexttest/contexttest.go b/pkg/sentry/context/contexttest/contexttest.go
new file mode 100644
index 000000000..193ce3440
--- /dev/null
+++ b/pkg/sentry/context/contexttest/contexttest.go
@@ -0,0 +1,133 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package contexttest builds a test context.Context.
+package contexttest
+
+import (
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ptrace"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+)
+
+// Context returns a Context that may be used in tests. Uses ptrace as the
+// platform.Platform.
+func Context(tb testing.TB) context.Context {
+	p, err := ptrace.New()
+	if err != nil {
+		tb.Fatal(err)
+	}
+	// Test usage of context.Background is fine.
+	return &testContext{
+		Context:  context.Background(),
+		l:        limits.NewLimitSet(),
+		platform: p,
+	}
+}
+
+type testContext struct {
+	context.Context
+	l        *limits.LimitSet
+	platform platform.Platform
+}
+
+// globalUniqueID tracks incremental unique identifiers for tests.
+var globalUniqueID uint64
+
+// lastInotifyCookie is a monotonically increasing counter for generating unique
+// inotify cookies. Must be accessed using atomic ops.
+var lastInotifyCookie uint32
+
+// hostClock implements ktime.Clock.
+type hostClock struct {
+	ktime.WallRateClock
+	ktime.NoClockEvents
+}
+
+// Now implements ktime.Clock.Now.
+func (hostClock) Now() ktime.Time {
+	return ktime.FromNanoseconds(time.Now().UnixNano())
+}
+
+// Value implements context.Context.
+func (t *testContext) Value(key interface{}) interface{} {
+	switch key {
+	case limits.CtxLimits:
+		return t.l
+	case platform.CtxPlatform:
+		return t.platform
+	case uniqueid.CtxGlobalUniqueID:
+		return atomic.AddUint64(&globalUniqueID, 1)
+	case uniqueid.CtxInotifyCookie:
+		return atomic.AddUint32(&lastInotifyCookie, 1)
+	case ktime.CtxRealtimeClock:
+		return hostClock{}
+	default:
+		return t.Context.Value(key)
+	}
+}
+
+// RootContext returns a Context that may be used in tests that need root
+// credentials. Uses ptrace as the platform.Platform.
+func RootContext(tb testing.TB) context.Context {
+	return WithCreds(Context(tb), auth.NewRootCredentials(auth.NewRootUserNamespace()))
+}
+
+// WithCreds returns a copy of ctx carrying creds.
+func WithCreds(ctx context.Context, creds *auth.Credentials) context.Context {
+	return &authContext{ctx, creds}
+}
+
+type authContext struct {
+	context.Context
+	creds *auth.Credentials
+}
+
+// Value implements context.Context.
+func (ac *authContext) Value(key interface{}) interface{} {
+	switch key {
+	case auth.CtxCredentials:
+		return ac.creds
+	default:
+		return ac.Context.Value(key)
+	}
+}
+
+// WithLimitSet returns a copy of ctx carrying l.
+func WithLimitSet(ctx context.Context, l *limits.LimitSet) context.Context {
+	return limitContext{ctx, l}
+}
+
+type limitContext struct {
+	context.Context
+	l *limits.LimitSet
+}
+
+// Value implements context.Context.
+func (lc limitContext) Value(key interface{}) interface{} {
+	switch key {
+	case limits.CtxLimits:
+		return lc.l
+	default:
+		return lc.Context.Value(key)
+	}
+}
diff --git a/pkg/sentry/control/BUILD b/pkg/sentry/control/BUILD
new file mode 100644
index 000000000..4d1d0d019
--- /dev/null
+++ b/pkg/sentry/control/BUILD
@@ -0,0 +1,39 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "control",
+    srcs = [
+        "control.go",
+        "proc.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/control",
+    visibility = [
+        "//pkg/sentry:internal",
+    ],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/host",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/usage",
+        "//pkg/urpc",
+    ],
+)
+
+go_test(
+    name = "control_test",
+    size = "small",
+    srcs = ["proc_test.go"],
+    embed = [":control"],
+    deps = [
+        "//pkg/log",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usage",
+    ],
+)
diff --git a/pkg/sentry/control/control.go b/pkg/sentry/control/control.go
new file mode 100644
index 000000000..a6ee6e649
--- /dev/null
+++ b/pkg/sentry/control/control.go
@@ -0,0 +1,17 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package control contains types that expose control server methods, and can
+// be used to configure and interact with a running sandbox process.
+package control
diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go
new file mode 100644
index 000000000..7d06a1d04
--- /dev/null
+++ b/pkg/sentry/control/proc.go
@@ -0,0 +1,293 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package control
+
+import (
+	"bytes"
+	"encoding/json"
+	"fmt"
+	"syscall"
+	"text/tabwriter"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/host"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/urpc"
+)
+
+// Proc includes task-related functions.
+//
+// At the moment, this is limited to exec support.
+type Proc struct {
+	Kernel *kernel.Kernel
+}
+
+// ExecArgs is the set of arguments to exec.
+type ExecArgs struct {
+	// Filename is the filename to load.
+	//
+	// If this is provided as "", then the file will be guessed via Argv[0].
+	Filename string `json:"filename"`
+
+	// Argv is a list of arguments.
+	Argv []string `json:"argv"`
+
+	// Envv is a list of environment variables.
+	Envv []string `json:"envv"`
+
+	// WorkingDirectory defines the working directory for the new process.
+	WorkingDirectory string `json:"wd"`
+
+	// KUID is the UID to run with in the root user namespace. Defaults to
+	// root if not set explicitly.
+	KUID auth.KUID
+
+	// KGID is the GID to run with in the root user namespace. Defaults to
+	// the root group if not set explicitly.
+	KGID auth.KGID
+
+	// ExtraKGIDs is the list of additional groups to which the user
+	// belongs.
+	ExtraKGIDs []auth.KGID
+
+	// Capabilities is the list of capabilities to give to the process.
+	Capabilities *auth.TaskCapabilities
+
+	// Detach indicates whether Exec should detach once the process starts.
+	Detach bool
+
+	// FilePayload determines the files to give to the new process.
+	urpc.FilePayload
+}
+
+// Exec runs a new task.
+func (proc *Proc) Exec(args *ExecArgs, waitStatus *uint32) error {
+	// Import file descriptors.
+	l := limits.NewLimitSet()
+	fdm := proc.Kernel.NewFDMap()
+	defer fdm.DecRef()
+
+	creds := auth.NewUserCredentials(
+		args.KUID,
+		args.KGID,
+		args.ExtraKGIDs,
+		args.Capabilities,
+		proc.Kernel.RootUserNamespace())
+
+	initArgs := kernel.CreateProcessArgs{
+		Filename:             args.Filename,
+		Argv:                 args.Argv,
+		Envv:                 args.Envv,
+		WorkingDirectory:     args.WorkingDirectory,
+		Credentials:          creds,
+		FDMap:                fdm,
+		Umask:                0022,
+		Limits:               l,
+		MaxSymlinkTraversals: linux.MaxSymlinkTraversals,
+		UTSNamespace:         proc.Kernel.RootUTSNamespace(),
+		IPCNamespace:         proc.Kernel.RootIPCNamespace(),
+	}
+	ctx := initArgs.NewContext(proc.Kernel)
+	mounter := fs.FileOwnerFromContext(ctx)
+
+	for appFD, f := range args.FilePayload.Files {
+		// Copy the underlying FD.
+		newFD, err := syscall.Dup(int(f.Fd()))
+		if err != nil {
+			return err
+		}
+		f.Close()
+
+		// Install the given file as an FD.
+		file, err := host.NewFile(ctx, newFD, mounter)
+		if err != nil {
+			syscall.Close(newFD)
+			return err
+		}
+		defer file.DecRef()
+		if err := fdm.NewFDAt(kdefs.FD(appFD), file, kernel.FDFlags{}, l); err != nil {
+			return err
+		}
+	}
+
+	// Start the new task.
+	newTG, err := proc.Kernel.CreateProcess(initArgs)
+	if err != nil {
+		return err
+	}
+
+	// If we're supposed to detach, don't wait for the process to exit.
+	if args.Detach {
+		*waitStatus = 0
+		return nil
+	}
+
+	// Wait for completion.
+	newTG.WaitExited()
+	*waitStatus = newTG.ExitStatus().Status()
+	return nil
+}
+
+// PsArgs is the set of arguments to ps.
+type PsArgs struct {
+	// JSON will force calls to Ps to return the result as a JSON payload.
+	JSON bool
+}
+
+// Ps provides a process listing for the running kernel.
+func (proc *Proc) Ps(args *PsArgs, out *string) error {
+	var p []*Process
+	if e := Processes(proc.Kernel, &p); e != nil {
+		return e
+	}
+	if !args.JSON {
+		*out = ProcessListToTable(p)
+	} else {
+		s, e := ProcessListToJSON(p)
+		if e != nil {
+			return e
+		}
+		*out = s
+	}
+	return nil
+}
+
+// Process contains information about a single process in a Sandbox.
+// TODO: Implement TTY field.
+type Process struct {
+	UID auth.KUID       `json:"uid"`
+	PID kernel.ThreadID `json:"pid"`
+	// Parent PID
+	PPID kernel.ThreadID `json:"ppid"`
+	// Processor utilization
+	C int32 `json:"c"`
+	// Start time
+	STime string `json:"stime"`
+	// CPU time
+	Time string `json:"time"`
+	// Executable shortname (e.g. "sh" for /bin/sh)
+	Cmd string `json:"cmd"`
+}
+
+// ProcessListToTable prints a table with the following format:
+// UID       PID       PPID      C         STIME     TIME       CMD
+// 0         1         0         0         14:04     505262ns   tail
+func ProcessListToTable(pl []*Process) string {
+	var buf bytes.Buffer
+	tw := tabwriter.NewWriter(&buf, 10, 1, 3, ' ', 0)
+	fmt.Fprint(tw, "UID\tPID\tPPID\tC\tSTIME\tTIME\tCMD")
+	for _, d := range pl {
+		fmt.Fprintf(tw, "\n%d\t%d\t%d\t%d\t%s\t%s\t%s",
+			d.UID,
+			d.PID,
+			d.PPID,
+			d.C,
+			d.STime,
+			d.Time,
+			d.Cmd)
+	}
+	tw.Flush()
+	return buf.String()
+}
+
+// ProcessListToJSON will return the JSON representation of ps.
+func ProcessListToJSON(pl []*Process) (string, error) {
+	b, err := json.Marshal(pl)
+	if err != nil {
+		return "", fmt.Errorf("couldn't marshal process list %v: %v", pl, err)
+	}
+	return string(b), nil
+}
+
+// PrintPIDsJSON prints a JSON object containing only the PIDs in pl. This
+// behavior is the same as runc's.
+func PrintPIDsJSON(pl []*Process) (string, error) {
+	pids := make([]kernel.ThreadID, 0, len(pl))
+	for _, d := range pl {
+		pids = append(pids, d.PID)
+	}
+	b, err := json.Marshal(pids)
+	if err != nil {
+		return "", fmt.Errorf("couldn't marshal PIDs %v: %v", pids, err)
+	}
+	return string(b), nil
+}
+
+// Processes retrieves information about processes running in the sandbox.
+func Processes(k *kernel.Kernel, out *[]*Process) error {
+	ts := k.TaskSet()
+	now := k.RealtimeClock().Now()
+	for _, tg := range ts.Root.ThreadGroups() {
+		pid := ts.Root.IDOfThreadGroup(tg)
+		// If tg has already been reaped ignore it.
+		if pid == 0 {
+			continue
+		}
+
+		*out = append(*out, &Process{
+			UID: tg.Leader().Credentials().EffectiveKUID,
+			PID: pid,
+			// If Parent is null (i.e. tg is the init process), PPID will be 0.
+			PPID:  ts.Root.IDOfTask(tg.Leader().Parent()),
+			STime: formatStartTime(now, tg.Leader().StartTime()),
+			C:     percentCPU(tg.CPUStats(), tg.Leader().StartTime(), now),
+			Time:  tg.CPUStats().SysTime.String(),
+			Cmd:   tg.Leader().Name(),
+		})
+	}
+	return nil
+}
+
+// formatStartTime formats startTime depending on the current time:
+// - If startTime was today, HH:MM is used.
+// - If startTime was not today but was this year, MonDD is used (e.g. Jan02)
+// - If startTime was not this year, the year is used.
+func formatStartTime(now, startTime ktime.Time) string {
+	nowS, nowNs := now.Unix()
+	n := time.Unix(nowS, nowNs)
+	startTimeS, startTimeNs := startTime.Unix()
+	st := time.Unix(startTimeS, startTimeNs)
+	format := "15:04"
+	if st.YearDay() != n.YearDay() {
+		format = "Jan02"
+	}
+	if st.Year() != n.Year() {
+		format = "2006"
+	}
+	return st.Format(format)
+}
+
+func percentCPU(stats usage.CPUStats, startTime, now ktime.Time) int32 {
+	// Note: In procps, there is an option to include child CPU stats. As
+	// it is disabled by default, we do not include them.
+	total := stats.UserTime + stats.SysTime
+	lifetime := now.Sub(startTime)
+	if lifetime <= 0 {
+		return 0
+	}
+	percentCPU := total * 100 / lifetime
+	// Cap at 99% since procps does the same.
+	if percentCPU > 99 {
+		percentCPU = 99
+	}
+	return int32(percentCPU)
+}
diff --git a/pkg/sentry/control/proc_test.go b/pkg/sentry/control/proc_test.go
new file mode 100644
index 000000000..18286496f
--- /dev/null
+++ b/pkg/sentry/control/proc_test.go
@@ -0,0 +1,164 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package control
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+)
+
+func init() {
+	log.SetLevel(log.Debug)
+}
+
+// Tests that ProcessData.Table() prints with the correct format.
+func TestProcessListTable(t *testing.T) {
+	testCases := []struct {
+		pl       []*Process
+		expected string
+	}{
+		{
+			pl:       []*Process{},
+			expected: "UID       PID       PPID      C         STIME     TIME      CMD",
+		},
+		{
+			pl: []*Process{
+				&Process{
+					UID:   0,
+					PID:   0,
+					PPID:  0,
+					C:     0,
+					STime: "0",
+					Time:  "0",
+					Cmd:   "zero",
+				},
+				&Process{
+					UID:   1,
+					PID:   1,
+					PPID:  1,
+					C:     1,
+					STime: "1",
+					Time:  "1",
+					Cmd:   "one",
+				},
+			},
+			expected: `UID       PID       PPID      C         STIME     TIME      CMD
+0         0         0         0         0         0         zero
+1         1         1         1         1         1         one`,
+		},
+	}
+
+	for _, tc := range testCases {
+		output := ProcessListToTable(tc.pl)
+
+		if tc.expected != output {
+			t.Errorf("PrintTable(%v): got:\n%s\nwant:\n%s", tc.pl, output, tc.expected)
+		}
+	}
+}
+
+func TestProcessListJSON(t *testing.T) {
+	testCases := []struct {
+		pl       []*Process
+		expected string
+	}{
+		{
+			pl:       []*Process{},
+			expected: "[]",
+		},
+		{
+			pl: []*Process{
+				&Process{
+					UID:   0,
+					PID:   0,
+					PPID:  0,
+					C:     0,
+					STime: "0",
+					Time:  "0",
+					Cmd:   "zero",
+				},
+				&Process{
+					UID:   1,
+					PID:   1,
+					PPID:  1,
+					C:     1,
+					STime: "1",
+					Time:  "1",
+					Cmd:   "one",
+				},
+			},
+			expected: "[0,1]",
+		},
+	}
+
+	for _, tc := range testCases {
+		output, err := PrintPIDsJSON(tc.pl)
+		if err != nil {
+			t.Errorf("failed to generate JSON: %v", err)
+		}
+
+		if tc.expected != output {
+			t.Errorf("PrintJSON(%v): got:\n%s\nwant:\n%s", tc.pl, output, tc.expected)
+		}
+	}
+}
+
+func TestPercentCPU(t *testing.T) {
+	testCases := []struct {
+		stats     usage.CPUStats
+		startTime ktime.Time
+		now       ktime.Time
+		expected  int32
+	}{
+		{
+			// Verify that 100% use is capped at 99.
+			stats:     usage.CPUStats{UserTime: 1e9, SysTime: 1e9},
+			startTime: ktime.FromNanoseconds(7e9),
+			now:       ktime.FromNanoseconds(9e9),
+			expected:  99,
+		},
+		{
+			// Verify that if usage > lifetime, we get at most 99%
+			// usage.
+			stats:     usage.CPUStats{UserTime: 2e9, SysTime: 2e9},
+			startTime: ktime.FromNanoseconds(7e9),
+			now:       ktime.FromNanoseconds(9e9),
+			expected:  99,
+		},
+		{
+			// Verify that 50% usage is reported correctly.
+			stats:     usage.CPUStats{UserTime: 1e9, SysTime: 1e9},
+			startTime: ktime.FromNanoseconds(12e9),
+			now:       ktime.FromNanoseconds(16e9),
+			expected:  50,
+		},
+		{
+			// Verify that 0% usage is reported correctly.
+			stats:     usage.CPUStats{UserTime: 0, SysTime: 0},
+			startTime: ktime.FromNanoseconds(12e9),
+			now:       ktime.FromNanoseconds(14e9),
+			expected:  0,
+		},
+	}
+
+	for _, tc := range testCases {
+		if pcpu := percentCPU(tc.stats, tc.startTime, tc.now); pcpu != tc.expected {
+			t.Errorf("percentCPU(%v, %v, %v): got %d, want %d", tc.stats, tc.startTime, tc.now, pcpu, tc.expected)
+		}
+	}
+}
diff --git a/pkg/sentry/device/BUILD b/pkg/sentry/device/BUILD
new file mode 100644
index 000000000..1a8b461ba
--- /dev/null
+++ b/pkg/sentry/device/BUILD
@@ -0,0 +1,18 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "device",
+    srcs = ["device.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/device",
+    visibility = ["//pkg/sentry:internal"],
+    deps = ["//pkg/abi/linux"],
+)
+
+go_test(
+    name = "device_test",
+    size = "small",
+    srcs = ["device_test.go"],
+    embed = [":device"],
+)
diff --git a/pkg/sentry/device/device.go b/pkg/sentry/device/device.go
new file mode 100644
index 000000000..a5514c72f
--- /dev/null
+++ b/pkg/sentry/device/device.go
@@ -0,0 +1,193 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package device defines reserved virtual kernel devices and structures
+// for managing them.
+//
+// Saving and restoring devices is not necessary if the devices are initialized
+// as package global variables. Package initialization happens in a single goroutine
+// and in a deterministic order, so minor device numbers will be assigned in the
+// same order as packages are loaded.
+package device
+
+import (
+	"bytes"
+	"fmt"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// ID identifies a device.
+type ID struct {
+	Major uint64
+	Minor uint64
+}
+
+// DeviceID formats a major and minor device number into a standard device number.
+func (i *ID) DeviceID() uint64 {
+	return uint64(linux.MakeDeviceID(uint16(i.Major), uint32(i.Minor)))
+}
+
+// nextAnonDeviceMinor is the next minor number for a new anonymous device.
+// Must be accessed atomically.
+var nextAnonDeviceMinor uint64
+
+// NewAnonDevice creates a new anonymous device. Packages that require an anonymous
+// device should initialize the device in a global variable in a file called device.go:
+//
+// var myDevice = device.NewAnonDevice()
+func NewAnonDevice() *Device {
+	return &Device{
+		ID: newAnonID(),
+	}
+}
+
+// NewAnonMultiDevice creates a new multi-keyed anonymous device. Packages that require
+// a multi-key anonymous device should initialize the device in a global variable in a
+// file called device.go:
+//
+// var myDevice = device.NewAnonMultiDevice()
+func NewAnonMultiDevice() *MultiDevice {
+	return &MultiDevice{
+		ID: newAnonID(),
+	}
+}
+
+// newAnonID assigns a major and minor number to an anonymous device ID.
+func newAnonID() ID {
+	return ID{
+		// Anon devices always have a major number of 0.
+		Major: 0,
+		// Use the next minor number.
+		Minor: atomic.AddUint64(&nextAnonDeviceMinor, 1),
+	}
+}
+
+// Device is a simple virtual kernel device.
+type Device struct {
+	ID
+
+	// last is the last generated inode.
+	last uint64
+}
+
+// NextIno generates a new inode number
+func (d *Device) NextIno() uint64 {
+	return atomic.AddUint64(&d.last, 1)
+}
+
+// MultiDeviceKey provides a hashable key for a MultiDevice. The key consists
+// of a raw device and inode for a resource, which must consistently identify
+// the unique resource.  It may optionally include a secondary device if
+// appropriate.
+//
+// Note that using the path is not enough, because filesystems may rename a file
+// to a different backing resource, at which point the path points to a different
+// entity.  Using only the inode is also not enough because the inode is assumed
+// to be unique only within the device on which the resource exists.
+type MultiDeviceKey struct {
+	Device          uint64
+	SecondaryDevice string
+	Inode           uint64
+}
+
+// String stringifies the key.
+func (m MultiDeviceKey) String() string {
+	return fmt.Sprintf("key{device: %d, sdevice: %s, inode: %d}", m.Device, m.SecondaryDevice, m.Inode)
+}
+
+// MultiDevice allows for remapping resources that come from a variety of raw
+// devices into a single device.  The device ID should be one of the static
+// Device IDs above and cannot be reused.
+type MultiDevice struct {
+	ID
+
+	mu     sync.Mutex
+	last   uint64
+	cache  map[MultiDeviceKey]uint64
+	rcache map[uint64]MultiDeviceKey
+}
+
+// String stringifies MultiDevice.
+func (m *MultiDevice) String() string {
+	buf := bytes.NewBuffer(nil)
+	buf.WriteString("cache{")
+	for k, v := range m.cache {
+		buf.WriteString(fmt.Sprintf("%s -> %d, ", k, v))
+	}
+	buf.WriteString("}")
+	return buf.String()
+}
+
+// Map maps a raw device and inode into the inode space of MultiDevice,
+// returning a virtualized inode.  Raw devices and inodes can be reused;
+// in this case, the same virtual inode will be returned.
+func (m *MultiDevice) Map(key MultiDeviceKey) uint64 {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if m.cache == nil {
+		m.cache = make(map[MultiDeviceKey]uint64)
+		m.rcache = make(map[uint64]MultiDeviceKey)
+	}
+
+	id, ok := m.cache[key]
+	if ok {
+		return id
+	}
+	// Step over reserved entries that may have been loaded.
+	idx := m.last + 1
+	for {
+		if _, ok := m.rcache[idx]; !ok {
+			break
+		}
+		idx++
+	}
+	// We found a non-reserved entry, use it.
+	m.last = idx
+	m.cache[key] = m.last
+	m.rcache[m.last] = key
+	return m.last
+}
+
+// Load loads a raw device and inode into MultiDevice inode mappings
+// with value as the virtual inode.
+//
+// By design, inodes start from 1 and continue until max uint64.  This means
+// that the zero value, which is often the uninitialized value, can be rejected
+// as invalid.
+func (m *MultiDevice) Load(key MultiDeviceKey, value uint64) bool {
+	// Reject the uninitialized value; see comment above.
+	if value == 0 {
+		return false
+	}
+
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	if m.cache == nil {
+		m.cache = make(map[MultiDeviceKey]uint64)
+		m.rcache = make(map[uint64]MultiDeviceKey)
+	}
+
+	// Cache value at key.
+	m.cache[key] = value
+
+	// Prevent value from being used by new inode mappings.
+	m.rcache[value] = key
+
+	return true
+}
diff --git a/pkg/sentry/device/device_test.go b/pkg/sentry/device/device_test.go
new file mode 100644
index 000000000..dfec45046
--- /dev/null
+++ b/pkg/sentry/device/device_test.go
@@ -0,0 +1,59 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package device
+
+import (
+	"testing"
+)
+
+func TestMultiDevice(t *testing.T) {
+	device := &MultiDevice{}
+
+	// Check that Load fails to install virtual inodes that are
+	// uninitialized.
+	if device.Load(MultiDeviceKey{}, 0) {
+		t.Fatalf("got load of invalid virtual inode 0, want unsuccessful")
+	}
+
+	inode := device.Map(MultiDeviceKey{})
+
+	// Assert that the same raw device and inode map to
+	// a consistent virtual inode.
+	if i := device.Map(MultiDeviceKey{}); i != inode {
+		t.Fatalf("got inode %d, want %d in %s", i, inode, device)
+	}
+
+	// Assert that a new inode or new device does not conflict.
+	if i := device.Map(MultiDeviceKey{Device: 0, Inode: 1}); i == inode {
+		t.Fatalf("got reused inode %d, want new distinct inode in %s", i, device)
+	}
+	last := device.Map(MultiDeviceKey{Device: 1, Inode: 0})
+	if last == inode {
+		t.Fatalf("got reused inode %d, want new distinct inode in %s", last, device)
+	}
+
+	// Virtual is the virtual inode we want to load.
+	virtual := last + 1
+
+	// Assert that we can load a virtual inode at a new place.
+	if !device.Load(MultiDeviceKey{Device: 0, Inode: 2}, virtual) {
+		t.Fatalf("got load of virtual inode %d failed, want success in %s", virtual, device)
+	}
+
+	// Assert that the next inode skips over the loaded one.
+	if i := device.Map(MultiDeviceKey{Device: 0, Inode: 3}); i != virtual+1 {
+		t.Fatalf("got inode %d, want %d in %s", i, virtual+1, device)
+	}
+}
diff --git a/pkg/sentry/fs/BUILD b/pkg/sentry/fs/BUILD
new file mode 100644
index 000000000..9b7264753
--- /dev/null
+++ b/pkg/sentry/fs/BUILD
@@ -0,0 +1,154 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "fs_state",
+    srcs = [
+        "attr.go",
+        "dentry.go",
+        "dirent.go",
+        "dirent_cache.go",
+        "dirent_list.go",
+        "dirent_state.go",
+        "file.go",
+        "file_overlay.go",
+        "file_state.go",
+        "filesystems.go",
+        "flags.go",
+        "inode.go",
+        "inode_inotify.go",
+        "inode_operations.go",
+        "inode_overlay.go",
+        "inotify.go",
+        "inotify_event.go",
+        "inotify_watch.go",
+        "mock.go",
+        "mount.go",
+        "mount_overlay.go",
+        "mount_state.go",
+        "mounts.go",
+        "overlay.go",
+        "path.go",
+    ],
+    out = "fs_state.go",
+    package = "fs",
+)
+
+go_library(
+    name = "fs",
+    srcs = [
+        "attr.go",
+        "context.go",
+        "copy_up.go",
+        "dentry.go",
+        "dirent.go",
+        "dirent_cache.go",
+        "dirent_list.go",
+        "dirent_state.go",
+        "file.go",
+        "file_operations.go",
+        "file_overlay.go",
+        "file_state.go",
+        "filesystems.go",
+        "flags.go",
+        "fs.go",
+        "fs_state.go",
+        "inode.go",
+        "inode_inotify.go",
+        "inode_operations.go",
+        "inode_overlay.go",
+        "inotify.go",
+        "inotify_event.go",
+        "inotify_watch.go",
+        "mock.go",
+        "mount.go",
+        "mount_overlay.go",
+        "mount_state.go",
+        "mounts.go",
+        "offset.go",
+        "overlay.go",
+        "path.go",
+        "restore.go",
+        "save.go",
+        "seek.go",
+        "sync.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/ilist",
+        "//pkg/log",
+        "//pkg/p9",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_template_instance(
+    name = "dirent_list",
+    out = "dirent_list.go",
+    package = "fs",
+    prefix = "dirent",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*Dirent",
+    },
+)
+
+go_test(
+    name = "fs_x_test",
+    size = "small",
+    srcs = [
+        "copy_up_test.go",
+        "file_overlay_test.go",
+        "inode_overlay_test.go",
+        "mounts_test.go",
+    ],
+    deps = [
+        ":fs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs/ramfs/test",
+        "//pkg/sentry/fs/tmpfs",
+        "//pkg/sentry/usermem",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "fs_test",
+    size = "small",
+    srcs = [
+        "dirent_cache_test.go",
+        "dirent_refs_test.go",
+        "file_test.go",
+        "mount_test.go",
+        "path_test.go",
+    ],
+    embed = [":fs"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+    ],
+)
diff --git a/pkg/sentry/fs/README.md b/pkg/sentry/fs/README.md
new file mode 100644
index 000000000..898271ee8
--- /dev/null
+++ b/pkg/sentry/fs/README.md
@@ -0,0 +1,217 @@
+This package provides an implementation of the Linux virtual filesystem.
+
+[TOC]
+
+## Overview
+
+-   An `fs.Dirent` caches an `fs.Inode` in memory at a path in the VFS, giving
+    the `fs.Inode` a relative position with respect to other `fs.Inode`s.
+
+-   If an `fs.Dirent` is referenced by two file descriptors, then those file
+    descriptors are coherent with each other: they depend on the same
+    `fs.Inode`.
+
+-   A mount point is an `fs.Dirent` for which `fs.Dirent.mounted` is true. It
+    exposes the root of a mounted filesystem.
+
+-   The `fs.Inode` produced by a registered filesystem on mount(2) owns an
+    `fs.MountedFilesystem` from which other `fs.Inode`s will be looked up. For a
+    remote filesystem, the `fs.MountedFilesystem` owns the connection to that
+    remote filesystem.
+
+-   In general:
+
+```
+fs.Inode <------------------------------
+|                                      |
+|                                      |
+produced by                            |
+exactly one                            |
+|                             responsible for the
+|                             virtual identity of
+v                                      |
+fs.MountedFilesystem -------------------
+```
+
+Glossary:
+
+-   VFS: virtual filesystem.
+
+-   inode: a virtual file object holding a cached view of a file on a backing
+    filesystem (includes metadata and page caches).
+
+-   superblock: the virtual state of a mounted filesystem (e.g. the virtual
+    inode number set).
+
+-   mount namespace: a view of the mounts under a root (during path traversal,
+    the VFS makes visible/follows the mount point that is in the current task's
+    mount namespace).
+
+## Save and restore
+
+An application's hard dependencies on filesystem state can be broken down into
+two categories:
+
+-   The state necessary to execute a traversal on or view the *virtual*
+    filesystem hierarchy, regardless of what files an application has open.
+
+-   The state necessary to represent open files.
+
+The first is always necessary to save and restore. An application may never have
+any open file descriptors, but across save and restore it should see a coherent
+view of any mount namespace. NOTE: Currently only one "initial"
+mount namespace is supported.
+
+The second is so that system calls across save and restore are coherent with
+each other (e.g. so that unintended re-reads or overwrites do not occur).
+
+Specifically this state is:
+
+-   An `fs.MountManager` containing mount points.
+
+-   A `kernel.FDMap` containing pointers to open files.
+
+Anything else managed by the VFS that can be easily loaded into memory from a
+filesystem is synced back to those filesystems and is no saved. Examples are
+pages in page caches used for optimizations (i.e. readahead and writeback), and
+directory entries used to accelerate path lookups.
+
+### Mount points
+
+Saving and restoring a mount point means saving and restoring:
+
+-   The root of the mounted filesystem.
+
+-   Mount flags, which control how the VFS interacts with the mounted
+    filesystem.
+
+-   Any relevant metadata about the mounted filesystem.
+
+-   All `fs.Inode`s referenced by the application that reside under the mount
+    point.
+
+`fs.MountedFilesystem` is metadata about a filesystem that is mounted. It is
+referenced by every `fs.Inode` loaded into memory under the mount point
+including the `fs.Inode` of the mount point itself. The `fs.MountedFilesystem`
+maps file objects on the filesystem to a virtualized `fs.Inode` number and vice
+versa.
+
+To restore all `fs.Inode`s under a given mount point, each `fs.Inode` leverages
+its dependency on an `fs.MountedFilesystem`. Since the `fs.MountedFilesystem`
+knows how an `fs.Inode` maps to a file object on a backing filesystem, this
+mapping can be trivially consulted by each `fs.Inode` when the `fs.Inode` is
+restored.
+
+In detail, a mount point is saved in two steps:
+
+-   First, after the kernel is paused but before state.Save, we walk all mount
+    namespaces and install a mapping from `fs.Inode` numbers to file paths
+    relative to the root of the mounted filesystem in each
+    `fs.MountedFilesystem`. This is subsequently called the set of `fs.Inode`
+    mappings.
+
+-   Second, during state.Save, each `fs.MountedFilesystem` decides whether to
+    save the set of `fs.Inode` mappings. In-memory filesystems, like tmpfs, have
+    no need to save a set of `fs.Inode` mappings, since the `fs.Inode`s can be
+    entirely encoded in state file. Each `fs.MountedFilesystem` also optionally
+    saves the device name from when the filesystem was originally mounted. Each
+    `fs.Inode` saves its virtual identifier and a reference to a
+    `fs.MountedFilesystem`.
+
+A mount point is restored in two steps:
+
+-   First, before state.Load, all mount configurations are stored in a global
+    `fs.RestoreEnvironment`. This tells us what mount points the user wants to
+    restore and how to re-establish pointers to backing filesystems.
+
+-   Second, during state.Load, each `fs.MountedFilesystem` optionally searches
+    for a mount in the `fs.RestoreEnvironment` that matches its saved device
+    name. The `fs.MountedFilesystem` then restablishes a pointer to the root of
+    the mounted filesystem. For example, the mount specification provides the
+    network connection for a mounted remote filesystem client to communicate
+    with its remote file server. The `fs.MountedFilesystem` also trivially loads
+    its set of `fs.Inode` mappings. When an `fs.Inode` is encountered, the
+    `fs.Inode` loads its virtual identifier and its reference a
+    `fs.MountedFilesystem`. It uses the `fs.MountedFilesystem` to obtain the
+    root of the mounted filesystem and the `fs.Inode` mappings to obtain the
+    relative file path to its data. With these, the `fs.Inode` re-establishes a
+    pointer to its file object.
+
+A mount point can trivially restore its `fs.Inode`s in parallel since
+`fs.Inode`s have a restore dependency on their `fs.MountedFilesystem` and not on
+each other.
+
+### Open files
+
+An `fs.File` references the following filesystem objects:
+
+```go
+fs.File -> fs.Dirent -> fs.Inode -> fs.MountedFilesystem
+```
+
+The `fs.Inode` is restored using its `fs.MountedFilesystem`. The [Mount
+points](#mount-points) section above describes how this happens in detail. The
+`fs.Dirent` restores its pointer to an `fs.Inode`, pointers to parent and
+children `fs.Dirents`, and the basename of the file.
+
+Otherwise an `fs.File` restores flags, an offset, and a unique identifier (only
+used internally).
+
+It may use the `fs.Inode`, which it indirectly holds a reference on through the
+`fs.Dirent`, to restablish an open file handle on the backing filesystem (e.g.
+to continue reading and writing).
+
+## Overlay
+
+The overlay implementation in the fs package takes Linux overlayfs as a frame of
+reference but corrects for several POSIX consistency errors.
+
+In Linux overlayfs, the `struct inode` used for reading and writing to the same
+file may be different. This is because the `struct inode` is dissociated with
+the process of copying up the file from the upper to the lower directory. Since
+flock(2) and fcntl(2) locks, inotify(7) watches, page caches, and a file's
+identity are all stored directly or indirectly off the `struct inode`, these
+properties of the `struct inode` may be stale after the first modification. This
+can lead to file locking bugs, missed inotify events, and inconsistent data in
+shared memory mappings of files, to name a few problems.
+
+The fs package maintains a single `fs.Inode` to represent a directory entry in
+an overlay and defines operations on this `fs.Inode` which synchronize with the
+copy up process. This achieves several things:
+
++   File locks, inotify watches, and the identity of the file need not be copied
+    at all.
+
++   Memory mappings of files coordinate with the copy up process so that if a
+    file in the lower directory is memory mapped, all references to it are
+    invalidated, forcing the application to re-fault on memory mappings of the
+    file under the upper directory.
+
+The `fs.Inode` holds metadata about files in the upper and/or lower directories
+via an `fs.overlayEntry`. The `fs.overlayEntry` implements the `fs.Mappable`
+interface. It multiplexes between upper and lower directory memory mappings and
+stores a copy of memory references so they can be transferred to the upper
+directory `fs.Mappable` when the file is copied up.
+
+The `fs.Inode` also holds a reference to a `fs.MountedFilesystem` that
+normalizes across the mounted filesystem state of the upper and lower
+directories.
+
+When a file is copied from the lower to the upper directory, attempts to
+interact with the file block until the copy completes. All copying synchronizes
+with rename(2).
+
+## Future Work
+
+### Overlay
+
+When a file is copied from a lower directory to an upper directory, several
+locks are taken: the global renamuMu and the copyMu of the `fs.Inode` being
+copied. This blocks operations on the file, including fault handling of memory
+mappings. Performance could be improved by copying files into a temporary
+directory that resides on the same filesystem as the upper directory and doing
+an atomic rename, holding locks only during the rename operation.
+
+Additionally files are copied up synchronously. For large files, this causes a
+noticeable latency. Performance could be improved by pipelining copies at
+non-overlapping file offsets.
diff --git a/pkg/sentry/fs/anon/BUILD b/pkg/sentry/fs/anon/BUILD
new file mode 100644
index 000000000..6b18aee47
--- /dev/null
+++ b/pkg/sentry/fs/anon/BUILD
@@ -0,0 +1,21 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "anon",
+    srcs = [
+        "anon.go",
+        "device.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/anon/anon.go b/pkg/sentry/fs/anon/anon.go
new file mode 100644
index 000000000..ddc2c0985
--- /dev/null
+++ b/pkg/sentry/fs/anon/anon.go
@@ -0,0 +1,46 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package anon implements an anonymous inode, useful for implementing
+// inodes for pseudo filesystems.
+package anon
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// NewInode constructs an anonymous Inode that is not associated
+// with any real filesystem. Some types depend on completely pseudo
+// "anon" inodes (eventfds, epollfds, etc).
+func NewInode(ctx context.Context) *fs.Inode {
+	return fs.NewInode(fsutil.NewSimpleInodeOperations(fsutil.InodeSimpleAttributes{
+		FSType: linux.ANON_INODE_FS_MAGIC,
+		UAttr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: fs.FileOwnerFromContext(ctx),
+			Perms: fs.FilePermissions{
+				User: fs.PermMask{Read: true, Write: true},
+			},
+			Links: 1,
+		}),
+	}), fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{}), fs.StableAttr{
+		Type:      fs.Anonymous,
+		DeviceID:  PseudoDevice.DeviceID(),
+		InodeID:   PseudoDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+	})
+}
diff --git a/pkg/sentry/fs/anon/device.go b/pkg/sentry/fs/anon/device.go
new file mode 100644
index 000000000..1c666729c
--- /dev/null
+++ b/pkg/sentry/fs/anon/device.go
@@ -0,0 +1,22 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package anon
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+)
+
+// PseudoDevice is the device on which all anonymous inodes reside.
+var PseudoDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/ashmem/BUILD b/pkg/sentry/fs/ashmem/BUILD
new file mode 100644
index 000000000..e20e22a0f
--- /dev/null
+++ b/pkg/sentry/fs/ashmem/BUILD
@@ -0,0 +1,83 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+go_stateify(
+    name = "ashmem_state",
+    srcs = [
+        "area.go",
+        "device.go",
+        "pin_board.go",
+        "uint64_range.go",
+        "uint64_set.go",
+    ],
+    out = "ashmem_state.go",
+    package = "ashmem",
+)
+
+go_library(
+    name = "ashmem",
+    srcs = [
+        "area.go",
+        "ashmem_state.go",
+        "device.go",
+        "pin_board.go",
+        "uint64_range.go",
+        "uint64_set.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ashmem",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/tmpfs",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+    ],
+)
+
+go_test(
+    name = "ashmem_test",
+    size = "small",
+    srcs = ["pin_board_test.go"],
+    embed = [":ashmem"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/usermem",
+    ],
+)
+
+go_template_instance(
+    name = "uint64_range",
+    out = "uint64_range.go",
+    package = "ashmem",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "uint64",
+    },
+)
+
+go_template_instance(
+    name = "uint64_set",
+    out = "uint64_set.go",
+    package = "ashmem",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "Range",
+        "Value": "noValue",
+        "Functions": "setFunctions",
+    },
+)
diff --git a/pkg/sentry/fs/ashmem/area.go b/pkg/sentry/fs/ashmem/area.go
new file mode 100644
index 000000000..e4f76f0d0
--- /dev/null
+++ b/pkg/sentry/fs/ashmem/area.go
@@ -0,0 +1,313 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ashmem
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/tmpfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	// namePrefix is the name prefix assumed and forced by the Linux implementation.
+	namePrefix = "dev/ashmem"
+
+	// nameLen is the maximum name length.
+	nameLen = 256
+)
+
+// Area implements fs.FileOperations.
+type Area struct {
+	fsutil.NoFsync
+	fsutil.DeprecatedFileOperations
+	fsutil.NotDirReaddir
+
+	ad *Device
+
+	// mu protects fields below.
+	mu        sync.Mutex `state:"nosave"`
+	tmpfsFile *fs.File
+	name      string
+	size      uint64
+	perms     usermem.AccessType
+	pb        *PinBoard
+}
+
+// Release implements fs.FileOperations.Release.
+func (a *Area) Release() {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	if a.tmpfsFile != nil {
+		a.tmpfsFile.DecRef()
+		a.tmpfsFile = nil
+	}
+}
+
+// Seek implements fs.FileOperations.Seek.
+func (a *Area) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	if a.size == 0 {
+		return 0, syserror.EINVAL
+	}
+	if a.tmpfsFile == nil {
+		return 0, syserror.EBADF
+	}
+	return a.tmpfsFile.FileOperations.Seek(ctx, file, whence, offset)
+}
+
+// Read implements fs.FileOperations.Read.
+func (a *Area) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	if a.size == 0 {
+		return 0, nil
+	}
+	if a.tmpfsFile == nil {
+		return 0, syserror.EBADF
+	}
+	return a.tmpfsFile.FileOperations.Read(ctx, file, dst, offset)
+}
+
+// Write implements fs.FileOperations.Write.
+func (a *Area) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	return 0, syserror.ENOSYS
+}
+
+// Flush implements fs.FileOperations.Flush.
+func (a *Area) Flush(ctx context.Context, file *fs.File) error {
+	return nil
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (a *Area) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	if a.size == 0 {
+		return syserror.EINVAL
+	}
+
+	if !a.perms.SupersetOf(opts.Perms) {
+		return syserror.EPERM
+	}
+	opts.MaxPerms = opts.MaxPerms.Intersect(a.perms)
+
+	if a.tmpfsFile == nil {
+		p := platform.FromContext(ctx)
+		if p == nil {
+			return syserror.ENOMEM
+		}
+		tmpfsInodeOps := tmpfs.NewInMemoryFile(ctx, usage.Tmpfs, fs.UnstableAttr{}, p)
+		// This is not backed by a real filesystem, so we pass in nil.
+		tmpfsInode := fs.NewInode(tmpfsInodeOps, fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{}), fs.StableAttr{})
+		dirent := fs.NewDirent(tmpfsInode, namePrefix+"/"+a.name)
+		tmpfsFile, err := tmpfsInode.GetFile(ctx, dirent, fs.FileFlags{Read: true, Write: true})
+		// Drop the extra reference on the Dirent.
+		dirent.DecRef()
+
+		if err != nil {
+			return err
+		}
+
+		// Truncate to the size set by ASHMEM_SET_SIZE ioctl.
+		err = tmpfsInodeOps.Truncate(ctx, tmpfsInode, int64(a.size))
+		if err != nil {
+			return err
+		}
+		a.tmpfsFile = tmpfsFile
+		a.pb = NewPinBoard()
+	}
+
+	return a.tmpfsFile.ConfigureMMap(ctx, opts)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (a *Area) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// Switch on ioctl request.
+	switch args[1].Uint() {
+	case linux.AshmemSetNameIoctl:
+		name, err := usermem.CopyStringIn(ctx, io, args[2].Pointer(), nameLen-1, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		if err != nil {
+			return 0, err
+		}
+
+		a.mu.Lock()
+		defer a.mu.Unlock()
+
+		// Cannot set name for already mapped ashmem.
+		if a.tmpfsFile != nil {
+			return 0, syserror.EINVAL
+		}
+		a.name = name
+		return 0, nil
+
+	case linux.AshmemGetNameIoctl:
+		a.mu.Lock()
+		var local []byte
+		if a.name != "" {
+			nameLen := len([]byte(a.name))
+			local = make([]byte, nameLen, nameLen+1)
+			copy(local, []byte(a.name))
+			local = append(local, 0)
+		} else {
+			nameLen := len([]byte(namePrefix))
+			local = make([]byte, nameLen, nameLen+1)
+			copy(local, []byte(namePrefix))
+			local = append(local, 0)
+		}
+		a.mu.Unlock()
+
+		if _, err := io.CopyOut(ctx, args[2].Pointer(), local, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, syserror.EFAULT
+		}
+		return 0, nil
+
+	case linux.AshmemSetSizeIoctl:
+		a.mu.Lock()
+		defer a.mu.Unlock()
+
+		// Cannot set size for already mapped ashmem.
+		if a.tmpfsFile != nil {
+			return 0, syserror.EINVAL
+		}
+		a.size = uint64(args[2].SizeT())
+		return 0, nil
+
+	case linux.AshmemGetSizeIoctl:
+		return uintptr(a.size), nil
+
+	case linux.AshmemPinIoctl, linux.AshmemUnpinIoctl, linux.AshmemGetPinStatusIoctl:
+		// Locking and unlocking is ok since once tmpfsFile is set, it won't be nil again
+		// even after unmapping! Unlocking is needed in order to avoid a deadlock on
+		// usermem.CopyObjectIn.
+
+		// Cannot execute pin-related ioctls before mapping.
+		a.mu.Lock()
+		if a.tmpfsFile == nil {
+			a.mu.Unlock()
+			return 0, syserror.EINVAL
+		}
+		a.mu.Unlock()
+
+		var pin linux.AshmemPin
+		_, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pin, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		if err != nil {
+			return 0, syserror.EFAULT
+		}
+
+		a.mu.Lock()
+		defer a.mu.Unlock()
+		return a.pinOperation(pin, args[1].Uint())
+
+	case linux.AshmemPurgeAllCachesIoctl:
+		return 0, nil
+
+	case linux.AshmemSetProtMaskIoctl:
+		prot := uint64(args[2].ModeT())
+		perms := usermem.AccessType{
+			Read:    prot&linux.PROT_READ != 0,
+			Write:   prot&linux.PROT_WRITE != 0,
+			Execute: prot&linux.PROT_EXEC != 0,
+		}
+
+		a.mu.Lock()
+		defer a.mu.Unlock()
+
+		// Can only narrow prot mask.
+		if !a.perms.SupersetOf(perms) {
+			return 0, syserror.EINVAL
+		}
+
+		// TODO: If personality flag
+		// READ_IMPLIES_EXEC is set, set PROT_EXEC if PORT_READ is set.
+
+		a.perms = perms
+		return 0, nil
+
+	case linux.AshmemGetProtMaskIoctl:
+		return uintptr(a.perms.Prot()), nil
+	default:
+		// Ioctls irrelevant to Ashmem.
+		return 0, syserror.EINVAL
+	}
+}
+
+// pinOperation should only be called while holding a.mu.
+func (a *Area) pinOperation(pin linux.AshmemPin, op uint32) (uintptr, error) {
+	// Page-align a.size for checks.
+	pageAlignedSize, ok := usermem.Addr(a.size).RoundUp()
+	if !ok {
+		return 0, syserror.EINVAL
+	}
+	// Len 0 means everything onward.
+	if pin.Len == 0 {
+		pin.Len = uint32(pageAlignedSize) - pin.Offset
+	}
+	// Both Offset and Len have to be page-aligned.
+	if pin.Offset%uint32(usermem.PageSize) != 0 {
+		return 0, syserror.EINVAL
+	}
+	if pin.Len%uint32(usermem.PageSize) != 0 {
+		return 0, syserror.EINVAL
+	}
+	// Adding Offset and Len must not cause an uint32 overflow.
+	if end := pin.Offset + pin.Len; end < pin.Offset {
+		return 0, syserror.EINVAL
+	}
+	// Pin range must not exceed a's size.
+	if uint32(pageAlignedSize) < pin.Offset+pin.Len {
+		return 0, syserror.EINVAL
+	}
+	// Handle each operation.
+	r := RangeFromAshmemPin(pin)
+	switch op {
+	case linux.AshmemPinIoctl:
+		if a.pb.PinRange(r) {
+			return linux.AshmemWasPurged, nil
+		}
+		return linux.AshmemNotPurged, nil
+
+	case linux.AshmemUnpinIoctl:
+		// TODO: Implement purge on unpin.
+		a.pb.UnpinRange(r)
+		return 0, nil
+
+	case linux.AshmemGetPinStatusIoctl:
+		if a.pb.RangePinnedStatus(r) {
+			return linux.AshmemIsPinned, nil
+		}
+		return linux.AshmemIsUnpinned, nil
+
+	default:
+		panic("unreachable")
+	}
+
+}
diff --git a/pkg/sentry/fs/ashmem/device.go b/pkg/sentry/fs/ashmem/device.go
new file mode 100644
index 000000000..c5b51d4a7
--- /dev/null
+++ b/pkg/sentry/fs/ashmem/device.go
@@ -0,0 +1,169 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package ashmem implements Android ashmem module (Anonymus Shared Memory).
+package ashmem
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Device implements fs.InodeOperations.
+type Device struct {
+	fsutil.DeprecatedFileOperations
+	fsutil.InodeNoExtendedAttributes
+	fsutil.InodeNotDirectory
+	fsutil.InodeNotRenameable
+	fsutil.InodeNotSocket
+	fsutil.InodeNotSymlink
+	fsutil.NoFsync
+	fsutil.NoMappable
+	fsutil.NoopWriteOut
+	fsutil.NotDirReaddir
+
+	mu       sync.Mutex `state:"nosave"`
+	unstable fs.UnstableAttr
+}
+
+// NewDevice creates and intializes a Device structure.
+func NewDevice(ctx context.Context, owner fs.FileOwner, fp fs.FilePermissions) *Device {
+	return &Device{
+		unstable: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: owner,
+			Perms: fp,
+			Links: 1,
+		}),
+	}
+}
+
+// Release implements fs.InodeOperations.Release.
+func (ad *Device) Release(context.Context) {}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (ad *Device) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return fs.NewFile(ctx, d, flags, &Area{
+		ad:        ad,
+		tmpfsFile: nil,
+		perms:     usermem.AnyAccess,
+	}), nil
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (ad *Device) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	ad.mu.Lock()
+	defer ad.mu.Unlock()
+	return ad.unstable, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (ad *Device) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (ad *Device) SetPermissions(ctx context.Context, inode *fs.Inode, fp fs.FilePermissions) bool {
+	ad.mu.Lock()
+	defer ad.mu.Unlock()
+	ad.unstable.Perms = fp
+	ad.unstable.StatusChangeTime = time.NowFromContext(ctx)
+	return true
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (ad *Device) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	ad.mu.Lock()
+	defer ad.mu.Unlock()
+	if owner.UID.Ok() {
+		ad.unstable.Owner.UID = owner.UID
+	}
+	if owner.GID.Ok() {
+		ad.unstable.Owner.GID = owner.GID
+	}
+	return nil
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (ad *Device) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	ad.mu.Lock()
+	defer ad.mu.Unlock()
+
+	now := time.NowFromContext(ctx)
+	if !ts.ATimeOmit {
+		if ts.ATimeSetSystemTime {
+			ad.unstable.AccessTime = now
+		} else {
+			ad.unstable.AccessTime = ts.ATime
+		}
+	}
+	if !ts.MTimeOmit {
+		if ts.MTimeSetSystemTime {
+			ad.unstable.ModificationTime = now
+		} else {
+			ad.unstable.ModificationTime = ts.MTime
+		}
+	}
+	ad.unstable.StatusChangeTime = now
+	return nil
+}
+
+// Truncate implements fs.InodeOperations.WriteOut.
+//
+// Ignored by ashmem.
+func (ad *Device) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	return nil
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+//
+// Ashmem doesn't support links, no-op.
+func (ad *Device) AddLink() {}
+
+// DropLink implements fs.InodeOperations.DropLink.
+//
+// Ashmem doesn't support links, no-op.
+func (ad *Device) DropLink() {}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+func (ad *Device) NotifyStatusChange(ctx context.Context) {
+	ad.mu.Lock()
+	defer ad.mu.Unlock()
+	now := time.NowFromContext(ctx)
+	ad.unstable.ModificationTime = now
+	ad.unstable.StatusChangeTime = now
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+//
+// Ashmem is virtual.
+func (ad *Device) IsVirtual() bool {
+	return true
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+//
+// Ashmem doesn't support querying for filesystem info.
+func (ad *Device) StatFS(context.Context) (fs.Info, error) {
+	return fs.Info{}, syserror.ENOSYS
+}
diff --git a/pkg/sentry/fs/ashmem/pin_board.go b/pkg/sentry/fs/ashmem/pin_board.go
new file mode 100644
index 000000000..c7fb3822c
--- /dev/null
+++ b/pkg/sentry/fs/ashmem/pin_board.go
@@ -0,0 +1,125 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ashmem
+
+import "gvisor.googlesource.com/gvisor/pkg/abi/linux"
+
+const maxUint64 = ^uint64(0)
+
+// setFunctions implements segment.Functions generated from segment.Functions for
+// uint64 Key and noValue Value. For more information, see the build file and
+// segment set implementation at pkg/segment/set.go.
+type setFunctions struct{}
+
+// noValue is a type of range attached value, which is irrelevant here.
+type noValue struct{}
+
+// MinKey implements segment.Functions.MinKey.
+func (setFunctions) MinKey() uint64 {
+	return 0
+}
+
+// MaxKey implements segment.Functions.MaxKey.
+func (setFunctions) MaxKey() uint64 {
+	return maxUint64
+}
+
+// ClearValue implements segment.Functions.ClearValue.
+func (setFunctions) ClearValue(*noValue) {
+	return
+}
+
+// Merge implements segment.Functions.Merge.
+func (setFunctions) Merge(Range, noValue, Range, noValue) (noValue, bool) {
+	return noValue{}, true
+}
+
+// Split implements segment.Functions.Split.
+func (setFunctions) Split(Range, noValue, uint64) (noValue, noValue) {
+	return noValue{}, noValue{}
+}
+
+// PinBoard represents a set of pinned ranges in ashmem.
+//
+// segment.Set is used for implementation where segments represent
+// ranges of pinned bytes, while gaps represent ranges of unpinned
+// bytes. All ranges are page-aligned.
+type PinBoard struct {
+	Set
+}
+
+// NewPinBoard creates a new pin board with all pages pinned.
+func NewPinBoard() *PinBoard {
+	var pb PinBoard
+	pb.PinRange(Range{0, maxUint64})
+	return &pb
+}
+
+// PinRange pins all pages in the specified range and returns true
+// if there are any newly pinned pages.
+func (pb *PinBoard) PinRange(r Range) bool {
+	pinnedPages := false
+	for gap := pb.LowerBoundGap(r.Start); gap.Ok() && gap.Start() < r.End; {
+		common := gap.Range().Intersect(r)
+		if common.Length() == 0 {
+			gap = gap.NextGap()
+			continue
+		}
+		pinnedPages = true
+		gap = pb.Insert(gap, common, noValue{}).NextGap()
+	}
+	return pinnedPages
+}
+
+// UnpinRange unpins all pages in the specified range.
+func (pb *PinBoard) UnpinRange(r Range) {
+	for seg := pb.LowerBoundSegment(r.Start); seg.Ok() && seg.Start() < r.End; {
+		common := seg.Range().Intersect(r)
+		if common.Length() == 0 {
+			seg = seg.NextSegment()
+			continue
+		}
+		seg = pb.RemoveRange(common).NextSegment()
+	}
+}
+
+// RangePinnedStatus returns false if there's at least one unpinned page in the
+// specified range.
+func (pb *PinBoard) RangePinnedStatus(r Range) bool {
+	for gap := pb.LowerBoundGap(r.Start); gap.Ok() && gap.Start() < r.End; {
+		common := gap.Range().Intersect(r)
+		if common.Length() == 0 {
+			gap = gap.NextGap()
+			continue
+		}
+		return false
+	}
+	return true
+}
+
+// RangeFromAshmemPin converts ashmem's original pin structure
+// to Range.
+func RangeFromAshmemPin(ap linux.AshmemPin) Range {
+	if ap.Len == 0 {
+		return Range{
+			uint64(ap.Offset),
+			maxUint64,
+		}
+	}
+	return Range{
+		uint64(ap.Offset),
+		uint64(ap.Offset) + uint64(ap.Len),
+	}
+}
diff --git a/pkg/sentry/fs/ashmem/pin_board_test.go b/pkg/sentry/fs/ashmem/pin_board_test.go
new file mode 100644
index 000000000..f4ea5de6d
--- /dev/null
+++ b/pkg/sentry/fs/ashmem/pin_board_test.go
@@ -0,0 +1,130 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ashmem
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func TestPinBoard(t *testing.T) {
+	pb := NewPinBoard()
+
+	// Confirm that all pages are pinned.
+	if !pb.RangePinnedStatus(RangeFromAshmemPin(linux.AshmemPin{0, 0})) {
+		t.Errorf("RangePinnedStatus(all pages) returned false (unpinned) at start.")
+	}
+
+	// Unpin pages [1, 11) (counting from 0)
+	pb.UnpinRange(RangeFromAshmemPin(linux.AshmemPin{
+		usermem.PageSize,
+		usermem.PageSize * 10,
+	}))
+
+	// Confirm that pages [1, 11) are unpinned and that page 0 and pages
+	// larger than 10 are pinned.
+	pinned := []linux.AshmemPin{
+		{
+			0,
+			usermem.PageSize,
+		}, {
+			usermem.PageSize * 11,
+			0,
+		},
+	}
+
+	for _, pin := range pinned {
+		if !pb.RangePinnedStatus(RangeFromAshmemPin(pin)) {
+			t.Errorf("RangePinnedStatus(AshmemPin{offset (pages): %v, len (pages): %v}) returned false (unpinned).",
+				pin.Offset, pin.Len)
+		}
+	}
+
+	unpinned := []linux.AshmemPin{
+		{
+			usermem.PageSize,
+			usermem.PageSize * 10,
+		},
+	}
+
+	for _, pin := range unpinned {
+		if pb.RangePinnedStatus(RangeFromAshmemPin(pin)) {
+			t.Errorf("RangePinnedStatus(AshmemPin{offset (pages): %v, len (pages): %v}) returned true (pinned).",
+				pin.Offset, pin.Len)
+		}
+	}
+
+	// Pin pages [2, 6).
+	pb.PinRange(RangeFromAshmemPin(linux.AshmemPin{
+		usermem.PageSize * 2,
+		usermem.PageSize * 4,
+	}))
+
+	// Confirm that pages 0, [2, 6) and pages larger than 10 are pinned
+	// while others remain unpinned.
+	pinned = []linux.AshmemPin{
+		{
+			0,
+			usermem.PageSize,
+		},
+		{
+			usermem.PageSize * 2,
+			usermem.PageSize * 4,
+		},
+		{
+			usermem.PageSize * 11,
+			0,
+		},
+	}
+
+	for _, pin := range pinned {
+		if !pb.RangePinnedStatus(RangeFromAshmemPin(pin)) {
+			t.Errorf("RangePinnedStatus(AshmemPin{offset (pages): %v, len (pages): %v}) returned false (unpinned).",
+				pin.Offset, pin.Len)
+		}
+	}
+
+	unpinned = []linux.AshmemPin{
+		{
+			usermem.PageSize,
+			usermem.PageSize,
+		}, {
+			usermem.PageSize * 6,
+			usermem.PageSize * 5,
+		},
+	}
+
+	for _, pin := range unpinned {
+		if pb.RangePinnedStatus(RangeFromAshmemPin(pin)) {
+			t.Errorf("RangePinnedStatus(AshmemPin{offset (pages): %v, len (pages): %v}) returned true (pinned).",
+				pin.Offset, pin.Len)
+		}
+	}
+
+	// Status of a partially pinned range is unpinned.
+	if pb.RangePinnedStatus(RangeFromAshmemPin(linux.AshmemPin{0, 0})) {
+		t.Errorf("RangePinnedStatus(all pages) returned true (pinned).")
+	}
+
+	// Pin the whole range again.
+	pb.PinRange(RangeFromAshmemPin(linux.AshmemPin{0, 0}))
+
+	// Confirm that all pages are pinned.
+	if !pb.RangePinnedStatus(RangeFromAshmemPin(linux.AshmemPin{0, 0})) {
+		t.Errorf("RangePinnedStatus(all pages) returned false (unpinned) at start.")
+	}
+}
diff --git a/pkg/sentry/fs/attr.go b/pkg/sentry/fs/attr.go
new file mode 100644
index 000000000..56a2ad6f7
--- /dev/null
+++ b/pkg/sentry/fs/attr.go
@@ -0,0 +1,382 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+)
+
+// InodeType enumerates types of Inodes.
+type InodeType int
+
+const (
+	// RegularFile is a regular file.
+	RegularFile InodeType = iota
+
+	// SpecialFile is a file that doesn't support SeekEnd. It is used for
+	// things like proc files.
+	SpecialFile
+
+	// Directory is a directory.
+	Directory
+
+	// SpecialDirectory is a directory that *does* support SeekEnd. It's
+	// the opposite of the SpecialFile scenario above. It similarly
+	// supports proc files.
+	SpecialDirectory
+
+	// Symlink is a symbolic link.
+	Symlink
+
+	// Pipe is a pipe (named or regular).
+	Pipe
+
+	// Socket is a socket.
+	Socket
+
+	// CharacterDevice is a character device.
+	CharacterDevice
+
+	// BlockDevice is a block device.
+	BlockDevice
+
+	// Anonymous is an anonymous type when none of the above apply.
+	// Epoll fds and event-driven fds fit this category.
+	Anonymous
+)
+
+// String returns a human-readable representation of the InodeType.
+func (n InodeType) String() string {
+	switch n {
+	case RegularFile, SpecialFile:
+		return "file"
+	case Directory, SpecialDirectory:
+		return "directory"
+	case Symlink:
+		return "symlink"
+	case Pipe:
+		return "pipe"
+	case Socket:
+		return "socket"
+	case CharacterDevice:
+		return "character-device"
+	case BlockDevice:
+		return "block-device"
+	case Anonymous:
+		return "anonymous"
+	default:
+		return "unknown"
+	}
+}
+
+// StableAttr contains Inode attributes that will be stable throughout the
+// lifetime of the Inode.
+type StableAttr struct {
+	// Type is the InodeType of a InodeOperations.
+	Type InodeType
+
+	// DeviceID is the device on which a InodeOperations resides.
+	DeviceID uint64
+
+	// InodeID uniquely identifies InodeOperations on its device.
+	InodeID uint64
+
+	// BlockSize is the block size of data backing this InodeOperations.
+	BlockSize int64
+
+	// DeviceFileMajor is the major device number of this Node, if it is a
+	// device file.
+	DeviceFileMajor uint16
+
+	// DeviceFileMinor is the minor device number of this Node, if it is a
+	// device file.
+	DeviceFileMinor uint32
+}
+
+// IsRegular returns true if StableAttr.Type matches a regular file.
+func IsRegular(s StableAttr) bool {
+	return s.Type == RegularFile
+}
+
+// IsFile returns true if StableAttr.Type matches any type of file.
+func IsFile(s StableAttr) bool {
+	return s.Type == RegularFile || s.Type == SpecialFile
+}
+
+// IsDir returns true if StableAttr.Type matches any type of directory.
+func IsDir(s StableAttr) bool {
+	return s.Type == Directory || s.Type == SpecialDirectory
+}
+
+// IsSymlink returns true if StableAttr.Type matches a symlink.
+func IsSymlink(s StableAttr) bool {
+	return s.Type == Symlink
+}
+
+// IsPipe returns true if StableAttr.Type matches any type of pipe.
+func IsPipe(s StableAttr) bool {
+	return s.Type == Pipe
+}
+
+// IsSocket returns true if StableAttr.Type matches any type of socket.
+func IsSocket(s StableAttr) bool {
+	return s.Type == Socket
+}
+
+// IsCharDevice returns true if StableAttr.Type matches a character device.
+func IsCharDevice(s StableAttr) bool {
+	return s.Type == CharacterDevice
+}
+
+// UnstableAttr contains Inode attributes that may change over the lifetime
+// of the Inode.
+type UnstableAttr struct {
+	// Size is the file size in bytes.
+	Size int64
+
+	// Usage is the actual data usage in bytes.
+	Usage int64
+
+	// Perms is the protection (read/write/execute for user/group/other).
+	Perms FilePermissions
+
+	// Owner describes the ownership of this file.
+	Owner FileOwner
+
+	// AccessTime is the time of last access
+	AccessTime ktime.Time
+
+	// ModificationTime is the time of last modification.
+	ModificationTime ktime.Time
+
+	// StatusChangeTime is the time of last attribute modification.
+	StatusChangeTime ktime.Time
+
+	// Links is the number of hard links.
+	Links uint64
+}
+
+// WithCurrentTime returns u with AccessTime == ModificationTime == current time.
+func WithCurrentTime(ctx context.Context, u UnstableAttr) UnstableAttr {
+	t := ktime.NowFromContext(ctx)
+	u.AccessTime = t
+	u.ModificationTime = t
+	u.StatusChangeTime = t
+	return u
+}
+
+// AttrMask contains fields to mask StableAttr and UnstableAttr.
+type AttrMask struct {
+	Type             bool
+	DeviceID         bool
+	InodeID          bool
+	BlockSize        bool
+	Size             bool
+	Usage            bool
+	Perms            bool
+	UID              bool
+	GID              bool
+	AccessTime       bool
+	ModificationTime bool
+	StatusChangeTime bool
+	Links            bool
+}
+
+// Empty returns true if all fields in AttrMask are false.
+func (a AttrMask) Empty() bool {
+	return a == AttrMask{}
+}
+
+// Union returns an AttrMask containing the inclusive disjunction of fields in a and b.
+func (a AttrMask) Union(b AttrMask) AttrMask {
+	return AttrMask{
+		Type:             a.Type || b.Type,
+		DeviceID:         a.DeviceID || b.DeviceID,
+		InodeID:          a.InodeID || b.InodeID,
+		BlockSize:        a.BlockSize || b.BlockSize,
+		Size:             a.Size || b.Size,
+		Usage:            a.Usage || b.Usage,
+		Perms:            a.Perms || b.Perms,
+		UID:              a.UID || b.UID,
+		GID:              a.GID || b.GID,
+		AccessTime:       a.AccessTime || b.AccessTime,
+		ModificationTime: a.ModificationTime || b.ModificationTime,
+		StatusChangeTime: a.StatusChangeTime || b.StatusChangeTime,
+		Links:            a.Links || b.Links,
+	}
+}
+
+// PermMask are file access permissions.
+type PermMask struct {
+	// Read indicates reading is permitted.
+	Read bool
+
+	// Write indicates writing is permitted.
+	Write bool
+
+	// Execute indicates execution is permitted.
+	Execute bool
+}
+
+// OnlyRead returns true when only the read bit is set.
+func (p PermMask) OnlyRead() bool {
+	return p.Read && !p.Write && !p.Execute
+}
+
+// String implements the fmt.Stringer interface for PermMask.
+func (p PermMask) String() string {
+	return fmt.Sprintf("PermMask{Read: %v, Write: %v, Execute: %v}", p.Read, p.Write, p.Execute)
+}
+
+// Mode returns the system mode (syscall.S_IXOTH, etc.) for these permissions
+// in the "other" bits.
+func (p PermMask) Mode() (mode os.FileMode) {
+	if p.Read {
+		mode |= syscall.S_IROTH
+	}
+	if p.Write {
+		mode |= syscall.S_IWOTH
+	}
+	if p.Execute {
+		mode |= syscall.S_IXOTH
+	}
+	return
+}
+
+// SupersetOf returns true iff the permissions in p are a superset of the
+// permissions in other.
+func (p PermMask) SupersetOf(other PermMask) bool {
+	if !p.Read && other.Read {
+		return false
+	}
+	if !p.Write && other.Write {
+		return false
+	}
+	if !p.Execute && other.Execute {
+		return false
+	}
+	return true
+}
+
+// FilePermissions represents the permissions of a file, with
+// Read/Write/Execute bits for user, group, and other.
+type FilePermissions struct {
+	User  PermMask
+	Group PermMask
+	Other PermMask
+
+	// Sticky, if set on directories, restricts renaming and deletion of
+	// files in those directories to the directory owner, file owner, or
+	// CAP_FOWNER. The sticky bit is ignored when set on other files.
+	Sticky bool
+
+	// SetUID executables can call UID-setting syscalls without CAP_SETUID.
+	SetUID bool
+
+	// SetGID executables can call GID-setting syscalls without CAP_SETGID.
+	SetGID bool
+}
+
+// PermsFromMode takes the Other permissions (last 3 bits) of a FileMode and
+// returns a set of PermMask.
+func PermsFromMode(mode linux.FileMode) (perms PermMask) {
+	perms.Read = mode&linux.ModeOtherRead != 0
+	perms.Write = mode&linux.ModeOtherWrite != 0
+	perms.Execute = mode&linux.ModeOtherExec != 0
+	return
+}
+
+// FilePermsFromP9 converts a p9.FileMode to a FilePermissions struct.
+func FilePermsFromP9(mode p9.FileMode) FilePermissions {
+	return FilePermsFromMode(linux.FileMode(mode))
+}
+
+// FilePermsFromMode converts a system file mode to a FilePermissions struct.
+func FilePermsFromMode(mode linux.FileMode) (fp FilePermissions) {
+	perm := mode.Permissions()
+	fp.Other = PermsFromMode(perm)
+	fp.Group = PermsFromMode(perm >> 3)
+	fp.User = PermsFromMode(perm >> 6)
+	fp.Sticky = mode&linux.ModeSticky == linux.ModeSticky
+	fp.SetUID = mode&linux.ModeSetUID == linux.ModeSetUID
+	fp.SetGID = mode&linux.ModeSetGID == linux.ModeSetGID
+	return
+}
+
+// LinuxMode returns the linux mode_t representation of these permissions.
+func (f FilePermissions) LinuxMode() linux.FileMode {
+	m := linux.FileMode(f.User.Mode()<<6 | f.Group.Mode()<<3 | f.Other.Mode())
+	if f.SetUID {
+		m |= linux.ModeSetUID
+	}
+	if f.SetGID {
+		m |= linux.ModeSetGID
+	}
+	if f.Sticky {
+		m |= linux.ModeSticky
+	}
+	return m
+}
+
+// OSMode returns the Go runtime's OS independent os.FileMode representation of
+// these permissions.
+func (f FilePermissions) OSMode() os.FileMode {
+	m := os.FileMode(f.User.Mode()<<6 | f.Group.Mode()<<3 | f.Other.Mode())
+	if f.SetUID {
+		m |= os.ModeSetuid
+	}
+	if f.SetGID {
+		m |= os.ModeSetgid
+	}
+	if f.Sticky {
+		m |= os.ModeSticky
+	}
+	return m
+}
+
+// AnyExecute returns true if any of U/G/O have the execute bit set.
+func (f FilePermissions) AnyExecute() bool {
+	return f.User.Execute || f.Group.Execute || f.Other.Execute
+}
+
+// AnyWrite returns true if any of U/G/O have the write bit set.
+func (f FilePermissions) AnyWrite() bool {
+	return f.User.Write || f.Group.Write || f.Other.Write
+}
+
+// AnyRead returns true if any of U/G/O have the read bit set.
+func (f FilePermissions) AnyRead() bool {
+	return f.User.Read || f.Group.Read || f.Other.Read
+}
+
+// FileOwner represents ownership of a file.
+type FileOwner struct {
+	UID auth.KUID
+	GID auth.KGID
+}
+
+// RootOwner corresponds to KUID/KGID 0/0.
+var RootOwner = FileOwner{
+	UID: auth.RootKUID,
+	GID: auth.RootKGID,
+}
diff --git a/pkg/sentry/fs/binder/BUILD b/pkg/sentry/fs/binder/BUILD
new file mode 100644
index 000000000..15f91699f
--- /dev/null
+++ b/pkg/sentry/fs/binder/BUILD
@@ -0,0 +1,38 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "binder_state",
+    srcs = ["binder.go"],
+    out = "binder_state.go",
+    package = "binder",
+)
+
+go_library(
+    name = "binder",
+    srcs = [
+        "binder.go",
+        "binder_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/binder",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+    ],
+)
diff --git a/pkg/sentry/fs/binder/binder.go b/pkg/sentry/fs/binder/binder.go
new file mode 100644
index 000000000..3f87b6b08
--- /dev/null
+++ b/pkg/sentry/fs/binder/binder.go
@@ -0,0 +1,358 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package binder implements Android Binder IPC module.
+package binder
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	currentProtocolVersion = 8
+
+	// mmapSizeLimit is the upper limit for mapped memory size in Binder.
+	mmapSizeLimit = 4 * 1024 * 1024 // 4MB
+)
+
+// Device implements fs.InodeOperations.
+type Device struct {
+	fsutil.InodeNoExtendedAttributes
+	fsutil.InodeNotDirectory
+	fsutil.InodeNotRenameable
+	fsutil.InodeNotSocket
+	fsutil.InodeNotSymlink
+	fsutil.NoMappable
+	fsutil.NoopWriteOut
+	fsutil.DeprecatedFileOperations
+
+	// mu protects unstable.
+	mu       sync.Mutex `state:"nosave"`
+	unstable fs.UnstableAttr
+}
+
+// NewDevice creates and intializes a Device structure.
+func NewDevice(ctx context.Context, owner fs.FileOwner, fp fs.FilePermissions) *Device {
+	return &Device{
+		unstable: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: owner,
+			Perms: fp,
+			Links: 1,
+		}),
+	}
+}
+
+// Release implements fs.InodeOperations.Release.
+func (bd *Device) Release(context.Context) {}
+
+// GetFile implements fs.InodeOperations.GetFile.
+//
+// TODO: Add functionality to GetFile: Additional fields will be
+// needed in the Device structure, initialize them here. Also, Device will need
+// to keep track of the created Procs in order to implement BINDER_READ_WRITE
+// ioctl.
+func (bd *Device) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return fs.NewFile(ctx, d, flags, &Proc{
+		bd:       bd,
+		task:     kernel.TaskFromContext(ctx),
+		platform: platform.FromContext(ctx),
+	}), nil
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (bd *Device) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	bd.mu.Lock()
+	defer bd.mu.Unlock()
+	return bd.unstable, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (bd *Device) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (bd *Device) SetPermissions(ctx context.Context, inode *fs.Inode, fp fs.FilePermissions) bool {
+	bd.mu.Lock()
+	defer bd.mu.Unlock()
+	bd.unstable.Perms = fp
+	bd.unstable.StatusChangeTime = time.NowFromContext(ctx)
+	return true
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (bd *Device) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	bd.mu.Lock()
+	defer bd.mu.Unlock()
+	if owner.UID.Ok() {
+		bd.unstable.Owner.UID = owner.UID
+	}
+	if owner.GID.Ok() {
+		bd.unstable.Owner.GID = owner.GID
+	}
+	return nil
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (bd *Device) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	bd.mu.Lock()
+	defer bd.mu.Unlock()
+
+	now := time.NowFromContext(ctx)
+	if !ts.ATimeOmit {
+		if ts.ATimeSetSystemTime {
+			bd.unstable.AccessTime = now
+		} else {
+			bd.unstable.AccessTime = ts.ATime
+		}
+	}
+	if !ts.MTimeOmit {
+		if ts.MTimeSetSystemTime {
+			bd.unstable.ModificationTime = now
+		} else {
+			bd.unstable.ModificationTime = ts.MTime
+		}
+	}
+	bd.unstable.StatusChangeTime = now
+	return nil
+}
+
+// Truncate implements fs.InodeOperations.WriteOut.
+//
+// Ignored for a character device, such as Binder.
+func (bd *Device) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	return nil
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+//
+// Binder doesn't support links, no-op.
+func (bd *Device) AddLink() {}
+
+// DropLink implements fs.InodeOperations.DropLink.
+//
+// Binder doesn't support links, no-op.
+func (bd *Device) DropLink() {}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+func (bd *Device) NotifyStatusChange(ctx context.Context) {
+	bd.mu.Lock()
+	defer bd.mu.Unlock()
+	now := time.NowFromContext(ctx)
+	bd.unstable.ModificationTime = now
+	bd.unstable.StatusChangeTime = now
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+//
+// Binder is virtual.
+func (bd *Device) IsVirtual() bool {
+	return true
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+//
+// Binder doesn't support querying for filesystem info.
+func (bd *Device) StatFS(context.Context) (fs.Info, error) {
+	return fs.Info{}, syserror.ENOSYS
+}
+
+// Proc implements fs.FileOperations and fs.IoctlGetter.
+type Proc struct {
+	fsutil.NoFsync
+	fsutil.DeprecatedFileOperations
+	fsutil.NotDirReaddir
+
+	bd       *Device
+	task     *kernel.Task
+	platform platform.Platform
+
+	// mu protects fr.
+	mu sync.Mutex `state:"nosave"`
+
+	// mapped is memory allocated from platform.Memory() by AddMapping.
+	mapped platform.FileRange
+}
+
+// Release implements fs.FileOperations.Release.
+func (bp *Proc) Release() {
+	bp.mu.Lock()
+	defer bp.mu.Unlock()
+	if bp.mapped.Length() != 0 {
+		bp.platform.Memory().DecRef(bp.mapped)
+	}
+}
+
+// Seek implements fs.FileOperations.Seek.
+//
+// Binder doesn't support seek operation (unless in debug mode).
+func (bp *Proc) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	return offset, syserror.EOPNOTSUPP
+}
+
+// Read implements fs.FileOperations.Read.
+//
+// Binder doesn't support read operation (unless in debug mode).
+func (bp *Proc) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	return 0, syserror.EOPNOTSUPP
+}
+
+// Write implements fs.FileOperations.Write.
+//
+// Binder doesn't support write operation.
+func (bp *Proc) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	return 0, syserror.EOPNOTSUPP
+}
+
+// Flush implements fs.FileOperations.Flush.
+//
+// TODO: Implement.
+func (bp *Proc) Flush(ctx context.Context, file *fs.File) error {
+	return nil
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (bp *Proc) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	// Compare drivers/android/binder.c:binder_mmap().
+	if caller := kernel.TaskFromContext(ctx); caller != bp.task {
+		return syserror.EINVAL
+	}
+	if opts.Length > mmapSizeLimit {
+		opts.Length = mmapSizeLimit
+	}
+	opts.MaxPerms.Write = false
+
+	// TODO: Binder sets VM_DONTCOPY, preventing the created vma
+	// from being copied across fork(), but we don't support this yet. As
+	// a result, MMs containing a Binder mapping cannot be forked (MM.Fork will
+	// fail when AddMapping returns EBUSY).
+
+	return fsutil.GenericConfigureMMap(file, bp, opts)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+//
+// TODO: Implement.
+func (bp *Proc) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// Switch on ioctl request.
+	switch uint32(args[1].Int()) {
+	case linux.BinderVersionIoctl:
+		ver := &linux.BinderVersion{
+			ProtocolVersion: currentProtocolVersion,
+		}
+		// Copy result to user-space.
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), ver, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	case linux.BinderWriteReadIoctl:
+		// TODO: Implement.
+		fallthrough
+	case linux.BinderSetIdleTimeoutIoctl:
+		// TODO: Implement.
+		fallthrough
+	case linux.BinderSetMaxThreadsIoctl:
+		// TODO: Implement.
+		fallthrough
+	case linux.BinderSetIdlePriorityIoctl:
+		// TODO: Implement.
+		fallthrough
+	case linux.BinderSetContextMgrIoctl:
+		// TODO: Implement.
+		fallthrough
+	case linux.BinderThreadExitIoctl:
+		// TODO: Implement.
+		return 0, syserror.ENOSYS
+	default:
+		// Ioctls irrelevant to Binder.
+		return 0, syserror.EINVAL
+	}
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (bp *Proc) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	bp.mu.Lock()
+	defer bp.mu.Unlock()
+	if bp.mapped.Length() != 0 {
+		// mmap has been called before, which binder_mmap() doesn't like.
+		return syserror.EBUSY
+	}
+	// Binder only allocates and maps a single page up-front
+	// (drivers/android/binder.c:binder_mmap() => binder_update_page_range()).
+	fr, err := bp.platform.Memory().Allocate(usermem.PageSize, usage.Anonymous)
+	if err != nil {
+		return err
+	}
+	bp.mapped = fr
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (bp *Proc) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+	// Nothing to do. Notably, we don't free bp.mapped to allow another mmap.
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (bp *Proc) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	// Nothing to do. Notably, this is one case where CopyMapping isn't
+	// equivalent to AddMapping, as AddMapping would return EBUSY.
+	return nil
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (bp *Proc) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	// TODO: In addition to the page initially allocated and mapped
+	// in AddMapping (Linux: binder_mmap), Binder allocates and maps pages for
+	// each transaction (Linux: binder_ioctl => binder_ioctl_write_read =>
+	// binder_thread_write => binder_transaction => binder_alloc_buf =>
+	// binder_update_page_range). Since we don't actually implement
+	// BinderWriteReadIoctl (Linux: BINDER_WRITE_READ), we only ever have the
+	// first page.
+	var err error
+	if required.End > usermem.PageSize {
+		err = &memmap.BusError{syserror.EFAULT}
+	}
+	if required.Start == 0 {
+		return []memmap.Translation{
+			{
+				Source: memmap.MappableRange{0, usermem.PageSize},
+				File:   bp.platform.Memory(),
+				Offset: bp.mapped.Start,
+			},
+		}, err
+	}
+	return nil, err
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (bp *Proc) InvalidateUnsavable(ctx context.Context) error {
+	return nil
+}
diff --git a/pkg/sentry/fs/context.go b/pkg/sentry/fs/context.go
new file mode 100644
index 000000000..b521bce75
--- /dev/null
+++ b/pkg/sentry/fs/context.go
@@ -0,0 +1,97 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+)
+
+// contextID is the kernel package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxRoot is a Context.Value key for a Dirent.
+	CtxRoot contextID = iota
+)
+
+// ContextCanAccessFile determines whether `file` can be accessed in the requested way
+// (for reading, writing, or execution) using the caller's credentials and user
+// namespace, as does Linux's fs/namei.c:generic_permission.
+func ContextCanAccessFile(ctx context.Context, inode *Inode, reqPerms PermMask) bool {
+	creds := auth.CredentialsFromContext(ctx)
+	uattr, err := inode.UnstableAttr(ctx)
+	if err != nil {
+		return false
+	}
+
+	p := uattr.Perms.Other
+	// Are we owner or in group?
+	if uattr.Owner.UID == creds.EffectiveKUID {
+		p = uattr.Perms.User
+	} else if creds.InGroup(uattr.Owner.GID) {
+		p = uattr.Perms.Group
+	}
+
+	// Are permissions satisfied without capability checks?
+	if p.SupersetOf(reqPerms) {
+		return true
+	}
+
+	if IsDir(inode.StableAttr) {
+		// CAP_DAC_OVERRIDE can override any perms on directories.
+		if inode.CheckCapability(ctx, linux.CAP_DAC_OVERRIDE) {
+			return true
+		}
+
+		// CAP_DAC_READ_SEARCH can normally only override Read perms,
+		// but for directories it can also override execution.
+		if !reqPerms.Write && inode.CheckCapability(ctx, linux.CAP_DAC_READ_SEARCH) {
+			return true
+		}
+	}
+
+	// CAP_DAC_OVERRIDE can always override Read/Write.
+	// Can override executable only when at least one execute bit is set.
+	if !reqPerms.Execute || uattr.Perms.AnyExecute() {
+		if inode.CheckCapability(ctx, linux.CAP_DAC_OVERRIDE) {
+			return true
+		}
+	}
+
+	// Read perms can be overridden by CAP_DAC_READ_SEARCH.
+	if reqPerms.OnlyRead() && inode.CheckCapability(ctx, linux.CAP_DAC_READ_SEARCH) {
+		return true
+	}
+	return false
+}
+
+// FileOwnerFromContext returns a FileOwner using the effective user and group
+// IDs used by ctx.
+func FileOwnerFromContext(ctx context.Context) FileOwner {
+	creds := auth.CredentialsFromContext(ctx)
+	return FileOwner{creds.EffectiveKUID, creds.EffectiveKGID}
+}
+
+// RootFromContext returns the root of the virtual filesystem observed by ctx,
+// or nil if ctx is not associated with a virtual filesystem. If
+// RootFromContext returns a non-nil fs.Dirent, a reference is taken on it.
+func RootFromContext(ctx context.Context) *Dirent {
+	if v := ctx.Value(CtxRoot); v != nil {
+		return v.(*Dirent)
+	}
+	return nil
+}
diff --git a/pkg/sentry/fs/copy_up.go b/pkg/sentry/fs/copy_up.go
new file mode 100644
index 000000000..ea74d0efd
--- /dev/null
+++ b/pkg/sentry/fs/copy_up.go
@@ -0,0 +1,414 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"io"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// copyUp copies a file in an overlay from a lower filesystem to an
+// upper filesytem so that the file can be modified in the upper
+// filesystem. Copying a file involves several steps:
+//
+// - All parent directories of the file are created in the upper
+//   filesystem if they don't exist there. For instance:
+//
+//     upper /dir0
+//     lower /dir0/dir1/file
+//
+//   copyUp of /dir0/dir1/file creates /dir0/dir1 in order to create
+//   /dir0/dir1/file.
+//
+// - The file content is copied from the lower file to the upper
+//   file. For symlinks this is the symlink target. For directories,
+//   upper directory entries are merged with lower directory entries
+//   so there is no need to copy any entries.
+//
+// - A subset of file attributes of the lower file are set on the
+//   upper file. These are the file owner, the file timestamps,
+//   and all non-overlay extended attributes. copyUp will fail if
+//   the upper filesystem does not support the setting of these
+//   attributes.
+//
+//   The file's permissions are set when the file is created and its
+//   size will be brought up to date when its contents are copied.
+//   Notably no attempt is made to bring link count up to date because
+//   hard links are currently not preserved across overlay filesystems.
+//
+// - Memory mappings of the lower file are invalidated and memory
+//   references are transferred to the upper file. From this point on,
+//   memory mappings of the file will be backed by content in the upper
+//   filesystem.
+//
+// Synchronization:
+//
+// copyUp synchronizes with rename(2) using renameMu to ensure that
+// parentage does not change while a file is being copied. In the context
+// of rename(2), copyUpLockedForRename should be used to avoid deadlock on
+// renameMu.
+//
+// The following operations synchronize with copyUp using copyMu:
+//
+// - InodeOperations, i.e. to ensure that looking up a directory takes
+//   into account new upper filesystem directories created by copy up,
+//   which subsequently can be modified.
+//
+// - FileOperations, i.e. to ensure that reading from a file does not
+//   continue using a stale, lower filesystem handle when the file is
+//   written to.
+//
+// Lock ordering: Dirent.mu -> Inode.overlay.copyMu -> Inode.mu.
+//
+// Caveats:
+//
+// If any step in copying up a file fails, copyUp cleans the upper
+// filesystem of any partially up-to-date file. If this cleanup fails,
+// the overlay may be in an unacceptable, inconsistent state, so copyUp
+// panics. If copyUp fails because any step (above) fails, a generic
+// error is returned.
+//
+// copyUp currently makes no attempt to optimize copying up file content.
+// For large files, this means that copyUp blocks until the entire file
+// is copied synchronously.
+func copyUp(ctx context.Context, d *Dirent) error {
+	renameMu.RLock()
+	defer renameMu.RUnlock()
+	return copyUpLockedForRename(ctx, d)
+}
+
+// copyUpLockedForRename is the same as copyUp except that it does not lock
+// renameMu.
+//
+// It copies each component of d that does not yet exist in the upper
+// filesystem. If d already exists in the upper filesystem, it is a no-op.
+//
+// Any error returned indicates a failure to copy all of d. This may
+// leave the upper filesystem filled with any number of parent directories
+// but the upper filesystem will never be in an inconsistent state.
+//
+// Preconditions:
+// - d.Inode.overlay is non-nil.
+func copyUpLockedForRename(ctx context.Context, d *Dirent) error {
+	for {
+		// Did we race with another copy up or does there
+		// already exist something in the upper filesystem
+		// for d?
+		d.Inode.overlay.copyMu.Lock()
+		if d.Inode.overlay.upper != nil {
+			d.Inode.overlay.copyMu.Unlock()
+			// Done, d is in the upper filesystem.
+			return nil
+		}
+		d.Inode.overlay.copyMu.Unlock()
+
+		// Find the next component to copy up. We will work our way
+		// down to the last component of d and finally copy it.
+		next := findNextCopyUp(ctx, d)
+
+		// Attempt to copy.
+		if err := doCopyUp(ctx, next); err != nil {
+			return err
+		}
+	}
+}
+
+// findNextCopyUp finds the next component of d from root that does not
+// yet exist in the upper filesystem. The parent of this component is
+// also returned, which is the root of the overlay in the worst case.
+func findNextCopyUp(ctx context.Context, d *Dirent) *Dirent {
+	next := d
+	for parent := next.parent; ; /* checked in-loop */ /* updated in-loop */ {
+		// Does this parent have a non-nil upper Inode?
+		parent.Inode.overlay.copyMu.RLock()
+		if parent.Inode.overlay.upper != nil {
+			parent.Inode.overlay.copyMu.RUnlock()
+			// Note that since we found an upper, it is stable.
+			return next
+		}
+		parent.Inode.overlay.copyMu.RUnlock()
+
+		// Continue searching for a parent with a non-nil
+		// upper Inode.
+		next = parent
+		parent = next.parent
+	}
+}
+
+func doCopyUp(ctx context.Context, d *Dirent) error {
+	// Wait to get exclusive access to the upper Inode.
+	d.Inode.overlay.copyMu.Lock()
+	defer d.Inode.overlay.copyMu.Unlock()
+	if d.Inode.overlay.upper != nil {
+		// We raced with another doCopyUp, no problem.
+		return nil
+	}
+
+	// Perform the copy.
+	return copyUpLocked(ctx, d.parent, d)
+}
+
+// copyUpLocked creates a copy of next in the upper filesystem of parent.
+//
+// copyUpLocked must be called with d.Inode.overlay.copyMu locked.
+//
+// Returns a generic error on failure.
+//
+// Preconditions:
+// - parent.Inode.overlay.upper must be non-nil.
+// - next.Inode.overlay.copyMu must be locked writable.
+// - next.Inode.overlay.lower must be non-nil.
+// - upper filesystem must support setting file ownership and timestamps.
+func copyUpLocked(ctx context.Context, parent *Dirent, next *Dirent) error {
+	// Extract the attributes of the file we wish to copy.
+	attrs, err := next.Inode.overlay.lower.UnstableAttr(ctx)
+	if err != nil {
+		log.Warningf("copy up failed to get lower attributes: %v", err)
+		return syserror.EIO
+	}
+
+	var childUpperInode *Inode
+	parentUpper := parent.Inode.overlay.upper
+
+	// Create the file in the upper filesystem and get an Inode for it.
+	switch next.Inode.StableAttr.Type {
+	case RegularFile:
+		childFile, err := parentUpper.Create(ctx, RootFromContext(ctx), next.name, FileFlags{Read: true, Write: true}, attrs.Perms)
+		if err != nil {
+			log.Warningf("copy up failed to create file: %v", err)
+			return syserror.EIO
+		}
+		defer childFile.DecRef()
+		childUpperInode = childFile.Dirent.Inode
+
+	case Directory:
+		if err := parentUpper.CreateDirectory(ctx, RootFromContext(ctx), next.name, attrs.Perms); err != nil {
+			log.Warningf("copy up failed to create directory: %v", err)
+			return syserror.EIO
+		}
+		childUpper, err := parentUpper.Lookup(ctx, next.name)
+		if err != nil {
+			log.Warningf("copy up failed to lookup directory: %v", err)
+			cleanupUpper(ctx, parentUpper, next.name)
+			return syserror.EIO
+		}
+		defer childUpper.DecRef()
+		childUpperInode = childUpper.Inode
+
+	case Symlink:
+		childLower := next.Inode.overlay.lower
+		link, err := childLower.Readlink(ctx)
+		if err != nil {
+			log.Warningf("copy up failed to read symlink value: %v", err)
+			return syserror.EIO
+		}
+		if err := parentUpper.CreateLink(ctx, RootFromContext(ctx), link, next.name); err != nil {
+			log.Warningf("copy up failed to create symlink: %v", err)
+			return syserror.EIO
+		}
+		childUpper, err := parentUpper.Lookup(ctx, next.name)
+		if err != nil {
+			log.Warningf("copy up failed to lookup symlink: %v", err)
+			cleanupUpper(ctx, parentUpper, next.name)
+			return syserror.EIO
+		}
+		defer childUpper.DecRef()
+		childUpperInode = childUpper.Inode
+
+	default:
+		return syserror.EINVAL
+	}
+
+	// Bring file attributes up to date. This does not include size, which will be
+	// brought up to date with copyContentsLocked.
+	if err := copyAttributesLocked(ctx, childUpperInode, next.Inode.overlay.lower); err != nil {
+		log.Warningf("copy up failed to copy up attributes: %v", err)
+		cleanupUpper(ctx, parentUpper, next.name)
+		return syserror.EIO
+	}
+
+	// Copy the entire file.
+	if err := copyContentsLocked(ctx, childUpperInode, next.Inode.overlay.lower, attrs.Size); err != nil {
+		log.Warningf("copy up failed to copy up contents: %v", err)
+		cleanupUpper(ctx, parentUpper, next.name)
+		return syserror.EIO
+	}
+
+	lowerMappable := next.Inode.overlay.lower.Mappable()
+	upperMappable := childUpperInode.Mappable()
+	if lowerMappable != nil && upperMappable == nil {
+		log.Warningf("copy up failed: cannot ensure memory mapping coherence")
+		cleanupUpper(ctx, parentUpper, next.name)
+		return syserror.EIO
+	}
+
+	// Propagate memory mappings to the upper Inode.
+	next.Inode.overlay.mapsMu.Lock()
+	defer next.Inode.overlay.mapsMu.Unlock()
+	if upperMappable != nil {
+		// Remember which mappings we added so we can remove them on failure.
+		allAdded := make(map[memmap.MappableRange]memmap.MappingsOfRange)
+		for seg := next.Inode.overlay.mappings.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+			added := make(memmap.MappingsOfRange)
+			for m := range seg.Value() {
+				if err := upperMappable.AddMapping(ctx, m.MappingSpace, m.AddrRange, seg.Start()); err != nil {
+					for m := range added {
+						upperMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, seg.Start())
+					}
+					for mr, mappings := range allAdded {
+						for m := range mappings {
+							upperMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, mr.Start)
+						}
+					}
+					return err
+				}
+				added[m] = struct{}{}
+			}
+			allAdded[seg.Range()] = added
+		}
+	}
+
+	// Take a reference on the upper Inode (transferred to
+	// next.Inode.overlay.upper) and make new translations use it.
+	next.Inode.overlay.dataMu.Lock()
+	childUpperInode.IncRef()
+	next.Inode.overlay.upper = childUpperInode
+	next.Inode.overlay.dataMu.Unlock()
+
+	// Invalidate existing translations through the lower Inode.
+	next.Inode.overlay.mappings.InvalidateAll(memmap.InvalidateOpts{})
+
+	// Remove existing memory mappings from the lower Inode.
+	if lowerMappable != nil {
+		for seg := next.Inode.overlay.mappings.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+			for m := range seg.Value() {
+				lowerMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, seg.Start())
+			}
+		}
+	}
+
+	return nil
+}
+
+// cleanupUpper removes name from parent, and panics if it is unsuccessful.
+func cleanupUpper(ctx context.Context, parent *Inode, name string) {
+	if err := parent.InodeOperations.Remove(ctx, parent, name); err != nil {
+		// Unfortunately we don't have much choice. We shouldn't
+		// willingly give the caller access to a nonsense filesystem.
+		panic(fmt.Sprintf("overlay filesystem is in an inconsistent state: failed to remove %q from upper filesystem: %v", name, err))
+	}
+}
+
+// copyUpBuffers is a buffer pool for copying file content. The buffer
+// size is the same used by io.Copy.
+var copyUpBuffers = sync.Pool{New: func() interface{} { return make([]byte, 8*usermem.PageSize) }}
+
+// copyContentsLocked copies the contents of lower to upper. It panics if
+// less than size bytes can be copied.
+func copyContentsLocked(ctx context.Context, upper *Inode, lower *Inode, size int64) error {
+	// We don't support copying up for anything other than regular files.
+	if lower.StableAttr.Type != RegularFile {
+		return nil
+	}
+
+	// Get a handle to the upper filesystem, which we will write to.
+	upperFile, err := overlayFile(ctx, upper, FileFlags{Write: true})
+	if err != nil {
+		return err
+	}
+	defer upperFile.DecRef()
+
+	// Get a handle to the lower filesystem, which we will read from.
+	lowerFile, err := overlayFile(ctx, lower, FileFlags{Read: true})
+	if err != nil {
+		return err
+	}
+	defer lowerFile.DecRef()
+
+	// Use a buffer pool to minimize allocations.
+	buf := copyUpBuffers.Get().([]byte)
+	defer copyUpBuffers.Put(buf)
+
+	// Transfer the contents.
+	//
+	// One might be able to optimize this by doing parallel reads, parallel writes and reads, larger
+	// buffers, etc. But we really don't know anything about the underlying implementation, so these
+	// optimizations could be self-defeating. So we leave this as simple as possible.
+	var offset int64
+	for {
+		nr, err := lowerFile.FileOperations.Read(ctx, lowerFile, usermem.BytesIOSequence(buf), offset)
+		if err != nil && err != io.EOF {
+			return err
+		}
+		if nr == 0 {
+			if offset != size {
+				// Same as in cleanupUpper, we cannot live
+				// with ourselves if we do anything less.
+				panic(fmt.Sprintf("filesystem is in an inconsistent state: wrote only %d bytes of %d sized file", offset, size))
+			}
+			return nil
+		}
+		nw, err := upperFile.FileOperations.Write(ctx, upperFile, usermem.BytesIOSequence(buf[:nr]), offset)
+		if err != nil {
+			return err
+		}
+		offset += nw
+	}
+}
+
+// copyAttributesLocked copies a subset of lower's attributes to upper,
+// specifically owner, timestamps (except of status change time), and
+// extended attributes. Notably no attempt is made to copy link count.
+// Size and permissions are set on upper when the file content is copied
+// and when the file is created respectively.
+func copyAttributesLocked(ctx context.Context, upper *Inode, lower *Inode) error {
+	// Extract attributes fro the lower filesystem.
+	lowerAttr, err := lower.UnstableAttr(ctx)
+	if err != nil {
+		return err
+	}
+	lowerXattr, err := lower.Listxattr()
+	if err != nil && err != syserror.EOPNOTSUPP {
+		return err
+	}
+
+	// Set the attributes on the upper filesystem.
+	if err := upper.InodeOperations.SetOwner(ctx, upper, lowerAttr.Owner); err != nil {
+		return err
+	}
+	if err := upper.InodeOperations.SetTimestamps(ctx, upper, TimeSpec{
+		ATime: lowerAttr.AccessTime,
+		MTime: lowerAttr.ModificationTime,
+	}); err != nil {
+		return err
+	}
+	for name := range lowerXattr {
+		value, err := lower.Getxattr(name)
+		if err != nil {
+			return err
+		}
+		if err := upper.InodeOperations.Setxattr(upper, name, value); err != nil {
+			return err
+		}
+	}
+	return nil
+}
diff --git a/pkg/sentry/fs/copy_up_test.go b/pkg/sentry/fs/copy_up_test.go
new file mode 100644
index 000000000..c3c9d963d
--- /dev/null
+++ b/pkg/sentry/fs/copy_up_test.go
@@ -0,0 +1,182 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs_test
+
+import (
+	"bytes"
+	"crypto/rand"
+	"fmt"
+	"io"
+	"sync"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	_ "gvisor.googlesource.com/gvisor/pkg/sentry/fs/tmpfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// origFileSize is the original file size. This many bytes should be
+	// copied up before the test file is modified.
+	origFileSize = 4096
+
+	// truncatedFileSize is the size to truncate all test files.
+	truncateFileSize = 10
+)
+
+// TestConcurrentCopyUp is a copy up stress test for an overlay.
+//
+// It creates a 64-level deep directory tree in the lower filesystem and
+// populates the last subdirectory with 64 files containing random content:
+//
+//    /lower
+//      /sudir0/.../subdir63/
+//                     /file0
+//                     ...
+//                     /file63
+//
+// The files are truncated concurrently by 4 goroutines per file.
+// These goroutines contend with copying up all parent 64 subdirectories
+// as well as the final file content.
+//
+// At the end of the test, we assert that the files respect the new truncated
+// size and contain the content we expect.
+func TestConcurrentCopyUp(t *testing.T) {
+	ctx := contexttest.Context(t)
+	files := makeOverlayTestFiles(t)
+
+	var wg sync.WaitGroup
+	for _, file := range files {
+		for i := 0; i < 4; i++ {
+			wg.Add(1)
+			go func(o *overlayTestFile) {
+				if err := o.File.Dirent.Inode.Truncate(ctx, o.File.Dirent, truncateFileSize); err != nil {
+					t.Fatalf("failed to copy up: %v", err)
+				}
+				wg.Done()
+			}(file)
+		}
+	}
+	wg.Wait()
+
+	for _, file := range files {
+		got := make([]byte, origFileSize)
+		n, err := file.File.Readv(ctx, usermem.BytesIOSequence(got))
+		if int(n) != truncateFileSize {
+			t.Fatalf("read %d bytes from file, want %d", n, truncateFileSize)
+		}
+		if err != nil && err != io.EOF {
+			t.Fatalf("read got error %v, want nil", err)
+		}
+		if !bytes.Equal(got[:n], file.content[:truncateFileSize]) {
+			t.Fatalf("file content is %v, want %v", got[:n], file.content[:truncateFileSize])
+		}
+	}
+}
+
+type overlayTestFile struct {
+	File    *fs.File
+	name    string
+	content []byte
+}
+
+func makeOverlayTestFiles(t *testing.T) []*overlayTestFile {
+	ctx := contexttest.Context(t)
+
+	// Create a lower tmpfs mount.
+	fsys, _ := fs.FindFilesystem("tmpfs")
+	lower, err := fsys.Mount(contexttest.Context(t), "", fs.MountSourceFlags{}, "")
+	if err != nil {
+		t.Fatalf("failed to mount tmpfs: %v", err)
+	}
+	lowerRoot := fs.NewDirent(lower, "")
+
+	// Make a deep set of subdirectories that everyone shares.
+	next := lowerRoot
+	for i := 0; i < 64; i++ {
+		name := fmt.Sprintf("subdir%d", i)
+		err := next.CreateDirectory(ctx, lowerRoot, name, fs.FilePermsFromMode(0777))
+		if err != nil {
+			t.Fatalf("failed to create dir %q: %v", name, err)
+		}
+		next, err = next.Walk(ctx, lowerRoot, name)
+		if err != nil {
+			t.Fatalf("failed to walk to %q: %v", name, err)
+		}
+	}
+
+	// Make a bunch of files in the last directory.
+	var files []*overlayTestFile
+	for i := 0; i < 64; i++ {
+		name := fmt.Sprintf("file%d", i)
+		f, err := next.Create(ctx, next, name, fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+		if err != nil {
+			t.Fatalf("failed to create file %q: %v", name, err)
+		}
+		defer f.DecRef()
+
+		relname, _ := f.Dirent.FullName(lowerRoot)
+
+		o := &overlayTestFile{
+			name:    relname,
+			content: make([]byte, origFileSize),
+		}
+
+		if _, err := rand.Read(o.content); err != nil {
+			t.Fatalf("failed to read from /dev/urandom: %v", err)
+		}
+
+		if _, err := f.Writev(ctx, usermem.BytesIOSequence(o.content)); err != nil {
+			t.Fatalf("failed to write content to file %q: %v", name, err)
+		}
+
+		files = append(files, o)
+	}
+
+	// Create an empty upper tmpfs mount which we will copy up into.
+	upper, err := fsys.Mount(ctx, "", fs.MountSourceFlags{}, "")
+	if err != nil {
+		t.Fatalf("failed to mount tmpfs: %v", err)
+	}
+
+	// Construct an overlay root.
+	overlay, err := fs.NewOverlayRoot(ctx, upper, lower, fs.MountSourceFlags{})
+	if err != nil {
+		t.Fatalf("failed to construct overlay root: %v", err)
+	}
+
+	// Create a MountNamespace to traverse the file system.
+	mns, err := fs.NewMountNamespace(ctx, overlay)
+	if err != nil {
+		t.Fatalf("failed to construct mount manager: %v", err)
+	}
+
+	// Walk to all of the files in the overlay, open them readable.
+	for _, f := range files {
+		d, err := mns.FindInode(ctx, mns.Root(), mns.Root(), f.name, 0)
+		if err != nil {
+			t.Fatalf("failed to find %q: %v", f.name, err)
+		}
+		defer d.DecRef()
+
+		f.File, err = d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true})
+		if err != nil {
+			t.Fatalf("failed to open file %q readable: %v", f.name, err)
+		}
+	}
+
+	return files
+}
diff --git a/pkg/sentry/fs/dentry.go b/pkg/sentry/fs/dentry.go
new file mode 100644
index 000000000..d42e8da81
--- /dev/null
+++ b/pkg/sentry/fs/dentry.go
@@ -0,0 +1,232 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"sort"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+)
+
+// DentAttr is the metadata of a directory entry. It is a subset of StableAttr.
+type DentAttr struct {
+	// Type is the InodeType of an Inode.
+	Type InodeType
+
+	// InodeID uniquely identifies an Inode on a device.
+	InodeID uint64
+}
+
+// GenericDentAttr returns a generic DentAttr where:
+//
+// Type == nt
+// InodeID == the inode id of a new inode on device.
+func GenericDentAttr(nt InodeType, device *device.Device) DentAttr {
+	return DentAttr{
+		Type:    nt,
+		InodeID: device.NextIno(),
+	}
+}
+
+// DentrySerializer serializes a directory entry.
+type DentrySerializer interface {
+	// CopyOut serializes a directory entry based on its name and attributes.
+	CopyOut(name string, attributes DentAttr) error
+
+	// Written returns the number of bytes written.
+	Written() int
+}
+
+// CollectEntriesSerializer copies DentAttrs to Entries. The order in
+// which entries are encountered is preserved in Order.
+type CollectEntriesSerializer struct {
+	Entries map[string]DentAttr
+	Order   []string
+}
+
+// CopyOut implements DentrySerializer.CopyOut.
+func (c *CollectEntriesSerializer) CopyOut(name string, attr DentAttr) error {
+	if c.Entries == nil {
+		c.Entries = make(map[string]DentAttr)
+	}
+	c.Entries[name] = attr
+	c.Order = append(c.Order, name)
+	return nil
+}
+
+// Written implements DentrySerializer.Written.
+func (c *CollectEntriesSerializer) Written() int {
+	return len(c.Entries)
+}
+
+// DirCtx is used by node.Readdir to emit directory entries.  It is not
+// thread-safe.
+type DirCtx struct {
+	// Serializer is used to serialize the node attributes.
+	Serializer DentrySerializer
+
+	// attrs are DentAttrs
+	attrs map[string]DentAttr
+
+	// DirCursor is the directory cursor.
+	// TODO: Once Handles are removed this can just live in the
+	// respective FileOperations implementations and not need to get
+	// plumbed everywhere.
+	DirCursor *string
+}
+
+// DirEmit is called for each directory entry.
+func (c *DirCtx) DirEmit(name string, attr DentAttr) error {
+	if c.Serializer != nil {
+		if err := c.Serializer.CopyOut(name, attr); err != nil {
+			return err
+		}
+	}
+	if c.attrs == nil {
+		c.attrs = make(map[string]DentAttr)
+	}
+	c.attrs[name] = attr
+	return nil
+}
+
+// DentAttrs returns a map of DentAttrs corresponding to the emitted directory
+// entries.
+func (c *DirCtx) DentAttrs() map[string]DentAttr {
+	if c.attrs == nil {
+		c.attrs = make(map[string]DentAttr)
+	}
+	return c.attrs
+}
+
+// GenericReaddir serializes DentAttrs based on a SortedDentryMap that must
+// contain _all_ up-to-date DentAttrs under a directory. If ctx.DirCursor is
+// not nil, it is updated to the name of the last DentAttr that was
+// successfully serialized.
+//
+// Returns the number of entries serialized.
+func GenericReaddir(ctx *DirCtx, s *SortedDentryMap) (int, error) {
+	// Retrieve the next directory entries.
+	var names []string
+	var entries map[string]DentAttr
+	if ctx.DirCursor != nil {
+		names, entries = s.GetNext(*ctx.DirCursor)
+	} else {
+		names, entries = s.GetAll()
+	}
+
+	// Try to serialize each entry.
+	var serialized int
+	for _, name := range names {
+		// Skip "" per POSIX. Skip "." and ".." which will be added by Dirent.Readdir.
+		if name == "" || name == "." || name == ".." {
+			continue
+		}
+
+		// Emit the directory entry.
+		if err := ctx.DirEmit(name, entries[name]); err != nil {
+			// Return potentially a partial serialized count.
+			return serialized, err
+		}
+
+		// We successfully serialized this entry.
+		serialized++
+
+		// Update the cursor with the name of the entry last serialized.
+		if ctx.DirCursor != nil {
+			*ctx.DirCursor = name
+		}
+	}
+
+	// Everything was serialized.
+	return serialized, nil
+}
+
+// SortedDentryMap is a sorted map of names and fs.DentAttr entries.
+type SortedDentryMap struct {
+	// names is always kept in sorted-order.
+	names []string
+
+	// entries maps names to fs.DentAttrs.
+	entries map[string]DentAttr
+}
+
+// NewSortedDentryMap maintains entries in name sorted order.
+func NewSortedDentryMap(entries map[string]DentAttr) *SortedDentryMap {
+	s := &SortedDentryMap{
+		names:   make([]string, 0, len(entries)),
+		entries: entries,
+	}
+	// Don't allow s.entries to be nil, because nil maps arn't Saveable.
+	if s.entries == nil {
+		s.entries = make(map[string]DentAttr)
+	}
+
+	// Collect names from entries and sort them.
+	for name := range s.entries {
+		s.names = append(s.names, name)
+	}
+	sort.Strings(s.names)
+	return s
+}
+
+// GetAll returns all names and entries in s.
+func (s *SortedDentryMap) GetAll() ([]string, map[string]DentAttr) {
+	return s.names, s.entries
+}
+
+// GetNext returns names after cursor in s and all entries.
+func (s *SortedDentryMap) GetNext(cursor string) ([]string, map[string]DentAttr) {
+	i := sort.SearchStrings(s.names, cursor)
+	if i == len(s.names) {
+		return nil, s.entries
+	}
+
+	// Return everything strictly after the cursor.
+	if s.names[i] == cursor {
+		i++
+	}
+	return s.names[i:], s.entries
+}
+
+// Add adds an entry with the given name to the map, preserving sort order.  If
+// name already exists in the map, its entry will be overwritten.
+func (s *SortedDentryMap) Add(name string, entry DentAttr) {
+	if _, ok := s.entries[name]; !ok {
+		// Map does not yet contain an entry with this name.  We must
+		// insert it in s.names at the appropriate spot.
+		i := sort.SearchStrings(s.names, name)
+		s.names = append(s.names, "")
+		copy(s.names[i+1:], s.names[i:])
+		s.names[i] = name
+	}
+	s.entries[name] = entry
+}
+
+// Remove removes an entry with the given name from the map, preserving sort order.
+func (s *SortedDentryMap) Remove(name string) {
+	if _, ok := s.entries[name]; !ok {
+		return
+	}
+	i := sort.SearchStrings(s.names, name)
+	copy(s.names[i:], s.names[i+1:])
+	s.names = s.names[:len(s.names)-1]
+	delete(s.entries, name)
+}
+
+// Contains reports whether the map contains an entry with the given name.
+func (s *SortedDentryMap) Contains(name string) bool {
+	_, ok := s.entries[name]
+	return ok
+}
diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD
new file mode 100644
index 000000000..42049ecb5
--- /dev/null
+++ b/pkg/sentry/fs/dev/BUILD
@@ -0,0 +1,53 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "dev_state",
+    srcs = [
+        "dev.go",
+        "fs.go",
+        "full.go",
+        "null.go",
+        "random.go",
+    ],
+    out = "dev_state.go",
+    package = "dev",
+)
+
+go_library(
+    name = "dev",
+    srcs = [
+        "dev.go",
+        "dev_state.go",
+        "device.go",
+        "fs.go",
+        "full.go",
+        "null.go",
+        "random.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/dev",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/ashmem",
+        "//pkg/sentry/fs/binder",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/fs/tmpfs",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/mm",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/fs/dev/dev.go b/pkg/sentry/fs/dev/dev.go
new file mode 100644
index 000000000..36c61bfc2
--- /dev/null
+++ b/pkg/sentry/fs/dev/dev.go
@@ -0,0 +1,122 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package dev provides a filesystem with simple devices.
+package dev
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ashmem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/binder"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/tmpfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Dev is the root node.
+type Dev struct {
+	ramfs.Dir
+}
+
+func newCharacterDevice(iops fs.InodeOperations, msrc *fs.MountSource) *fs.Inode {
+	return fs.NewInode(iops, msrc, fs.StableAttr{
+		DeviceID:  devDevice.DeviceID(),
+		InodeID:   devDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.CharacterDevice,
+	})
+}
+
+func newDirectory(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	iops := &ramfs.Dir{}
+	iops.InitDir(ctx, map[string]*fs.Inode{}, fs.RootOwner, fs.FilePermsFromMode(0555))
+	return fs.NewInode(iops, msrc, fs.StableAttr{
+		DeviceID:  devDevice.DeviceID(),
+		InodeID:   devDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Directory,
+	})
+}
+
+func newSymlink(ctx context.Context, target string, msrc *fs.MountSource) *fs.Inode {
+	iops := &ramfs.Symlink{}
+	iops.InitSymlink(ctx, fs.RootOwner, target)
+	return fs.NewInode(iops, msrc, fs.StableAttr{
+		DeviceID:  devDevice.DeviceID(),
+		InodeID:   devDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Symlink,
+	})
+}
+
+// New returns the root node of a device filesystem.
+func New(ctx context.Context, msrc *fs.MountSource, binderEnabled bool, ashmemEnabled bool) *fs.Inode {
+	d := &Dev{}
+
+	contents := map[string]*fs.Inode{
+		"fd":     newSymlink(ctx, "/proc/self/fd", msrc),
+		"stdin":  newSymlink(ctx, "/proc/self/fd/0", msrc),
+		"stdout": newSymlink(ctx, "/proc/self/fd/1", msrc),
+		"stderr": newSymlink(ctx, "/proc/self/fd/2", msrc),
+
+		"null": newCharacterDevice(newNullDevice(ctx, fs.RootOwner, 0666), msrc),
+		"zero": newCharacterDevice(newZeroDevice(ctx, fs.RootOwner, 0666), msrc),
+		"full": newCharacterDevice(newFullDevice(ctx, fs.RootOwner, 0666), msrc),
+
+		// This is not as good as /dev/random in linux because go
+		// runtime uses sys_random and /dev/urandom internally.
+		// According to 'man 4 random', this will be sufficient unless
+		// application uses this to generate long-lived GPG/SSL/SSH
+		// keys.
+		"random":  newCharacterDevice(newRandomDevice(ctx, fs.RootOwner, 0444), msrc),
+		"urandom": newCharacterDevice(newRandomDevice(ctx, fs.RootOwner, 0444), msrc),
+
+		"shm": tmpfs.NewDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0777), msrc, platform.FromContext(ctx)),
+
+		// A devpts is typically mounted at /dev/pts to provide
+		// pseudoterminal support. Place an empty directory there for
+		// the devpts to be mounted over.
+		"pts": newDirectory(ctx, msrc),
+		// Similarly, applications expect a ptmx device at /dev/ptmx
+		// connected to the terminals provided by /dev/pts/. Rather
+		// than creating a device directly (which requires a hairy
+		// lookup on open to determine if a devpts exists), just create
+		// a symlink to the ptmx provided by devpts. (The Linux devpts
+		// documentation recommends this).
+		//
+		// If no devpts is mounted, this will simply be a dangling
+		// symlink, which is fine.
+		"ptmx": newSymlink(ctx, "pts/ptmx", msrc),
+	}
+
+	if binderEnabled {
+		binder := binder.NewDevice(ctx, fs.RootOwner, fs.FilePermsFromMode(0666))
+		contents["binder"] = newCharacterDevice(binder, msrc)
+	}
+
+	if ashmemEnabled {
+		ashmem := ashmem.NewDevice(ctx, fs.RootOwner, fs.FilePermsFromMode(0666))
+		contents["ashmem"] = newCharacterDevice(ashmem, msrc)
+	}
+
+	d.InitDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555))
+	return fs.NewInode(d, msrc, fs.StableAttr{
+		DeviceID:  devDevice.DeviceID(),
+		InodeID:   devDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Directory,
+	})
+}
diff --git a/pkg/sentry/fs/dev/device.go b/pkg/sentry/fs/dev/device.go
new file mode 100644
index 000000000..9d935e008
--- /dev/null
+++ b/pkg/sentry/fs/dev/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dev
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// devDevice is the pseudo-filesystem device.
+var devDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/dev/fs.go b/pkg/sentry/fs/dev/fs.go
new file mode 100644
index 000000000..4945ac962
--- /dev/null
+++ b/pkg/sentry/fs/dev/fs.go
@@ -0,0 +1,90 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dev
+
+import (
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Optional key containing boolean flag which specifies if Android Binder IPC should be enabled.
+const binderEnabledKey = "binder_enabled"
+
+// Optional key containing boolean flag which specifies if Android ashmem should be enabled.
+const ashmemEnabledKey = "ashmem_enabled"
+
+// filesystem is a devtmpfs.
+type filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&filesystem{})
+}
+
+// FilesystemName is the name underwhich the filesystem is registered.
+// Name matches drivers/base/devtmpfs.c:dev_fs_type.name.
+const FilesystemName = "devtmpfs"
+
+// Name is the name of the file system.
+func (*filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount allows users to mount(2) this file system.
+func (*filesystem) AllowUserMount() bool {
+	return true
+}
+
+// Flags returns that there is nothing special about this file system.
+//
+// In Linux, devtmpfs does the same thing.
+func (*filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns a devtmpfs root that can be positioned in the vfs.
+func (f *filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// device is always ignored.
+	// devtmpfs backed by ramfs ignores bad options. See fs/ramfs/inode.c:ramfs_parse_options.
+	//  -> we should consider parsing the mode and backing devtmpfs by this.
+
+	// Parse generic comma-separated key=value options.
+	options := fs.GenericMountSourceOptions(data)
+
+	// binerEnabledKey is optional and binder is disabled by default.
+	binderEnabled := false
+	if beStr, exists := options[binderEnabledKey]; exists {
+		var err error
+		binderEnabled, err = strconv.ParseBool(beStr)
+		if err != nil {
+			return nil, syserror.EINVAL
+		}
+	}
+
+	// ashmemEnabledKey is optional and ashmem is disabled by default.
+	ashmemEnabled := false
+	if aeStr, exists := options[ashmemEnabledKey]; exists {
+		var err error
+		ashmemEnabled, err = strconv.ParseBool(aeStr)
+		if err != nil {
+			return nil, syserror.EINVAL
+		}
+	}
+
+	// Construct the devtmpfs root.
+	return New(ctx, fs.NewNonCachingMountSource(f, flags), binderEnabled, ashmemEnabled), nil
+}
diff --git a/pkg/sentry/fs/dev/full.go b/pkg/sentry/fs/dev/full.go
new file mode 100644
index 000000000..e13eb6c03
--- /dev/null
+++ b/pkg/sentry/fs/dev/full.go
@@ -0,0 +1,53 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dev
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// fullDevice is used to implement /dev/full.
+type fullDevice struct {
+	ramfs.Entry
+}
+
+func newFullDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *fullDevice {
+	f := &fullDevice{}
+	f.InitEntry(ctx, owner, fs.FilePermsFromMode(mode))
+	return f
+}
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev by
+// returining ENOSPC.
+func (f *fullDevice) DeprecatedPwritev(_ context.Context, _ usermem.IOSequence, _ int64) (int64, error) {
+	return 0, syserror.ENOSPC
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (f *fullDevice) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, _ int64) (int64, error) {
+	return dst.ZeroOut(ctx, math.MaxInt64)
+}
+
+// Truncate should be simply ignored for character devices on linux.
+func (f *fullDevice) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
diff --git a/pkg/sentry/fs/dev/null.go b/pkg/sentry/fs/dev/null.go
new file mode 100644
index 000000000..66b8ba967
--- /dev/null
+++ b/pkg/sentry/fs/dev/null.go
@@ -0,0 +1,96 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dev
+
+import (
+	"io"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type nullDevice struct {
+	ramfs.Entry
+}
+
+func newNullDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *nullDevice {
+	n := &nullDevice{}
+	n.InitEntry(ctx, owner, fs.FilePermsFromMode(mode))
+	return n
+}
+
+// DeprecatedPreadv reads data from the device.
+func (n *nullDevice) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	return 0, io.EOF
+}
+
+// DeprecatedPwritev discards writes.
+func (n *nullDevice) DeprecatedPwritev(_ context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	return src.NumBytes(), nil
+}
+
+// Truncate should be simply ignored for character devices on linux.
+func (n *nullDevice) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
+
+type zeroDevice struct {
+	nullDevice
+}
+
+func newZeroDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *zeroDevice {
+	zd := &zeroDevice{}
+	zd.InitEntry(ctx, owner, fs.FilePermsFromMode(mode))
+	return zd
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (zd *zeroDevice) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	return dst.ZeroOut(ctx, math.MaxInt64)
+}
+
+// GetFile overrides ramfs.Entry.GetFile and returns a zeroFile instead.
+func (zd *zeroDevice) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	// Allow pread(2) and pwrite(2) on this file.
+	flags.Pread = true
+	flags.Pwrite = true
+
+	return fs.NewFile(ctx, dirent, flags, &zeroFileOperations{
+		FileOperations: &fsutil.Handle{HandleOperations: dirent.Inode.HandleOps()},
+	}), nil
+}
+
+type zeroFileOperations struct {
+	fs.FileOperations
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (*zeroFileOperations) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	m, err := mm.NewSharedAnonMappable(opts.Length, platform.FromContext(ctx))
+	if err != nil {
+		return err
+	}
+	opts.MappingIdentity = m
+	opts.Mappable = m
+	return nil
+}
diff --git a/pkg/sentry/fs/dev/random.go b/pkg/sentry/fs/dev/random.go
new file mode 100644
index 000000000..0402f9355
--- /dev/null
+++ b/pkg/sentry/fs/dev/random.go
@@ -0,0 +1,55 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package dev
+
+import (
+	"crypto/rand"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type randomDevice struct {
+	ramfs.Entry
+}
+
+func newRandomDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMode) *randomDevice {
+	r := &randomDevice{}
+	r.InitEntry(ctx, owner, fs.FilePermsFromMode(mode))
+	return r
+}
+
+// DeprecatedPreadv reads random data.
+func (*randomDevice) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	return dst.CopyOutFrom(ctx, safemem.FromIOReader{rand.Reader})
+}
+
+// DeprecatedPwritev implements fs.HandleOperations.DeprecatedPwritev.
+func (*randomDevice) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	// On Linux, "Writing to /dev/random or /dev/urandom will update the
+	// entropy pool with the data written, but this will not result in a higher
+	// entropy count" - random(4). We don't need to support this, but we do
+	// need to support the write, so just make it a no-op a la /dev/null.
+	return src.NumBytes(), nil
+}
+
+// Truncate should be simply ignored for character devices on linux.
+func (r *randomDevice) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go
new file mode 100644
index 000000000..a75c7ea7e
--- /dev/null
+++ b/pkg/sentry/fs/dirent.go
@@ -0,0 +1,1605 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"path"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+type globalDirentMap struct {
+	mu      sync.Mutex
+	dirents map[*Dirent]struct{}
+}
+
+func (g *globalDirentMap) add(d *Dirent) {
+	g.mu.Lock()
+	g.dirents[d] = struct{}{}
+	g.mu.Unlock()
+}
+
+func (g *globalDirentMap) remove(d *Dirent) {
+	g.mu.Lock()
+	delete(g.dirents, d)
+	g.mu.Unlock()
+}
+
+// allDirents keeps track of all Dirents that need to be considered in
+// Save/Restore for inode mappings.
+//
+// Because inodes do not hold paths, but inodes for external file systems map
+// to an external path, every user-visible Dirent is stored in this map and
+// iterated through upon save to keep inode ID -> restore path mappings.
+var allDirents = globalDirentMap{
+	dirents: map[*Dirent]struct{}{},
+}
+
+// renameMu protects the parent of *all* Dirents. (See explanation in
+// lockForRename.)
+//
+// See fs.go for lock ordering.
+var renameMu sync.RWMutex
+
+// Dirent holds an Inode in memory.
+//
+// A Dirent may be negative or positive:
+//
+// A negative Dirent contains a nil Inode and indicates that a path does not exist. This
+// is a convention taken from the Linux dcache, see fs/dcache.c. A negative Dirent remains
+// cached until a create operation replaces it with a positive Dirent. A negative Dirent
+// always has one reference owned by its parent and takes _no_ reference on its parent. This
+// ensures that its parent can be unhashed regardless of negative children.
+//
+// A positive Dirent contains a non-nil Inode. It remains cached for as long as there remain
+// references to it. A positive Dirent always takes a reference on its parent.
+//
+// A Dirent may be a root Dirent (parent is nil) or be parented (non-nil parent).
+//
+// Dirents currently do not attempt to free entries that lack application references under
+// memory pressure.
+type Dirent struct {
+	// AtomicRefCount is our reference count.
+	refs.AtomicRefCount
+
+	// userVisible indicates whether the Dirent is visible to the user or
+	// not.  Only user-visible Dirents should save inode mappings in
+	// save/restore, as only they hold the real path to the underlying
+	// inode.
+	//
+	// See newDirent and Dirent.afterLoad.
+	userVisible bool
+
+	// Inode is the underlying file object.
+	//
+	// Inode is exported currently to assist in implementing overlay Inodes (where a
+	// Inode.InodeOperations.Lookup may need to merge the Inode contained in a positive Dirent with
+	// another Inode). This is normally done before the Dirent is parented (there are
+	// no external references to it).
+	//
+	// Other objects in the VFS may take a reference to this Inode but only while holding
+	// a reference to this Dirent.
+	Inode *Inode
+
+	// name is the name (i.e. basename) of this entry.
+	//
+	// N.B. name is protected by parent.mu, not this node's mu!
+	name string
+
+	// parent is the parent directory.
+	//
+	// We hold a hard reference to the parent.
+	//
+	// parent is protected by renameMu.
+	parent *Dirent
+
+	// deleted may be set atomically when removed.
+	deleted int32 `state:"nosave"`
+
+	// frozen indicates this entry can't walk to unknown nodes.
+	frozen bool
+
+	// mounted is true if Dirent is a mount point, similar to include/linux/dcache.h:DCACHE_MOUNTED.
+	mounted bool
+
+	// direntEntry identifies this Dirent as an element in a DirentCache. DirentCaches
+	// and their contents are not saved.
+	direntEntry `state:"nosave"`
+
+	// dirMu is a read-write mutex that protects caching decisions made by directory operations.
+	// Lock ordering: dirMu must be taken before mu (see below). Details:
+	//
+	// dirMu does not participate in Rename; instead mu and renameMu are used, see lockForRename.
+	//
+	// Creation and Removal operations must be synchronized with Walk to prevent stale negative
+	// caching. Note that this requirement is not specific to a _Dirent_ doing negative caching.
+	// The following race exists at any level of the VFS:
+	//
+	// For an object D that represents a directory, containing a cache of non-existent paths,
+	// protected by D.cacheMu:
+	//
+	// T1:                       T2:
+	//                           D.lookup(name)
+	//                           --> ENOENT
+	// D.create(name)
+	// --> success
+	// D.cacheMu.Lock
+	//   delete(D.cache, name)
+	// D.cacheMu.Unlock
+	//                           D.cacheMu.Lock
+	//                             D.cache[name] = true
+	//                           D.cacheMu.Unlock
+	//
+	// D.lookup(name)
+	// D.cacheMu.Lock
+	//   if D.cache[name] {
+	//   --> ENOENT (wrong)
+	//   }
+	// D.cacheMu.Lock
+	//
+	// Correct:
+	//
+	// T1:                       T2:
+	//                           D.cacheMu.Lock
+	//                             D.lookup(name)
+	//                             --> ENOENT
+	//                             D.cache[name] = true
+	//                           D.cacheMu.Unlock
+	// D.cacheMu.Lock
+	//   D.create(name)
+	//   --> success
+	//   delete(D.cache, name)
+	// D.cacheMu.Unlock
+	//
+	// D.cacheMu.Lock
+	//   D.lookup(name)
+	//   --> EXISTS (right)
+	// D.cacheMu.Unlock
+	//
+	// Note that the above "correct" solution causes too much lock contention: all lookups are
+	// synchronized with each other. This is a problem because lookups are involved in any VFS
+	// path operation.
+	//
+	// A Dirent diverges from the single D.cacheMu and instead uses two locks: dirMu to protect
+	// concurrent creation/removal/lookup caching, and mu to protect the Dirent's children map
+	// in general.
+	//
+	// This allows for concurrent Walks to be executed in order to pipeline lookups. For instance
+	// for a hot directory /a/b, threads T1, T2, T3 will only block on each other update the
+	// children map of /a/b when their individual lookups complete.
+	//
+	// T1:           T2:           T3:
+	// stat(/a/b/c)  stat(/a/b/d)  stat(/a/b/e)
+	dirMu sync.RWMutex `state:"nosave"`
+
+	// mu protects the below fields. Lock ordering: mu must be taken after dirMu.
+	mu sync.Mutex `state:"nosave"`
+
+	// children are cached via weak references.
+	children map[string]*refs.WeakRef
+}
+
+// NewDirent returns a new root Dirent, taking the caller's reference on inode. The caller
+// holds the only reference to the Dirent. Parents may call hashChild to parent this Dirent.
+func NewDirent(inode *Inode, name string) *Dirent {
+	d := newDirent(inode, name)
+	allDirents.add(d)
+	d.userVisible = true
+	return d
+}
+
+// NewTransientDirent creates a transient Dirent that shouldn't actually be
+// visible to users.
+func NewTransientDirent(inode *Inode) *Dirent {
+	return newDirent(inode, "transient")
+}
+
+func newDirent(inode *Inode, name string) *Dirent {
+	// The Dirent needs to maintain one reference to MountSource.
+	if inode != nil {
+		inode.MountSource.IncDirentRefs()
+	}
+	return &Dirent{
+		Inode:    inode,
+		name:     name,
+		children: make(map[string]*refs.WeakRef),
+	}
+}
+
+// NewNegativeDirent returns a new root negative Dirent. Otherwise same as NewDirent.
+func NewNegativeDirent(name string) *Dirent {
+	return newDirent(nil, name)
+}
+
+// IsRoot returns true if d is a root Dirent.
+func (d *Dirent) IsRoot() bool {
+	return d.parent == nil
+}
+
+// IsNegative returns true if d represents a path that does not exist.
+func (d *Dirent) IsNegative() bool {
+	return d.Inode == nil
+}
+
+// hashChild will hash child into the children list of its new parent d, carrying over
+// any "frozen" state from d.
+//
+// Returns (*WeakRef, true) if hashing child caused a Dirent to be unhashed. The caller must
+// validate the returned unhashed weak reference. Common cases:
+//
+// * Remove: hashing a negative Dirent unhashes a positive Dirent (unimplemented).
+// * Create: hashing a positive Dirent unhashes a negative Dirent.
+// * Lookup: hashing any Dirent should not unhash any other Dirent.
+//
+// Preconditions:
+// * d.mu must be held.
+// * child must be a root Dirent.
+func (d *Dirent) hashChild(child *Dirent) (*refs.WeakRef, bool) {
+	if !child.IsRoot() {
+		panic("hashChild must be a root Dirent")
+	}
+
+	// Assign parentage.
+	child.parent = d
+
+	// Avoid letting negative Dirents take a reference on their parent; these Dirents
+	// don't have a role outside of the Dirent cache and should not keep their parent
+	// indefinitely pinned.
+	if !child.IsNegative() {
+		// Positive dirents must take a reference on their parent.
+		d.IncRef()
+	}
+
+	// Carry over parent's frozen state.
+	child.frozen = d.frozen
+
+	return d.hashChildParentSet(child)
+}
+
+// hashChildParentSet will rehash child into the children list of its parent d.
+//
+// Assumes that child.parent = d already.
+func (d *Dirent) hashChildParentSet(child *Dirent) (*refs.WeakRef, bool) {
+	if child.parent != d {
+		panic("hashChildParentSet assumes the child already belongs to the parent")
+	}
+
+	// Save any replaced child so our caller can validate it.
+	old, ok := d.children[child.name]
+
+	// Hash the child.
+	d.children[child.name] = refs.NewWeakRef(child, nil)
+
+	// Return any replaced child.
+	return old, ok
+}
+
+// SyncAll iterates through mount points under d and writes back their buffered
+// modifications to filesystems.
+func (d *Dirent) SyncAll(ctx context.Context) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	// For negative Dirents there is nothing to sync. By definition these are
+	// leaves (there is nothing left to traverse).
+	if d.IsNegative() {
+		return
+	}
+
+	// There is nothing to sync for a read-only filesystem.
+	if !d.Inode.MountSource.Flags.ReadOnly {
+		// FIXME: This should be a mount traversal, not a
+		// Dirent traversal, because some Inodes that need to be synced
+		// may no longer be reachable by name (after sys_unlink).
+		//
+		// Write out metadata, dirty page cached pages, and sync disk/remote
+		// caches.
+		d.Inode.WriteOut(ctx)
+	}
+
+	// Continue iterating through other mounted filesystems.
+	for _, w := range d.children {
+		if child := w.Get(); child != nil {
+			child.(*Dirent).SyncAll(ctx)
+			child.DecRef()
+		}
+	}
+}
+
+// FullName returns the fully-qualified name and a boolean value representing
+// whether this Dirent was a descendant of root.
+// If the root argument is nil it is assumed to be the root of the Dirent tree.
+func (d *Dirent) FullName(root *Dirent) (string, bool) {
+	renameMu.RLock()
+	defer renameMu.RUnlock()
+	return d.fullName(root)
+}
+
+// fullName returns the fully-qualified name and a boolean value representing
+// if the root node was reachable from this Dirent.
+func (d *Dirent) fullName(root *Dirent) (string, bool) {
+	if d == root {
+		return "/", true
+	}
+
+	if d.IsRoot() {
+		if root != nil {
+			// We reached the top of the Dirent tree but did not encounter
+			// the given root. Return false for reachable so the caller
+			// can handle this situation accordingly.
+			return d.name, false
+		}
+		return d.name, true
+	}
+
+	// Traverse up to parent.
+	d.parent.mu.Lock()
+	name := d.name
+	d.parent.mu.Unlock()
+	parentName, reachable := d.parent.fullName(root)
+	s := path.Join(parentName, name)
+	if atomic.LoadInt32(&d.deleted) != 0 {
+		return s + " (deleted)", reachable
+	}
+	return s, reachable
+}
+
+func (d *Dirent) freeze() {
+	if d.frozen {
+		// Already frozen.
+		return
+	}
+	d.frozen = true
+
+	// Take a reference when freezing.
+	for _, w := range d.children {
+		if child := w.Get(); child != nil {
+			// NOTE: We would normally drop the reference here. But
+			// instead we're hanging on to it.
+			ch := child.(*Dirent)
+			ch.Freeze()
+		}
+	}
+
+	// Drop all expired weak references.
+	d.flush()
+}
+
+// Freeze prevents this dirent from walking to more nodes. Freeze is applied
+// recursively to all children.
+//
+// If this particular Dirent represents a Virtual node, then Walks and Creates
+// may proceed as before.
+//
+// Freeze can only be called before the application starts running, otherwise
+// the root it might be out of sync with the application root if modified by
+// sys_chroot.
+func (d *Dirent) Freeze() {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.freeze()
+}
+
+// descendantOf returns true if the receiver dirent is equal to, or a
+// descendant of, the argument dirent.
+//
+// d.mu must be held.
+func (d *Dirent) descendantOf(p *Dirent) bool {
+	if d == p {
+		return true
+	}
+	if d.IsRoot() {
+		return false
+	}
+	return d.parent.descendantOf(p)
+}
+
+// walk walks to path name starting at the dirent, and will not traverse above
+// root Dirent.
+//
+// If walkMayUnlock is true then walk can unlock d.mu to execute a slow
+// Inode.Lookup, otherwise walk will keep d.mu locked.
+//
+// Preconditions:
+// - d.mu must be held.
+// - name must must not contain "/"s.
+func (d *Dirent) walk(ctx context.Context, root *Dirent, name string, walkMayUnlock bool) (*Dirent, error) {
+	if !IsDir(d.Inode.StableAttr) {
+		return nil, syscall.ENOTDIR
+	}
+	if name == "" || name == "." {
+		d.IncRef()
+		return d, nil
+	} else if name == ".." {
+		renameMu.RLock()
+		// Respect the chroot. Note that in Linux there is no check to enforce
+		// that d is a descendant of root.
+		if d == root {
+			d.IncRef()
+			renameMu.RUnlock()
+			return d, nil
+		}
+		// Are we already at the root? Then ".." is ".".
+		if d.IsRoot() {
+			d.IncRef()
+			renameMu.RUnlock()
+			return d, nil
+		}
+		d.parent.IncRef()
+		renameMu.RUnlock()
+		return d.parent, nil
+	}
+
+	if w, ok := d.children[name]; ok {
+		// Try to resolve the weak reference to a hard reference.
+		if child := w.Get(); child != nil {
+			cd := child.(*Dirent)
+
+			// Is this a negative Dirent?
+			if cd.IsNegative() {
+				// Don't leak a reference; this doesn't matter as much for negative Dirents,
+				// which don't hold a hard reference on their parent (their parent holds a
+				// hard reference on them, and they contain virtually no state). But this is
+				// good house-keeping.
+				child.DecRef()
+				return nil, syscall.ENOENT
+			}
+
+			// Do we need to revalidate this child?
+			//
+			// We never allow the file system to revalidate mounts, that could cause them
+			// to unexpectedly drop out before umount.
+			if cd.mounted || !cd.Inode.MountSource.Revalidate(cd) {
+				// Good to go. This is the fast-path.
+				return cd, nil
+			}
+
+			// If we're revalidating a child, we must ensure all inotify watches release
+			// their pins on the child. Inotify doesn't properly support filesystems that
+			// revalidate dirents (since watches are lost on revalidation), but if we fail
+			// to unpin the watches child will never be GCed.
+			cd.Inode.Watches.Unpin(cd)
+
+			// This child needs to be revalidated, fallthrough to unhash it. Make sure
+			// to not leak a reference from Get().
+			//
+			// Note that previous lookups may still have a reference to this stale child;
+			// this can't be helped, but we can ensure that *new* lookups are up-to-date.
+			child.DecRef()
+		}
+
+		// Either our weak reference expired or we need to revalidate it. Unhash child first, we're
+		// about to replace it.
+		delete(d.children, name)
+		w.Drop()
+	}
+
+	// Are we allowed to do the lookup?
+	if d.frozen && !d.Inode.IsVirtual() {
+		return nil, syscall.ENOENT
+	}
+
+	// Slow path: load the InodeOperations into memory. Since this is a hot path and the lookup may be expensive,
+	// if possible release the lock and re-acquire it.
+	if walkMayUnlock {
+		d.mu.Unlock()
+	}
+	c, err := d.Inode.Lookup(ctx, name)
+	if walkMayUnlock {
+		d.mu.Lock()
+	}
+	// No dice.
+	if err != nil {
+		return nil, err
+	}
+
+	// Sanity check c, its name must be consistent.
+	if c.name != name {
+		panic(fmt.Sprintf("lookup from %q to %q returned unexpected name %q", d.name, name, c.name))
+	}
+
+	// Now that we have the lock again, check if we raced.
+	if w, ok := d.children[name]; ok {
+		// Someone else looked up or created a child at name before us.
+		if child := w.Get(); child != nil {
+			cd := child.(*Dirent)
+
+			// There are active references to the existing child, prefer it to the one we
+			// retrieved from Lookup. Likely the Lookup happened very close to the insertion
+			// of child, so considering one stale over the other is fairly arbitrary.
+			c.DecRef()
+
+			// The child that was installed could be negative.
+			if cd.IsNegative() {
+				// If so, don't leak a reference and short circuit.
+				child.DecRef()
+				return nil, syscall.ENOENT
+			}
+
+			// We make the judgement call that if c raced with cd they are close enough to have
+			// the same staleness, so we don't attempt to revalidate cd. In Linux revalidations
+			// can continue indefinitely (see fs/namei.c, retry_estale); we try to avoid this.
+			return cd, nil
+		}
+
+		// Weak reference expired. We went through a full cycle of create/destroy in the time
+		// we did the Inode.Lookup. Fully drop the weak reference and fallback to using the child
+		// we looked up.
+		delete(d.children, name)
+		w.Drop()
+	}
+
+	// Give the looked up child a parent. We cannot kick out entries, since we just checked above
+	// that there is nothing at name in d's children list.
+	if _, kicked := d.hashChild(c); kicked {
+		// Yell loudly.
+		panic(fmt.Sprintf("hashed child %q over existing child", c.name))
+	}
+
+	// Is this a negative Dirent?
+	if c.IsNegative() {
+		// Don't drop a reference on the negative Dirent, it was just installed and this is the
+		// only reference we'll ever get. d owns the reference.
+		return nil, syscall.ENOENT
+	}
+
+	// Return the positive Dirent.
+	return c, nil
+}
+
+// Walk walks to a new dirent, and will not walk higher than the given root
+// Dirent, which must not be nil.
+func (d *Dirent) Walk(ctx context.Context, root *Dirent, name string) (*Dirent, error) {
+	if root == nil {
+		panic("Dirent.Walk: root must not be nil")
+	}
+
+	d.dirMu.RLock()
+	d.mu.Lock()
+	child, err := d.walk(ctx, root, name, true /* may unlock */)
+	d.mu.Unlock()
+	d.dirMu.RUnlock()
+
+	return child, err
+}
+
+// exists returns true if name exists in relation to d.
+//
+// Preconditions: d.mu must be held.
+func (d *Dirent) exists(ctx context.Context, root *Dirent, name string) bool {
+	child, err := d.walk(ctx, root, name, true /* may unlock */)
+	if err != nil {
+		// Child may not exist.
+		return false
+	}
+	// Child exists.
+	child.DecRef()
+	return true
+}
+
+// lockDirectory should be called for any operation that changes this `d`s
+// children (creating or removing them).
+func (d *Dirent) lockDirectory() func() {
+	if d.Inode.overlay != nil {
+		// overlay copyUp may need to look at Dirent parents, and hence
+		// may need renameMu.
+		renameMu.RLock()
+		d.dirMu.Lock()
+		d.mu.Lock()
+		return func() {
+			d.mu.Unlock()
+			d.dirMu.Unlock()
+			renameMu.RUnlock()
+		}
+	}
+
+	d.dirMu.Lock()
+	d.mu.Lock()
+	return func() {
+		d.mu.Unlock()
+		d.dirMu.Unlock()
+	}
+}
+
+// Create creates a new regular file in this directory.
+func (d *Dirent) Create(ctx context.Context, root *Dirent, name string, flags FileFlags, perms FilePermissions) (*File, error) {
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	// Does something already exist?
+	if d.exists(ctx, root, name) {
+		return nil, syscall.EEXIST
+	}
+
+	// Are we frozen?
+	if d.frozen && !d.Inode.IsVirtual() {
+		return nil, syscall.ENOENT
+	}
+
+	// Try the create. We need to trust the file system to return EEXIST (or something
+	// that will translate to EEXIST) if name already exists.
+	file, err := d.Inode.Create(ctx, d, name, flags, perms)
+	if err != nil {
+		return nil, err
+	}
+	child := file.Dirent
+
+	// Sanity check c, its name must be consistent.
+	if child.name != name {
+		panic(fmt.Sprintf("create from %q to %q returned unexpected name %q", d.name, name, child.name))
+	}
+
+	// File systems cannot return a negative Dirent on Create, that makes no sense.
+	if child.IsNegative() {
+		panic(fmt.Sprintf("create from %q to %q returned negative Dirent", d.name, name))
+	}
+
+	// Hash the child into its parent. We can only kick out a Dirent if it is negative
+	// (we are replacing something that does not exist with something that now does).
+	if w, kicked := d.hashChild(child); kicked {
+		if old := w.Get(); old != nil {
+			if !old.(*Dirent).IsNegative() {
+				panic(fmt.Sprintf("hashed child %q over a positive child", child.name))
+			}
+			// Don't leak a reference.
+			old.DecRef()
+
+			// Drop d's reference.
+			old.DecRef()
+		}
+
+		// Finally drop the useless weak reference on the floor.
+		w.Drop()
+	}
+
+	d.Inode.Watches.Notify(name, linux.IN_CREATE, 0)
+
+	// Allow the file system to take extra references on c.
+	child.maybeExtendReference()
+
+	// Return the reference and the new file. When the last reference to
+	// the file is dropped, file.Dirent may no longer be cached.
+	return file, nil
+}
+
+// genericCreate executes create if name does not exist. Removes a negative Dirent at name if
+// create succeeds.
+//
+// Preconditions: d.mu must be held.
+func (d *Dirent) genericCreate(ctx context.Context, root *Dirent, name string, create func() error) error {
+	// Does something already exist?
+	if d.exists(ctx, root, name) {
+		return syscall.EEXIST
+	}
+
+	// Are we frozen?
+	if d.frozen && !d.Inode.IsVirtual() {
+		return syscall.ENOENT
+	}
+
+	// Execute the create operation.
+	if err := create(); err != nil {
+		return err
+	}
+
+	// Remove any negative Dirent. We've already asserted above with d.exists
+	// that the only thing remaining here can be a negative Dirent.
+	if w, ok := d.children[name]; ok {
+		// Same as Create.
+		if old := w.Get(); old != nil {
+			if !old.(*Dirent).IsNegative() {
+				panic(fmt.Sprintf("hashed over a positive child %q", old.(*Dirent).name))
+			}
+			// Don't leak a reference.
+			old.DecRef()
+
+			// Drop d's reference.
+			old.DecRef()
+		}
+
+		// Unhash the negative Dirent, name needs to exist now.
+		delete(d.children, name)
+
+		// Finally drop the useless weak reference on the floor.
+		w.Drop()
+	}
+
+	return nil
+}
+
+// CreateLink creates a new link in this directory.
+func (d *Dirent) CreateLink(ctx context.Context, root *Dirent, oldname, newname string) error {
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	return d.genericCreate(ctx, root, newname, func() error {
+		if err := d.Inode.CreateLink(ctx, d, oldname, newname); err != nil {
+			return err
+		}
+		d.Inode.Watches.Notify(newname, linux.IN_CREATE, 0)
+		return nil
+	})
+}
+
+// CreateHardLink creates a new hard link in this directory.
+func (d *Dirent) CreateHardLink(ctx context.Context, root *Dirent, target *Dirent, name string) error {
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	// Make sure that target does not span filesystems.
+	if d.Inode.MountSource != target.Inode.MountSource {
+		return syscall.EXDEV
+	}
+
+	return d.genericCreate(ctx, root, name, func() error {
+		if err := d.Inode.CreateHardLink(ctx, d, target, name); err != nil {
+			return err
+		}
+		target.Inode.Watches.Notify("", linux.IN_ATTRIB, 0) // Link count change.
+		d.Inode.Watches.Notify(name, linux.IN_CREATE, 0)
+		return nil
+	})
+}
+
+// CreateDirectory creates a new directory under this dirent.
+func (d *Dirent) CreateDirectory(ctx context.Context, root *Dirent, name string, perms FilePermissions) error {
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	return d.genericCreate(ctx, root, name, func() error {
+		if err := d.Inode.CreateDirectory(ctx, d, name, perms); err != nil {
+			return err
+		}
+		d.Inode.Watches.Notify(name, linux.IN_ISDIR|linux.IN_CREATE, 0)
+		return nil
+	})
+}
+
+// Bind satisfies the InodeOperations interface; otherwise same as GetFile.
+func (d *Dirent) Bind(ctx context.Context, root *Dirent, name string, socket unix.BoundEndpoint, perms FilePermissions) error {
+	d.dirMu.Lock()
+	defer d.dirMu.Unlock()
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	err := d.genericCreate(ctx, root, name, func() error {
+		if err := d.Inode.Bind(ctx, name, socket, perms); err != nil {
+			return err
+		}
+		d.Inode.Watches.Notify(name, linux.IN_CREATE, 0)
+		return nil
+	})
+	if err == syscall.EEXIST {
+		return syscall.EADDRINUSE
+	}
+	return err
+}
+
+// CreateFifo creates a new named pipe under this dirent.
+func (d *Dirent) CreateFifo(ctx context.Context, root *Dirent, name string, perms FilePermissions) error {
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	return d.genericCreate(ctx, root, name, func() error {
+		if err := d.Inode.CreateFifo(ctx, d, name, perms); err != nil {
+			return err
+		}
+		d.Inode.Watches.Notify(name, linux.IN_CREATE, 0)
+		return nil
+	})
+}
+
+// getDotAttrs returns the DentAttrs corresponding to "." and ".." directories.
+func (d *Dirent) getDotAttrs(root *Dirent) (DentAttr, DentAttr) {
+	// Get '.'.
+	sattr := d.Inode.StableAttr
+	dot := DentAttr{
+		Type:    sattr.Type,
+		InodeID: sattr.InodeID,
+	}
+
+	// Get '..'.
+	if !d.IsRoot() && d.descendantOf(root) {
+		// Dirent is a descendant of the root.  Get its parent's attrs.
+		psattr := d.parent.Inode.StableAttr
+		dotdot := DentAttr{
+			Type:    psattr.Type,
+			InodeID: psattr.InodeID,
+		}
+		return dot, dotdot
+	}
+	// Dirent is either root or not a descendant of the root.  ".." is the
+	// same as ".".
+	return dot, dot
+}
+
+// readdirFrozen returns readdir results based solely on the frozen children.
+func (d *Dirent) readdirFrozen(root *Dirent, offset int64, dirCtx *DirCtx) (int64, error) {
+	// Collect attrs for "." and  "..".
+	attrs := make(map[string]DentAttr)
+	names := []string{".", ".."}
+	attrs["."], attrs[".."] = d.getDotAttrs(root)
+
+	// Get info from all children.
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	for name, w := range d.children {
+		if child := w.Get(); child != nil {
+			defer child.DecRef()
+
+			// Skip negative children.
+			if child.(*Dirent).IsNegative() {
+				continue
+			}
+
+			sattr := child.(*Dirent).Inode.StableAttr
+			attrs[name] = DentAttr{
+				Type:    sattr.Type,
+				InodeID: sattr.InodeID,
+			}
+			names = append(names, name)
+		}
+	}
+
+	sort.Strings(names)
+
+	if int(offset) >= len(names) {
+		return offset, nil
+	}
+	names = names[int(offset):]
+	for _, name := range names {
+		if err := dirCtx.DirEmit(name, attrs[name]); err != nil {
+			return offset, err
+		}
+		offset++
+	}
+	return offset, nil
+}
+
+// DirIterator is an open directory containing directory entries that can be read.
+type DirIterator interface {
+	// IterateDir emits directory entries by calling dirCtx.EmitDir, beginning
+	// with the entry at offset and returning the next directory offset.
+	//
+	// Entries for "." and ".." must *not* be included.
+	//
+	// If the offset returned is the same as the argument offset, then
+	// nothing has been serialized.  This is equivalent to reaching EOF.
+	// In this case serializer.Written() should return 0.
+	//
+	// The order of entries to emit must be consistent between Readdir
+	// calls, and must start with the given offset.
+	//
+	// The caller must ensure that this operation is permitted.
+	IterateDir(ctx context.Context, dirCtx *DirCtx, offset int) (int, error)
+}
+
+// DirentReaddir serializes the directory entries of d including "." and "..".
+//
+// Arguments:
+//
+// * d:		the Dirent of the directory being read; required to provide "." and "..".
+// * it:	the directory iterator; which represents an open directory handle.
+// * root: 	fs root; if d is equal to the root, then '..' will refer to d.
+// * ctx: 	context provided to file systems in order to select and serialize entries.
+// * offset:	the current directory offset.
+//
+// Returns the offset of the *next* element which was not serialized.
+func DirentReaddir(ctx context.Context, d *Dirent, it DirIterator, root *Dirent, dirCtx *DirCtx, offset int64) (int64, error) {
+	offset, err := direntReaddir(ctx, d, it, root, dirCtx, offset)
+	// Serializing any directory entries at all means success.
+	if dirCtx.Serializer.Written() > 0 {
+		return offset, nil
+	}
+	return offset, err
+}
+
+func direntReaddir(ctx context.Context, d *Dirent, it DirIterator, root *Dirent, dirCtx *DirCtx, offset int64) (int64, error) {
+	if root == nil {
+		panic("Dirent.Readdir: root must not be nil")
+	}
+	if dirCtx.Serializer == nil {
+		panic("Dirent.Readdir: serializer must not be nil")
+	}
+	if d.frozen {
+		return d.readdirFrozen(root, offset, dirCtx)
+	}
+
+	// Check that this is actually a directory before emitting anything.
+	// Once we have written entries for "." and "..", future errors from
+	// IterateDir will be hidden.
+	if !IsDir(d.Inode.StableAttr) {
+		return 0, syserror.ENOTDIR
+	}
+
+	// Collect attrs for "." and "..".
+	dot, dotdot := d.getDotAttrs(root)
+
+	// Emit "." and ".." if the offset is low enough.
+	if offset == 0 {
+		// Serialize ".".
+		if err := dirCtx.DirEmit(".", dot); err != nil {
+			return offset, err
+		}
+		offset++
+	}
+	if offset == 1 {
+		// Serialize "..".
+		if err := dirCtx.DirEmit("..", dotdot); err != nil {
+			return offset, err
+		}
+		offset++
+	}
+
+	// it.IterateDir should be passed an offset that does not include the
+	// initial dot elements.  We will add them back later.
+	offset -= 2
+	newOffset, err := it.IterateDir(ctx, dirCtx, int(offset))
+	if int64(newOffset) < offset {
+		panic(fmt.Sprintf("node.Readdir returned offset %v less that input offset %v", offset, newOffset))
+	}
+	// Add the initial nodes back to the offset count.
+	newOffset += 2
+	return int64(newOffset), err
+}
+
+// flush flushes all weak references recursively, and removes any cached
+// references to children.
+//
+// Preconditions: d.mu must be held.
+func (d *Dirent) flush() {
+	expired := make(map[string]*refs.WeakRef)
+	for n, w := range d.children {
+		// Call flush recursively on each child before removing our
+		// reference on it, and removing the cache's reference.
+		if child := w.Get(); child != nil {
+			cd := child.(*Dirent)
+
+			if !cd.IsNegative() {
+				// Flush the child.
+				cd.mu.Lock()
+				cd.flush()
+				cd.mu.Unlock()
+
+				// Allow the file system to drop extra references on child.
+				cd.dropExtendedReference()
+			}
+
+			// Don't leak a reference.
+			child.DecRef()
+		}
+		// Check if the child dirent is closed, and mark it as expired if it is.
+		// We must call w.Get() again here, since the child could have been closed
+		// by the calls to flush() and cache.Remove() in the above if-block.
+		if child := w.Get(); child != nil {
+			child.DecRef()
+		} else {
+			expired[n] = w
+		}
+	}
+
+	// Remove expired entries.
+	for n, w := range expired {
+		delete(d.children, n)
+		w.Drop()
+	}
+}
+
+// Busy indicates whether this Dirent is a mount point or root dirent, or has
+// active positive children.
+//
+// This is expensive, since it flushes the children cache.
+//
+// TODO: Fix this busy-ness check.
+func (d *Dirent) Busy() bool {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	if d.mounted || d.parent == nil {
+		return true
+	}
+
+	// Flush any cached references to children that are doomed.
+	d.flush()
+
+	// Count positive children.
+	var nonNegative int
+	for _, w := range d.children {
+		if child := w.Get(); child != nil {
+			if !child.(*Dirent).IsNegative() {
+				nonNegative++
+			}
+			child.DecRef()
+		}
+	}
+	return nonNegative > 0
+}
+
+// mount mounts a new dirent with the given inode over d.
+//
+// Precondition: must be called with mm.withMountLocked held on `d`.
+func (d *Dirent) mount(ctx context.Context, inode *Inode) (newChild *Dirent, err error) {
+	// Did we race with deletion?
+	if atomic.LoadInt32(&d.deleted) != 0 {
+		return nil, syserror.ENOENT
+	}
+
+	// Refuse to mount a symlink.
+	//
+	// See Linux equivalent in fs/namespace.c:do_add_mount.
+	if IsSymlink(inode.StableAttr) {
+		return nil, syserror.EINVAL
+	}
+
+	// Are we frozen?
+	if d.parent.frozen && !d.parent.Inode.IsVirtual() {
+		return nil, syserror.ENOENT
+	}
+
+	// Dirent that'll replace d.
+	//
+	// Note that NewDirent returns with one reference taken; the reference
+	// is donated to the caller as the mount reference.
+	replacement := NewDirent(inode, d.name)
+	replacement.mounted = true
+
+	weakRef, ok := d.parent.hashChild(replacement)
+	if !ok {
+		panic("mount must mount over an existing dirent")
+	}
+	weakRef.Drop()
+
+	// Note that even though `d` is now hidden, it still holds a reference
+	// to its parent.
+	return replacement, nil
+}
+
+// unmount unmounts `d` and replaces it with the last Dirent that was in its
+// place, supplied by the MountNamespace as `replacement`.
+//
+// Precondition: must be called with mm.withMountLocked held on `d`.
+func (d *Dirent) unmount(ctx context.Context, replacement *Dirent) error {
+	// Did we race with deletion?
+	if atomic.LoadInt32(&d.deleted) != 0 {
+		return syserror.ENOENT
+	}
+
+	// Are we frozen?
+	if d.parent.frozen && !d.parent.Inode.IsVirtual() {
+		return syserror.ENOENT
+	}
+
+	// Remount our former child in its place.
+	//
+	// As replacement used to be our child, it must already have the right
+	// parent.
+	weakRef, ok := d.parent.hashChildParentSet(replacement)
+	if !ok {
+		panic("mount must mount over an existing dirent")
+	}
+	weakRef.Drop()
+
+	// d is not reachable anymore, and hence not mounted anymore.
+	d.mounted = false
+
+	// Drop mount reference.
+	d.DecRef()
+	return nil
+}
+
+// Remove removes the given file or symlink.  The root dirent is used to
+// resolve name, and must not be nil.
+func (d *Dirent) Remove(ctx context.Context, root *Dirent, name string) error {
+	// Check the root.
+	if root == nil {
+		panic("Dirent.Remove: root must not be nil")
+	}
+
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	// Are we frozen?
+	if d.frozen && !d.Inode.IsVirtual() {
+		return syscall.ENOENT
+	}
+
+	// Try to walk to the node.
+	child, err := d.walk(ctx, root, name, false /* may unlock */)
+	if err != nil {
+		// Child does not exist.
+		return err
+	}
+	defer child.DecRef()
+
+	// Remove cannot remove directories.
+	if IsDir(child.Inode.StableAttr) {
+		return syscall.EISDIR
+	}
+
+	// Remove cannot remove a mount point.
+	if child.Busy() {
+		return syscall.EBUSY
+	}
+
+	// Try to remove name on the file system.
+	if err := d.Inode.Remove(ctx, d, child); err != nil {
+		return err
+	}
+
+	// Link count changed, this only applies to non-directory nodes.
+	child.Inode.Watches.Notify("", linux.IN_ATTRIB, 0)
+
+	// Mark name as deleted and remove from children.
+	atomic.StoreInt32(&child.deleted, 1)
+	if w, ok := d.children[name]; ok {
+		delete(d.children, name)
+		w.Drop()
+	}
+
+	// Allow the file system to drop extra references on child.
+	child.dropExtendedReference()
+
+	// Finally, let inotify know the child is being unlinked. Drop any extra
+	// refs from inotify to this child dirent. This doesn't necessarily mean the
+	// watches on the underlying inode will be destroyed, since the underlying
+	// inode may have other links. If this was the last link, the events for the
+	// watch removal will be queued by the inode destructor.
+	child.Inode.Watches.MarkUnlinked()
+	child.Inode.Watches.Unpin(child)
+	d.Inode.Watches.Notify(name, linux.IN_DELETE, 0)
+
+	return nil
+}
+
+// RemoveDirectory removes the given directory.  The root dirent is used to
+// resolve name, and must not be nil.
+func (d *Dirent) RemoveDirectory(ctx context.Context, root *Dirent, name string) error {
+	// Check the root.
+	if root == nil {
+		panic("Dirent.Remove: root must not be nil")
+	}
+
+	unlock := d.lockDirectory()
+	defer unlock()
+
+	// Are we frozen?
+	if d.frozen && !d.Inode.IsVirtual() {
+		return syscall.ENOENT
+	}
+
+	// Check for dots.
+	if name == "." {
+		// Rejected as the last component by rmdir(2).
+		return syscall.EINVAL
+	}
+	if name == ".." {
+		// If d was found, then its parent is not empty.
+		return syscall.ENOTEMPTY
+	}
+
+	// Try to walk to the node.
+	child, err := d.walk(ctx, root, name, false /* may unlock */)
+	if err != nil {
+		// Child does not exist.
+		return err
+	}
+	defer child.DecRef()
+
+	// RemoveDirectory can only remove directories.
+	if !IsDir(child.Inode.StableAttr) {
+		return syscall.ENOTDIR
+	}
+
+	// Remove cannot remove a mount point.
+	if child.Busy() {
+		return syscall.EBUSY
+	}
+
+	// Try to remove name on the file system.
+	if err := d.Inode.Remove(ctx, d, child); err != nil {
+		return err
+	}
+
+	// Mark name as deleted and remove from children.
+	atomic.StoreInt32(&child.deleted, 1)
+	if w, ok := d.children[name]; ok {
+		delete(d.children, name)
+		w.Drop()
+	}
+
+	// Allow the file system to drop extra references on child.
+	child.dropExtendedReference()
+
+	// Finally, let inotify know the child is being unlinked. Drop any extra
+	// refs from inotify to this child dirent.
+	child.Inode.Watches.MarkUnlinked()
+	child.Inode.Watches.Unpin(child)
+	d.Inode.Watches.Notify(name, linux.IN_ISDIR|linux.IN_DELETE, 0)
+
+	return nil
+}
+
+// destroy closes this node and all children.
+func (d *Dirent) destroy() {
+	if d.IsNegative() {
+		// Nothing to tear-down and no parent references to drop, since a negative
+		// Dirent does not take a references on its parent, has no Inode and no children.
+		return
+	}
+
+	var wg sync.WaitGroup
+	defer wg.Wait()
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	// Drop all weak references.
+	for _, w := range d.children {
+		w.Drop()
+	}
+	d.children = nil
+
+	allDirents.remove(d)
+
+	// Drop our reference to the Inode.
+	d.Inode.DecRef()
+
+	// Allow the Dirent to be GC'ed after this point, since the Inode may still
+	// be referenced after the Dirent is destroyed (for instance by filesystem
+	// internal caches or hard links).
+	d.Inode = nil
+
+	// Drop the reference we have on our parent if we took one. renameMu doesn't need to be
+	// held because d can't be reparented without any references to it left.
+	if d.parent != nil {
+		d.parent.DecRef()
+	}
+}
+
+// IncRef increases the Dirent's refcount as well as its mount's refcount.
+//
+// IncRef implements RefCounter.IncRef.
+func (d *Dirent) IncRef() {
+	if d.Inode != nil {
+		d.Inode.MountSource.IncDirentRefs()
+	}
+	d.AtomicRefCount.IncRef()
+}
+
+// TryIncRef implements RefCounter.TryIncRef.
+func (d *Dirent) TryIncRef() bool {
+	ok := d.AtomicRefCount.TryIncRef()
+	if ok && d.Inode != nil {
+		d.Inode.MountSource.IncDirentRefs()
+	}
+	return ok
+}
+
+// DecRef decreases the Dirent's refcount and drops its reference on its mount.
+//
+// DecRef implements RefCounter.DecRef with destructor d.destroy.
+func (d *Dirent) DecRef() {
+	if d.Inode != nil {
+		// Keep mount around, since DecRef may destroy d.Inode.
+		msrc := d.Inode.MountSource
+		d.DecRefWithDestructor(d.destroy)
+		msrc.DecDirentRefs()
+	} else {
+		d.DecRefWithDestructor(d.destroy)
+	}
+}
+
+// InotifyEvent notifies all watches on the inode for this dirent and its parent
+// of potential events. The events may not actually propagate up to the user,
+// depending on the event masks. InotifyEvent automatically provides the name of
+// the current dirent as the subject of the event as required, and adds the
+// IN_ISDIR flag for dirents that refer to directories.
+func (d *Dirent) InotifyEvent(events, cookie uint32) {
+	// N.B. We don't defer the unlocks because InotifyEvent is in the hot
+	// path of all IO operations, and the defers cost too much for small IO
+	// operations.
+	renameMu.RLock()
+
+	if IsDir(d.Inode.StableAttr) {
+		events |= linux.IN_ISDIR
+	}
+
+	// The ordering below is important, Linux always notifies the parent first.
+	if d.parent != nil {
+		d.parent.Inode.Watches.Notify(d.name, events, cookie)
+	}
+	d.Inode.Watches.Notify("", events, cookie)
+
+	renameMu.RUnlock()
+}
+
+// maybeExtendReference caches a reference on this Dirent if
+// MountSourceOperations.Keep returns true.
+func (d *Dirent) maybeExtendReference() {
+	if msrc := d.Inode.MountSource; msrc.Keep(d) {
+		msrc.fscache.Add(d)
+	}
+}
+
+// dropExtendedReference drops any cached reference held by the
+// MountSource on the dirent.
+func (d *Dirent) dropExtendedReference() {
+	d.Inode.MountSource.fscache.Remove(d)
+}
+
+// lockForRename takes locks on oldParent and newParent as required by Rename
+// and returns a function that will unlock the locks taken. The returned
+// function must be called even if a non-nil error is returned.
+func lockForRename(oldParent *Dirent, oldName string, newParent *Dirent, newName string) (func(), error) {
+	if oldParent == newParent {
+		oldParent.mu.Lock()
+		return oldParent.mu.Unlock, nil
+	}
+
+	// Renaming between directories is a bit subtle:
+	//
+	// - A concurrent cross-directory Rename may try to lock in the opposite
+	// order; take renameMu to prevent this from happening.
+	//
+	// - If either directory is an ancestor of the other, then a concurrent
+	// Remove may lock the descendant (in DecRef -> closeAll) while holding a
+	// lock on the ancestor; to avoid this, ensure we take locks in the same
+	// ancestor-to-descendant order. (Holding renameMu prevents this
+	// relationship from changing.)
+	renameMu.Lock()
+
+	// First check if newParent is a descendant of oldParent.
+	child := newParent
+	for p := newParent.parent; p != nil; p = p.parent {
+		if p == oldParent {
+			oldParent.mu.Lock()
+			newParent.mu.Lock()
+			var err error
+			if child.name == oldName {
+				// newParent is not just a descendant of oldParent, but
+				// more specifically of oldParent/oldName. That is, we're
+				// trying to rename something into a subdirectory of
+				// itself.
+				err = syscall.EINVAL
+			}
+			return func() {
+				newParent.mu.Unlock()
+				oldParent.mu.Unlock()
+				renameMu.Unlock()
+			}, err
+		}
+		child = p
+	}
+
+	// Otherwise, either oldParent is a descendant of newParent or the two
+	// have no relationship; in either case we can do this:
+	newParent.mu.Lock()
+	oldParent.mu.Lock()
+	return func() {
+		oldParent.mu.Unlock()
+		newParent.mu.Unlock()
+		renameMu.Unlock()
+	}, nil
+}
+
+func checkSticky(ctx context.Context, dir *Dirent, victim *Dirent) error {
+	uattr, err := dir.Inode.UnstableAttr(ctx)
+	if err != nil {
+		return syserror.EPERM
+	}
+	if !uattr.Perms.Sticky {
+		return nil
+	}
+
+	creds := auth.CredentialsFromContext(ctx)
+	if uattr.Owner.UID == creds.EffectiveKUID {
+		return nil
+	}
+
+	vuattr, err := victim.Inode.UnstableAttr(ctx)
+	if err != nil {
+		return syserror.EPERM
+	}
+	if vuattr.Owner.UID == creds.EffectiveKUID {
+		return nil
+	}
+	if victim.Inode.CheckCapability(ctx, linux.CAP_FOWNER) {
+		return nil
+	}
+	return syserror.EPERM
+}
+
+// MayDelete determines whether `name`, a child of `dir`, can be deleted or
+// renamed by `ctx`.
+//
+// Compare Linux kernel fs/namei.c:may_delete.
+func MayDelete(ctx context.Context, root, dir *Dirent, name string) error {
+	victim, err := dir.Walk(ctx, root, name)
+	if err != nil {
+		return err
+	}
+	defer victim.DecRef()
+
+	return mayDelete(ctx, dir, victim)
+}
+
+func mayDelete(ctx context.Context, dir *Dirent, victim *Dirent) error {
+	if err := dir.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil {
+		return err
+	}
+
+	return checkSticky(ctx, dir, victim)
+}
+
+// Rename atomically converts the child of oldParent named oldName to a
+// child of newParent named newName.
+func Rename(ctx context.Context, root *Dirent, oldParent *Dirent, oldName string, newParent *Dirent, newName string) error {
+	if root == nil {
+		panic("Rename: root must not be nil")
+	}
+	if oldParent == newParent && oldName == newName {
+		return nil
+	}
+
+	// Acquire global renameMu lock, and mu locks on oldParent/newParent.
+	unlock, err := lockForRename(oldParent, oldName, newParent, newName)
+	defer unlock()
+	if err != nil {
+		return err
+	}
+
+	// Are we frozen?
+	// TODO: Is this the right errno?
+	if oldParent.frozen && !oldParent.Inode.IsVirtual() {
+		return syscall.ENOENT
+	}
+	if newParent.frozen && !newParent.Inode.IsVirtual() {
+		return syscall.ENOENT
+	}
+
+	// Check constraints on the object being renamed.
+	renamed, err := oldParent.walk(ctx, root, oldName, false /* may unlock */)
+	if err != nil {
+		return err
+	}
+	defer renamed.DecRef()
+
+	// Make sure we have write permissions on old and new parent.
+	if err := mayDelete(ctx, oldParent, renamed); err != nil {
+		return err
+	}
+	if newParent != oldParent {
+		if err := newParent.Inode.CheckPermission(ctx, PermMask{Write: true, Execute: true}); err != nil {
+			return err
+		}
+	}
+
+	// Source should not be an ancestor of the target.
+	if renamed == newParent {
+		return syscall.EINVAL
+	}
+
+	// Is the thing we're trying to rename busy?
+	if renamed.Busy() {
+		return syscall.EBUSY
+	}
+
+	// Per rename(2): "... EACCES: ... or oldpath is a directory and does not
+	// allow write permission (needed to update the .. entry)."
+	if IsDir(renamed.Inode.StableAttr) {
+		if err := renamed.Inode.CheckPermission(ctx, PermMask{Write: true}); err != nil {
+			return err
+		}
+	}
+
+	// Check constraints on the object being replaced, if any.
+	replaced, err := newParent.walk(ctx, root, newName, false /* may unlock */)
+	if err == nil {
+		defer replaced.DecRef()
+
+		// Target should not be an ancestor of source.
+		if replaced == oldParent {
+			// Why is this not EINVAL? See fs/namei.c.
+			return syscall.ENOTEMPTY
+		}
+
+		// Is the thing we're trying to replace busy?
+		if replaced.Busy() {
+			return syscall.EBUSY
+		}
+
+		// Require that a directory is replaced by a directory.
+		oldIsDir := IsDir(renamed.Inode.StableAttr)
+		newIsDir := IsDir(replaced.Inode.StableAttr)
+		if !newIsDir && oldIsDir {
+			return syscall.ENOTDIR
+		}
+		if !oldIsDir && newIsDir {
+			return syscall.EISDIR
+		}
+
+		// Allow the file system to drop extra references on replaced.
+		replaced.dropExtendedReference()
+
+		// NOTE: Keeping a dirent
+		// open across renames is currently broken for multiple
+		// reasons, so we flush all references on the replaced node and
+		// its children.
+		replaced.Inode.Watches.Unpin(replaced)
+		replaced.flush()
+	}
+
+	if err := renamed.Inode.Rename(ctx, oldParent, renamed, newParent, newName); err != nil {
+		return err
+	}
+
+	renamed.name = newName
+	renamed.parent = newParent
+	if oldParent != newParent {
+		// Reparent the reference held by renamed.parent. oldParent.DecRef
+		// can't destroy oldParent (and try to retake its lock) because
+		// Rename's caller must be holding a reference.
+		newParent.IncRef()
+		oldParent.DecRef()
+	}
+	if w, ok := newParent.children[newName]; ok {
+		w.Drop()
+		delete(newParent.children, newName)
+	}
+	if w, ok := oldParent.children[oldName]; ok {
+		w.Drop()
+		delete(oldParent.children, oldName)
+	}
+
+	// Add a weak reference from the new parent.  This ensures that the child
+	// can still be found from the new parent if a prior hard reference is
+	// held on renamed.
+	//
+	// This is required for file lock correctness because file locks are per-Dirent
+	// and without maintaining the a cached child (via a weak reference) for renamed,
+	// multiple Dirents can correspond to the same resource (by virtue of the renamed
+	// Dirent being unreachable by its parent and it being looked up).
+	newParent.children[newName] = refs.NewWeakRef(renamed, nil)
+
+	// Queue inotify events for the rename.
+	var ev uint32
+	if IsDir(renamed.Inode.StableAttr) {
+		ev |= linux.IN_ISDIR
+	}
+
+	cookie := uniqueid.InotifyCookie(ctx)
+	oldParent.Inode.Watches.Notify(oldName, ev|linux.IN_MOVED_FROM, cookie)
+	newParent.Inode.Watches.Notify(newName, ev|linux.IN_MOVED_TO, cookie)
+	// Somewhat surprisingly, self move events do not have a cookie.
+	renamed.Inode.Watches.Notify("", linux.IN_MOVE_SELF, 0)
+
+	// Allow the file system to drop extra references on renamed.
+	renamed.dropExtendedReference()
+
+	// Same as replaced.flush above.
+	renamed.flush()
+
+	return nil
+}
diff --git a/pkg/sentry/fs/dirent_cache.go b/pkg/sentry/fs/dirent_cache.go
new file mode 100644
index 000000000..e786e4f65
--- /dev/null
+++ b/pkg/sentry/fs/dirent_cache.go
@@ -0,0 +1,142 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sync"
+)
+
+// DirentCache is an LRU cache of Dirents. The Dirent's refCount is
+// incremented when it is added to the cache, and decremented when it is
+// removed.
+//
+// A nil DirentCache corresponds to a cache with size 0. All methods can be
+// called, but nothing is actually cached.
+type DirentCache struct {
+	// Maximum size of the cache. This must be saved manually, to handle the case
+	// when cache is nil.
+	maxSize uint64
+
+	// mu protects currentSize and direntList.
+	mu sync.Mutex `state:"nosave"`
+
+	// currentSize is the number of elements in the cache. It must be zero (i.e.
+	// the cache must be empty) on Save.
+	currentSize uint64 `state:"zerovalue"`
+
+	// list is a direntList, an ilist of Dirents. New Dirents are added
+	// to the front of the list. Old Dirents are removed from the back of
+	// the list. It must be zerovalue (i.e. the cache must be empty) on Save.
+	list direntList `state:"zerovalue"`
+}
+
+// NewDirentCache returns a new DirentCache with the given maxSize. If maxSize
+// is 0, nil is returned.
+func NewDirentCache(maxSize uint64) *DirentCache {
+	return &DirentCache{
+		maxSize: maxSize,
+	}
+}
+
+// Add adds the element to the cache and increments the refCount. If the
+// argument is already in the cache, it is moved to the front. An element is
+// removed from the back if the cache is over capacity.
+func (c *DirentCache) Add(d *Dirent) {
+	if c == nil || c.maxSize == 0 {
+		return
+	}
+
+	c.mu.Lock()
+	if c.contains(d) {
+		// d is already in cache. Bump it to the front.
+		// currentSize and refCount are unaffected.
+		c.list.Remove(d)
+		c.list.PushFront(d)
+		c.mu.Unlock()
+		return
+	}
+
+	// d is not in cache. Add it and take a reference.
+	c.list.PushFront(d)
+	d.IncRef()
+	c.currentSize++
+
+	// Remove the oldest until we are under the size limit.
+	for c.maxSize > 0 && c.currentSize > c.maxSize {
+		c.remove(c.list.Back())
+	}
+	c.mu.Unlock()
+}
+
+func (c *DirentCache) remove(d *Dirent) {
+	if !c.contains(d) {
+		panic(fmt.Sprintf("trying to remove %v, which is not in the dirent cache", d))
+	}
+	c.list.Remove(d)
+	d.SetPrev(nil)
+	d.SetNext(nil)
+	d.DecRef()
+	c.currentSize--
+}
+
+// Remove removes the element from the cache and decrements its refCount. It
+// also sets the previous and next elements to nil, which allows us to
+// determine if a given element is in the cache.
+func (c *DirentCache) Remove(d *Dirent) {
+	if c == nil || c.maxSize == 0 {
+		return
+	}
+	c.mu.Lock()
+	if !c.contains(d) {
+		c.mu.Unlock()
+		return
+	}
+	c.remove(d)
+	c.mu.Unlock()
+}
+
+// Size returns the number of elements in the cache.
+func (c *DirentCache) Size() uint64 {
+	if c == nil {
+		return 0
+	}
+	c.mu.Lock()
+	size := c.currentSize
+	c.mu.Unlock()
+	return size
+}
+
+func (c *DirentCache) contains(d *Dirent) bool {
+	// If d has a Prev or Next element, then it is in the cache.
+	if d.Prev() != nil || d.Next() != nil {
+		return true
+	}
+	// Otherwise, d is in the cache if it is the only element (and thus the
+	// first element).
+	return c.list.Front() == d
+}
+
+// Invalidate removes all Dirents from the cache, caling DecRef on each.
+func (c *DirentCache) Invalidate() {
+	if c == nil {
+		return
+	}
+	c.mu.Lock()
+	for c.list.Front() != nil {
+		c.remove(c.list.Front())
+	}
+	c.mu.Unlock()
+}
diff --git a/pkg/sentry/fs/dirent_cache_test.go b/pkg/sentry/fs/dirent_cache_test.go
new file mode 100644
index 000000000..82b7f6bd5
--- /dev/null
+++ b/pkg/sentry/fs/dirent_cache_test.go
@@ -0,0 +1,157 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"testing"
+)
+
+func TestDirentCache(t *testing.T) {
+	const maxSize = 5
+
+	c := NewDirentCache(maxSize)
+
+	// Size starts at 0.
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// Create a Dirent d.
+	d := NewNegativeDirent("")
+
+	// c does not contain d.
+	if got, want := c.contains(d), false; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// Add d to the cache.
+	c.Add(d)
+
+	// Size is now 1.
+	if got, want := c.Size(), uint64(1); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// c contains d.
+	if got, want := c.contains(d), true; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// Add maxSize-1 more elements.  d should be oldest element.
+	for i := 0; i < maxSize-1; i++ {
+		c.Add(NewNegativeDirent(""))
+	}
+
+	// Size is maxSize.
+	if got, want := c.Size(), uint64(maxSize); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// c contains d.
+	if got, want := c.contains(d), true; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// "Bump" d to the front by re-adding it.
+	c.Add(d)
+
+	// Size is maxSize.
+	if got, want := c.Size(), uint64(maxSize); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// c contains d.
+	if got, want := c.contains(d), true; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// Add maxSize-1 more elements.  d should again be oldest element.
+	for i := 0; i < maxSize-1; i++ {
+		c.Add(NewNegativeDirent(""))
+	}
+
+	// Size is maxSize.
+	if got, want := c.Size(), uint64(maxSize); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// c contains d.
+	if got, want := c.contains(d), true; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// Add one more element, which will bump d from the cache.
+	c.Add(NewNegativeDirent(""))
+
+	// Size is maxSize.
+	if got, want := c.Size(), uint64(maxSize); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// c does not contain d.
+	if got, want := c.contains(d), false; got != want {
+		t.Errorf("c.contains(d) got %v want %v", got, want)
+	}
+
+	// Invalidating causes size to be 0 and list to be empty.
+	c.Invalidate()
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+	if got, want := c.list.Empty(), true; got != want {
+		t.Errorf("c.list.Empty() got %v, want %v", got, want)
+	}
+
+	// Fill cache with maxSize dirents.
+	for i := 0; i < maxSize; i++ {
+		c.Add(NewNegativeDirent(""))
+	}
+}
+
+// TestNilDirentCache tests that a nil cache supports all cache operations, but
+// treats them as noop.
+func TestNilDirentCache(t *testing.T) {
+	// Create a nil cache.
+	var c *DirentCache
+
+	// Size is zero.
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// Call Add.
+	c.Add(NewNegativeDirent(""))
+
+	// Size is zero.
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// Call Remove.
+	c.Remove(NewNegativeDirent(""))
+
+	// Size is zero.
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+
+	// Call Invalidate.
+	c.Invalidate()
+
+	// Size is zero.
+	if got, want := c.Size(), uint64(0); got != want {
+		t.Errorf("c.Size() got %v, want %v", got, want)
+	}
+}
diff --git a/pkg/sentry/fs/dirent_refs_test.go b/pkg/sentry/fs/dirent_refs_test.go
new file mode 100644
index 000000000..8ce9ba02d
--- /dev/null
+++ b/pkg/sentry/fs/dirent_refs_test.go
@@ -0,0 +1,417 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+)
+
+func newMockDirInode(ctx context.Context, cache *DirentCache) *Inode {
+	return NewMockInode(ctx, NewMockMountSource(cache), StableAttr{Type: Directory})
+}
+
+func TestWalkPositive(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	root := NewDirent(newMockDirInode(ctx, nil), "root")
+
+	if got := root.TestReadRefs(); got != 0 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 0)
+	}
+
+	name := "d"
+	d, err := root.walk(ctx, root, name, false)
+	if err != nil {
+		t.Fatalf("root.walk(root, %q) got %v, want nil", name, err)
+	}
+
+	if got := root.TestReadRefs(); got != 1 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 1)
+	}
+
+	if got := d.TestReadRefs(); got != 0 {
+		t.Fatalf("child name = %q has a ref count of %d, want %d", d.name, got, 0)
+	}
+
+	d.DecRef()
+
+	if got := root.TestReadRefs(); got != 0 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 0)
+	}
+
+	if got := d.TestReadRefs(); got != -1 {
+		t.Fatalf("child name = %q has a ref count of %d, want %d", d.name, got, -1)
+	}
+
+	root.flush()
+
+	if got := len(root.children); got != 0 {
+		t.Fatalf("root has %d children, want %d", got, 0)
+	}
+}
+
+func TestWalkNegative(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	root := NewDirent(NewEmptyDir(ctx, nil), "root")
+	mn := root.Inode.InodeOperations.(*mockInodeOperationsLookupNegative)
+
+	if got := root.TestReadRefs(); got != 0 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 0)
+	}
+
+	name := "d"
+	for i := 0; i < 100; i++ {
+		_, err := root.walk(ctx, root, name, false)
+		if err != syscall.ENOENT {
+			t.Fatalf("root.walk(root, %q) got %v, want %v", name, err, syscall.ENOENT)
+		}
+	}
+
+	if got := root.TestReadRefs(); got != 0 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 1)
+	}
+
+	if got := len(root.children); got != 1 {
+		t.Fatalf("root has %d children, want %d", got, 1)
+	}
+
+	w, ok := root.children[name]
+	if !ok {
+		t.Fatalf("root wants child at %q", name)
+	}
+
+	child := w.Get()
+	if child == nil {
+		t.Fatalf("root wants to resolve weak reference")
+	}
+
+	if !child.(*Dirent).IsNegative() {
+		t.Fatalf("root found positive child at %q, want negative", name)
+	}
+
+	if got := child.(*Dirent).TestReadRefs(); got != 1 {
+		t.Fatalf("child has a ref count of %d, want %d", got, 1)
+	}
+
+	child.DecRef()
+
+	if got := child.(*Dirent).TestReadRefs(); got != 0 {
+		t.Fatalf("child has a ref count of %d, want %d", got, 0)
+	}
+
+	if got := len(root.children); got != 1 {
+		t.Fatalf("root has %d children, want %d", got, 1)
+	}
+
+	root.DecRef()
+
+	if got := root.TestReadRefs(); got != -1 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 0)
+	}
+
+	AsyncBarrier()
+
+	if got := mn.releaseCalled; got != true {
+		t.Fatalf("root.Close was called %v, want true", got)
+	}
+}
+
+type mockInodeOperationsLookupNegative struct {
+	*MockInodeOperations
+	releaseCalled bool
+}
+
+func NewEmptyDir(ctx context.Context, cache *DirentCache) *Inode {
+	m := NewMockMountSource(cache)
+	return NewInode(&mockInodeOperationsLookupNegative{
+		MockInodeOperations: NewMockInodeOperations(ctx),
+	}, m, StableAttr{Type: Directory})
+}
+
+func (m *mockInodeOperationsLookupNegative) Lookup(ctx context.Context, dir *Inode, p string) (*Dirent, error) {
+	return NewNegativeDirent(p), nil
+}
+
+func (m *mockInodeOperationsLookupNegative) Release(context.Context) {
+	m.releaseCalled = true
+}
+
+func TestHashNegativeToPositive(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	root := NewDirent(NewEmptyDir(ctx, nil), "root")
+
+	name := "d"
+	_, err := root.walk(ctx, root, name, false)
+	if err != syscall.ENOENT {
+		t.Fatalf("root.walk(root, %q) got %v, want %v", name, err, syscall.ENOENT)
+	}
+
+	if got := root.exists(ctx, root, name); got != false {
+		t.Fatalf("got %q exists, want does not exist", name)
+	}
+
+	f, err := root.Create(ctx, root, name, FileFlags{}, FilePermissions{})
+	if err != nil {
+		t.Fatalf("root.Create(%q, _), got error %v, want nil", name, err)
+	}
+	d := f.Dirent
+
+	if d.IsNegative() {
+		t.Fatalf("got negative Dirent, want positive")
+	}
+
+	if got := d.TestReadRefs(); got != 0 {
+		t.Fatalf("child %q has a ref count of %d, want %d", name, got, 0)
+	}
+
+	if got := root.TestReadRefs(); got != 1 {
+		t.Fatalf("root has a ref count of %d, want %d", got, 1)
+	}
+
+	if got := len(root.children); got != 1 {
+		t.Fatalf("got %d children, want %d", got, 1)
+	}
+
+	w, ok := root.children[name]
+	if !ok {
+		t.Fatalf("failed to find weak reference to %q", name)
+	}
+
+	child := w.Get()
+	if child == nil {
+		t.Fatalf("want to resolve weak reference")
+	}
+
+	if child.(*Dirent) != d {
+		t.Fatalf("got foreign child")
+	}
+}
+
+func TestRevalidate(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// Whether to make negative Dirents.
+		makeNegative bool
+	}{
+		{
+			desc:         "Revalidate negative Dirent",
+			makeNegative: true,
+		},
+		{
+			desc:         "Revalidate positive Dirent",
+			makeNegative: false,
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			root := NewDirent(NewMockInodeRevalidate(ctx, test.makeNegative), "root")
+
+			name := "d"
+			d1, err := root.walk(ctx, root, name, false)
+			if !test.makeNegative && err != nil {
+				t.Fatalf("root.walk(root, %q) got %v, want nil", name, err)
+			}
+			d2, err := root.walk(ctx, root, name, false)
+			if !test.makeNegative && err != nil {
+				t.Fatalf("root.walk(root, %q) got %v, want nil", name, err)
+			}
+			if !test.makeNegative && d1 == d2 {
+				t.Fatalf("revalidating walk got same *Dirent, want different")
+			}
+			if got := len(root.children); got != 1 {
+				t.Errorf("revalidating walk got %d children, want %d", got, 1)
+			}
+		})
+	}
+}
+
+type MockInodeOperationsRevalidate struct {
+	*MockInodeOperations
+	makeNegative bool
+}
+
+func NewMockInodeRevalidate(ctx context.Context, makeNegative bool) *Inode {
+	mn := NewMockInodeOperations(ctx)
+	m := NewMockMountSource(nil)
+	m.MountSourceOperations.(*MockMountSourceOps).revalidate = true
+	return NewInode(&MockInodeOperationsRevalidate{MockInodeOperations: mn, makeNegative: makeNegative}, m, StableAttr{Type: Directory})
+}
+
+func (m *MockInodeOperationsRevalidate) Lookup(ctx context.Context, dir *Inode, p string) (*Dirent, error) {
+	if !m.makeNegative {
+		return m.MockInodeOperations.Lookup(ctx, dir, p)
+	}
+	return NewNegativeDirent(p), nil
+}
+
+func TestCreateExtraRefs(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// root is the Dirent to create from.
+		root *Dirent
+
+		// expected references on walked Dirent.
+		refs int64
+	}{
+		{
+			desc: "Create caching",
+			root: NewDirent(NewEmptyDir(ctx, NewDirentCache(1)), "root"),
+			refs: 1,
+		},
+		{
+			desc: "Create not caching",
+			root: NewDirent(NewEmptyDir(ctx, nil), "root"),
+			refs: 0,
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			name := "d"
+			f, err := test.root.Create(ctx, test.root, name, FileFlags{}, FilePermissions{})
+			if err != nil {
+				t.Fatalf("root.Create(root, %q) failed: %v", name, err)
+			}
+			d := f.Dirent
+
+			if got := d.TestReadRefs(); got != test.refs {
+				t.Errorf("dirent has a ref count of %d, want %d", got, test.refs)
+			}
+		})
+	}
+}
+
+func TestRemoveExtraRefs(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// root is the Dirent to make and remove from.
+		root *Dirent
+	}{
+		{
+			desc: "Remove caching",
+			root: NewDirent(NewEmptyDir(ctx, NewDirentCache(1)), "root"),
+		},
+		{
+			desc: "Remove not caching",
+			root: NewDirent(NewEmptyDir(ctx, nil), "root"),
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			name := "d"
+			f, err := test.root.Create(ctx, test.root, name, FileFlags{}, FilePermissions{})
+			if err != nil {
+				t.Fatalf("root.Create(%q, _) failed: %v", name, err)
+			}
+			d := f.Dirent
+
+			if err := test.root.Remove(contexttest.Context(t), test.root, name); err != nil {
+				t.Fatalf("root.Remove(root, %q) failed: %v", name, err)
+			}
+
+			if got := d.TestReadRefs(); got != 0 {
+				t.Fatalf("dirent has a ref count of %d, want %d", got, 0)
+			}
+
+			d.DecRef()
+
+			test.root.flush()
+
+			if got := len(test.root.children); got != 0 {
+				t.Errorf("root has %d children, want %d", got, 0)
+			}
+		})
+	}
+}
+
+func TestRenameExtraRefs(t *testing.T) {
+	// refs == 0 -> one reference.
+	// refs == -1 -> has been destroyed.
+
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// cache of extra Dirent references, may be nil.
+		cache *DirentCache
+	}{
+		{
+			desc:  "Rename no caching",
+			cache: nil,
+		},
+		{
+			desc:  "Rename caching",
+			cache: NewDirentCache(5),
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			dirAttr := StableAttr{Type: Directory}
+
+			oldParent := NewDirent(NewMockInode(ctx, NewMockMountSource(test.cache), dirAttr), "old_parent")
+			newParent := NewDirent(NewMockInode(ctx, NewMockMountSource(test.cache), dirAttr), "new_parent")
+
+			renamed, err := oldParent.Walk(ctx, oldParent, "old_child")
+			if err != nil {
+				t.Fatalf("Walk(oldParent, %q) got error %v, want nil", "old_child", err)
+			}
+			replaced, err := newParent.Walk(ctx, oldParent, "new_child")
+			if err != nil {
+				t.Fatalf("Walk(newParent, %q) got error %v, want nil", "new_child", err)
+			}
+
+			if err := Rename(contexttest.RootContext(t), oldParent /*root */, oldParent, "old_child", newParent, "new_child"); err != nil {
+				t.Fatalf("Rename got error %v, want nil", err)
+			}
+
+			oldParent.flush()
+			newParent.flush()
+
+			// Expect to have only active references.
+			if got := renamed.TestReadRefs(); got != 0 {
+				t.Errorf("renamed has ref count %d, want only active references %d", got, 0)
+			}
+			if got := replaced.TestReadRefs(); got != 0 {
+				t.Errorf("replaced has ref count %d, want only active references %d", got, 0)
+			}
+		})
+	}
+}
diff --git a/pkg/sentry/fs/dirent_state.go b/pkg/sentry/fs/dirent_state.go
new file mode 100644
index 000000000..c6a1b5e38
--- /dev/null
+++ b/pkg/sentry/fs/dirent_state.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sync/atomic"
+)
+
+// beforeSave is invoked by stateify.
+func (d *Dirent) beforeSave() {
+	// Refuse to save if the file has already been deleted (but still has
+	// open fds, which is why the Dirent is still accessible). We know the
+	// the restore opening of the file will always fail. This condition will
+	// last until all the open fds and this Dirent are closed and released.
+	//
+	// Note that this is rejection rather than failure---it would be
+	// perfectly OK to save---we are simply disallowing it here to prevent
+	// generating non-restorable state dumps. As the program continues its
+	// execution, it may become allowed to save again.
+	if atomic.LoadInt32(&d.deleted) != 0 {
+		n, _ := d.FullName(nil /* root */)
+		panic(ErrSaveRejection{fmt.Errorf("deleted file %q still has open fds", n)})
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (d *Dirent) afterLoad() {
+	if d.userVisible {
+		allDirents.add(d)
+	}
+}
diff --git a/pkg/sentry/fs/fdpipe/BUILD b/pkg/sentry/fs/fdpipe/BUILD
new file mode 100644
index 000000000..9e1f65d3e
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/BUILD
@@ -0,0 +1,76 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "pipe_state",
+    srcs = [
+        "pipe.go",
+        "pipe_state.go",
+    ],
+    out = "pipe_autogen_state.go",
+    imports = ["gvisor.googlesource.com/gvisor/pkg/sentry/fs"],
+    package = "fdpipe",
+)
+
+go_library(
+    name = "fdpipe",
+    srcs = [
+        "pipe.go",
+        "pipe_autogen_state.go",
+        "pipe_opener.go",
+        "pipe_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/fdpipe",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/fd",
+        "//pkg/log",
+        "//pkg/metric",
+        "//pkg/p9",
+        "//pkg/refs",
+        "//pkg/secio",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/unet",
+        "//pkg/waiter",
+        "//pkg/waiter/fdnotifier",
+    ],
+)
+
+go_test(
+    name = "fdpipe_test",
+    size = "small",
+    srcs = [
+        "pipe_opener_test.go",
+        "pipe_test.go",
+    ],
+    embed = [":fdpipe"],
+    deps = [
+        "//pkg/fd",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/usermem",
+        "//pkg/syserror",
+        "//pkg/waiter/fdnotifier",
+        "@com_github_google_uuid//:go_default_library",
+    ],
+)
diff --git a/pkg/sentry/fs/fdpipe/pipe.go b/pkg/sentry/fs/fdpipe/pipe.go
new file mode 100644
index 000000000..f7bbd4aff
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/pipe.go
@@ -0,0 +1,167 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package fdpipe implements common namedpipe opening and accessing logic.
+package fdpipe
+
+import (
+	"os"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/secio"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+// pipeOperations are the fs.FileOperations of a host pipe.
+type pipeOperations struct {
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	fsutil.NoIoctl       `state:"nosave"`
+	waiter.Queue         `state:"nosave"`
+
+	// flags are the flags used to open the pipe.
+	flags fs.FileFlags `state:".(fs.FileFlags)"`
+
+	// opener is how the pipe was opened.
+	opener NonBlockingOpener `state:"wait"`
+
+	// file represents the host pipe.
+	file *fd.FD `state:"nosave"`
+
+	// mu protects readAheadBuffer access below.
+	mu sync.Mutex `state:"nosave"`
+
+	// readAheadBuffer contains read bytes that have not yet been read
+	// by the application but need to be buffered for save-restore for correct
+	// opening semantics.  The readAheadBuffer will only be non-empty when the
+	// is first opened and will be drained by subsequent reads on the pipe.
+	readAheadBuffer []byte
+}
+
+// newPipeOperations returns an implementation of fs.FileOperations for a pipe.
+func newPipeOperations(ctx context.Context, opener NonBlockingOpener, flags fs.FileFlags, file *fd.FD, readAheadBuffer []byte) (*pipeOperations, error) {
+	pipeOps := &pipeOperations{
+		flags:           flags,
+		opener:          opener,
+		file:            file,
+		readAheadBuffer: readAheadBuffer,
+	}
+	if err := pipeOps.init(); err != nil {
+		return nil, err
+	}
+	return pipeOps, nil
+}
+
+// init initializes p.file.
+func (p *pipeOperations) init() error {
+	var s syscall.Stat_t
+	if err := syscall.Fstat(p.file.FD(), &s); err != nil {
+		log.Warningf("pipe: cannot stat fd %d: %v", p.file.FD(), err)
+		return syscall.EINVAL
+	}
+	if s.Mode&syscall.S_IFIFO != syscall.S_IFIFO {
+		log.Warningf("pipe: cannot load fd %d as pipe, file type: %o", p.file.FD(), s.Mode)
+		return syscall.EINVAL
+	}
+	if err := syscall.SetNonblock(p.file.FD(), true); err != nil {
+		return err
+	}
+	if err := fdnotifier.AddFD(int32(p.file.FD()), &p.Queue); err != nil {
+		return err
+	}
+	return nil
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (p *pipeOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	p.Queue.EventRegister(e, mask)
+	fdnotifier.UpdateFD(int32(p.file.FD()))
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (p *pipeOperations) EventUnregister(e *waiter.Entry) {
+	p.Queue.EventUnregister(e)
+	fdnotifier.UpdateFD(int32(p.file.FD()))
+}
+
+// Readiness returns a mask of ready events for stream.
+func (p *pipeOperations) Readiness(mask waiter.EventMask) (eventMask waiter.EventMask) {
+	return fdnotifier.NonBlockingPoll(int32(p.file.FD()), mask)
+}
+
+// Release implements fs.FileOperations.Release.
+func (p *pipeOperations) Release() {
+	fdnotifier.RemoveFD(int32(p.file.FD()))
+	p.file.Close()
+	p.file = nil
+}
+
+// Read implements fs.FileOperations.Read.
+func (p *pipeOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	// Drain the read ahead buffer, if it contains anything first.
+	var bufN int
+	var bufErr error
+	p.mu.Lock()
+	if len(p.readAheadBuffer) > 0 {
+		bufN, bufErr = dst.CopyOut(ctx, p.readAheadBuffer)
+		p.readAheadBuffer = p.readAheadBuffer[bufN:]
+		dst = dst.DropFirst(bufN)
+	}
+	p.mu.Unlock()
+	if dst.NumBytes() == 0 || bufErr != nil {
+		return int64(bufN), bufErr
+	}
+
+	// Pipes expect full reads.
+	n, err := dst.CopyOutFrom(ctx, safemem.FromIOReader{secio.FullReader{p.file}})
+	total := int64(bufN) + n
+	if err != nil && isBlockError(err) {
+		return total, syserror.ErrWouldBlock
+	}
+	return total, err
+}
+
+// Write implements fs.FileOperations.Write.
+func (p *pipeOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	n, err := src.CopyInTo(ctx, safemem.FromIOWriter{p.file})
+	if err != nil && isBlockError(err) {
+		return n, syserror.ErrWouldBlock
+	}
+	return n, err
+}
+
+// isBlockError unwraps os errors and checks if they are caused by EAGAIN or
+// EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock.
+func isBlockError(err error) bool {
+	if err == syserror.EAGAIN || err == syserror.EWOULDBLOCK {
+		return true
+	}
+	if pe, ok := err.(*os.PathError); ok {
+		return isBlockError(pe.Err)
+	}
+	return false
+}
diff --git a/pkg/sentry/fs/fdpipe/pipe_opener.go b/pkg/sentry/fs/fdpipe/pipe_opener.go
new file mode 100644
index 000000000..a0d59575f
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/pipe_opener.go
@@ -0,0 +1,193 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fdpipe
+
+import (
+	"io"
+	"os"
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// NonBlockingOpener is a generic host file opener used to retry opening host
+// pipes if necessary.
+type NonBlockingOpener interface {
+	// NonBlockingOpen tries to open a host pipe in a non-blocking way,
+	// and otherwise returns an error. Implementations should be idempotent.
+	NonBlockingOpen(context.Context, fs.PermMask) (*fd.FD, error)
+}
+
+// Open blocks until a host pipe can be opened or the action was cancelled.
+// On success, returns fs.FileOperations wrapping the opened host pipe.
+func Open(ctx context.Context, opener NonBlockingOpener, flags fs.FileFlags) (fs.FileOperations, error) {
+	p := &pipeOpenState{}
+	canceled := false
+	for {
+		if file, err := p.TryOpen(ctx, opener, flags); err != syserror.ErrWouldBlock {
+			return file, err
+		}
+
+		// Honor the cancellation request if open still blocks.
+		if canceled {
+			// If we were canceled but we have a handle to a host
+			// file, we need to close it.
+			if p.hostFile != nil {
+				p.hostFile.Close()
+			}
+			return nil, syserror.ErrInterrupted
+		}
+
+		cancel := ctx.SleepStart()
+		select {
+		case <-cancel:
+			// The cancellation request received here really says
+			// "cancel from now on (or ASAP)". Any environmental
+			// changes happened before receiving it, that might have
+			// caused open to not block anymore, should still be
+			// respected. So we cannot just return here. We have to
+			// give open another try below first.
+			canceled = true
+			ctx.SleepFinish(false)
+		case <-time.After(100 * time.Millisecond):
+			// If we would block, then delay retrying for a bit, since there
+			// is no way to know when the pipe would be ready to be
+			// re-opened. This is identical to sending an event notification
+			// to stop blocking in Task.Block, given that this routine will
+			// stop retrying if a cancelation is received.
+			ctx.SleepFinish(true)
+		}
+	}
+}
+
+// pipeOpenState holds state needed to open a blocking named pipe read only, for instance the
+// file that has been opened but doesn't yet have a corresponding writer.
+type pipeOpenState struct {
+	// hostFile is the read only named pipe which lacks a corresponding writer.
+	hostFile *fd.FD
+}
+
+// unwrapError is needed to match against ENXIO primarily.
+func unwrapError(err error) error {
+	if pe, ok := err.(*os.PathError); ok {
+		return pe.Err
+	}
+	return err
+}
+
+// TryOpen uses a NonBlockingOpener to try to open a host pipe, respecting the fs.FileFlags.
+func (p *pipeOpenState) TryOpen(ctx context.Context, opener NonBlockingOpener, flags fs.FileFlags) (*pipeOperations, error) {
+	switch {
+	// Reject invalid configurations so they don't accidently succeed below.
+	case !flags.Read && !flags.Write:
+		return nil, syscall.EINVAL
+
+	// Handle opening RDWR or with O_NONBLOCK: will never block, so try only once.
+	case (flags.Read && flags.Write) || flags.NonBlocking:
+		f, err := opener.NonBlockingOpen(ctx, fs.PermMask{Read: flags.Read, Write: flags.Write})
+		if err != nil {
+			return nil, err
+		}
+		return newPipeOperations(ctx, opener, flags, f, nil)
+
+	// Handle opening O_WRONLY blocking: convert ENXIO to syserror.ErrWouldBlock.
+	// See TryOpenWriteOnly for more details.
+	case flags.Write:
+		return p.TryOpenWriteOnly(ctx, opener)
+
+	default:
+		// Handle opening O_RDONLY blocking: convert EOF from read to syserror.ErrWouldBlock.
+		// See TryOpenReadOnly for more details.
+		return p.TryOpenReadOnly(ctx, opener)
+	}
+}
+
+// TryOpenReadOnly tries to open a host pipe read only but only returns a fs.File when
+// there is a coordinating writer.  Call TryOpenReadOnly repeatedly on the same pipeOpenState
+// until syserror.ErrWouldBlock is no longer returned.
+//
+// How it works:
+//
+// Opening a pipe read only will return no error, but each non zero Read will return EOF
+// until a writer becomes available, then EWOULDBLOCK.  This is the only state change
+// available to us.  We keep a read ahead buffer in case we read bytes instead of getting
+// EWOULDBLOCK, to be read from on the first read request to this fs.File.
+func (p *pipeOpenState) TryOpenReadOnly(ctx context.Context, opener NonBlockingOpener) (*pipeOperations, error) {
+	// Waiting for a blocking read only open involves reading from the host pipe until
+	// bytes or other writers are available, so instead of retrying opening the pipe,
+	// it's necessary to retry reading from the pipe. To do this we need to keep around
+	// the read only pipe we opened, until success or an irrecoverable read error (at
+	// which point it must be closed).
+	if p.hostFile == nil {
+		var err error
+		p.hostFile, err = opener.NonBlockingOpen(ctx, fs.PermMask{Read: true})
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	// Try to read from the pipe to see if writers are around.
+	tryReadBuffer := make([]byte, 1)
+	n, rerr := p.hostFile.Read(tryReadBuffer)
+
+	// No bytes were read.
+	if n == 0 {
+		// EOF means that we're not ready yet.
+		if rerr == nil || rerr == io.EOF {
+			return nil, syserror.ErrWouldBlock
+		}
+		// Any error that is not EWOULDBLOCK also means we're not
+		// ready yet, and probably never will be ready.  In this
+		// case we need to close the host pipe we opened.
+		if unwrapError(rerr) != syscall.EWOULDBLOCK {
+			p.hostFile.Close()
+			return nil, rerr
+		}
+	}
+
+	// If any bytes were read, no matter the corresponding error, we need
+	// to keep them around so they can be read by the application.
+	var readAheadBuffer []byte
+	if n > 0 {
+		readAheadBuffer = tryReadBuffer
+	}
+
+	// Successfully opened read only blocking pipe with either bytes available
+	// to read and/or a writer available.
+	return newPipeOperations(ctx, opener, fs.FileFlags{Read: true}, p.hostFile, readAheadBuffer)
+}
+
+// TryOpenWriteOnly tries to open a host pipe write only but only returns a fs.File when
+// there is a coordinating reader.  Call TryOpenWriteOnly repeatedly on the same pipeOpenState
+// until syserror.ErrWouldBlock is no longer returned.
+//
+// How it works:
+//
+// Opening a pipe write only will return ENXIO until readers are available.  Converts the ENXIO
+// to an syserror.ErrWouldBlock, to tell callers to retry.
+func (*pipeOpenState) TryOpenWriteOnly(ctx context.Context, opener NonBlockingOpener) (*pipeOperations, error) {
+	hostFile, err := opener.NonBlockingOpen(ctx, fs.PermMask{Write: true})
+	if unwrapError(err) == syscall.ENXIO {
+		return nil, syserror.ErrWouldBlock
+	}
+	if err != nil {
+		return nil, err
+	}
+	return newPipeOperations(ctx, opener, fs.FileFlags{Write: true}, hostFile, nil)
+}
diff --git a/pkg/sentry/fs/fdpipe/pipe_opener_test.go b/pkg/sentry/fs/fdpipe/pipe_opener_test.go
new file mode 100644
index 000000000..83f6c1986
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/pipe_opener_test.go
@@ -0,0 +1,522 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fdpipe
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"os"
+	"path"
+	"syscall"
+	"testing"
+	"time"
+
+	"github.com/google/uuid"
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+type hostOpener struct {
+	name string
+}
+
+func (h *hostOpener) NonBlockingOpen(_ context.Context, p fs.PermMask) (*fd.FD, error) {
+	var flags int
+	switch {
+	case p.Read && p.Write:
+		flags = syscall.O_RDWR
+	case p.Write:
+		flags = syscall.O_WRONLY
+	case p.Read:
+		flags = syscall.O_RDONLY
+	default:
+		return nil, syscall.EINVAL
+	}
+	f, err := syscall.Open(h.name, flags|syscall.O_NONBLOCK, 0666)
+	if err != nil {
+		return nil, err
+	}
+	return fd.New(f), nil
+}
+
+func pipename() string {
+	return fmt.Sprintf(path.Join(os.TempDir(), "test-named-pipe-%s"), uuid.New())
+}
+
+func mkpipe(name string) error {
+	return syscall.Mknod(name, syscall.S_IFIFO|0666, 0)
+}
+
+func TestTryOpen(t *testing.T) {
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// makePipe is true if the test case should create the pipe.
+		makePipe bool
+
+		// flags are the fs.FileFlags used to open the pipe.
+		flags fs.FileFlags
+
+		// expectFile is true if a fs.File is expected.
+		expectFile bool
+
+		// err is the expected error
+		err error
+	}{
+		{
+			desc:       "FileFlags lacking Read and Write are invalid",
+			makePipe:   false,
+			flags:      fs.FileFlags{}, /* bogus */
+			expectFile: false,
+			err:        syscall.EINVAL,
+		},
+		{
+			desc:       "NonBlocking Read only error returns immediately",
+			makePipe:   false, /* causes the error */
+			flags:      fs.FileFlags{Read: true, NonBlocking: true},
+			expectFile: false,
+			err:        syscall.ENOENT,
+		},
+		{
+			desc:       "NonBlocking Read only success returns immediately",
+			makePipe:   true,
+			flags:      fs.FileFlags{Read: true, NonBlocking: true},
+			expectFile: true,
+			err:        nil,
+		},
+		{
+			desc:       "NonBlocking Write only error returns immediately",
+			makePipe:   false, /* causes the error */
+			flags:      fs.FileFlags{Write: true, NonBlocking: true},
+			expectFile: false,
+			err:        syscall.ENOENT,
+		},
+		{
+			desc:       "NonBlocking Write only no reader error returns immediately",
+			makePipe:   true,
+			flags:      fs.FileFlags{Write: true, NonBlocking: true},
+			expectFile: false,
+			err:        syscall.ENXIO,
+		},
+		{
+			desc:       "ReadWrite error returns immediately",
+			makePipe:   false, /* causes the error */
+			flags:      fs.FileFlags{Read: true, Write: true},
+			expectFile: false,
+			err:        syscall.ENOENT,
+		},
+		{
+			desc:       "ReadWrite returns immediately",
+			makePipe:   true,
+			flags:      fs.FileFlags{Read: true, Write: true},
+			expectFile: true,
+			err:        nil,
+		},
+		{
+			desc:       "Blocking Write only returns open error",
+			makePipe:   false, /* causes the error */
+			flags:      fs.FileFlags{Write: true},
+			expectFile: false,
+			err:        syscall.ENOENT, /* from bogus perms */
+		},
+		{
+			desc:       "Blocking Read only returns open error",
+			makePipe:   false, /* causes the error */
+			flags:      fs.FileFlags{Read: true},
+			expectFile: false,
+			err:        syscall.ENOENT,
+		},
+		{
+			desc:       "Blocking Write only returns with syserror.ErrWouldBlock",
+			makePipe:   true,
+			flags:      fs.FileFlags{Write: true},
+			expectFile: false,
+			err:        syserror.ErrWouldBlock,
+		},
+		{
+			desc:       "Blocking Read only returns with syserror.ErrWouldBlock",
+			makePipe:   true,
+			flags:      fs.FileFlags{Read: true},
+			expectFile: false,
+			err:        syserror.ErrWouldBlock,
+		},
+	} {
+		name := pipename()
+		if test.makePipe {
+			// Create the pipe.  We do this per-test case to keep tests independent.
+			if err := mkpipe(name); err != nil {
+				t.Errorf("%s: failed to make host pipe: %v", test.desc, err)
+				continue
+			}
+			defer syscall.Unlink(name)
+		}
+
+		// Use a host opener to keep things simple.
+		opener := &hostOpener{name: name}
+
+		pipeOpenState := &pipeOpenState{}
+		ctx := contexttest.Context(t)
+		pipeOps, err := pipeOpenState.TryOpen(ctx, opener, test.flags)
+		if unwrapError(err) != test.err {
+			t.Errorf("%s: got error %v, want %v", test.desc, err, test.err)
+			if pipeOps != nil {
+				// Cleanup the state of the pipe, and remove the fd from the
+				// fdnotifier.  Sadly this needed to maintain the correctness
+				// of other tests because the fdnotifier is global.
+				pipeOps.Release()
+			}
+			continue
+		}
+		if (pipeOps != nil) != test.expectFile {
+			t.Errorf("%s: got non-nil file %v, want %v", test.desc, pipeOps != nil, test.expectFile)
+		}
+		if pipeOps != nil {
+			// Same as above.
+			pipeOps.Release()
+		}
+	}
+}
+
+func TestPipeOpenUnblocksEventually(t *testing.T) {
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// partnerIsReader is true if the goroutine opening the same pipe as the test case
+		// should open the pipe read only.  Otherwise write only.  This also means that the
+		// test case will open the pipe in the opposite way.
+		partnerIsReader bool
+
+		// partnerIsBlocking is true if the goroutine opening the same pipe as the test case
+		// should do so without the O_NONBLOCK flag, otherwise opens the pipe with O_NONBLOCK
+		// until ENXIO is not returned.
+		partnerIsBlocking bool
+	}{
+		{
+			desc:              "Blocking Read with blocking writer partner opens eventually",
+			partnerIsReader:   false,
+			partnerIsBlocking: true,
+		},
+		{
+			desc:              "Blocking Write with blocking reader partner opens eventually",
+			partnerIsReader:   true,
+			partnerIsBlocking: true,
+		},
+		{
+			desc:              "Blocking Read with non-blocking writer partner opens eventually",
+			partnerIsReader:   false,
+			partnerIsBlocking: false,
+		},
+		{
+			desc:              "Blocking Write with non-blocking reader partner opens eventually",
+			partnerIsReader:   true,
+			partnerIsBlocking: false,
+		},
+	} {
+		// Create the pipe.  We do this per-test case to keep tests independent.
+		name := pipename()
+		if err := mkpipe(name); err != nil {
+			t.Errorf("%s: failed to make host pipe: %v", test.desc, err)
+			continue
+		}
+		defer syscall.Unlink(name)
+
+		// Spawn the partner.
+		type fderr struct {
+			fd  int
+			err error
+		}
+		errch := make(chan fderr, 1)
+		go func() {
+			var flags int
+			if test.partnerIsReader {
+				flags = syscall.O_RDONLY
+			} else {
+				flags = syscall.O_WRONLY
+			}
+			if test.partnerIsBlocking {
+				fd, err := syscall.Open(name, flags, 0666)
+				errch <- fderr{fd: fd, err: err}
+			} else {
+				var fd int
+				err := error(syscall.ENXIO)
+				for err == syscall.ENXIO {
+					fd, err = syscall.Open(name, flags|syscall.O_NONBLOCK, 0666)
+					time.Sleep(1 * time.Second)
+				}
+				errch <- fderr{fd: fd, err: err}
+			}
+		}()
+
+		// Setup file flags for either a read only or write only open.
+		flags := fs.FileFlags{
+			Read:  !test.partnerIsReader,
+			Write: test.partnerIsReader,
+		}
+
+		// Open the pipe in a blocking way, which should succeed eventually.
+		opener := &hostOpener{name: name}
+		ctx := contexttest.Context(t)
+		pipeOps, err := Open(ctx, opener, flags)
+		if pipeOps != nil {
+			// Same as TestTryOpen.
+			pipeOps.Release()
+		}
+
+		// Check that the partner opened the file successfully.
+		e := <-errch
+		if e.err != nil {
+			t.Errorf("%s: partner got error %v, wanted nil", test.desc, e.err)
+			continue
+		}
+		// If so, then close the partner fd to avoid leaking an fd.
+		syscall.Close(e.fd)
+
+		// Check that our blocking open was successful.
+		if err != nil {
+			t.Errorf("%s: blocking open got error %v, wanted nil", test.desc, err)
+			continue
+		}
+		if pipeOps == nil {
+			t.Errorf("%s: blocking open got nil file, wanted non-nil", test.desc)
+			continue
+		}
+	}
+}
+
+func TestCopiedReadAheadBuffer(t *testing.T) {
+	// Create the pipe.
+	name := pipename()
+	if err := mkpipe(name); err != nil {
+		t.Fatalf("failed to make host pipe: %v", err)
+	}
+	defer syscall.Unlink(name)
+
+	// We're taking advantage of the fact that pipes opened read only always return
+	// success, but internally they are not deemed "opened" until we're sure that
+	// another writer comes along.  This means we can open the same pipe write only
+	// with no problems + write to it, given that opener.Open already tried to open
+	// the pipe RDONLY and succeeded, which we know happened if TryOpen returns
+	// syserror.ErrwouldBlock.
+	//
+	// This simulates the open(RDONLY) <-> open(WRONLY)+write race we care about, but
+	// does not cause our test to be racy (which would be terrible).
+	opener := &hostOpener{name: name}
+	pipeOpenState := &pipeOpenState{}
+	ctx := contexttest.Context(t)
+	pipeOps, err := pipeOpenState.TryOpen(ctx, opener, fs.FileFlags{Read: true})
+	if pipeOps != nil {
+		pipeOps.Release()
+		t.Fatalf("open(%s, %o) got file, want nil", name, syscall.O_RDONLY)
+	}
+	if err != syserror.ErrWouldBlock {
+		t.Fatalf("open(%s, %o) got error %v, want %v", name, syscall.O_RDONLY, err, syserror.ErrWouldBlock)
+	}
+
+	// Then open the same pipe write only and write some bytes to it.  The next
+	// time we try to open the pipe read only again via the pipeOpenState, we should
+	// succeed and buffer some of the bytes written.
+	fd, err := syscall.Open(name, syscall.O_WRONLY, 0666)
+	if err != nil {
+		t.Fatalf("open(%s, %o) got error %v, want nil", name, syscall.O_WRONLY, err)
+	}
+	defer syscall.Close(fd)
+
+	data := []byte("hello")
+	if n, err := syscall.Write(fd, data); n != len(data) || err != nil {
+		t.Fatalf("write(%v) got (%d, %v), want (%d, nil)", data, n, err, len(data))
+	}
+
+	// Try the read again, knowing that it should succeed this time.
+	pipeOps, err = pipeOpenState.TryOpen(ctx, opener, fs.FileFlags{Read: true})
+	if pipeOps == nil {
+		t.Fatalf("open(%s, %o) got nil file, want not nil", name, syscall.O_RDONLY)
+	}
+	defer pipeOps.Release()
+
+	if err != nil {
+		t.Fatalf("open(%s, %o) got error %v, want nil", name, syscall.O_RDONLY, err)
+	}
+
+	inode := fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{
+		Type: fs.Pipe,
+	})
+	file := fs.NewFile(ctx, fs.NewDirent(inode, "pipe"), fs.FileFlags{Read: true}, pipeOps)
+
+	// Check that the file we opened points to a pipe with a non-empty read ahead buffer.
+	bufsize := len(pipeOps.readAheadBuffer)
+	if bufsize != 1 {
+		t.Fatalf("read ahead buffer got %d bytes, want %d", bufsize, 1)
+	}
+
+	// Now for the final test, try to read everything in, expecting to get back all of
+	// the bytes that were written at once.  Note that in the wild there is no atomic
+	// read size so expecting to get all bytes from a single writer when there are
+	// multiple readers is a bad expectation.
+	buf := make([]byte, len(data))
+	ioseq := usermem.BytesIOSequence(buf)
+	n, err := pipeOps.Read(ctx, file, ioseq, 0)
+	if err != nil {
+		t.Fatalf("read request got error %v, want nil", err)
+	}
+	if n != int64(len(data)) {
+		t.Fatalf("read request got %d bytes, want %d", n, len(data))
+	}
+	if !bytes.Equal(buf, data) {
+		t.Errorf("read request got bytes [%v], want [%v]", buf, data)
+	}
+}
+
+func TestPipeHangup(t *testing.T) {
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// flags control how we open our end of the pipe and must be read
+		// only or write only.  They also dicate how a coordinating partner
+		// fd is opened, which is their inverse (read only -> write only, etc).
+		flags fs.FileFlags
+
+		// hangupSelf if true causes the test case to close our end of the pipe
+		// and causes hangup errors to be asserted on our coordinating partner's
+		// fd.  If hangupSelf is false, then our partner's fd is closed and the
+		// hangup errors are expected on our end of the pipe.
+		hangupSelf bool
+	}{
+		{
+			desc:  "Read only gets hangup error",
+			flags: fs.FileFlags{Read: true},
+		},
+		{
+			desc:  "Write only gets hangup error",
+			flags: fs.FileFlags{Write: true},
+		},
+		{
+			desc:       "Read only generates hangup error",
+			flags:      fs.FileFlags{Read: true},
+			hangupSelf: true,
+		},
+		{
+			desc:       "Write only generates hangup error",
+			flags:      fs.FileFlags{Write: true},
+			hangupSelf: true,
+		},
+	} {
+		if test.flags.Read == test.flags.Write {
+			t.Errorf("%s: test requires a single reader or writer", test.desc)
+			continue
+		}
+
+		// Create the pipe.  We do this per-test case to keep tests independent.
+		name := pipename()
+		if err := mkpipe(name); err != nil {
+			t.Errorf("%s: failed to make host pipe: %v", test.desc, err)
+			continue
+		}
+		defer syscall.Unlink(name)
+
+		// Fire off a partner routine which tries to open the same pipe blocking,
+		// which will synchronize with us.  The channel allows us to get back the
+		// fd once we expect this partner routine to succeed, so we can manifest
+		// hangup events more directly.
+		fdchan := make(chan int, 1)
+		go func() {
+			// Be explicit about the flags to protect the test from
+			// misconfiguration.
+			var flags int
+			if test.flags.Read {
+				flags = syscall.O_WRONLY
+			} else {
+				flags = syscall.O_RDONLY
+			}
+			fd, err := syscall.Open(name, flags, 0666)
+			if err != nil {
+				t.Logf("Open(%q, %o, 0666) partner failed: %v", name, flags, err)
+			}
+			fdchan <- fd
+		}()
+
+		// Open our end in a blocking way to ensure that we coordinate.
+		opener := &hostOpener{name: name}
+		ctx := contexttest.Context(t)
+		pipeOps, err := Open(ctx, opener, test.flags)
+		if err != nil {
+			t.Errorf("%s: Open got error %v, want nil", test.desc, err)
+			continue
+		}
+		// Don't defer file.DecRef here because that causes the hangup we're
+		// trying to test for.
+
+		// Expect the partner routine to have coordinated with us and get back
+		// its open fd.
+		f := <-fdchan
+		if f < 0 {
+			t.Errorf("%s: partner routine got fd %d, want > 0", test.desc, f)
+			pipeOps.Release()
+			continue
+		}
+
+		if test.hangupSelf {
+			// Hangup self and assert that our partner got the expected hangup
+			// error.
+			pipeOps.Release()
+
+			if test.flags.Read {
+				// Partner is writer.
+				assertWriterHungup(t, test.desc, fd.NewReadWriter(f))
+			} else {
+				// Partner is reader.
+				assertReaderHungup(t, test.desc, fd.NewReadWriter(f))
+			}
+		} else {
+			// Hangup our partner and expect us to get the hangup error.
+			syscall.Close(f)
+			defer pipeOps.Release()
+
+			if test.flags.Read {
+				assertReaderHungup(t, test.desc, pipeOps.(*pipeOperations).file)
+			} else {
+				assertWriterHungup(t, test.desc, pipeOps.(*pipeOperations).file)
+			}
+		}
+	}
+}
+
+func assertReaderHungup(t *testing.T, desc string, reader io.Reader) bool {
+	// Drain the pipe completely, it might have crap in it, but expect EOF eventually.
+	var err error
+	for err == nil {
+		_, err = reader.Read(make([]byte, 10))
+	}
+	if err != io.EOF {
+		t.Errorf("%s: read from self after hangup got error %v, want %v", desc, err, io.EOF)
+		return false
+	}
+	return true
+}
+
+func assertWriterHungup(t *testing.T, desc string, writer io.Writer) bool {
+	if _, err := writer.Write([]byte("hello")); unwrapError(err) != syscall.EPIPE {
+		t.Errorf("%s: write to self after hangup got error %v, want %v", desc, err, syscall.EPIPE)
+		return false
+	}
+	return true
+}
diff --git a/pkg/sentry/fs/fdpipe/pipe_state.go b/pkg/sentry/fs/fdpipe/pipe_state.go
new file mode 100644
index 000000000..8996a2178
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/pipe_state.go
@@ -0,0 +1,88 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fdpipe
+
+import (
+	"fmt"
+	"io/ioutil"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// beforeSave is invoked by stateify.
+func (p *pipeOperations) beforeSave() {
+	if p.flags.Read {
+		data, err := ioutil.ReadAll(p.file)
+		if err != nil && !isBlockError(err) {
+			panic(fmt.Sprintf("failed to read from pipe: %v", err))
+		}
+		p.readAheadBuffer = append(p.readAheadBuffer, data...)
+	} else if p.flags.Write {
+		file, err := p.opener.NonBlockingOpen(context.Background(), fs.PermMask{Write: true})
+		if err != nil {
+			panic(fs.ErrSaveRejection{fmt.Errorf("write-only pipe end cannot be re-opened as %v: %v", p, err)})
+		}
+		file.Close()
+	}
+}
+
+// saveFlags is invoked by stateify.
+func (p *pipeOperations) saveFlags() fs.FileFlags {
+	return p.flags
+}
+
+// readPipeOperationsLoading is used to ensure that write-only pipe fds are
+// opened after read/write and read-only pipe fds, to avoid ENXIO when
+// multiple pipe fds refer to different ends of the same pipe.
+var readPipeOperationsLoading sync.WaitGroup
+
+// loadFlags is invoked by stateify.
+func (p *pipeOperations) loadFlags(flags fs.FileFlags) {
+	// This is a hack to ensure that readPipeOperationsLoading includes all
+	// readable pipe fds before any asynchronous calls to
+	// readPipeOperationsLoading.Wait().
+	if flags.Read {
+		readPipeOperationsLoading.Add(1)
+	}
+	p.flags = flags
+}
+
+// afterLoad is invoked by stateify.
+func (p *pipeOperations) afterLoad() {
+	load := func() {
+		if !p.flags.Read {
+			readPipeOperationsLoading.Wait()
+		} else {
+			defer readPipeOperationsLoading.Done()
+		}
+		var err error
+		p.file, err = p.opener.NonBlockingOpen(context.Background(), fs.PermMask{
+			Read:  p.flags.Read,
+			Write: p.flags.Write,
+		})
+		if err != nil {
+			panic(fmt.Sprintf("unable to open pipe %v: %v", p, err))
+		}
+		if err := p.init(); err != nil {
+			panic(fmt.Sprintf("unable to initialize pipe %v: %v", p, err))
+		}
+	}
+
+	// Do background opening of pipe ends. Note for write-only pipe ends we
+	// have to do it asynchronously to avoid blocking the restore.
+	fs.Async(load)
+}
diff --git a/pkg/sentry/fs/fdpipe/pipe_test.go b/pkg/sentry/fs/fdpipe/pipe_test.go
new file mode 100644
index 000000000..6cd314f5b
--- /dev/null
+++ b/pkg/sentry/fs/fdpipe/pipe_test.go
@@ -0,0 +1,489 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fdpipe
+
+import (
+	"bytes"
+	"io"
+	"os"
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+func singlePipeFD() (int, error) {
+	fds := make([]int, 2)
+	if err := syscall.Pipe(fds); err != nil {
+		return -1, err
+	}
+	syscall.Close(fds[1])
+	return fds[0], nil
+}
+
+func singleDirFD() (int, error) {
+	return syscall.Open(os.TempDir(), syscall.O_RDONLY, 0666)
+}
+
+func mockPipeDirent(t *testing.T) *fs.Dirent {
+	ctx := contexttest.Context(t)
+	node := fs.NewMockInodeOperations(ctx)
+	node.UAttr = fs.UnstableAttr{
+		Perms: fs.FilePermissions{
+			User: fs.PermMask{Read: true, Write: true},
+		},
+	}
+	inode := fs.NewInode(node, fs.NewMockMountSource(nil), fs.StableAttr{
+		Type:      fs.Pipe,
+		BlockSize: usermem.PageSize,
+	})
+	return fs.NewDirent(inode, "")
+}
+
+func TestNewPipe(t *testing.T) {
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// getfd generates the fd to pass to newPipeOperations.
+		getfd func() (int, error)
+
+		// flags are the fs.FileFlags passed to newPipeOperations.
+		flags fs.FileFlags
+
+		// readAheadBuffer is the buffer passed to newPipeOperations.
+		readAheadBuffer []byte
+
+		// err is the expected error.
+		err error
+	}{
+		{
+			desc:  "Cannot make new pipe from bad fd",
+			getfd: func() (int, error) { return -1, nil },
+			err:   syscall.EINVAL,
+		},
+		{
+			desc:  "Cannot make new pipe from non-pipe fd",
+			getfd: singleDirFD,
+			err:   syscall.EINVAL,
+		},
+		{
+			desc:            "Can make new pipe from pipe fd",
+			getfd:           singlePipeFD,
+			flags:           fs.FileFlags{Read: true},
+			readAheadBuffer: []byte("hello"),
+		},
+	} {
+		gfd, err := test.getfd()
+		if err != nil {
+			t.Errorf("%s: getfd got (%d, %v), want (fd, nil)", test.desc, gfd, err)
+			continue
+		}
+		f := fd.New(gfd)
+
+		p, err := newPipeOperations(contexttest.Context(t), nil, test.flags, f, test.readAheadBuffer)
+		if p != nil {
+			// This is necessary to remove the fd from the global fd notifier.
+			defer p.Release()
+		} else {
+			// If there is no p to DecRef on, because newPipeOperations failed, then the
+			// file still needs to be closed.
+			defer f.Close()
+		}
+
+		if err != test.err {
+			t.Errorf("%s: got error %v, want %v", test.desc, err, test.err)
+			continue
+		}
+		// Check the state of the pipe given that it was successfully opened.
+		if err == nil {
+			if p == nil {
+				t.Errorf("%s: got nil pipe and nil error, want (pipe, nil)", test.desc)
+				continue
+			}
+			if flags := p.flags; test.flags != flags {
+				t.Errorf("%s: got file flags %s, want %s", test.desc, flags, test.flags)
+				continue
+			}
+			if len(test.readAheadBuffer) != len(p.readAheadBuffer) {
+				t.Errorf("%s: got read ahead buffer length %d, want %d", test.desc, len(p.readAheadBuffer), len(test.readAheadBuffer))
+				continue
+			}
+			fileFlags, _, errno := syscall.Syscall(syscall.SYS_FCNTL, uintptr(p.file.FD()), syscall.F_GETFL, 0)
+			if errno != 0 {
+				t.Errorf("%s: failed to get file flags for fd %d, got %v, want 0", test.desc, p.file.FD(), errno)
+				continue
+			}
+			if fileFlags&syscall.O_NONBLOCK == 0 {
+				t.Errorf("%s: pipe is blocking, expected non-blocking", test.desc)
+				continue
+			}
+			if !fdnotifier.HasFD(int32(f.FD())) {
+				t.Errorf("%s: pipe fd %d is not registered for events", test.desc, f.FD)
+			}
+		}
+	}
+}
+
+func TestPipeDestruction(t *testing.T) {
+	fds := make([]int, 2)
+	if err := syscall.Pipe(fds); err != nil {
+		t.Fatalf("failed to create pipes: got %v, want nil", err)
+	}
+	f := fd.New(fds[0])
+
+	// We don't care about the other end, just use the read end.
+	syscall.Close(fds[1])
+
+	// Test the read end, but it doesn't really matter which.
+	p, err := newPipeOperations(contexttest.Context(t), nil, fs.FileFlags{Read: true}, f, nil)
+	if err != nil {
+		f.Close()
+		t.Fatalf("newPipeOperations got error %v, want nil", err)
+	}
+	// Drop our only reference, which should trigger the destructor.
+	p.Release()
+
+	if fdnotifier.HasFD(int32(fds[0])) {
+		t.Fatalf("after DecRef fdnotifier has fd %d, want no longer registered", fds[0])
+	}
+	if p.file != nil {
+		t.Errorf("after DecRef got file, want nil")
+	}
+}
+
+type Seek struct{}
+
+type ReadDir struct{}
+
+type Writev struct {
+	Src usermem.IOSequence
+}
+
+type Readv struct {
+	Dst usermem.IOSequence
+}
+
+type Fsync struct{}
+
+func TestPipeRequest(t *testing.T) {
+	for _, test := range []struct {
+		// desc is the test's description.
+		desc string
+
+		// request to execute.
+		context interface{}
+
+		// flags determines whether to use the read or write end
+		// of the pipe, for this test it can only be Read or Write.
+		flags fs.FileFlags
+
+		// keepOpenPartner if false closes the other end of the pipe,
+		// otherwise this is delayed until the end of the test.
+		keepOpenPartner bool
+
+		// expected error
+		err error
+	}{
+		{
+			desc:    "ReadDir on pipe returns ENOTDIR",
+			context: &ReadDir{},
+			err:     syscall.ENOTDIR,
+		},
+		{
+			desc:    "Fsync on pipe returns EINVAL",
+			context: &Fsync{},
+			err:     syscall.EINVAL,
+		},
+		{
+			desc:    "Seek on pipe returns ESPIPE",
+			context: &Seek{},
+			err:     syscall.ESPIPE,
+		},
+		{
+			desc:    "Readv on pipe from empty buffer returns nil",
+			context: &Readv{Dst: usermem.BytesIOSequence(nil)},
+			flags:   fs.FileFlags{Read: true},
+		},
+		{
+			desc:    "Readv on pipe from non-empty buffer and closed partner returns EOF",
+			context: &Readv{Dst: usermem.BytesIOSequence(make([]byte, 10))},
+			flags:   fs.FileFlags{Read: true},
+			err:     io.EOF,
+		},
+		{
+			desc:            "Readv on pipe from non-empty buffer and open partner returns EWOULDBLOCK",
+			context:         &Readv{Dst: usermem.BytesIOSequence(make([]byte, 10))},
+			flags:           fs.FileFlags{Read: true},
+			keepOpenPartner: true,
+			err:             syserror.ErrWouldBlock,
+		},
+		{
+			desc:    "Writev on pipe from empty buffer returns nil",
+			context: &Writev{Src: usermem.BytesIOSequence(nil)},
+			flags:   fs.FileFlags{Write: true},
+		},
+		{
+			desc:    "Writev on pipe from non-empty buffer and closed partner returns EPIPE",
+			context: &Writev{Src: usermem.BytesIOSequence([]byte("hello"))},
+			flags:   fs.FileFlags{Write: true},
+			err:     syscall.EPIPE,
+		},
+		{
+			desc:            "Writev on pipe from non-empty buffer and open partner succeeds",
+			context:         &Writev{Src: usermem.BytesIOSequence([]byte("hello"))},
+			flags:           fs.FileFlags{Write: true},
+			keepOpenPartner: true,
+		},
+	} {
+		if test.flags.Read && test.flags.Write {
+			panic("both read and write not supported for this test")
+		}
+
+		fds := make([]int, 2)
+		if err := syscall.Pipe(fds); err != nil {
+			t.Errorf("%s: failed to create pipes: got %v, want nil", test.desc, err)
+			continue
+		}
+
+		// Configure the fd and partner fd based on the file flags.
+		testFd, partnerFd := fds[0], fds[1]
+		if test.flags.Write {
+			testFd, partnerFd = fds[1], fds[0]
+		}
+
+		// Configure closing the fds.
+		if test.keepOpenPartner {
+			defer syscall.Close(partnerFd)
+		} else {
+			syscall.Close(partnerFd)
+		}
+
+		// Create the pipe.
+		ctx := contexttest.Context(t)
+		p, err := newPipeOperations(ctx, nil, test.flags, fd.New(testFd), nil)
+		if err != nil {
+			t.Fatalf("%s: newPipeOperations got error %v, want nil", test.desc, err)
+		}
+		defer p.Release()
+
+		inode := fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{Type: fs.Pipe})
+		file := fs.NewFile(ctx, fs.NewDirent(inode, "pipe"), fs.FileFlags{Read: true}, p)
+
+		// Issue request via the appropriate function.
+		switch c := test.context.(type) {
+		case *Seek:
+			_, err = p.Seek(ctx, file, 0, 0)
+		case *ReadDir:
+			_, err = p.Readdir(ctx, file, nil)
+		case *Readv:
+			_, err = p.Read(ctx, file, c.Dst, 0)
+		case *Writev:
+			_, err = p.Write(ctx, file, c.Src, 0)
+		case *Fsync:
+			err = p.Fsync(ctx, file, 0, fs.FileMaxOffset, fs.SyncAll)
+		default:
+			t.Errorf("%s: unknown request type %T", test.desc, test.context)
+		}
+
+		if unwrapError(err) != test.err {
+			t.Errorf("%s: got error %v, want %v", test.desc, err, test.err)
+		}
+	}
+}
+
+func TestPipeReadAheadBuffer(t *testing.T) {
+	fds := make([]int, 2)
+	if err := syscall.Pipe(fds); err != nil {
+		t.Fatalf("failed to create pipes: got %v, want nil", err)
+	}
+	rfile := fd.New(fds[0])
+
+	// Eventually close the write end, which is not wrapped in a pipe object.
+	defer syscall.Close(fds[1])
+
+	// Write some bytes to this end.
+	data := []byte("world")
+	if n, err := syscall.Write(fds[1], data); n != len(data) || err != nil {
+		rfile.Close()
+		t.Fatalf("write to pipe got (%d, %v), want (%d, nil)", n, err, len(data))
+	}
+	// Close the write end immediately, we don't care about it.
+
+	buffered := []byte("hello ")
+	ctx := contexttest.Context(t)
+	p, err := newPipeOperations(ctx, nil, fs.FileFlags{Read: true}, rfile, buffered)
+	if err != nil {
+		rfile.Close()
+		t.Fatalf("newPipeOperations got error %v, want nil", err)
+	}
+	defer p.Release()
+
+	inode := fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{
+		Type: fs.Pipe,
+	})
+	file := fs.NewFile(ctx, fs.NewDirent(inode, "pipe"), fs.FileFlags{Read: true}, p)
+
+	// In total we expect to read data + buffered.
+	total := append(buffered, data...)
+
+	buf := make([]byte, len(total))
+	iov := usermem.BytesIOSequence(buf)
+	n, err := p.Read(contexttest.Context(t), file, iov, 0)
+	if err != nil {
+		t.Fatalf("read request got error %v, want nil", err)
+	}
+	if n != int64(len(total)) {
+		t.Fatalf("read request got %d bytes, want %d", n, len(total))
+	}
+	if !bytes.Equal(buf, total) {
+		t.Errorf("read request got bytes [%v], want [%v]", buf, total)
+	}
+}
+
+// This is very important for pipes in general because they can return EWOULDBLOCK and for
+// those that block they must continue until they have read all of the data (and report it
+// as such.
+func TestPipeReadsAccumulate(t *testing.T) {
+	fds := make([]int, 2)
+	if err := syscall.Pipe(fds); err != nil {
+		t.Fatalf("failed to create pipes: got %v, want nil", err)
+	}
+	rfile := fd.New(fds[0])
+
+	// Eventually close the write end, it doesn't depend on a pipe object.
+	defer syscall.Close(fds[1])
+
+	// Get a new read only pipe reference.
+	ctx := contexttest.Context(t)
+	p, err := newPipeOperations(ctx, nil, fs.FileFlags{Read: true}, rfile, nil)
+	if err != nil {
+		rfile.Close()
+		t.Fatalf("newPipeOperations got error %v, want nil", err)
+	}
+	// Don't forget to remove the fd from the fd notifier.  Otherwise other tests will
+	// likely be borked, because it's global :(
+	defer p.Release()
+
+	inode := fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{
+		Type: fs.Pipe,
+	})
+	file := fs.NewFile(ctx, fs.NewDirent(inode, "pipe"), fs.FileFlags{Read: true}, p)
+
+	// Write some some bytes to the pipe.
+	data := []byte("some message")
+	if n, err := syscall.Write(fds[1], data); n != len(data) || err != nil {
+		t.Fatalf("write to pipe got (%d, %v), want (%d, nil)", n, err, len(data))
+	}
+
+	// Construct a segment vec that is a bit more than we have written so we trigger
+	// an EWOULDBLOCK.
+	wantBytes := len(data) + 1
+	readBuffer := make([]byte, wantBytes)
+	iov := usermem.BytesIOSequence(readBuffer)
+	n, err := p.Read(ctx, file, iov, 0)
+	total := n
+	iov = iov.DropFirst64(n)
+	if err != syserror.ErrWouldBlock {
+		t.Fatalf("Readv got error %v, want %v", err, syserror.ErrWouldBlock)
+	}
+
+	// Write a few more bytes to allow us to read more/accumulate.
+	extra := []byte("extra")
+	if n, err := syscall.Write(fds[1], extra); n != len(extra) || err != nil {
+		t.Fatalf("write to pipe got (%d, %v), want (%d, nil)", n, err, len(extra))
+	}
+
+	// This time, using the same request, we should not block.
+	n, err = p.Read(ctx, file, iov, 0)
+	total += n
+	if err != nil {
+		t.Fatalf("Readv got error %v, want nil", err)
+	}
+
+	// Assert that the result we got back is cumulative.
+	if total != int64(wantBytes) {
+		t.Fatalf("Readv sequence got %d bytes, want %d", total, wantBytes)
+	}
+
+	if want := append(data, extra[0]); !bytes.Equal(readBuffer, want) {
+		t.Errorf("Readv sequence got %v, want %v", readBuffer, want)
+	}
+}
+
+// Same as TestReadsAccumulate.
+func TestPipeWritesAccumulate(t *testing.T) {
+	fds := make([]int, 2)
+	if err := syscall.Pipe(fds); err != nil {
+		t.Fatalf("failed to create pipes: got %v, want nil", err)
+	}
+	wfile := fd.New(fds[1])
+
+	// Eventually close the read end, it doesn't depend on a pipe object.
+	defer syscall.Close(fds[0])
+
+	// Get a new write only pipe reference.
+	ctx := contexttest.Context(t)
+	p, err := newPipeOperations(ctx, nil, fs.FileFlags{Write: true}, wfile, nil)
+	if err != nil {
+		wfile.Close()
+		t.Fatalf("newPipeOperations got error %v, want nil", err)
+	}
+	// Don't forget to remove the fd from the fd notifier.  Otherwise other tests will
+	// likely be borked, because it's global :(
+	defer p.Release()
+
+	inode := fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{
+		Type: fs.Pipe,
+	})
+	file := fs.NewFile(ctx, fs.NewDirent(inode, "pipe"), fs.FileFlags{Read: true}, p)
+
+	// Construct a segment vec that is larger than the pipe size to trigger an EWOULDBLOCK.
+	wantBytes := 65536 * 2
+	writeBuffer := make([]byte, wantBytes)
+	for i := 0; i < wantBytes; i++ {
+		writeBuffer[i] = 'a'
+	}
+	iov := usermem.BytesIOSequence(writeBuffer)
+	n, err := p.Write(ctx, file, iov, 0)
+	total := n
+	iov = iov.DropFirst64(n)
+	if err != syserror.ErrWouldBlock {
+		t.Fatalf("Writev got error %v, want %v", err, syserror.ErrWouldBlock)
+	}
+
+	// Read the entire pipe buf size to make space for the second half.
+	throwAway := make([]byte, 65536)
+	if n, err := syscall.Read(fds[0], throwAway); n != len(throwAway) || err != nil {
+		t.Fatalf("write to pipe got (%d, %v), want (%d, nil)", n, err, len(throwAway))
+	}
+
+	// This time we should not block.
+	n, err = p.Write(ctx, file, iov, 0)
+	total += n
+	if err != nil {
+		t.Fatalf("Writev got error %v, want nil", err)
+	}
+
+	// Assert that the result we got back is cumulative.
+	if total != int64(wantBytes) {
+		t.Fatalf("Writev sequence got %d bytes, want %d", total, wantBytes)
+	}
+}
diff --git a/pkg/sentry/fs/file.go b/pkg/sentry/fs/file.go
new file mode 100644
index 000000000..de2e80bf0
--- /dev/null
+++ b/pkg/sentry/fs/file.go
@@ -0,0 +1,404 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"math"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/amutex"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/lock"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// FileMaxOffset is the maximum possible file offset.
+const FileMaxOffset = math.MaxInt64
+
+// File is an open file handle. It is thread-safe.
+//
+// File provides stronger synchronization guarantees than Linux. Linux
+// synchronizes lseek(2), read(2), and write(2) with respect to the file
+// offset for regular files and only for those interfaces. See
+// fs/read_write.c:fdget_pos, fs.read_write.c:fdput_pos and FMODE_ATOMIC_POS.
+//
+// In contrast, File synchronizes any operation that could take a long time
+// under a single abortable mutex which also synchronizes lseek(2), read(2),
+// and write(2).
+//
+// FIXME: Split synchronization from cancellation.
+type File struct {
+	refs.AtomicRefCount
+
+	// UniqueID is the globally unique identifier of the File.
+	UniqueID uint64
+
+	// Dirent is the Dirent backing this File. This encodes the name
+	// of the File via Dirent.FullName() as well as its identity via the
+	// Dirent's Inode. The Dirent is non-nil.
+	//
+	// A File holds a reference to this Dirent. Using the returned Dirent is
+	// only safe as long as a reference on the File is held. The association
+	// between a File and a Dirent is immutable.
+	//
+	// Files that are not parented in a filesystem return a root Dirent
+	// that holds a reference to their Inode.
+	//
+	// The name of the Dirent may reflect parentage if the Dirent is not a
+	// root Dirent or the identity of the File on a pseudo filesystem (pipefs,
+	// sockfs, etc).
+	//
+	// Multiple Files may hold a reference to the same Dirent. This is the
+	// common case for Files that are parented and maintain consistency with
+	// other files via the Dirent cache.
+	Dirent *Dirent
+
+	// flags are the File's flags. Setting or getting flags is fully atomic
+	// and is not protected by mu (below).
+	flags atomic.Value `state:".(FileFlags)"`
+
+	// mu is dual-purpose: first, to make read(2) and write(2) thread-safe
+	// in conformity with POSIX, and second, to cancel operations before they
+	// begin in response to interruptions (i.e. signals).
+	mu amutex.AbortableMutex `state:"nosave"`
+
+	// FileOperations implements file system specific behavior for this File.
+	FileOperations FileOperations
+
+	// offset is the File's offset. Updating offset is protected by mu but
+	// can be read atomically via File.Offset() outside of mu.
+	offset int64
+}
+
+// NewFile returns a File. It takes a reference on the Dirent and owns the
+// lifetime of the FileOperations. Files that do not support reading and
+// writing at an arbitrary offset should set flags.Pread and flags.Pwrite
+// to false respectively.
+func NewFile(ctx context.Context, dirent *Dirent, flags FileFlags, fops FileOperations) *File {
+	dirent.IncRef()
+	f := &File{
+		UniqueID:       uniqueid.GlobalFromContext(ctx),
+		Dirent:         dirent,
+		FileOperations: fops,
+	}
+	f.flags.Store(flags)
+	f.mu.Init()
+	return f
+}
+
+// DecRef destroys the File when it is no longer referenced.
+func (f *File) DecRef() {
+	f.DecRefWithDestructor(func() {
+		// Drop BSD style locks.
+		lockRng := lock.LockRange{Start: 0, End: lock.LockEOF}
+		f.Dirent.Inode.LockCtx.BSD.UnlockRegion(lock.UniqueID(f.UniqueID), lockRng)
+
+		// Release resources held by the FileOperations.
+		f.FileOperations.Release()
+
+		// Release a reference on the Dirent.
+		f.Dirent.DecRef()
+	})
+}
+
+// Flags atomically loads the File's flags.
+func (f *File) Flags() FileFlags {
+	return f.flags.Load().(FileFlags)
+}
+
+// SetFlags atomically changes the File's flags to the values contained
+// in newFlags. See SettableFileFlags for values that can be set.
+func (f *File) SetFlags(newFlags SettableFileFlags) {
+	flags := f.flags.Load().(FileFlags)
+	flags.Direct = newFlags.Direct
+	flags.NonBlocking = newFlags.NonBlocking
+	flags.Append = newFlags.Append
+	f.flags.Store(flags)
+}
+
+// Offset atomically loads the File's offset.
+func (f *File) Offset() int64 {
+	return atomic.LoadInt64(&f.offset)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (f *File) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return f.FileOperations.Readiness(mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (f *File) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	f.FileOperations.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (f *File) EventUnregister(e *waiter.Entry) {
+	f.FileOperations.EventUnregister(e)
+}
+
+// Seek calls f.FileOperations.Seek with f as the File, updating the file
+// offset to the value returned by f.FileOperations.Seek if the operation
+// is successful.
+//
+// Returns syserror.ErrInterrupted if seeking was interrupted.
+func (f *File) Seek(ctx context.Context, whence SeekWhence, offset int64) (int64, error) {
+	if !f.mu.Lock(ctx) {
+		return 0, syserror.ErrInterrupted
+	}
+	defer f.mu.Unlock()
+
+	newOffset, err := f.FileOperations.Seek(ctx, f, whence, offset)
+	if err == nil {
+		atomic.StoreInt64(&f.offset, newOffset)
+	}
+	return newOffset, err
+}
+
+// Readdir reads the directory entries of this File and writes them out
+// to the DentrySerializer until entries can no longer be written. If even
+// a single directory entry is written then Readdir returns a nil error
+// and the directory offset is advanced.
+//
+// Readdir unconditionally updates the access time on the File's Inode,
+// see fs/readdir.c:iterate_dir.
+//
+// Returns syserror.ErrInterrupted if reading was interrupted.
+func (f *File) Readdir(ctx context.Context, serializer DentrySerializer) error {
+	if !f.mu.Lock(ctx) {
+		return syserror.ErrInterrupted
+	}
+	defer f.mu.Unlock()
+
+	offset, err := f.FileOperations.Readdir(ctx, f, serializer)
+	atomic.StoreInt64(&f.offset, offset)
+	return err
+}
+
+// Readv calls f.FileOperations.Read with f as the File, advancing the file
+// offset if f.FileOperations.Read returns bytes read > 0.
+//
+// Returns syserror.ErrInterrupted if reading was interrupted.
+func (f *File) Readv(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	if !f.mu.Lock(ctx) {
+		return 0, syserror.ErrInterrupted
+	}
+
+	n, err := f.FileOperations.Read(ctx, f, dst, f.offset)
+	if n > 0 {
+		atomic.AddInt64(&f.offset, n)
+	}
+	f.mu.Unlock()
+	return n, err
+}
+
+// Preadv calls f.FileOperations.Read with f as the File. It does not
+// advance the file offset. If !f.Flags().Pread, Preadv should not be
+// called.
+//
+// Otherwise same as Readv.
+func (f *File) Preadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if !f.mu.Lock(ctx) {
+		return 0, syserror.ErrInterrupted
+	}
+
+	n, err := f.FileOperations.Read(ctx, f, dst, offset)
+	f.mu.Unlock()
+	return n, err
+}
+
+// Writev calls f.FileOperations.Write with f as the File, advancing the
+// file offset if f.FileOperations.Write returns bytes written > 0.
+//
+// Writev positions the write offset at EOF if f.Flags().Append. This is
+// unavoidably racy for network file systems. Writev also truncates src
+// to avoid overrunning the current file size limit if necessary.
+//
+// Returns syserror.ErrInterrupted if writing was interrupted.
+func (f *File) Writev(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	if !f.mu.Lock(ctx) {
+		return 0, syserror.ErrInterrupted
+	}
+
+	offset, err := f.checkWriteLocked(ctx, &src, f.offset)
+	if err != nil {
+		f.mu.Unlock()
+		return 0, err
+	}
+	n, err := f.FileOperations.Write(ctx, f, src, offset)
+	if n >= 0 {
+		atomic.StoreInt64(&f.offset, offset+n)
+	}
+	f.mu.Unlock()
+	return n, err
+}
+
+// Pwritev calls f.FileOperations.Write with f as the File. It does not
+// advance the file offset. If !f.Flags().Pwritev, Pwritev should not be
+// called.
+//
+// Otherwise same as Writev.
+func (f *File) Pwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if !f.mu.Lock(ctx) {
+		return 0, syserror.ErrInterrupted
+	}
+
+	offset, err := f.checkWriteLocked(ctx, &src, offset)
+	if err != nil {
+		f.mu.Unlock()
+		return 0, err
+	}
+	n, err := f.FileOperations.Write(ctx, f, src, offset)
+	f.mu.Unlock()
+	return n, err
+}
+
+// checkWriteLocked returns the offset to write at or an error if the write
+// would not succeed. May update src to fit a write operation into a file
+// size limit.
+func (f *File) checkWriteLocked(ctx context.Context, src *usermem.IOSequence, offset int64) (int64, error) {
+	// Handle append only files. Note that this is still racy for network
+	// filesystems.
+	if f.Flags().Append {
+		uattr, err := f.Dirent.Inode.UnstableAttr(ctx)
+		if err != nil {
+			// This is an odd error, most likely it is evidence
+			// that something is terribly wrong with the filesystem.
+			// Return a generic EIO error.
+			log.Warningf("Failed to check write of inode %#v: %v", f.Dirent.Inode.StableAttr, err)
+			return offset, syserror.EIO
+		}
+		offset = uattr.Size
+	}
+
+	// Is this a regular file?
+	if IsRegular(f.Dirent.Inode.StableAttr) {
+		// Enforce size limits.
+		fileSizeLimit := limits.FromContext(ctx).Get(limits.FileSize).Cur
+		if fileSizeLimit <= math.MaxInt64 {
+			if offset >= int64(fileSizeLimit) {
+				return offset, syserror.ErrExceedsFileSizeLimit
+			}
+			*src = src.TakeFirst64(int64(fileSizeLimit) - offset)
+		}
+	}
+
+	return offset, nil
+}
+
+// Fsync calls f.FileOperations.Fsync with f as the File.
+//
+// Returns syserror.ErrInterrupted if syncing was interrupted.
+func (f *File) Fsync(ctx context.Context, start int64, end int64, syncType SyncType) error {
+	if !f.mu.Lock(ctx) {
+		return syserror.ErrInterrupted
+	}
+	defer f.mu.Unlock()
+
+	return f.FileOperations.Fsync(ctx, f, start, end, syncType)
+}
+
+// Flush calls f.FileOperations.Flush with f as the File.
+//
+// Returns syserror.ErrInterrupted if syncing was interrupted.
+func (f *File) Flush(ctx context.Context) error {
+	if !f.mu.Lock(ctx) {
+		return syserror.ErrInterrupted
+	}
+	defer f.mu.Unlock()
+
+	return f.FileOperations.Flush(ctx, f)
+}
+
+// ConfigureMMap calls f.FileOperations.ConfigureMMap with f as the File.
+//
+// Returns syserror.ErrInterrupted if interrupted.
+func (f *File) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error {
+	if !f.mu.Lock(ctx) {
+		return syserror.ErrInterrupted
+	}
+	defer f.mu.Unlock()
+
+	return f.FileOperations.ConfigureMMap(ctx, f, opts)
+}
+
+// MappedName implements memmap.MappingIdentity.MappedName.
+func (f *File) MappedName(ctx context.Context) string {
+	name, _ := f.Dirent.FullName(RootFromContext(ctx))
+	return name
+}
+
+// DeviceID implements memmap.MappingIdentity.DeviceID.
+func (f *File) DeviceID() uint64 {
+	return f.Dirent.Inode.StableAttr.DeviceID
+}
+
+// InodeID implements memmap.MappingIdentity.InodeID.
+func (f *File) InodeID() uint64 {
+	return f.Dirent.Inode.StableAttr.InodeID
+}
+
+// Msync implements memmap.MappingIdentity.Msync.
+func (f *File) Msync(ctx context.Context, mr memmap.MappableRange) error {
+	return f.Fsync(ctx, int64(mr.Start), int64(mr.End-1), SyncData)
+}
+
+// FileReader implements io.Reader and io.ReaderAt.
+type FileReader struct {
+	// Ctx is the context for the file reader.
+	Ctx context.Context
+
+	// File is the file to read from.
+	File *File
+}
+
+// Read implements io.Reader.Read.
+func (r *FileReader) Read(buf []byte) (int, error) {
+	n, err := r.File.Readv(r.Ctx, usermem.BytesIOSequence(buf))
+	return int(n), err
+}
+
+// ReadAt implementes io.Reader.ReadAt.
+func (r *FileReader) ReadAt(buf []byte, offset int64) (int, error) {
+	n, err := r.File.Preadv(r.Ctx, usermem.BytesIOSequence(buf), offset)
+	return int(n), err
+}
+
+// FileWriter implements io.Writer and io.WriterAt.
+type FileWriter struct {
+	// Ctx is the context for the file writer.
+	Ctx context.Context
+
+	// File is the file to write to.
+	File *File
+}
+
+// Write implements io.Writer.Write.
+func (w *FileWriter) Write(buf []byte) (int, error) {
+	n, err := w.File.Writev(w.Ctx, usermem.BytesIOSequence(buf))
+	return int(n), err
+}
+
+// WriteAt implements io.Writer.WriteAt.
+func (w *FileWriter) WriteAt(buf []byte, offset int64) (int, error) {
+	n, err := w.File.Pwritev(w.Ctx, usermem.BytesIOSequence(buf), offset)
+	return int(n), err
+}
diff --git a/pkg/sentry/fs/file_operations.go b/pkg/sentry/fs/file_operations.go
new file mode 100644
index 000000000..d223bb5c7
--- /dev/null
+++ b/pkg/sentry/fs/file_operations.go
@@ -0,0 +1,106 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// FileOperations are operations on a File that diverge per file system.
+//
+// Operations that take a *File may use only the following interfaces:
+//
+// - File.UniqueID:	Operations may only read this value.
+// - File.Dirent:	Operations must not take or drop a reference.
+// - File.Offset(): 	This value is guaranteed to not change for the
+//			duration of the operation.
+// - File.Flags():	This value may change during the operation.
+type FileOperations interface {
+	// Release release resources held by FileOperations.
+	Release()
+
+	// Waitable defines how this File can be waited on for read and
+	// write readiness.
+	waiter.Waitable
+
+	// Seek seeks to offset based on SeekWhence. Returns the new
+	// offset or no change in the offset and an error.
+	Seek(ctx context.Context, file *File, whence SeekWhence, offset int64) (int64, error)
+
+	// Readdir reads the directory entries of file and serializes them
+	// using serializer.
+	//
+	// Returns the new directory offset or no change in the offset and
+	// an error. The offset returned must not be less than file.Offset().
+	//
+	// Serialization of directory entries must not happen asynchronously.
+	Readdir(ctx context.Context, file *File, serializer DentrySerializer) (int64, error)
+
+	// Read reads from file into dst at offset and returns the number
+	// of bytes read which must be greater than or equal to 0. File
+	// systems that do not support reading at an offset, (i.e. pipefs,
+	// sockfs) may ignore the offset. These file systems are expected
+	// to construct Files with !FileFlags.Pread.
+	//
+	// Read may return a nil error and only partially fill dst (at or
+	// before EOF). If the file represents a symlink, Read reads the target
+	// value of the symlink.
+	//
+	// Read does not check permissions nor flags.
+	//
+	// Read must not be called if !FileFlags.Read.
+	Read(ctx context.Context, file *File, dst usermem.IOSequence, offset int64) (int64, error)
+
+	// Write writes src to file at offset and returns the number of bytes
+	// written which must be greater than or equal to 0. Like Read, file
+	// systems that do not support writing at an offset (i.e. pipefs, sockfs)
+	// may ignore the offset. These file systems are expected to construct
+	// Files with !FileFlags.Pwrite.
+	//
+	// If only part of src could be written, Write must return an error
+	// indicating why (e.g. syserror.ErrWouldBlock).
+	//
+	// Write does not check permissions nor flags.
+	//
+	// Write must not be called if !FileFlags.Write.
+	Write(ctx context.Context, file *File, src usermem.IOSequence, offset int64) (int64, error)
+
+	// Fsync writes buffered modifications of file and/or flushes in-flight
+	// operations to backing storage based on syncType. The range to sync is
+	// [start, end]. The end is inclusive so that the last byte of a maximally
+	// sized file can be synced.
+	Fsync(ctx context.Context, file *File, start, end int64, syncType SyncType) error
+
+	// Flush this file's buffers/state (on close(2)).
+	Flush(ctx context.Context, file *File) error
+
+	// ConfigureMMap mutates opts to implement mmap(2) for the file. Most
+	// implementations can either embed fsutil.NoMMap (if they don't support
+	// memory mapping) or call fsutil.GenericConfigureMMap with the appropriate
+	// memmap.Mappable.
+	ConfigureMMap(ctx context.Context, file *File, opts *memmap.MMapOpts) error
+
+	// Ioctl implements the ioctl(2) linux syscall.
+	//
+	// io provides access to the virtual memory space to which pointers in args
+	// refer.
+	//
+	// Preconditions: The AddressSpace (if any) that io refers to is activated.
+	Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error)
+}
diff --git a/pkg/sentry/fs/file_overlay.go b/pkg/sentry/fs/file_overlay.go
new file mode 100644
index 000000000..0c6e622b9
--- /dev/null
+++ b/pkg/sentry/fs/file_overlay.go
@@ -0,0 +1,345 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// overlayFile gets a handle to a file from the upper or lower filesystem
+// in an overlay. The caller is responsible for calling File.DecRef on
+// the returned file.
+func overlayFile(ctx context.Context, inode *Inode, flags FileFlags) (*File, error) {
+	// Do a song and dance to eventually get to:
+	//
+	//   File -> single reference
+	//   Dirent -> single reference
+	//   Inode -> multiple references
+	//
+	// So that File.DecRef() -> File.destroy -> Dirent.DecRef -> Dirent.destroy,
+	// and both the transitory File and Dirent can be GC'ed but the Inode
+	// remains.
+
+	// Take another reference on the Inode.
+	inode.IncRef()
+
+	// Start with a single reference on the Dirent. It inherits the reference
+	// we just took on the Inode above.
+	dirent := NewTransientDirent(inode)
+
+	// Get a File. This will take another reference on the Dirent.
+	f, err := inode.GetFile(ctx, dirent, flags)
+
+	// Drop the extra reference on the Dirent. Now there's only one reference
+	// on the dirent, either owned by f (if non-nil), or the Dirent is about
+	// to be destroyed (if GetFile failed).
+	dirent.DecRef()
+
+	return f, err
+}
+
+// overlayFileOperations implements FileOperations for a file in an overlay.
+type overlayFileOperations struct {
+	// upperMu protects upper below. In contrast lower is stable.
+	upperMu sync.Mutex `state:"nosave"`
+
+	// We can't share Files in upper and lower filesystems between all Files
+	// in an overlay because some file systems expect to get distinct handles
+	// that are not consistent with each other on open(2).
+	//
+	// So we lazily acquire an upper File when the overlayEntry acquires an
+	// upper Inode (it might have one from the start). This synchronizes with
+	// copy up.
+	//
+	// If upper is non-nil and this is not a directory, then lower is ignored.
+	//
+	// For directories, upper and lower are ignored because it is always
+	// necessary to acquire new directory handles so that the directory cursors
+	// of the upper and lower Files are not exhausted.
+	upper *File
+	lower *File
+
+	// dirCursor is a directory cursor for a directory in an overlay.
+	dirCursor string
+
+	// dirCache is cache of DentAttrs from upper and lower Inodes.
+	dirCache *SortedDentryMap
+}
+
+// Release implements FileOperations.Release.
+func (f *overlayFileOperations) Release() {
+	if f.upper != nil {
+		f.upper.DecRef()
+	}
+	if f.lower != nil {
+		f.lower.DecRef()
+	}
+}
+
+// EventRegister implements FileOperations.EventRegister.
+func (f *overlayFileOperations) EventRegister(we *waiter.Entry, mask waiter.EventMask) {
+	f.upperMu.Lock()
+	defer f.upperMu.Unlock()
+	if f.upper != nil {
+		f.upper.EventRegister(we, mask)
+		return
+	}
+	f.lower.EventRegister(we, mask)
+}
+
+// EventUnregister implements FileOperations.Unregister.
+func (f *overlayFileOperations) EventUnregister(we *waiter.Entry) {
+	f.upperMu.Lock()
+	defer f.upperMu.Unlock()
+	if f.upper != nil {
+		f.upper.EventUnregister(we)
+		return
+	}
+	f.lower.EventUnregister(we)
+}
+
+// Readiness implements FileOperations.Readiness.
+func (f *overlayFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	f.upperMu.Lock()
+	defer f.upperMu.Unlock()
+	if f.upper != nil {
+		return f.upper.Readiness(mask)
+	}
+	return f.lower.Readiness(mask)
+}
+
+// Seek implements FileOperations.Seek.
+func (f *overlayFileOperations) Seek(ctx context.Context, file *File, whence SeekWhence, offset int64) (int64, error) {
+	f.upperMu.Lock()
+	defer f.upperMu.Unlock()
+
+	var seekDir bool
+	var n int64
+	if f.upper != nil {
+		var err error
+		if n, err = f.upper.FileOperations.Seek(ctx, file, whence, offset); err != nil {
+			return n, err
+		}
+		seekDir = IsDir(f.upper.Dirent.Inode.StableAttr)
+	} else {
+		var err error
+		if n, err = f.lower.FileOperations.Seek(ctx, file, whence, offset); err != nil {
+			return n, err
+		}
+		seekDir = IsDir(f.lower.Dirent.Inode.StableAttr)
+	}
+
+	// If this was a seek on a directory, we must update the cursor.
+	if seekDir && whence == SeekSet && offset == 0 {
+		// Currenly only seeking to 0 on a directory is supported.
+		// FIXME: Lift directory seeking limitations.
+		f.dirCursor = ""
+	}
+	return n, nil
+}
+
+// Readdir implements FileOperations.Readdir.
+func (f *overlayFileOperations) Readdir(ctx context.Context, file *File, serializer DentrySerializer) (int64, error) {
+	o := file.Dirent.Inode.overlay
+
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	var err error
+	f.dirCache, err = readdirEntries(ctx, o)
+	if err != nil {
+		return file.Offset(), err
+	}
+
+	root := RootFromContext(ctx)
+	defer root.DecRef()
+
+	dirCtx := &DirCtx{
+		Serializer: serializer,
+		DirCursor:  &f.dirCursor,
+	}
+	return DirentReaddir(ctx, file.Dirent, f, root, dirCtx, file.Offset())
+}
+
+// IterateDir implements DirIterator.IterateDir.
+func (f *overlayFileOperations) IterateDir(ctx context.Context, dirCtx *DirCtx, offset int) (int, error) {
+	n, err := GenericReaddir(dirCtx, f.dirCache)
+	return offset + n, err
+}
+
+// Read implements FileOperations.Read.
+func (f *overlayFileOperations) Read(ctx context.Context, file *File, dst usermem.IOSequence, offset int64) (int64, error) {
+	o := file.Dirent.Inode.overlay
+
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	if o.upper != nil {
+		// We may need to acquire an open file handle to read from if
+		// copy up has occurred. Otherwise we risk reading from the
+		// wrong source.
+		f.upperMu.Lock()
+		if f.upper == nil {
+			var err error
+			f.upper, err = overlayFile(ctx, o.upper, file.Flags())
+			if err != nil {
+				f.upperMu.Unlock()
+				log.Warningf("failed to acquire handle with flags %v: %v", file.Flags(), err)
+				return 0, syserror.EIO
+			}
+		}
+		f.upperMu.Unlock()
+		return f.upper.FileOperations.Read(ctx, f.upper, dst, offset)
+	}
+	return f.lower.FileOperations.Read(ctx, f.lower, dst, offset)
+}
+
+// Write implements FileOperations.Write.
+func (f *overlayFileOperations) Write(ctx context.Context, file *File, src usermem.IOSequence, offset int64) (int64, error) {
+	// f.upper must be non-nil. See inode_overlay.go:overlayGetFile, where the
+	// file is copied up and opened in the upper filesystem if FileFlags.Write.
+	// Write cannot be called if !FileFlags.Write, see FileOperations.Write.
+	return f.upper.FileOperations.Write(ctx, f.upper, src, offset)
+}
+
+// Fsync implements FileOperations.Fsync.
+func (f *overlayFileOperations) Fsync(ctx context.Context, file *File, start, end int64, syncType SyncType) error {
+	var err error
+	f.upperMu.Lock()
+	if f.upper != nil {
+		err = f.upper.FileOperations.Fsync(ctx, f.upper, start, end, syncType)
+	}
+	f.upperMu.Unlock()
+	if f.lower != nil {
+		// N.B. Fsync on the lower filesystem can cause writes of file
+		// attributes (i.e. access time) despite the fact that we must
+		// treat the lower filesystem as read-only.
+		//
+		// This matches the semantics of fsync(2) in Linux overlayfs.
+		err = f.lower.FileOperations.Fsync(ctx, f.lower, start, end, syncType)
+	}
+	return err
+}
+
+// Flush implements FileOperations.Flush.
+func (f *overlayFileOperations) Flush(ctx context.Context, file *File) error {
+	// Flush whatever handles we have.
+	var err error
+	f.upperMu.Lock()
+	if f.upper != nil {
+		err = f.upper.FileOperations.Flush(ctx, f.upper)
+	}
+	f.upperMu.Unlock()
+	if f.lower != nil {
+		err = f.lower.FileOperations.Flush(ctx, f.lower)
+	}
+	return err
+}
+
+// ConfigureMMap implements FileOperations.ConfigureMMap.
+func (*overlayFileOperations) ConfigureMMap(ctx context.Context, file *File, opts *memmap.MMapOpts) error {
+	o := file.Dirent.Inode.overlay
+
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	if !o.isMappableLocked() {
+		return syserror.ENODEV
+	}
+	// FIXME: This is a copy/paste of fsutil.GenericConfigureMMap,
+	// which we can't use because the overlay implementation is in package fs,
+	// so depending on fs/fsutil would create a circular dependency. Move
+	// overlay to fs/overlay.
+	opts.Mappable = o
+	opts.MappingIdentity = file
+	file.IncRef()
+	return nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl and always returns ENOTTY.
+func (*overlayFileOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return 0, syserror.ENOTTY
+}
+
+// readdirEntries returns a sorted map of directory entries from the
+// upper and/or lower filesystem.
+func readdirEntries(ctx context.Context, o *overlayEntry) (*SortedDentryMap, error) {
+	// Assert that there is at least one upper or lower entry.
+	if o.upper == nil && o.lower == nil {
+		panic("invalid overlayEntry, needs at least one Inode")
+	}
+	entries := make(map[string]DentAttr)
+
+	// Try the upper filesystem first.
+	if o.upper != nil {
+		var err error
+		entries, err = readdirOne(ctx, NewTransientDirent(o.upper))
+		if err != nil {
+			return nil, err
+		}
+	}
+
+	// Try the lower filesystem next.
+	if o.lower != nil {
+		lowerEntries, err := readdirOne(ctx, NewTransientDirent(o.lower))
+		if err != nil {
+			return nil, err
+		}
+		for name, entry := range lowerEntries {
+			// Skip this name if it is a negative entry in the
+			// upper or there exists a whiteout for it.
+			if o.upper != nil {
+				if overlayHasWhiteout(o.upper, name) {
+					continue
+				}
+			}
+			// Prefer the entries from the upper filesystem
+			// when names overlap.
+			if _, ok := entries[name]; !ok {
+				entries[name] = entry
+			}
+		}
+	}
+
+	// Sort and return the entries.
+	return NewSortedDentryMap(entries), nil
+}
+
+// readdirOne reads all of the directory entries from d.
+func readdirOne(ctx context.Context, d *Dirent) (map[string]DentAttr, error) {
+	dir, err := d.Inode.GetFile(ctx, d, FileFlags{Read: true})
+	if err != nil {
+		return nil, err
+	}
+	defer dir.DecRef()
+
+	// Use a stub serializer to read the entries into memory.
+	stubSerializer := &CollectEntriesSerializer{}
+	if err := dir.Readdir(ctx, stubSerializer); err != nil {
+		return nil, err
+	}
+	// The "." and ".." entries are from the overlay Inode's Dirent, not the stub.
+	delete(stubSerializer.Entries, ".")
+	delete(stubSerializer.Entries, "..")
+	return stubSerializer.Entries, nil
+}
diff --git a/pkg/sentry/fs/file_overlay_test.go b/pkg/sentry/fs/file_overlay_test.go
new file mode 100644
index 000000000..407ba8562
--- /dev/null
+++ b/pkg/sentry/fs/file_overlay_test.go
@@ -0,0 +1,137 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs_test
+
+import (
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+func TestReaddir(t *testing.T) {
+	ctx := contexttest.Context(t)
+	ctx = &rootContext{
+		Context: ctx,
+		root:    fs.NewDirent(newTestRamfsDir(ctx, nil, nil), "root"),
+	}
+	for _, test := range []struct {
+		// Test description.
+		desc string
+
+		// Lookup parameters.
+		dir *fs.Inode
+
+		// Want from lookup.
+		err   error
+		names []string
+	}{
+		{
+			desc: "no upper, lower has entries",
+			dir: fs.NewTestOverlayDir(ctx,
+				nil, /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "a"},
+					{name: "b"},
+				}, nil), /* lower */
+			),
+			names: []string{".", "..", "a", "b"},
+		},
+		{
+			desc: "upper has entries, no lower",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "a"},
+					{name: "b"},
+				}, nil), /* upper */
+				nil, /* lower */
+			),
+			names: []string{".", "..", "a", "b"},
+		},
+		{
+			desc: "upper and lower, entries combine",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "a"},
+				}, nil), /* lower */
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "b"},
+				}, nil), /* lower */
+			),
+			names: []string{".", "..", "a", "b"},
+		},
+		{
+			desc: "upper and lower, entries combine, none are masked",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "a"},
+				}, []string{"b"}), /* lower */
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "c"},
+				}, nil), /* lower */
+			),
+			names: []string{".", "..", "a", "c"},
+		},
+		{
+			desc: "upper and lower, entries combine, upper masks some of lower",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "a"},
+				}, []string{"b"}), /* lower */
+				newTestRamfsDir(ctx, []dirContent{
+					{name: "b"}, /* will be masked */
+					{name: "c"},
+				}, nil), /* lower */
+			),
+			names: []string{".", "..", "a", "c"},
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			openDir, err := test.dir.GetFile(ctx, fs.NewDirent(test.dir, "stub"), fs.FileFlags{Read: true})
+			if err != nil {
+				t.Fatalf("GetFile got error %v, want nil", err)
+			}
+			stubSerializer := &fs.CollectEntriesSerializer{}
+			err = openDir.Readdir(ctx, stubSerializer)
+			if err != test.err {
+				t.Fatalf("Readdir got error %v, want nil", err)
+			}
+			if err != nil {
+				return
+			}
+			if !reflect.DeepEqual(stubSerializer.Order, test.names) {
+				t.Errorf("Readdir got names %v, want %v", stubSerializer.Order, test.names)
+			}
+		})
+	}
+}
+
+type rootContext struct {
+	context.Context
+	root *fs.Dirent
+}
+
+// Value implements context.Context.
+func (r *rootContext) Value(key interface{}) interface{} {
+	switch key {
+	case fs.CtxRoot:
+		r.root.IncRef()
+		return r.root
+	default:
+		return r.Context.Value(key)
+	}
+}
diff --git a/pkg/sentry/fs/file_state.go b/pkg/sentry/fs/file_state.go
new file mode 100644
index 000000000..341cbda0b
--- /dev/null
+++ b/pkg/sentry/fs/file_state.go
@@ -0,0 +1,30 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// afterLoad is invoked by stateify.
+func (f *File) afterLoad() {
+	f.mu.Init()
+}
+
+// saveFlags is invoked by stateify.
+func (f *File) saveFlags() FileFlags {
+	return f.flags.Load().(FileFlags)
+}
+
+// loadFlags is invoked by stateify.
+func (f *File) loadFlags(flags FileFlags) {
+	f.flags.Store(flags)
+}
diff --git a/pkg/sentry/fs/file_test.go b/pkg/sentry/fs/file_test.go
new file mode 100644
index 000000000..18aee7101
--- /dev/null
+++ b/pkg/sentry/fs/file_test.go
@@ -0,0 +1,24 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import "io"
+
+var (
+	_ = io.Reader(&FileReader{})
+	_ = io.ReaderAt(&FileReader{})
+	_ = io.Writer(&FileWriter{})
+	_ = io.WriterAt(&FileWriter{})
+)
diff --git a/pkg/sentry/fs/filesystems.go b/pkg/sentry/fs/filesystems.go
new file mode 100644
index 000000000..7cd76dfe9
--- /dev/null
+++ b/pkg/sentry/fs/filesystems.go
@@ -0,0 +1,162 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// FilesystemFlags matches include/linux/fs.h:file_system_type.fs_flags.
+type FilesystemFlags int
+
+const (
+	// FilesystemRequiresDev indicates that the file system requires a device name
+	// on mount. It is used to construct the output of /proc/filesystems.
+	FilesystemRequiresDev FilesystemFlags = 1
+
+	// Currently other flags are not used, but can be pulled in from
+	// include/linux/fs.h:file_system_type as needed.
+)
+
+// Filesystem is a mountable file system.
+type Filesystem interface {
+	// Name is the unique identifier of the file system. It corresponds to the
+	// filesystemtype argument of sys_mount and will appear in the output of
+	// /proc/filesystems.
+	Name() string
+
+	// Flags indicate common properties of the file system.
+	Flags() FilesystemFlags
+
+	// Mount generates a mountable Inode backed by device and configured
+	// using file system independent flags and file system dependent
+	// data options.
+	Mount(ctx context.Context, device string, flags MountSourceFlags, data string) (*Inode, error)
+
+	// AllowUserMount determines whether mount(2) is allowed to mount a
+	// file system of this type.
+	AllowUserMount() bool
+}
+
+// filesystems is the global set of registered file systems. It does not need
+// to be saved. Packages registering and unregistering file systems must do so
+// before calling save/restore methods.
+var filesystems = struct {
+	// mu protects registered below.
+	mu sync.Mutex
+
+	// registered is a set of registered Filesystems.
+	registered map[string]Filesystem
+}{
+	registered: make(map[string]Filesystem),
+}
+
+// RegisterFilesystem registers a new file system that is visible to mount and
+// the /proc/filesystems list. Packages implementing Filesystem should call
+// RegisterFilesystem in init().
+func RegisterFilesystem(f Filesystem) {
+	filesystems.mu.Lock()
+	defer filesystems.mu.Unlock()
+
+	if _, ok := filesystems.registered[f.Name()]; ok {
+		panic(fmt.Sprintf("filesystem already registered at %q", f.Name()))
+	}
+	filesystems.registered[f.Name()] = f
+}
+
+// UnregisterFilesystem removes a file system from the global set. To keep the
+// file system set compatible with save/restore, UnregisterFilesystem must be
+// called before save/restore methods.
+//
+// For instance, packages may unregister their file system after it is mounted.
+// This makes sense for pseudo file systems that should not be visible or
+// mountable. See whitelistfs in fs/host/fs.go for one example.
+func UnregisterFilesystem(name string) {
+	filesystems.mu.Lock()
+	defer filesystems.mu.Unlock()
+
+	delete(filesystems.registered, name)
+}
+
+// FindFilesystem returns a Filesystem registered at name or (nil, false) if name
+// is not a file system type that can be found in /proc/filesystems.
+func FindFilesystem(name string) (Filesystem, bool) {
+	filesystems.mu.Lock()
+	defer filesystems.mu.Unlock()
+
+	f, ok := filesystems.registered[name]
+	return f, ok
+}
+
+// GetFilesystems returns the set of registered filesystems in a consistent order.
+func GetFilesystems() []Filesystem {
+	filesystems.mu.Lock()
+	defer filesystems.mu.Unlock()
+
+	var ss []Filesystem
+	for _, s := range filesystems.registered {
+		ss = append(ss, s)
+	}
+	sort.Slice(ss, func(i, j int) bool { return ss[i].Name() < ss[j].Name() })
+	return ss
+}
+
+// MountSourceFlags represents all mount option flags as a struct.
+type MountSourceFlags struct {
+	// ReadOnly corresponds to mount(2)'s "MS_RDONLY" and indicates that
+	// the filesystem should be mounted read-only.
+	ReadOnly bool
+
+	// NoAtime corresponds to mount(2)'s "MS_NOATIME" and indicates that
+	// the filesystem should not update access time in-place.
+	NoAtime bool
+
+	// ForcePageCache causes all filesystem I/O operations to use the page
+	// cache, even when the platform supports direct mapped I/O. This
+	// doesn't correspond to any Linux mount options.
+	ForcePageCache bool
+}
+
+// GenericMountSourceOptions splits a string containing comma separated tokens of the
+// format 'key=value' or 'key' into a map of keys and values. For example:
+//
+// data = "key0=value0,key1,key2=value2" -> map{'key0':'value0','key1':'','key2':'value2'}
+//
+// If data contains duplicate keys, then the last token wins.
+func GenericMountSourceOptions(data string) map[string]string {
+	options := make(map[string]string)
+	if len(data) == 0 {
+		// Don't return a nil map, callers might not be expecting that.
+		return options
+	}
+
+	// Parse options and skip empty ones.
+	for _, opt := range strings.Split(data, ",") {
+		if len(opt) > 0 {
+			res := strings.SplitN(opt, "=", 2)
+			if len(res) == 2 {
+				options[res[0]] = res[1]
+			} else {
+				options[opt] = ""
+			}
+		}
+	}
+	return options
+}
diff --git a/pkg/sentry/fs/filetest/BUILD b/pkg/sentry/fs/filetest/BUILD
new file mode 100644
index 000000000..51a390d77
--- /dev/null
+++ b/pkg/sentry/fs/filetest/BUILD
@@ -0,0 +1,35 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "filetest_state",
+    srcs = [
+        "filetest.go",
+    ],
+    out = "filetest_state.go",
+    package = "filetest",
+)
+
+go_library(
+    name = "filetest",
+    testonly = 1,
+    srcs = [
+        "filetest.go",
+        "filetest_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/filetest",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/refs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/fs/filetest/filetest.go b/pkg/sentry/fs/filetest/filetest.go
new file mode 100644
index 000000000..1831aa82f
--- /dev/null
+++ b/pkg/sentry/fs/filetest/filetest.go
@@ -0,0 +1,59 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package filetest provides a test implementation of an fs.File.
+package filetest
+
+import (
+	"fmt"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// TestFileOperations is an implementation of the File interface. It provides all
+// required methods.
+type TestFileOperations struct {
+	fsutil.NoopRelease   `state:"nosave"`
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	fsutil.NoIoctl       `state:"nosave"`
+	waiter.AlwaysReady   `state:"nosave"`
+}
+
+// NewTestFile creates and initializes a new test file.
+func NewTestFile(tb testing.TB) *fs.File {
+	ctx := contexttest.Context(tb)
+	dirent := fs.NewDirent(anon.NewInode(ctx), "test")
+	return fs.NewFile(ctx, dirent, fs.FileFlags{}, &TestFileOperations{})
+}
+
+// Read just fails the request.
+func (*TestFileOperations) Read(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, fmt.Errorf("Readv not implemented")
+}
+
+// Write just fails the request.
+func (*TestFileOperations) Write(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, fmt.Errorf("Writev not implemented")
+}
diff --git a/pkg/sentry/fs/flags.go b/pkg/sentry/fs/flags.go
new file mode 100644
index 000000000..dfa6a3d62
--- /dev/null
+++ b/pkg/sentry/fs/flags.go
@@ -0,0 +1,67 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// FileFlags encodes file flags.
+type FileFlags struct {
+	// Direct indicates that I/O should be done directly.
+	Direct bool
+
+	// NonBlocking indicates that I/O should not block.
+	NonBlocking bool
+
+	// Sync indicates that any writes should be synchronous.
+	Sync bool
+
+	// Append indicates this file is append only.
+	Append bool
+
+	// Read indicates this file is readable.
+	Read bool
+
+	// Write indicates this file is writeable.
+	Write bool
+
+	// Pread indicates this file is readable at an arbitrary offset.
+	Pread bool
+
+	// Pwrite indicates this file is writable at an arbitrary offset.
+	Pwrite bool
+
+	// Directory indicates that this file must be a directory.
+	Directory bool
+}
+
+// SettableFileFlags is a subset of FileFlags above that can be changed
+// via fcntl(2) using the F_SETFL command.
+type SettableFileFlags struct {
+	// Direct indicates that I/O should be done directly.
+	Direct bool
+
+	// NonBlocking indicates that I/O should not block.
+	NonBlocking bool
+
+	// Append indicates this file is append only.
+	Append bool
+}
+
+// Settable returns the subset of f that are settable.
+func (f FileFlags) Settable() SettableFileFlags {
+	return SettableFileFlags{
+		Direct:      f.Direct,
+		NonBlocking: f.NonBlocking,
+		Append:      f.Append,
+	}
+}
diff --git a/pkg/sentry/fs/fs.go b/pkg/sentry/fs/fs.go
new file mode 100644
index 000000000..f54f767d3
--- /dev/null
+++ b/pkg/sentry/fs/fs.go
@@ -0,0 +1,88 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package fs implements a virtual filesystem layer.
+//
+// Specific filesystem implementations must implement the InodeOperations
+// interface (inode.go).
+//
+// The MountNamespace (mounts.go) is used to create a collection of mounts in
+// a filesystem rooted at a given Inode.
+//
+// MountSources (mount.go) form a tree, with each mount holding pointers to its
+// parent and children.
+//
+// Dirents (dirents.go) wrap Inodes in a caching layer.
+//
+// When multiple locks are to be held at the same time, they should be acquired
+// in the following order.
+//
+// Either:
+//   File.mu
+//     Locks in FileOperations implementations
+//       goto Dirent-Locks
+//
+// Or:
+//   MountNamespace.mu
+//     goto Dirent-Locks
+//
+// Dirent-Locks:
+//   renameMu
+//     Dirent.dirMu
+//       Dirent.mu
+//         DirentCache.mu
+//         Locks in InodeOperations implementations or overlayEntry
+//         Inode.Watches.mu (see `Inotify` for other lock ordering)
+//	   MountSource.mu
+//
+// If multiple Dirent or MountSource locks must be taken, locks in the parent must be
+// taken before locks in their children.
+//
+// If locks must be taken on multiple unrelated Dirents, renameMu must be taken
+// first. See lockForRename.
+package fs
+
+import (
+	"sync"
+)
+
+// work is a sync.WaitGroup that can be used to queue asynchronous operations
+// via Do. Callers can use Barrier to ensure no operations are outstanding.
+var work sync.WaitGroup
+
+// AsyncBarrier waits for all outstanding asynchronous work to complete.
+func AsyncBarrier() {
+	work.Wait()
+}
+
+// Async executes a function asynchronously.
+func Async(f func()) {
+	work.Add(1)
+	go func() { // S/R-SAFE: Barrier must be called.
+		defer work.Done() // Ensure Done in case of panic.
+		f()
+	}()
+}
+
+// ErrSaveRejection indicates a failed save due to unsupported file system state
+// such as dangling open fd, etc.
+type ErrSaveRejection struct {
+	// Err is the wrapped error.
+	Err error
+}
+
+// Error returns a sensible description of the save rejection error.
+func (e ErrSaveRejection) Error() string {
+	return "save rejected due to unsupported file system state: " + e.Err.Error()
+}
diff --git a/pkg/sentry/fs/fsutil/BUILD b/pkg/sentry/fs/fsutil/BUILD
new file mode 100644
index 000000000..4fa6395f7
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/BUILD
@@ -0,0 +1,149 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "fsutil_state",
+    srcs = [
+        "dirty_set_impl.go",
+        "file.go",
+        "file_range_set_impl.go",
+        "frame_ref_set_impl.go",
+        "handle.go",
+        "host_file_mapper.go",
+        "host_file_mapper_state.go",
+        "inode.go",
+        "inode_cached.go",
+    ],
+    out = "fsutil_state.go",
+    package = "fsutil",
+)
+
+go_template_instance(
+    name = "dirty_set_impl",
+    out = "dirty_set_impl.go",
+    imports = {
+        "memmap": "gvisor.googlesource.com/gvisor/pkg/sentry/memmap",
+        "platform": "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    },
+    package = "fsutil",
+    prefix = "Dirty",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "memmap.MappableRange",
+        "Value": "DirtyInfo",
+        "Functions": "dirtySetFunctions",
+    },
+)
+
+go_template_instance(
+    name = "frame_ref_set_impl",
+    out = "frame_ref_set_impl.go",
+    imports = {
+        "platform": "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    },
+    package = "fsutil",
+    prefix = "frameRef",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "platform.FileRange",
+        "Value": "uint64",
+        "Functions": "frameRefSetFunctions",
+    },
+)
+
+go_template_instance(
+    name = "file_range_set_impl",
+    out = "file_range_set_impl.go",
+    imports = {
+        "memmap": "gvisor.googlesource.com/gvisor/pkg/sentry/memmap",
+        "platform": "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    },
+    package = "fsutil",
+    prefix = "FileRange",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "memmap.MappableRange",
+        "Value": "uint64",
+        "Functions": "fileRangeSetFunctions",
+    },
+)
+
+go_library(
+    name = "fsutil",
+    srcs = [
+        "dirty_set.go",
+        "dirty_set_impl.go",
+        "file.go",
+        "file_range_set.go",
+        "file_range_set_impl.go",
+        "frame_ref_set.go",
+        "frame_ref_set_impl.go",
+        "fsutil.go",
+        "fsutil_state.go",
+        "handle.go",
+        "host_file_mapper.go",
+        "host_file_mapper_state.go",
+        "host_file_mapper_unsafe.go",
+        "inode.go",
+        "inode_cached.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "fsutil_x_test",
+    size = "small",
+    srcs = ["handle_test.go"],
+    deps = [
+        ":fsutil",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/ramfs/test",
+        "//pkg/sentry/usermem",
+    ],
+)
+
+go_test(
+    name = "fsutil_test",
+    size = "small",
+    srcs = [
+        "dirty_set_test.go",
+        "inode_cached_test.go",
+    ],
+    embed = [":fsutil"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/fsutil/README.md b/pkg/sentry/fs/fsutil/README.md
new file mode 100644
index 000000000..d3780e9fa
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/README.md
@@ -0,0 +1,207 @@
+This package provides utilities for implementing virtual filesystem objects.
+
+[TOC]
+
+## Page cache
+
+`CachingInodeOperations` implements a page cache for files that cannot use the
+host page cache. Normally these are files that store their data in a remote
+filesystem. This also applies to files that are accessed on a platform that does
+not support directly memory mapping host file descriptors (e.g. the ptrace
+platform).
+
+An `CachingInodeOperations` buffers regions of a single file into memory. It is
+owned by an `fs.Inode`, the in-memory representation of a file (all open file
+descriptors are backed by an `fs.Inode`). The `fs.Inode` provides operations for
+reading memory into an `CachingInodeOperations`, to represent the contents of
+the file in-memory, and for writing memory out, to relieve memory pressure on
+the kernel and to synchronize in-memory changes to filesystems.
+
+An `CachingInodeOperations` enables readable and/or writable memory access to
+file content. Files can be mapped shared or private, see mmap(2). When a file is
+mapped shared, changes to the file via write(2) and truncate(2) are reflected in
+the shared memory region. Conversely, when the shared memory region is modified,
+changes to the file are visible via read(2). Multiple shared mappings of the
+same file are coherent with each other. This is consistent with Linux.
+
+When a file is mapped private, updates to the mapped memory are not visible to
+other memory mappings. Updates to the mapped memory are also not reflected in
+the file content as seen by read(2). If the file is changed after a private
+mapping is created, for instance by write(2), the change to the file may or may
+not be reflected in the private mapping. This is consistent with Linux.
+
+An `CachingInodeOperations` keeps track of ranges of memory that were modified
+(or "dirtied"). When the file is explicitly synced via fsync(2), only the dirty
+ranges are written out to the filesystem. Any error returned indicates a failure
+to write all dirty memory of an `CachingInodeOperations` to the filesystem. In
+this case the filesystem may be in an inconsistent state. The same operation can
+be performed on the shared memory itself using msync(2). If neither fsync(2) nor
+msync(2) is performed, then the dirty memory is written out in accordance with
+the `CachingInodeOperations` eviction strategy (see below) and there is no
+guarantee that memory will be written out successfully in full.
+
+### Memory allocation and eviction
+
+An `CachingInodeOperations` implements the following allocation and eviction
+strategy:
+
+-   Memory is allocated and brought up to date with the contents of a file when
+    a region of mapped memory is accessed (or "faulted on").
+
+-   Dirty memory is written out to filesystems when an fsync(2) or msync(2)
+    operation is performed on a memory mapped file, for all memory mapped files
+    when saved, and/or when there are no longer any memory mappings of a range
+    of a file, see munmap(2). As the latter implies, in the absence of a panic
+    or SIGKILL, dirty memory is written out for all memory mapped files when an
+    application exits.
+
+-   Memory is freed when there are no longer any memory mappings of a range of a
+    file (e.g. when an application exits). This behavior is consistent with
+    Linux for shared memory that has been locked via mlock(2).
+
+Notably, memory is not allocated for read(2) or write(2) operations. This means
+that reads and writes to the file are only accelerated by an
+`CachingInodeOperations` if the file being read or written has been memory
+mapped *and* if the shared memory has been accessed at the region being read or
+written. This diverges from Linux which buffers memory into a page cache on
+read(2) proactively (i.e. readahead) and delays writing it out to filesystems on
+write(2) (i.e. writeback). The absence of these optimizations is not visible to
+applications beyond less than optimal performance when repeatedly reading and/or
+writing to same region of a file. See [Future Work](#future-work) for plans to
+implement these optimizations.
+
+Additionally, memory held by `CachingInodeOperationss` is currently unbounded in
+size. An `CachingInodeOperations` does not write out dirty memory and free it
+under system memory pressure. This can cause pathological memory usage.
+
+When memory is written back, an `CachingInodeOperations` may write regions of
+shared memory that were never modified. This is due to the strategy of
+minimizing page faults (see below) and handling only a subset of memory write
+faults. In the absence of an application or sentry crash, it is guaranteed that
+if a region of shared memory was written to, it is written back to a filesystem.
+
+### Life of a shared memory mapping
+
+A file is memory mapped via mmap(2). For example, if `A` is an address, an
+application may execute:
+
+```
+mmap(A, 0x1000, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
+```
+
+This creates a shared mapping of fd that reflects 4k of the contents of fd
+starting at offset 0, accessible at address `A`. This in turn creates a virtual
+memory area region ("vma") which indicates that [`A`, `A`+0x1000) is now a valid
+address range for this application to access.
+
+At this point, memory has not been allocated in the file's
+`CachingInodeOperations`. It is also the case that the address range [`A`,
+`A`+0x1000) has not been mapped on the host on behalf of the application. If the
+application then tries to modify 8 bytes of the shared memory:
+
+```
+char buffer[] = "aaaaaaaa";
+memcpy(A, buffer, 8);
+```
+
+The host then sends a `SIGSEGV` to the sentry because the address range [`A`,
+`A`+8) is not mapped on the host. The `SIGSEGV` indicates that the memory was
+accessed writable. The sentry looks up the vma associated with [`A`, `A`+8),
+finds the file that was mapped and its `CachingInodeOperations`. It then calls
+`CachingInodeOperations.MapInto` which allocates memory to back [`A`, `A`+8). It
+may choose to allocate more memory (i.e. do "readahead") to minimize subsequent
+faults.
+
+Memory that is allocated comes from a host tmpfs file (see `filemem.FileMem`).
+The host tmpfs file memory is brought up to date with the contents of the mapped
+file on its filesystem. The region of the host tmpfs file that reflects the
+mapped file is then mapped into the host address space of the application so
+that subsequent memory accesses do not repeatedly generate a `SIGSEGV`.
+
+The range that was allocated, including any extra memory allocation to minimize
+faults, is marked dirty due to the write fault. This overcounts dirty memory if
+the extra memory allocated is never modified.
+
+To make the scenario more interesting, imagine that this application spawns
+another process and maps the same file in the exact same way:
+
+```
+mmap(A, 0x1000, PROT_READ|PROT_WRITE, MAP_SHARED, fd, 0);
+```
+
+Imagine that this process then tries to modify the file again but with only 4
+bytes:
+
+```
+char buffer[] = "bbbb";
+memcpy(A, buffer, 4);
+```
+
+Since the first process has already mapped and accessed the same region of the
+file writable, `CachingInodeOperations.MapInto` is called but re-maps the memory
+that has already been allocated (because the host mapping can be invalidated at
+any time) rather than allocating new memory. The address range [`A`, `A`+0x1000)
+reflects the same cached view of the file as the first process sees. For
+example, reading 8 bytes from the file from either process via read(2) starting
+at offset 0 returns a consistent "bbbbaaaa".
+
+When this process no longer needs the shared memory, it may do:
+
+```
+munmap(A, 0x1000);
+```
+
+At this point, the modified memory cached by the `CachingInodeOperations` is not
+written back to the file because it is still in use by the first process that
+mapped it. When the first process also does:
+
+```
+munmap(A, 0x1000);
+```
+
+Then the last memory mapping of the file at the range [0, 0x1000) is gone. The
+file's `CachingInodeOperations` then starts writing back memory marked dirty to
+the file on its filesystem. Once writing completes, regardless of whether it was
+successful, the `CachingInodeOperations` frees the memory cached at the range
+[0, 0x1000).
+
+Subsequent read(2) or write(2) operations on the file go directly to the
+filesystem since there no longer exists memory for it in its
+`CachingInodeOperations`.
+
+## Future Work
+
+### Page cache
+
+The sentry does not yet implement the readahead and writeback optimizations for
+read(2) and write(2) respectively. To do so, on read(2) and/or write(2) the
+sentry must ensure that memory is allocated in a page cache to read or write
+into. However, the sentry cannot boundlessly allocate memory. If it did, the
+host would eventually OOM-kill the sentry+application process. This means that
+the sentry must implement a page cache memory allocation strategy that is
+bounded by a global user or container imposed limit. When this limit is
+approached, the sentry must decide from which page cache memory should be freed
+so that it can allocate more memory. If it makes a poor decision, the sentry may
+end up freeing and re-allocating memory to back regions of files that are
+frequently used, nullifying the optimization (and in some cases causing worse
+performance due to the overhead of memory allocation and general management).
+This is a form of "cache thrashing".
+
+In Linux, much research has been done to select and implement a lightweight but
+optimal page cache eviction algorithm. Linux makes use of hardware page bits to
+keep track of whether memory has been accessed. The sentry does not have direct
+access to hardware. Implementing a similarly lightweight and optimal page cache
+eviction algorithm will need to either introduce a kernel interface to obtain
+these page bits or find a suitable alternative proxy for access events.
+
+In Linux, readahead happens by default but is not always ideal. For instance,
+for files that are not read sequentially, it would be more ideal to simply read
+from only those regions of the file rather than to optimistically cache some
+number of bytes ahead of the read (up to 2MB in Linux) if the bytes cached won't
+be accessed. Linux implements the fadvise64(2) system call for applications to
+specify that a range of a file will not be accessed sequentially. The advice bit
+FADV_RANDOM turns off the readahead optimization for the given range in the
+given file. However fadvise64 is rarely used by applications so Linux implements
+a readahead backoff strategy if reads are not sequential. To ensure that
+application performance is not degraded, the sentry must implement a similar
+backoff strategy.
diff --git a/pkg/sentry/fs/fsutil/dirty_set.go b/pkg/sentry/fs/fsutil/dirty_set.go
new file mode 100644
index 000000000..9c6c98542
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/dirty_set.go
@@ -0,0 +1,213 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// DirtySet maps offsets into a memmap.Mappable to DirtyInfo. It is used to
+// implement Mappables that cache data from another source.
+//
+// type DirtySet <generated by go_generics>
+
+// DirtyInfo is the value type of DirtySet, and represents information about a
+// Mappable offset that is dirty (the cached data for that offset is newer than
+// its source).
+type DirtyInfo struct {
+	// Keep is true if the represented offset is concurrently writable, such
+	// that writing the data for that offset back to the source does not
+	// guarantee that the offset is clean (since it may be concurrently
+	// rewritten after the writeback).
+	Keep bool
+}
+
+// dirtySetFunctions implements segment.Functions for DirtySet.
+type dirtySetFunctions struct{}
+
+// MinKey implements segment.Functions.MinKey.
+func (dirtySetFunctions) MinKey() uint64 {
+	return 0
+}
+
+// MaxKey implements segment.Functions.MaxKey.
+func (dirtySetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+// ClearValue implements segment.Functions.ClearValue.
+func (dirtySetFunctions) ClearValue(val *DirtyInfo) {
+}
+
+// Merge implements segment.Functions.Merge.
+func (dirtySetFunctions) Merge(_ memmap.MappableRange, val1 DirtyInfo, _ memmap.MappableRange, val2 DirtyInfo) (DirtyInfo, bool) {
+	if val1 != val2 {
+		return DirtyInfo{}, false
+	}
+	return val1, true
+}
+
+// Split implements segment.Functions.Split.
+func (dirtySetFunctions) Split(_ memmap.MappableRange, val DirtyInfo, _ uint64) (DirtyInfo, DirtyInfo) {
+	return val, val
+}
+
+// MarkClean marks all offsets in mr as not dirty, except for those to which
+// KeepDirty has been applied.
+func (ds *DirtySet) MarkClean(mr memmap.MappableRange) {
+	seg := ds.LowerBoundSegment(mr.Start)
+	for seg.Ok() && seg.Start() < mr.End {
+		if seg.Value().Keep {
+			seg = seg.NextSegment()
+			continue
+		}
+		seg = ds.Isolate(seg, mr)
+		seg = ds.Remove(seg).NextSegment()
+	}
+}
+
+// KeepClean marks all offsets in mr as not dirty, even those that were
+// previously kept dirty by KeepDirty.
+func (ds *DirtySet) KeepClean(mr memmap.MappableRange) {
+	ds.RemoveRange(mr)
+}
+
+// MarkDirty marks all offsets in mr as dirty.
+func (ds *DirtySet) MarkDirty(mr memmap.MappableRange) {
+	ds.setDirty(mr, false)
+}
+
+// KeepDirty marks all offsets in mr as dirty and prevents them from being
+// marked as clean by MarkClean.
+func (ds *DirtySet) KeepDirty(mr memmap.MappableRange) {
+	ds.setDirty(mr, true)
+}
+
+func (ds *DirtySet) setDirty(mr memmap.MappableRange, keep bool) {
+	var changedAny bool
+	defer func() {
+		if changedAny {
+			ds.MergeRange(mr)
+		}
+	}()
+	seg, gap := ds.Find(mr.Start)
+	for {
+		switch {
+		case seg.Ok() && seg.Start() < mr.End:
+			if keep && !seg.Value().Keep {
+				changedAny = true
+				seg = ds.Isolate(seg, mr)
+				seg.ValuePtr().Keep = true
+			}
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok() && gap.Start() < mr.End:
+			changedAny = true
+			seg = ds.Insert(gap, gap.Range().Intersect(mr), DirtyInfo{keep})
+			seg, gap = seg.NextNonEmpty()
+
+		default:
+			return
+		}
+	}
+}
+
+// SyncDirty passes pages in the range mr that are stored in cache and
+// identified as dirty to writeAt, updating dirty to reflect successful writes.
+// If writeAt returns a successful partial write, SyncDirty will call it
+// repeatedly until all bytes have been written. max is the true size of the
+// cached object; offsets beyond max will not be passed to writeAt, even if
+// they are marked dirty.
+func SyncDirty(ctx context.Context, mr memmap.MappableRange, cache *FileRangeSet, dirty *DirtySet, max uint64, mem platform.File, writeAt func(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error)) error {
+	var changedDirty bool
+	defer func() {
+		if changedDirty {
+			dirty.MergeRange(mr)
+		}
+	}()
+	dseg := dirty.LowerBoundSegment(mr.Start)
+	for dseg.Ok() && dseg.Start() < mr.End {
+		var dr memmap.MappableRange
+		if dseg.Value().Keep {
+			dr = dseg.Range().Intersect(mr)
+		} else {
+			changedDirty = true
+			dseg = dirty.Isolate(dseg, mr)
+			dr = dseg.Range()
+		}
+		if err := syncDirtyRange(ctx, dr, cache, max, mem, writeAt); err != nil {
+			return err
+		}
+		if dseg.Value().Keep {
+			dseg = dseg.NextSegment()
+		} else {
+			dseg = dirty.Remove(dseg).NextSegment()
+		}
+	}
+	return nil
+}
+
+// SyncDirtyAll passes all pages stored in cache identified as dirty to
+// writeAt, updating dirty to reflect successful writes. If writeAt returns a
+// successful partial write, SyncDirtyAll will call it repeatedly until all
+// bytes have been written. max is the true size of the cached object; offsets
+// beyond max will not be passed to writeAt, even if they are marked dirty.
+func SyncDirtyAll(ctx context.Context, cache *FileRangeSet, dirty *DirtySet, max uint64, mem platform.File, writeAt func(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error)) error {
+	dseg := dirty.FirstSegment()
+	for dseg.Ok() {
+		if err := syncDirtyRange(ctx, dseg.Range(), cache, max, mem, writeAt); err != nil {
+			return err
+		}
+		if dseg.Value().Keep {
+			dseg = dseg.NextSegment()
+		} else {
+			dseg = dirty.Remove(dseg).NextSegment()
+		}
+	}
+	return nil
+}
+
+// Preconditions: mr must be page-aligned.
+func syncDirtyRange(ctx context.Context, mr memmap.MappableRange, cache *FileRangeSet, max uint64, mem platform.File, writeAt func(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error)) error {
+	for cseg := cache.LowerBoundSegment(mr.Start); cseg.Ok() && cseg.Start() < mr.End; cseg = cseg.NextSegment() {
+		wbr := cseg.Range().Intersect(mr)
+		if max < wbr.Start {
+			break
+		}
+		ims, err := mem.MapInternal(cseg.FileRangeOf(wbr), usermem.Read)
+		if err != nil {
+			return err
+		}
+		if max < wbr.End {
+			ims = ims.TakeFirst64(max - wbr.Start)
+		}
+		offset := wbr.Start
+		for !ims.IsEmpty() {
+			n, err := writeAt(ctx, ims, offset)
+			if err != nil {
+				return err
+			}
+			offset += n
+			ims = ims.DropFirst64(n)
+		}
+	}
+	return nil
+}
diff --git a/pkg/sentry/fs/fsutil/dirty_set_test.go b/pkg/sentry/fs/fsutil/dirty_set_test.go
new file mode 100644
index 000000000..f7693cb19
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/dirty_set_test.go
@@ -0,0 +1,38 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func TestDirtySet(t *testing.T) {
+	var set DirtySet
+	set.MarkDirty(memmap.MappableRange{0, 2 * usermem.PageSize})
+	set.KeepDirty(memmap.MappableRange{usermem.PageSize, 2 * usermem.PageSize})
+	set.MarkClean(memmap.MappableRange{0, 2 * usermem.PageSize})
+	want := &DirtySegmentDataSlices{
+		Start:  []uint64{usermem.PageSize},
+		End:    []uint64{2 * usermem.PageSize},
+		Values: []DirtyInfo{{Keep: true}},
+	}
+	if got := set.ExportSortedSlices(); !reflect.DeepEqual(got, want) {
+		t.Errorf("set:\n\tgot %v,\n\twant %v", got, want)
+	}
+}
diff --git a/pkg/sentry/fs/fsutil/file.go b/pkg/sentry/fs/fsutil/file.go
new file mode 100644
index 000000000..a7329f1c9
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/file.go
@@ -0,0 +1,267 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// NoopRelease implements FileOperations.Release for files that have no
+// resources to release.
+type NoopRelease struct{}
+
+// Release is a no-op.
+func (NoopRelease) Release() {}
+
+// SeekWithDirCursor is used to implement fs.FileOperations.Seek.  If dirCursor
+// is not nil and the seek was on a directory, the cursor will be updated.
+//
+// Currenly only seeking to 0 on a directory is supported.
+//
+// FIXME: Lift directory seeking limitations.
+func SeekWithDirCursor(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64, dirCursor *string) (int64, error) {
+	inode := file.Dirent.Inode
+	current := file.Offset()
+
+	// Does the Inode represents a non-seekable type?
+	if fs.IsPipe(inode.StableAttr) || fs.IsSocket(inode.StableAttr) {
+		return current, syserror.ESPIPE
+	}
+
+	// Does the Inode represent a character device?
+	if fs.IsCharDevice(inode.StableAttr) {
+		// Ignore seek requests.
+		//
+		// FIXME: This preserves existing
+		// behavior but is not universally correct.
+		return 0, nil
+	}
+
+	// Otherwise compute the new offset.
+	switch whence {
+	case fs.SeekSet:
+		switch inode.StableAttr.Type {
+		case fs.RegularFile, fs.SpecialFile, fs.BlockDevice:
+			if offset < 0 {
+				return current, syserror.EINVAL
+			}
+			return offset, nil
+		case fs.Directory, fs.SpecialDirectory:
+			if offset != 0 {
+				return current, syserror.EINVAL
+			}
+			// SEEK_SET to 0 moves the directory "cursor" to the beginning.
+			if dirCursor != nil {
+				*dirCursor = ""
+			}
+			return 0, nil
+		default:
+			return current, syserror.EINVAL
+		}
+	case fs.SeekCurrent:
+		switch inode.StableAttr.Type {
+		case fs.RegularFile, fs.SpecialFile, fs.BlockDevice:
+			if current+offset < 0 {
+				return current, syserror.EINVAL
+			}
+			return current + offset, nil
+		case fs.Directory, fs.SpecialDirectory:
+			if offset != 0 {
+				return current, syserror.EINVAL
+			}
+			return current, nil
+		default:
+			return current, syserror.EINVAL
+		}
+	case fs.SeekEnd:
+		switch inode.StableAttr.Type {
+		case fs.RegularFile, fs.BlockDevice:
+			// Allow the file to determine the end.
+			uattr, err := inode.UnstableAttr(ctx)
+			if err != nil {
+				return current, err
+			}
+			sz := uattr.Size
+			if sz+offset < 0 {
+				return current, syserror.EINVAL
+			}
+			return sz + offset, nil
+		// FIXME: This is not universally correct.
+		// Remove SpecialDirectory.
+		case fs.SpecialDirectory:
+			if offset != 0 {
+				return current, syserror.EINVAL
+			}
+			// SEEK_END to 0 moves the directory "cursor" to the end.
+			//
+			// FIXME: The ensures that after the seek,
+			// reading on the directory will get EOF. But it is not
+			// correct in general because the directory can grow in
+			// size; attempting to read those new entries will be
+			// futile (EOF will always be the result).
+			return fs.FileMaxOffset, nil
+		default:
+			return current, syserror.EINVAL
+		}
+	}
+
+	// Not a valid seek request.
+	return current, syserror.EINVAL
+}
+
+// GenericSeek implements FileOperations.Seek for files that use a generic
+// seek implementation.
+type GenericSeek struct{}
+
+// Seek implements fs.FileOperations.Seek.
+func (GenericSeek) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	return SeekWithDirCursor(ctx, file, whence, offset, nil)
+}
+
+// ZeroSeek implements FileOperations.Seek for files that maintain a constant
+// zero-value offset and require a no-op Seek.
+type ZeroSeek struct{}
+
+// Seek implements FileOperations.Seek.
+func (ZeroSeek) Seek(context.Context, *fs.File, fs.SeekWhence, int64) (int64, error) {
+	return 0, nil
+}
+
+// PipeSeek implements FileOperations.Seek and can be used for files that behave
+// like pipes (seeking is not supported).
+type PipeSeek struct{}
+
+// Seek implements FileOperations.Seek.
+func (PipeSeek) Seek(context.Context, *fs.File, fs.SeekWhence, int64) (int64, error) {
+	return 0, syserror.ESPIPE
+}
+
+// NotDirReaddir implements FileOperations.Readdir for non-directories.
+type NotDirReaddir struct{}
+
+// Readdir implements FileOperations.NotDirReaddir.
+func (NotDirReaddir) Readdir(context.Context, *fs.File, fs.DentrySerializer) (int64, error) {
+	return 0, syserror.ENOTDIR
+}
+
+// NoFsync implements FileOperations.Fsync for files that don't support syncing.
+type NoFsync struct{}
+
+// Fsync implements FileOperations.Fsync.
+func (NoFsync) Fsync(context.Context, *fs.File, int64, int64, fs.SyncType) error {
+	return syserror.EINVAL
+}
+
+// NoopFsync implements FileOperations.Fsync for files that don't need to synced.
+type NoopFsync struct{}
+
+// Fsync implements FileOperations.Fsync.
+func (NoopFsync) Fsync(context.Context, *fs.File, int64, int64, fs.SyncType) error {
+	return nil
+}
+
+// NoopFlush implements FileOperations.Flush as a no-op.
+type NoopFlush struct{}
+
+// Flush implements FileOperations.Flush.
+func (NoopFlush) Flush(context.Context, *fs.File) error {
+	return nil
+}
+
+// NoMMap implements fs.FileOperations.Mappable for files that cannot
+// be memory mapped.
+type NoMMap struct{}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (NoMMap) ConfigureMMap(context.Context, *fs.File, *memmap.MMapOpts) error {
+	return syserror.ENODEV
+}
+
+// GenericConfigureMMap implements fs.FileOperations.ConfigureMMap for most
+// filesystems that support memory mapping.
+func GenericConfigureMMap(file *fs.File, m memmap.Mappable, opts *memmap.MMapOpts) error {
+	opts.Mappable = m
+	opts.MappingIdentity = file
+	file.IncRef()
+	return nil
+}
+
+// NoIoctl implements fs.FileOperations.Ioctl for files that don't implement
+// the ioctl syscall.
+type NoIoctl struct{}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (NoIoctl) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return 0, syserror.ENOTTY
+}
+
+// DirFileOperations implements FileOperations for directories.
+type DirFileOperations struct {
+	waiter.AlwaysReady `state:"nosave"`
+	NoopRelease        `state:"nosave"`
+	GenericSeek        `state:"nosave"`
+	NoFsync            `state:"nosave"`
+	NoopFlush          `state:"nosave"`
+	NoMMap             `state:"nosave"`
+	NoIoctl            `state:"nosave"`
+
+	// dentryMap is a SortedDentryMap used to implement Readdir.
+	dentryMap *fs.SortedDentryMap
+
+	// dirCursor contains the name of the last directory entry that was
+	// serialized.
+	dirCursor string
+}
+
+// NewDirFileOperations returns a new DirFileOperations that will iterate the
+// given denty map.
+func NewDirFileOperations(dentries *fs.SortedDentryMap) *DirFileOperations {
+	return &DirFileOperations{
+		dentryMap: dentries,
+	}
+}
+
+// IterateDir implements DirIterator.IterateDir.
+func (dfo *DirFileOperations) IterateDir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	n, err := fs.GenericReaddir(dirCtx, dfo.dentryMap)
+	return offset + n, err
+}
+
+// Readdir implements FileOperations.Readdir.
+func (dfo *DirFileOperations) Readdir(ctx context.Context, file *fs.File, serializer fs.DentrySerializer) (int64, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+	dirCtx := &fs.DirCtx{
+		Serializer: serializer,
+		DirCursor:  &dfo.dirCursor,
+	}
+	return fs.DirentReaddir(ctx, file.Dirent, dfo, root, dirCtx, file.Offset())
+}
+
+// Read implements FileOperations.Read
+func (*DirFileOperations) Read(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// Write implements FileOperations.Write.
+func (*DirFileOperations) Write(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EISDIR
+}
diff --git a/pkg/sentry/fs/fsutil/file_range_set.go b/pkg/sentry/fs/fsutil/file_range_set.go
new file mode 100644
index 000000000..da6949ccb
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/file_range_set.go
@@ -0,0 +1,208 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"fmt"
+	"io"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// FileRangeSet maps offsets into a memmap.Mappable to offsets into a
+// platform.File. It is used to implement Mappables that store data in
+// sparsely-allocated memory.
+//
+// type FileRangeSet <generated by go_generics>
+
+// fileRangeSetFunctions implements segment.Functions for FileRangeSet.
+type fileRangeSetFunctions struct{}
+
+// MinKey implements segment.Functions.MinKey.
+func (fileRangeSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+// MaxKey implements segment.Functions.MaxKey.
+func (fileRangeSetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+// ClearValue implements segment.Functions.ClearValue.
+func (fileRangeSetFunctions) ClearValue(_ *uint64) {
+}
+
+// Merge implements segment.Functions.Merge.
+func (fileRangeSetFunctions) Merge(mr1 memmap.MappableRange, frstart1 uint64, _ memmap.MappableRange, frstart2 uint64) (uint64, bool) {
+	if frstart1+mr1.Length() != frstart2 {
+		return 0, false
+	}
+	return frstart1, true
+}
+
+// Split implements segment.Functions.Split.
+func (fileRangeSetFunctions) Split(mr memmap.MappableRange, frstart uint64, split uint64) (uint64, uint64) {
+	return frstart, frstart + (split - mr.Start)
+}
+
+// FileRange returns the FileRange mapped by seg.
+func (seg FileRangeIterator) FileRange() platform.FileRange {
+	return seg.FileRangeOf(seg.Range())
+}
+
+// FileRangeOf returns the FileRange mapped by mr.
+//
+// Preconditions: seg.Range().IsSupersetOf(mr). mr.Length() != 0.
+func (seg FileRangeIterator) FileRangeOf(mr memmap.MappableRange) platform.FileRange {
+	frstart := seg.Value() + (mr.Start - seg.Start())
+	return platform.FileRange{frstart, frstart + mr.Length()}
+}
+
+// Fill attempts to ensure that all memmap.Mappable offsets in required are
+// mapped to a platform.File offset, by allocating from mem with the given
+// memory usage kind and invoking readAt to store data into memory. (If readAt
+// returns a successful partial read, Fill will call it repeatedly until all
+// bytes have been read.) EOF is handled consistently with the requirements of
+// mmap(2): bytes after EOF on the same page are zeroed; pages after EOF are
+// invalid.
+//
+// Fill may read offsets outside of required, but will never read offsets
+// outside of optional. It returns a non-nil error if any error occurs, even
+// if the error only affects offsets in optional, but not in required.
+//
+// Preconditions: required.Length() > 0. optional.IsSupersetOf(required).
+// required and optional must be page-aligned.
+func (frs *FileRangeSet) Fill(ctx context.Context, required, optional memmap.MappableRange, mem platform.Memory, kind usage.MemoryKind, readAt func(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error)) error {
+	gap := frs.LowerBoundGap(required.Start)
+	for gap.Ok() && gap.Start() < required.End {
+		if gap.Range().Length() == 0 {
+			gap = gap.NextGap()
+			continue
+		}
+		gr := gap.Range().Intersect(optional)
+
+		// Read data into the gap.
+		fr, err := platform.AllocateAndFill(mem, gr.Length(), kind, safemem.ReaderFunc(func(dsts safemem.BlockSeq) (uint64, error) {
+			var done uint64
+			for !dsts.IsEmpty() {
+				n, err := readAt(ctx, dsts, gr.Start+done)
+				done += n
+				dsts = dsts.DropFirst64(n)
+				if err != nil {
+					if err == io.EOF {
+						// platform.AllocateAndFill truncates down to a page
+						// boundary, but FileRangeSet.Fill is supposed to
+						// zero-fill to the end of the page in this case.
+						donepgaddr, ok := usermem.Addr(done).RoundUp()
+						if donepg := uint64(donepgaddr); ok && donepg != done {
+							dsts.DropFirst64(donepg - done)
+							done = donepg
+							if dsts.IsEmpty() {
+								return done, nil
+							}
+						}
+					}
+					return done, err
+				}
+			}
+			return done, nil
+		}))
+
+		// Store anything we managed to read into the cache.
+		if done := fr.Length(); done != 0 {
+			gr.End = gr.Start + done
+			gap = frs.Insert(gap, gr, fr.Start).NextGap()
+		}
+
+		if err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// Drop removes segments for memmap.Mappable offsets in mr, freeing the
+// corresponding platform.FileRanges.
+//
+// Preconditions: mr must be page-aligned.
+func (frs *FileRangeSet) Drop(mr memmap.MappableRange, mem platform.Memory) {
+	seg := frs.LowerBoundSegment(mr.Start)
+	for seg.Ok() && seg.Start() < mr.End {
+		seg = frs.Isolate(seg, mr)
+		mem.DecRef(seg.FileRange())
+		seg = frs.Remove(seg).NextSegment()
+	}
+}
+
+// DropAll removes all segments in mr, freeing the corresponding
+// platform.FileRanges.
+func (frs *FileRangeSet) DropAll(mem platform.Memory) {
+	for seg := frs.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		mem.DecRef(seg.FileRange())
+	}
+	frs.RemoveAll()
+}
+
+// Truncate updates frs to reflect Mappable truncation to the given length:
+// bytes after the new EOF on the same page are zeroed, and pages after the new
+// EOF are freed.
+func (frs *FileRangeSet) Truncate(end uint64, mem platform.Memory) {
+	pgendaddr, ok := usermem.Addr(end).RoundUp()
+	if ok {
+		pgend := uint64(pgendaddr)
+
+		// Free truncated pages.
+		frs.SplitAt(pgend)
+		seg := frs.LowerBoundSegment(pgend)
+		for seg.Ok() {
+			mem.DecRef(seg.FileRange())
+			seg = frs.Remove(seg).NextSegment()
+		}
+
+		if end == pgend {
+			return
+		}
+	}
+
+	// Here we know end < end.RoundUp(). If the new EOF lands in the
+	// middle of a page that we have, zero out its contents beyond the new
+	// length.
+	seg := frs.FindSegment(end)
+	if seg.Ok() {
+		fr := seg.FileRange()
+		fr.Start += end - seg.Start()
+		ims, err := mem.MapInternal(fr, usermem.Write)
+		if err != nil {
+			// There's no good recourse from here. This means
+			// that we can't keep cached memory consistent with
+			// the new end of file. The caller may have already
+			// updated the file size on their backing file system.
+			//
+			// We don't want to risk blindly continuing onward,
+			// so in the extremely rare cases this does happen,
+			// we abandon ship.
+			panic(fmt.Sprintf("Failed to map %v: %v", fr, err))
+		}
+		if _, err := safemem.ZeroSeq(ims); err != nil {
+			panic(fmt.Sprintf("Zeroing %v failed: %v", fr, err))
+		}
+	}
+}
diff --git a/pkg/sentry/fs/fsutil/frame_ref_set.go b/pkg/sentry/fs/fsutil/frame_ref_set.go
new file mode 100644
index 000000000..14dece315
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/frame_ref_set.go
@@ -0,0 +1,50 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+)
+
+type frameRefSetFunctions struct{}
+
+// MinKey implements segment.Functions.MinKey.
+func (frameRefSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+// MaxKey implements segment.Functions.MaxKey.
+func (frameRefSetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+// ClearValue implements segment.Functions.ClearValue.
+func (frameRefSetFunctions) ClearValue(val *uint64) {
+}
+
+// Merge implements segment.Functions.Merge.
+func (frameRefSetFunctions) Merge(_ platform.FileRange, val1 uint64, _ platform.FileRange, val2 uint64) (uint64, bool) {
+	if val1 != val2 {
+		return 0, false
+	}
+	return val1, true
+}
+
+// Split implements segment.Functions.Split.
+func (frameRefSetFunctions) Split(_ platform.FileRange, val uint64, _ uint64) (uint64, uint64) {
+	return val, val
+}
diff --git a/pkg/sentry/fs/fsutil/fsutil.go b/pkg/sentry/fs/fsutil/fsutil.go
new file mode 100644
index 000000000..6fe4ef13d
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/fsutil.go
@@ -0,0 +1,26 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package fsutil provides utilities for implementing fs.InodeOperations
+// and fs.FileOperations:
+//
+// - For embeddable utilities, see inode.go and file.go.
+//
+// - For fs.Inodes that require a page cache to be memory mapped, see
+//   inode_cache.go.
+//
+// - For fs.Files that implement fs.HandleOps, see handle.go.
+//
+// - For anon fs.Inodes, see anon.go.
+package fsutil
diff --git a/pkg/sentry/fs/fsutil/handle.go b/pkg/sentry/fs/fsutil/handle.go
new file mode 100644
index 000000000..149c0f84a
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/handle.go
@@ -0,0 +1,126 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Handle implements FileOperations.
+//
+// FIXME: Remove Handle entirely in favor of individual fs.File
+// implementations using simple generic utilities.
+type Handle struct {
+	NoopRelease      `state:"nosave"`
+	NoIoctl          `state:"nosave"`
+	HandleOperations fs.HandleOperations
+
+	// dirCursor is the directory cursor.
+	dirCursor string
+}
+
+// NewHandle returns a File backed by the Dirent and FileFlags.
+func NewHandle(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags, hops fs.HandleOperations) *fs.File {
+	if !fs.IsPipe(dirent.Inode.StableAttr) && !fs.IsSocket(dirent.Inode.StableAttr) {
+		// Allow reading/writing at an arbitrary offset for non-pipes
+		// and non-sockets.
+		flags.Pread = true
+		flags.Pwrite = true
+	}
+
+	return fs.NewFile(ctx, dirent, flags, &Handle{HandleOperations: hops})
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (h *Handle) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return h.HandleOperations.Readiness(mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (h *Handle) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	h.HandleOperations.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (h *Handle) EventUnregister(e *waiter.Entry) {
+	h.HandleOperations.EventUnregister(e)
+}
+
+// Readdir implements FileOperations.Readdir.
+func (h *Handle) Readdir(ctx context.Context, file *fs.File, serializer fs.DentrySerializer) (int64, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+	dirCtx := &fs.DirCtx{
+		Serializer: serializer,
+		DirCursor:  &h.dirCursor,
+	}
+	n, err := fs.DirentReaddir(ctx, file.Dirent, h, root, dirCtx, file.Offset())
+	return n, err
+}
+
+// Seek implements FileOperations.Seek.
+func (h *Handle) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	return SeekWithDirCursor(ctx, file, whence, offset, &h.dirCursor)
+}
+
+// IterateDir implements DirIterator.IterateDir.
+func (h *Handle) IterateDir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	return h.HandleOperations.DeprecatedReaddir(ctx, dirCtx, offset)
+}
+
+// Read implements FileOperations.Read.
+func (h *Handle) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	return h.HandleOperations.DeprecatedPreadv(ctx, dst, offset)
+}
+
+// Write implements FileOperations.Write.
+func (h *Handle) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	return h.HandleOperations.DeprecatedPwritev(ctx, src, offset)
+}
+
+// Fsync implements FileOperations.Fsync.
+func (h *Handle) Fsync(ctx context.Context, file *fs.File, start int64, end int64, syncType fs.SyncType) error {
+	switch syncType {
+	case fs.SyncAll, fs.SyncData:
+		// Write out metadata.
+		if err := file.Dirent.Inode.WriteOut(ctx); err != nil {
+			return err
+		}
+		fallthrough
+	case fs.SyncBackingStorage:
+		// Use DeprecatedFsync to sync disks.
+		return h.HandleOperations.DeprecatedFsync()
+	}
+	panic("invalid sync type")
+}
+
+// Flush implements FileOperations.Flush.
+func (h *Handle) Flush(context.Context, *fs.File) error {
+	return h.HandleOperations.DeprecatedFlush()
+}
+
+// ConfigureMMap implements FileOperations.ConfigureMMap.
+func (h *Handle) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	mappable := file.Dirent.Inode.Mappable()
+	if mappable == nil {
+		return syserror.ENODEV
+	}
+	return GenericConfigureMMap(file, mappable, opts)
+}
diff --git a/pkg/sentry/fs/fsutil/handle_test.go b/pkg/sentry/fs/fsutil/handle_test.go
new file mode 100644
index 000000000..d94c3eb0d
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/handle_test.go
@@ -0,0 +1,227 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil_test
+
+import (
+	"io"
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	ramfstest "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs/test"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type testInodeOperations struct {
+	fs.InodeOperations
+	fs.InodeType
+	FileSize int64
+	writes   uint
+	reads    uint
+}
+
+func (t *testInodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	return fs.UnstableAttr{Size: t.FileSize}, nil
+}
+
+// Check implements InodeOperations.Check.
+func (t *testInodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+func (t *testInodeOperations) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	t.reads++
+	return t.InodeOperations.DeprecatedPreadv(ctx, dst, offset)
+}
+
+func (t *testInodeOperations) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	t.writes++
+	return t.InodeOperations.DeprecatedPwritev(ctx, src, offset)
+}
+
+// testHandle returns a handle for a test node.
+//
+// The size of the node is fixed at 20 bytes.
+func testHandle(t *testing.T, flags fs.FileFlags, nt fs.InodeType) (*fs.File, *testInodeOperations) {
+	ctx := contexttest.Context(t)
+	m := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+	n := &testInodeOperations{
+		InodeOperations: ramfstest.NewFile(ctx, fs.FilePermissions{User: fs.PermMask{Read: true, Write: true}}),
+		FileSize:        20,
+	}
+	d := fs.NewDirent(fs.NewInode(n, m, fs.StableAttr{Type: nt}), "test")
+	return fsutil.NewHandle(ctx, d, flags, d.Inode.HandleOps()), n
+}
+
+func TestHandleOps(t *testing.T) {
+	h, n := testHandle(t, fs.FileFlags{Read: true, Write: true}, fs.RegularFile)
+	defer h.DecRef()
+
+	// Make sure a write request works.
+	if n, err := h.Writev(contexttest.Context(t), usermem.BytesIOSequence([]byte("a"))); n != 1 || err != nil {
+		t.Fatalf("Writev: got (%d, %v), wanted (1, nil)", n, err)
+	}
+	if n.writes != 1 {
+		t.Errorf("found %d writes, expected 1", n.writes)
+	}
+
+	// Make sure a read request works.
+	dst := make([]byte, 1)
+	if n, err := h.Preadv(contexttest.Context(t), usermem.BytesIOSequence(dst), 0); n != 1 || (err != nil && err != io.EOF) {
+		t.Errorf("Preadv: got (%d, %v), wanted (1, nil or EOF)", n, err)
+	}
+	if dst[0] != 'a' {
+		t.Errorf("Preadv: read %q, wanted 'a'", dst[0])
+	}
+	if n.reads != 1 {
+		t.Errorf("found %d reads, expected 1", n.reads)
+	}
+}
+
+type seekTest struct {
+	whence fs.SeekWhence
+	offset int64
+	result int64
+	err    error
+}
+
+type seekSuite struct {
+	nodeType fs.InodeType
+	cases    []seekTest
+}
+
+// FIXME: This is currently missing fs.SeekEnd tests due to the
+// fact that NullInodeOperations returns an error on stat.
+func TestHandleSeek(t *testing.T) {
+	ts := []seekSuite{
+		{
+			nodeType: fs.RegularFile,
+			cases: []seekTest{
+				{fs.SeekSet, 0, 0, nil},
+				{fs.SeekSet, 10, 10, nil},
+				{fs.SeekSet, -5, 10, syscall.EINVAL},
+				{fs.SeekCurrent, -1, 9, nil},
+				{fs.SeekCurrent, 2, 11, nil},
+				{fs.SeekCurrent, -12, 11, syscall.EINVAL},
+				{fs.SeekEnd, -1, 19, nil},
+				{fs.SeekEnd, 0, 20, nil},
+				{fs.SeekEnd, 2, 22, nil},
+			},
+		},
+		{
+			nodeType: fs.Directory,
+			cases: []seekTest{
+				{fs.SeekSet, 0, 0, nil},
+				{fs.SeekSet, 10, 0, syscall.EINVAL},
+				{fs.SeekSet, -5, 0, syscall.EINVAL},
+				{fs.SeekCurrent, 0, 0, nil},
+				{fs.SeekCurrent, 11, 0, syscall.EINVAL},
+				{fs.SeekCurrent, -6, 0, syscall.EINVAL},
+				{fs.SeekEnd, 0, 0, syscall.EINVAL},
+				{fs.SeekEnd, -1, 0, syscall.EINVAL},
+				{fs.SeekEnd, 2, 0, syscall.EINVAL},
+			},
+		},
+		{
+			nodeType: fs.Symlink,
+			cases: []seekTest{
+				{fs.SeekSet, 5, 0, syscall.EINVAL},
+				{fs.SeekSet, -5, 0, syscall.EINVAL},
+				{fs.SeekSet, 0, 0, syscall.EINVAL},
+				{fs.SeekCurrent, 5, 0, syscall.EINVAL},
+				{fs.SeekCurrent, -5, 0, syscall.EINVAL},
+				{fs.SeekCurrent, 0, 0, syscall.EINVAL},
+				{fs.SeekEnd, 5, 0, syscall.EINVAL},
+				{fs.SeekEnd, -5, 0, syscall.EINVAL},
+				{fs.SeekEnd, 0, 0, syscall.EINVAL},
+			},
+		},
+		{
+			nodeType: fs.Pipe,
+			cases: []seekTest{
+				{fs.SeekSet, 5, 0, syscall.ESPIPE},
+				{fs.SeekSet, -5, 0, syscall.ESPIPE},
+				{fs.SeekSet, 0, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, 5, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, -5, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, 0, 0, syscall.ESPIPE},
+				{fs.SeekEnd, 5, 0, syscall.ESPIPE},
+				{fs.SeekEnd, -5, 0, syscall.ESPIPE},
+				{fs.SeekEnd, 0, 0, syscall.ESPIPE},
+			},
+		},
+		{
+			nodeType: fs.Socket,
+			cases: []seekTest{
+				{fs.SeekSet, 5, 0, syscall.ESPIPE},
+				{fs.SeekSet, -5, 0, syscall.ESPIPE},
+				{fs.SeekSet, 0, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, 5, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, -5, 0, syscall.ESPIPE},
+				{fs.SeekCurrent, 0, 0, syscall.ESPIPE},
+				{fs.SeekEnd, 5, 0, syscall.ESPIPE},
+				{fs.SeekEnd, -5, 0, syscall.ESPIPE},
+				{fs.SeekEnd, 0, 0, syscall.ESPIPE},
+			},
+		},
+		{
+			nodeType: fs.CharacterDevice,
+			cases: []seekTest{
+				{fs.SeekSet, 5, 0, nil},
+				{fs.SeekSet, -5, 0, nil},
+				{fs.SeekSet, 0, 0, nil},
+				{fs.SeekCurrent, 5, 0, nil},
+				{fs.SeekCurrent, -5, 0, nil},
+				{fs.SeekCurrent, 0, 0, nil},
+				{fs.SeekEnd, 5, 0, nil},
+				{fs.SeekEnd, -5, 0, nil},
+				{fs.SeekEnd, 0, 0, nil},
+			},
+		},
+		{
+			nodeType: fs.BlockDevice,
+			cases: []seekTest{
+				{fs.SeekSet, 0, 0, nil},
+				{fs.SeekSet, 10, 10, nil},
+				{fs.SeekSet, -5, 10, syscall.EINVAL},
+				{fs.SeekCurrent, -1, 9, nil},
+				{fs.SeekCurrent, 2, 11, nil},
+				{fs.SeekCurrent, -12, 11, syscall.EINVAL},
+				{fs.SeekEnd, -1, 19, nil},
+				{fs.SeekEnd, 0, 20, nil},
+				{fs.SeekEnd, 2, 22, nil},
+			},
+		},
+	}
+
+	for _, s := range ts {
+		h, _ := testHandle(t, fs.FileFlags{Read: true, Write: true}, s.nodeType)
+		defer h.DecRef()
+
+		for _, c := range s.cases {
+			// Try the given seek.
+			offset, err := h.Seek(contexttest.Context(t), c.whence, c.offset)
+			if err != c.err {
+				t.Errorf("seek(%s, %d) on %s had unexpected error: expected %v, got %v", c.whence, c.offset, s.nodeType, c.err, err)
+			}
+			if err == nil && offset != c.result {
+				t.Errorf("seek(%s, %d) on %s had bad result: expected %v, got %v", c.whence, c.offset, s.nodeType, c.result, offset)
+			}
+		}
+	}
+}
diff --git a/pkg/sentry/fs/fsutil/host_file_mapper.go b/pkg/sentry/fs/fsutil/host_file_mapper.go
new file mode 100644
index 000000000..d0a27fc1c
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/host_file_mapper.go
@@ -0,0 +1,209 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// HostFileMapper caches mappings of an arbitrary host file descriptor. It is
+// used by implementations of memmap.Mappable that represent a host file
+// descriptor.
+type HostFileMapper struct {
+	// HostFile conceptually breaks the file into pieces called chunks, of
+	// size and alignment chunkSize, and caches mappings of the file on a chunk
+	// granularity.
+
+	refsMu sync.Mutex `state:"nosave"`
+
+	// refs maps chunk start offsets to the sum of reference counts for all
+	// pages in that chunk. refs is protected by refsMu.
+	refs map[uint64]int32
+
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings maps chunk start offsets to mappings of those chunks,
+	// obtained by calling syscall.Mmap. mappings is protected by
+	// mapsMu.
+	mappings map[uint64]mapping `state:"nosave"`
+}
+
+const (
+	chunkShift = usermem.HugePageShift
+	chunkSize  = 1 << chunkShift
+	chunkMask  = chunkSize - 1
+)
+
+func pagesInChunk(mr memmap.MappableRange, chunkStart uint64) int32 {
+	return int32(mr.Intersect(memmap.MappableRange{chunkStart, chunkStart + chunkSize}).Length() / usermem.PageSize)
+}
+
+type mapping struct {
+	addr     uintptr
+	writable bool
+}
+
+// NewHostFileMapper returns a HostFileMapper with no references or cached
+// mappings.
+func NewHostFileMapper() *HostFileMapper {
+	return &HostFileMapper{
+		refs:     make(map[uint64]int32),
+		mappings: make(map[uint64]mapping),
+	}
+}
+
+// IncRefOn increments the reference count on all offsets in mr.
+//
+// Preconditions: mr.Length() != 0. mr.Start and mr.End must be page-aligned.
+func (f *HostFileMapper) IncRefOn(mr memmap.MappableRange) {
+	f.refsMu.Lock()
+	defer f.refsMu.Unlock()
+	for chunkStart := mr.Start &^ chunkMask; chunkStart < mr.End; chunkStart += chunkSize {
+		refs := f.refs[chunkStart]
+		pgs := pagesInChunk(mr, chunkStart)
+		if refs+pgs < refs {
+			// Would overflow.
+			panic(fmt.Sprintf("HostFileMapper.IncRefOn(%v): adding %d page references to chunk %#x, which has %d page references", mr, pgs, chunkStart, refs))
+		}
+		f.refs[chunkStart] = refs + pgs
+	}
+}
+
+// DecRefOn decrements the reference count on all offsets in mr.
+//
+// Preconditions: mr.Length() != 0. mr.Start and mr.End must be page-aligned.
+func (f *HostFileMapper) DecRefOn(mr memmap.MappableRange) {
+	f.refsMu.Lock()
+	defer f.refsMu.Unlock()
+	for chunkStart := mr.Start &^ chunkMask; chunkStart < mr.End; chunkStart += chunkSize {
+		refs := f.refs[chunkStart]
+		pgs := pagesInChunk(mr, chunkStart)
+		switch {
+		case refs > pgs:
+			f.refs[chunkStart] = refs - pgs
+		case refs == pgs:
+			f.mapsMu.Lock()
+			delete(f.refs, chunkStart)
+			if m, ok := f.mappings[chunkStart]; ok {
+				f.unmapAndRemoveLocked(chunkStart, m)
+			}
+			f.mapsMu.Unlock()
+		case refs < pgs:
+			panic(fmt.Sprintf("HostFileMapper.DecRefOn(%v): removing %d page references from chunk %#x, which has %d page references", mr, pgs, chunkStart, refs))
+		}
+	}
+}
+
+// MapInternal returns a mapping of offsets in fr from fd. The returned
+// safemem.BlockSeq is valid as long as at least one reference is held on all
+// offsets in fr or until the next call to UnmapAll.
+//
+// Preconditions: The caller must hold a reference on all offsets in fr.
+func (f *HostFileMapper) MapInternal(fr platform.FileRange, fd int, write bool) (safemem.BlockSeq, error) {
+	chunks := ((fr.End + chunkMask) >> chunkShift) - (fr.Start >> chunkShift)
+	f.mapsMu.Lock()
+	defer f.mapsMu.Unlock()
+	if chunks == 1 {
+		// Avoid an unnecessary slice allocation.
+		var seq safemem.BlockSeq
+		err := f.forEachMappingBlockLocked(fr, fd, write, func(b safemem.Block) {
+			seq = safemem.BlockSeqOf(b)
+		})
+		return seq, err
+	}
+	blocks := make([]safemem.Block, 0, chunks)
+	err := f.forEachMappingBlockLocked(fr, fd, write, func(b safemem.Block) {
+		blocks = append(blocks, b)
+	})
+	return safemem.BlockSeqFromSlice(blocks), err
+}
+
+// Preconditions: f.mapsMu must be locked.
+func (f *HostFileMapper) forEachMappingBlockLocked(fr platform.FileRange, fd int, write bool, fn func(safemem.Block)) error {
+	prot := syscall.PROT_READ
+	if write {
+		prot |= syscall.PROT_WRITE
+	}
+	for chunkStart := fr.Start &^ chunkMask; chunkStart < fr.End; chunkStart += chunkSize {
+		m, ok := f.mappings[chunkStart]
+		if !ok {
+			addr, _, errno := syscall.Syscall6(
+				syscall.SYS_MMAP,
+				0,
+				chunkSize,
+				uintptr(prot),
+				syscall.MAP_SHARED,
+				uintptr(fd),
+				uintptr(chunkStart))
+			if errno != 0 {
+				return errno
+			}
+			m = mapping{addr, write}
+			f.mappings[chunkStart] = m
+		} else if write && !m.writable {
+			addr, _, errno := syscall.Syscall6(
+				syscall.SYS_MMAP,
+				m.addr,
+				chunkSize,
+				uintptr(prot),
+				syscall.MAP_SHARED|syscall.MAP_FIXED,
+				uintptr(fd),
+				uintptr(chunkStart))
+			if errno != 0 {
+				return errno
+			}
+			m = mapping{addr, write}
+			f.mappings[chunkStart] = m
+		}
+		var startOff uint64
+		if chunkStart < fr.Start {
+			startOff = fr.Start - chunkStart
+		}
+		endOff := uint64(chunkSize)
+		if chunkStart+chunkSize > fr.End {
+			endOff = fr.End - chunkStart
+		}
+		fn(f.unsafeBlockFromChunkMapping(m.addr).TakeFirst64(endOff).DropFirst64(startOff))
+	}
+	return nil
+}
+
+// UnmapAll unmaps all cached mappings. Callers are responsible for
+// synchronization with mappings returned by previous calls to MapInternal.
+func (f *HostFileMapper) UnmapAll() {
+	f.mapsMu.Lock()
+	defer f.mapsMu.Unlock()
+	for chunkStart, m := range f.mappings {
+		f.unmapAndRemoveLocked(chunkStart, m)
+	}
+}
+
+// Preconditions: f.mapsMu must be locked. f.mappings[chunkStart] == m.
+func (f *HostFileMapper) unmapAndRemoveLocked(chunkStart uint64, m mapping) {
+	if _, _, errno := syscall.Syscall(syscall.SYS_MUNMAP, m.addr, chunkSize, 0); errno != 0 {
+		// This leaks address space and is unexpected, but is otherwise
+		// harmless, so complain but don't panic.
+		log.Warningf("HostFileMapper: failed to unmap mapping %#x for chunk %#x: %v", m.addr, chunkStart, errno)
+	}
+	delete(f.mappings, chunkStart)
+}
diff --git a/pkg/sentry/fs/fsutil/host_file_mapper_state.go b/pkg/sentry/fs/fsutil/host_file_mapper_state.go
new file mode 100644
index 000000000..57705decd
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/host_file_mapper_state.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+// afterLoad is invoked by stateify.
+func (f *HostFileMapper) afterLoad() {
+	f.mappings = make(map[uint64]mapping)
+}
diff --git a/pkg/sentry/fs/fsutil/host_file_mapper_unsafe.go b/pkg/sentry/fs/fsutil/host_file_mapper_unsafe.go
new file mode 100644
index 000000000..790f3a5a6
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/host_file_mapper_unsafe.go
@@ -0,0 +1,27 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+)
+
+func (*HostFileMapper) unsafeBlockFromChunkMapping(addr uintptr) safemem.Block {
+	// We don't control the host file's length, so touching its mappings may
+	// raise SIGBUS. Thus accesses to it must use safecopy.
+	return safemem.BlockFromUnsafePointer((unsafe.Pointer)(addr), chunkSize)
+}
diff --git a/pkg/sentry/fs/fsutil/inode.go b/pkg/sentry/fs/fsutil/inode.go
new file mode 100644
index 000000000..e1ad07df2
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/inode.go
@@ -0,0 +1,380 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// NewSimpleInodeOperations constructs fs.InodeOperations from InodeSimpleAttributes.
+func NewSimpleInodeOperations(i InodeSimpleAttributes) fs.InodeOperations {
+	return &simpleInodeOperations{InodeSimpleAttributes: i}
+}
+
+// simpleInodeOperations is a simple implementation of Inode.
+type simpleInodeOperations struct {
+	DeprecatedFileOperations  `state:"nosave"`
+	InodeNotDirectory         `state:"nosave"`
+	InodeNotSocket            `state:"nosave"`
+	InodeNotRenameable        `state:"nosave"`
+	InodeNotOpenable          `state:"nosave"`
+	InodeNotVirtual           `state:"nosave"`
+	InodeNotSymlink           `state:"nosave"`
+	InodeNoExtendedAttributes `state:"nosave"`
+	NoMappable                `state:"nosave"`
+	NoopWriteOut              `state:"nosave"`
+
+	InodeSimpleAttributes
+}
+
+// InodeSimpleAttributes implements a subset of the Inode interface. It provides
+// read-only access to attributes.
+type InodeSimpleAttributes struct {
+	// FSType is the filesystem type reported by StatFS.
+	FSType uint64
+
+	// UAttr are the unstable attributes of the Inode.
+	UAttr fs.UnstableAttr
+}
+
+// Release implements fs.InodeOperations.Release.
+func (i *InodeSimpleAttributes) Release(context.Context) {}
+
+// StatFS implements fs.InodeOperations.StatFS.
+func (i *InodeSimpleAttributes) StatFS(context.Context) (fs.Info, error) {
+	return fs.Info{Type: i.FSType}, nil
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (i *InodeSimpleAttributes) UnstableAttr(context.Context, *fs.Inode) (fs.UnstableAttr, error) {
+	return i.UAttr, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (i *InodeSimpleAttributes) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+func (*InodeSimpleAttributes) AddLink() {}
+
+// DropLink implements fs.InodeOperations.DropLink.
+func (*InodeSimpleAttributes) DropLink() {}
+
+// NotifyStatusChange implements fs.fs.InodeOperations.
+func (i *InodeSimpleAttributes) NotifyStatusChange(ctx context.Context) {
+	i.UAttr.StatusChangeTime = ktime.NowFromContext(ctx)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (*InodeSimpleAttributes) SetPermissions(context.Context, *fs.Inode, fs.FilePermissions) bool {
+	return false
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (*InodeSimpleAttributes) SetOwner(context.Context, *fs.Inode, fs.FileOwner) error {
+	return syserror.EINVAL
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (*InodeSimpleAttributes) SetTimestamps(context.Context, *fs.Inode, fs.TimeSpec) error {
+	return syserror.EINVAL
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (*InodeSimpleAttributes) Truncate(context.Context, *fs.Inode, int64) error {
+	return syserror.EINVAL
+}
+
+// InMemoryAttributes implements utilities for updating in-memory unstable
+// attributes and extended attributes. It is not thread-safe.
+//
+// Users need not initialize Xattrs to non-nil (it will be initialized
+// when the first extended attribute is set.
+type InMemoryAttributes struct {
+	Unstable fs.UnstableAttr
+	Xattrs   map[string][]byte
+}
+
+// SetPermissions updates the permissions to p.
+func (i *InMemoryAttributes) SetPermissions(ctx context.Context, p fs.FilePermissions) bool {
+	i.Unstable.Perms = p
+	i.Unstable.StatusChangeTime = ktime.NowFromContext(ctx)
+	return true
+}
+
+// SetOwner updates the file owner to owner.
+func (i *InMemoryAttributes) SetOwner(ctx context.Context, owner fs.FileOwner) error {
+	if owner.UID.Ok() {
+		i.Unstable.Owner.UID = owner.UID
+	}
+	if owner.GID.Ok() {
+		i.Unstable.Owner.GID = owner.GID
+	}
+	return nil
+}
+
+// SetTimestamps sets the timestamps to ts.
+func (i *InMemoryAttributes) SetTimestamps(ctx context.Context, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	now := ktime.NowFromContext(ctx)
+	if !ts.ATimeOmit {
+		if ts.ATimeSetSystemTime {
+			i.Unstable.AccessTime = now
+		} else {
+			i.Unstable.AccessTime = ts.ATime
+		}
+	}
+	if !ts.MTimeOmit {
+		if ts.MTimeSetSystemTime {
+			i.Unstable.ModificationTime = now
+		} else {
+			i.Unstable.ModificationTime = ts.MTime
+		}
+	}
+	i.Unstable.StatusChangeTime = now
+	return nil
+}
+
+// TouchAccessTime updates access time to the current time.
+func (i *InMemoryAttributes) TouchAccessTime(ctx context.Context) {
+	i.Unstable.AccessTime = ktime.NowFromContext(ctx)
+}
+
+// TouchModificationTime updates modification and status change
+// time to the current time.
+func (i *InMemoryAttributes) TouchModificationTime(ctx context.Context) {
+	now := ktime.NowFromContext(ctx)
+	i.Unstable.ModificationTime = now
+	i.Unstable.StatusChangeTime = now
+}
+
+// TouchStatusChangeTime updates status change time to the current time.
+func (i *InMemoryAttributes) TouchStatusChangeTime(ctx context.Context) {
+	i.Unstable.StatusChangeTime = ktime.NowFromContext(ctx)
+}
+
+// Getxattr returns the extended attribute at name or ENOATTR if
+// it isn't set.
+func (i *InMemoryAttributes) Getxattr(name string) ([]byte, error) {
+	if value, ok := i.Xattrs[name]; ok {
+		return value, nil
+	}
+	return nil, syserror.ENOATTR
+}
+
+// Setxattr sets the extended attribute at name to value.
+func (i *InMemoryAttributes) Setxattr(name string, value []byte) error {
+	if i.Xattrs == nil {
+		i.Xattrs = make(map[string][]byte)
+	}
+	i.Xattrs[name] = value
+	return nil
+}
+
+// Listxattr returns the set of all currently set extended attributes.
+func (i *InMemoryAttributes) Listxattr() (map[string]struct{}, error) {
+	names := make(map[string]struct{}, len(i.Xattrs))
+	for name := range i.Xattrs {
+		names[name] = struct{}{}
+	}
+	return names, nil
+}
+
+// NoMappable returns a nil memmap.Mappable.
+type NoMappable struct{}
+
+// Mappable implements fs.InodeOperations.Mappable.
+func (NoMappable) Mappable(*fs.Inode) memmap.Mappable {
+	return nil
+}
+
+// NoopWriteOut is a no-op implementation of Inode.WriteOut.
+type NoopWriteOut struct{}
+
+// WriteOut is a no-op.
+func (NoopWriteOut) WriteOut(context.Context, *fs.Inode) error {
+	return nil
+}
+
+// InodeNotDirectory can be used by Inodes that are not directories.
+type InodeNotDirectory struct{}
+
+// Lookup implements fs.InodeOperations.Lookup.
+func (InodeNotDirectory) Lookup(context.Context, *fs.Inode, string) (*fs.Dirent, error) {
+	return nil, syserror.ENOTDIR
+}
+
+// Create implements fs.InodeOperations.Create.
+func (InodeNotDirectory) Create(context.Context, *fs.Inode, string, fs.FileFlags, fs.FilePermissions) (*fs.File, error) {
+	return nil, syserror.ENOTDIR
+}
+
+// CreateLink implements fs.InodeOperations.CreateLink.
+func (InodeNotDirectory) CreateLink(context.Context, *fs.Inode, string, string) error {
+	return syserror.ENOTDIR
+}
+
+// CreateHardLink implements fs.InodeOperations.CreateHardLink.
+func (InodeNotDirectory) CreateHardLink(context.Context, *fs.Inode, *fs.Inode, string) error {
+	return syserror.ENOTDIR
+}
+
+// CreateDirectory implements fs.InodeOperations.CreateDirectory.
+func (InodeNotDirectory) CreateDirectory(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return syserror.ENOTDIR
+}
+
+// Bind implements fs.InodeOperations.Bind.
+func (InodeNotDirectory) Bind(context.Context, *fs.Inode, string, unix.BoundEndpoint, fs.FilePermissions) error {
+	return syserror.ENOTDIR
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo.
+func (InodeNotDirectory) CreateFifo(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return syserror.ENOTDIR
+}
+
+// Remove implements fs.InodeOperations.Remove.
+func (InodeNotDirectory) Remove(context.Context, *fs.Inode, string) error {
+	return syserror.ENOTDIR
+}
+
+// RemoveDirectory implements fs.InodeOperations.RemoveDirectory.
+func (InodeNotDirectory) RemoveDirectory(context.Context, *fs.Inode, string) error {
+	return syserror.ENOTDIR
+}
+
+// InodeNotSocket can be used by Inodes that are not sockets.
+type InodeNotSocket struct{}
+
+// BoundEndpoint implements fs.InodeOperations.BoundEndpoint.
+func (InodeNotSocket) BoundEndpoint(*fs.Inode, string) unix.BoundEndpoint {
+	return nil
+}
+
+// InodeNotRenameable can be used by Inodes that cannot be renamed.
+type InodeNotRenameable struct{}
+
+// Rename implements fs.InodeOperations.Rename.
+func (InodeNotRenameable) Rename(context.Context, *fs.Inode, string, *fs.Inode, string) error {
+	return syserror.EINVAL
+}
+
+// InodeNotOpenable can be used by Inodes that cannot be opened.
+type InodeNotOpenable struct{}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (InodeNotOpenable) GetFile(context.Context, *fs.Dirent, fs.FileFlags) (*fs.File, error) {
+	return nil, syserror.EIO
+}
+
+// InodeNotVirtual can be used by Inodes that are not virtual.
+type InodeNotVirtual struct{}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+func (InodeNotVirtual) IsVirtual() bool {
+	return false
+}
+
+// InodeNotSymlink can be used by Inodes that are not symlinks.
+type InodeNotSymlink struct{}
+
+// Readlink implements fs.InodeOperations.Readlink.
+func (InodeNotSymlink) Readlink(context.Context, *fs.Inode) (string, error) {
+	return "", syserror.ENOLINK
+}
+
+// Getlink implements fs.InodeOperations.Getlink.
+func (InodeNotSymlink) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	return nil, syserror.ENOLINK
+}
+
+// InodeNoExtendedAttributes can be used by Inodes that do not support
+// extended attributes.
+type InodeNoExtendedAttributes struct{}
+
+// Getxattr implements fs.InodeOperations.Getxattr.
+func (InodeNoExtendedAttributes) Getxattr(*fs.Inode, string) ([]byte, error) {
+	return nil, syserror.EOPNOTSUPP
+}
+
+// Setxattr implements fs.InodeOperations.Setxattr.
+func (InodeNoExtendedAttributes) Setxattr(*fs.Inode, string, []byte) error {
+	return syserror.EOPNOTSUPP
+}
+
+// Listxattr implements fs.InodeOperations.Listxattr.
+func (InodeNoExtendedAttributes) Listxattr(*fs.Inode) (map[string]struct{}, error) {
+	return nil, syserror.EOPNOTSUPP
+}
+
+// DeprecatedFileOperations panics if any deprecated Inode method is called.
+type DeprecatedFileOperations struct{}
+
+// Readiness implements fs.InodeOperations.Waitable.Readiness.
+func (DeprecatedFileOperations) Readiness(waiter.EventMask) waiter.EventMask {
+	panic("not implemented")
+}
+
+// EventRegister implements fs.InodeOperations.Waitable.EventRegister.
+func (DeprecatedFileOperations) EventRegister(*waiter.Entry, waiter.EventMask) {
+	panic("not implemented")
+}
+
+// EventUnregister implements fs.InodeOperations.Waitable.EventUnregister.
+func (DeprecatedFileOperations) EventUnregister(*waiter.Entry) {
+	panic("not implemented")
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (DeprecatedFileOperations) DeprecatedPreadv(context.Context, usermem.IOSequence, int64) (int64, error) {
+	panic("not implemented")
+}
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev.
+func (DeprecatedFileOperations) DeprecatedPwritev(context.Context, usermem.IOSequence, int64) (int64, error) {
+	panic("not implemented")
+}
+
+// DeprecatedReaddir implements fs.InodeOperations.DeprecatedReaddir.
+func (DeprecatedFileOperations) DeprecatedReaddir(context.Context, *fs.DirCtx, int) (int, error) {
+	panic("not implemented")
+}
+
+// DeprecatedFsync implements fs.InodeOperations.DeprecatedFsync.
+func (DeprecatedFileOperations) DeprecatedFsync() error {
+	panic("not implemented")
+}
+
+// DeprecatedFlush implements fs.InodeOperations.DeprecatedFlush.
+func (DeprecatedFileOperations) DeprecatedFlush() error {
+	panic("not implemented")
+}
+
+// DeprecatedMappable implements fs.InodeOperations.DeprecatedMappable.
+func (DeprecatedFileOperations) DeprecatedMappable(context.Context, *fs.Inode) (memmap.Mappable, bool) {
+	panic("not implemented")
+}
diff --git a/pkg/sentry/fs/fsutil/inode_cached.go b/pkg/sentry/fs/fsutil/inode_cached.go
new file mode 100644
index 000000000..484668735
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/inode_cached.go
@@ -0,0 +1,845 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"fmt"
+	"io"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Lock order (compare the lock order model in mm/mm.go):
+//
+// CachingInodeOperations.attrMu ("fs locks")
+//   CachingInodeOperations.mapsMu ("memmap.Mappable locks not taken by Translate")
+//     CachingInodeOperations.dataMu ("memmap.Mappable locks taken by Translate")
+//       CachedFileObject locks
+
+// CachingInodeOperations caches the metadata and content of a CachedFileObject.
+// It implements a subset of InodeOperations. As a utility it can be used to
+// implement the full set of InodeOperations. Generally it should not be
+// embedded to avoid unexpected inherited behavior.
+//
+// CachingInodeOperations implements Mappable for the CachedFileObject:
+//
+// - If CachedFileObject.FD returns a value >= 0 and the current platform shares
+//   a host fd table with the sentry, then the value of CachedFileObject.FD
+//   will be memory mapped on the host.
+//
+// - Otherwise, the contents of CachedFileObject are buffered into memory
+//   managed by the CachingInodeOperations.
+//
+// Implementations of FileOperations for a CachedFileObject must read and
+// write through CachingInodeOperations using Read and Write respectively.
+//
+// Implementations of InodeOperations.WriteOut must call Sync to write out
+// in-memory modifications of data and metadata to the CachedFileObject.
+type CachingInodeOperations struct {
+	// backingFile is a handle to a cached file object.
+	backingFile CachedFileObject
+
+	// platform is used to allocate memory that caches backingFile's contents.
+	platform platform.Platform
+
+	// forcePageCache indicates the sentry page cache should be used regardless
+	// of whether the platform supports host mapped I/O or not. This must not be
+	// modified after inode creation.
+	forcePageCache bool
+
+	attrMu sync.Mutex `state:"nosave"`
+
+	// attr is unstable cached metadata.
+	//
+	// attr is protected by attrMu. attr.Size is protected by both attrMu and
+	// dataMu; reading it requires locking either mutex, while mutating it
+	// requires locking both.
+	attr fs.UnstableAttr
+
+	// dirtyAttr is metadata that was updated in-place but hasn't yet
+	// been successfully written out.
+	//
+	// dirtyAttr is protected by attrMu.
+	dirtyAttr fs.AttrMask
+
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings tracks mappings of the cached file object into
+	// memmap.MappingSpaces.
+	//
+	// mappings is protected by mapsMu.
+	mappings memmap.MappingSet
+
+	dataMu sync.RWMutex `state:"nosave"`
+
+	// cache maps offsets into the cached file to offsets into
+	// platform.Memory() that store the file's data.
+	//
+	// cache is protected by dataMu.
+	cache FileRangeSet
+
+	// dirty tracks dirty segments in cache.
+	//
+	// dirty is protected by dataMu.
+	dirty DirtySet
+
+	// hostFileMapper caches internal mappings of backingFile.FD().
+	hostFileMapper *HostFileMapper
+
+	// refs tracks active references to data in the cache.
+	//
+	// refs is protected by dataMu.
+	refs frameRefSet
+}
+
+// CachedFileObject is a file that may require caching.
+type CachedFileObject interface {
+	// ReadToBlocksAt reads up to dsts.NumBytes() bytes from the file to dsts,
+	// starting at offset, and returns the number of bytes read. ReadToBlocksAt
+	// may return a partial read without an error.
+	ReadToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error)
+
+	// WriteFromBlocksAt writes up to srcs.NumBytes() bytes from srcs to the
+	// file, starting at offset, and returns the number of bytes written.
+	// WriteFromBlocksAt may return a partial write without an error.
+	WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error)
+
+	// SetMaskedAttributes sets the attributes in attr that are true in mask
+	// on the backing file.
+	//
+	// SetMaskedAttributes may be called at any point, regardless of whether
+	// the file was opened.
+	SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error
+
+	// Sync instructs the remote filesystem to sync the file to stable storage.
+	Sync(ctx context.Context) error
+
+	// FD returns a host file descriptor. Return value must be -1 or not -1
+	// for the lifetime of the CachedFileObject.
+	//
+	// FD is called iff the file has been memory mapped. This implies that
+	// the file was opened (see fs.InodeOperations.GetFile).
+	//
+	// FIXME: This interface seems to be
+	// fundamentally broken.  We should clarify CachingInodeOperation's
+	// behavior with metadata.
+	FD() int
+}
+
+// NewCachingInodeOperations returns a new CachingInodeOperations backed by
+// a CachedFileObject and its initial unstable attributes.
+func NewCachingInodeOperations(ctx context.Context, backingFile CachedFileObject, uattr fs.UnstableAttr, forcePageCache bool) *CachingInodeOperations {
+	p := platform.FromContext(ctx)
+	if p == nil {
+		panic(fmt.Sprintf("context.Context %T lacks non-nil value for key %T", ctx, platform.CtxPlatform))
+	}
+	return &CachingInodeOperations{
+		backingFile:    backingFile,
+		platform:       p,
+		forcePageCache: forcePageCache,
+		attr:           uattr,
+		hostFileMapper: NewHostFileMapper(),
+	}
+}
+
+// Release implements fs.InodeOperations.Release.
+func (c *CachingInodeOperations) Release() {
+	c.mapsMu.Lock()
+	defer c.mapsMu.Unlock()
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+	// The cache should be empty (something has gone terribly wrong if we're
+	// releasing an inode that is still memory-mapped).
+	if !c.mappings.IsEmpty() || !c.cache.IsEmpty() || !c.dirty.IsEmpty() {
+		panic(fmt.Sprintf("Releasing CachingInodeOperations with mappings:\n%s\ncache contents:\n%s\ndirty segments:\n%s", &c.mappings, &c.cache, &c.dirty))
+	}
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (c *CachingInodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+	return c.attr, nil
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (c *CachingInodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, perms fs.FilePermissions) bool {
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+
+	masked := fs.AttrMask{Perms: true}
+	if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Perms: perms}); err != nil {
+		return false
+	}
+	c.attr.Perms = perms
+	// FIXME: Clarify CachingInodeOperations behavior with metadata.
+	c.dirtyAttr.Perms = true
+	c.touchStatusChangeTimeLocked(ctx)
+	return true
+
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (c *CachingInodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	if !owner.UID.Ok() && !owner.GID.Ok() {
+		return nil
+	}
+
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+
+	masked := fs.AttrMask{
+		UID: owner.UID.Ok(),
+		GID: owner.GID.Ok(),
+	}
+	if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{Owner: owner}); err != nil {
+		return err
+	}
+	if owner.UID.Ok() {
+		c.attr.Owner.UID = owner.UID
+		// FIXME: Clarify CachingInodeOperations behavior with metadata.
+		c.dirtyAttr.UID = true
+	}
+	if owner.GID.Ok() {
+		c.attr.Owner.GID = owner.GID
+		// FIXME: Clarify CachingInodeOperations behavior with metadata.
+		c.dirtyAttr.GID = true
+	}
+	c.touchStatusChangeTimeLocked(ctx)
+	return nil
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (c *CachingInodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+
+	// Replace requests to use the "system time" with the current time to
+	// ensure that cached timestamps remain consistent with the remote
+	// filesystem.
+	now := ktime.NowFromContext(ctx)
+	if ts.ATimeSetSystemTime {
+		ts.ATime = now
+	}
+	if ts.MTimeSetSystemTime {
+		ts.MTime = now
+	}
+	masked := fs.AttrMask{
+		AccessTime:       !ts.ATimeOmit,
+		ModificationTime: !ts.MTimeOmit,
+	}
+	if err := c.backingFile.SetMaskedAttributes(ctx, masked, fs.UnstableAttr{AccessTime: ts.ATime, ModificationTime: ts.MTime}); err != nil {
+		return err
+	}
+	if !ts.ATimeOmit {
+		c.attr.AccessTime = ts.ATime
+		// FIXME: Clarify CachingInodeOperations behavior with metadata.
+		c.dirtyAttr.AccessTime = true
+	}
+	if !ts.MTimeOmit {
+		c.attr.ModificationTime = ts.MTime
+		// FIXME: Clarify CachingInodeOperations behavior with metadata.
+		c.dirtyAttr.ModificationTime = true
+	}
+	c.touchStatusChangeTimeLocked(ctx)
+	return nil
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (c *CachingInodeOperations) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+
+	// c.attr.Size is protected by both c.attrMu and c.dataMu.
+	c.dataMu.Lock()
+	if err := c.backingFile.SetMaskedAttributes(ctx, fs.AttrMask{
+		Size: true,
+	}, fs.UnstableAttr{
+		Size: size,
+	}); err != nil {
+		c.dataMu.Unlock()
+		return err
+	}
+	oldSize := c.attr.Size
+	if oldSize != size {
+		c.attr.Size = size
+		// FIXME: Clarify CachingInodeOperations behavior with metadata.
+		c.dirtyAttr.Size = true
+		c.touchModificationTimeLocked(ctx)
+	}
+	// We drop c.dataMu here so that we can lock c.mapsMu and invalidate
+	// mappings below. This allows concurrent calls to Read/Translate/etc.
+	// These functions synchronize with an in-progress Truncate by refusing to
+	// use cache contents beyond the new c.attr.Size. (We are still holding
+	// c.attrMu, so we can't race with Truncate/Write.)
+	c.dataMu.Unlock()
+
+	// Nothing left to do unless shrinking the file.
+	if size >= oldSize {
+		return nil
+	}
+
+	oldpgend := fs.OffsetPageEnd(oldSize)
+	newpgend := fs.OffsetPageEnd(size)
+
+	// Invalidate past translations of truncated pages.
+	if newpgend != oldpgend {
+		c.mapsMu.Lock()
+		c.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{
+			// Compare Linux's mm/truncate.c:truncate_setsize() =>
+			// truncate_pagecache() =>
+			// mm/memory.c:unmap_mapping_range(evencows=1).
+			InvalidatePrivate: true,
+		})
+		c.mapsMu.Unlock()
+	}
+
+	// We are now guaranteed that there are no translations of truncated pages,
+	// and can remove them from the cache. Since truncated pages have been
+	// removed from the backing file, they should be dropped without being
+	// written back.
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+	c.cache.Truncate(uint64(size), c.platform.Memory())
+	c.dirty.KeepClean(memmap.MappableRange{uint64(size), oldpgend})
+
+	return nil
+}
+
+// WriteOut implements fs.InodeOperations.WriteOut.
+func (c *CachingInodeOperations) WriteOut(ctx context.Context, inode *fs.Inode) error {
+	c.attrMu.Lock()
+
+	// Write dirty pages back.
+	c.dataMu.RLock()
+	err := SyncDirtyAll(ctx, &c.cache, &c.dirty, uint64(c.attr.Size), c.platform.Memory(), c.backingFile.WriteFromBlocksAt)
+	c.dataMu.RUnlock()
+	if err != nil {
+		c.attrMu.Unlock()
+		return err
+	}
+
+	// Write out cached attributes.
+	if err := c.backingFile.SetMaskedAttributes(ctx, c.dirtyAttr, c.attr); err != nil {
+		c.attrMu.Unlock()
+		return err
+	}
+	c.dirtyAttr = fs.AttrMask{}
+
+	c.attrMu.Unlock()
+
+	// Fsync the remote file.
+	return c.backingFile.Sync(ctx)
+}
+
+// IncLinks increases the link count and updates cached access time.
+func (c *CachingInodeOperations) IncLinks(ctx context.Context) {
+	c.attrMu.Lock()
+	c.attr.Links++
+	c.touchModificationTimeLocked(ctx)
+	c.attrMu.Unlock()
+}
+
+// DecLinks decreases the link count and updates cached access time.
+func (c *CachingInodeOperations) DecLinks(ctx context.Context) {
+	c.attrMu.Lock()
+	c.attr.Links--
+	c.touchModificationTimeLocked(ctx)
+	c.attrMu.Unlock()
+}
+
+// TouchAccessTime updates the cached access time in-place to the
+// current time. It does not update status change time in-place. See
+// mm/filemap.c:do_generic_file_read -> include/linux/h:file_accessed.
+func (c *CachingInodeOperations) TouchAccessTime(ctx context.Context, inode *fs.Inode) {
+	if inode.MountSource.Flags.NoAtime {
+		return
+	}
+
+	c.attrMu.Lock()
+	c.touchAccessTimeLocked(ctx)
+	c.attrMu.Unlock()
+}
+
+// touchAccesstimeLocked updates the cached access time in-place to the current
+// time.
+//
+// Preconditions: c.attrMu is locked for writing.
+func (c *CachingInodeOperations) touchAccessTimeLocked(ctx context.Context) {
+	c.attr.AccessTime = ktime.NowFromContext(ctx)
+	c.dirtyAttr.AccessTime = true
+}
+
+// TouchModificationTime updates the cached modification and status change time
+// in-place to the current time.
+func (c *CachingInodeOperations) TouchModificationTime(ctx context.Context) {
+	c.attrMu.Lock()
+	c.touchModificationTimeLocked(ctx)
+	c.attrMu.Unlock()
+}
+
+// touchModificationTimeLocked updates the cached modification and status
+// change time in-place to the current time.
+//
+// Preconditions: c.attrMu is locked for writing.
+func (c *CachingInodeOperations) touchModificationTimeLocked(ctx context.Context) {
+	now := ktime.NowFromContext(ctx)
+	c.attr.ModificationTime = now
+	c.dirtyAttr.ModificationTime = true
+	c.attr.StatusChangeTime = now
+	c.dirtyAttr.StatusChangeTime = true
+}
+
+// touchStatusChangeTimeLocked updates the cached status change time
+// in-place to the current time.
+//
+// Preconditions: c.attrMu is locked for writing.
+func (c *CachingInodeOperations) touchStatusChangeTimeLocked(ctx context.Context) {
+	now := ktime.NowFromContext(ctx)
+	c.attr.StatusChangeTime = now
+	c.dirtyAttr.StatusChangeTime = true
+}
+
+// Read reads from frames and otherwise directly from the backing file
+// into dst starting at offset until dst is full, EOF is reached, or an
+// error is encountered.
+//
+// Read may partially fill dst and return a nil error.
+func (c *CachingInodeOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Have we reached EOF? We check for this again in
+	// inodeReadWriter.ReadToBlocks to avoid holding c.attrMu (which would
+	// serialize reads) or c.dataMu (which would violate lock ordering), but
+	// check here first (before calling into MM) since reading at EOF is
+	// common: getting a return value of 0 from a read syscall is the only way
+	// to detect EOF.
+	//
+	// TODO: Separate out c.attr.Size and use atomics instead of
+	// c.dataMu.
+	c.dataMu.RLock()
+	size := c.attr.Size
+	c.dataMu.RUnlock()
+	if offset >= size {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOutFrom(ctx, &inodeReadWriter{ctx, c, offset})
+	// Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed().
+	c.TouchAccessTime(ctx, file.Dirent.Inode)
+	return n, err
+}
+
+// Write writes to frames and otherwise directly to the backing file
+// from src starting at offset and until src is empty or an error is
+// encountered.
+//
+// If Write partially fills src, a non-nil error is returned.
+func (c *CachingInodeOperations) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+	// Compare Linux's mm/filemap.c:__generic_file_write_iter() => file_update_time().
+	c.touchModificationTimeLocked(ctx)
+	return src.CopyInTo(ctx, &inodeReadWriter{ctx, c, offset})
+}
+
+type inodeReadWriter struct {
+	ctx    context.Context
+	c      *CachingInodeOperations
+	offset int64
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *inodeReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	rw.c.dataMu.RLock()
+	defer rw.c.dataMu.RUnlock()
+
+	// Compute the range to read.
+	if rw.offset >= rw.c.attr.Size {
+		return 0, io.EOF
+	}
+	end := fs.ReadEndOffset(rw.offset, int64(dsts.NumBytes()), rw.c.attr.Size)
+	if end == rw.offset { // dsts.NumBytes() == 0?
+		return 0, nil
+	}
+
+	mem := rw.c.platform.Memory()
+	var done uint64
+	seg, gap := rw.c.cache.Find(uint64(rw.offset))
+	for rw.offset < end {
+		mr := memmap.MappableRange{uint64(rw.offset), uint64(end)}
+		switch {
+		case seg.Ok():
+			// Get internal mappings from the cache.
+			ims, err := mem.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read)
+			if err != nil {
+				return done, err
+			}
+
+			// Copy from internal mappings.
+			n, err := safemem.CopySeq(dsts, ims)
+			done += n
+			rw.offset += int64(n)
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Read directly from the backing file.
+			gapmr := gap.Range().Intersect(mr)
+			dst := dsts.TakeFirst64(gapmr.Length())
+			n, err := rw.c.backingFile.ReadToBlocksAt(rw.ctx, dst, gapmr.Start)
+			done += n
+			rw.offset += int64(n)
+			dsts = dsts.DropFirst64(n)
+			// Partial reads are fine. But we must stop reading.
+			if n != dst.NumBytes() || err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = gap.NextSegment(), FileRangeGapIterator{}
+
+		default:
+			break
+		}
+	}
+	return done, nil
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+//
+// Preconditions: rw.c.attrMu must be locked.
+func (rw *inodeReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	rw.c.dataMu.Lock()
+	defer rw.c.dataMu.Unlock()
+
+	// Compute the range to write.
+	end := fs.WriteEndOffset(rw.offset, int64(srcs.NumBytes()))
+	if end == rw.offset { // srcs.NumBytes() == 0?
+		return 0, nil
+	}
+
+	defer func() {
+		// If the write ends beyond the file's previous size, it causes the
+		// file to grow.
+		if rw.offset > rw.c.attr.Size {
+			rw.c.attr.Size = rw.offset
+			rw.c.dirtyAttr.Size = true
+		}
+		if rw.offset > rw.c.attr.Usage {
+			// This is incorrect if CachingInodeOperations is caching a sparse
+			// file. (In Linux, keeping inode::i_blocks up to date is the
+			// filesystem's responsibility.)
+			rw.c.attr.Usage = rw.offset
+			rw.c.dirtyAttr.Usage = true
+		}
+	}()
+
+	mem := rw.c.platform.Memory()
+	var done uint64
+	seg, gap := rw.c.cache.Find(uint64(rw.offset))
+	for rw.offset < end {
+		mr := memmap.MappableRange{uint64(rw.offset), uint64(end)}
+		switch {
+		case seg.Ok() && seg.Start() < mr.End:
+			// Get internal mappings from the cache.
+			segMR := seg.Range().Intersect(mr)
+			ims, err := mem.MapInternal(seg.FileRangeOf(segMR), usermem.Write)
+			if err != nil {
+				return done, err
+			}
+
+			// Copy to internal mappings.
+			n, err := safemem.CopySeq(ims, srcs)
+			done += n
+			rw.offset += int64(n)
+			srcs = srcs.DropFirst64(n)
+			rw.c.dirty.MarkDirty(segMR)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok() && gap.Start() < mr.End:
+			// Write directly to the backing file.
+			gapmr := gap.Range().Intersect(mr)
+			src := srcs.TakeFirst64(gapmr.Length())
+			n, err := rw.c.backingFile.WriteFromBlocksAt(rw.ctx, src, gapmr.Start)
+			done += n
+			rw.offset += int64(n)
+			srcs = srcs.DropFirst64(n)
+			// Partial writes are fine. But we must stop writing.
+			if n != src.NumBytes() || err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = gap.NextSegment(), FileRangeGapIterator{}
+
+		default:
+			break
+		}
+	}
+	return done, nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (c *CachingInodeOperations) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	c.mapsMu.Lock()
+	defer c.mapsMu.Unlock()
+	mapped := c.mappings.AddMapping(ms, ar, offset)
+	// Do this unconditionally since whether we have c.backingFile.FD() >= 0
+	// can change across save/restore.
+	for _, r := range mapped {
+		c.hostFileMapper.IncRefOn(r)
+	}
+	if !usage.IncrementalMappedAccounting && !c.forcePageCache && c.backingFile.FD() >= 0 {
+		for _, r := range mapped {
+			usage.MemoryAccounting.Inc(r.Length(), usage.Mapped)
+		}
+	}
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (c *CachingInodeOperations) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+	c.mapsMu.Lock()
+	defer c.mapsMu.Unlock()
+	unmapped := c.mappings.RemoveMapping(ms, ar, offset)
+	for _, r := range unmapped {
+		c.hostFileMapper.DecRefOn(r)
+	}
+	if !c.forcePageCache && c.backingFile.FD() >= 0 {
+		if !usage.IncrementalMappedAccounting {
+			for _, r := range unmapped {
+				usage.MemoryAccounting.Dec(r.Length(), usage.Mapped)
+			}
+		}
+		return
+	}
+
+	// Writeback dirty mapped memory now that there are no longer any
+	// mappings that reference it. This is our naive memory eviction
+	// strategy.
+	mem := c.platform.Memory()
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+	for _, r := range unmapped {
+		if err := SyncDirty(ctx, r, &c.cache, &c.dirty, uint64(c.attr.Size), c.platform.Memory(), c.backingFile.WriteFromBlocksAt); err != nil {
+			log.Warningf("Failed to writeback cached data %v: %v", r, err)
+		}
+		c.cache.Drop(r, mem)
+		c.dirty.KeepClean(r)
+	}
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (c *CachingInodeOperations) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	return c.AddMapping(ctx, ms, dstAR, offset)
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (c *CachingInodeOperations) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	if !c.forcePageCache && c.backingFile.FD() >= 0 {
+		return []memmap.Translation{
+			{
+				Source: optional,
+				File:   c,
+				Offset: optional.Start,
+			},
+		}, nil
+	}
+
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+
+	// Constrain translations to c.attr.Size (rounded up) to prevent
+	// translation to pages that may be concurrently truncated.
+	pgend := fs.OffsetPageEnd(c.attr.Size)
+	var beyondEOF bool
+	if required.End > pgend {
+		if required.Start >= pgend {
+			return nil, &memmap.BusError{io.EOF}
+		}
+		beyondEOF = true
+		required.End = pgend
+	}
+	if optional.End > pgend {
+		optional.End = pgend
+	}
+
+	mem := c.platform.Memory()
+	cerr := c.cache.Fill(ctx, required, maxFillRange(required, optional), mem, usage.PageCache, c.backingFile.ReadToBlocksAt)
+
+	var ts []memmap.Translation
+	var translatedEnd uint64
+	for seg := c.cache.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() {
+		segMR := seg.Range().Intersect(optional)
+		ts = append(ts, memmap.Translation{
+			Source: segMR,
+			File:   mem,
+			Offset: seg.FileRangeOf(segMR).Start,
+		})
+		if at.Write {
+			// From this point forward, this memory can be dirtied through the
+			// mapping at any time.
+			c.dirty.KeepDirty(segMR)
+		}
+		translatedEnd = segMR.End
+	}
+
+	// Don't return the error returned by c.cache.Fill if it occurred outside
+	// of required.
+	if translatedEnd < required.End && cerr != nil {
+		return ts, &memmap.BusError{cerr}
+	}
+	if beyondEOF {
+		return ts, &memmap.BusError{io.EOF}
+	}
+	return ts, nil
+}
+
+func maxFillRange(required, optional memmap.MappableRange) memmap.MappableRange {
+	const maxReadahead = 64 << 10 // 64 KB, chosen arbitrarily
+	if required.Length() >= maxReadahead {
+		return required
+	}
+	if optional.Length() <= maxReadahead {
+		return optional
+	}
+	optional.Start = required.Start
+	if optional.Length() <= maxReadahead {
+		return optional
+	}
+	optional.End = optional.Start + maxReadahead
+	return optional
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (c *CachingInodeOperations) InvalidateUnsavable(ctx context.Context) error {
+	// Whether we have a host fd (and consequently what platform.File is
+	// mapped) can change across save/restore, so invalidate all translations
+	// unconditionally.
+	c.mapsMu.Lock()
+	defer c.mapsMu.Unlock()
+	c.mappings.InvalidateAll(memmap.InvalidateOpts{})
+
+	// Sync the cache's contents so that if we have a host fd after restore,
+	// the remote file's contents are coherent.
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+	if err := SyncDirtyAll(ctx, &c.cache, &c.dirty, uint64(c.attr.Size), c.platform.Memory(), c.backingFile.WriteFromBlocksAt); err != nil {
+		return err
+	}
+
+	// Discard the cache so that it's not stored in saved state. This is safe
+	// because per InvalidateUnsavable invariants, no new translations can have
+	// been returned after we invalidated all existing translations above.
+	c.cache.DropAll(c.platform.Memory())
+	c.dirty.RemoveAll()
+
+	return nil
+}
+
+// MapInto implements platform.File.MapInto. This is used when we directly map
+// an underlying host fd and CachingInodeOperations is used as the platform.File
+// during translation.
+func (c *CachingInodeOperations) MapInto(as platform.AddressSpace, addr usermem.Addr, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+	return as.MapFile(addr, c.backingFile.FD(), fr, at, precommit)
+}
+
+// MapInternal implements platform.File.MapInternal. This is used when we
+// directly map an underlying host fd and CachingInodeOperations is used as the
+// platform.File during translation.
+func (c *CachingInodeOperations) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) {
+	return c.hostFileMapper.MapInternal(fr, c.backingFile.FD(), at.Write)
+}
+
+// IncRef implements platform.File.IncRef. This is used when we directly map an
+// underlying host fd and CachingInodeOperations is used as the platform.File
+// during translation.
+func (c *CachingInodeOperations) IncRef(fr platform.FileRange) {
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+
+	seg, gap := c.refs.Find(fr.Start)
+	for {
+		switch {
+		case seg.Ok() && seg.Start() < fr.End:
+			seg = c.refs.Isolate(seg, fr)
+			seg.SetValue(seg.Value() + 1)
+			seg, gap = seg.NextNonEmpty()
+		case gap.Ok() && gap.Start() < fr.End:
+			newRange := gap.Range().Intersect(fr)
+			if usage.IncrementalMappedAccounting {
+				usage.MemoryAccounting.Inc(newRange.Length(), usage.Mapped)
+			}
+			seg, gap = c.refs.InsertWithoutMerging(gap, newRange, 1).NextNonEmpty()
+		default:
+			c.refs.MergeAdjacent(fr)
+			return
+		}
+	}
+}
+
+// DecRef implements platform.File.DecRef. This is used when we directly map an
+// underlying host fd and CachingInodeOperations is used as the platform.File
+// during translation.
+func (c *CachingInodeOperations) DecRef(fr platform.FileRange) {
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+
+	seg := c.refs.FindSegment(fr.Start)
+
+	for seg.Ok() && seg.Start() < fr.End {
+		seg = c.refs.Isolate(seg, fr)
+		if old := seg.Value(); old == 1 {
+			if usage.IncrementalMappedAccounting {
+				usage.MemoryAccounting.Dec(seg.Range().Length(), usage.Mapped)
+			}
+			seg = c.refs.Remove(seg).NextSegment()
+		} else {
+			seg.SetValue(old - 1)
+			seg = seg.NextSegment()
+		}
+	}
+	c.refs.MergeAdjacent(fr)
+}
diff --git a/pkg/sentry/fs/fsutil/inode_cached_test.go b/pkg/sentry/fs/fsutil/inode_cached_test.go
new file mode 100644
index 000000000..996c91849
--- /dev/null
+++ b/pkg/sentry/fs/fsutil/inode_cached_test.go
@@ -0,0 +1,403 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fsutil
+
+import (
+	"bytes"
+	"io"
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type noopBackingFile struct{}
+
+func (noopBackingFile) ReadToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) {
+	return dsts.NumBytes(), nil
+}
+
+func (noopBackingFile) WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) {
+	return srcs.NumBytes(), nil
+}
+
+func (noopBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr) error {
+	return nil
+}
+
+func (noopBackingFile) Sync(context.Context) error {
+	return nil
+}
+
+func (noopBackingFile) FD() int {
+	return -1
+}
+
+func TestSetPermissions(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	uattr := fs.WithCurrentTime(ctx, fs.UnstableAttr{
+		Perms: fs.FilePermsFromMode(0444),
+	})
+	iops := NewCachingInodeOperations(ctx, noopBackingFile{}, uattr, false /*forcePageCache*/)
+	defer iops.Release()
+
+	perms := fs.FilePermsFromMode(0777)
+	if !iops.SetPermissions(ctx, nil, perms) {
+		t.Fatalf("SetPermissions failed, want success")
+	}
+
+	// Did permissions change?
+	if !iops.dirtyAttr.Perms {
+		t.Fatalf("got perms not dirty, want dirty")
+	}
+	if iops.attr.Perms != perms {
+		t.Fatalf("got perms +%v, want +%v", iops.attr.Perms, perms)
+	}
+
+	// Did status change time change?
+	if !iops.dirtyAttr.StatusChangeTime {
+		t.Fatalf("got status change time not dirty, want dirty")
+	}
+	if iops.attr.StatusChangeTime.Equal(uattr.StatusChangeTime) {
+		t.Fatalf("got status change time unchanged")
+	}
+}
+
+func TestSetTimestamps(t *testing.T) {
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		desc      string
+		ts        fs.TimeSpec
+		wantDirty fs.AttrMask
+	}{
+		{
+			desc: "noop",
+			ts: fs.TimeSpec{
+				ATimeOmit: true,
+				MTimeOmit: true,
+			},
+			wantDirty: fs.AttrMask{},
+		},
+		{
+			desc: "access time only",
+			ts: fs.TimeSpec{
+				ATime:     ktime.NowFromContext(ctx),
+				MTimeOmit: true,
+			},
+			wantDirty: fs.AttrMask{
+				AccessTime:       true,
+				StatusChangeTime: true,
+			},
+		},
+		{
+			desc: "modification time only",
+			ts: fs.TimeSpec{
+				ATimeOmit: true,
+				MTime:     ktime.NowFromContext(ctx),
+			},
+			wantDirty: fs.AttrMask{
+				ModificationTime: true,
+				StatusChangeTime: true,
+			},
+		},
+		{
+			desc: "access and modification time",
+			ts: fs.TimeSpec{
+				ATime: ktime.NowFromContext(ctx),
+				MTime: ktime.NowFromContext(ctx),
+			},
+			wantDirty: fs.AttrMask{
+				AccessTime:       true,
+				ModificationTime: true,
+				StatusChangeTime: true,
+			},
+		},
+		{
+			desc: "system time access and modification time",
+			ts: fs.TimeSpec{
+				ATimeSetSystemTime: true,
+				MTimeSetSystemTime: true,
+			},
+			wantDirty: fs.AttrMask{
+				AccessTime:       true,
+				ModificationTime: true,
+				StatusChangeTime: true,
+			},
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			ctx := contexttest.Context(t)
+
+			epoch := ktime.ZeroTime
+			uattr := fs.UnstableAttr{
+				AccessTime:       epoch,
+				ModificationTime: epoch,
+				StatusChangeTime: epoch,
+			}
+			iops := NewCachingInodeOperations(ctx, noopBackingFile{}, uattr, false /*forcePageCache*/)
+			defer iops.Release()
+
+			if err := iops.SetTimestamps(ctx, nil, test.ts); err != nil {
+				t.Fatalf("SetTimestamps got error %v, want nil", err)
+			}
+			if !reflect.DeepEqual(iops.dirtyAttr, test.wantDirty) {
+				t.Fatalf("dirty got %+v, want %+v", iops.dirtyAttr, test.wantDirty)
+			}
+			if iops.dirtyAttr.AccessTime {
+				if !iops.attr.AccessTime.After(uattr.AccessTime) {
+					t.Fatalf("diritied access time did not advance, want %v > %v", iops.attr.AccessTime, uattr.AccessTime)
+				}
+				if !iops.dirtyAttr.StatusChangeTime {
+					t.Fatalf("dirty access time requires dirty status change time")
+				}
+				if !iops.attr.StatusChangeTime.After(uattr.StatusChangeTime) {
+					t.Fatalf("dirtied status change time did not advance")
+				}
+			}
+			if iops.dirtyAttr.ModificationTime {
+				if !iops.attr.ModificationTime.After(uattr.ModificationTime) {
+					t.Fatalf("diritied modification time did not advance")
+				}
+				if !iops.dirtyAttr.StatusChangeTime {
+					t.Fatalf("dirty modification time requires dirty status change time")
+				}
+				if !iops.attr.StatusChangeTime.After(uattr.StatusChangeTime) {
+					t.Fatalf("dirtied status change time did not advance")
+				}
+			}
+		})
+	}
+}
+
+func TestTruncate(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	uattr := fs.UnstableAttr{
+		Size: 0,
+	}
+	iops := NewCachingInodeOperations(ctx, noopBackingFile{}, uattr, false /*forcePageCache*/)
+	defer iops.Release()
+
+	if err := iops.Truncate(ctx, nil, uattr.Size); err != nil {
+		t.Fatalf("Truncate got error %v, want nil", err)
+	}
+	if iops.dirtyAttr.Size {
+		t.Fatalf("Truncate caused size to be dirtied")
+	}
+	var size int64 = 4096
+	if err := iops.Truncate(ctx, nil, size); err != nil {
+		t.Fatalf("Truncate got error %v, want nil", err)
+	}
+	if !iops.dirtyAttr.Size {
+		t.Fatalf("Truncate caused size to not be dirtied")
+	}
+	if iops.attr.Size != size {
+		t.Fatalf("Truncate got %d, want %d", iops.attr.Size, size)
+	}
+	if !iops.dirtyAttr.ModificationTime || !iops.dirtyAttr.StatusChangeTime {
+		t.Fatalf("Truncate did not dirty modification and status change time")
+	}
+	if !iops.attr.ModificationTime.After(uattr.ModificationTime) {
+		t.Fatalf("dirtied modification time did not change")
+	}
+	if !iops.attr.StatusChangeTime.After(uattr.StatusChangeTime) {
+		t.Fatalf("dirtied status change time did not change")
+	}
+}
+
+type sliceBackingFile struct {
+	data []byte
+}
+
+func newSliceBackingFile(data []byte) *sliceBackingFile {
+	return &sliceBackingFile{data}
+}
+
+func (f *sliceBackingFile) ReadToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) {
+	r := safemem.BlockSeqReader{safemem.BlockSeqOf(safemem.BlockFromSafeSlice(f.data)).DropFirst64(offset)}
+	return r.ReadToBlocks(dsts)
+}
+
+func (f *sliceBackingFile) WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) {
+	w := safemem.BlockSeqWriter{safemem.BlockSeqOf(safemem.BlockFromSafeSlice(f.data)).DropFirst64(offset)}
+	return w.WriteFromBlocks(srcs)
+}
+
+func (*sliceBackingFile) SetMaskedAttributes(context.Context, fs.AttrMask, fs.UnstableAttr) error {
+	return nil
+}
+
+func (*sliceBackingFile) Sync(context.Context) error {
+	return nil
+}
+
+func (*sliceBackingFile) FD() int {
+	return -1
+}
+
+type noopMappingSpace struct{}
+
+// Invalidate implements memmap.MappingSpace.Invalidate.
+func (noopMappingSpace) Invalidate(ar usermem.AddrRange, opts memmap.InvalidateOpts) {
+}
+
+func anonInode(ctx context.Context) *fs.Inode {
+	return fs.NewInode(NewSimpleInodeOperations(InodeSimpleAttributes{
+		UAttr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: fs.FileOwnerFromContext(ctx),
+			Perms: fs.FilePermissions{
+				User: fs.PermMask{Read: true, Write: true},
+			},
+			Links: 1,
+		}),
+	}), fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{}), fs.StableAttr{
+		Type:      fs.Anonymous,
+		BlockSize: usermem.PageSize,
+	})
+}
+
+func pagesOf(bs ...byte) []byte {
+	buf := make([]byte, 0, len(bs)*usermem.PageSize)
+	for _, b := range bs {
+		buf = append(buf, bytes.Repeat([]byte{b}, usermem.PageSize)...)
+	}
+	return buf
+}
+
+func TestRead(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	// Construct a 3-page file.
+	buf := pagesOf('a', 'b', 'c')
+	file := fs.NewFile(ctx, fs.NewDirent(anonInode(ctx), "anon"), fs.FileFlags{}, nil)
+	uattr := fs.UnstableAttr{
+		Size: int64(len(buf)),
+	}
+	iops := NewCachingInodeOperations(ctx, newSliceBackingFile(buf), uattr, false /*forcePageCache*/)
+	defer iops.Release()
+
+	// Expect the cache to be initially empty.
+	if cached := iops.cache.Span(); cached != 0 {
+		t.Errorf("Span got %d, want 0", cached)
+	}
+
+	// Create a memory mapping of the second page (as CachingInodeOperations
+	// expects to only cache mapped pages), then call Translate to force it to
+	// be cached.
+	var ms noopMappingSpace
+	ar := usermem.AddrRange{usermem.PageSize, 2 * usermem.PageSize}
+	if err := iops.AddMapping(ctx, ms, ar, usermem.PageSize); err != nil {
+		t.Fatalf("AddMapping got %v, want nil", err)
+	}
+	mr := memmap.MappableRange{usermem.PageSize, 2 * usermem.PageSize}
+	if _, err := iops.Translate(ctx, mr, mr, usermem.Read); err != nil {
+		t.Fatalf("Translate got %v, want nil", err)
+	}
+	if cached := iops.cache.Span(); cached != usermem.PageSize {
+		t.Errorf("SpanRange got %d, want %d", cached, usermem.PageSize)
+	}
+
+	// Try to read 4 pages. The first and third pages should be read directly
+	// from the "file", the second page should be read from the cache, and only
+	// 3 pages (the size of the file) should be readable.
+	rbuf := make([]byte, 4*usermem.PageSize)
+	dst := usermem.BytesIOSequence(rbuf)
+	n, err := iops.Read(ctx, file, dst, 0)
+	if n != 3*usermem.PageSize || (err != nil && err != io.EOF) {
+		t.Fatalf("Read got (%d, %v), want (%d, nil or EOF)", n, err, 3*usermem.PageSize)
+	}
+	rbuf = rbuf[:3*usermem.PageSize]
+
+	// Did we get the bytes we expect?
+	if !bytes.Equal(rbuf, buf) {
+		t.Errorf("Read back bytes %v, want %v", rbuf, buf)
+	}
+
+	// Delete the memory mapping and expect it to cause the cached page to be
+	// uncached.
+	iops.RemoveMapping(ctx, ms, ar, usermem.PageSize)
+	if cached := iops.cache.Span(); cached != 0 {
+		t.Fatalf("Span got %d, want 0", cached)
+	}
+}
+
+func TestWrite(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	// Construct a 4-page file.
+	buf := pagesOf('a', 'b', 'c', 'd')
+	orig := append([]byte(nil), buf...)
+	inode := anonInode(ctx)
+	uattr := fs.UnstableAttr{
+		Size: int64(len(buf)),
+	}
+	iops := NewCachingInodeOperations(ctx, newSliceBackingFile(buf), uattr, false /*forcePageCache*/)
+	defer iops.Release()
+
+	// Expect the cache to be initially empty.
+	if cached := iops.cache.Span(); cached != 0 {
+		t.Errorf("Span got %d, want 0", cached)
+	}
+
+	// Create a memory mapping of the second and third pages (as
+	// CachingInodeOperations expects to only cache mapped pages), then call
+	// Translate to force them to be cached.
+	var ms noopMappingSpace
+	ar := usermem.AddrRange{usermem.PageSize, 3 * usermem.PageSize}
+	if err := iops.AddMapping(ctx, ms, ar, usermem.PageSize); err != nil {
+		t.Fatalf("AddMapping got %v, want nil", err)
+	}
+	defer iops.RemoveMapping(ctx, ms, ar, usermem.PageSize)
+	mr := memmap.MappableRange{usermem.PageSize, 3 * usermem.PageSize}
+	if _, err := iops.Translate(ctx, mr, mr, usermem.Read); err != nil {
+		t.Fatalf("Translate got %v, want nil", err)
+	}
+	if cached := iops.cache.Span(); cached != 2*usermem.PageSize {
+		t.Errorf("SpanRange got %d, want %d", cached, 2*usermem.PageSize)
+	}
+
+	// Write to the first 2 pages.
+	wbuf := pagesOf('e', 'f')
+	src := usermem.BytesIOSequence(wbuf)
+	n, err := iops.Write(ctx, src, 0)
+	if n != 2*usermem.PageSize || err != nil {
+		t.Fatalf("Write got (%d, %v), want (%d, nil)", n, err, 2*usermem.PageSize)
+	}
+
+	// The first page should have been written directly, since it was not cached.
+	want := append([]byte(nil), orig...)
+	copy(want, pagesOf('e'))
+	if !bytes.Equal(buf, want) {
+		t.Errorf("File contents are %v, want %v", buf, want)
+	}
+
+	// Sync back to the "backing file".
+	if err := iops.WriteOut(ctx, inode); err != nil {
+		t.Errorf("Sync got %v, want nil", err)
+	}
+
+	// Now the second page should have been written as well.
+	copy(want[usermem.PageSize:], pagesOf('f'))
+	if !bytes.Equal(buf, want) {
+		t.Errorf("File contents are %v, want %v", buf, want)
+	}
+}
diff --git a/pkg/sentry/fs/g3doc/inotify.md b/pkg/sentry/fs/g3doc/inotify.md
new file mode 100644
index 000000000..1e99a3357
--- /dev/null
+++ b/pkg/sentry/fs/g3doc/inotify.md
@@ -0,0 +1,122 @@
+# Inotify
+
+Inotify implements the like-named filesystem event notification system for the
+sentry, see `inotify(7)`.
+
+## Architecture
+
+For the most part, the sentry implementation of inotify mirrors the Linux
+architecture. Inotify instances (i.e. the fd returned by inotify_init(2)) are
+backed by a pseudo-filesystem. Events are generated from various places in the
+sentry, including the [syscall layer][syscall_dir], the [vfs layer][dirent] and
+the [process fd table][fd_map]. Watches are stored in inodes and generated
+events are queued to the inotify instance owning the watches for delivery to the
+user.
+
+## Objects
+
+Here is a brief description of the existing and new objects involved in the
+sentry inotify mechanism, and how they interact:
+
+### [`fs.Inotify`][inotify]
+
+-   An inotify instances, created by inotify_init(2)/inotify_init1(2).
+-   The inotify fd has a `fs.Dirent`, supports filesystem syscalls to read
+    events.
+-   Has multiple `fs.Watch`es, with at most one watch per target inode, per
+    inotify instance.
+-   Has an instance `id` which is globally unique. This is *not* the fd number
+    for this instance, since the fd can be duped. This `id` is not externally
+    visible.
+
+### [`fs.Watch`][watch]
+
+-   An inotify watch, created/deleted by
+    inotify_add_watch(2)/inotify_rm_watch(2).
+-   Owned by an `fs.Inotify` instance, each watch keeps a pointer to the
+    `owner`.
+-   Associated with a single `fs.Inode`, which is the watch `target`. While the
+    watch is active, it indirectly pins `target` to memory. See the "Reference
+    Model" section for a detailed explanation.
+-   Filesystem operations on `target` generate `fs.Event`s.
+
+### [`fs.Event`][event]
+
+-   A simple struct encapsulating all the fields for an inotify event.
+-   Generated by `fs.Watch`es and forwarded to the watches' `owner`s.
+-   Serialized to the user during read(2) syscalls on the associated
+    `fs.Inotify`'s fd.
+
+### [`fs.Dirent`][dirent]
+
+-   Many inotify events are generated inside dirent methods. Events are
+    generated in the dirent methods rather than `fs.Inode` methods because some
+    events carry the name of the subject node, and node names are generally
+    unavailable in an `fs.Inode`.
+-   Dirents do not directly contain state for any watches. Instead, they forward
+    notifications to the underlying `fs.Inode`.
+
+### [`fs.Inode`][inode]
+
+-   Interacts with inotify through `fs.Watch`es.
+-   Inodes contain a map of all active `fs.Watch`es on them.
+-   An `fs.Inotify` instance can have at most one `fs.Watch` per inode.
+    `fs.Watch`es on an inode are indexed by their `owner`'s `id`.
+-   All inotify logic is encapsulated in the [`Watches`][inode_watches] struct
+    in an inode. Logically, `Watches` is the set of inotify watches on the
+    inode.
+
+## Reference Model
+
+The sentry inotify implementation has a complex reference model. An inotify
+watch observes a single inode. For efficient lookup, the state for a watch is
+stored directly on the target inode. This state needs to be persistent for the
+lifetime of watch. Unlike usual filesystem metadata, the watch state has no
+"on-disk" representation, so they cannot be reconstructed by the filesystem if
+the inode is flushed from memory. This effectively means we need to keep any
+inodes with actives watches pinned to memory.
+
+We can't just hold an extra ref on the inode to pin it to memory because some
+filesystems (such as gofer-based filesystems) don't have persistent inodes. In
+such a filesystem, if we just pin the inode, nothing prevents the enclosing
+dirent from being GCed. Once the dirent is GCed, the pinned inode is
+unreachable -- these filesystems generate a new inode by re-reading the node
+state on the next walk. Incidentally, hardlinks also don't work on these
+filesystems for this reason.
+
+To prevent the above scenario, when a new watch is added on an inode, we *pin*
+the dirent we used to reach the inode. Note that due to hardlinks, this dirent
+may not be the only dirent pointing to the inode. Attempting to set an inotify
+watch via multiple hardlinks to the same file results in the same watch being
+returned for both links. However, for each new dirent we use to reach the same
+inode, we add a new pin. We need a new pin for each new dirent used to reach the
+inode because we have no guarantees about the deletion order of the different
+links to the inode.
+
+## Lock Ordering
+
+There are 4 locks related to the inotify implementation:
+
+-   `Inotify.mu`: the inotify instance lock.
+-   `Inotify.evMu`: the inotify event queue lock.
+-   `Watch.mu`: the watch lock, used to protect pins.
+-   `fs.Watches.mu`: the inode watch set mu, used to protect the collection of
+    watches on the inode.
+
+The correct lock ordering for inotify code is:
+
+`Inotify.mu` -> `fs.Watches.mu` -> `Watch.mu` -> `Inotify.evMu`.
+
+We need a distinct lock for the event queue because by the time a goroutine
+attempts to queue a new event, it is already holding `fs.Watches.mu`. If we used
+`Inotify.mu` to also protect the event queue, this would violate the above lock
+ordering.
+
+[dirent]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/dirent.go
+[event]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/inotify_event.go
+[fd_map]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/kernel/fd_map.go
+[inode]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/inode.go
+[inode_watches]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/inode_inotify.go
+[inotify]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/inotify.go
+[syscall_dir]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/syscalls/linux/
+[watch]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/fs/inotify_watch.go
diff --git a/pkg/sentry/fs/gofer/BUILD b/pkg/sentry/fs/gofer/BUILD
new file mode 100644
index 000000000..ca42b0a54
--- /dev/null
+++ b/pkg/sentry/fs/gofer/BUILD
@@ -0,0 +1,90 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "gofer_state",
+    srcs = [
+        "file.go",
+        "file_state.go",
+        "fs.go",
+        "inode.go",
+        "inode_state.go",
+        "session.go",
+        "session_state.go",
+    ],
+    out = "gofer_state.go",
+    package = "gofer",
+)
+
+go_library(
+    name = "gofer",
+    srcs = [
+        "attr.go",
+        "context_file.go",
+        "device.go",
+        "file.go",
+        "file_state.go",
+        "fs.go",
+        "gofer_state.go",
+        "handles.go",
+        "inode.go",
+        "inode_state.go",
+        "path.go",
+        "session.go",
+        "session_state.go",
+        "socket.go",
+        "util.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/gofer",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/fd",
+        "//pkg/log",
+        "//pkg/metric",
+        "//pkg/p9",
+        "//pkg/refs",
+        "//pkg/secio",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fdpipe",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/host",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/unet",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "gofer_test",
+    size = "small",
+    srcs = ["gofer_test.go"],
+    embed = [":gofer"],
+    deps = [
+        "//pkg/log",
+        "//pkg/p9",
+        "//pkg/p9/p9test",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/unet",
+    ],
+)
diff --git a/pkg/sentry/fs/gofer/attr.go b/pkg/sentry/fs/gofer/attr.go
new file mode 100644
index 000000000..5e24767f9
--- /dev/null
+++ b/pkg/sentry/fs/gofer/attr.go
@@ -0,0 +1,162 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// getattr returns the 9p attributes of the p9.File. On success, Mode, Size, and RDev
+// are guaranteed to be masked as valid.
+func getattr(ctx context.Context, file contextFile) (p9.QID, p9.AttrMask, p9.Attr, error) {
+	// Retrieve attributes over the wire.
+	qid, valid, attr, err := file.getAttr(ctx, p9.AttrMaskAll())
+	if err != nil {
+		return qid, valid, attr, err
+	}
+
+	// Require mode, size, and raw device id.
+	if !valid.Mode || !valid.Size || !valid.RDev {
+		return qid, valid, attr, syscall.EIO
+	}
+
+	return qid, valid, attr, nil
+}
+
+func unstable(ctx context.Context, valid p9.AttrMask, pattr p9.Attr, mounter fs.FileOwner, client *p9.Client) fs.UnstableAttr {
+	return fs.UnstableAttr{
+		Size:             int64(pattr.Size),
+		Usage:            int64(pattr.Size),
+		Perms:            perms(valid, pattr, client),
+		Owner:            owner(mounter, valid, pattr),
+		AccessTime:       atime(ctx, valid, pattr),
+		ModificationTime: mtime(ctx, valid, pattr),
+		StatusChangeTime: ctime(ctx, valid, pattr),
+		Links:            links(valid, pattr),
+	}
+}
+
+func perms(valid p9.AttrMask, pattr p9.Attr, client *p9.Client) fs.FilePermissions {
+	if pattr.Mode.IsDir() && !p9.VersionSupportsMultiUser(client.Version()) {
+		// If user and group permissions bits are not supplied, use
+		// "other" bits to supplement them.
+		//
+		// Older Gofer's fake directories only have "other" permission,
+		// but will often be accessed via user or group permissions.
+		if pattr.Mode&0770 == 0 {
+			other := pattr.Mode & 07
+			pattr.Mode = pattr.Mode | other<<3 | other<<6
+		}
+	}
+	return fs.FilePermsFromP9(pattr.Mode)
+}
+
+func owner(mounter fs.FileOwner, valid p9.AttrMask, pattr p9.Attr) fs.FileOwner {
+	// Unless the file returned its UID and GID, it belongs to the mounting
+	// task's EUID/EGID.
+	owner := mounter
+	if valid.UID {
+		owner.UID = auth.KUID(pattr.UID)
+	}
+	if valid.GID {
+		owner.GID = auth.KGID(pattr.GID)
+	}
+	return owner
+}
+
+// bsize returns a block size from 9p attributes.
+func bsize(pattr p9.Attr) int64 {
+	if pattr.BlockSize > 0 {
+		return int64(pattr.BlockSize)
+	}
+	// Some files may have no clue of their block size. Better not to report
+	// something misleading or buggy and have a safe default.
+	return usermem.PageSize
+}
+
+// ntype returns an fs.InodeType from 9p attributes.
+func ntype(pattr p9.Attr) fs.InodeType {
+	switch {
+	case pattr.Mode.IsNamedPipe():
+		return fs.Pipe
+	case pattr.Mode.IsDir():
+		return fs.Directory
+	case pattr.Mode.IsSymlink():
+		return fs.Symlink
+	case pattr.Mode.IsCharacterDevice():
+		return fs.CharacterDevice
+	case pattr.Mode.IsBlockDevice():
+		return fs.BlockDevice
+	case pattr.Mode.IsSocket():
+		return fs.Socket
+	case pattr.Mode.IsRegular():
+		fallthrough
+	default:
+		return fs.RegularFile
+	}
+}
+
+// ctime returns a change time from 9p attributes.
+func ctime(ctx context.Context, valid p9.AttrMask, pattr p9.Attr) ktime.Time {
+	if valid.CTime {
+		return ktime.FromUnix(int64(pattr.CTimeSeconds), int64(pattr.CTimeNanoSeconds))
+	}
+	// Approximate ctime with mtime if ctime isn't available.
+	return mtime(ctx, valid, pattr)
+}
+
+// atime returns an access time from 9p attributes.
+func atime(ctx context.Context, valid p9.AttrMask, pattr p9.Attr) ktime.Time {
+	if valid.ATime {
+		return ktime.FromUnix(int64(pattr.ATimeSeconds), int64(pattr.ATimeNanoSeconds))
+	}
+	return ktime.NowFromContext(ctx)
+}
+
+// mtime returns a modification time from 9p attributes.
+func mtime(ctx context.Context, valid p9.AttrMask, pattr p9.Attr) ktime.Time {
+	if valid.MTime {
+		return ktime.FromUnix(int64(pattr.MTimeSeconds), int64(pattr.MTimeNanoSeconds))
+	}
+	return ktime.NowFromContext(ctx)
+}
+
+// links returns a hard link count from 9p attributes.
+func links(valid p9.AttrMask, pattr p9.Attr) uint64 {
+	// For gofer file systems that support link count (such as a local file gofer),
+	// we return the link count reported by the underlying file system.
+	if valid.NLink {
+		return pattr.NLink
+	}
+
+	// This node is likely backed by a file system that doesn't support links.
+	// We could readdir() and count children directories to provide an accurate
+	// link count. However this may be expensive since the gofer may be backed by remote
+	// storage. Instead, simply return 2 links for directories and 1 for everything else
+	// since no one relies on an accurate link count for gofer-based file systems.
+	switch ntype(pattr) {
+	case fs.Directory:
+		return 2
+	default:
+		return 1
+	}
+}
diff --git a/pkg/sentry/fs/gofer/context_file.go b/pkg/sentry/fs/gofer/context_file.go
new file mode 100644
index 000000000..d4b6f6eb7
--- /dev/null
+++ b/pkg/sentry/fs/gofer/context_file.go
@@ -0,0 +1,190 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextFile is a wrapper around p9.File that notifies the context that
+// it's about to sleep before calling the Gofer over P9.
+type contextFile struct {
+	file p9.File
+}
+
+func (c *contextFile) walk(ctx context.Context, names []string) ([]p9.QID, contextFile, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	q, f, err := c.file.Walk(names)
+	if err != nil {
+		return nil, contextFile{}, err
+	}
+	return q, contextFile{file: f}, nil
+}
+
+func (c *contextFile) statFS(ctx context.Context) (p9.FSStat, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.StatFS()
+}
+
+func (c *contextFile) getAttr(ctx context.Context, req p9.AttrMask) (p9.QID, p9.AttrMask, p9.Attr, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.GetAttr(req)
+}
+
+func (c *contextFile) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAttr) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.SetAttr(valid, attr)
+}
+
+func (c *contextFile) remove(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Remove()
+}
+
+func (c *contextFile) rename(ctx context.Context, directory contextFile, name string) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Rename(directory.file, name)
+}
+
+func (c *contextFile) close(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Close()
+}
+
+func (c *contextFile) open(ctx context.Context, mode p9.OpenFlags) (*fd.FD, p9.QID, uint32, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Open(mode)
+}
+
+func (c *contextFile) readAt(ctx context.Context, p []byte, offset uint64) (int, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.ReadAt(p, offset)
+}
+
+func (c *contextFile) writeAt(ctx context.Context, p []byte, offset uint64) (int, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.WriteAt(p, offset)
+}
+
+func (c *contextFile) fsync(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.FSync()
+}
+
+func (c *contextFile) create(ctx context.Context, name string, flags p9.OpenFlags, permissions p9.FileMode, uid p9.UID, gid p9.GID) (*fd.FD, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	fd, _, _, _, err := c.file.Create(name, flags, permissions, uid, gid)
+	return fd, err
+}
+
+func (c *contextFile) mkdir(ctx context.Context, name string, permissions p9.FileMode, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Mkdir(name, permissions, uid, gid)
+}
+
+func (c *contextFile) symlink(ctx context.Context, oldName string, newName string, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Symlink(oldName, newName, uid, gid)
+}
+
+func (c *contextFile) link(ctx context.Context, target *contextFile, newName string) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Link(target.file, newName)
+}
+
+func (c *contextFile) mknod(ctx context.Context, name string, permissions p9.FileMode, major uint32, minor uint32, uid p9.UID, gid p9.GID) (p9.QID, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Mknod(name, permissions, major, minor, uid, gid)
+}
+
+func (c *contextFile) unlinkAt(ctx context.Context, name string, flags uint32) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.UnlinkAt(name, flags)
+}
+
+func (c *contextFile) readdir(ctx context.Context, offset uint64, count uint32) ([]p9.Dirent, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Readdir(offset, count)
+}
+
+func (c *contextFile) readlink(ctx context.Context) (string, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Readlink()
+}
+
+func (c *contextFile) flush(ctx context.Context) error {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Flush()
+}
+
+func (c *contextFile) walkGetAttr(ctx context.Context, names []string) ([]p9.QID, contextFile, p9.AttrMask, p9.Attr, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	q, f, m, a, err := c.file.WalkGetAttr(names)
+	if err != nil {
+		return nil, contextFile{}, p9.AttrMask{}, p9.Attr{}, err
+	}
+	return q, contextFile{file: f}, m, a, nil
+}
+
+func (c *contextFile) connect(ctx context.Context, flags p9.ConnectFlags) (*fd.FD, error) {
+	ctx.UninterruptibleSleepStart(false)
+	defer ctx.UninterruptibleSleepFinish(false)
+
+	return c.file.Connect(flags)
+}
diff --git a/pkg/sentry/fs/gofer/device.go b/pkg/sentry/fs/gofer/device.go
new file mode 100644
index 000000000..fac7306d4
--- /dev/null
+++ b/pkg/sentry/fs/gofer/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// goferDevice is the gofer virtual device.
+var goferDevice = device.NewAnonMultiDevice()
diff --git a/pkg/sentry/fs/gofer/file.go b/pkg/sentry/fs/gofer/file.go
new file mode 100644
index 000000000..07c9bf01d
--- /dev/null
+++ b/pkg/sentry/fs/gofer/file.go
@@ -0,0 +1,255 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/metric"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+var openedWX = metric.MustCreateNewUint64Metric("/gofer/opened_write_execute_file", true /* sync */, "Number of times a writable+executable file was opened from a gofer.")
+
+// fileOperations implements fs.FileOperations for a remote file system.
+type fileOperations struct {
+	fsutil.NoIoctl     `state:"nosave"`
+	waiter.AlwaysReady `state:"nosave"`
+
+	// inodeOperations is the inodeOperations backing the file. It is protected
+	// by a reference held by File.Dirent.Inode which is stable until
+	// FileOperations.Release is called.
+	inodeOperations *inodeOperations `state:"wait"`
+
+	// dirCursor is the directory cursor.
+	dirCursor string
+
+	// handles are the opened remote file system handles, which may
+	// be shared with other files.
+	handles *handles `state:"nosave"`
+
+	// flags are the flags used to open handles.
+	flags fs.FileFlags `state:"wait"`
+}
+
+// fileOperations implements fs.FileOperations.
+var _ fs.FileOperations = (*fileOperations)(nil)
+
+// NewFile returns a file. NewFile is not appropriate with host pipes and sockets.
+func NewFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags, i *inodeOperations, handles *handles) *fs.File {
+	// Remote file systems enforce readability/writability at an offset,
+	// see fs/9p/vfs_inode.c:v9fs_vfs_atomic_open -> fs/open.c:finish_open.
+	flags.Pread = true
+	flags.Pwrite = true
+
+	f := &fileOperations{
+		inodeOperations: i,
+		handles:         handles,
+		flags:           flags,
+	}
+	if flags.Write {
+		if err := dirent.Inode.CheckPermission(ctx, fs.PermMask{Execute: true}); err == nil {
+			name, _ := dirent.FullName(fs.RootFromContext(ctx))
+			openedWX.Increment()
+			log.Warningf("Opened a writable executable: %q", name)
+		}
+	}
+	return fs.NewFile(ctx, dirent, flags, f)
+}
+
+// Release implements fs.FileOpeations.Release.
+func (f *fileOperations) Release() {
+	f.handles.DecRef()
+}
+
+// Readdir implements fs.FileOperations.Readdir.
+func (f *fileOperations) Readdir(ctx context.Context, file *fs.File, serializer fs.DentrySerializer) (int64, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+
+	dirCtx := &fs.DirCtx{
+		Serializer: serializer,
+		DirCursor:  &f.dirCursor,
+	}
+	n, err := fs.DirentReaddir(ctx, file.Dirent, f, root, dirCtx, file.Offset())
+	if f.inodeOperations.session().cachePolicy != cacheNone {
+		f.inodeOperations.cachingInodeOps.TouchAccessTime(ctx, file.Dirent.Inode)
+	}
+	return n, err
+}
+
+// IterateDir implements fs.DirIterator.IterateDir.
+func (f *fileOperations) IterateDir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	f.inodeOperations.readdirMu.Lock()
+	defer f.inodeOperations.readdirMu.Unlock()
+
+	// Fetch directory entries if needed.
+	if f.inodeOperations.readdirCache == nil || f.inodeOperations.session().cachePolicy == cacheNone {
+		entries, err := f.readdirAll(ctx)
+		if err != nil {
+			return offset, err
+		}
+
+		// Cache the readdir result.
+		f.inodeOperations.readdirCache = fs.NewSortedDentryMap(entries)
+	}
+
+	// Serialize the entries.
+	n, err := fs.GenericReaddir(dirCtx, f.inodeOperations.readdirCache)
+	return offset + n, err
+}
+
+// readdirAll fetches fs.DentAttrs for f, using the attributes of g.
+func (f *fileOperations) readdirAll(ctx context.Context) (map[string]fs.DentAttr, error) {
+	entries := make(map[string]fs.DentAttr)
+	var readOffset uint64
+	for {
+		// We choose some arbitrary high number of directory entries (64k) and call
+		// Readdir until we've exhausted them all.
+		dirents, err := f.handles.File.readdir(ctx, readOffset, 64*1024)
+		if err != nil {
+			return nil, err
+		}
+		if len(dirents) == 0 {
+			// We're done, we reached EOF.
+			break
+		}
+
+		// The last dirent contains the offset into the next set of dirents.  The gofer
+		// returns the offset as an index into directories, not as a byte offset, because
+		// converting a byte offset to an index into directories entries is a huge pain.
+		// But everything is fine if we're consistent.
+		readOffset = dirents[len(dirents)-1].Offset
+
+		for _, dirent := range dirents {
+			if dirent.Name == "." || dirent.Name == ".." {
+				// These must not be included in Readdir results.
+				continue
+			}
+
+			// Find a best approximation of the type.
+			var nt fs.InodeType
+			switch dirent.Type {
+			case p9.TypeDir:
+				nt = fs.Directory
+			case p9.TypeSymlink:
+				nt = fs.Symlink
+			default:
+				nt = fs.RegularFile
+			}
+
+			// Install the DentAttr.
+			entries[dirent.Name] = fs.DentAttr{
+				Type: nt,
+				// Construct the key to find the virtual inode.
+				// Directory entries reside on the same Device
+				// and SecondaryDevice as their parent.
+				InodeID: goferDevice.Map(device.MultiDeviceKey{
+					Device:          f.inodeOperations.fileState.key.Device,
+					SecondaryDevice: f.inodeOperations.fileState.key.SecondaryDevice,
+					Inode:           dirent.QID.Path,
+				}),
+			}
+		}
+	}
+
+	return entries, nil
+}
+
+// Write implements fs.FileOperations.Write.
+func (f *fileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	if fs.IsDir(file.Dirent.Inode.StableAttr) {
+		// Not all remote file systems enforce this so this client does.
+		return 0, syserror.EISDIR
+	}
+
+	// Do cached IO for regular files only. Some character devices expect no caching.
+	isFile := fs.IsFile(file.Dirent.Inode.StableAttr)
+	if f.inodeOperations.session().cachePolicy == cacheNone || !isFile {
+		return src.CopyInTo(ctx, f.handles.readWriterAt(ctx, offset))
+	}
+	return f.inodeOperations.cachingInodeOps.Write(ctx, src, offset)
+}
+
+// Read implements fs.FileOperations.Read.
+func (f *fileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	if fs.IsDir(file.Dirent.Inode.StableAttr) {
+		// Not all remote file systems enforce this so this client does.
+		return 0, syserror.EISDIR
+	}
+
+	// Do cached IO for regular files only. Some character devices expect no caching.
+	isFile := fs.IsFile(file.Dirent.Inode.StableAttr)
+	if f.inodeOperations.session().cachePolicy == cacheNone || !isFile {
+		return dst.CopyOutFrom(ctx, f.handles.readWriterAt(ctx, offset))
+	}
+	return f.inodeOperations.cachingInodeOps.Read(ctx, file, dst, offset)
+}
+
+// Fsync implements fs.FileOperations.Fsync.
+func (f *fileOperations) Fsync(ctx context.Context, file *fs.File, start int64, end int64, syncType fs.SyncType) error {
+	switch syncType {
+	case fs.SyncAll, fs.SyncData:
+		if err := file.Dirent.Inode.WriteOut(ctx); err != nil {
+			return err
+		}
+		fallthrough
+	case fs.SyncBackingStorage:
+		// Sync remote caches.
+		if f.handles.Host != nil {
+			// Sync the host fd directly.
+			return syscall.Fsync(f.handles.Host.FD())
+		}
+		// Otherwise sync on the p9.File handle.
+		return f.handles.File.fsync(ctx)
+	}
+	panic("invalid sync type")
+}
+
+// Flush implements fs.FileOperations.Flush.
+func (f *fileOperations) Flush(ctx context.Context, file *fs.File) error {
+	// If this file is not opened writable then there is nothing to flush.
+	// We do this because some p9 server implementations of Flush are
+	// over-zealous.
+	//
+	// FIXME: weaken these implementations and remove this check.
+	if !file.Flags().Write {
+		return nil
+	}
+	// Execute the flush.
+	return f.handles.File.flush(ctx)
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (f *fileOperations) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	if !isFileCachable(f.inodeOperations.session(), file.Dirent.Inode) {
+		return syserror.ENODEV
+	}
+	return fsutil.GenericConfigureMMap(file, f.inodeOperations.cachingInodeOps, opts)
+}
+
+// Seek implements fs.FileOperations.Seek.
+func (f *fileOperations) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	return fsutil.SeekWithDirCursor(ctx, file, whence, offset, &f.dirCursor)
+}
diff --git a/pkg/sentry/fs/gofer/file_state.go b/pkg/sentry/fs/gofer/file_state.go
new file mode 100644
index 000000000..1d63e33ec
--- /dev/null
+++ b/pkg/sentry/fs/gofer/file_state.go
@@ -0,0 +1,37 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// afterLoad is invoked by stateify.
+func (f *fileOperations) afterLoad() {
+	load := func() {
+		f.inodeOperations.fileState.waitForLoad()
+
+		// Manually load the open handles.
+		var err error
+		// TODO: Context is not plumbed to save/restore.
+		f.handles, err = newHandles(context.Background(), f.inodeOperations.fileState.file, f.flags)
+		if err != nil {
+			panic("failed to re-open handle: " + err.Error())
+		}
+		f.inodeOperations.fileState.setHandlesForCachedIO(f.flags, f.handles)
+	}
+	fs.Async(load)
+}
diff --git a/pkg/sentry/fs/gofer/fs.go b/pkg/sentry/fs/gofer/fs.go
new file mode 100644
index 000000000..0a1a49bbd
--- /dev/null
+++ b/pkg/sentry/fs/gofer/fs.go
@@ -0,0 +1,252 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package gofer implements a remote 9p filesystem.
+package gofer
+
+import (
+	"errors"
+	"fmt"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// The following are options defined by the Linux 9p client that we support,
+// see Documentation/filesystems/9p.txt.
+const (
+	// The transport method.
+	transportKey = "trans"
+
+	// The file tree to access when the file server
+	// is exporting several file systems. Stands for "attach name".
+	anameKey = "aname"
+
+	// The caching policy.
+	cacheKey = "cache"
+
+	// The file descriptor for reading with trans=fd.
+	readFDKey = "rfdno"
+
+	// The file descriptor for writing with trans=fd.
+	writeFDKey = "wfdno"
+
+	// The number of bytes to use for a 9p packet payload.
+	msizeKey = "msize"
+
+	// The 9p protocol version.
+	versionKey = "version"
+
+	// If set to true allows the creation of unix domain sockets inside the
+	// sandbox using files backed by the gofer. If set to false, unix sockets
+	// cannot be bound to gofer files without an overlay on top.
+	privateUnixSocketKey = "privateunixsocket"
+)
+
+// cachePolicy is a 9p cache policy.
+type cachePolicy string
+
+const (
+	// Use virtual file system cache.
+	cacheAll cachePolicy = "fscache"
+
+	// TODO: fully support cache=none.
+	cacheNone cachePolicy = "none"
+
+	// defaultCache is cacheAll. Note this diverges from the 9p Linux
+	// client whose default is "none".  See TODO above.
+	defaultCache = cacheAll
+)
+
+// defaultAname is the default attach name.
+const defaultAname = "/"
+
+// defaultMSize is the message size used for chunking large read and write requests.
+// This has been tested to give good enough performance up to 64M.
+const defaultMSize = 1024 * 1024 // 1M
+
+// defaultVersion is the default 9p protocol version. Will negotiate downwards with
+// file server if needed.
+var defaultVersion = p9.HighestVersionString()
+
+// Number of names of non-children to cache, preventing unneeded walks.  64 is
+// plenty for nodejs, which seems to stat about 4 children on every require().
+const nonChildrenCacheSize = 64
+
+var (
+	// ErrNoTransport is returned when there is no 'trans' option.
+	ErrNoTransport = errors.New("missing required option: 'trans='")
+
+	// ErrNoReadFD is returned when there is no 'rfdno' option.
+	ErrNoReadFD = errors.New("missing required option: 'rfdno='")
+
+	// ErrNoWriteFD is returned when there is no 'wfdno' option.
+	ErrNoWriteFD = errors.New("missing required option: 'wfdno='")
+)
+
+// filesystem is a 9p client.
+type filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&filesystem{})
+}
+
+// FilesystemName is the name under which the filesystem is registered.
+// The name matches fs/9p/vfs_super.c:v9fs_fs_type.name.
+const FilesystemName = "9p"
+
+// Name is the name of the filesystem.
+func (*filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount prohibits users from using mount(2) with this file system.
+func (*filesystem) AllowUserMount() bool {
+	return false
+}
+
+// Flags returns that there is nothing special about this file system.
+//
+// The 9p Linux client returns FS_RENAME_DOES_D_MOVE, see fs/9p/vfs_super.c.
+func (*filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns an attached 9p client that can be positioned in the vfs.
+func (f *filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// Parse and validate the mount options.
+	o, err := options(data)
+	if err != nil {
+		return nil, err
+	}
+
+	// Construct the 9p root to mount. We intentionally diverge from Linux in that
+	// the first Tversion and Tattach requests are done lazily.
+	return Root(ctx, device, f, flags, o)
+}
+
+// opts are parsed 9p mount options.
+type opts struct {
+	fd                int
+	aname             string
+	policy            cachePolicy
+	msize             uint32
+	version           string
+	privateunixsocket bool
+}
+
+// options parses mount(2) data into structured options.
+func options(data string) (opts, error) {
+	var o opts
+
+	// Parse generic comma-separated key=value options, this file system expects them.
+	options := fs.GenericMountSourceOptions(data)
+
+	// Check for the required 'trans=fd' option.
+	trans, ok := options[transportKey]
+	if !ok {
+		return o, ErrNoTransport
+	}
+	if trans != "fd" {
+		return o, fmt.Errorf("unsupported transport: 'trans=%s'", trans)
+	}
+	delete(options, transportKey)
+
+	// Check for the required 'rfdno=' option.
+	srfd, ok := options[readFDKey]
+	if !ok {
+		return o, ErrNoReadFD
+	}
+	delete(options, readFDKey)
+
+	// Check for the required 'wfdno=' option.
+	swfd, ok := options[writeFDKey]
+	if !ok {
+		return o, ErrNoWriteFD
+	}
+	delete(options, writeFDKey)
+
+	// Parse the read fd.
+	rfd, err := strconv.Atoi(srfd)
+	if err != nil {
+		return o, fmt.Errorf("invalid fd for 'rfdno=%s': %v", srfd, err)
+	}
+
+	// Parse the write fd.
+	wfd, err := strconv.Atoi(swfd)
+	if err != nil {
+		return o, fmt.Errorf("invalid fd for 'wfdno=%s': %v", swfd, err)
+	}
+
+	// Require that the read and write fd are the same.
+	if rfd != wfd {
+		return o, fmt.Errorf("fd in 'rfdno=%d' and 'wfdno=%d' must match", rfd, wfd)
+	}
+	o.fd = rfd
+
+	// Parse the attach name.
+	o.aname = defaultAname
+	if an, ok := options[anameKey]; ok {
+		o.aname = an
+		delete(options, anameKey)
+	}
+
+	// Parse the cache policy. Reject unsupported policies.
+	o.policy = cacheAll
+	if cp, ok := options[cacheKey]; ok {
+		if cachePolicy(cp) != cacheAll && cachePolicy(cp) != cacheNone {
+			return o, fmt.Errorf("unsupported cache mode: 'cache=%s'", cp)
+		}
+		o.policy = cachePolicy(cp)
+		delete(options, cacheKey)
+	}
+
+	// Parse the message size. Reject malformed options.
+	o.msize = uint32(defaultMSize)
+	if m, ok := options[msizeKey]; ok {
+		i, err := strconv.ParseUint(m, 10, 32)
+		if err != nil {
+			return o, fmt.Errorf("invalid message size for 'msize=%s': %v", m, err)
+		}
+		o.msize = uint32(i)
+		delete(options, msizeKey)
+	}
+
+	// Parse the protocol version.
+	o.version = defaultVersion
+	if v, ok := options[versionKey]; ok {
+		o.version = v
+		delete(options, versionKey)
+	}
+
+	// Parse the unix socket policy. Reject non-booleans.
+	if v, ok := options[privateUnixSocketKey]; ok {
+		b, err := strconv.ParseBool(v)
+		if err != nil {
+			return o, fmt.Errorf("invalid boolean value for '%s=%s': %v", privateUnixSocketKey, v, err)
+		}
+		o.privateunixsocket = b
+		delete(options, privateUnixSocketKey)
+	}
+
+	// Fail to attach if the caller wanted us to do something that we
+	// don't support.
+	if len(options) > 0 {
+		return o, fmt.Errorf("unsupported mount options: %v", options)
+	}
+
+	return o, nil
+}
diff --git a/pkg/sentry/fs/gofer/gofer_test.go b/pkg/sentry/fs/gofer/gofer_test.go
new file mode 100644
index 000000000..58a2e2ef5
--- /dev/null
+++ b/pkg/sentry/fs/gofer/gofer_test.go
@@ -0,0 +1,776 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/p9/p9test"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+)
+
+// A errMock is an error that comes from bad usage of the mock.
+var errMock = errors.New("mock error")
+
+// goodMockFile returns a file that can be Walk'ed to and created.
+func goodMockFile(mode p9.FileMode, size uint64) *p9test.FileMock {
+	return &p9test.FileMock{
+		GetAttrMock: p9test.GetAttrMock{
+			Valid: p9.AttrMask{Mode: true, Size: true, RDev: true},
+			Attr:  p9.Attr{Mode: mode, Size: size, RDev: 0},
+		},
+	}
+}
+
+func newClosedSocket() (*unet.Socket, error) {
+	fd, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		return nil, err
+	}
+
+	s, err := unet.NewSocket(fd)
+	if err != nil {
+		syscall.Close(fd)
+		return nil, err
+	}
+
+	return s, s.Close()
+}
+
+// root returns a p9 file mock and an fs.InodeOperations created from that file.  Any
+// functions performed on fs.InodeOperations will use the p9 file mock.
+func root(ctx context.Context, mode p9.FileMode, size uint64) (*p9test.FileMock, *fs.Inode, error) {
+	sock, err := newClosedSocket()
+	if err != nil {
+		return nil, nil, err
+	}
+
+	// Construct a dummy session that we can destruct.
+	s := &session{
+		conn:        sock,
+		mounter:     fs.RootOwner,
+		cachePolicy: cacheNone,
+	}
+
+	rootFile := goodMockFile(mode, size)
+	sattr, rootInodeOperations := newInodeOperations(ctx, s, contextFile{file: rootFile}, p9.QID{}, rootFile.GetAttrMock.Valid, rootFile.GetAttrMock.Attr)
+	m := fs.NewMountSource(s, &filesystem{}, fs.MountSourceFlags{})
+	return rootFile, fs.NewInode(rootInodeOperations, m, sattr), nil
+}
+
+func TestLookup(t *testing.T) {
+	// Test parameters.
+	type lookupTest struct {
+		// Name of the test.
+		name string
+
+		// Function input parameters.
+		fileName string
+
+		// Expected return value.
+		want error
+	}
+
+	tests := []lookupTest{
+		{
+			name:     "mock Walk passes (function succeeds)",
+			fileName: "ppp",
+			want:     nil,
+		},
+		{
+			name:     "mock Walk fails (function fails)",
+			fileName: "ppp",
+			want:     syscall.ENOENT,
+		},
+	}
+
+	ctx := contexttest.Context(t)
+	for _, test := range tests {
+		// Set up mock.
+		rootFile, rootInode, err := root(ctx, p9.PermissionsMask, 0)
+		if err != nil {
+			t.Errorf("TestWalk %s failed: root error got %v, want nil", test.name, err)
+		}
+
+		rootFile.WalkGetAttrMock.QIDs = []p9.QID{{}}
+		rootFile.WalkGetAttrMock.Err = test.want
+		rootFile.WalkGetAttrMock.File = goodMockFile(p9.PermissionsMask, 0)
+
+		// Call function.
+		dirent, err := rootInode.Lookup(ctx, test.fileName)
+
+		// Unwrap the InodeOperations.
+		var newInodeOperations fs.InodeOperations
+		if dirent != nil {
+			if dirent.IsNegative() {
+				err = syscall.ENOENT
+			} else {
+				newInodeOperations = dirent.Inode.InodeOperations
+			}
+		}
+
+		// Check return values.
+		if err != test.want {
+			t.Errorf("TestWalk %s failed: got %v, want %v", test.name, err, test.want)
+		}
+		if err == nil && newInodeOperations == nil {
+			t.Errorf("TestWalk %s failed: expected either non-nil err or non-nil node, but both are nil", test.name)
+		}
+
+		// Check mock parameters.
+		if !rootFile.WalkGetAttrMock.Called {
+			t.Errorf("TestWalk %s failed: GetAttr not called; error: %v", test.name, err)
+		} else if rootFile.WalkGetAttrMock.Names[0] != test.fileName {
+			t.Errorf("TestWalk %s failed: file name not set", test.name)
+		}
+	}
+}
+
+func TestSetTimestamps(t *testing.T) {
+	// Test parameters.
+	type setTimestampsTest struct {
+		// Name of the test.
+		name string
+
+		// Function input parameters.
+		ts fs.TimeSpec
+	}
+
+	ctx := contexttest.Context(t)
+	now := ktime.NowFromContext(ctx)
+	tests := []setTimestampsTest{
+		{
+			name: "mock SetAttr passes (function succeeds)",
+			ts: fs.TimeSpec{
+				ATime: now,
+				MTime: now,
+			},
+		},
+		{
+			name: "mock SetAttr passes, times are 0 (function succeeds)",
+			ts:   fs.TimeSpec{},
+		},
+		{
+			name: "mock SetAttr passes, times are 0 and not system time (function succeeds)",
+			ts: fs.TimeSpec{
+				ATimeSetSystemTime: false,
+				MTimeSetSystemTime: false,
+			},
+		},
+		{
+			name: "mock SetAttr passes, times are set to system time (function succeeds)",
+			ts: fs.TimeSpec{
+				ATimeSetSystemTime: true,
+				MTimeSetSystemTime: true,
+			},
+		},
+		{
+			name: "mock SetAttr passes, times are omitted (function succeeds)",
+			ts: fs.TimeSpec{
+				ATimeOmit: true,
+				MTimeOmit: true,
+			},
+		},
+	}
+
+	for _, test := range tests {
+		// Set up mock.
+		rootFile, rootInode, err := root(ctx, p9.PermissionsMask, 0)
+		if err != nil {
+			t.Errorf("TestSetTimestamps %s failed: root error got %v, want nil", test.name, err)
+		}
+
+		// Call function.
+		err = rootInode.SetTimestamps(ctx, nil /* Dirent */, test.ts)
+
+		// Check return values.
+		if err != nil {
+			t.Errorf("TestSetTimestamps %s failed: got %v, want nil", test.name, err)
+		}
+
+		// Check mock parameters.
+		if !(test.ts.ATimeOmit && test.ts.MTimeOmit) && !rootFile.SetAttrMock.Called {
+			t.Errorf("TestSetTimestamps %s failed: SetAttr not called", test.name)
+			continue
+		}
+
+		// Check what was passed to the mock function.
+		attr := rootFile.SetAttrMock.Attr
+		atimeGiven := ktime.FromUnix(int64(attr.ATimeSeconds), int64(attr.ATimeNanoSeconds))
+		if test.ts.ATimeOmit {
+			if rootFile.SetAttrMock.Valid.ATime {
+				t.Errorf("TestSetTimestamps %s failed: ATime got set true in mask, wanted false", test.name)
+			}
+		} else {
+			if got, want := rootFile.SetAttrMock.Valid.ATimeNotSystemTime, !test.ts.ATimeSetSystemTime; got != want {
+				t.Errorf("TestSetTimestamps %s failed: got ATimeNotSystemTime %v, want %v", test.name, got, want)
+			}
+			if !test.ts.ATimeSetSystemTime && !test.ts.ATime.Equal(atimeGiven) {
+				t.Errorf("TestSetTimestamps %s failed: ATime got %v, want %v", test.name, atimeGiven, test.ts.ATime)
+			}
+		}
+
+		mtimeGiven := ktime.FromUnix(int64(attr.MTimeSeconds), int64(attr.MTimeNanoSeconds))
+		if test.ts.MTimeOmit {
+			if rootFile.SetAttrMock.Valid.MTime {
+				t.Errorf("TestSetTimestamps %s failed: MTime got set true in mask, wanted false", test.name)
+			}
+		} else {
+			if got, want := rootFile.SetAttrMock.Valid.MTimeNotSystemTime, !test.ts.MTimeSetSystemTime; got != want {
+				t.Errorf("TestSetTimestamps %s failed: got MTimeNotSystemTime %v, want %v", test.name, got, want)
+			}
+			if !test.ts.MTimeSetSystemTime && !test.ts.MTime.Equal(mtimeGiven) {
+				t.Errorf("TestSetTimestamps %s failed: MTime got %v, want %v", test.name, mtimeGiven, test.ts.MTime)
+			}
+		}
+
+	}
+}
+
+func TestSetPermissions(t *testing.T) {
+	// Test parameters.
+	type setPermissionsTest struct {
+		// Name of the test.
+		name string
+
+		// SetPermissions input parameters.
+		perms fs.FilePermissions
+
+		// Error that SetAttr mock should return.
+		setAttrErr error
+
+		// Expected return value.
+		want bool
+	}
+
+	tests := []setPermissionsTest{
+		{
+			name:       "SetAttr mock succeeds (function succeeds)",
+			perms:      fs.FilePermissions{User: fs.PermMask{Read: true, Write: true, Execute: true}},
+			want:       true,
+			setAttrErr: nil,
+		},
+		{
+			name:       "SetAttr mock fails (function fails)",
+			perms:      fs.FilePermissions{User: fs.PermMask{Read: true, Write: true}},
+			want:       false,
+			setAttrErr: syscall.ENOENT,
+		},
+	}
+
+	ctx := contexttest.Context(t)
+	for _, test := range tests {
+		// Set up mock.
+		rootFile, rootInode, err := root(ctx, 0, 0)
+		if err != nil {
+			t.Errorf("TestSetPermissions %s failed: root error got %v, want nil", test.name, err)
+		}
+		rootFile.SetAttrMock.Err = test.setAttrErr
+
+		ok := rootInode.SetPermissions(ctx, nil /* Dirent */, test.perms)
+
+		// Check return value.
+		if ok != test.want {
+			t.Errorf("TestSetPermissions %s failed: got %v, want %v", test.name, ok, test.want)
+		}
+
+		// Check mock parameters.
+		pattr := rootFile.SetAttrMock.Attr
+		if !rootFile.SetAttrMock.Called {
+			t.Errorf("TestSetPermissions %s failed: SetAttr not called", test.name)
+			continue
+		}
+		if !rootFile.SetAttrMock.Valid.Permissions {
+			t.Errorf("TestSetPermissions %s failed: SetAttr did not get right request (got false, expected SetAttrMask.Permissions true)",
+				test.name)
+		}
+		if got := fs.FilePermsFromP9(pattr.Permissions); got != test.perms {
+			t.Errorf("TestSetPermissions %s failed: SetAttr did not get right permissions -- got %v, want %v",
+				test.name, got, test.perms)
+		}
+	}
+}
+
+func TestClose(t *testing.T) {
+	ctx := contexttest.Context(t)
+	// Set up mock.
+	rootFile, rootInode, err := root(ctx, p9.PermissionsMask, 0)
+	if err != nil {
+		t.Errorf("TestClose failed: root error got %v, want nil", err)
+	}
+
+	// Call function.
+	rootInode.InodeOperations.Release(ctx)
+
+	// Check mock parameters.
+	if !rootFile.CloseMock.Called {
+		t.Errorf("TestClose failed: Close not called")
+	}
+}
+
+func TestRename(t *testing.T) {
+	// Test parameters.
+	type renameTest struct {
+		// Name of the test.
+		name string
+
+		// Input parameters.
+		newParent *fs.Inode
+		newName   string
+
+		// Rename mock parameters.
+		renameErr    error
+		renameCalled bool
+
+		// Error want to return given the parameters. (Same as what
+		// we expect and tell rename to return.)
+		want error
+	}
+	ctx := contexttest.Context(t)
+	rootFile, rootInode, err := root(ctx, p9.PermissionsMask, 0)
+	if err != nil {
+		t.Errorf("TestRename failed: root error got %v, want nil", err)
+	}
+
+	tests := []renameTest{
+		{
+			name:         "mock Rename succeeds (function succeeds)",
+			newParent:    rootInode,
+			newName:      "foo2",
+			want:         nil,
+			renameErr:    nil,
+			renameCalled: true,
+		},
+		{
+			name:         "mock Rename fails (function fails)",
+			newParent:    rootInode,
+			newName:      "foo2",
+			want:         syscall.ENOENT,
+			renameErr:    syscall.ENOENT,
+			renameCalled: true,
+		},
+		{
+			name:         "newParent is not inodeOperations but should be (function fails)",
+			newParent:    fs.NewMockInode(ctx, fs.NewMockMountSource(nil), fs.StableAttr{Type: fs.Directory}),
+			newName:      "foo2",
+			want:         syscall.EXDEV,
+			renameErr:    nil,
+			renameCalled: false,
+		},
+	}
+
+	for _, test := range tests {
+		mockFile := goodMockFile(p9.PermissionsMask, 0)
+		rootFile.WalkGetAttrMock.QIDs = []p9.QID{{}}
+		rootFile.WalkGetAttrMock.File = mockFile
+
+		dirent, err := rootInode.Lookup(ctx, "foo")
+		if err != nil {
+			t.Fatalf("root.Walk failed: %v", err)
+		}
+		mockFile.RenameMock.Err = test.renameErr
+		mockFile.RenameMock.Called = false
+
+		// Use a dummy oldParent to acquire write access to that directory.
+		oldParent := &inodeOperations{
+			readdirCache: fs.NewSortedDentryMap(nil),
+		}
+		oldInode := fs.NewInode(oldParent, fs.NewMockMountSource(nil), fs.StableAttr{Type: fs.Directory})
+
+		// Call function.
+		err = dirent.Inode.InodeOperations.Rename(ctx, oldInode, "", test.newParent, test.newName)
+
+		// Check return value.
+		if err != test.want {
+			t.Errorf("TestRename %s failed: got %v, want %v", test.name, err, test.want)
+		}
+
+		// Check mock parameters.
+		if got, want := mockFile.RenameMock.Called, test.renameCalled; got != want {
+			t.Errorf("TestRename %s failed: renameCalled got %v want %v", test.name, got, want)
+		}
+	}
+}
+
+// This file is read from in TestPreadv.
+type readAtFileFake struct {
+	p9test.FileMock
+
+	// Parameters for faking ReadAt.
+	FileLength int
+	Err        error
+	ChunkSize  int
+	Called     bool
+	LengthRead int
+}
+
+func (r *readAtFileFake) ReadAt(p []byte, offset uint64) (int, error) {
+	r.Called = true
+	log.Warningf("ReadAt fake: length read so far = %d, len(p) = %d, offset = %d", r.LengthRead, len(p), offset)
+	if int(offset) != r.LengthRead {
+		return 0, fmt.Errorf("offset got %d; expected %d", offset, r.LengthRead)
+	}
+
+	if r.Err != nil {
+		return 0, r.Err
+	}
+
+	if r.LengthRead >= r.FileLength {
+		return 0, io.EOF
+	}
+
+	// Read at most ChunkSize and read at most what's left in the file.
+	toBeRead := len(p)
+	if r.LengthRead+toBeRead >= r.FileLength {
+		toBeRead = r.FileLength - int(offset)
+	}
+	if toBeRead > r.ChunkSize {
+		toBeRead = r.ChunkSize
+	}
+
+	r.LengthRead += toBeRead
+	if r.LengthRead == r.FileLength {
+		return toBeRead, io.EOF
+	}
+	return toBeRead, nil
+}
+
+func TestPreadv(t *testing.T) {
+	// Test parameters.
+	type preadvTest struct {
+		// Name of the test.
+		name string
+
+		// Mock parameters
+		mode p9.FileMode
+
+		// Buffer to read into.
+		buffer    [512]byte
+		sliceSize int
+
+		// How much readAt returns at a time.
+		chunkSize int
+
+		// Whether or not we expect ReadAt to be called.
+		readAtCalled bool
+		readAtErr    error
+
+		// Expected return values.
+		want error
+	}
+
+	tests := []preadvTest{
+		{
+			name:         "fake ReadAt succeeds, 512 bytes requested, 512 byte chunks (function succeeds)",
+			want:         nil,
+			readAtErr:    nil,
+			mode:         p9.PermissionsMask,
+			readAtCalled: true,
+			sliceSize:    512,
+			chunkSize:    512,
+		},
+		{
+			name:         "fake ReadAt succeeds, 512 bytes requested, 200 byte chunks (function succeeds)",
+			want:         nil,
+			readAtErr:    nil,
+			mode:         p9.PermissionsMask,
+			readAtCalled: true,
+			sliceSize:    512,
+			chunkSize:    200,
+		},
+		{
+			name:         "fake ReadAt succeeds, 0 bytes requested (function succeeds)",
+			want:         nil,
+			readAtErr:    nil,
+			mode:         p9.PermissionsMask,
+			readAtCalled: false,
+			sliceSize:    0,
+			chunkSize:    100,
+		},
+		{
+			name:         "fake ReadAt returns 0 bytes and EOF (function fails)",
+			want:         io.EOF,
+			readAtErr:    io.EOF,
+			mode:         p9.PermissionsMask,
+			readAtCalled: true,
+			sliceSize:    512,
+			chunkSize:    512,
+		},
+	}
+
+	ctx := contexttest.Context(t)
+	for _, test := range tests {
+		// Set up mock.
+		rootFile, rootInode, err := root(ctx, test.mode, 1024)
+		if err != nil {
+			t.Errorf("TestPreadv %s failed: root error got %v, want nil", test.name, err)
+		}
+
+		// Set up the read buffer.
+		dst := usermem.BytesIOSequence(test.buffer[:test.sliceSize])
+
+		// This file will be read from.
+		openFile := &readAtFileFake{
+			Err:        test.readAtErr,
+			FileLength: test.sliceSize,
+			ChunkSize:  test.chunkSize,
+		}
+		rootFile.WalkGetAttrMock.File = openFile
+		rootFile.WalkGetAttrMock.Attr.Mode = test.mode
+		rootFile.WalkGetAttrMock.Valid.Mode = true
+
+		f := NewFile(
+			ctx,
+			fs.NewDirent(rootInode, ""),
+			fs.FileFlags{Read: true},
+			rootInode.InodeOperations.(*inodeOperations),
+			&handles{File: contextFile{file: openFile}},
+		)
+
+		// Call function.
+		_, err = f.Preadv(ctx, dst, 0)
+
+		// Check return value.
+		if err != test.want {
+			t.Errorf("TestPreadv %s failed: got %v, want %v", test.name, err, test.want)
+		}
+
+		// Check mock parameters.
+		if test.readAtCalled != openFile.Called {
+			t.Errorf("TestPreadv %s failed: ReadAt called: %v, but expected opposite", test.name, openFile.Called)
+		}
+	}
+}
+
+func TestReadlink(t *testing.T) {
+	// Test parameters.
+	type readlinkTest struct {
+		// Name of the test.
+		name string
+
+		// Mock parameters
+		mode p9.FileMode
+
+		// Whether or not we expect ReadAt to be called and what error
+		// it shall return.
+		readlinkCalled bool
+		readlinkErr    error
+
+		// Expected return values.
+		want error
+	}
+
+	tests := []readlinkTest{
+		{
+			name:           "file is not symlink (function fails)",
+			want:           syscall.ENOLINK,
+			mode:           p9.PermissionsMask,
+			readlinkCalled: false,
+			readlinkErr:    nil,
+		},
+		{
+			name:           "mock Readlink succeeds (function succeeds)",
+			want:           nil,
+			mode:           p9.PermissionsMask | p9.ModeSymlink,
+			readlinkCalled: true,
+			readlinkErr:    nil,
+		},
+		{
+			name:           "mock Readlink fails (function fails)",
+			want:           syscall.ENOENT,
+			mode:           p9.PermissionsMask | p9.ModeSymlink,
+			readlinkCalled: true,
+			readlinkErr:    syscall.ENOENT,
+		},
+	}
+
+	ctx := contexttest.Context(t)
+	for _, test := range tests {
+		// Set up mock.
+		rootFile, rootInode, err := root(ctx, test.mode, 0)
+		if err != nil {
+			t.Errorf("TestReadlink %s failed: root error got %v, want nil", test.name, err)
+		}
+
+		openFile := goodMockFile(test.mode, 0)
+		rootFile.WalkMock.File = openFile
+		rootFile.ReadlinkMock.Err = test.readlinkErr
+
+		// Call function.
+		_, err = rootInode.Readlink(ctx)
+
+		// Check return value.
+		if err != test.want {
+			t.Errorf("TestReadlink %s failed: got %v, want %v", test.name, err, test.want)
+		}
+
+		// Check mock parameters.
+		if test.readlinkCalled && !rootFile.ReadlinkMock.Called {
+			t.Errorf("TestReadlink %s failed: Readlink not called", test.name)
+		}
+	}
+}
+
+// This file is write from in TestPwritev.
+type writeAtFileFake struct {
+	p9test.FileMock
+
+	// Parameters for faking WriteAt.
+	Err           error
+	ChunkSize     int
+	Called        bool
+	LengthWritten int
+}
+
+func (r *writeAtFileFake) WriteAt(p []byte, offset uint64) (int, error) {
+	r.Called = true
+	log.Warningf("WriteAt fake: length written so far = %d, len(p) = %d, offset = %d", r.LengthWritten, len(p), offset)
+	if int(offset) != r.LengthWritten {
+		return 0, fmt.Errorf("offset got %d; want %d", offset, r.LengthWritten)
+	}
+
+	if r.Err != nil {
+		return 0, r.Err
+	}
+
+	// Write at most ChunkSize.
+	toBeWritten := len(p)
+	if toBeWritten > r.ChunkSize {
+		toBeWritten = r.ChunkSize
+	}
+	r.LengthWritten += toBeWritten
+	return toBeWritten, nil
+}
+
+func TestPwritev(t *testing.T) {
+	// Test parameters.
+	type pwritevTest struct {
+		// Name of the test.
+		name string
+
+		// Mock parameters
+		mode p9.FileMode
+
+		allowWrite bool
+
+		// Buffer to write into.
+		buffer    [512]byte
+		sliceSize int
+		chunkSize int
+
+		// Whether or not we expect writeAt to be called.
+		writeAtCalled bool
+		writeAtErr    error
+
+		// Expected return values.
+		want error
+	}
+
+	tests := []pwritevTest{
+		{
+			name:          "fake writeAt succeeds, one chunk (function succeeds)",
+			want:          nil,
+			writeAtErr:    nil,
+			mode:          p9.PermissionsMask,
+			allowWrite:    true,
+			writeAtCalled: true,
+			sliceSize:     512,
+			chunkSize:     512,
+		},
+		{
+			name:          "fake writeAt fails, short write (function fails)",
+			want:          io.ErrShortWrite,
+			writeAtErr:    nil,
+			mode:          p9.PermissionsMask,
+			allowWrite:    true,
+			writeAtCalled: true,
+			sliceSize:     512,
+			chunkSize:     200,
+		},
+		{
+			name:          "fake writeAt succeeds, len 0 (function succeeds)",
+			want:          nil,
+			writeAtErr:    nil,
+			mode:          p9.PermissionsMask,
+			allowWrite:    true,
+			writeAtCalled: false,
+			sliceSize:     0,
+			chunkSize:     0,
+		},
+		{
+			name:          "writeAt can still write despite file permissions read only (function succeeds)",
+			want:          nil,
+			writeAtErr:    nil,
+			mode:          p9.PermissionsMask,
+			allowWrite:    false,
+			writeAtCalled: true,
+			sliceSize:     512,
+			chunkSize:     512,
+		},
+	}
+
+	ctx := contexttest.Context(t)
+	for _, test := range tests {
+		// Set up mock.
+		_, rootInode, err := root(ctx, test.mode, 0)
+		if err != nil {
+			t.Errorf("TestPwritev %s failed: root error got %v, want nil", test.name, err)
+		}
+
+		src := usermem.BytesIOSequence(test.buffer[:test.sliceSize])
+
+		// This is the file that will be used for writing.
+		openFile := &writeAtFileFake{
+			Err:       test.writeAtErr,
+			ChunkSize: test.chunkSize,
+		}
+
+		f := NewFile(
+			ctx,
+			fs.NewDirent(rootInode, ""),
+			fs.FileFlags{Write: true},
+			rootInode.InodeOperations.(*inodeOperations),
+			&handles{File: contextFile{file: openFile}},
+		)
+
+		// Call function.
+		_, err = f.Pwritev(ctx, src, 0)
+
+		// Check return value.
+		if err != test.want {
+			t.Errorf("TestPwritev %s failed: got %v, want %v", test.name, err, test.want)
+		}
+
+		// Check mock parameters.
+		if test.writeAtCalled != openFile.Called {
+			t.Errorf("TestPwritev %s failed: WriteAt called: %v, but expected opposite", test.name, openFile.Called)
+			continue
+		}
+		if openFile.Called && test.writeAtErr != nil && openFile.LengthWritten != test.sliceSize {
+			t.Errorf("TestPwritev %s failed: wrote %d bytes, expected %d bytes written", test.name, openFile.LengthWritten, test.sliceSize)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/gofer/handles.go b/pkg/sentry/fs/gofer/handles.go
new file mode 100644
index 000000000..a660c9230
--- /dev/null
+++ b/pkg/sentry/fs/gofer/handles.go
@@ -0,0 +1,144 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/secio"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+)
+
+// handles are the open handles of a gofer file. They are reference counted to
+// support open handle sharing between files for read only filesystems.
+//
+// If Host != nil then it will be used exclusively over File.
+type handles struct {
+	refs.AtomicRefCount
+
+	// File is a p9.File handle. Must not be nil.
+	File contextFile
+
+	// Host is an *fd.FD handle. May be nil.
+	Host *fd.FD
+}
+
+// DecRef drops a reference on handles.
+func (h *handles) DecRef() {
+	h.DecRefWithDestructor(func() {
+		if h.Host != nil {
+			if err := h.Host.Close(); err != nil {
+				log.Warningf("error closing host file: %v", err)
+			}
+		}
+		// FIXME: Context is not plumbed here.
+		if err := h.File.close(context.Background()); err != nil {
+			log.Warningf("error closing p9 file: %v", err)
+		}
+	})
+}
+
+func newHandles(ctx context.Context, file contextFile, flags fs.FileFlags) (*handles, error) {
+	_, newFile, err := file.walk(ctx, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	switch {
+	case flags.Read && flags.Write:
+		hostFile, _, _, err := newFile.open(ctx, p9.ReadWrite)
+		if err != nil {
+			newFile.close(ctx)
+			return nil, err
+		}
+		h := &handles{
+			File: newFile,
+			Host: hostFile,
+		}
+		return h, nil
+	case flags.Read && !flags.Write:
+		hostFile, _, _, err := newFile.open(ctx, p9.ReadOnly)
+		if err != nil {
+			newFile.close(ctx)
+			return nil, err
+		}
+		h := &handles{
+			File: newFile,
+			Host: hostFile,
+		}
+		return h, nil
+	case !flags.Read && flags.Write:
+		hostFile, _, _, err := newFile.open(ctx, p9.WriteOnly)
+		if err != nil {
+			newFile.close(ctx)
+			return nil, err
+		}
+		h := &handles{
+			File: newFile,
+			Host: hostFile,
+		}
+		return h, nil
+	default:
+		panic("impossible fs.FileFlags")
+	}
+}
+
+type handleReadWriter struct {
+	ctx context.Context
+	h   *handles
+	off int64
+}
+
+func (h *handles) readWriterAt(ctx context.Context, offset int64) *handleReadWriter {
+	return &handleReadWriter{ctx, h, offset}
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *handleReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	var r io.Reader
+	if rw.h.Host != nil {
+		r = secio.NewOffsetReader(rw.h.Host, rw.off)
+	} else {
+		r = &p9.ReadWriterFile{File: rw.h.File.file, Offset: uint64(rw.off)}
+	}
+
+	rw.ctx.UninterruptibleSleepStart(false)
+	defer rw.ctx.UninterruptibleSleepFinish(false)
+	n, err := safemem.FromIOReader{r}.ReadToBlocks(dsts)
+	rw.off += int64(n)
+	return n, err
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+func (rw *handleReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	var w io.Writer
+	if rw.h.Host != nil {
+		w = secio.NewOffsetWriter(rw.h.Host, rw.off)
+	} else {
+		w = &p9.ReadWriterFile{File: rw.h.File.file, Offset: uint64(rw.off)}
+	}
+
+	rw.ctx.UninterruptibleSleepStart(false)
+	defer rw.ctx.UninterruptibleSleepFinish(false)
+	n, err := safemem.FromIOWriter{w}.WriteFromBlocks(srcs)
+	rw.off += int64(n)
+	return n, err
+}
diff --git a/pkg/sentry/fs/gofer/inode.go b/pkg/sentry/fs/gofer/inode.go
new file mode 100644
index 000000000..454242923
--- /dev/null
+++ b/pkg/sentry/fs/gofer/inode.go
@@ -0,0 +1,554 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"errors"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fdpipe"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/host"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// inodeOperations implements fs.InodeOperations.
+type inodeOperations struct {
+	fsutil.InodeNotVirtual           `state:"nosave"`
+	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.DeprecatedFileOperations  `state:"nosave"`
+
+	// fileState implements fs.CachedFileObject. It exists
+	// to break a circular load dependency between inodeOperations
+	// and cachingInodeOps (below).
+	fileState *inodeFileState `state:"wait"`
+
+	// cachingInodeOps implement memmap.Mappable for inodeOperations.
+	cachingInodeOps *fsutil.CachingInodeOperations
+
+	// readdirMu protects readdirCache and concurrent Readdirs.
+	readdirMu sync.Mutex `state:"nosave"`
+
+	// readdirCache is a cache of readdir results in the form of
+	// a fs.SortedDentryMap.
+	//
+	// Starts out as nil, and is initialized under readdirMu lazily;
+	// invalidating the cache means setting it to nil.
+	readdirCache *fs.SortedDentryMap `state:"nosave"`
+}
+
+// inodeFileState implements fs.CachedFileObject and otherwise fully
+// encapsulates state that needs to be manually loaded on restore for
+// this file object.
+//
+// This unfortunate structure exists because fs.CachingInodeOperations
+// defines afterLoad and therefore cannot be lazily loaded (to break a
+// circular load dependency between it and inodeOperations). Even with
+// lazy loading, this approach defines the dependencies between objects
+// and the expected load behavior more concretely.
+type inodeFileState struct {
+	// s is common file system state for Gofers.
+	s *session `state:"wait"`
+
+	// MultiDeviceKey consists of:
+	//
+	// * Device:          file system device from a specific gofer.
+	// * SecondaryDevice: unique identifier of the attach point.
+	// * Inode:           the inode of this resource, unique per Device.=
+	//
+	// These fields combined enable consistent hashing of virtual inodes
+	// on goferDevice.
+	key device.MultiDeviceKey `state:"nosave"`
+
+	// file is the p9 file that contains a single unopened fid.
+	file contextFile `state:"nosave"`
+
+	// sattr caches the stable attributes.
+	sattr fs.StableAttr `state:"wait"`
+
+	// handlesMu protects the below fields.
+	handlesMu sync.RWMutex `state:"nosave"`
+
+	// Do minimal open handle caching: only for read only filesystems.
+	readonly *handles `state:"nosave"`
+
+	// Maintain readthrough handles for populating page caches.
+	readthrough *handles `state:"nosave"`
+
+	// Maintain writeback handles for syncing from page caches.
+	writeback *handles `state:"nosave"`
+
+	// writebackRW indicates whether writeback is opened read-write. If
+	// it is not and a read-write handle could replace writeback (above),
+	// then writeback is replaced with the read-write handle. This
+	// ensures that files that were first opened write-only and then
+	// later are opened read-write to be mapped can in fact be mapped.
+	writebackRW bool
+
+	// loading is acquired when the inodeFileState begins an asynchronous
+	// load. It releases when the load is complete. Callers that require all
+	// state to be available should call waitForLoad() to ensure that.
+	loading sync.Mutex `state:".(struct{})"`
+
+	// savedUAttr is only allocated during S/R. It points to the save-time
+	// unstable attributes and is used to validate restore-time ones.
+	//
+	// Note that these unstable attributes are only used to detect cross-S/R
+	// external file system metadata changes. They may differ from the
+	// cached unstable attributes in cachingInodeOps, as that might differ
+	// from the external file system attributes if there had been WriteOut
+	// failures. S/R is transparent to Sentry and the latter will continue
+	// using its cached values after restore.
+	savedUAttr *fs.UnstableAttr
+}
+
+// Release releases file handles.
+func (i *inodeFileState) Release(ctx context.Context) {
+	i.file.close(ctx)
+	if i.readonly != nil {
+		i.readonly.DecRef()
+	}
+	if i.readthrough != nil {
+		i.readthrough.DecRef()
+	}
+	if i.writeback != nil {
+		i.writeback.DecRef()
+	}
+}
+
+// setHandlesForCachedIO installs file handles for reading and writing
+// through fs.CachingInodeOperations.
+func (i *inodeFileState) setHandlesForCachedIO(flags fs.FileFlags, h *handles) {
+	i.handlesMu.Lock()
+	defer i.handlesMu.Unlock()
+
+	if flags.Read {
+		if i.readthrough == nil {
+			h.IncRef()
+			i.readthrough = h
+		}
+	}
+	if flags.Write {
+		if i.writeback == nil {
+			h.IncRef()
+			i.writeback = h
+		} else if !i.writebackRW && flags.Read {
+			i.writeback.DecRef()
+			h.IncRef()
+			i.writeback = h
+		}
+		if flags.Read {
+			i.writebackRW = true
+		}
+	}
+}
+
+// getCachedHandles returns any cached handles which would accelerate
+// performance generally. These handles should only be used if the mount
+// supports caching. This is distinct from fs.CachingInodeOperations
+// which is used for a limited set of file types (those that can be mapped).
+func (i *inodeFileState) getCachedHandles(ctx context.Context, flags fs.FileFlags, msrc *fs.MountSource) (*handles, bool) {
+	i.handlesMu.Lock()
+	defer i.handlesMu.Unlock()
+
+	if flags.Read && !flags.Write && msrc.Flags.ReadOnly {
+		if i.readonly != nil {
+			i.readonly.IncRef()
+			return i.readonly, true
+		}
+		h, err := newHandles(ctx, i.file, flags)
+		if err != nil {
+			return nil, false
+		}
+		i.readonly = h
+		i.readonly.IncRef()
+		return i.readonly, true
+	}
+
+	return nil, false
+}
+
+// ReadToBlocksAt implements fsutil.CachedFileObject.ReadToBlocksAt.
+func (i *inodeFileState) ReadToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) {
+	i.handlesMu.RLock()
+	defer i.handlesMu.RUnlock()
+	return i.readthrough.readWriterAt(ctx, int64(offset)).ReadToBlocks(dsts)
+}
+
+// WriteFromBlocksAt implements fsutil.CachedFileObject.WriteFromBlocksAt.
+func (i *inodeFileState) WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) {
+	i.handlesMu.RLock()
+	defer i.handlesMu.RUnlock()
+	return i.writeback.readWriterAt(ctx, int64(offset)).WriteFromBlocks(srcs)
+}
+
+// SetMaskedAttributes implements fsutil.CachedFileObject.SetMaskedAttributes.
+func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error {
+	if mask.Empty() {
+		return nil
+	}
+	as, ans := attr.AccessTime.Unix()
+	ms, mns := attr.ModificationTime.Unix()
+	// An update of status change time is implied by mask.AccessTime
+	// or mask.ModificationTime. Updating status change time to a
+	// time earlier than the system time is not possible.
+	return i.file.setAttr(
+		ctx,
+		p9.SetAttrMask{
+			Permissions:        mask.Perms,
+			Size:               mask.Size,
+			UID:                mask.UID,
+			GID:                mask.GID,
+			ATime:              mask.AccessTime,
+			ATimeNotSystemTime: true,
+			MTime:              mask.ModificationTime,
+			MTimeNotSystemTime: true,
+		}, p9.SetAttr{
+			Permissions:      p9.FileMode(attr.Perms.LinuxMode()),
+			UID:              p9.UID(attr.Owner.UID),
+			GID:              p9.GID(attr.Owner.GID),
+			Size:             uint64(attr.Size),
+			ATimeSeconds:     uint64(as),
+			ATimeNanoSeconds: uint64(ans),
+			MTimeSeconds:     uint64(ms),
+			MTimeNanoSeconds: uint64(mns),
+		})
+}
+
+// Sync implements fsutil.CachedFileObject.Sync.
+func (i *inodeFileState) Sync(ctx context.Context) error {
+	i.handlesMu.RLock()
+	defer i.handlesMu.RUnlock()
+	if i.writeback == nil {
+		return nil
+	}
+	return i.writeback.File.fsync(ctx)
+}
+
+// FD implements fsutil.CachedFileObject.FD.
+//
+// FD meets the requirements of fsutil.CachedFileObject.FD because p9.File.Open
+// returns a host file descriptor to back _both_ readthrough and writeback or
+// not at all (e.g. both are nil).
+func (i *inodeFileState) FD() int {
+	i.handlesMu.RLock()
+	defer i.handlesMu.RUnlock()
+
+	// Assert that the file was actually opened.
+	if i.writeback == nil && i.readthrough == nil {
+		panic("cannot get host FD for a file that was never opened")
+	}
+	// If this file is mapped, then it must have been opened
+	// read-write and i.writeback was upgraded to a read-write
+	// handle. Prefer that to map.
+	if i.writeback != nil {
+		if i.writeback.Host == nil {
+			return -1
+		}
+		return int(i.writeback.Host.FD())
+	}
+	// Otherwise the file may only have been opened readable
+	// so far. That's the only way it can be accessed.
+	if i.readthrough.Host == nil {
+		return -1
+	}
+	return int(i.readthrough.Host.FD())
+}
+
+// waitForLoad makes sure any restore-issued loading is done.
+func (i *inodeFileState) waitForLoad() {
+	// This is not a no-op. The loading mutex is hold upon restore until
+	// all loading actions are done.
+	i.loading.Lock()
+	i.loading.Unlock()
+}
+
+func (i *inodeFileState) unstableAttr(ctx context.Context) (fs.UnstableAttr, error) {
+	_, valid, pattr, err := getattr(ctx, i.file)
+	if err != nil {
+		return fs.UnstableAttr{}, err
+	}
+	return unstable(ctx, valid, pattr, i.s.mounter, i.s.client), nil
+}
+
+// session extracts the gofer's session from the MountSource.
+func (i *inodeOperations) session() *session {
+	return i.fileState.s
+}
+
+// Release implements fs.InodeOperations.Release.
+func (i *inodeOperations) Release(ctx context.Context) {
+	i.fileState.Release(ctx)
+	i.cachingInodeOps.Release()
+}
+
+// Mappable implements fs.InodeOperations.Mappable.
+func (i *inodeOperations) Mappable(inode *fs.Inode) memmap.Mappable {
+	if i.session().cachePolicy == cacheNone || !fs.IsFile(inode.StableAttr) {
+		return nil
+	}
+	return i.cachingInodeOps
+}
+
+func isCachable(session *session, inode *fs.Inode) bool {
+	return session.cachePolicy != cacheNone && (fs.IsFile(inode.StableAttr) || fs.IsDir(inode.StableAttr))
+}
+
+func isFileCachable(session *session, inode *fs.Inode) bool {
+	return session.cachePolicy != cacheNone && fs.IsFile(inode.StableAttr)
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (i *inodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	if isCachable(i.session(), inode) {
+		return i.cachingInodeOps.UnstableAttr(ctx, inode)
+	}
+	return i.fileState.unstableAttr(ctx)
+}
+
+// Check implements fs.InodeOperations.Check.
+func (i *inodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (i *inodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	switch d.Inode.StableAttr.Type {
+	case fs.Socket:
+		return i.getFileSocket(ctx, d, flags)
+	case fs.Pipe:
+		return i.getFilePipe(ctx, d, flags)
+	default:
+		return i.getFileDefault(ctx, d, flags)
+	}
+}
+
+func (i *inodeOperations) getFileSocket(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	f, err := i.fileState.file.connect(ctx, p9.AnonymousSocket)
+	if err != nil {
+		return nil, syscall.EIO
+	}
+	fsf, err := host.NewSocketWithDirent(ctx, d, f, flags)
+	if err != nil {
+		f.Close()
+		return nil, err
+	}
+	return fsf, nil
+}
+
+func (i *inodeOperations) getFilePipe(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	// Try to open as a host pipe.
+	if pipeOps, err := fdpipe.Open(ctx, i, flags); err != errNotHostFile {
+		return fs.NewFile(ctx, d, flags, pipeOps), err
+	}
+
+	// If the error is due to the fact that this was never a host pipe, then back
+	// this file with its dirent.
+	h, err := newHandles(ctx, i.fileState.file, flags)
+	if err != nil {
+		return nil, err
+	}
+	return NewFile(ctx, d, flags, i, h), nil
+}
+
+// errNotHostFile indicates that the file is not a host file.
+var errNotHostFile = errors.New("not a host file")
+
+// NonBlockingOpen implements fdpipe.NonBlockingOpener for opening host named pipes.
+func (i *inodeOperations) NonBlockingOpen(ctx context.Context, p fs.PermMask) (*fd.FD, error) {
+	i.fileState.waitForLoad()
+
+	// Get a cloned fid which we will open.
+	_, newFile, err := i.fileState.file.walk(ctx, nil)
+	if err != nil {
+		log.Warningf("Open Walk failed: %v", err)
+		return nil, err
+	}
+	defer newFile.close(ctx)
+
+	flags, err := openFlagsFromPerms(p)
+	if err != nil {
+		log.Warningf("Open flags %s parsing failed: %v", p, err)
+		return nil, err
+	}
+	hostFile, _, _, err := newFile.open(ctx, flags)
+	// If the host file returned is nil and the error is nil,
+	// then this was never a host file to begin with, and should
+	// be treated like a remote file.
+	if hostFile == nil && err == nil {
+		return nil, errNotHostFile
+	}
+	return hostFile, err
+}
+
+func (i *inodeOperations) getFileDefault(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	if !isFileCachable(i.session(), d.Inode) {
+		h, err := newHandles(ctx, i.fileState.file, flags)
+		if err != nil {
+			return nil, err
+		}
+		return NewFile(ctx, d, flags, i, h), nil
+	}
+
+	h, ok := i.fileState.getCachedHandles(ctx, flags, d.Inode.MountSource)
+	if !ok {
+		var err error
+		h, err = newHandles(ctx, i.fileState.file, flags)
+		if err != nil {
+			return nil, err
+		}
+	}
+	i.fileState.setHandlesForCachedIO(flags, h)
+
+	return NewFile(ctx, d, flags, i, h), nil
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (i *inodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, p fs.FilePermissions) bool {
+	if isCachable(i.session(), inode) {
+		return i.cachingInodeOps.SetPermissions(ctx, inode, p)
+	}
+
+	mask := p9.SetAttrMask{Permissions: true}
+	pattr := p9.SetAttr{Permissions: p9.FileMode(p.LinuxMode())}
+	// Execute the chmod.
+	return i.fileState.file.setAttr(ctx, mask, pattr) == nil
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (i *inodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	// Save the roundtrip.
+	if !owner.UID.Ok() && !owner.GID.Ok() {
+		return nil
+	}
+
+	if isCachable(i.session(), inode) {
+		return i.cachingInodeOps.SetOwner(ctx, inode, owner)
+	}
+
+	var mask p9.SetAttrMask
+	var attr p9.SetAttr
+	if owner.UID.Ok() {
+		mask.UID = true
+		attr.UID = p9.UID(owner.UID)
+	}
+	if owner.GID.Ok() {
+		mask.GID = true
+		attr.GID = p9.GID(owner.GID)
+	}
+	return i.fileState.file.setAttr(ctx, mask, attr)
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (i *inodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if isCachable(i.session(), inode) {
+		return i.cachingInodeOps.SetTimestamps(ctx, inode, ts)
+	}
+
+	return utimes(ctx, i.fileState.file, ts)
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (i *inodeOperations) Truncate(ctx context.Context, inode *fs.Inode, length int64) error {
+	// This can only be called for files anyway.
+	if isFileCachable(i.session(), inode) {
+		return i.cachingInodeOps.Truncate(ctx, inode, length)
+	}
+
+	return i.fileState.file.setAttr(ctx, p9.SetAttrMask{Size: true}, p9.SetAttr{Size: uint64(length)})
+}
+
+// WriteOut implements fs.InodeOperations.WriteOut.
+func (i *inodeOperations) WriteOut(ctx context.Context, inode *fs.Inode) error {
+	if !isCachable(i.session(), inode) {
+		return nil
+	}
+
+	return i.cachingInodeOps.WriteOut(ctx, inode)
+}
+
+// Readlink implements fs.InodeOperations.Readlink.
+func (i *inodeOperations) Readlink(ctx context.Context, inode *fs.Inode) (string, error) {
+	if !fs.IsSymlink(inode.StableAttr) {
+		return "", syscall.ENOLINK
+	}
+	return i.fileState.file.readlink(ctx)
+}
+
+// Getlink implementfs fs.InodeOperations.Getlink.
+func (i *inodeOperations) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	if !fs.IsSymlink(i.fileState.sattr) {
+		return nil, syserror.ENOLINK
+	}
+	return nil, fs.ErrResolveViaReadlink
+}
+
+// StatFS makes a StatFS request.
+func (i *inodeOperations) StatFS(ctx context.Context) (fs.Info, error) {
+	fsstat, err := i.fileState.file.statFS(ctx)
+	if err != nil {
+		return fs.Info{}, err
+	}
+
+	info := fs.Info{
+		// This is primarily for distinguishing a gofer file system in
+		// tests. Testing is important, so instead of defining
+		// something completely random, use a standard value.
+		Type:        linux.V9FS_MAGIC,
+		TotalBlocks: fsstat.Blocks,
+		FreeBlocks:  fsstat.BlocksFree,
+		TotalFiles:  fsstat.Files,
+		FreeFiles:   fsstat.FilesFree,
+	}
+
+	// If blocks available is non-zero, prefer that.
+	if fsstat.BlocksAvailable != 0 {
+		info.FreeBlocks = fsstat.BlocksAvailable
+	}
+
+	return info, nil
+}
+
+func init() {
+	syserror.AddErrorUnwrapper(func(err error) (syscall.Errno, bool) {
+		if _, ok := err.(p9.ErrSocket); ok {
+			// Treat as an I/O error.
+			return syscall.EIO, true
+		}
+		return 0, false
+	})
+}
+
+// AddLink implements InodeOperations.AddLink, but is currently a noop.
+// FIXME: Remove this from InodeOperations altogether.
+func (*inodeOperations) AddLink() {}
+
+// DropLink implements InodeOperations.DropLink, but is currently a noop.
+// FIXME: Remove this from InodeOperations altogether.
+func (*inodeOperations) DropLink() {}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+// FIXME: Remove this from InodeOperations altogether.
+func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {}
diff --git a/pkg/sentry/fs/gofer/inode_state.go b/pkg/sentry/fs/gofer/inode_state.go
new file mode 100644
index 000000000..997a7d1c1
--- /dev/null
+++ b/pkg/sentry/fs/gofer/inode_state.go
@@ -0,0 +1,141 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"fmt"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+)
+
+// Some fs implementations may not support atime, ctime, or mtime in getattr.
+// The unstable() logic would try to use clock time for them. However, we do not
+// want to use such time during S/R as that would cause restore timestamp
+// checking failure. Hence a dummy stable-time clock is needed.
+//
+// Note that application-visible UnstableAttrs either come from CachingInodeOps
+// (in which case they are saved), or they are requested from the gofer on each
+// stat (for non-caching), so the dummy time only affects the modification
+// timestamp check.
+type dummyClock struct {
+	time.Clock
+}
+
+// Now returns a stable dummy time.
+func (d *dummyClock) Now() time.Time {
+	return time.Time{}
+}
+
+type dummyClockContext struct {
+	context.Context
+}
+
+// Value implements context.Context
+func (d *dummyClockContext) Value(key interface{}) interface{} {
+	switch key {
+	case time.CtxRealtimeClock:
+		return &dummyClock{}
+	default:
+		return d.Context.Value(key)
+	}
+}
+
+// beforeSave is invoked by stateify.
+func (i *inodeFileState) beforeSave() {
+	if _, ok := i.s.inodeMappings[i.sattr.InodeID]; !ok {
+		panic(fmt.Sprintf("failed to find path for inode number %d. Device %s contains %s", i.sattr.InodeID, i.s.connID, fs.InodeMappings(i.s.inodeMappings)))
+	}
+	if i.sattr.Type == fs.RegularFile {
+		uattr, err := i.unstableAttr(&dummyClockContext{context.Background()})
+		if err != nil {
+			panic(fmt.Sprintf("failed to get unstable atttribute of %s: %v", i.s.inodeMappings[i.sattr.InodeID], err))
+		}
+		i.savedUAttr = &uattr
+	}
+}
+
+// saveLoading is invoked by stateify.
+func (i *inodeFileState) saveLoading() struct{} {
+	return struct{}{}
+}
+
+// loadLoading is invoked by stateify.
+func (i *inodeFileState) loadLoading(_ struct{}) {
+	i.loading.Lock()
+}
+
+// afterLoad is invoked by stateify.
+func (i *inodeFileState) afterLoad() {
+	load := func() {
+		// See comment on i.loading().
+		defer i.loading.Unlock()
+
+		// Manually restore the p9.File.
+		name, ok := i.s.inodeMappings[i.sattr.InodeID]
+		if !ok {
+			// This should be impossible, see assertion in
+			// beforeSave.
+			panic(fmt.Sprintf("failed to find path for inode number %d. Device %s contains %s", i.sattr.InodeID, i.s.connID, fs.InodeMappings(i.s.inodeMappings)))
+		}
+		// TODO: Context is not plumbed to save/restore.
+		ctx := &dummyClockContext{context.Background()}
+		var err error
+		_, i.file, err = i.s.attach.walk(ctx, strings.Split(name, "/"))
+		if err != nil {
+			panic(fmt.Sprintf("failed to walk to %q: %v", name, err))
+		}
+
+		// Remap the saved inode number into the gofer device using the
+		// actual device and actual inode that exists in our new
+		// environment.
+		qid, mask, attrs, err := i.file.getAttr(ctx, p9.AttrMaskAll())
+		if err != nil {
+			panic(fmt.Sprintf("failed to get file attributes of %s: %v", name, err))
+		}
+		if !mask.RDev {
+			panic(fmt.Sprintf("file %s lacks device", name))
+		}
+		i.key = device.MultiDeviceKey{
+			Device:          attrs.RDev,
+			SecondaryDevice: i.s.connID,
+			Inode:           qid.Path,
+		}
+		if !goferDevice.Load(i.key, i.sattr.InodeID) {
+			panic(fmt.Sprintf("gofer device %s -> %d conflict in gofer device mappings: %s", i.key, i.sattr.InodeID, goferDevice))
+		}
+
+		if i.sattr.Type == fs.RegularFile {
+			env, ok := fs.CurrentRestoreEnvironment()
+			if !ok {
+				panic("missing restore environment")
+			}
+			uattr := unstable(ctx, mask, attrs, i.s.mounter, i.s.client)
+			if env.ValidateFileSize && uattr.Size != i.savedUAttr.Size {
+				panic(fmt.Errorf("file size has changed for %s: previously %d, now %d", i.s.inodeMappings[i.sattr.InodeID], i.savedUAttr.Size, uattr.Size))
+			}
+			if env.ValidateFileTimestamp && uattr.ModificationTime != i.savedUAttr.ModificationTime {
+				panic(fmt.Errorf("file modification time has changed for %s: previously %v, now %v", i.s.inodeMappings[i.sattr.InodeID], i.savedUAttr.ModificationTime, uattr.ModificationTime))
+			}
+			i.savedUAttr = nil
+		}
+	}
+
+	fs.Async(load)
+}
diff --git a/pkg/sentry/fs/gofer/path.go b/pkg/sentry/fs/gofer/path.go
new file mode 100644
index 000000000..d696f1561
--- /dev/null
+++ b/pkg/sentry/fs/gofer/path.go
@@ -0,0 +1,331 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// Lookup loads an Inode at name into a Dirent based on the session's cache
+// policy.
+func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) {
+	if i.session().cachePolicy != cacheNone {
+		// Check to see if we have readdirCache that indicates the
+		// child does not exist.  Avoid holding readdirMu longer than
+		// we need to.
+		i.readdirMu.Lock()
+		if i.readdirCache != nil && !i.readdirCache.Contains(name) {
+			// No such child.  Return a negative dirent.
+			i.readdirMu.Unlock()
+			return fs.NewNegativeDirent(name), nil
+		}
+		i.readdirMu.Unlock()
+	}
+
+	// Get a p9.File for name.
+	qids, newFile, mask, p9attr, err := i.fileState.file.walkGetAttr(ctx, []string{name})
+	if err != nil {
+		if err == syscall.ENOENT {
+			if i.session().cachePolicy != cacheNone {
+				// Return a negative Dirent. It will stay cached until something
+				// is created over it.
+				return fs.NewNegativeDirent(name), nil
+			}
+			return nil, syserror.ENOENT
+		}
+		return nil, err
+	}
+
+	// Construct the Inode operations.
+	sattr, node := newInodeOperations(ctx, i.fileState.s, newFile, qids[0], mask, p9attr)
+
+	// Construct a positive Dirent.
+	return fs.NewDirent(fs.NewInode(node, dir.MountSource, sattr), name), nil
+}
+
+// Creates a new Inode at name and returns its File based on the session's cache policy.
+//
+// Ownership is currently ignored.
+func (i *inodeOperations) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perm fs.FilePermissions) (*fs.File, error) {
+	// Create replaces the directory fid with the newly created/opened
+	// file, so clone this directory so it doesn't change out from under
+	// this node.
+	_, newFile, err := i.fileState.file.walk(ctx, nil)
+	if err != nil {
+		return nil, err
+	}
+
+	// Map the FileFlags to p9 OpenFlags.
+	var openFlags p9.OpenFlags
+	switch {
+	case flags.Read && flags.Write:
+		openFlags = p9.ReadWrite
+	case flags.Read:
+		openFlags = p9.ReadOnly
+	case flags.Write:
+		openFlags = p9.WriteOnly
+	default:
+		panic(fmt.Sprintf("Create called with unknown or unset open flags: %v", flags))
+	}
+
+	owner := fs.FileOwnerFromContext(ctx)
+	hostFile, err := newFile.create(ctx, name, openFlags, p9.FileMode(perm.LinuxMode()), p9.UID(owner.UID), p9.GID(owner.GID))
+	if err != nil {
+		// Could not create the file.
+		return nil, err
+	}
+
+	i.touchModificationTime(ctx)
+
+	// Get the attributes of the file.
+	qid, mask, p9attr, err := getattr(ctx, newFile)
+	if err != nil {
+		newFile.close(ctx)
+		return nil, err
+	}
+
+	// Get an unopened p9.File for the file we created so that it can be
+	// cloned and re-opened multiple times after creation.
+	_, unopened, err := i.fileState.file.walk(ctx, []string{name})
+	if err != nil {
+		newFile.close(ctx)
+		return nil, err
+	}
+
+	// Construct the InodeOperations.
+	sattr, iops := newInodeOperations(ctx, i.fileState.s, unopened, qid, mask, p9attr)
+
+	// Construct the positive Dirent.
+	d := fs.NewDirent(fs.NewInode(iops, dir.MountSource, sattr), name)
+	defer d.DecRef()
+
+	// Construct the new file, caching the handles if allowed.
+	h := &handles{
+		File: newFile,
+		Host: hostFile,
+	}
+	if isFileCachable(iops.session(), d.Inode) {
+		iops.fileState.setHandlesForCachedIO(flags, h)
+	}
+	return NewFile(ctx, d, flags, iops, h), nil
+}
+
+// CreateLink uses Create to create a symlink between oldname and newname.
+func (i *inodeOperations) CreateLink(ctx context.Context, dir *fs.Inode, oldname string, newname string) error {
+	owner := fs.FileOwnerFromContext(ctx)
+	if _, err := i.fileState.file.symlink(ctx, oldname, newname, p9.UID(owner.UID), p9.GID(owner.GID)); err != nil {
+		return err
+	}
+	i.touchModificationTime(ctx)
+	return nil
+}
+
+// CreateHardLink implements InodeOperations.CreateHardLink.
+func (i *inodeOperations) CreateHardLink(ctx context.Context, _ *fs.Inode, target *fs.Inode, newName string) error {
+	targetOpts, ok := target.InodeOperations.(*inodeOperations)
+	if !ok {
+		return syscall.EXDEV
+	}
+
+	if err := i.fileState.file.link(ctx, &targetOpts.fileState.file, newName); err != nil {
+		return err
+	}
+	// TODO: Don't increase link count because we can't properly accounts for links
+	// with gofers.
+	i.touchModificationTime(ctx)
+	return nil
+}
+
+// CreateDirectory uses Create to create a directory named s under inodeOperations.
+func (i *inodeOperations) CreateDirectory(ctx context.Context, dir *fs.Inode, s string, perm fs.FilePermissions) error {
+	owner := fs.FileOwnerFromContext(ctx)
+	if _, err := i.fileState.file.mkdir(ctx, s, p9.FileMode(perm.LinuxMode()), p9.UID(owner.UID), p9.GID(owner.GID)); err != nil {
+		return err
+	}
+	if i.session().cachePolicy == cacheAll {
+		// Increase link count.
+		i.cachingInodeOps.IncLinks(ctx)
+
+		// Invalidate readdir cache.
+		i.markDirectoryDirty()
+	}
+	return nil
+}
+
+// Bind implements InodeOperations.
+func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, ep unix.BoundEndpoint, perm fs.FilePermissions) error {
+	if i.session().endpoints == nil {
+		return syscall.EOPNOTSUPP
+	}
+
+	// Create replaces the directory fid with the newly created/opened
+	// file, so clone this directory so it doesn't change out from under
+	// this node.
+	_, newFile, err := i.fileState.file.walk(ctx, nil)
+	if err != nil {
+		return err
+	}
+
+	// Stabilize the endpoint map while creation is in progress.
+	unlock := i.session().endpoints.lock()
+	defer unlock()
+
+	// Create a regular file in the gofer and then mark it as a socket by
+	// adding this inode key in the 'endpoints' map.
+	owner := fs.FileOwnerFromContext(ctx)
+	hostFile, err := newFile.create(ctx, name, p9.ReadWrite, p9.FileMode(perm.LinuxMode()), p9.UID(owner.UID), p9.GID(owner.GID))
+	if err != nil {
+		return err
+	}
+	// We're not going to use this file.
+	hostFile.Close()
+
+	i.touchModificationTime(ctx)
+
+	// Get the attributes of the file to create inode key.
+	qid, _, attr, err := getattr(ctx, newFile)
+	if err != nil {
+		newFile.close(ctx)
+		return err
+	}
+
+	key := device.MultiDeviceKey{
+		Device:          attr.RDev,
+		SecondaryDevice: i.session().connID,
+		Inode:           qid.Path,
+	}
+	i.session().endpoints.add(key, ep)
+
+	return nil
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo. Gofer nodes do not support the
+// creation of fifos and always returns EOPNOTSUPP.
+func (*inodeOperations) CreateFifo(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return syscall.EOPNOTSUPP
+}
+
+// Remove implements InodeOperations.Remove.
+func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string) error {
+	var key device.MultiDeviceKey
+	removeSocket := false
+	if i.session().endpoints != nil {
+		// Find out if file being deleted is a socket that needs to be
+		// removed from endpoint map.
+		if d, err := i.Lookup(ctx, dir, name); err == nil {
+			defer d.DecRef()
+			if fs.IsSocket(d.Inode.StableAttr) {
+				child := d.Inode.InodeOperations.(*inodeOperations)
+				key = child.fileState.key
+				removeSocket = true
+
+				// Stabilize the endpoint map while deletion is in progress.
+				unlock := i.session().endpoints.lock()
+				defer unlock()
+			}
+		}
+	}
+
+	if err := i.fileState.file.unlinkAt(ctx, name, 0); err != nil {
+		return err
+	}
+	if removeSocket {
+		i.session().endpoints.remove(key)
+	}
+	i.touchModificationTime(ctx)
+
+	return nil
+}
+
+// Remove implements InodeOperations.RemoveDirectory.
+func (i *inodeOperations) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error {
+	// 0x200 = AT_REMOVEDIR.
+	if err := i.fileState.file.unlinkAt(ctx, name, 0x200); err != nil {
+		return err
+	}
+	if i.session().cachePolicy == cacheAll {
+		// Decrease link count and updates atime.
+		i.cachingInodeOps.DecLinks(ctx)
+
+		// Invalidate readdir cache.
+		i.markDirectoryDirty()
+	}
+	return nil
+}
+
+// Rename renames this node.
+func (i *inodeOperations) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	// Unwrap the new parent to a *inodeOperations.
+	newParentInodeOperations, ok := newParent.InodeOperations.(*inodeOperations)
+	if !ok {
+		return syscall.EXDEV
+	}
+
+	// Unwrap the old parent to a *inodeOperations.
+	oldParentInodeOperations, ok := oldParent.InodeOperations.(*inodeOperations)
+	if !ok {
+		return syscall.EXDEV
+	}
+
+	// Do the rename.
+	if err := i.fileState.file.rename(ctx, newParentInodeOperations.fileState.file, newName); err != nil {
+		return err
+	}
+
+	// Update cached state.
+	if i.session().cachePolicy == cacheAll {
+		// Is the renamed entity a directory? Fix link counts.
+		if fs.IsDir(i.fileState.sattr) {
+			oldParentInodeOperations.cachingInodeOps.DecLinks(ctx)
+			newParentInodeOperations.cachingInodeOps.IncLinks(ctx)
+		}
+
+		// Mark old directory dirty.
+		oldParentInodeOperations.markDirectoryDirty()
+		if oldParent != newParent {
+			// Mark new directory dirty.
+			newParentInodeOperations.markDirectoryDirty()
+		}
+	}
+	return nil
+}
+
+func (i *inodeOperations) touchModificationTime(ctx context.Context) {
+	if i.session().cachePolicy == cacheAll {
+		i.cachingInodeOps.TouchModificationTime(ctx)
+
+		// Invalidate readdir cache.
+		i.markDirectoryDirty()
+	}
+}
+
+// markDirectoryDirty marks any cached data dirty for this directory. This is necessary in order
+// to ensure that this node does not retain stale state throughout its lifetime across multiple
+// open directory handles.
+//
+// Currently this means invalidating any readdir caches.
+func (i *inodeOperations) markDirectoryDirty() {
+	i.readdirMu.Lock()
+	defer i.readdirMu.Unlock()
+	i.readdirCache = nil
+}
diff --git a/pkg/sentry/fs/gofer/session.go b/pkg/sentry/fs/gofer/session.go
new file mode 100644
index 000000000..ab3b964e0
--- /dev/null
+++ b/pkg/sentry/fs/gofer/session.go
@@ -0,0 +1,251 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+)
+
+type endpointMap struct {
+	mu sync.RWMutex
+	m  map[device.MultiDeviceKey]unix.BoundEndpoint
+}
+
+// add adds the endpoint to the map.
+//
+// Precondition: map must have been locked with 'lock'.
+func (e *endpointMap) add(key device.MultiDeviceKey, ep unix.BoundEndpoint) {
+	e.m[key] = ep
+}
+
+// remove deletes the key from the map.
+//
+// Precondition: map must have been locked with 'lock'.
+func (e *endpointMap) remove(key device.MultiDeviceKey) {
+	delete(e.m, key)
+}
+
+// lock blocks other addition and removal operations from happening while
+// the backing file is being created or deleted. Returns a function that unlocks
+// the endpoint map.
+func (e *endpointMap) lock() func() {
+	e.mu.Lock()
+	return func() { e.mu.Unlock() }
+}
+
+func (e *endpointMap) get(key device.MultiDeviceKey) unix.BoundEndpoint {
+	e.mu.RLock()
+	ep := e.m[key]
+	e.mu.RUnlock()
+	return ep
+}
+
+// session holds state for each 9p session established during sys_mount.
+type session struct {
+	refs.AtomicRefCount
+
+	// conn is a unet.Socket that wraps the readFD/writeFD mount option,
+	// see fs/gofer/fs.go.
+	conn *unet.Socket `state:"nosave"`
+
+	// msize is the value of the msize mount option, see fs/gofer/fs.go.
+	msize uint32 `state:"wait"`
+
+	// version is the value of the version mount option, see fs/gofer/fs.go.
+	version string `state:"wait"`
+
+	// cachePolicy is the cache policy. It may be either cacheAll or cacheNone.
+	cachePolicy cachePolicy `state:"wait"`
+
+	// aname is the value of the aname mount option, see fs/gofer/fs.go.
+	aname string `state:"wait"`
+
+	// The client associated with this session. This will be initialized lazily.
+	client *p9.Client `state:"nosave"`
+
+	// The p9.File pointing to attachName via the client. This will be initialized
+	// lazily.
+	attach contextFile `state:"nosave"`
+
+	// Flags provided to the mount.
+	superBlockFlags fs.MountSourceFlags `state:"wait"`
+
+	// connID is a unique identifier for the session connection.
+	connID string `state:"wait"`
+
+	// inodeMappings contains mappings of fs.Inodes associated with this session
+	// to paths relative to the attach point, where inodeMappings is keyed by
+	// Inode.StableAttr.InodeID.
+	inodeMappings map[uint64]string `state:"wait"`
+
+	// mounter is the EUID/EGID that mounted this file system.
+	mounter fs.FileOwner `state:"wait"`
+
+	// endpoints is used to map inodes that represent socket files to their
+	// corresponding endpoint. Socket files are created as regular files in the
+	// gofer and their presence in this map indicate that they should indeed be
+	// socket files. This allows unix domain sockets to be used with paths that
+	// belong to a gofer.
+	//
+	// TODO: there are few possible races with someone stat'ing the
+	// file and another deleting it concurrently, where the file will not be
+	// reported as socket file.
+	endpoints *endpointMap `state:"wait"`
+}
+
+// Destroy tears down the session.
+func (s *session) Destroy() {
+	s.conn.Close()
+}
+
+// Revalidate returns true if the cache policy is does not allow for VFS caching.
+func (s *session) Revalidate(*fs.Dirent) bool {
+	return s.cachePolicy == cacheNone
+}
+
+// TakeRefs takes an extra reference on dirent if possible.
+func (s *session) Keep(dirent *fs.Dirent) bool {
+	// NOTE: Only cache files and directories.
+	sattr := dirent.Inode.StableAttr
+	return s.cachePolicy != cacheNone && (fs.IsFile(sattr) || fs.IsDir(sattr))
+}
+
+// ResetInodeMappings implements fs.MountSourceOperations.ResetInodeMappings.
+func (s *session) ResetInodeMappings() {
+	s.inodeMappings = make(map[uint64]string)
+}
+
+// SaveInodeMapping implements fs.MountSourceOperations.SaveInodeMapping.
+func (s *session) SaveInodeMapping(inode *fs.Inode, path string) {
+	// This is very unintuitive. We *CANNOT* trust the inode's StableAttrs,
+	// because overlay copyUp may have changed them out from under us.
+	// So much for "immutable".
+	sattr := inode.InodeOperations.(*inodeOperations).fileState.sattr
+	s.inodeMappings[sattr.InodeID] = path
+}
+
+// newInodeOperations creates a new 9p fs.InodeOperations backed by a p9.File and attributes
+// (p9.QID, p9.AttrMask, p9.Attr).
+func newInodeOperations(ctx context.Context, s *session, file contextFile, qid p9.QID, valid p9.AttrMask, attr p9.Attr) (fs.StableAttr, *inodeOperations) {
+	deviceKey := device.MultiDeviceKey{
+		Device:          attr.RDev,
+		SecondaryDevice: s.connID,
+		Inode:           qid.Path,
+	}
+
+	sattr := fs.StableAttr{
+		Type:      ntype(attr),
+		DeviceID:  goferDevice.DeviceID(),
+		InodeID:   goferDevice.Map(deviceKey),
+		BlockSize: bsize(attr),
+	}
+
+	if s.endpoints != nil {
+		// If unix sockets are allowed on this filesystem, check if this file is
+		// supposed to be a socket file.
+		if s.endpoints.get(deviceKey) != nil {
+			sattr.Type = fs.Socket
+		}
+	}
+
+	fileState := &inodeFileState{
+		s:     s,
+		file:  file,
+		sattr: sattr,
+		key:   deviceKey,
+	}
+
+	uattr := unstable(ctx, valid, attr, s.mounter, s.client)
+	return sattr, &inodeOperations{
+		fileState:       fileState,
+		cachingInodeOps: fsutil.NewCachingInodeOperations(ctx, fileState, uattr, s.superBlockFlags.ForcePageCache),
+	}
+}
+
+// Root returns the root of a 9p mount. This mount is bound to a 9p server
+// based on conn. Otherwise configuration parameters are:
+//
+// * dev:         connection id
+// * filesystem:  the filesystem backing the mount
+// * superBlockFlags:  the mount flags describing general mount options
+// * opts:        parsed 9p mount options
+func Root(ctx context.Context, dev string, filesystem fs.Filesystem, superBlockFlags fs.MountSourceFlags, o opts) (*fs.Inode, error) {
+	// The mounting EUID/EGID will be cached by this file system. This will
+	// be used to assign ownership to files that the Gofer owns.
+	mounter := fs.FileOwnerFromContext(ctx)
+
+	conn, err := unet.NewSocket(o.fd)
+	if err != nil {
+		return nil, err
+	}
+
+	// Construct the session.
+	s := &session{
+		connID:          dev,
+		conn:            conn,
+		msize:           o.msize,
+		version:         o.version,
+		cachePolicy:     o.policy,
+		aname:           o.aname,
+		superBlockFlags: superBlockFlags,
+		mounter:         mounter,
+	}
+
+	if o.privateunixsocket {
+		s.endpoints = &endpointMap{m: make(map[device.MultiDeviceKey]unix.BoundEndpoint)}
+	}
+
+	// Construct the MountSource with the session and superBlockFlags.
+	m := fs.NewMountSource(s, filesystem, superBlockFlags)
+
+	// Send the Tversion request.
+	s.client, err = p9.NewClient(s.conn, s.msize, s.version)
+	if err != nil {
+		// Drop our reference on the session, it needs to be torn down.
+		s.DecRef()
+		return nil, err
+	}
+
+	// Notify that we're about to call the Gofer and block.
+	ctx.UninterruptibleSleepStart(false)
+	// Send the Tattach request.
+	s.attach.file, err = s.client.Attach(s.aname)
+	ctx.UninterruptibleSleepFinish(false)
+	if err != nil {
+		// Same as above.
+		s.DecRef()
+		return nil, err
+	}
+
+	qid, valid, attr, err := s.attach.getAttr(ctx, p9.AttrMaskAll())
+	if err != nil {
+		s.attach.close(ctx)
+		// Same as above, but after we execute the Close request.
+		s.DecRef()
+		return nil, err
+	}
+
+	sattr, iops := newInodeOperations(ctx, s, s.attach, qid, valid, attr)
+	return fs.NewInode(iops, m, sattr), nil
+}
diff --git a/pkg/sentry/fs/gofer/session_state.go b/pkg/sentry/fs/gofer/session_state.go
new file mode 100644
index 000000000..4d993a219
--- /dev/null
+++ b/pkg/sentry/fs/gofer/session_state.go
@@ -0,0 +1,90 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+)
+
+// afterLoad is invoked by stateify.
+func (s *session) afterLoad() {
+	// The restore environment contains the 9p connection of this mount.
+	fsys := filesystem{}
+	env, ok := fs.CurrentRestoreEnvironment()
+	if !ok {
+		panic("failed to find restore environment")
+	}
+	mounts, ok := env.MountSources[fsys.Name()]
+	if !ok {
+		panic("failed to find mounts for filesystem type " + fsys.Name())
+	}
+	var args fs.MountArgs
+	var found bool
+	for _, mount := range mounts {
+		if mount.Dev == s.connID {
+			args = mount
+			found = true
+		}
+	}
+	if !found {
+		panic(fmt.Sprintf("no connection for connection id %q", s.connID))
+	}
+
+	// Validate the mount flags and options.
+	opts, err := options(args.Data)
+	if err != nil {
+		panic("failed to parse mount options: " + err.Error())
+	}
+	if opts.msize != s.msize {
+		panic(fmt.Sprintf("new message size %v, want %v", opts.msize, s.msize))
+	}
+	if opts.version != s.version {
+		panic(fmt.Sprintf("new version %v, want %v", opts.version, s.version))
+	}
+	if opts.policy != s.cachePolicy {
+		panic(fmt.Sprintf("new cache policy %v, want %v", opts.policy, s.cachePolicy))
+	}
+	if opts.aname != s.aname {
+		panic(fmt.Sprintf("new attach name %v, want %v", opts.aname, s.aname))
+	}
+	if opts.privateunixsocket != (s.endpoints != nil) {
+		panic(fmt.Sprintf("new privateunixsocket option %v, want %v", opts.privateunixsocket, s.endpoints != nil))
+	}
+	if args.Flags != s.superBlockFlags {
+		panic(fmt.Sprintf("new mount flags %v, want %v", args.Flags, s.superBlockFlags))
+	}
+
+	// Manually restore the connection.
+	s.conn, err = unet.NewSocket(opts.fd)
+	if err != nil {
+		panic(fmt.Sprintf("failed to create Socket for FD %d: %v", opts.fd, err))
+	}
+
+	// Manually restore the client.
+	s.client, err = p9.NewClient(s.conn, s.msize, s.version)
+	if err != nil {
+		panic(fmt.Sprintf("failed to connect client to server: %v", err))
+	}
+
+	// Manually restore the attach point.
+	s.attach.file, err = s.client.Attach(s.aname)
+	if err != nil {
+		panic(fmt.Sprintf("failed to attach to aname: %v", err))
+	}
+}
diff --git a/pkg/sentry/fs/gofer/socket.go b/pkg/sentry/fs/gofer/socket.go
new file mode 100644
index 000000000..954000ef0
--- /dev/null
+++ b/pkg/sentry/fs/gofer/socket.go
@@ -0,0 +1,127 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/host"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// BoundEndpoint returns a gofer-backed unix.BoundEndpoint.
+func (i *inodeOperations) BoundEndpoint(inode *fs.Inode, path string) unix.BoundEndpoint {
+	if !fs.IsSocket(i.fileState.sattr) {
+		return nil
+	}
+
+	if i.session().endpoints != nil {
+		ep := i.session().endpoints.get(i.fileState.key)
+		if ep != nil {
+			return ep
+		}
+
+		// Not found in endpoints map, it may be a gofer backed unix socket...
+	}
+
+	inode.IncRef()
+	return &endpoint{inode, i.fileState.file.file, path}
+}
+
+// endpoint is a Gofer-backed unix.BoundEndpoint.
+//
+// An endpoint's lifetime is the time between when InodeOperations.BoundEndpoint()
+// is called and either BoundEndpoint.BidirectionalConnect or
+// BoundEndpoint.UnidirectionalConnect is called.
+type endpoint struct {
+	// inode is the filesystem inode which produced this endpoint.
+	inode *fs.Inode
+
+	// file is the p9 file that contains a single unopened fid.
+	file p9.File
+
+	// path is the sentry path where this endpoint is bound.
+	path string
+}
+
+func unixSockToP9(t unix.SockType) (p9.ConnectFlags, bool) {
+	switch t {
+	case unix.SockStream:
+		return p9.StreamSocket, true
+	case unix.SockSeqpacket:
+		return p9.SeqpacketSocket, true
+	case unix.SockDgram:
+		return p9.DgramSocket, true
+	}
+	return 0, false
+}
+
+// BidirectionalConnect implements ConnectableEndpoint.BidirectionalConnect.
+func (e *endpoint) BidirectionalConnect(ce unix.ConnectingEndpoint, returnConnect func(unix.Receiver, unix.ConnectedEndpoint)) *tcpip.Error {
+	cf, ok := unixSockToP9(ce.Type())
+	if !ok {
+		return tcpip.ErrConnectionRefused
+	}
+
+	// No lock ordering required as only the ConnectingEndpoint has a mutex.
+	ce.Lock()
+	defer ce.Unlock()
+
+	// Check connecting state.
+	if ce.Connected() {
+		return tcpip.ErrAlreadyConnected
+	}
+	if ce.Listening() {
+		return tcpip.ErrInvalidEndpointState
+	}
+
+	hostFile, err := e.file.Connect(cf)
+	if err != nil {
+		return tcpip.ErrConnectionRefused
+	}
+
+	r, c, terr := host.NewConnectedEndpoint(hostFile, ce.WaiterQueue(), e.path)
+	if terr != nil {
+		return terr
+	}
+	returnConnect(r, c)
+	return nil
+}
+
+// UnidirectionalConnect implements unix.BoundEndpoint.UnidirectionalConnect.
+func (e *endpoint) UnidirectionalConnect() (unix.ConnectedEndpoint, *tcpip.Error) {
+	hostFile, err := e.file.Connect(p9.DgramSocket)
+	if err != nil {
+		return nil, tcpip.ErrConnectionRefused
+	}
+
+	r, c, terr := host.NewConnectedEndpoint(hostFile, &waiter.Queue{}, e.path)
+	if terr != nil {
+		return nil, terr
+	}
+
+	// We don't need the receiver.
+	r.CloseRecv()
+	r.Release()
+
+	return c, nil
+}
+
+// Release implements unix.BoundEndpoint.Release.
+func (e *endpoint) Release() {
+	e.inode.DecRef()
+}
diff --git a/pkg/sentry/fs/gofer/util.go b/pkg/sentry/fs/gofer/util.go
new file mode 100644
index 000000000..d9ed8c81e
--- /dev/null
+++ b/pkg/sentry/fs/gofer/util.go
@@ -0,0 +1,60 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package gofer
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/p9"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+func utimes(ctx context.Context, file contextFile, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+	mask := p9.SetAttrMask{
+		ATime:              !ts.ATimeOmit,
+		ATimeNotSystemTime: !ts.ATimeSetSystemTime,
+		MTime:              !ts.MTimeOmit,
+		MTimeNotSystemTime: !ts.MTimeSetSystemTime,
+	}
+	as, ans := ts.ATime.Unix()
+	ms, mns := ts.MTime.Unix()
+	attr := p9.SetAttr{
+		ATimeSeconds:     uint64(as),
+		ATimeNanoSeconds: uint64(ans),
+		MTimeSeconds:     uint64(ms),
+		MTimeNanoSeconds: uint64(mns),
+	}
+	// 9p2000.L SetAttr: "If a time bit is set without the corresponding SET bit,
+	// the current system time on the server is used instead of the value sent
+	// in the request."
+	return file.setAttr(ctx, mask, attr)
+}
+
+func openFlagsFromPerms(p fs.PermMask) (p9.OpenFlags, error) {
+	switch {
+	case p.Read && p.Write:
+		return p9.ReadWrite, nil
+	case p.Write:
+		return p9.WriteOnly, nil
+	case p.Read:
+		return p9.ReadOnly, nil
+	default:
+		return 0, syscall.EINVAL
+	}
+}
diff --git a/pkg/sentry/fs/host/BUILD b/pkg/sentry/fs/host/BUILD
new file mode 100644
index 000000000..97b64daed
--- /dev/null
+++ b/pkg/sentry/fs/host/BUILD
@@ -0,0 +1,104 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "host_state",
+    srcs = [
+        "control.go",
+        "descriptor.go",
+        "descriptor_state.go",
+        "file.go",
+        "fs.go",
+        "inode.go",
+        "inode_state.go",
+        "socket.go",
+        "socket_state.go",
+    ],
+    out = "host_state.go",
+    package = "host",
+)
+
+go_library(
+    name = "host",
+    srcs = [
+        "control.go",
+        "descriptor.go",
+        "descriptor_state.go",
+        "device.go",
+        "file.go",
+        "fs.go",
+        "host_state.go",
+        "inode.go",
+        "inode_state.go",
+        "ioctl_unsafe.go",
+        "socket.go",
+        "socket_state.go",
+        "socket_unsafe.go",
+        "util.go",
+        "util_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/host",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/fd",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/secio",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/socket/control",
+        "//pkg/sentry/socket/unix",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/link/rawfile",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/unet",
+        "//pkg/waiter",
+        "//pkg/waiter/fdnotifier",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
+
+go_test(
+    name = "host_test",
+    size = "small",
+    srcs = [
+        "fs_test.go",
+        "inode_test.go",
+        "socket_test.go",
+        "wait_test.go",
+    ],
+    embed = [":host"],
+    deps = [
+        "//pkg/fd",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/usermem",
+        "//pkg/syserr",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+        "//pkg/waiter/fdnotifier",
+    ],
+)
diff --git a/pkg/sentry/fs/host/control.go b/pkg/sentry/fs/host/control.go
new file mode 100644
index 000000000..d2b007ab2
--- /dev/null
+++ b/pkg/sentry/fs/host/control.go
@@ -0,0 +1,90 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/control"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+type scmRights struct {
+	fds []int
+}
+
+func newSCMRights(fds []int) control.SCMRights {
+	return &scmRights{fds}
+}
+
+// Files implements control.SCMRights.Files.
+func (c *scmRights) Files(ctx context.Context, max int) control.RightsFiles {
+	n := max
+	if l := len(c.fds); n > l {
+		n = l
+	}
+
+	rf := control.RightsFiles(fdsToFiles(ctx, c.fds[:n]))
+
+	// Only consume converted FDs (fdsToFiles may convert fewer than n FDs).
+	c.fds = c.fds[len(rf):]
+	return rf
+}
+
+// Clone implements unix.RightsControlMessage.Clone.
+func (c *scmRights) Clone() unix.RightsControlMessage {
+	// Host rights never need to be cloned.
+	return nil
+}
+
+// Release implements unix.RightsControlMessage.Release.
+func (c *scmRights) Release() {
+	for _, fd := range c.fds {
+		syscall.Close(fd)
+	}
+	c.fds = nil
+}
+
+// If an error is encountered, only files created before the error will be
+// returned. This is what Linux does.
+func fdsToFiles(ctx context.Context, fds []int) []*fs.File {
+	files := make([]*fs.File, 0, len(fds))
+	for _, fd := range fds {
+		// Get flags. We do it here because they may be modified
+		// by subsequent functions.
+		fileFlags, _, errno := syscall.Syscall(syscall.SYS_FCNTL, uintptr(fd), syscall.F_GETFL, 0)
+		if errno != 0 {
+			ctx.Warningf("Error retrieving host FD flags: %v", error(errno))
+			break
+		}
+
+		// Create the file backed by hostFD.
+		file, err := NewFile(ctx, fd, fs.FileOwnerFromContext(ctx))
+		if err != nil {
+			ctx.Warningf("Error creating file from host FD: %v", err)
+			break
+		}
+
+		// Set known flags.
+		file.SetFlags(fs.SettableFileFlags{
+			NonBlocking: fileFlags&syscall.O_NONBLOCK != 0,
+		})
+
+		files = append(files, file)
+	}
+	return files
+}
diff --git a/pkg/sentry/fs/host/descriptor.go b/pkg/sentry/fs/host/descriptor.go
new file mode 100644
index 000000000..613bd06e8
--- /dev/null
+++ b/pkg/sentry/fs/host/descriptor.go
@@ -0,0 +1,118 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"path"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+// descriptor wraps a host fd.
+type descriptor struct {
+	// donated is true if the host fd was donated by another process.
+	donated bool
+
+	// If origFD >= 0, it is the host fd that this file was
+	// originally created from, which must be available at time
+	// of restore. Only valid if donated is true.
+	origFD int
+
+	// wouldBlock is true if value (below) points to a file that can
+	// return EWOULDBLOCK for operations that would block.
+	wouldBlock bool
+
+	// value is the wrapped host fd. It is never saved or restored
+	// directly. How it is restored depends on whether it was
+	// donated and the fs.MountSource it was originally
+	// opened/created from.
+	value int `state:"nosave"`
+}
+
+// newDescriptor returns a wrapped host file descriptor. On success,
+// the descriptor is registered for event notifications with queue.
+func newDescriptor(fd int, donated bool, saveable bool, wouldBlock bool, queue *waiter.Queue) (*descriptor, error) {
+	ownedFD := fd
+	origFD := -1
+	if saveable {
+		var err error
+		ownedFD, err = syscall.Dup(fd)
+		if err != nil {
+			return nil, err
+		}
+		origFD = fd
+	}
+	if wouldBlock {
+		if err := syscall.SetNonblock(ownedFD, true); err != nil {
+			return nil, err
+		}
+		if err := fdnotifier.AddFD(int32(ownedFD), queue); err != nil {
+			return nil, err
+		}
+	}
+	return &descriptor{
+		donated:    donated,
+		origFD:     origFD,
+		wouldBlock: wouldBlock,
+		value:      ownedFD,
+	}, nil
+}
+
+// initAfterLoad initializes the value of the descriptor after Load.
+func (d *descriptor) initAfterLoad(mo *superOperations, id uint64, queue *waiter.Queue) error {
+	if d.donated {
+		var err error
+		d.value, err = syscall.Dup(d.origFD)
+		if err != nil {
+			return fmt.Errorf("failed to dup restored fd %d: %v", d.origFD, err)
+		}
+	} else {
+		name, ok := mo.inodeMappings[id]
+		if !ok {
+			return fmt.Errorf("failed to find path for inode number %d", id)
+		}
+		fullpath := path.Join(mo.root, name)
+
+		var err error
+		d.value, err = open(nil, fullpath)
+		if err != nil {
+			return fmt.Errorf("failed to open %q: %v", fullpath, err)
+		}
+	}
+	if d.wouldBlock {
+		if err := syscall.SetNonblock(d.value, true); err != nil {
+			return err
+		}
+		if err := fdnotifier.AddFD(int32(d.value), queue); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// Release releases all resources held by descriptor.
+func (d *descriptor) Release() {
+	if d.wouldBlock {
+		fdnotifier.RemoveFD(int32(d.value))
+	}
+	if err := syscall.Close(d.value); err != nil {
+		log.Warningf("error closing fd %d: %v", d.value, err)
+	}
+	d.value = -1
+}
diff --git a/pkg/sentry/fs/host/descriptor_state.go b/pkg/sentry/fs/host/descriptor_state.go
new file mode 100644
index 000000000..7fb274451
--- /dev/null
+++ b/pkg/sentry/fs/host/descriptor_state.go
@@ -0,0 +1,29 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+// beforeSave is invoked by stateify.
+func (d *descriptor) beforeSave() {
+	if d.donated && d.origFD < 0 {
+		panic("donated file descriptor cannot be saved")
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (d *descriptor) afterLoad() {
+	// value must be manually restored by the descriptor's parent using
+	// initAfterLoad.
+	d.value = -1
+}
diff --git a/pkg/sentry/fs/host/device.go b/pkg/sentry/fs/host/device.go
new file mode 100644
index 000000000..f2a0b6b15
--- /dev/null
+++ b/pkg/sentry/fs/host/device.go
@@ -0,0 +1,25 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+)
+
+// hostFileDevice is the host file virtual device.
+var hostFileDevice = device.NewAnonMultiDevice()
+
+// hostPipeDevice is the host pipe virtual device.
+var hostPipeDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/host/file.go b/pkg/sentry/fs/host/file.go
new file mode 100644
index 000000000..bdf844337
--- /dev/null
+++ b/pkg/sentry/fs/host/file.go
@@ -0,0 +1,371 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"syscall"
+
+	"golang.org/x/sys/unix"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/secio"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+// fileOperations implements fs.FileOperations for a host file descriptor.
+type fileOperations struct {
+	fsutil.NoopRelease `state:"nosave"`
+
+	// iops are the Inode operations for this file.
+	iops *inodeOperations `state:"wait"`
+
+	// a scratch buffer for reading directory entries.
+	dirinfo *dirInfo `state:"nosave"`
+
+	// dirCursor is the directory cursor.
+	dirCursor string
+
+	// allowIoctl determines whether ioctls should be passed through to the
+	// host.
+	allowIoctl bool
+}
+
+// fileOperations implements fs.FileOperations.
+var _ fs.FileOperations = (*fileOperations)(nil)
+
+// NewFile creates a new File backed by the provided host file descriptor. If
+// NewFile succeeds, ownership of the fd is transferred to the returned File.
+//
+// The returned File cannot be saved, since there is no guarantee that the same
+// fd will exist or represent the same file at time of restore. If such a
+// guarantee does exist, use ImportFile instead.
+func NewFile(ctx context.Context, fd int, mounter fs.FileOwner) (*fs.File, error) {
+	return newFileFromDonatedFD(ctx, fd, mounter, false, false)
+}
+
+// ImportFile creates a new File backed by the provided host file descriptor.
+// Unlike NewFile, the file descriptor used by the File is duped from fd to
+// ensure that later changes to fd are not reflected by the fs.File.
+//
+// If the returned file is saved, it will be restored by re-importing the fd
+// originally passed to ImportFile. It is the restorer's responsibility to
+// ensure that the fd represents the same file.
+func ImportFile(ctx context.Context, fd int, mounter fs.FileOwner, allowIoctl bool) (*fs.File, error) {
+	return newFileFromDonatedFD(ctx, fd, mounter, true, allowIoctl)
+}
+
+// newFileFromDonatedFD returns an fs.File from a donated fd. If the fd is
+// saveable, then saveable is true.
+func newFileFromDonatedFD(ctx context.Context, donated int, mounter fs.FileOwner, saveable, allowIoctl bool) (*fs.File, error) {
+	var s syscall.Stat_t
+	if err := syscall.Fstat(donated, &s); err != nil {
+		return nil, err
+	}
+	switch s.Mode & syscall.S_IFMT {
+	case syscall.S_IFSOCK:
+		flags, err := fileFlagsFromDonatedFD(donated)
+		if err != nil {
+			return nil, err
+		}
+		s, err := newSocket(ctx, donated, saveable)
+		if err != nil {
+			return nil, err
+		}
+		s.SetFlags(fs.SettableFileFlags{
+			NonBlocking: flags.NonBlocking,
+		})
+		return s, nil
+	default:
+		flags, err := fileFlagsFromDonatedFD(donated)
+		if err != nil {
+			return nil, err
+		}
+		msrc := newMountSource(ctx, "/", mounter, &Filesystem{}, fs.MountSourceFlags{}, false /* dontTranslateOwnership */)
+		inode, err := newInode(ctx, msrc, donated, saveable, true /* donated */)
+		if err != nil {
+			return nil, err
+		}
+		iops := inode.InodeOperations.(*inodeOperations)
+
+		name := fmt.Sprintf("host:[%d]", inode.StableAttr.InodeID)
+		dirent := fs.NewDirent(inode, name)
+		defer dirent.DecRef()
+
+		return newFile(ctx, dirent, flags, iops, allowIoctl), nil
+	}
+}
+
+func fileFlagsFromDonatedFD(donated int) (fs.FileFlags, error) {
+	flags, _, errno := syscall.Syscall(syscall.SYS_FCNTL, uintptr(donated), syscall.F_GETFL, 0)
+	if errno != 0 {
+		log.Warningf("Failed to get file flags for donated fd %d (errno=%d)", donated, errno)
+		return fs.FileFlags{}, syscall.EIO
+	}
+	accmode := flags & syscall.O_ACCMODE
+	return fs.FileFlags{
+		Direct:      flags&syscall.O_DIRECT != 0,
+		NonBlocking: flags&syscall.O_NONBLOCK != 0,
+		Sync:        flags&syscall.O_SYNC != 0,
+		Append:      flags&syscall.O_APPEND != 0,
+		Read:        accmode == syscall.O_RDONLY || accmode == syscall.O_RDWR,
+		Write:       accmode == syscall.O_WRONLY || accmode == syscall.O_RDWR,
+	}, nil
+}
+
+// newFile returns a new fs.File.
+func newFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags, iops *inodeOperations, allowIoctl bool) *fs.File {
+	if !iops.ReturnsWouldBlock() {
+		// Allow reading/writing at an arbitrary offset for files
+		// that support it.
+		flags.Pread = true
+		flags.Pwrite = true
+	}
+	return fs.NewFile(ctx, dirent, flags, &fileOperations{
+		iops:       iops,
+		allowIoctl: allowIoctl,
+	})
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (f *fileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	f.iops.fileState.queue.EventRegister(e, mask)
+	fdnotifier.UpdateFD(int32(f.iops.fileState.FD()))
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (f *fileOperations) EventUnregister(e *waiter.Entry) {
+	f.iops.fileState.queue.EventUnregister(e)
+	fdnotifier.UpdateFD(int32(f.iops.fileState.FD()))
+}
+
+// Readiness uses the poll() syscall to check the status of the underlying FD.
+func (f *fileOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return fdnotifier.NonBlockingPoll(int32(f.iops.fileState.FD()), mask)
+}
+
+// Readdir implements fs.FileOperations.Readdir.
+func (f *fileOperations) Readdir(ctx context.Context, file *fs.File, serializer fs.DentrySerializer) (int64, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+	dirCtx := &fs.DirCtx{
+		Serializer: serializer,
+		DirCursor:  &f.dirCursor,
+	}
+	return fs.DirentReaddir(ctx, file.Dirent, f, root, dirCtx, file.Offset())
+}
+
+// IterateDir implements fs.DirIterator.IterateDir.
+func (f *fileOperations) IterateDir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	if f.dirinfo == nil {
+		f.dirinfo = new(dirInfo)
+		f.dirinfo.buf = make([]byte, usermem.PageSize)
+	}
+	entries, err := f.iops.readdirAll(f.dirinfo)
+	if err != nil {
+		return offset, err
+	}
+	count, err := fs.GenericReaddir(dirCtx, fs.NewSortedDentryMap(entries))
+	return offset + count, err
+}
+
+// Write implements fs.FileOperations.Write.
+func (f *fileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	// Would this file block?
+	if f.iops.ReturnsWouldBlock() {
+		// These files can't be memory mapped, assert this. This also
+		// means that writes do not need to synchronize with memory
+		// mappings nor metadata cached by this file's fs.Inode.
+		if canMap(file.Dirent.Inode) {
+			panic("files that can return EWOULDBLOCK cannot be memory mapped")
+		}
+		// Ignore the offset, these files don't support writing at
+		// an arbitrary offset.
+		writer := fd.NewReadWriter(f.iops.fileState.FD())
+		n, err := src.CopyInTo(ctx, safemem.FromIOWriter{writer})
+		if isBlockError(err) {
+			err = syserror.ErrWouldBlock
+		}
+		return n, err
+	}
+	if !file.Dirent.Inode.MountSource.Flags.ForcePageCache {
+		writer := secio.NewOffsetWriter(fd.NewReadWriter(f.iops.fileState.FD()), offset)
+		return src.CopyInTo(ctx, safemem.FromIOWriter{writer})
+	}
+	return f.iops.cachingInodeOps.Write(ctx, src, offset)
+}
+
+// Read implements fs.FileOperations.Read.
+func (f *fileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	// Would this file block?
+	if f.iops.ReturnsWouldBlock() {
+		// These files can't be memory mapped, assert this. This also
+		// means that reads do not need to synchronize with memory
+		// mappings nor metadata cached by this file's fs.Inode.
+		if canMap(file.Dirent.Inode) {
+			panic("files that can return EWOULDBLOCK cannot be memory mapped")
+		}
+		// Ignore the offset, these files don't support reading at
+		// an arbitrary offset.
+		reader := fd.NewReadWriter(f.iops.fileState.FD())
+		n, err := dst.CopyOutFrom(ctx, safemem.FromIOReader{reader})
+		if isBlockError(err) {
+			// If we got any data at all, return it as a "completed" partial read
+			// rather than retrying until complete.
+			if n != 0 {
+				err = nil
+			} else {
+				err = syserror.ErrWouldBlock
+			}
+		}
+		return n, err
+	}
+	if !file.Dirent.Inode.MountSource.Flags.ForcePageCache {
+		reader := secio.NewOffsetReader(fd.NewReadWriter(f.iops.fileState.FD()), offset)
+		return dst.CopyOutFrom(ctx, safemem.FromIOReader{reader})
+	}
+	return f.iops.cachingInodeOps.Read(ctx, file, dst, offset)
+}
+
+// Fsync implements fs.FileOperations.Fsync.
+func (f *fileOperations) Fsync(ctx context.Context, file *fs.File, start int64, end int64, syncType fs.SyncType) error {
+	switch syncType {
+	case fs.SyncAll, fs.SyncData:
+		if err := file.Dirent.Inode.WriteOut(ctx); err != nil {
+			return err
+		}
+		fallthrough
+	case fs.SyncBackingStorage:
+		return syscall.Fsync(f.iops.fileState.FD())
+	}
+	panic("invalid sync type")
+}
+
+// Flush implements fs.FileOperations.Flush.
+func (f *fileOperations) Flush(context.Context, *fs.File) error {
+	// This is a no-op because flushing the resource backing this
+	// file would mean closing it. We can't do that because other
+	// open files may depend on the backing host fd.
+	return nil
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (f *fileOperations) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	if !canMap(file.Dirent.Inode) {
+		return syserror.ENODEV
+	}
+	return fsutil.GenericConfigureMMap(file, f.iops.cachingInodeOps, opts)
+}
+
+// Seek implements fs.FileOperations.Seek.
+func (f *fileOperations) Seek(ctx context.Context, file *fs.File, whence fs.SeekWhence, offset int64) (int64, error) {
+	return fsutil.SeekWithDirCursor(ctx, file, whence, offset, &f.dirCursor)
+}
+
+// Ioctl implements fs.FileOperations.Iocotl.
+func (f *fileOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	if !f.allowIoctl {
+		return 0, syserror.ENOTTY
+	}
+	// Ignore arg[0].  This is the real FD:
+	fd := f.iops.fileState.FD()
+	ioctl := args[1].Uint64()
+	switch ioctl {
+	case unix.TCGETS:
+		termios, err := ioctlGetTermios(fd)
+		if err != nil {
+			return 0, err
+		}
+		_, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), termios, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+
+	case unix.TCSETS, unix.TCSETSW:
+		var termios linux.Termios
+		if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &termios, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, err
+		}
+		err := ioctlSetTermios(fd, ioctl, &termios)
+		return 0, err
+
+	case unix.TIOCGPGRP:
+		// Args: pid_t *argp
+		// When successful, equivalent to *argp = tcgetpgrp(fd).
+		// Get the process group ID of the foreground process group on
+		// this terminal.
+
+		t := kernel.TaskFromContext(ctx)
+		if t == nil {
+			panic(fmt.Sprintf("cannot get thread group from context %v", ctx))
+		}
+		tid := t.ThreadID()
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &tid, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+
+	case unix.TIOCSPGRP:
+		// Args: const pid_t *argp
+		// Equivalent to tcsetpgrp(fd, *argp).
+		// Set the foreground process group ID of this terminal.
+
+		// Not much we can do with this one at the moment, so we just
+		// lie and pretend everything is great. Bash and Sh seem fine
+		// with this.
+		log.Warningf("Ignoring application ioctl(TIOCSPGRP) call")
+		return 0, nil
+
+	case unix.TIOCGWINSZ:
+		// Args: struct winsize *argp
+		// Get window size.
+		winsize, err := unix.IoctlGetWinsize(fd, unix.TIOCGWINSZ)
+		if err != nil {
+			return 0, err
+		}
+		_, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), winsize, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+
+	case unix.TIOCSWINSZ:
+		// Args: const struct winsize *argp
+		// Set window size.
+		var winsize unix.Winsize
+		if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &winsize, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, err
+		}
+		err := unix.IoctlSetWinsize(fd, unix.TIOCSWINSZ, &winsize)
+		return 0, err
+
+	default:
+		return 0, syserror.ENOTTY
+	}
+}
diff --git a/pkg/sentry/fs/host/fs.go b/pkg/sentry/fs/host/fs.go
new file mode 100644
index 000000000..ffd55a5ab
--- /dev/null
+++ b/pkg/sentry/fs/host/fs.go
@@ -0,0 +1,327 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package host implements an fs.Filesystem for files backed by host
+// file descriptors.
+package host
+
+import (
+	"fmt"
+	"path"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// FilesystemName is the name under which Filesystem is registered.
+const FilesystemName = "whitelistfs"
+
+const (
+	// whitelistKey is the mount option containing a comma-separated list
+	// of host paths to whitelist.
+	whitelistKey = "whitelist"
+
+	// rootPathKey is the mount option containing the root path of the
+	// mount.
+	rootPathKey = "root"
+
+	// dontTranslateOwnershipKey is the key to superOperations.dontTranslateOwnership.
+	dontTranslateOwnershipKey = "dont_translate_ownership"
+)
+
+// maxTraversals determines link traversals in building the whitelist.
+const maxTraversals = 10
+
+// Filesystem is a pseudo file system that is only available during the setup
+// to lock down the configurations. This filesystem should only be mounted at root.
+//
+// Think twice before exposing this to applications.
+type Filesystem struct {
+	// whitelist is a set of host paths to whitelist.
+	paths []string
+}
+
+// Name is the identifier of this file system.
+func (*Filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount prohibits users from using mount(2) with this file system.
+func (*Filesystem) AllowUserMount() bool {
+	return false
+}
+
+// Flags returns that there is nothing special about this file system.
+func (*Filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns an fs.Inode exposing the host file system.  It is intended to be locked
+// down in PreExec below.
+func (f *Filesystem) Mount(ctx context.Context, _ string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// Parse generic comma-separated key=value options.
+	options := fs.GenericMountSourceOptions(data)
+
+	// Grab the whitelist if one was specified.
+	// TODO: require another option "testonly" in order to allow
+	// no whitelist.
+	if wl, ok := options[whitelistKey]; ok {
+		f.paths = strings.Split(wl, "|")
+		delete(options, whitelistKey)
+	}
+
+	// If the rootPath was set, use it. Othewise default to the root of the
+	// host fs.
+	rootPath := "/"
+	if rp, ok := options[rootPathKey]; ok {
+		rootPath = rp
+		delete(options, rootPathKey)
+
+		// We must relativize the whitelisted paths to the new root.
+		for i, p := range f.paths {
+			rel, err := filepath.Rel(rootPath, p)
+			if err != nil {
+				return nil, fmt.Errorf("whitelist path %q must be a child of root path %q", p, rootPath)
+			}
+			f.paths[i] = path.Join("/", rel)
+		}
+	}
+	fd, err := open(nil, rootPath)
+	if err != nil {
+		return nil, fmt.Errorf("failed to find root: %v", err)
+	}
+
+	var dontTranslateOwnership bool
+	if v, ok := options[dontTranslateOwnershipKey]; ok {
+		b, err := strconv.ParseBool(v)
+		if err != nil {
+			return nil, fmt.Errorf("invalid value for %q: %v", dontTranslateOwnershipKey, err)
+		}
+		dontTranslateOwnership = b
+		delete(options, dontTranslateOwnershipKey)
+	}
+
+	// Fail if the caller passed us more options than we know about.
+	if len(options) > 0 {
+		return nil, fmt.Errorf("unsupported mount options: %v", options)
+	}
+
+	// The mounting EUID/EGID will be cached by this file system. This will
+	// be used to assign ownership to files that we own.
+	owner := fs.FileOwnerFromContext(ctx)
+
+	// Construct the host file system mount and inode.
+	msrc := newMountSource(ctx, rootPath, owner, f, flags, dontTranslateOwnership)
+	return newInode(ctx, msrc, fd, false /* saveable */, false /* donated */)
+}
+
+// InstallWhitelist locks down the MountNamespace to only the currently installed
+// Dirents and the given paths.
+func (f *Filesystem) InstallWhitelist(ctx context.Context, m *fs.MountNamespace) error {
+	return installWhitelist(ctx, m, f.paths)
+}
+
+func installWhitelist(ctx context.Context, m *fs.MountNamespace, paths []string) error {
+	if len(paths) == 0 || (len(paths) == 1 && paths[0] == "") {
+		// Warning will be logged during filter installation if the empty
+		// whitelist matters (allows for host file access).
+		return nil
+	}
+
+	// Done tracks entries already added.
+	done := make(map[string]bool)
+	root := m.Root()
+	defer root.DecRef()
+
+	for i := 0; i < len(paths); i++ {
+		// Make sure the path is absolute. This is a sanity check.
+		if !path.IsAbs(paths[i]) {
+			return fmt.Errorf("path %q is not absolute", paths[i])
+		}
+
+		// We need to add all the intermediate paths, in case one of
+		// them is a symlink that needs to be resolved.
+		for j := 1; j <= len(paths[i]); j++ {
+			if j < len(paths[i]) && paths[i][j] != '/' {
+				continue
+			}
+			current := paths[i][:j]
+
+			// Lookup the given component in the tree.
+			d, err := m.FindLink(ctx, root, nil, current, maxTraversals)
+			if err != nil {
+				log.Warningf("populate failed for %q: %v", current, err)
+				continue
+			}
+
+			// It's critical that this DecRef happens after the
+			// freeze below. This ensures that the dentry is in
+			// place to be frozen. Otherwise, we freeze without
+			// these entries.
+			defer d.DecRef()
+
+			// Expand the last component if necessary.
+			if current == paths[i] {
+				// Is it a directory or symlink?
+				sattr := d.Inode.StableAttr
+				if fs.IsDir(sattr) {
+					for name := range childDentAttrs(ctx, d) {
+						paths = append(paths, path.Join(current, name))
+					}
+				}
+				if fs.IsSymlink(sattr) {
+					// Only expand symlinks once. The
+					// folder structure may contain
+					// recursive symlinks and we don't want
+					// to end up infinitely expanding this
+					// symlink. This is safe because this
+					// is the last component. If a later
+					// path wants to symlink something
+					// beneath this symlink that will still
+					// be handled by the FindLink above.
+					if done[current] {
+						continue
+					}
+
+					s, err := d.Inode.Readlink(ctx)
+					if err != nil {
+						log.Warningf("readlink failed for %q: %v", current, err)
+						continue
+					}
+					if path.IsAbs(s) {
+						paths = append(paths, s)
+					} else {
+						target := path.Join(path.Dir(current), s)
+						paths = append(paths, target)
+					}
+				}
+			}
+
+			// Only report this one once even though we may look
+			// it up more than once. If we whitelist /a/b,/a then
+			// /a will be "done" when it is looked up for /a/b,
+			// however we still need to expand all of its contents
+			// when whitelisting /a.
+			if !done[current] {
+				log.Debugf("whitelisted: %s", current)
+			}
+			done[current] = true
+		}
+	}
+
+	// Freeze the mount tree in place. This prevents any new paths from
+	// being opened and any old ones from being removed. If we do provide
+	// tmpfs mounts, we'll want to freeze/thaw those separately.
+	m.Freeze()
+	return nil
+}
+
+func childDentAttrs(ctx context.Context, d *fs.Dirent) map[string]fs.DentAttr {
+	dirname, _ := d.FullName(nil /* root */)
+	dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true})
+	if err != nil {
+		log.Warningf("failed to open directory %q: %v", dirname, err)
+		return nil
+	}
+	dir.DecRef()
+	var stubSerializer fs.CollectEntriesSerializer
+	if err := dir.Readdir(ctx, &stubSerializer); err != nil {
+		log.Warningf("failed to iterate on host directory %q: %v", dirname, err)
+		return nil
+	}
+	delete(stubSerializer.Entries, ".")
+	delete(stubSerializer.Entries, "..")
+	return stubSerializer.Entries
+}
+
+// newMountSource constructs a new host fs.MountSource
+// relative to a root path. The root should match the mount point.
+func newMountSource(ctx context.Context, root string, mounter fs.FileOwner, filesystem fs.Filesystem, flags fs.MountSourceFlags, dontTranslateOwnership bool) *fs.MountSource {
+	return fs.NewMountSource(&superOperations{
+		root:                   root,
+		inodeMappings:          make(map[uint64]string),
+		mounter:                mounter,
+		dontTranslateOwnership: dontTranslateOwnership,
+	}, filesystem, flags)
+}
+
+// superOperations implements fs.MountSourceOperations.
+type superOperations struct {
+	fs.SimpleMountSourceOperations `state:"nosave"`
+
+	// root is the path of the mount point. All inode mappings
+	// are relative to this root.
+	root string
+
+	// inodeMappings contains mappings of fs.Inodes associated
+	// with this MountSource to paths under root.
+	inodeMappings map[uint64]string
+
+	// mounter is the cached EUID/EGID that mounted this file system.
+	mounter fs.FileOwner
+
+	// dontTranslateOwnership indicates whether to not translate file
+	// ownership.
+	//
+	// By default, files/directories owned by the sandbox uses UID/GID
+	// of the mounter. For files/directories that are not owned by the
+	// sandbox, file UID/GID is translated to a UID/GID which cannot
+	// be mapped in the sandboxed application's user namespace. The
+	// UID/GID will look like the nobody UID/GID (65534) but is not
+	// strictly owned by the user "nobody".
+	//
+	// If whitelistfs is a lower filesystem in an overlay, set
+	// dont_translate_ownership=true in mount options.
+	dontTranslateOwnership bool
+}
+
+var _ fs.MountSourceOperations = (*superOperations)(nil)
+
+// ResetInodeMappings implements fs.MountSourceOperations.ResetInodeMappings.
+func (m *superOperations) ResetInodeMappings() {
+	m.inodeMappings = make(map[uint64]string)
+}
+
+// SaveInodeMapping implements fs.MountSourceOperations.SaveInodeMapping.
+func (m *superOperations) SaveInodeMapping(inode *fs.Inode, path string) {
+	// This is very unintuitive. We *CANNOT* trust the inode's StableAttrs,
+	// because overlay copyUp may have changed them out from under us.
+	// So much for "immutable".
+	sattr := inode.InodeOperations.(*inodeOperations).fileState.sattr
+	m.inodeMappings[sattr.InodeID] = path
+}
+
+// Keep implements fs.MountSourceOperations.Keep.
+//
+// TODO: It is possible to change the permissions on a
+// host file while it is in the dirent cache (say from RO to RW), but it is not
+// possible to re-open the file with more relaxed permissions, since the host
+// FD is already open and stored in the inode.
+//
+// Using the dirent LRU cache increases the odds that this bug is encountered.
+// Since host file access is relatively fast anyways, we disable the LRU cache
+// for host fs files.  Once we can properly deal with permissions changes and
+// re-opening host files, we should revisit whether or not to make use of the
+// LRU cache.
+func (*superOperations) Keep(*fs.Dirent) bool {
+	return false
+}
+
+func init() {
+	fs.RegisterFilesystem(&Filesystem{})
+}
diff --git a/pkg/sentry/fs/host/fs_test.go b/pkg/sentry/fs/host/fs_test.go
new file mode 100644
index 000000000..c000afc49
--- /dev/null
+++ b/pkg/sentry/fs/host/fs_test.go
@@ -0,0 +1,383 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"path"
+	"reflect"
+	"sort"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// newTestMountNamespace creates a MountNamespace with a ramfs root.
+// It returns the host folder created, which should be removed when done.
+func newTestMountNamespace(t *testing.T) (*fs.MountNamespace, string, error) {
+	p, err := ioutil.TempDir("", "root")
+	if err != nil {
+		return nil, "", err
+	}
+
+	fd, err := open(nil, p)
+	if err != nil {
+		os.RemoveAll(p)
+		return nil, "", err
+	}
+	ctx := contexttest.Context(t)
+	root, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false)
+	if err != nil {
+		os.RemoveAll(p)
+		return nil, "", err
+	}
+	mm, err := fs.NewMountNamespace(ctx, root)
+	if err != nil {
+		os.RemoveAll(p)
+		return nil, "", err
+	}
+	return mm, p, nil
+}
+
+// createTestDirs populates the root with some test files and directories.
+// /a/a1.txt
+// /a/a2.txt
+// /b/b1.txt
+// /b/c/c1.txt
+// /symlinks/normal.txt
+// /symlinks/to_normal.txt -> /symlinks/normal.txt
+// /symlinks/recursive -> /symlinks
+func createTestDirs(ctx context.Context, t *testing.T, m *fs.MountNamespace) error {
+	r := m.Root()
+	defer r.DecRef()
+
+	if err := r.CreateDirectory(ctx, r, "a", fs.FilePermsFromMode(0777)); err != nil {
+		return err
+	}
+
+	a, err := r.Walk(ctx, r, "a")
+	if err != nil {
+		return err
+	}
+	defer a.DecRef()
+
+	a1, err := a.Create(ctx, r, "a1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+	if err != nil {
+		return err
+	}
+	a1.DecRef()
+
+	a2, err := a.Create(ctx, r, "a2.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+	if err != nil {
+		return err
+	}
+	a2.DecRef()
+
+	if err := r.CreateDirectory(ctx, r, "b", fs.FilePermsFromMode(0777)); err != nil {
+		return err
+	}
+
+	b, err := r.Walk(ctx, r, "b")
+	if err != nil {
+		return err
+	}
+	defer b.DecRef()
+
+	b1, err := b.Create(ctx, r, "b1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+	if err != nil {
+		return err
+	}
+	b1.DecRef()
+
+	if err := b.CreateDirectory(ctx, r, "c", fs.FilePermsFromMode(0777)); err != nil {
+		return err
+	}
+
+	c, err := b.Walk(ctx, r, "c")
+	if err != nil {
+		return err
+	}
+	defer c.DecRef()
+
+	c1, err := c.Create(ctx, r, "c1.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+	if err != nil {
+		return err
+	}
+	c1.DecRef()
+
+	if err := r.CreateDirectory(ctx, r, "symlinks", fs.FilePermsFromMode(0777)); err != nil {
+		return err
+	}
+
+	symlinks, err := r.Walk(ctx, r, "symlinks")
+	if err != nil {
+		return err
+	}
+	defer symlinks.DecRef()
+
+	normal, err := symlinks.Create(ctx, r, "normal.txt", fs.FileFlags{Read: true, Write: true}, fs.FilePermsFromMode(0666))
+	if err != nil {
+		return err
+	}
+	normal.DecRef()
+
+	if err := symlinks.CreateLink(ctx, r, "/symlinks/normal.txt", "to_normal.txt"); err != nil {
+		return err
+	}
+
+	if err := symlinks.CreateLink(ctx, r, "/symlinks", "recursive"); err != nil {
+		return err
+	}
+
+	return nil
+}
+
+// allPaths returns a slice of all paths of entries visible in the rootfs.
+func allPaths(ctx context.Context, t *testing.T, m *fs.MountNamespace, base string) ([]string, error) {
+	var paths []string
+	root := m.Root()
+	defer root.DecRef()
+
+	d, err := m.FindLink(ctx, root, nil, base, 1)
+	if err != nil {
+		t.Logf("FindLink failed for %q", base)
+		return paths, err
+	}
+	defer d.DecRef()
+
+	if fs.IsDir(d.Inode.StableAttr) {
+		dir, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true})
+		if err != nil {
+			return nil, fmt.Errorf("failed to open directory %q: %v", base, err)
+		}
+		iter, ok := dir.FileOperations.(fs.DirIterator)
+		if !ok {
+			return nil, fmt.Errorf("cannot directly iterate on host directory %q", base)
+		}
+		dirCtx := &fs.DirCtx{
+			Serializer: noopDentrySerializer{},
+		}
+		if _, err := fs.DirentReaddir(ctx, d, iter, root, dirCtx, 0); err != nil {
+			return nil, err
+		}
+		for name := range dirCtx.DentAttrs() {
+			if name == "." || name == ".." {
+				continue
+			}
+
+			fullName := path.Join(base, name)
+			paths = append(paths, fullName)
+
+			// Recurse.
+			subpaths, err := allPaths(ctx, t, m, fullName)
+			if err != nil {
+				return paths, err
+			}
+			paths = append(paths, subpaths...)
+		}
+	}
+
+	return paths, nil
+}
+
+type noopDentrySerializer struct{}
+
+func (noopDentrySerializer) CopyOut(string, fs.DentAttr) error {
+	return nil
+}
+func (noopDentrySerializer) Written() int {
+	return 4096
+}
+
+// pathsEqual returns true if the two string slices contain the same entries.
+func pathsEqual(got, want []string) bool {
+	sort.Strings(got)
+	sort.Strings(want)
+
+	if len(got) != len(want) {
+		return false
+	}
+
+	for i := range got {
+		if got[i] != want[i] {
+			return false
+		}
+	}
+
+	return true
+}
+
+func TestWhitelist(t *testing.T) {
+	for _, test := range []struct {
+		// description of the test.
+		desc string
+		// paths are the paths to whitelist
+		paths []string
+		// want are all of the directory entries that should be
+		// visible (nothing beyond this set should be visible).
+		want []string
+	}{
+		{
+			desc:  "root",
+			paths: []string{"/"},
+			want:  []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt", "/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt", "/symlinks/recursive"},
+		},
+		{
+			desc:  "top-level directories",
+			paths: []string{"/a", "/b"},
+			want:  []string{"/a", "/a/a1.txt", "/a/a2.txt", "/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"},
+		},
+		{
+			desc:  "nested directories (1/2)",
+			paths: []string{"/b", "/b/c"},
+			want:  []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"},
+		},
+		{
+			desc:  "nested directories (2/2)",
+			paths: []string{"/b/c", "/b"},
+			want:  []string{"/b", "/b/b1.txt", "/b/c", "/b/c/c1.txt"},
+		},
+		{
+			desc:  "single file",
+			paths: []string{"/b/c/c1.txt"},
+			want:  []string{"/b", "/b/c", "/b/c/c1.txt"},
+		},
+		{
+			desc:  "single file and directory",
+			paths: []string{"/a/a1.txt", "/b/c"},
+			want:  []string{"/a", "/a/a1.txt", "/b", "/b/c", "/b/c/c1.txt"},
+		},
+		{
+			desc:  "symlink",
+			paths: []string{"/symlinks/to_normal.txt"},
+			want:  []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/to_normal.txt"},
+		},
+		{
+			desc:  "recursive symlink",
+			paths: []string{"/symlinks/recursive/normal.txt"},
+			want:  []string{"/symlinks", "/symlinks/normal.txt", "/symlinks/recursive"},
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			m, p, err := newTestMountNamespace(t)
+			if err != nil {
+				t.Errorf("Failed to create MountNamespace: %v", err)
+			}
+			defer os.RemoveAll(p)
+
+			ctx := withRoot(contexttest.RootContext(t), m.Root())
+			if err := createTestDirs(ctx, t, m); err != nil {
+				t.Errorf("Failed to create test dirs: %v", err)
+			}
+
+			if err := installWhitelist(ctx, m, test.paths); err != nil {
+				t.Errorf("installWhitelist(%v) err got %v want nil", test.paths, err)
+			}
+
+			got, err := allPaths(ctx, t, m, "/")
+			if err != nil {
+				t.Fatalf("Failed to lookup paths (whitelisted: %v): %v", test.paths, err)
+			}
+
+			if !pathsEqual(got, test.want) {
+				t.Errorf("For paths %v got %v want %v", test.paths, got, test.want)
+			}
+		})
+	}
+}
+
+func TestRootPath(t *testing.T) {
+	// Create a temp dir, which will be the root of our mounted fs.
+	rootPath, err := ioutil.TempDir(os.TempDir(), "root")
+	if err != nil {
+		t.Fatalf("TempDir failed: %v", err)
+	}
+	defer os.RemoveAll(rootPath)
+
+	// Create two files inside the new root, one which will be whitelisted
+	// and one not.
+	whitelisted, err := ioutil.TempFile(rootPath, "white")
+	if err != nil {
+		t.Fatalf("TempFile failed: %v", err)
+	}
+	if _, err := ioutil.TempFile(rootPath, "black"); err != nil {
+		t.Fatalf("TempFile failed: %v", err)
+	}
+
+	// Create a mount with a root path and single whitelisted file.
+	hostFS := &Filesystem{}
+	ctx := contexttest.Context(t)
+	data := fmt.Sprintf("%s=%s,%s=%s", rootPathKey, rootPath, whitelistKey, whitelisted.Name())
+	inode, err := hostFS.Mount(ctx, "", fs.MountSourceFlags{}, data)
+	if err != nil {
+		t.Fatalf("Mount failed: %v", err)
+	}
+	mm, err := fs.NewMountNamespace(ctx, inode)
+	if err != nil {
+		t.Fatalf("NewMountNamespace failed: %v", err)
+	}
+	if err := hostFS.InstallWhitelist(ctx, mm); err != nil {
+		t.Fatalf("InstallWhitelist failed: %v", err)
+	}
+
+	// Get the contents of the root directory.
+	rootDir := mm.Root()
+	rctx := withRoot(ctx, rootDir)
+	f, err := rootDir.Inode.GetFile(rctx, rootDir, fs.FileFlags{})
+	if err != nil {
+		t.Fatalf("GetFile failed: %v", err)
+	}
+	c := &fs.CollectEntriesSerializer{}
+	if err := f.Readdir(rctx, c); err != nil {
+		t.Fatalf("Readdir failed: %v", err)
+	}
+
+	// We should have only our whitelisted file, plus the dots.
+	want := []string{path.Base(whitelisted.Name()), ".", ".."}
+	got := c.Order
+	sort.Strings(want)
+	sort.Strings(got)
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("Readdir got %v, wanted %v", got, want)
+	}
+}
+
+type rootContext struct {
+	context.Context
+	root *fs.Dirent
+}
+
+// withRoot returns a copy of ctx with the given root.
+func withRoot(ctx context.Context, root *fs.Dirent) context.Context {
+	return &rootContext{
+		Context: ctx,
+		root:    root,
+	}
+}
+
+// Value implements Context.Value.
+func (rc rootContext) Value(key interface{}) interface{} {
+	switch key {
+	case fs.CtxRoot:
+		rc.root.IncRef()
+		return rc.root
+	default:
+		return rc.Context.Value(key)
+	}
+}
diff --git a/pkg/sentry/fs/host/inode.go b/pkg/sentry/fs/host/inode.go
new file mode 100644
index 000000000..226bc5164
--- /dev/null
+++ b/pkg/sentry/fs/host/inode.go
@@ -0,0 +1,506 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/secio"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// inodeOperations implements fs.InodeOperations for an fs.Inodes backed
+// by a host file descriptor.
+type inodeOperations struct {
+	fsutil.InodeNotVirtual           `state:"nosave"`
+	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.DeprecatedFileOperations  `state:"nosave"`
+
+	// fileState implements fs.CachedFileObject. It exists
+	// to break a circular load dependency between inodeOperations
+	// and cachingInodeOps (below).
+	fileState *inodeFileState `state:"wait"`
+
+	// cachedInodeOps implements memmap.Mappable.
+	cachingInodeOps *fsutil.CachingInodeOperations
+
+	// readdirMu protects the file offset on the host FD. This is needed
+	// for readdir because getdents must use the kernel offset, so
+	// concurrent readdirs must be exclusive.
+	//
+	// All read/write functions pass the offset directly to the kernel and
+	// thus don't need a lock.
+	readdirMu sync.Mutex `state:"nosave"`
+}
+
+// inodeFileState implements fs.CachedFileObject and otherwise fully
+// encapsulates state that needs to be manually loaded on restore for
+// this file object.
+//
+// This unfortunate structure exists because fs.CachingInodeOperations
+// defines afterLoad and therefore cannot be lazily loaded (to break a
+// circular load dependency between it and inodeOperations). Even with
+// lazy loading, this approach defines the dependencies between objects
+// and the expected load behavior more concretely.
+type inodeFileState struct {
+	// Common file system state.
+	mops *superOperations `state:"wait"`
+
+	// descriptor is the backing host fd.
+	descriptor *descriptor `state:"wait"`
+
+	// Event queue for blocking operations.
+	queue waiter.Queue `state:"nosave"`
+
+	// sattr is used to restore the inodeOperations.
+	sattr fs.StableAttr `state:"wait"`
+
+	// savedUAttr is only allocated during S/R. It points to the save-time
+	// unstable attributes and is used to validate restore-time ones.
+	//
+	// Note that these unstable attributes are only used to detect cross-S/R
+	// external file system metadata changes. They may differ from the
+	// cached unstable attributes in cachingInodeOps, as that might differ
+	// from the external file system attributes if there had been WriteOut
+	// failures. S/R is transparent to Sentry and the latter will continue
+	// using its cached values after restore.
+	savedUAttr *fs.UnstableAttr
+}
+
+// ReadToBlocksAt implements fsutil.CachedFileObject.ReadToBlocksAt.
+func (i *inodeFileState) ReadToBlocksAt(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error) {
+	// TODO: Using safemem.FromIOReader here is wasteful for two
+	// reasons:
+	//
+	// - Using preadv instead of iterated preads saves on host system calls.
+	//
+	// - Host system calls can handle destination memory that would fault in
+	// gr3 (i.e. they can accept safemem.Blocks with NeedSafecopy() == true),
+	// so the buffering performed by FromIOReader is unnecessary.
+	//
+	// This also applies to the write path below.
+	return safemem.FromIOReader{secio.NewOffsetReader(fd.NewReadWriter(i.FD()), int64(offset))}.ReadToBlocks(dsts)
+}
+
+// WriteFromBlocksAt implements fsutil.CachedFileObject.WriteFromBlocksAt.
+func (i *inodeFileState) WriteFromBlocksAt(ctx context.Context, srcs safemem.BlockSeq, offset uint64) (uint64, error) {
+	return safemem.FromIOWriter{secio.NewOffsetWriter(fd.NewReadWriter(i.FD()), int64(offset))}.WriteFromBlocks(srcs)
+}
+
+// SetMaskedAttributes implements fsutil.CachedFileObject.SetMaskedAttributes.
+func (i *inodeFileState) SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error {
+	if mask.Empty() {
+		return nil
+	}
+	if mask.UID || mask.GID {
+		return syserror.EPERM
+	}
+	if mask.Perms {
+		if err := syscall.Fchmod(i.FD(), uint32(attr.Perms.LinuxMode())); err != nil {
+			return err
+		}
+	}
+	if mask.Size {
+		if err := syscall.Ftruncate(i.FD(), attr.Size); err != nil {
+			return err
+		}
+	}
+	if mask.AccessTime || mask.ModificationTime {
+		ts := fs.TimeSpec{
+			ATime:     attr.AccessTime,
+			ATimeOmit: !mask.AccessTime,
+			MTime:     attr.ModificationTime,
+			MTimeOmit: !mask.ModificationTime,
+		}
+		if err := setTimestamps(i.FD(), ts); err != nil {
+			return err
+		}
+	}
+	return nil
+}
+
+// Sync implements fsutil.CachedFileObject.Sync.
+func (i *inodeFileState) Sync(ctx context.Context) error {
+	return syscall.Fsync(i.FD())
+}
+
+// FD implements fsutil.CachedFileObject.FD.
+func (i *inodeFileState) FD() int {
+	return i.descriptor.value
+}
+
+func (i *inodeFileState) unstableAttr(ctx context.Context) (fs.UnstableAttr, error) {
+	var s syscall.Stat_t
+	if err := syscall.Fstat(i.FD(), &s); err != nil {
+		return fs.UnstableAttr{}, err
+	}
+	return unstableAttr(i.mops, &s), nil
+}
+
+// inodeOperations implements fs.InodeOperations.
+var _ fs.InodeOperations = (*inodeOperations)(nil)
+
+// newInode returns a new fs.Inode backed by the host fd.
+func newInode(ctx context.Context, msrc *fs.MountSource, fd int, saveable bool, donated bool) (*fs.Inode, error) {
+	// Retrieve metadata.
+	var s syscall.Stat_t
+	err := syscall.Fstat(fd, &s)
+	if err != nil {
+		return nil, err
+	}
+
+	fileState := &inodeFileState{
+		mops:  msrc.MountSourceOperations.(*superOperations),
+		sattr: stableAttr(&s),
+	}
+
+	// Initialize the wrapped host file descriptor.
+	fileState.descriptor, err = newDescriptor(
+		fd,
+		donated,
+		saveable,
+		wouldBlock(&s),
+		&fileState.queue,
+	)
+	if err != nil {
+		return nil, err
+	}
+
+	// Build the fs.InodeOperations.
+	uattr := unstableAttr(msrc.MountSourceOperations.(*superOperations), &s)
+	iops := &inodeOperations{
+		fileState:       fileState,
+		cachingInodeOps: fsutil.NewCachingInodeOperations(ctx, fileState, uattr, msrc.Flags.ForcePageCache),
+	}
+
+	// Return the fs.Inode.
+	return fs.NewInode(iops, msrc, fileState.sattr), nil
+}
+
+// Mappable implements fs.InodeOperations.Mappable.
+func (i *inodeOperations) Mappable(inode *fs.Inode) memmap.Mappable {
+	if !canMap(inode) {
+		return nil
+	}
+	return i.cachingInodeOps
+}
+
+// ReturnsWouldBlock returns true if this host fd can return EWOULDBLOCK
+// for operations that would block.
+func (i *inodeOperations) ReturnsWouldBlock() bool {
+	return i.fileState.descriptor.wouldBlock
+}
+
+// Release implements fs.InodeOperations.Release.
+func (i *inodeOperations) Release(context.Context) {
+	i.fileState.descriptor.Release()
+	i.cachingInodeOps.Release()
+}
+
+// Lookup implements fs.InodeOperations.Lookup.
+func (i *inodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) {
+	// Get a new fd relative to i at name.
+	fd, err := open(i, name)
+	if err != nil {
+		if err == syserror.ENOENT {
+			return nil, syserror.ENOENT
+		}
+		return nil, err
+	}
+
+	inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */)
+	if err != nil {
+		return nil, err
+	}
+
+	// Return the fs.Dirent.
+	return fs.NewDirent(inode, name), nil
+}
+
+// Create implements fs.InodeOperations.Create.
+func (i *inodeOperations) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perm fs.FilePermissions) (*fs.File, error) {
+	// Create a file relative to i at name.
+	//
+	// N.B. We always open this file O_RDWR regardless of flags because a
+	// future GetFile might want more access. Open allows this regardless
+	// of perm.
+	fd, err := openAt(i, name, syscall.O_RDWR|syscall.O_CREAT|syscall.O_EXCL, perm.LinuxMode())
+	if err != nil {
+		return nil, err
+	}
+
+	inode, err := newInode(ctx, dir.MountSource, fd, false /* saveable */, false /* donated */)
+	if err != nil {
+		return nil, err
+	}
+
+	d := fs.NewDirent(inode, name)
+	defer d.DecRef()
+	return inode.GetFile(ctx, d, flags)
+}
+
+// CreateDirectory implements fs.InodeOperations.CreateDirectory.
+func (i *inodeOperations) CreateDirectory(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error {
+	return syscall.Mkdirat(i.fileState.FD(), name, uint32(perm.LinuxMode()))
+}
+
+// CreateLink implements fs.InodeOperations.CreateLink.
+func (i *inodeOperations) CreateLink(ctx context.Context, dir *fs.Inode, oldname string, newname string) error {
+	return createLink(i.fileState.FD(), oldname, newname)
+}
+
+// CreateHardLink implements fs.InodeOperations.CreateHardLink.
+func (*inodeOperations) CreateHardLink(context.Context, *fs.Inode, *fs.Inode, string) error {
+	return syserror.EPERM
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo.
+func (*inodeOperations) CreateFifo(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return syserror.EOPNOTSUPP
+}
+
+// Remove implements fs.InodeOperations.Remove.
+func (i *inodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string) error {
+	return unlinkAt(i.fileState.FD(), name, false /* dir */)
+}
+
+// RemoveDirectory implements fs.InodeOperations.RemoveDirectory.
+func (i *inodeOperations) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error {
+	return unlinkAt(i.fileState.FD(), name, true /* dir */)
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (i *inodeOperations) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	op, ok := oldParent.InodeOperations.(*inodeOperations)
+	if !ok {
+		return syscall.EXDEV
+	}
+	np, ok := newParent.InodeOperations.(*inodeOperations)
+	if !ok {
+		return syscall.EXDEV
+	}
+	return syscall.Renameat(op.fileState.FD(), oldName, np.fileState.FD(), newName)
+}
+
+// Bind implements fs.InodeOperations.Bind.
+func (i *inodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, data unix.BoundEndpoint, perm fs.FilePermissions) error {
+	return syserror.EOPNOTSUPP
+}
+
+// BoundEndpoint implements fs.InodeOperations.BoundEndpoint.
+func (i *inodeOperations) BoundEndpoint(inode *fs.Inode, path string) unix.BoundEndpoint {
+	return nil
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (i *inodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return newFile(ctx, d, flags, i, false), nil
+}
+
+// canMap returns true if this fs.Inode can be memory mapped.
+func canMap(inode *fs.Inode) bool {
+	// FIXME: Some obscure character devices can be mapped.
+	return fs.IsFile(inode.StableAttr)
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (i *inodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	// When the kernel supports mapping host FDs, we do so to take
+	// advantage of the host page cache. We forego updating fs.Inodes
+	// because the host manages consistency of its own inode structures.
+	//
+	// For fs.Inodes that can never be mapped we take advantage of
+	// synchronizing metadata updates through host caches.
+	//
+	// So can we use host kernel metadata caches?
+	if !inode.MountSource.Flags.ForcePageCache || !canMap(inode) {
+		// Then just obtain the attributes.
+		return i.fileState.unstableAttr(ctx)
+	}
+	// No, we're maintaining consistency of metadata ourselves.
+	return i.cachingInodeOps.UnstableAttr(ctx, inode)
+}
+
+// Check implements fs.InodeOperations.Check.
+func (i *inodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (i *inodeOperations) SetOwner(context.Context, *fs.Inode, fs.FileOwner) error {
+	return syserror.EPERM
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (i *inodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, f fs.FilePermissions) bool {
+	// Can we use host kernel metadata caches?
+	if !inode.MountSource.Flags.ForcePageCache || !canMap(inode) {
+		// Then just change the timestamps on the fd, the host
+		// will synchronize the metadata update with any host
+		// inode and page cache.
+		return syscall.Fchmod(i.fileState.FD(), uint32(f.LinuxMode())) == nil
+	}
+	// Otherwise update our cached metadata.
+	return i.cachingInodeOps.SetPermissions(ctx, inode, f)
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (i *inodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	// Can we use host kernel metadata caches?
+	if !inode.MountSource.Flags.ForcePageCache || !canMap(inode) {
+		// Then just change the timestamps on the fd, the host
+		// will synchronize the metadata update with any host
+		// inode and page cache.
+		return setTimestamps(i.fileState.FD(), ts)
+	}
+	// Otherwise update our cached metadata.
+	return i.cachingInodeOps.SetTimestamps(ctx, inode, ts)
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (i *inodeOperations) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	// Is the file not memory-mappable?
+	if !canMap(inode) {
+		// Then just change the file size on the fd, the host
+		// will synchronize the metadata update with any host
+		// inode and page cache.
+		return syscall.Ftruncate(i.fileState.FD(), size)
+	}
+	// Otherwise we need to go through cachingInodeOps, even if the host page
+	// cache is in use, to invalidate private copies of truncated pages.
+	return i.cachingInodeOps.Truncate(ctx, inode, size)
+}
+
+// WriteOut implements fs.InodeOperations.WriteOut.
+func (i *inodeOperations) WriteOut(ctx context.Context, inode *fs.Inode) error {
+	// Have we been using host kernel metadata caches?
+	if !inode.MountSource.Flags.ForcePageCache || !canMap(inode) {
+		// Then the metadata is already up to date on the host.
+		return nil
+	}
+	// Otherwise we need to write out cached pages and attributes
+	// that are dirty.
+	return i.cachingInodeOps.WriteOut(ctx, inode)
+}
+
+// Readlink implements fs.InodeOperations.Readlink.
+func (i *inodeOperations) Readlink(ctx context.Context, inode *fs.Inode) (string, error) {
+	return readLink(i.fileState.FD())
+}
+
+// Getlink implements fs.InodeOperations.Getlink.
+func (i *inodeOperations) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	if !fs.IsSymlink(i.fileState.sattr) {
+		return nil, syserror.ENOLINK
+	}
+	return nil, fs.ErrResolveViaReadlink
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+func (i *inodeOperations) StatFS(context.Context) (fs.Info, error) {
+	return fs.Info{}, syserror.ENOSYS
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+// FIXME: Remove this from InodeOperations altogether.
+func (i *inodeOperations) AddLink() {}
+
+// DropLink implements fs.InodeOperations.DropLink.
+// FIXME: Remove this from InodeOperations altogether.
+func (i *inodeOperations) DropLink() {}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+// FIXME: Remove this from InodeOperations altogether.
+func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {}
+
+// readdirAll returns all of the directory entries in i.
+func (i *inodeOperations) readdirAll(d *dirInfo) (map[string]fs.DentAttr, error) {
+	i.readdirMu.Lock()
+	defer i.readdirMu.Unlock()
+
+	fd := i.fileState.FD()
+
+	// syscall.ReadDirent will use getdents, which will seek the file past
+	// the last directory entry. To read the directory entries a second
+	// time, we need to seek back to the beginning.
+	if _, err := syscall.Seek(fd, 0, 0); err != nil {
+		if err == syscall.ESPIPE {
+			// All directories should be seekable. If this file
+			// isn't seekable, it is not a directory and we should
+			// return that more sane error.
+			err = syscall.ENOTDIR
+		}
+		return nil, err
+	}
+
+	names := make([]string, 0, 100)
+	for {
+		// Refill the buffer if necessary
+		if d.bufp >= d.nbuf {
+			d.bufp = 0
+			// ReadDirent will just do a sys_getdents64 to the kernel.
+			n, err := syscall.ReadDirent(fd, d.buf)
+			if err != nil {
+				return nil, err
+			}
+			if n == 0 {
+				break // EOF
+			}
+			d.nbuf = n
+		}
+
+		var nb int
+		// Parse the dirent buffer we just get and return the directory names along
+		// with the number of bytes consumed in the buffer.
+		nb, _, names = syscall.ParseDirent(d.buf[d.bufp:d.nbuf], -1, names)
+		d.bufp += nb
+	}
+
+	entries := make(map[string]fs.DentAttr)
+	for _, filename := range names {
+		// Lookup the type and host device and inode.
+		stat, lerr := fstatat(fd, filename, linux.AT_SYMLINK_NOFOLLOW)
+		if lerr == syscall.ENOENT {
+			// File disappeared between readdir and lstat.
+			// Just treat it as if it didn't exist.
+			continue
+		}
+
+		// There was a serious problem, we should probably report it.
+		if lerr != nil {
+			return nil, lerr
+		}
+
+		entries[filename] = fs.DentAttr{
+			Type: nodeType(&stat),
+			InodeID: hostFileDevice.Map(device.MultiDeviceKey{
+				Device: stat.Dev,
+				Inode:  stat.Ino,
+			}),
+		}
+	}
+	return entries, nil
+}
diff --git a/pkg/sentry/fs/host/inode_state.go b/pkg/sentry/fs/host/inode_state.go
new file mode 100644
index 000000000..80066512a
--- /dev/null
+++ b/pkg/sentry/fs/host/inode_state.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// beforeSave is invoked by stateify.
+func (i *inodeFileState) beforeSave() {
+	if !i.queue.IsEmpty() {
+		panic("event queue must be empty")
+	}
+	if !i.descriptor.donated && i.sattr.Type == fs.RegularFile {
+		uattr, err := i.unstableAttr(context.Background())
+		if err != nil {
+			panic(fmt.Sprintf("failed to get unstable atttribute of %s: %v", i.mops.inodeMappings[i.sattr.InodeID], err))
+		}
+		i.savedUAttr = &uattr
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (i *inodeFileState) afterLoad() {
+	// Initialize the descriptor value.
+	if err := i.descriptor.initAfterLoad(i.mops, i.sattr.InodeID, &i.queue); err != nil {
+		panic(fmt.Sprintf("failed to load value of descriptor: %v", err))
+	}
+
+	// Remap the inode number.
+	var s syscall.Stat_t
+	if err := syscall.Fstat(i.FD(), &s); err != nil {
+		panic(fmt.Sprintf("failed to get metadata for fd %d: %v", i.FD(), err))
+	}
+	key := device.MultiDeviceKey{
+		Device: s.Dev,
+		Inode:  s.Ino,
+	}
+	if !hostFileDevice.Load(key, i.sattr.InodeID) {
+		// This means there was a conflict at s.Dev and s.Ino with
+		// another inode mapping: two files that were unique on the
+		// saved filesystem are no longer unique on this filesystem.
+		// Since this violates the contract that filesystems cannot
+		// change across save and restore, error out.
+		panic(fmt.Sprintf("host %s conflict in host device mappings: %s", key, hostFileDevice))
+	}
+
+	if !i.descriptor.donated && i.sattr.Type == fs.RegularFile {
+		env, ok := fs.CurrentRestoreEnvironment()
+		if !ok {
+			panic("missing restore environment")
+		}
+		uattr := unstableAttr(i.mops, &s)
+		if env.ValidateFileSize && uattr.Size != i.savedUAttr.Size {
+			panic(fmt.Errorf("file size has changed for %s: previously %d, now %d", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.Size, uattr.Size))
+		}
+		if env.ValidateFileTimestamp && uattr.ModificationTime != i.savedUAttr.ModificationTime {
+			panic(fmt.Errorf("file modification time has changed for %s: previously %v, now %v", i.mops.inodeMappings[i.sattr.InodeID], i.savedUAttr.ModificationTime, uattr.ModificationTime))
+		}
+		i.savedUAttr = nil
+	}
+}
diff --git a/pkg/sentry/fs/host/inode_test.go b/pkg/sentry/fs/host/inode_test.go
new file mode 100644
index 000000000..0ff87c418
--- /dev/null
+++ b/pkg/sentry/fs/host/inode_test.go
@@ -0,0 +1,112 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"io/ioutil"
+	"os"
+	"path"
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// TestMultipleReaddir verifies that multiple Readdir calls return the same
+// thing if they use different dir contexts.
+func TestMultipleReaddir(t *testing.T) {
+	p, err := ioutil.TempDir("", "readdir")
+	if err != nil {
+		t.Fatalf("Failed to create test dir: %v", err)
+	}
+	defer os.RemoveAll(p)
+
+	f, err := os.Create(path.Join(p, "a.txt"))
+	if err != nil {
+		t.Fatalf("Failed to create a.txt: %v", err)
+	}
+	f.Close()
+
+	f, err = os.Create(path.Join(p, "b.txt"))
+	if err != nil {
+		t.Fatalf("Failed to create b.txt: %v", err)
+	}
+	f.Close()
+
+	fd, err := open(nil, p)
+	if err != nil {
+		t.Fatalf("Failed to open %q: %v", p, err)
+	}
+	ctx := contexttest.Context(t)
+	n, err := newInode(ctx, newMountSource(ctx, p, fs.RootOwner, &Filesystem{}, fs.MountSourceFlags{}, false), fd, false, false)
+	if err != nil {
+		t.Fatalf("Failed to create inode: %v", err)
+	}
+
+	dirent := fs.NewDirent(n, "readdir")
+	openFile, err := n.GetFile(ctx, dirent, fs.FileFlags{Read: true})
+	if err != nil {
+		t.Fatalf("Failed to get file: %v", err)
+	}
+	defer openFile.DecRef()
+
+	c1 := &fs.DirCtx{DirCursor: new(string)}
+	if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, c1, 0); err != nil {
+		t.Fatalf("First Readdir failed: %v", err)
+	}
+
+	c2 := &fs.DirCtx{DirCursor: new(string)}
+	if _, err := openFile.FileOperations.(*fileOperations).IterateDir(ctx, c2, 0); err != nil {
+		t.Errorf("Second Readdir failed: %v", err)
+	}
+
+	if _, ok := c1.DentAttrs()["a.txt"]; !ok {
+		t.Errorf("want a.txt in first Readdir, got %v", c1.DentAttrs())
+	}
+	if _, ok := c1.DentAttrs()["b.txt"]; !ok {
+		t.Errorf("want b.txt in first Readdir, got %v", c1.DentAttrs())
+	}
+
+	if _, ok := c2.DentAttrs()["a.txt"]; !ok {
+		t.Errorf("want a.txt in second Readdir, got %v", c2.DentAttrs())
+	}
+	if _, ok := c2.DentAttrs()["b.txt"]; !ok {
+		t.Errorf("want b.txt in second Readdir, got %v", c2.DentAttrs())
+	}
+}
+
+// TestCloseFD verifies fds will be closed.
+func TestCloseFD(t *testing.T) {
+	var p [2]int
+	if err := syscall.Pipe(p[0:]); err != nil {
+		t.Fatalf("Failed to create pipe %v", err)
+	}
+	defer syscall.Close(p[0])
+	defer syscall.Close(p[1])
+
+	// Use the write-end because we will detect if it's closed on the read end.
+	ctx := contexttest.Context(t)
+	file, err := NewFile(ctx, p[1], fs.RootOwner)
+	if err != nil {
+		t.Fatalf("Failed to create File: %v", err)
+	}
+	file.DecRef()
+
+	s := make([]byte, 10)
+	if c, err := syscall.Read(p[0], s); c != 0 || err != nil {
+		t.Errorf("want 0, nil (EOF) from read end, got %v, %v", c, err)
+	}
+}
diff --git a/pkg/sentry/fs/host/ioctl_unsafe.go b/pkg/sentry/fs/host/ioctl_unsafe.go
new file mode 100644
index 000000000..3c07c3850
--- /dev/null
+++ b/pkg/sentry/fs/host/ioctl_unsafe.go
@@ -0,0 +1,39 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+func ioctlGetTermios(fd int) (*linux.Termios, error) {
+	var t linux.Termios
+	_, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), syscall.TCGETS, uintptr(unsafe.Pointer(&t)))
+	if errno != 0 {
+		return nil, errno
+	}
+	return &t, nil
+}
+
+func ioctlSetTermios(fd int, req uint64, t *linux.Termios) error {
+	_, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(fd), uintptr(req), uintptr(unsafe.Pointer(t)))
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
diff --git a/pkg/sentry/fs/host/socket.go b/pkg/sentry/fs/host/socket.go
new file mode 100644
index 000000000..8e36ed7ee
--- /dev/null
+++ b/pkg/sentry/fs/host/socket.go
@@ -0,0 +1,471 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/control"
+	unixsocket "gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/link/rawfile"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+// endpoint encapsulates the state needed to represent a host Unix socket.
+type endpoint struct {
+	queue waiter.Queue `state:"nosave"`
+
+	// stype is the type of Unix socket. (Ex: unix.SockStream,
+	// unix.SockSeqpacket, unix.SockDgram)
+	stype unix.SockType `state:"nosave"`
+
+	// fd is the host fd backing this file.
+	fd int `state:"nosave"`
+
+	// If srfd >= 0, it is the host fd that fd was imported from.
+	srfd int `state:"wait"`
+}
+
+func (e *endpoint) init() error {
+	family, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_DOMAIN)
+	if err != nil {
+		return err
+	}
+
+	if family != syscall.AF_UNIX {
+		// We only allow Unix sockets.
+		return syserror.EINVAL
+	}
+
+	stype, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_TYPE)
+	if err != nil {
+		return err
+	}
+
+	if err := syscall.SetNonblock(e.fd, true); err != nil {
+		return err
+	}
+
+	e.stype = unix.SockType(stype)
+	if err := fdnotifier.AddFD(int32(e.fd), &e.queue); err != nil {
+		return err
+	}
+	return nil
+}
+
+// newEndpoint creates a new host endpoint.
+func newEndpoint(fd int, srfd int) (*endpoint, error) {
+	ep := &endpoint{fd: fd, srfd: srfd}
+	if err := ep.init(); err != nil {
+		return nil, err
+	}
+	return ep, nil
+}
+
+// newSocket allocates a new unix socket with host endpoint.
+func newSocket(ctx context.Context, fd int, saveable bool) (*fs.File, error) {
+	ownedfd := fd
+	srfd := -1
+	if saveable {
+		var err error
+		ownedfd, err = syscall.Dup(fd)
+		if err != nil {
+			return nil, err
+		}
+		srfd = fd
+	}
+	ep, err := newEndpoint(ownedfd, srfd)
+	if err != nil {
+		if saveable {
+			syscall.Close(ownedfd)
+		}
+		return nil, err
+	}
+	return unixsocket.New(ctx, ep), nil
+}
+
+// NewSocketWithDirent allocates a new unix socket with host endpoint.
+//
+// This is currently only used by unsaveable Gofer nodes.
+//
+// NewSocketWithDirent takes ownership of f on success.
+func NewSocketWithDirent(ctx context.Context, d *fs.Dirent, f *fd.FD, flags fs.FileFlags) (*fs.File, error) {
+	ep, err := newEndpoint(f.FD(), -1)
+	if err != nil {
+		return nil, err
+	}
+
+	// Take ownship of the FD.
+	f.Release()
+
+	return unixsocket.NewWithDirent(ctx, d, ep, flags), nil
+}
+
+// Close implements unix.Endpoint.Close.
+func (e *endpoint) Close() {
+	fdnotifier.RemoveFD(int32(e.fd))
+	syscall.Close(e.fd)
+	e.fd = -1
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (e *endpoint) EventRegister(we *waiter.Entry, mask waiter.EventMask) {
+	e.queue.EventRegister(we, mask)
+	fdnotifier.UpdateFD(int32(e.fd))
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (e *endpoint) EventUnregister(we *waiter.Entry) {
+	e.queue.EventUnregister(we)
+	fdnotifier.UpdateFD(int32(e.fd))
+}
+
+// Readiness implements unix.Endpoint.Readiness.
+func (e *endpoint) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return fdnotifier.NonBlockingPoll(int32(e.fd), mask)
+}
+
+// Type implements unix.Endpoint.Type.
+func (e *endpoint) Type() unix.SockType {
+	return e.stype
+}
+
+// Connect implements unix.Endpoint.Connect.
+func (e *endpoint) Connect(server unix.BoundEndpoint) *tcpip.Error {
+	return tcpip.ErrInvalidEndpointState
+}
+
+// Bind implements unix.Endpoint.Bind.
+func (e *endpoint) Bind(address tcpip.FullAddress, commit func() *tcpip.Error) *tcpip.Error {
+	return tcpip.ErrInvalidEndpointState
+}
+
+// Listen implements unix.Endpoint.Listen.
+func (e *endpoint) Listen(backlog int) *tcpip.Error {
+	return tcpip.ErrInvalidEndpointState
+}
+
+// Accept implements unix.Endpoint.Accept.
+func (e *endpoint) Accept() (unix.Endpoint, *tcpip.Error) {
+	return nil, tcpip.ErrInvalidEndpointState
+}
+
+// Shutdown implements unix.Endpoint.Shutdown.
+func (e *endpoint) Shutdown(flags tcpip.ShutdownFlags) *tcpip.Error {
+	return tcpip.ErrInvalidEndpointState
+}
+
+// GetSockOpt implements unix.Endpoint.GetSockOpt.
+func (e *endpoint) GetSockOpt(opt interface{}) *tcpip.Error {
+	switch o := opt.(type) {
+	case tcpip.ErrorOption:
+		_, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_ERROR)
+		return translateError(err)
+	case *tcpip.PasscredOption:
+		// We don't support passcred on host sockets.
+		*o = 0
+		return nil
+	case *tcpip.SendBufferSizeOption:
+		v, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_SNDBUF)
+		*o = tcpip.SendBufferSizeOption(v)
+		return translateError(err)
+	case *tcpip.ReceiveBufferSizeOption:
+		v, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_RCVBUF)
+		*o = tcpip.ReceiveBufferSizeOption(v)
+		return translateError(err)
+	case *tcpip.ReuseAddressOption:
+		v, err := syscall.GetsockoptInt(e.fd, syscall.SOL_SOCKET, syscall.SO_REUSEADDR)
+		*o = tcpip.ReuseAddressOption(v)
+		return translateError(err)
+	case *tcpip.ReceiveQueueSizeOption:
+		return tcpip.ErrQueueSizeNotSupported
+	}
+	return tcpip.ErrInvalidEndpointState
+}
+
+// SetSockOpt implements unix.Endpoint.SetSockOpt.
+func (e *endpoint) SetSockOpt(opt interface{}) *tcpip.Error {
+	return nil
+}
+
+// GetLocalAddress implements unix.Endpoint.GetLocalAddress.
+func (e *endpoint) GetLocalAddress() (tcpip.FullAddress, *tcpip.Error) {
+	return tcpip.FullAddress{}, nil
+}
+
+// GetRemoteAddress implements unix.Endpoint.GetRemoteAddress.
+func (e *endpoint) GetRemoteAddress() (tcpip.FullAddress, *tcpip.Error) {
+	return tcpip.FullAddress{}, nil
+}
+
+// Passcred returns whether or not the SO_PASSCRED socket option is
+// enabled on this end.
+func (e *endpoint) Passcred() bool {
+	// We don't support credential passing for host sockets.
+	return false
+}
+
+// ConnectedPasscred returns whether or not the SO_PASSCRED socket option
+// is enabled on the connected end.
+func (e *endpoint) ConnectedPasscred() bool {
+	// We don't support credential passing for host sockets.
+	return false
+}
+
+// SendMsg implements unix.Endpoint.SendMsg.
+func (e *endpoint) SendMsg(data [][]byte, controlMessages unix.ControlMessages, to unix.BoundEndpoint) (uintptr, *tcpip.Error) {
+	if to != nil {
+		return 0, tcpip.ErrInvalidEndpointState
+	}
+	return sendMsg(e.fd, data, controlMessages)
+}
+
+func sendMsg(fd int, data [][]byte, controlMessages unix.ControlMessages) (uintptr, *tcpip.Error) {
+	if !controlMessages.Empty() {
+		return 0, tcpip.ErrInvalidEndpointState
+	}
+	n, err := fdWriteVec(fd, data)
+	return n, translateError(err)
+}
+
+// RecvMsg implements unix.Endpoint.RecvMsg.
+func (e *endpoint) RecvMsg(data [][]byte, creds bool, numRights uintptr, peek bool, addr *tcpip.FullAddress) (uintptr, uintptr, unix.ControlMessages, *tcpip.Error) {
+	return recvMsg(e.fd, data, numRights, peek, addr)
+}
+
+func recvMsg(fd int, data [][]byte, numRights uintptr, peek bool, addr *tcpip.FullAddress) (uintptr, uintptr, unix.ControlMessages, *tcpip.Error) {
+	var cm unet.ControlMessage
+	if numRights > 0 {
+		cm.EnableFDs(int(numRights))
+	}
+	rl, ml, cl, err := fdReadVec(fd, data, []byte(cm), peek)
+	if err == syscall.EAGAIN {
+		return 0, 0, unix.ControlMessages{}, tcpip.ErrWouldBlock
+	}
+	if err != nil {
+		return 0, 0, unix.ControlMessages{}, translateError(err)
+	}
+
+	// Trim the control data if we received less than the full amount.
+	if cl < uint64(len(cm)) {
+		cm = cm[:cl]
+	}
+
+	// Avoid extra allocations in the case where there isn't any control data.
+	if len(cm) == 0 {
+		return rl, ml, unix.ControlMessages{}, nil
+	}
+
+	fds, err := cm.ExtractFDs()
+	if err != nil {
+		return 0, 0, unix.ControlMessages{}, translateError(err)
+	}
+
+	if len(fds) == 0 {
+		return rl, ml, unix.ControlMessages{}, nil
+	}
+	return rl, ml, control.New(nil, nil, newSCMRights(fds)), nil
+}
+
+// NewConnectedEndpoint creates a new unix.Receiver and unix.ConnectedEndpoint
+// backed by a host FD that will pretend to be bound at a given sentry path.
+func NewConnectedEndpoint(file *fd.FD, queue *waiter.Queue, path string) (unix.Receiver, unix.ConnectedEndpoint, *tcpip.Error) {
+	if err := fdnotifier.AddFD(int32(file.FD()), queue); err != nil {
+		return nil, nil, translateError(err)
+	}
+
+	e := &connectedEndpoint{path: path, queue: queue, file: file}
+
+	// AtomicRefCounters start off with a single reference. We need two.
+	e.ref.IncRef()
+
+	return e, e, nil
+}
+
+// connectedEndpoint is a host FD backed implementation of
+// unix.ConnectedEndpoint and unix.Receiver.
+//
+// connectedEndpoint does not support save/restore for now.
+type connectedEndpoint struct {
+	queue *waiter.Queue
+	path  string
+
+	// ref keeps track of references to a connectedEndpoint.
+	ref refs.AtomicRefCount
+
+	// mu protects fd, readClosed and writeClosed.
+	mu sync.RWMutex
+
+	// file is an *fd.FD containing the FD backing this endpoint. It must be
+	// set to nil if it has been closed.
+	file *fd.FD
+
+	// readClosed is true if the FD has read shutdown or if it has been closed.
+	readClosed bool
+
+	// writeClosed is true if the FD has write shutdown or if it has been
+	// closed.
+	writeClosed bool
+}
+
+// Send implements unix.ConnectedEndpoint.Send.
+func (c *connectedEndpoint) Send(data [][]byte, controlMessages unix.ControlMessages, from tcpip.FullAddress) (uintptr, bool, *tcpip.Error) {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	if c.writeClosed {
+		return 0, false, tcpip.ErrClosedForSend
+	}
+	n, err := sendMsg(c.file.FD(), data, controlMessages)
+	// There is no need for the callee to call SendNotify because sendMsg uses
+	// the host's sendmsg(2) and the host kernel's queue.
+	return n, false, err
+}
+
+// SendNotify implements unix.ConnectedEndpoint.SendNotify.
+func (c *connectedEndpoint) SendNotify() {}
+
+// CloseSend implements unix.ConnectedEndpoint.CloseSend.
+func (c *connectedEndpoint) CloseSend() {
+	c.mu.Lock()
+	c.writeClosed = true
+	c.mu.Unlock()
+}
+
+// CloseNotify implements unix.ConnectedEndpoint.CloseNotify.
+func (c *connectedEndpoint) CloseNotify() {}
+
+// Writable implements unix.ConnectedEndpoint.Writable.
+func (c *connectedEndpoint) Writable() bool {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	if c.writeClosed {
+		return true
+	}
+	return fdnotifier.NonBlockingPoll(int32(c.file.FD()), waiter.EventOut)&waiter.EventOut != 0
+}
+
+// Passcred implements unix.ConnectedEndpoint.Passcred.
+func (c *connectedEndpoint) Passcred() bool {
+	// We don't support credential passing for host sockets.
+	return false
+}
+
+// GetLocalAddress implements unix.ConnectedEndpoint.GetLocalAddress.
+func (c *connectedEndpoint) GetLocalAddress() (tcpip.FullAddress, *tcpip.Error) {
+	return tcpip.FullAddress{Addr: tcpip.Address(c.path)}, nil
+}
+
+// EventUpdate implements unix.ConnectedEndpoint.EventUpdate.
+func (c *connectedEndpoint) EventUpdate() {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	if c.file.FD() != -1 {
+		fdnotifier.UpdateFD(int32(c.file.FD()))
+	}
+}
+
+// Recv implements unix.Receiver.Recv.
+func (c *connectedEndpoint) Recv(data [][]byte, creds bool, numRights uintptr, peek bool) (uintptr, uintptr, unix.ControlMessages, tcpip.FullAddress, bool, *tcpip.Error) {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	if c.readClosed {
+		return 0, 0, unix.ControlMessages{}, tcpip.FullAddress{}, false, tcpip.ErrClosedForReceive
+	}
+	rl, ml, cm, err := recvMsg(c.file.FD(), data, numRights, peek, nil)
+	// There is no need for the callee to call RecvNotify because recvMsg uses
+	// the host's recvmsg(2) and the host kernel's queue.
+	return rl, ml, cm, tcpip.FullAddress{Addr: tcpip.Address(c.path)}, false, err
+}
+
+// close releases all resources related to the endpoint.
+func (c *connectedEndpoint) close() {
+	fdnotifier.RemoveFD(int32(c.file.FD()))
+	c.file.Close()
+	c.file = nil
+}
+
+// RecvNotify implements unix.Receiver.RecvNotify.
+func (c *connectedEndpoint) RecvNotify() {}
+
+// CloseRecv implements unix.Receiver.CloseRecv.
+func (c *connectedEndpoint) CloseRecv() {
+	c.mu.Lock()
+	c.readClosed = true
+	c.mu.Unlock()
+}
+
+// Readable implements unix.Receiver.Readable.
+func (c *connectedEndpoint) Readable() bool {
+	c.mu.RLock()
+	defer c.mu.RUnlock()
+	if c.readClosed {
+		return true
+	}
+	return fdnotifier.NonBlockingPoll(int32(c.file.FD()), waiter.EventIn)&waiter.EventIn != 0
+}
+
+// SendQueuedSize implements unix.Receiver.SendQueuedSize.
+func (c *connectedEndpoint) SendQueuedSize() int64 {
+	// SendQueuedSize isn't supported for host sockets because we don't allow the
+	// sentry to call ioctl(2).
+	return -1
+}
+
+// RecvQueuedSize implements unix.Receiver.RecvQueuedSize.
+func (c *connectedEndpoint) RecvQueuedSize() int64 {
+	// RecvQueuedSize isn't supported for host sockets because we don't allow the
+	// sentry to call ioctl(2).
+	return -1
+}
+
+// SendMaxQueueSize implements unix.Receiver.SendMaxQueueSize.
+func (c *connectedEndpoint) SendMaxQueueSize() int64 {
+	v, err := syscall.GetsockoptInt(c.file.FD(), syscall.SOL_SOCKET, syscall.SO_SNDBUF)
+	if err != nil {
+		return -1
+	}
+	return int64(v)
+}
+
+// RecvMaxQueueSize implements unix.Receiver.RecvMaxQueueSize.
+func (c *connectedEndpoint) RecvMaxQueueSize() int64 {
+	v, err := syscall.GetsockoptInt(c.file.FD(), syscall.SOL_SOCKET, syscall.SO_RCVBUF)
+	if err != nil {
+		return -1
+	}
+	return int64(v)
+}
+
+// Release implements unix.ConnectedEndpoint.Release and unix.Receiver.Release.
+func (c *connectedEndpoint) Release() {
+	c.ref.DecRefWithDestructor(c.close)
+}
+
+func translateError(err error) *tcpip.Error {
+	if err == nil {
+		return nil
+	}
+	return rawfile.TranslateErrno(err.(syscall.Errno))
+}
diff --git a/pkg/sentry/fs/host/socket_state.go b/pkg/sentry/fs/host/socket_state.go
new file mode 100644
index 000000000..6acabd55a
--- /dev/null
+++ b/pkg/sentry/fs/host/socket_state.go
@@ -0,0 +1,39 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"fmt"
+	"syscall"
+)
+
+// beforeSave is invoked by stateify.
+func (ep *endpoint) beforeSave() {
+	if ep.srfd < 0 {
+		panic("only host file descriptors provided at sentry startup can be saved")
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (ep *endpoint) afterLoad() {
+	fd, err := syscall.Dup(ep.srfd)
+	if err != nil {
+		panic(fmt.Sprintf("failed to dup restored fd %d: %v", ep.srfd, err))
+	}
+	ep.fd = fd
+	if err := ep.init(); err != nil {
+		panic(fmt.Sprintf("Could not restore host socket fd %d: %v", ep.srfd, err))
+	}
+}
diff --git a/pkg/sentry/fs/host/socket_test.go b/pkg/sentry/fs/host/socket_test.go
new file mode 100644
index 000000000..80c46dcfa
--- /dev/null
+++ b/pkg/sentry/fs/host/socket_test.go
@@ -0,0 +1,401 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"reflect"
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/fd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+var (
+	// Make sure that connectedEndpoint implements unix.ConnectedEndpoint.
+	_ = unix.ConnectedEndpoint(new(connectedEndpoint))
+
+	// Make sure that connectedEndpoint implements unix.Receiver.
+	_ = unix.Receiver(new(connectedEndpoint))
+)
+
+func getFl(fd int) (uint32, error) {
+	fl, _, err := syscall.RawSyscall(syscall.SYS_FCNTL, uintptr(fd), syscall.F_GETFL, 0)
+	if err == 0 {
+		return uint32(fl), nil
+	}
+	return 0, err
+}
+
+func TestSocketIsBlocking(t *testing.T) {
+	// Using socketpair here because it's already connected.
+	pair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("host socket creation failed: %v", err)
+	}
+
+	fl, err := getFl(pair[0])
+	if err != nil {
+		t.Fatalf("getFl: fcntl(%v, GETFL) => %v", pair[0], err)
+	}
+	if fl&syscall.O_NONBLOCK == syscall.O_NONBLOCK {
+		t.Fatalf("Expected socket %v to be blocking", pair[0])
+	}
+	if fl, err = getFl(pair[1]); err != nil {
+		t.Fatalf("getFl: fcntl(%v, GETFL) => %v", pair[1], err)
+	}
+	if fl&syscall.O_NONBLOCK == syscall.O_NONBLOCK {
+		t.Fatalf("Expected socket %v to be blocking", pair[1])
+	}
+	sock, err := newSocket(contexttest.Context(t), pair[0], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) failed => %v", pair[0], err)
+	}
+	defer sock.DecRef()
+	// Test that the socket now is non blocking.
+	if fl, err = getFl(pair[0]); err != nil {
+		t.Fatalf("getFl: fcntl(%v, GETFL) => %v", pair[0], err)
+	}
+	if fl&syscall.O_NONBLOCK != syscall.O_NONBLOCK {
+		t.Errorf("Expected socket %v to have becoming non blocking", pair[0])
+	}
+	if fl, err = getFl(pair[1]); err != nil {
+		t.Fatalf("getFl: fcntl(%v, GETFL) => %v", pair[1], err)
+	}
+	if fl&syscall.O_NONBLOCK == syscall.O_NONBLOCK {
+		t.Errorf("Did not expect socket %v to become non blocking", pair[1])
+	}
+}
+
+func TestSocketWritev(t *testing.T) {
+	// Using socketpair here because it's already connected.
+	pair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("host socket creation failed: %v", err)
+	}
+	socket, err := newSocket(contexttest.Context(t), pair[0], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", pair[0], err)
+	}
+	defer socket.DecRef()
+	buf := []byte("hello world\n")
+	n, err := socket.Writev(contexttest.Context(t), usermem.BytesIOSequence(buf))
+	if err != nil {
+		t.Fatalf("socket writev failed: %v", err)
+	}
+
+	if n != int64(len(buf)) {
+		t.Fatalf("socket writev wrote incorrect bytes: %d", n)
+	}
+}
+
+func TestSocketWritevLen0(t *testing.T) {
+	// Using socketpair here because it's already connected.
+	pair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("host socket creation failed: %v", err)
+	}
+	socket, err := newSocket(contexttest.Context(t), pair[0], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", pair[0], err)
+	}
+	defer socket.DecRef()
+	n, err := socket.Writev(contexttest.Context(t), usermem.BytesIOSequence(nil))
+	if err != nil {
+		t.Fatalf("socket writev failed: %v", err)
+	}
+
+	if n != 0 {
+		t.Fatalf("socket writev wrote incorrect bytes: %d", n)
+	}
+}
+
+func TestSocketSendMsgLen0(t *testing.T) {
+	// Using socketpair here because it's already connected.
+	pair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("host socket creation failed: %v", err)
+	}
+	sfile, err := newSocket(contexttest.Context(t), pair[0], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", pair[0], err)
+	}
+	defer sfile.DecRef()
+
+	s := sfile.FileOperations.(socket.Socket)
+	n, terr := s.SendMsg(nil, usermem.BytesIOSequence(nil), []byte{}, 0, unix.ControlMessages{})
+	if n != 0 {
+		t.Fatalf("socket sendmsg() failed: %v wrote: %d", terr, n)
+	}
+
+	if terr != nil {
+		t.Fatalf("socket sendmsg() failed: %v", terr)
+	}
+}
+
+func TestListen(t *testing.T) {
+	pair, err := syscall.Socketpair(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM, 0) => %v", err)
+	}
+	sfile1, err := newSocket(contexttest.Context(t), pair[0], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", pair[0], err)
+	}
+	defer sfile1.DecRef()
+	socket1 := sfile1.FileOperations.(socket.Socket)
+
+	sfile2, err := newSocket(contexttest.Context(t), pair[1], false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", pair[1], err)
+	}
+	defer sfile2.DecRef()
+	socket2 := sfile2.FileOperations.(socket.Socket)
+
+	// Socketpairs can not be listened to.
+	if err := socket1.Listen(nil, 64); err != syserr.ErrInvalidEndpointState {
+		t.Fatalf("socket1.Listen(nil, 64) => %v, want syserr.ErrInvalidEndpointState", err)
+	}
+	if err := socket2.Listen(nil, 64); err != syserr.ErrInvalidEndpointState {
+		t.Fatalf("socket2.Listen(nil, 64) => %v, want syserr.ErrInvalidEndpointState", err)
+	}
+
+	// Create a Unix socket, do not bind it.
+	sock, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM, 0)
+	if err != nil {
+		t.Fatalf("syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM, 0) => %v", err)
+	}
+	sfile3, err := newSocket(contexttest.Context(t), sock, false)
+	if err != nil {
+		t.Fatalf("newSocket(%v) => %v", sock, err)
+	}
+	defer sfile3.DecRef()
+	socket3 := sfile3.FileOperations.(socket.Socket)
+
+	// This socket is not bound so we can't listen on it.
+	if err := socket3.Listen(nil, 64); err != syserr.ErrInvalidEndpointState {
+		t.Fatalf("socket3.Listen(nil, 64) => %v, want syserr.ErrInvalidEndpointState", err)
+	}
+}
+
+func TestSend(t *testing.T) {
+	e := connectedEndpoint{writeClosed: true}
+	if _, _, err := e.Send(nil, unix.ControlMessages{}, tcpip.FullAddress{}); err != tcpip.ErrClosedForSend {
+		t.Errorf("Got %#v.Send() = %v, want = %v", e, err, tcpip.ErrClosedForSend)
+	}
+}
+
+func TestRecv(t *testing.T) {
+	e := connectedEndpoint{readClosed: true}
+	if _, _, _, _, _, err := e.Recv(nil, false, 0, false); err != tcpip.ErrClosedForReceive {
+		t.Errorf("Got %#v.Recv() = %v, want = %v", e, err, tcpip.ErrClosedForReceive)
+	}
+}
+
+func TestPasscred(t *testing.T) {
+	e := connectedEndpoint{}
+	if got, want := e.Passcred(), false; got != want {
+		t.Errorf("Got %#v.Passcred() = %t, want = %t", e, got, want)
+	}
+}
+
+func TestGetLocalAddress(t *testing.T) {
+	e := connectedEndpoint{path: "foo"}
+	want := tcpip.FullAddress{Addr: tcpip.Address("foo")}
+	if got, err := e.GetLocalAddress(); err != nil || got != want {
+		t.Errorf("Got %#v.GetLocalAddress() = %#v, %v, want = %#v, %v", e, got, err, want, nil)
+	}
+}
+
+func TestQueuedSize(t *testing.T) {
+	e := connectedEndpoint{}
+	tests := []struct {
+		name string
+		f    func() int64
+	}{
+		{"SendQueuedSize", e.SendQueuedSize},
+		{"RecvQueuedSize", e.RecvQueuedSize},
+	}
+
+	for _, test := range tests {
+		if got, want := test.f(), int64(-1); got != want {
+			t.Errorf("Got %#v.%s() = %d, want = %d", e, test.name, got, want)
+		}
+	}
+}
+
+func TestReadable(t *testing.T) {
+	e := connectedEndpoint{readClosed: true}
+	if got, want := e.Readable(), true; got != want {
+		t.Errorf("Got %#v.Readable() = %t, want = %t", e, got, want)
+	}
+}
+
+func TestWritable(t *testing.T) {
+	e := connectedEndpoint{writeClosed: true}
+	if got, want := e.Writable(), true; got != want {
+		t.Errorf("Got %#v.Writable() = %t, want = %t", e, got, want)
+	}
+}
+
+func TestRelease(t *testing.T) {
+	f, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c := &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f)}
+	want := &connectedEndpoint{queue: c.queue}
+	want.ref.DecRef()
+	fdnotifier.AddFD(int32(c.file.FD()), nil)
+	c.Release()
+	if !reflect.DeepEqual(c, want) {
+		t.Errorf("got = %#v, want = %#v", c, want)
+	}
+}
+
+func TestClose(t *testing.T) {
+	type testCase struct {
+		name  string
+		cep   *connectedEndpoint
+		addFD bool
+		f     func()
+		want  *connectedEndpoint
+	}
+
+	var tests []testCase
+
+	// nil is the value used by connectedEndpoint to indicate a closed file.
+	// Non-nil files are used to check if the file gets closed.
+
+	f, err := syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c := &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f)}
+	tests = append(tests, testCase{
+		name:  "First CloseRecv",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseRecv,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), readClosed: true}
+	tests = append(tests, testCase{
+		name:  "Second CloseRecv",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseRecv,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f)}
+	tests = append(tests, testCase{
+		name:  "First CloseSend",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseSend,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, writeClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), writeClosed: true}
+	tests = append(tests, testCase{
+		name:  "Second CloseSend",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseSend,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, writeClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), writeClosed: true}
+	tests = append(tests, testCase{
+		name:  "CloseSend then CloseRecv",
+		cep:   c,
+		addFD: true,
+		f:     c.CloseRecv,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true, writeClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), readClosed: true}
+	tests = append(tests, testCase{
+		name:  "CloseRecv then CloseSend",
+		cep:   c,
+		addFD: true,
+		f:     c.CloseSend,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true, writeClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), readClosed: true, writeClosed: true}
+	tests = append(tests, testCase{
+		name:  "Full close then CloseRecv",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseRecv,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true, writeClosed: true},
+	})
+
+	f, err = syscall.Socket(syscall.AF_UNIX, syscall.SOCK_STREAM|syscall.SOCK_NONBLOCK|syscall.SOCK_CLOEXEC, 0)
+	if err != nil {
+		t.Fatal("Creating socket:", err)
+	}
+	c = &connectedEndpoint{queue: &waiter.Queue{}, file: fd.New(f), readClosed: true, writeClosed: true}
+	tests = append(tests, testCase{
+		name:  "Full close then CloseSend",
+		cep:   c,
+		addFD: false,
+		f:     c.CloseSend,
+		want:  &connectedEndpoint{queue: c.queue, file: c.file, readClosed: true, writeClosed: true},
+	})
+
+	for _, test := range tests {
+		if test.addFD {
+			fdnotifier.AddFD(int32(test.cep.file.FD()), nil)
+		}
+		if test.f(); !reflect.DeepEqual(test.cep, test.want) {
+			t.Errorf("%s: got = %#v, want = %#v", test.name, test.cep, test.want)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/host/socket_unsafe.go b/pkg/sentry/fs/host/socket_unsafe.go
new file mode 100644
index 000000000..bf8da6867
--- /dev/null
+++ b/pkg/sentry/fs/host/socket_unsafe.go
@@ -0,0 +1,82 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+// buildIovec builds an iovec slice from the given []byte slice.
+func buildIovec(bufs [][]byte) (uintptr, []syscall.Iovec) {
+	var length uintptr
+	iovecs := make([]syscall.Iovec, 0, 10)
+	for i := range bufs {
+		if l := len(bufs[i]); l > 0 {
+			length += uintptr(l)
+			iovecs = append(iovecs, syscall.Iovec{
+				Base: &bufs[i][0],
+				Len:  uint64(l),
+			})
+		}
+	}
+	return length, iovecs
+}
+
+func fdReadVec(fd int, bufs [][]byte, control []byte, peek bool) (readLen uintptr, msgLen uintptr, controlLen uint64, err error) {
+	flags := uintptr(syscall.MSG_DONTWAIT | syscall.MSG_TRUNC)
+	if peek {
+		flags |= syscall.MSG_PEEK
+	}
+
+	length, iovecs := buildIovec(bufs)
+
+	var msg syscall.Msghdr
+	if len(control) != 0 {
+		msg.Control = &control[0]
+		msg.Controllen = uint64(len(control))
+	}
+
+	if len(iovecs) != 0 {
+		msg.Iov = &iovecs[0]
+		msg.Iovlen = uint64(len(iovecs))
+	}
+	n, _, e := syscall.RawSyscall(syscall.SYS_RECVMSG, uintptr(fd), uintptr(unsafe.Pointer(&msg)), flags)
+	if e != 0 {
+		return 0, 0, 0, e
+	}
+
+	if n > length {
+		return length, n, msg.Controllen, nil
+	}
+
+	return n, n, msg.Controllen, nil
+}
+
+func fdWriteVec(fd int, bufs [][]byte) (uintptr, error) {
+	_, iovecs := buildIovec(bufs)
+
+	var msg syscall.Msghdr
+	if len(iovecs) > 0 {
+		msg.Iov = &iovecs[0]
+		msg.Iovlen = uint64(len(iovecs))
+	}
+	n, _, e := syscall.RawSyscall(syscall.SYS_SENDMSG, uintptr(fd), uintptr(unsafe.Pointer(&msg)), syscall.MSG_DONTWAIT|syscall.MSG_NOSIGNAL)
+	if e != 0 {
+		return 0, e
+	}
+
+	return n, nil
+}
diff --git a/pkg/sentry/fs/host/util.go b/pkg/sentry/fs/host/util.go
new file mode 100644
index 000000000..74c703eb7
--- /dev/null
+++ b/pkg/sentry/fs/host/util.go
@@ -0,0 +1,197 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"os"
+	"path"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func open(parent *inodeOperations, name string) (int, error) {
+	if parent == nil && !path.IsAbs(name) {
+		return -1, syserror.EINVAL
+	}
+	name = path.Clean(name)
+
+	// Don't follow through symlinks.
+	flags := syscall.O_NOFOLLOW
+
+	if fd, err := openAt(parent, name, flags|syscall.O_RDWR, 0); err == nil {
+		return fd, nil
+	}
+	// Retry as read-only.
+	if fd, err := openAt(parent, name, flags|syscall.O_RDONLY, 0); err == nil {
+		return fd, nil
+	}
+
+	// Retry as write-only.
+	if fd, err := openAt(parent, name, flags|syscall.O_WRONLY, 0); err == nil {
+		return fd, nil
+	}
+
+	// Retry as a symlink, by including O_PATH as an option.
+	fd, err := openAt(parent, name, linux.O_PATH|flags, 0)
+	if err == nil {
+		return fd, nil
+	}
+
+	// Everything failed.
+	return -1, err
+}
+
+func openAt(parent *inodeOperations, name string, flags int, perm linux.FileMode) (int, error) {
+	if parent == nil {
+		return syscall.Open(name, flags, uint32(perm))
+	}
+	return syscall.Openat(parent.fileState.FD(), name, flags, uint32(perm))
+}
+
+func nodeType(s *syscall.Stat_t) fs.InodeType {
+	switch x := (s.Mode & syscall.S_IFMT); x {
+	case syscall.S_IFLNK:
+		return fs.Symlink
+	case syscall.S_IFIFO:
+		return fs.Pipe
+	case syscall.S_IFCHR:
+		return fs.CharacterDevice
+	case syscall.S_IFBLK:
+		return fs.BlockDevice
+	case syscall.S_IFSOCK:
+		return fs.Socket
+	case syscall.S_IFDIR:
+		return fs.Directory
+	case syscall.S_IFREG:
+		return fs.RegularFile
+	default:
+		// This shouldn't happen, but just in case...
+		log.Warningf("unknown host file type %d: assuming regular", x)
+		return fs.RegularFile
+	}
+}
+
+func wouldBlock(s *syscall.Stat_t) bool {
+	typ := nodeType(s)
+	return typ == fs.Pipe || typ == fs.Socket || typ == fs.CharacterDevice
+}
+
+func stableAttr(s *syscall.Stat_t) fs.StableAttr {
+	return fs.StableAttr{
+		Type:     nodeType(s),
+		DeviceID: hostFileDevice.DeviceID(),
+		InodeID: hostFileDevice.Map(device.MultiDeviceKey{
+			Device: s.Dev,
+			Inode:  s.Ino,
+		}),
+		BlockSize: int64(s.Blksize),
+	}
+}
+
+func owner(mo *superOperations, s *syscall.Stat_t) fs.FileOwner {
+	// User requested no translation, just return actual owner.
+	if mo.dontTranslateOwnership {
+		return fs.FileOwner{auth.KUID(s.Uid), auth.KGID(s.Gid)}
+	}
+
+	// Show only IDs relevant to the sandboxed task. I.e. if we not own the
+	// file, no sandboxed task can own the file. In that case, we
+	// use OverflowID for UID, implying that the IDs are not mapped in the
+	// "root" user namespace.
+	//
+	// E.g.
+	// sandbox's host EUID/EGID is 1/1.
+	// some_dir's host UID/GID is 2/1.
+	// Task that mounted this fs has virtualized EUID/EGID 5/5.
+	//
+	// If you executed `ls -n` in the sandboxed task, it would show:
+	// drwxwrxwrx [...] 65534 5 [...] some_dir
+
+	// Files are owned by OverflowID by default.
+	owner := fs.FileOwner{auth.KUID(auth.OverflowUID), auth.KGID(auth.OverflowGID)}
+
+	// If we own file on host, let mounting task's initial EUID own
+	// the file.
+	if s.Uid == hostUID {
+		owner.UID = mo.mounter.UID
+	}
+
+	// If our group matches file's group, make file's group match
+	// the mounting task's initial EGID.
+	for _, gid := range hostGIDs {
+		if s.Gid == gid {
+			owner.GID = mo.mounter.GID
+			break
+		}
+	}
+	return owner
+}
+
+func unstableAttr(mo *superOperations, s *syscall.Stat_t) fs.UnstableAttr {
+	return fs.UnstableAttr{
+		Size:             s.Size,
+		Usage:            s.Blocks * 512,
+		Perms:            fs.FilePermsFromMode(linux.FileMode(s.Mode)),
+		Owner:            owner(mo, s),
+		AccessTime:       ktime.FromUnix(s.Atim.Sec, s.Atim.Nsec),
+		ModificationTime: ktime.FromUnix(s.Mtim.Sec, s.Mtim.Nsec),
+		StatusChangeTime: ktime.FromUnix(s.Ctim.Sec, s.Ctim.Nsec),
+		Links:            s.Nlink,
+	}
+}
+
+type dirInfo struct {
+	buf  []byte // buffer for directory I/O.
+	nbuf int    // length of buf; return value from ReadDirent.
+	bufp int    // location of next record in buf.
+}
+
+// isBlockError unwraps os errors and checks if they are caused by EAGAIN or
+// EWOULDBLOCK. This is so they can be transformed into syserror.ErrWouldBlock.
+func isBlockError(err error) bool {
+	if err == syserror.EAGAIN || err == syserror.EWOULDBLOCK {
+		return true
+	}
+	if pe, ok := err.(*os.PathError); ok {
+		return isBlockError(pe.Err)
+	}
+	return false
+}
+
+func hostEffectiveKIDs() (uint32, []uint32, error) {
+	gids, err := os.Getgroups()
+	if err != nil {
+		return 0, nil, err
+	}
+	egids := make([]uint32, len(gids))
+	for i, gid := range gids {
+		egids[i] = uint32(gid)
+	}
+	return uint32(os.Geteuid()), append(egids, uint32(os.Getegid())), nil
+}
+
+var hostUID uint32
+var hostGIDs []uint32
+
+func init() {
+	hostUID, hostGIDs, _ = hostEffectiveKIDs()
+}
diff --git a/pkg/sentry/fs/host/util_unsafe.go b/pkg/sentry/fs/host/util_unsafe.go
new file mode 100644
index 000000000..c38d2392d
--- /dev/null
+++ b/pkg/sentry/fs/host/util_unsafe.go
@@ -0,0 +1,137 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+)
+
+func createLink(fd int, name string, linkName string) error {
+	namePtr, err := syscall.BytePtrFromString(name)
+	if err != nil {
+		return err
+	}
+	linkNamePtr, err := syscall.BytePtrFromString(linkName)
+	if err != nil {
+		return err
+	}
+	_, _, errno := syscall.Syscall(
+		syscall.SYS_SYMLINKAT,
+		uintptr(unsafe.Pointer(namePtr)),
+		uintptr(fd),
+		uintptr(unsafe.Pointer(linkNamePtr)))
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+func readLink(fd int) (string, error) {
+	// Buffer sizing copied from os.Readlink.
+	for l := 128; ; l *= 2 {
+		b := make([]byte, l)
+		n, _, errno := syscall.Syscall6(
+			syscall.SYS_READLINKAT,
+			uintptr(fd),
+			uintptr(unsafe.Pointer(syscall.StringBytePtr(""))),
+			uintptr(unsafe.Pointer(&b[0])),
+			uintptr(l),
+			0, 0)
+		if n < 0 {
+			n = 0
+		}
+		if errno != 0 {
+			return "", errno
+		}
+		if n < uintptr(l) {
+			return string(b[:n]), nil
+		}
+	}
+}
+
+func unlinkAt(fd int, name string, dir bool) error {
+	namePtr, err := syscall.BytePtrFromString(name)
+	if err != nil {
+		return err
+	}
+	var flags uintptr
+	if dir {
+		flags = linux.AT_REMOVEDIR
+	}
+	_, _, errno := syscall.Syscall(
+		syscall.SYS_UNLINKAT,
+		uintptr(fd),
+		uintptr(unsafe.Pointer(namePtr)),
+		flags,
+	)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+func timespecFromTimestamp(t ktime.Time, omit, setSysTime bool) syscall.Timespec {
+	if omit {
+		return syscall.Timespec{0, linux.UTIME_OMIT}
+	}
+	if setSysTime {
+		return syscall.Timespec{0, linux.UTIME_NOW}
+	}
+	return syscall.NsecToTimespec(t.Nanoseconds())
+}
+
+func setTimestamps(fd int, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+	var sts [2]syscall.Timespec
+	sts[0] = timespecFromTimestamp(ts.ATime, ts.ATimeOmit, ts.ATimeSetSystemTime)
+	sts[1] = timespecFromTimestamp(ts.MTime, ts.MTimeOmit, ts.MTimeSetSystemTime)
+	_, _, errno := syscall.Syscall6(
+		syscall.SYS_UTIMENSAT,
+		uintptr(fd),
+		0, /* path */
+		uintptr(unsafe.Pointer(&sts)),
+		0, /* flags */
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+func fstatat(fd int, name string, flags int) (syscall.Stat_t, error) {
+	var stat syscall.Stat_t
+	namePtr, err := syscall.BytePtrFromString(name)
+	if err != nil {
+		return stat, err
+	}
+	_, _, errno := syscall.Syscall6(
+		syscall.SYS_NEWFSTATAT,
+		uintptr(fd),
+		uintptr(unsafe.Pointer(namePtr)),
+		uintptr(unsafe.Pointer(&stat)),
+		uintptr(flags),
+		0, 0)
+	if errno != 0 {
+		return stat, errno
+	}
+	return stat, nil
+}
diff --git a/pkg/sentry/fs/host/wait_test.go b/pkg/sentry/fs/host/wait_test.go
new file mode 100644
index 000000000..c5f5c9c0d
--- /dev/null
+++ b/pkg/sentry/fs/host/wait_test.go
@@ -0,0 +1,70 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package host
+
+import (
+	"syscall"
+	"testing"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+func TestWait(t *testing.T) {
+	var fds [2]int
+	err := syscall.Pipe(fds[:])
+	if err != nil {
+		t.Fatalf("Unable to create pipe: %v", err)
+	}
+
+	defer syscall.Close(fds[1])
+
+	ctx := contexttest.Context(t)
+	file, err := NewFile(ctx, fds[0], fs.RootOwner)
+	if err != nil {
+		syscall.Close(fds[0])
+		t.Fatalf("NewFile failed: %v", err)
+	}
+
+	defer file.DecRef()
+
+	r := file.Readiness(waiter.EventIn)
+	if r != 0 {
+		t.Fatalf("File is ready for read when it shouldn't be.")
+	}
+
+	e, ch := waiter.NewChannelEntry(nil)
+	file.EventRegister(&e, waiter.EventIn)
+	defer file.EventUnregister(&e)
+
+	// Check that there are no notifications yet.
+	if len(ch) != 0 {
+		t.Fatalf("Channel is non-empty")
+	}
+
+	// Write to the pipe, so it should be writable now.
+	syscall.Write(fds[1], []byte{1})
+
+	// Check that we get a notification. We need to yield the current thread
+	// so that the fdnotifier can deliver notifications, so we use a
+	// 1-second timeout instead of just checking the length of the channel.
+	select {
+	case <-ch:
+	case <-time.After(1 * time.Second):
+		t.Fatalf("Channel not notified")
+	}
+}
diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go
new file mode 100644
index 000000000..b624f4182
--- /dev/null
+++ b/pkg/sentry/fs/inode.go
@@ -0,0 +1,455 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/lock"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// Inode is a file system object that can be simulatenously referenced by different
+// components of the VFS (Dirent, fs.File, etc).
+type Inode struct {
+	// AtomicRefCount is our reference count.
+	refs.AtomicRefCount
+
+	// InodeOperations is the file system specific behavior of the Inode.
+	InodeOperations InodeOperations
+
+	// StableAttr are stable cached attributes of the Inode.
+	StableAttr StableAttr
+
+	// LockCtx is the file lock context. It manages its own sychronization and tracks
+	// regions of the Inode that have locks held.
+	LockCtx LockCtx
+
+	// Watches is the set of inotify watches for this inode.
+	Watches *Watches
+
+	// MountSource is the mount source this Inode is a part of.
+	MountSource *MountSource
+
+	// overlay is the overlay entry for this Inode.
+	overlay *overlayEntry
+}
+
+// LockCtx is an Inode's lock context and contains different personalities of locks; both
+// Posix and BSD style locks are supported.
+//
+// Note that in Linux fcntl(2) and flock(2) locks are _not_ cooperative, because race and
+// deadlock conditions make merging them prohibitive. We do the same and keep them oblivious
+// to each other but provide a "context" as a convenient container.
+type LockCtx struct {
+	// Posix is a set of POSIX-style regional advisory locks, see fcntl(2).
+	Posix lock.Locks
+
+	// BSD is a set of BSD-style advisory file wide locks, see flock(2).
+	BSD lock.Locks
+}
+
+// NewInode constructs an Inode from InodeOperations, a MountSource, and stable attributes.
+//
+// NewInode takes a reference on msrc.
+func NewInode(iops InodeOperations, msrc *MountSource, sattr StableAttr) *Inode {
+	msrc.IncRef()
+	return &Inode{
+		InodeOperations: iops,
+		StableAttr:      sattr,
+		Watches:         newWatches(),
+		MountSource:     msrc,
+	}
+}
+
+// DecRef drops a reference on the Inode.
+func (i *Inode) DecRef() {
+	i.DecRefWithDestructor(i.destroy)
+}
+
+// destroy releases the Inode and releases the msrc reference taken.
+func (i *Inode) destroy() {
+	// FIXME: Context is not plumbed here.
+	ctx := context.Background()
+	if err := i.WriteOut(ctx); err != nil {
+		// FIXME: Mark as warning again once noatime is
+		// properly supported.
+		log.Debugf("Inode %+v, failed to sync all metadata: %v", i.StableAttr, err)
+	}
+
+	// If this inode is being destroyed because it was unlinked, queue a
+	// deletion event. This may not be the case for inodes being revalidated.
+	if i.Watches.unlinked {
+		i.Watches.Notify("", linux.IN_DELETE_SELF, 0)
+	}
+
+	// Remove references from the watch owners to the watches on this inode,
+	// since the watches are about to be GCed. Note that we don't need to worry
+	// about the watch pins since if there were any active pins, this inode
+	// wouldn't be in the destructor.
+	i.Watches.targetDestroyed()
+
+	// Overlay resources should be released synchronously, since they may
+	// trigger more Inode.destroy calls which must themselves be handled
+	// synchronously, like the WriteOut call above.
+	if i.overlay != nil {
+		i.overlay.release()
+		i.MountSource.DecRef()
+		return
+	}
+
+	// Regular (non-overlay) resources may be released asynchronously.
+	Async(func() {
+		i.InodeOperations.Release(ctx)
+		i.MountSource.DecRef()
+	})
+}
+
+// Mappable calls i.InodeOperations.Mappable.
+func (i *Inode) Mappable() memmap.Mappable {
+	if i.overlay != nil {
+		// In an overlay, Mappable is always implemented by
+		// the overlayEntry metadata to synchronize memory
+		// access of files with copy up. But first check if
+		// the Inodes involved would be mappable in the first
+		// place.
+		i.overlay.copyMu.RLock()
+		ok := i.overlay.isMappableLocked()
+		i.overlay.copyMu.RUnlock()
+		if !ok {
+			return nil
+		}
+		return i.overlay
+	}
+	return i.InodeOperations.Mappable(i)
+}
+
+// WriteOut calls i.InodeOperations.WriteOut with i as the Inode.
+func (i *Inode) WriteOut(ctx context.Context) error {
+	if i.overlay != nil {
+		return overlayWriteOut(ctx, i.overlay)
+	}
+	return i.InodeOperations.WriteOut(ctx, i)
+}
+
+// Lookup calls i.InodeOperations.Lookup with i as the directory.
+func (i *Inode) Lookup(ctx context.Context, name string) (*Dirent, error) {
+	if i.overlay != nil {
+		return overlayLookup(ctx, i.overlay, i, name)
+	}
+	return i.InodeOperations.Lookup(ctx, i, name)
+}
+
+// Create calls i.InodeOperations.Create with i as the directory.
+func (i *Inode) Create(ctx context.Context, d *Dirent, name string, flags FileFlags, perm FilePermissions) (*File, error) {
+	if i.overlay != nil {
+		return overlayCreate(ctx, i.overlay, d, name, flags, perm)
+	}
+	return i.InodeOperations.Create(ctx, i, name, flags, perm)
+}
+
+// CreateDirectory calls i.InodeOperations.CreateDirectory with i as the directory.
+func (i *Inode) CreateDirectory(ctx context.Context, d *Dirent, name string, perm FilePermissions) error {
+	if i.overlay != nil {
+		return overlayCreateDirectory(ctx, i.overlay, d, name, perm)
+	}
+	return i.InodeOperations.CreateDirectory(ctx, i, name, perm)
+}
+
+// CreateLink calls i.InodeOperations.CreateLink with i as the directory.
+func (i *Inode) CreateLink(ctx context.Context, d *Dirent, oldname string, newname string) error {
+	if i.overlay != nil {
+		return overlayCreateLink(ctx, i.overlay, d, oldname, newname)
+	}
+	return i.InodeOperations.CreateLink(ctx, i, oldname, newname)
+}
+
+// CreateHardLink calls i.InodeOperations.CreateHardLink with i as the directory.
+func (i *Inode) CreateHardLink(ctx context.Context, d *Dirent, target *Dirent, name string) error {
+	if i.overlay != nil {
+		return overlayCreateHardLink(ctx, i.overlay, d, target, name)
+	}
+	return i.InodeOperations.CreateHardLink(ctx, i, target.Inode, name)
+}
+
+// CreateFifo calls i.InodeOperations.CreateFifo with i as the directory.
+func (i *Inode) CreateFifo(ctx context.Context, d *Dirent, name string, perm FilePermissions) error {
+	if i.overlay != nil {
+		return overlayCreateFifo(ctx, i.overlay, d, name, perm)
+	}
+	return i.InodeOperations.CreateFifo(ctx, i, name, perm)
+}
+
+// Remove calls i.InodeOperations.Remove/RemoveDirectory with i as the directory.
+func (i *Inode) Remove(ctx context.Context, d *Dirent, remove *Dirent) error {
+	if i.overlay != nil {
+		return overlayRemove(ctx, i.overlay, d, remove)
+	}
+	switch remove.Inode.StableAttr.Type {
+	case Directory, SpecialDirectory:
+		return i.InodeOperations.RemoveDirectory(ctx, i, remove.name)
+	default:
+		return i.InodeOperations.Remove(ctx, i, remove.name)
+	}
+}
+
+// Rename calls i.InodeOperations.Rename with the given arguments.
+func (i *Inode) Rename(ctx context.Context, oldParent *Dirent, renamed *Dirent, newParent *Dirent, newName string) error {
+	if i.overlay != nil {
+		return overlayRename(ctx, i.overlay, oldParent, renamed, newParent, newName)
+	}
+	return i.InodeOperations.Rename(ctx, oldParent.Inode, renamed.name, newParent.Inode, newName)
+}
+
+// Bind calls i.InodeOperations.Bind with i as the directory.
+func (i *Inode) Bind(ctx context.Context, name string, data unix.BoundEndpoint, perm FilePermissions) error {
+	if i.overlay != nil {
+		return overlayBind(ctx, i.overlay, name, data, perm)
+	}
+	return i.InodeOperations.Bind(ctx, i, name, data, perm)
+}
+
+// BoundEndpoint calls i.InodeOperations.BoundEndpoint with i as the Inode.
+func (i *Inode) BoundEndpoint(path string) unix.BoundEndpoint {
+	if i.overlay != nil {
+		return overlayBoundEndpoint(i.overlay, path)
+	}
+	return i.InodeOperations.BoundEndpoint(i, path)
+}
+
+// GetFile calls i.InodeOperations.GetFile with the given arguments.
+func (i *Inode) GetFile(ctx context.Context, d *Dirent, flags FileFlags) (*File, error) {
+	if i.overlay != nil {
+		return overlayGetFile(ctx, i.overlay, d, flags)
+	}
+	return i.InodeOperations.GetFile(ctx, d, flags)
+}
+
+// UnstableAttr calls i.InodeOperations.UnstableAttr with i as the Inode.
+func (i *Inode) UnstableAttr(ctx context.Context) (UnstableAttr, error) {
+	if i.overlay != nil {
+		return overlayUnstableAttr(ctx, i.overlay)
+	}
+	return i.InodeOperations.UnstableAttr(ctx, i)
+}
+
+// Getxattr calls i.InodeOperations.Getxattr with i as the Inode.
+func (i *Inode) Getxattr(name string) ([]byte, error) {
+	if i.overlay != nil {
+		return overlayGetxattr(i.overlay, name)
+	}
+	return i.InodeOperations.Getxattr(i, name)
+}
+
+// Listxattr calls i.InodeOperations.Listxattr with i as the Inode.
+func (i *Inode) Listxattr() (map[string]struct{}, error) {
+	if i.overlay != nil {
+		return overlayListxattr(i.overlay)
+	}
+	return i.InodeOperations.Listxattr(i)
+}
+
+// CheckPermission will check if the caller may access this file in the
+// requested way for reading, writing, or executing.
+//
+// CheckPermission is like Linux's fs/namei.c:inode_permission. It
+// - checks file system mount flags,
+// - and utilizes InodeOperations.Check to check capabilities and modes.
+func (i *Inode) CheckPermission(ctx context.Context, p PermMask) error {
+	// First check the outer-most mounted filesystem.
+	if p.Write && i.MountSource.Flags.ReadOnly {
+		return syserror.EROFS
+	}
+
+	if i.overlay != nil {
+		// CheckPermission requires some special handling for
+		// an overlay.
+		//
+		// Writes will always be redirected to an upper filesystem,
+		// so ignore all lower layers being read-only.
+		//
+		// But still honor the upper-most filesystem's mount flags;
+		// we should not attempt to modify the writable layer if it
+		// is mounted read-only.
+		if p.Write && overlayUpperMountSource(i.MountSource).Flags.ReadOnly {
+			return syserror.EROFS
+		}
+	}
+
+	return i.check(ctx, p)
+}
+
+func (i *Inode) check(ctx context.Context, p PermMask) error {
+	if i.overlay != nil {
+		return overlayCheck(ctx, i.overlay, p)
+	}
+	if !i.InodeOperations.Check(ctx, i, p) {
+		return syserror.EACCES
+	}
+	return nil
+}
+
+// SetPermissions calls i.InodeOperations.SetPermissions with i as the Inode.
+func (i *Inode) SetPermissions(ctx context.Context, d *Dirent, f FilePermissions) bool {
+	if i.overlay != nil {
+		return overlaySetPermissions(ctx, i.overlay, d, f)
+	}
+	return i.InodeOperations.SetPermissions(ctx, i, f)
+}
+
+// SetOwner calls i.InodeOperations.SetOwner with i as the Inode.
+func (i *Inode) SetOwner(ctx context.Context, d *Dirent, o FileOwner) error {
+	if i.overlay != nil {
+		return overlaySetOwner(ctx, i.overlay, d, o)
+	}
+	return i.InodeOperations.SetOwner(ctx, i, o)
+}
+
+// SetTimestamps calls i.InodeOperations.SetTimestamps with i as the Inode.
+func (i *Inode) SetTimestamps(ctx context.Context, d *Dirent, ts TimeSpec) error {
+	if i.overlay != nil {
+		return overlaySetTimestamps(ctx, i.overlay, d, ts)
+	}
+	return i.InodeOperations.SetTimestamps(ctx, i, ts)
+}
+
+// Truncate calls i.InodeOperations.Truncate with i as the Inode.
+func (i *Inode) Truncate(ctx context.Context, d *Dirent, size int64) error {
+	if i.overlay != nil {
+		return overlayTruncate(ctx, i.overlay, d, size)
+	}
+	return i.InodeOperations.Truncate(ctx, i, size)
+}
+
+// Readlink calls i.InodeOperations.Readlnk with i as the Inode.
+func (i *Inode) Readlink(ctx context.Context) (string, error) {
+	if i.overlay != nil {
+		return overlayReadlink(ctx, i.overlay)
+	}
+	return i.InodeOperations.Readlink(ctx, i)
+}
+
+// Getlink calls i.InodeOperations.Getlink.
+func (i *Inode) Getlink(ctx context.Context) (*Dirent, error) {
+	if i.overlay != nil {
+		return overlayGetlink(ctx, i.overlay)
+	}
+	return i.InodeOperations.Getlink(ctx, i)
+}
+
+// AddLink calls i.InodeOperations.AddLink.
+func (i *Inode) AddLink() {
+	if i.overlay != nil {
+		// FIXME: Remove this from InodeOperations altogether.
+		//
+		// This interface (including DropLink and NotifyStatusChange)
+		// is only used by ramfs to update metadata of children. These
+		// filesystems should _never_ have overlay Inodes cached as
+		// children. So explicitly disallow this scenario and avoid plumbing
+		// Dirents through to do copy up.
+		panic("overlay Inodes cached in ramfs directories are not supported")
+	}
+	i.InodeOperations.AddLink()
+}
+
+// DropLink calls i.InodeOperations.DropLink.
+func (i *Inode) DropLink() {
+	if i.overlay != nil {
+		// Same as AddLink.
+		panic("overlay Inodes cached in ramfs directories are not supported")
+	}
+	i.InodeOperations.DropLink()
+}
+
+// NotifyStatusChange calls i.InodeOperations.NotifyStatusChange.
+func (i *Inode) NotifyStatusChange(ctx context.Context) {
+	if i.overlay != nil {
+		// Same as AddLink.
+		panic("overlay Inodes cached in ramfs directories are not supported")
+	}
+	i.InodeOperations.NotifyStatusChange(ctx)
+}
+
+// IsVirtual calls i.InodeOperations.IsVirtual.
+func (i *Inode) IsVirtual() bool {
+	if i.overlay != nil {
+		// An overlay configuration does not support virtual files.
+		return false
+	}
+	return i.InodeOperations.IsVirtual()
+}
+
+// StatFS calls i.InodeOperations.StatFS.
+func (i *Inode) StatFS(ctx context.Context) (Info, error) {
+	if i.overlay != nil {
+		return overlayStatFS(ctx, i.overlay)
+	}
+	return i.InodeOperations.StatFS(ctx)
+}
+
+// HandleOps extracts HandleOperations from i.
+func (i *Inode) HandleOps() HandleOperations {
+	if i.overlay != nil {
+		return overlayHandleOps(i.overlay)
+	}
+	if h, ok := i.InodeOperations.(HandleOperations); ok {
+		return h
+	}
+	return nil
+}
+
+// CheckOwnership checks whether `ctx` owns this Inode or may act as its owner.
+// Compare Linux's fs/inode.c:inode_owner_or_capable().
+func (i *Inode) CheckOwnership(ctx context.Context) bool {
+	uattr, err := i.UnstableAttr(ctx)
+	if err != nil {
+		return false
+	}
+	creds := auth.CredentialsFromContext(ctx)
+	if uattr.Owner.UID == creds.EffectiveKUID {
+		return true
+	}
+	if creds.HasCapability(linux.CAP_FOWNER) && creds.UserNamespace.MapFromKUID(uattr.Owner.UID).Ok() {
+		return true
+	}
+	return false
+}
+
+// CheckCapability checks whether `ctx` has capability `cp` with respect to
+// operations on this Inode.
+//
+// Compare Linux's kernel/capability.c:capable_wrt_inode_uidgid(). Note that
+// this function didn't exist in Linux 3.11.10, but was added by upstream
+// 23adbe12ef7d "fs,userns: Change inode_capable to capable_wrt_inode_uidgid"
+// to fix local privilege escalation CVE-2014-4014.
+func (i *Inode) CheckCapability(ctx context.Context, cp linux.Capability) bool {
+	uattr, err := i.UnstableAttr(ctx)
+	if err != nil {
+		return false
+	}
+	creds := auth.CredentialsFromContext(ctx)
+	if !creds.UserNamespace.MapFromKUID(uattr.Owner.UID).Ok() {
+		return false
+	}
+	if !creds.UserNamespace.MapFromKGID(uattr.Owner.GID).Ok() {
+		return false
+	}
+	return creds.HasCapability(cp)
+}
diff --git a/pkg/sentry/fs/inode_inotify.go b/pkg/sentry/fs/inode_inotify.go
new file mode 100644
index 000000000..358bbecdf
--- /dev/null
+++ b/pkg/sentry/fs/inode_inotify.go
@@ -0,0 +1,166 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sync"
+)
+
+// Watches is the collection of inotify watches on an inode.
+type Watches struct {
+	// mu protects the fields below.
+	mu sync.RWMutex `state:"nosave"`
+
+	// ws is the map of active watches in this collection, keyed by the inotify
+	// instance id of the owner.
+	ws map[uint64]*Watch
+
+	// unlinked indicates whether the target inode was ever unlinked. This is a
+	// hack to figure out if we should queue a IN_DELETE_SELF event when this
+	// watches collection is being destroyed, since otherwise we have no way of
+	// knowing if the target inode is going down due to a deletion or
+	// revalidation.
+	unlinked bool
+}
+
+func newWatches() *Watches {
+	return &Watches{
+		ws: make(map[uint64]*Watch),
+	}
+}
+
+// MarkUnlinked indicates the target for this set of watches to be unlinked.
+// This has implications for the IN_EXCL_UNLINK flag.
+func (w *Watches) MarkUnlinked() {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	w.unlinked = true
+}
+
+// Lookup returns a matching watch with the given id. Returns nil if no such
+// watch exists. Note that the result returned by this method only remains valid
+// if the inotify instance owning the watch is locked, preventing modification
+// of the returned watch and preventing the replacement of the watch by another
+// one from the same instance (since there may be at most one watch per
+// instance, per target).
+func (w *Watches) Lookup(id uint64) *Watch {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	return w.ws[id]
+}
+
+// Add adds watch into this set of watches. The watch being added must be unique
+// - its ID() should not collide with any existing watches.
+func (w *Watches) Add(watch *Watch) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	// Sanity check, the new watch shouldn't collide with an existing
+	// watch. Silently replacing an existing watch would result in a ref leak on
+	// this inode. We could handle this collision by calling Unpin() on the
+	// existing watch, but then we end up leaking watch descriptor ids at the
+	// inotify level.
+	if _, exists := w.ws[watch.ID()]; exists {
+		panic(fmt.Sprintf("Watch collision with ID %+v", watch.ID()))
+	}
+	w.ws[watch.ID()] = watch
+}
+
+// Remove removes a watch with the given id from this set of watches. The caller
+// is responsible for generating any watch removal event, as appropriate. The
+// provided id must match an existing watch in this collection.
+func (w *Watches) Remove(id uint64) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+
+	if w.ws == nil {
+		// This watch set is being destroyed. The thread executing the
+		// destructor is already in the process of deleting all our watches. We
+		// got here with no refs on the inode because we raced with the
+		// destructor notifying all the watch owners of the inode's destruction.
+		// See the comment in Watches.TargetDestroyed for why this race exists.
+		return
+	}
+
+	watch, ok := w.ws[id]
+	if !ok {
+		// While there's technically no problem with silently ignoring a missing
+		// watch, this is almost certainly a bug.
+		panic(fmt.Sprintf("Attempt to remove a watch, but no watch found with provided id %+v.", id))
+	}
+	delete(w.ws, watch.ID())
+}
+
+// Notify queues a new event with all watches in this set.
+func (w *Watches) Notify(name string, events, cookie uint32) {
+	// N.B. We don't defer the unlocks because Notify is in the hot path of
+	// all IO operations, and the defer costs too much for small IO
+	// operations.
+	w.mu.RLock()
+	for _, watch := range w.ws {
+		if name != "" && w.unlinked && !watch.NotifyParentAfterUnlink() {
+			// IN_EXCL_UNLINK - By default, when watching events on the children
+			// of a directory, events are generated for children even after they
+			// have been unlinked from the directory. This can result in large
+			// numbers of uninteresting events for some applications (e.g., if
+			// watching /tmp, in which many applications create temporary files
+			// whose names are immediately unlinked). Specifying IN_EXCL_UNLINK
+			// changes the default behavior, so that events are not generated
+			// for children after they have been unlinked from the watched
+			// directory.  -- inotify(7)
+			//
+			// We know we're dealing with events for a parent when the name
+			// isn't empty.
+			continue
+		}
+		watch.Notify(name, events, cookie)
+	}
+	w.mu.RUnlock()
+}
+
+// Unpin unpins dirent from all watches in this set.
+func (w *Watches) Unpin(d *Dirent) {
+	w.mu.RLock()
+	defer w.mu.RUnlock()
+	for _, watch := range w.ws {
+		watch.Unpin(d)
+	}
+}
+
+// targetDestroyed is called by the inode destructor to notify the watch owners
+// of the impending destruction of the watch target.
+func (w *Watches) targetDestroyed() {
+	var ws map[uint64]*Watch
+
+	// We can't hold w.mu while calling watch.TargetDestroyed to preserve lock
+	// ordering w.r.t to the owner inotify instances. Instead, atomically move
+	// the watches map into a local variable so we can iterate over it safely.
+	//
+	// Because of this however, it is possible for the watches' owners to reach
+	// this inode while the inode has no refs. This is still safe because the
+	// owners can only reach the inode until this function finishes calling
+	// watch.TargetDestroyed() below and the inode is guaranteed to exist in the
+	// meanwhile. But we still have to be very careful not to rely on inode
+	// state that may have been already destroyed.
+	w.mu.Lock()
+	ws = w.ws
+	w.ws = nil
+	w.mu.Unlock()
+
+	for _, watch := range ws {
+		watch.TargetDestroyed()
+	}
+}
diff --git a/pkg/sentry/fs/inode_operations.go b/pkg/sentry/fs/inode_operations.go
new file mode 100644
index 000000000..b33980178
--- /dev/null
+++ b/pkg/sentry/fs/inode_operations.go
@@ -0,0 +1,385 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"errors"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+var (
+	// ErrResolveViaReadlink is a special error value returned by
+	// InodeOperations.Getlink() to indicate that a link should be
+	// resolved automatically by walking to the path returned by
+	// InodeOperations.Readlink().
+	ErrResolveViaReadlink = errors.New("link should be resolved via Readlink()")
+)
+
+// TimeSpec contains access and modification timestamps. If either ATimeOmit or
+// MTimeOmit is true, then the corresponding timestamp should not be updated.
+// If either ATimeSetSystemTime or MTimeSetSystemTime are set then the
+// corresponding timestamp should be ignored and the time will be set to the
+// current system time.
+type TimeSpec struct {
+	ATime              ktime.Time
+	ATimeOmit          bool
+	ATimeSetSystemTime bool
+	MTime              ktime.Time
+	MTimeOmit          bool
+	MTimeSetSystemTime bool
+}
+
+// InodeOperations are operations on an Inode that diverge per file system.
+//
+// Objects that implement InodeOperations may cache file system "private"
+// data that is useful for implementing these methods. In contrast, Inode
+// contains state that is common to all Inodes; this state may be optionally
+// used by InodeOperations. An object that implements InodeOperations may
+// not take a reference on an Inode.
+type InodeOperations interface {
+	// Release releases all private file system data held by this object.
+	// Once Release is called, this object is dead (no other methods will
+	// ever be called).
+	Release(context.Context)
+
+	// Lookup loads an Inode at name under dir into a Dirent. The name
+	// is a valid component path: it contains no "/"s nor is the empty
+	// string.
+	//
+	// Lookup may return one of:
+	//
+	// * A nil Dirent and a non-nil error. If the reason that Lookup failed
+	//   was because the name does not exist under Inode, then must return
+	//   syserror.ENOENT.
+	//
+	// * If name does not exist under dir and the file system wishes this
+	//   fact to be cached, a non-nil Dirent containing a nil Inode and a
+	//   nil error. This is a negative Dirent and must have exactly one
+	//   reference (at-construction reference).
+	//
+	// * If name does exist under this dir, a non-nil Dirent containing a
+	//   non-nil Inode, and a nil error. File systems that take extra
+	//   references on this Dirent should implement DirentOperations.
+	Lookup(ctx context.Context, dir *Inode, name string) (*Dirent, error)
+
+	// Create creates an Inode at name under dir and returns a new File
+	// whose Dirent backs the new Inode. Implementations must ensure that
+	// name does not already exist. Create may return one of:
+	//
+	// * A nil File and a non-nil error.
+	//
+	// * A non-nil File and a nil error. File.Dirent will be a new Dirent,
+	// with a single reference held by File. File systems that take extra
+	// references on this Dirent should implement DirentOperations.
+	//
+	// The caller must ensure that this operation is permitted.
+	Create(ctx context.Context, dir *Inode, name string, flags FileFlags, perm FilePermissions) (*File, error)
+
+	// CreateDirectory creates a new directory under this dir.
+	// CreateDirectory should otherwise do the same as Create.
+	//
+	// The caller must ensure that this operation is permitted.
+	CreateDirectory(ctx context.Context, dir *Inode, name string, perm FilePermissions) error
+
+	// CreateLink creates a symbolic link under dir between newname
+	// and oldname. CreateLink should otherwise do the same as Create.
+	//
+	// The caller must ensure that this operation is permitted.
+	CreateLink(ctx context.Context, dir *Inode, oldname string, newname string) error
+
+	// CreateHardLink creates a hard link under dir between the target
+	// Inode and name. Implementations must ensure that name does not
+	// already exist.
+	//
+	// The caller must ensure this operation is permitted.
+	CreateHardLink(ctx context.Context, dir *Inode, target *Inode, name string) error
+
+	// CreateFifo creates a new named pipe under dir at name.
+	// Implementations must ensure that an Inode at name does not
+	// already exist.
+	//
+	// The caller must ensure that this operation is permitted.
+	CreateFifo(ctx context.Context, dir *Inode, name string, perm FilePermissions) error
+
+	// Remove removes the given named non-directory under dir.
+	//
+	// The caller must ensure that this operation is permitted.
+	//
+	// TODO: merge Remove and RemoveDirectory, Remove
+	// just needs a type flag.
+	Remove(ctx context.Context, dir *Inode, name string) error
+
+	// RemoveDirectory removes the given named directory under dir.
+	//
+	// The caller must ensure that this operation is permitted.
+	//
+	// RemoveDirectory should check that the directory to be
+	// removed is empty.
+	RemoveDirectory(ctx context.Context, dir *Inode, name string) error
+
+	// Rename atomically renames oldName under oldParent to newName
+	// under newParent where oldParent and newParent are directories.
+	//
+	// Implementations are responsible for rejecting renames that
+	// replace non-empty directories.
+	Rename(ctx context.Context, oldParent *Inode, oldName string, newParent *Inode, newName string) error
+
+	// Bind binds a new socket under dir at the given name.
+	// Implementations must ensure that name does not already exist.
+	//
+	// The caller must ensure that this operation is permitted.
+	Bind(ctx context.Context, dir *Inode, name string, data unix.BoundEndpoint, perm FilePermissions) error
+
+	// BoundEndpoint returns the socket endpoint at path stored in
+	// or generated by an Inode.
+	//
+	// The path is only relevant for generated endpoint because stored
+	// endpoints already know their path. It is ok for the endpoint to
+	// hold onto their path because the only way to change a bind
+	// address is to rebind the socket.
+	//
+	// This is valid iff the type of the Inode is a Socket, which
+	// generally implies that this Inode was created via CreateSocket.
+	//
+	// If there is no socket endpoint available, nil will be returned.
+	BoundEndpoint(inode *Inode, path string) unix.BoundEndpoint
+
+	// GetFile returns a new open File backed by a Dirent and FileFlags.
+	// It may block as long as it is done with ctx.
+	//
+	// The returned File will uniquely back an application fd.
+	GetFile(ctx context.Context, d *Dirent, flags FileFlags) (*File, error)
+
+	// UnstableAttr returns the most up-to-date "unstable" attributes of
+	// an Inode, where "unstable" means that they change in response to
+	// file system events.
+	UnstableAttr(ctx context.Context, inode *Inode) (UnstableAttr, error)
+
+	// Getxattr retrieves the value of extended attribute name. Inodes that
+	// do not support extended attributes return EOPNOTSUPP. Inodes that
+	// support extended attributes but don't have a value at name return
+	// ENODATA.
+	Getxattr(inode *Inode, name string) ([]byte, error)
+
+	// Setxattr sets the value of extended attribute name. Inodes that
+	// do not support extended attributes return EOPNOTSUPP.
+	Setxattr(inode *Inode, name string, value []byte) error
+
+	// Listxattr returns the set of all extended attributes names that
+	// have values. Inodes that do not support extended attributes return
+	// EOPNOTSUPP.
+	Listxattr(inode *Inode) (map[string]struct{}, error)
+
+	// Check determines whether an Inode can be accessed with the
+	// requested permission mask using the context (which gives access
+	// to Credentials and UserNamespace).
+	Check(ctx context.Context, inode *Inode, p PermMask) bool
+
+	// SetPermissions sets new permissions for an Inode.  Returns false
+	// if it was not possible to set the new permissions.
+	//
+	// The caller must ensure that this operation is permitted.
+	SetPermissions(ctx context.Context, inode *Inode, f FilePermissions) bool
+
+	// SetOwner sets the ownership for this file.
+	//
+	// If either UID or GID are set to auth.NoID, its value will not be
+	// changed.
+	//
+	// The caller must ensure that this operation is permitted.
+	SetOwner(ctx context.Context, inode *Inode, owner FileOwner) error
+
+	// SetTimestamps sets the access and modification timestamps of an
+	// Inode according to the access and modification times in the TimeSpec.
+	//
+	// If either ATimeOmit or MTimeOmit is set, then the corresponding
+	// timestamp is not updated.
+	//
+	// If either ATimeSetSystemTime or MTimeSetSystemTime is true, that
+	// timestamp is set to the current time instead.
+	//
+	// The caller must ensure that this operation is permitted.
+	SetTimestamps(ctx context.Context, inode *Inode, ts TimeSpec) error
+
+	// Truncate changes the size of an Inode. Truncate should not check
+	// permissions internally, as it is used for both sys_truncate and
+	// sys_ftruncate.
+	//
+	// Implementations need not check that length >= 0.
+	Truncate(ctx context.Context, inode *Inode, size int64) error
+
+	// WriteOut writes cached Inode state to a backing filesystem in a
+	// synchronous manner.
+	//
+	// File systems that do not cache metadata or data via an Inode
+	// implement WriteOut as a no-op. File systems that are entirely in
+	// memory also implement WriteOut as a no-op. Otherwise file systems
+	// call Inode.Sync to write back page cached data and cached metadata
+	// followed by syncing writeback handles.
+	//
+	// It derives from include/linux/fs.h:super_operations->write_inode.
+	WriteOut(ctx context.Context, inode *Inode) error
+
+	// Readlink reads the symlink path of an Inode.
+	//
+	// Readlink is permitted to return a different path depending on ctx,
+	// the request originator.
+	//
+	// The caller must ensure that this operation is permitted.
+	//
+	// Readlink should check that Inode is a symlink and its content is
+	// at least readable.
+	Readlink(ctx context.Context, inode *Inode) (string, error)
+
+	// Getlink resolves a symlink to a target *Dirent.
+	//
+	// Filesystems that can resolve the link by walking to the path returned
+	// by Readlink should return (nil, ErrResolveViaReadlink), which
+	// triggers link resolution via Realink and Lookup.
+	//
+	// Some links cannot be followed by Lookup. In this case, Getlink can
+	// return the Dirent of the link target. The caller holds a reference
+	// to the Dirent. Filesystems that return a non-nil *Dirent from Getlink
+	// cannot participate in an overlay because it is impossible for the
+	// overlay to ascertain whether or not the *Dirent should contain an
+	// overlayEntry.
+	//
+	// Any error returned from Getlink other than ErrResolveViaReadlink
+	// indicates the caller's inability to traverse this Inode as a link
+	// (e.g. syserror.ENOLINK indicates that the Inode is not a link,
+	// syscall.EPERM indicates that traversing the link is not allowed, etc).
+	Getlink(context.Context, *Inode) (*Dirent, error)
+
+	// Mappable returns a memmap.Mappable that provides memory mappings of the
+	// Inode's data. Mappable may return nil if this is not supported. The
+	// returned Mappable must remain valid until InodeOperations.Release is
+	// called.
+	Mappable(*Inode) memmap.Mappable
+
+	// The below methods require cleanup.
+
+	// AddLink increments the hard link count of an Inode.
+	//
+	// Remove in favor of Inode.IncLink.
+	AddLink()
+
+	// DropLink decrements the hard link count of an Inode.
+	//
+	// Remove in favor of Inode.DecLink.
+	DropLink()
+
+	// NotifyStatusChange sets the status change time to the current time.
+	//
+	// Remove in favor of updating the Inode's cached status change time.
+	NotifyStatusChange(ctx context.Context)
+
+	// IsVirtual indicates whether or not this corresponds to a virtual
+	// resource.
+	//
+	// If IsVirtual returns true, then caching will be disabled for this
+	// node, and fs.Dirent.Freeze() will not stop operations on the node.
+	//
+	// Remove in favor of freezing specific mounts.
+	IsVirtual() bool
+
+	// StatFS returns a filesystem Info implementation or an error.  If
+	// the filesystem does not support this operation (maybe in the future
+	// it will), then ENOSYS should be returned.
+	//
+	// Move to MountSourceOperations.
+	StatFS(context.Context) (Info, error)
+
+	HandleOperations
+}
+
+// HandleOperations are extended InodeOperations that are only implemented
+// for file systems that use fs/handle.go:Handle to generate open Files.
+//
+// Handle is deprecated; these methods are deprecated as well.
+//
+// Filesystems are encouraged to implement the File interface directly
+// instead of using Handle. To indicate that the below methods should never
+// be called, embed DeprecatedFileOperations to satisfy this interface.
+type HandleOperations interface {
+	waiter.Waitable
+
+	// DeprecatedPreadv is deprecated in favor of filesystems
+	// implementing File.Preadv directly.
+	//
+	// DeprecatedPreadv reads up to dst.NumBytes() bytes into dst, starting at
+	// the given offset, and returns the number of bytes read.
+	//
+	// Preadv may return a partial read result before EOF is reached.
+	//
+	// If a symlink, Preadv reads the target value of the symlink.
+	//
+	// Preadv should not check for readable permissions.
+	DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error)
+
+	// DeprecatedPwritev is deprecated in favor of filesystems
+	// implementing File.Pwritev directly.
+	//
+	// DeprecatedPwritev writes up to src.NumBytes() bytes from src to the
+	// Inode, starting at the given offset and returns the number of bytes
+	// written.
+	//
+	// Pwritev should not check that the Inode has writable permissions.
+	DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error)
+
+	// DeprecatedReaddir is deprecated in favor of filesystems
+	// implementing File.Readdir directly.
+	//
+	// DeprecatedReaddir emits directory entries by calling dirCtx.EmitDir,
+	// beginning with the entry at offset.
+	//
+	// Entries for "." and ".." must *not* be included.
+	//
+	// If the offset returned is the same as the argument offset, then
+	// nothing has been serialized.  This is equivalent to reaching EOF.
+	// In this case serializer.Written() should return 0.
+	//
+	// The order of entries to emit must be consistent between Readdir
+	// calls, and must start with the given offset.
+	//
+	// The caller must ensure that this operation is permitted.
+	DeprecatedReaddir(ctx context.Context, dirCtx *DirCtx, offset int) (int, error)
+
+	// DeprecatedFsync is deprecated in favor of filesystems implementing
+	// File.Fsync directly.
+	//
+	// DeprecatedFsync syncs a file.
+	DeprecatedFsync() error
+
+	// DeprecatedMappable is deprecated in favor of filesystems implementing
+	// File.Mappable directly.
+	//
+	// DeprecatedMappable returns a Mappable if the Inode can be mapped.
+	DeprecatedMappable(ctx context.Context, inode *Inode) (memmap.Mappable, bool)
+
+	// DeprecatedFlush is deprecated in favor of filesystems implementing
+	// File.Flush directly.
+	//
+	// DeprecatedFlush flushes a file.
+	//
+	// Implementations may choose to free up memory or complete pending I/O
+	// but also may implement Flush as a no-op.
+	DeprecatedFlush() error
+}
diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go
new file mode 100644
index 000000000..343150bb8
--- /dev/null
+++ b/pkg/sentry/fs/inode_overlay.go
@@ -0,0 +1,555 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+func overlayHasWhiteout(parent *Inode, name string) bool {
+	buf, err := parent.Getxattr(XattrOverlayWhiteout(name))
+	return err == nil && string(buf) == "y"
+}
+
+func overlayCreateWhiteout(parent *Inode, name string) error {
+	return parent.InodeOperations.Setxattr(parent, XattrOverlayWhiteout(name), []byte("y"))
+}
+
+func overlayWriteOut(ctx context.Context, o *overlayEntry) error {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper == nil {
+		return nil
+	}
+	return o.upper.InodeOperations.WriteOut(ctx, o.upper)
+}
+
+func overlayLookup(ctx context.Context, parent *overlayEntry, inode *Inode, name string) (*Dirent, error) {
+	parent.copyMu.RLock()
+	defer parent.copyMu.RUnlock()
+
+	// Assert that there is at least one upper or lower entry.
+	if parent.upper == nil && parent.lower == nil {
+		panic("invalid overlayEntry, needs at least one Inode")
+	}
+
+	var upperInode *Inode
+	var lowerInode *Inode
+
+	// Does the parent directory exist in the upper file system?
+	if parent.upper != nil {
+		// First check if a file object exists in the upper file system.
+		// A file could have been created over a whiteout, so we need to
+		// check if something exists in the upper file system first.
+		child, err := parent.upper.Lookup(ctx, name)
+		if err != nil && err != syserror.ENOENT {
+			// We encountered an error that an overlay cannot handle,
+			// we must propagate it to the caller.
+			return nil, err
+		}
+		if child != nil {
+			defer child.DecRef()
+
+			// Is the child non-negative?
+			if !child.IsNegative() {
+				upperInode = child.Inode
+				upperInode.IncRef()
+			}
+		}
+
+		// Are we done?
+		if overlayHasWhiteout(parent.upper, name) {
+			if upperInode == nil {
+				return NewNegativeDirent(name), nil
+			}
+			entry, err := newOverlayEntry(ctx, upperInode, nil, false)
+			if err != nil {
+				// Don't leak resources.
+				upperInode.DecRef()
+				return nil, err
+			}
+			return NewDirent(newOverlayInode(ctx, entry, inode.MountSource), name), nil
+		}
+	}
+
+	// Check the lower file system. We do this unconditionally (even for
+	// non-directories) because we may need to use stable attributes from
+	// the lower filesystem (e.g. device number, inode number) that were
+	// visible before a copy up.
+	if parent.lower != nil {
+		// Check the lower file system.
+		child, err := parent.lower.Lookup(ctx, name)
+		// Same song and dance as above.
+		if err != nil && err != syserror.ENOENT {
+			// Don't leak resources.
+			if upperInode != nil {
+				upperInode.DecRef()
+			}
+			return nil, err
+		}
+		if child != nil {
+			defer child.DecRef()
+
+			// Is the child negative?
+			if !child.IsNegative() {
+				// Did we find something in the upper filesystem? We can
+				// only use it if the types match.
+				if upperInode == nil || upperInode.StableAttr.Type == child.Inode.StableAttr.Type {
+					lowerInode = child.Inode
+					lowerInode.IncRef()
+				}
+			}
+		}
+	}
+
+	// Was all of this for naught?
+	if upperInode == nil && lowerInode == nil {
+		// Return a negative Dirent indicating that nothing was found.
+		return NewNegativeDirent(name), nil
+	}
+
+	// Did we find a lower Inode? Remember this because we may decide we don't
+	// actually need the lower Inode (see below).
+	lowerExists := lowerInode != nil
+
+	// If we found something in the upper filesystem and the lower filesystem,
+	// use the stable attributes from the lower filesystem. If we don't do this,
+	// then it may appear that the file was magically recreated across copy up.
+	if upperInode != nil && lowerInode != nil {
+		// Steal attributes.
+		upperInode.StableAttr = lowerInode.StableAttr
+
+		// For non-directories, the lower filesystem resource is strictly
+		// unnecessary because we don't need to copy-up and we will always
+		// operate (e.g. read/write) on the upper Inode.
+		if !IsDir(upperInode.StableAttr) {
+			lowerInode.DecRef()
+			lowerInode = nil
+		}
+	}
+
+	// Phew, finally done.
+	entry, err := newOverlayEntry(ctx, upperInode, lowerInode, lowerExists)
+	if err != nil {
+		// Well, not quite, we failed at the last moment, how depressing.
+		// Be sure not to leak resources.
+		if upperInode != nil {
+			upperInode.DecRef()
+		}
+		if lowerInode != nil {
+			lowerInode.DecRef()
+		}
+		return nil, err
+	}
+	return NewDirent(newOverlayInode(ctx, entry, inode.MountSource), name), nil
+}
+
+func overlayCreate(ctx context.Context, o *overlayEntry, parent *Dirent, name string, flags FileFlags, perm FilePermissions) (*File, error) {
+	// Dirent.Create takes renameMu if the Inode is an overlay Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return nil, err
+	}
+
+	upperFile, err := o.upper.InodeOperations.Create(ctx, o.upper, name, flags, perm)
+	if err != nil {
+		return nil, err
+	}
+
+	// Take another reference on the upper file's inode, which will be
+	// owned by the overlay entry.
+	upperFile.Dirent.Inode.IncRef()
+	entry, err := newOverlayEntry(ctx, upperFile.Dirent.Inode, nil, false)
+	if err != nil {
+		cleanupUpper(ctx, o.upper, name)
+		return nil, err
+	}
+
+	// NOTE: Replace the Dirent with a transient Dirent, since
+	// we are about to create the real Dirent: an overlay Dirent.
+	//
+	// This ensures the *fs.File returned from overlayCreate is in the same
+	// state as the *fs.File returned by overlayGetFile, where the upper
+	// file has a transient Dirent.
+	//
+	// This is necessary for Save/Restore, as otherwise the upper Dirent
+	// (which has no path as it is unparented and never reachable by the
+	// user) will clobber the real path for the underlying Inode.
+	upperFile.Dirent.Inode.IncRef()
+	upperDirent := NewTransientDirent(upperFile.Dirent.Inode)
+	upperFile.Dirent.DecRef()
+	upperFile.Dirent = upperDirent
+
+	// Create the overlay inode and dirent.  We need this to construct the
+	// overlay file.
+	overlayInode := newOverlayInode(ctx, entry, parent.Inode.MountSource)
+	// d will own the inode reference.
+	overlayDirent := NewDirent(overlayInode, name)
+	// The overlay file created below with NewFile will take a reference on
+	// the overlayDirent, and it should be the only thing holding a
+	// reference at the time of creation, so we must drop this reference.
+	defer overlayDirent.DecRef()
+
+	// Create a new overlay file that wraps the upper file.
+	flags.Pread = upperFile.Flags().Pread
+	flags.Pwrite = upperFile.Flags().Pwrite
+	overlayFile := NewFile(ctx, overlayDirent, flags, &overlayFileOperations{upper: upperFile})
+
+	return overlayFile, nil
+}
+
+func overlayCreateDirectory(ctx context.Context, o *overlayEntry, parent *Dirent, name string, perm FilePermissions) error {
+	// Dirent.CreateDirectory takes renameMu if the Inode is an overlay
+	// Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.CreateDirectory(ctx, o.upper, name, perm)
+}
+
+func overlayCreateLink(ctx context.Context, o *overlayEntry, parent *Dirent, oldname string, newname string) error {
+	// Dirent.CreateLink takes renameMu if the Inode is an overlay Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.CreateLink(ctx, o.upper, oldname, newname)
+}
+
+func overlayCreateHardLink(ctx context.Context, o *overlayEntry, parent *Dirent, target *Dirent, name string) error {
+	// Dirent.CreateHardLink takes renameMu if the Inode is an overlay
+	// Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return err
+	}
+	if err := copyUpLockedForRename(ctx, target); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.CreateHardLink(ctx, o.upper, target.Inode.overlay.upper, name)
+}
+
+func overlayCreateFifo(ctx context.Context, o *overlayEntry, parent *Dirent, name string, perm FilePermissions) error {
+	// Dirent.CreateFifo takes renameMu if the Inode is an overlay Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.CreateFifo(ctx, o.upper, name, perm)
+}
+
+func overlayRemove(ctx context.Context, o *overlayEntry, parent *Dirent, child *Dirent) error {
+	// Dirent.Remove and Dirent.RemoveDirectory take renameMu if the Inode
+	// is an overlay Inode.
+	if err := copyUpLockedForRename(ctx, parent); err != nil {
+		return err
+	}
+	child.Inode.overlay.copyMu.RLock()
+	defer child.Inode.overlay.copyMu.RUnlock()
+	if child.Inode.overlay.upper != nil {
+		if child.Inode.StableAttr.Type == Directory {
+			if err := o.upper.InodeOperations.RemoveDirectory(ctx, o.upper, child.name); err != nil {
+				return err
+			}
+		} else {
+			if err := o.upper.InodeOperations.Remove(ctx, o.upper, child.name); err != nil {
+				return err
+			}
+		}
+	}
+	if child.Inode.overlay.lowerExists {
+		return overlayCreateWhiteout(o.upper, child.name)
+	}
+	return nil
+}
+
+func overlayRename(ctx context.Context, o *overlayEntry, oldParent *Dirent, renamed *Dirent, newParent *Dirent, newName string) error {
+	// To be able to copy these up below, they have to be part of an
+	// overlay file system.
+	//
+	// Maybe some day we can allow the more complicated case of
+	// non-overlay X overlay renames, but that's not necessary right now.
+	if renamed.Inode.overlay == nil || newParent.Inode.overlay == nil || oldParent.Inode.overlay == nil {
+		return syserror.EXDEV
+	}
+
+	// Check here if the file to be replaced exists and is a non-empty
+	// directory. If we copy up first, we may end up copying the directory
+	// but none of its children, so the directory will appear empty in the
+	// upper fs, which will then allow the rename to proceed when it should
+	// return ENOTEMPTY.
+	replaced, err := newParent.Inode.Lookup(ctx, newName)
+	if err != nil && err != syserror.ENOENT {
+		return err
+	}
+	if err == nil && !replaced.IsNegative() && IsDir(replaced.Inode.StableAttr) {
+		children, err := readdirOne(ctx, replaced)
+		if err != nil {
+			return err
+		}
+
+		// readdirOne ensures that "." and ".." are not
+		// included among the returned children, so we don't
+		// need to bother checking for them.
+		if len(children) > 0 {
+			return syserror.ENOTEMPTY
+		}
+	}
+	if err := copyUpLockedForRename(ctx, renamed); err != nil {
+		return err
+	}
+	if err := copyUpLockedForRename(ctx, newParent); err != nil {
+		return err
+	}
+	oldName := renamed.name
+	if err := o.upper.InodeOperations.Rename(ctx, oldParent.Inode.overlay.upper, oldName, newParent.Inode.overlay.upper, newName); err != nil {
+		return err
+	}
+	if renamed.Inode.overlay.lowerExists {
+		return overlayCreateWhiteout(oldParent.Inode.overlay.upper, oldName)
+	}
+	return nil
+}
+
+func overlayBind(ctx context.Context, o *overlayEntry, name string, data unix.BoundEndpoint, perm FilePermissions) error {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	// We do not support doing anything exciting with sockets unless there
+	// is already a directory in the upper filesystem.
+	if o.upper == nil {
+		return syserror.EOPNOTSUPP
+	}
+	return o.upper.InodeOperations.Bind(ctx, o.upper, name, data, perm)
+}
+
+func overlayBoundEndpoint(o *overlayEntry, path string) unix.BoundEndpoint {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	if o.upper != nil {
+		return o.upper.InodeOperations.BoundEndpoint(o.upper, path)
+	}
+	// If a socket is already in the lower file system, allow connections
+	// to it.
+	return o.lower.InodeOperations.BoundEndpoint(o.lower, path)
+}
+
+func overlayGetFile(ctx context.Context, o *overlayEntry, d *Dirent, flags FileFlags) (*File, error) {
+	if flags.Write {
+		if err := copyUp(ctx, d); err != nil {
+			return nil, err
+		}
+	}
+
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	if o.upper != nil {
+		upper, err := overlayFile(ctx, o.upper, flags)
+		if err != nil {
+			return nil, err
+		}
+		flags.Pread = upper.Flags().Pread
+		flags.Pwrite = upper.Flags().Pwrite
+		return NewFile(ctx, d, flags, &overlayFileOperations{upper: upper}), nil
+	}
+
+	lower, err := overlayFile(ctx, o.lower, flags)
+	if err != nil {
+		return nil, err
+	}
+	flags.Pread = lower.Flags().Pread
+	flags.Pwrite = lower.Flags().Pwrite
+	return NewFile(ctx, d, flags, &overlayFileOperations{lower: lower}), nil
+}
+
+func overlayUnstableAttr(ctx context.Context, o *overlayEntry) (UnstableAttr, error) {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper != nil {
+		return o.upper.UnstableAttr(ctx)
+	}
+	return o.lower.UnstableAttr(ctx)
+}
+
+func overlayGetxattr(o *overlayEntry, name string) ([]byte, error) {
+	// Don't forward the value of the extended attribute if it would
+	// unexpectedly change the behavior of a wrapping overlay layer.
+	if strings.HasPrefix(XattrOverlayPrefix, name) {
+		return nil, syserror.ENODATA
+	}
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper != nil {
+		return o.upper.Getxattr(name)
+	}
+	return o.lower.Getxattr(name)
+}
+
+func overlayListxattr(o *overlayEntry) (map[string]struct{}, error) {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	var names map[string]struct{}
+	var err error
+	if o.upper != nil {
+		names, err = o.upper.Listxattr()
+	} else {
+		names, err = o.lower.Listxattr()
+	}
+	for name := range names {
+		// Same as overlayGetxattr, we shouldn't forward along
+		// overlay attributes.
+		if strings.HasPrefix(XattrOverlayPrefix, name) {
+			delete(names, name)
+		}
+	}
+	return names, err
+}
+
+func overlayCheck(ctx context.Context, o *overlayEntry, p PermMask) error {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper != nil {
+		return o.upper.check(ctx, p)
+	}
+	if p.Write {
+		// Since writes will be redirected to the upper filesystem, the lower
+		// filesystem need not be writable, but must be readable for copy-up.
+		p.Write = false
+		p.Read = true
+	}
+	return o.lower.check(ctx, p)
+}
+
+func overlaySetPermissions(ctx context.Context, o *overlayEntry, d *Dirent, f FilePermissions) bool {
+	if err := copyUp(ctx, d); err != nil {
+		return false
+	}
+	return o.upper.InodeOperations.SetPermissions(ctx, o.upper, f)
+}
+
+func overlaySetOwner(ctx context.Context, o *overlayEntry, d *Dirent, owner FileOwner) error {
+	if err := copyUp(ctx, d); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.SetOwner(ctx, o.upper, owner)
+}
+
+func overlaySetTimestamps(ctx context.Context, o *overlayEntry, d *Dirent, ts TimeSpec) error {
+	if err := copyUp(ctx, d); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.SetTimestamps(ctx, o.upper, ts)
+}
+
+func overlayTruncate(ctx context.Context, o *overlayEntry, d *Dirent, size int64) error {
+	if err := copyUp(ctx, d); err != nil {
+		return err
+	}
+	return o.upper.InodeOperations.Truncate(ctx, o.upper, size)
+}
+
+func overlayReadlink(ctx context.Context, o *overlayEntry) (string, error) {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper != nil {
+		return o.upper.Readlink(ctx)
+	}
+	return o.lower.Readlink(ctx)
+}
+
+func overlayGetlink(ctx context.Context, o *overlayEntry) (*Dirent, error) {
+	var dirent *Dirent
+	var err error
+
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	if o.upper != nil {
+		dirent, err = o.upper.Getlink(ctx)
+	} else {
+		dirent, err = o.lower.Getlink(ctx)
+	}
+	if dirent != nil {
+		// This dirent is likely bogus (its Inode likely doesn't contain
+		// the right overlayEntry). So we're forced to drop it on the
+		// ground and claim that jumping around the filesystem like this
+		// is not supported.
+		name, _ := dirent.FullName(nil)
+		dirent.DecRef()
+
+		// Claim that the path is not accessible.
+		err = syserror.EACCES
+		log.Warningf("Getlink not supported in overlay for %q", name)
+	}
+	return nil, err
+}
+
+func overlayStatFS(ctx context.Context, o *overlayEntry) (Info, error) {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+
+	var i Info
+	var err error
+	if o.upper != nil {
+		i, err = o.upper.StatFS(ctx)
+	} else {
+		i, err = o.lower.StatFS(ctx)
+	}
+	if err != nil {
+		return Info{}, err
+	}
+
+	i.Type = linux.OVERLAYFS_SUPER_MAGIC
+
+	return i, nil
+}
+
+func overlayHandleOps(o *overlayEntry) HandleOperations {
+	o.copyMu.RLock()
+	defer o.copyMu.RUnlock()
+	if o.upper != nil {
+		return o.upper.HandleOps()
+	}
+	return o.lower.HandleOps()
+}
+
+// NewTestOverlayDir returns an overlay Inode for tests.
+func NewTestOverlayDir(ctx context.Context, upper *Inode, lower *Inode) *Inode {
+	fs := &overlayFilesystem{}
+	msrc := NewMountSource(&overlayMountSourceOperations{
+		upper: NewNonCachingMountSource(fs, MountSourceFlags{}),
+		lower: NewNonCachingMountSource(fs, MountSourceFlags{}),
+	}, fs, MountSourceFlags{})
+	overlay := &overlayEntry{
+		upper: upper,
+		lower: lower,
+	}
+	return newOverlayInode(ctx, overlay, msrc)
+}
+
+// TestHasUpperFS returns true if i is an overlay Inode and it has a pointer
+// to an Inode on an upper filesystem.
+func (i *Inode) TestHasUpperFS() bool {
+	return i.overlay != nil && i.overlay.upper != nil
+}
+
+// TestHasLowerFS returns true if i is an overlay Inode and it has a pointer
+// to an Inode on a lower filesystem.
+func (i *Inode) TestHasLowerFS() bool {
+	return i.overlay != nil && i.overlay.lower != nil
+}
diff --git a/pkg/sentry/fs/inode_overlay_test.go b/pkg/sentry/fs/inode_overlay_test.go
new file mode 100644
index 000000000..684d54bd2
--- /dev/null
+++ b/pkg/sentry/fs/inode_overlay_test.go
@@ -0,0 +1,251 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs_test
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ramfstest "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs/test"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func TestLookup(t *testing.T) {
+	ctx := contexttest.Context(t)
+	for _, test := range []struct {
+		// Test description.
+		desc string
+
+		// Lookup parameters.
+		dir  *fs.Inode
+		name string
+
+		// Want from lookup.
+		err      error
+		found    bool
+		hasUpper bool
+		hasLower bool
+	}{
+		{
+			desc: "no upper, lower has name",
+			dir: fs.NewTestOverlayDir(ctx,
+				nil, /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: false,
+			hasLower: true,
+		},
+		{
+			desc: "no lower, upper has name",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* upper */
+				nil, /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: true,
+			hasLower: false,
+		},
+		{
+			desc: "upper and lower, only lower has name",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "b",
+						dir:  false,
+					},
+				}, nil), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: false,
+			hasLower: true,
+		},
+		{
+			desc: "upper and lower, only upper has name",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "b",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: true,
+			hasLower: false,
+		},
+		{
+			desc: "upper and lower, both have file",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: true,
+			hasLower: false,
+		},
+		{
+			desc: "upper and lower, both have directory",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  true,
+					},
+				}, nil), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  true,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: true,
+			hasLower: true,
+		},
+		{
+			desc: "upper and lower, upper negative masks lower file",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, nil, []string{"a"}), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    false,
+			hasUpper: false,
+			hasLower: false,
+		},
+		{
+			desc: "upper and lower, upper negative does not mask lower file",
+			dir: fs.NewTestOverlayDir(ctx,
+				newTestRamfsDir(ctx, nil, []string{"b"}), /* upper */
+				newTestRamfsDir(ctx, []dirContent{
+					{
+						name: "a",
+						dir:  false,
+					},
+				}, nil), /* lower */
+			),
+			name:     "a",
+			found:    true,
+			hasUpper: false,
+			hasLower: true,
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			dirent, err := test.dir.Lookup(ctx, test.name)
+			if err != test.err {
+				t.Fatalf("lookup got error %v, want %v", err, test.err)
+			}
+			if test.found && dirent.IsNegative() {
+				t.Fatalf("lookup expected to find %q, got negative dirent", test.name)
+			}
+			if !test.found {
+				return
+			}
+			if hasUpper := dirent.Inode.TestHasUpperFS(); hasUpper != test.hasUpper {
+				t.Fatalf("lookup got upper filesystem %v, want %v", hasUpper, test.hasUpper)
+			}
+			if hasLower := dirent.Inode.TestHasLowerFS(); hasLower != test.hasLower {
+				t.Errorf("lookup got lower filesystem %v, want %v", hasLower, test.hasLower)
+			}
+		})
+	}
+}
+
+type dir struct {
+	fs.InodeOperations
+
+	// list of negative child names.
+	negative []string
+}
+
+func (d *dir) Getxattr(inode *fs.Inode, name string) ([]byte, error) {
+	for _, n := range d.negative {
+		if name == fs.XattrOverlayWhiteout(n) {
+			return []byte("y"), nil
+		}
+	}
+	return nil, syserror.ENOATTR
+}
+
+type dirContent struct {
+	name string
+	dir  bool
+}
+
+func newTestRamfsDir(ctx context.Context, contains []dirContent, negative []string) *fs.Inode {
+	msrc := fs.NewCachingMountSource(nil, fs.MountSourceFlags{})
+	contents := make(map[string]*fs.Inode)
+	for _, c := range contains {
+		if c.dir {
+			contents[c.name] = newTestRamfsDir(ctx, nil, nil)
+		} else {
+			contents[c.name] = fs.NewInode(ramfstest.NewFile(ctx, fs.FilePermissions{}), msrc, fs.StableAttr{Type: fs.RegularFile})
+		}
+	}
+	dops := ramfstest.NewDir(ctx, contents, fs.FilePermissions{
+		User: fs.PermMask{Read: true, Execute: true},
+	})
+	return fs.NewInode(&dir{
+		InodeOperations: dops,
+		negative:        negative,
+	}, msrc, fs.StableAttr{Type: fs.Directory})
+}
diff --git a/pkg/sentry/fs/inotify.go b/pkg/sentry/fs/inotify.go
new file mode 100644
index 000000000..9f50cb800
--- /dev/null
+++ b/pkg/sentry/fs/inotify.go
@@ -0,0 +1,329 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Inotify represents an inotify instance created by inotify_init(2) or
+// inotify_init1(2). Inotify implements the FileOperations interface.
+//
+// Lock ordering:
+//   Inotify.mu -> Inode.Watches.mu -> Watch.mu -> Inotify.evMu
+type Inotify struct {
+	// Unique identifier for this inotify instance. We don't just reuse the
+	// inotify fd because fds can be duped. These should not be exposed to the
+	// user, since we may aggressively reuse an id on S/R.
+	id uint64
+
+	// evMu *only* protects the event queue. We need a separate lock because
+	// while queuing events, a watch needs to lock the event queue, and using mu
+	// for that would violate lock ordering since at that point the calling
+	// goroutine already holds Watch.target.Watches.mu.
+	evMu sync.Mutex `state:"nosave"`
+
+	waiter.Queue `state:"nosave"`
+
+	// A list of pending events for this inotify instance. Protected by evMu.
+	events ilist.List
+
+	// A scratch buffer, use to serialize inotify events. Use allocate this
+	// ahead of time and reuse performance. Protected by evMu.
+	scratch []byte
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// The next watch descriptor number to use for this inotify instance. Note
+	// that Linux starts numbering watch descriptors from 1.
+	nextWatch int32
+
+	// Map from watch descriptors to watch objects.
+	watches map[int32]*Watch
+}
+
+// NewInotify constructs a new Inotify instance.
+func NewInotify(ctx context.Context) *Inotify {
+	return &Inotify{
+		id:        uniqueid.GlobalFromContext(ctx),
+		scratch:   make([]byte, inotifyEventBaseSize),
+		nextWatch: 1, // Linux starts numbering watch descriptors from 1.
+		watches:   make(map[int32]*Watch),
+	}
+}
+
+// Release implements FileOperations.Release. Release removes all watches and
+// frees all resources for an inotify instance.
+func (i *Inotify) Release() {
+	// We need to hold i.mu to avoid a race with concurrent calls to
+	// Inotify.targetDestroyed from Watches. There's no risk of Watches
+	// accessing this Inotify after the destructor ends, because we remove all
+	// references to it below.
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	for _, w := range i.watches {
+		// Remove references to the watch from the watch target. We don't need
+		// to worry about the references from the owner instance, since we're in
+		// the owner's destructor.
+		w.target.Watches.Remove(w.ID())
+		// Don't leak any references to the target, held by pins in the watch.
+		w.destroy()
+	}
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+//
+// Readiness indicates whether there are pending events for an inotify instance.
+func (i *Inotify) Readiness(mask waiter.EventMask) waiter.EventMask {
+	ready := waiter.EventMask(0)
+
+	i.evMu.Lock()
+	defer i.evMu.Unlock()
+
+	if !i.events.Empty() {
+		ready |= waiter.EventIn
+	}
+
+	return mask & ready
+}
+
+// Seek implements FileOperations.Seek.
+func (*Inotify) Seek(context.Context, *File, SeekWhence, int64) (int64, error) {
+	return 0, syserror.ESPIPE
+}
+
+// Readdir implements FileOperatons.Readdir.
+func (*Inotify) Readdir(context.Context, *File, DentrySerializer) (int64, error) {
+	return 0, syserror.ENOTDIR
+}
+
+// Write implements FileOperations.Write.
+func (*Inotify) Write(context.Context, *File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EBADF
+}
+
+// Read implements FileOperations.Read.
+func (i *Inotify) Read(ctx context.Context, _ *File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() < inotifyEventBaseSize {
+		return 0, syserror.EINVAL
+	}
+
+	i.evMu.Lock()
+	defer i.evMu.Unlock()
+
+	if i.events.Empty() {
+		// Nothing to read yet, tell caller to block.
+		return 0, syserror.ErrWouldBlock
+	}
+
+	var writeLen int64
+	for e := i.events.Front(); e != nil; e = e.Next() {
+		event := e.(*Event)
+
+		// Does the buffer have enough remaining space to hold the event we're
+		// about to write out?
+		if dst.NumBytes() < int64(event.sizeOf()) {
+			if writeLen > 0 {
+				// Buffer wasn't big enough for all pending events, but we did
+				// write some events out.
+				return writeLen, nil
+			}
+			return 0, syserror.EINVAL
+		}
+
+		// Linux always dequeues an available event as long as there's enough
+		// buffer space to copy it out, even if the copy below fails. Emulate
+		// this behaviour.
+		i.events.Remove(e)
+
+		// Buffer has enough space, copy event to the read buffer.
+		n, err := event.CopyTo(ctx, i.scratch, dst)
+		if err != nil {
+			return 0, err
+		}
+
+		writeLen += n
+		dst = dst.DropFirst64(n)
+	}
+	return writeLen, nil
+}
+
+// Fsync implements FileOperations.Fsync.
+func (*Inotify) Fsync(context.Context, *File, int64, int64, SyncType) error {
+	return syserror.EINVAL
+}
+
+// Flush implements FileOperations.Flush.
+func (*Inotify) Flush(context.Context, *File) error {
+	return nil
+}
+
+// ConfigureMMap implements FileOperations.ConfigureMMap.
+func (*Inotify) ConfigureMMap(context.Context, *File, *memmap.MMapOpts) error {
+	return syserror.ENODEV
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (i *Inotify) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch args[1].Int() {
+	case linux.FIONREAD:
+		i.evMu.Lock()
+		defer i.evMu.Unlock()
+		var n uint32
+		for e := i.events.Front(); e != nil; e = e.Next() {
+			event := e.(*Event)
+			n += uint32(event.sizeOf())
+		}
+		var buf [4]byte
+		usermem.ByteOrder.PutUint32(buf[:], n)
+		_, err := io.CopyOut(ctx, args[2].Pointer(), buf[:], usermem.IOOpts{})
+		return 0, err
+
+	default:
+		return 0, syserror.ENOTTY
+	}
+}
+
+func (i *Inotify) queueEvent(ev *Event) {
+	i.evMu.Lock()
+	defer i.evMu.Unlock()
+
+	// Check if we should coalesce the event we're about to queue with the last
+	// one currently in the queue. Events are coalesced if they are identical.
+	if last := i.events.Back(); last != nil {
+		if ev.equals(last.(*Event)) {
+			// "Coalesce" the two events by simply not queuing the new one. We
+			// don't need to raise a waiter.EventIn notification because no new
+			// data is available for reading.
+			return
+		}
+	}
+
+	i.events.PushBack(ev)
+	i.Queue.Notify(waiter.EventIn)
+}
+
+// newWatchLocked creates and adds a new watch to target.
+func (i *Inotify) newWatchLocked(target *Dirent, mask uint32) *Watch {
+	wd := i.nextWatch
+	i.nextWatch++
+
+	watch := &Watch{
+		owner:  i,
+		wd:     wd,
+		mask:   mask,
+		target: target.Inode,
+		pins:   make(map[*Dirent]bool),
+	}
+
+	i.watches[wd] = watch
+
+	// Grab an extra reference to target to prevent it from being evicted from
+	// memory. This ref is dropped during either watch removal, target
+	// destruction, or inotify instance destruction. See callers of Watch.Unpin.
+	watch.Pin(target)
+	target.Inode.Watches.Add(watch)
+
+	return watch
+}
+
+// targetDestroyed is called by w to notify i that w's target is gone. This
+// automatically generates a watch removal event.
+func (i *Inotify) targetDestroyed(w *Watch) {
+	i.mu.Lock()
+	_, found := i.watches[w.wd]
+	delete(i.watches, w.wd)
+	i.mu.Unlock()
+
+	if found {
+		i.queueEvent(newEvent(w.wd, "", linux.IN_IGNORED, 0))
+	}
+}
+
+// AddWatch constructs a new inotify watch and adds it to the target dirent. It
+// returns the watch descriptor returned by inotify_add_watch(2).
+func (i *Inotify) AddWatch(target *Dirent, mask uint32) int32 {
+	// Note: Locking this inotify instance protects the result returned by
+	// Lookup() below. With the lock held, we know for sure the lookup result
+	// won't become stale because it's impossible for *this* instance to
+	// add/remove watches on target.
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
+	// Does the target already have a watch from this inotify instance?
+	if existing := target.Inode.Watches.Lookup(i.id); existing != nil {
+		// This may be a watch on a different dirent pointing to the
+		// same inode. Obtain an extra reference if necessary.
+		existing.Pin(target)
+
+		if mergeMask := mask&linux.IN_MASK_ADD != 0; mergeMask {
+			// "Add (OR) events to watch mask for this pathname if it already
+			// exists (instead of replacing mask)." -- inotify(7)
+			existing.mask |= mask
+		} else {
+			existing.mask = mask
+		}
+		return existing.wd
+	}
+
+	// No existing watch, create a new watch.
+	watch := i.newWatchLocked(target, mask)
+	return watch.wd
+}
+
+// RmWatch implements watcher.Watchable.RmWatch.
+//
+// RmWatch looks up an inotify watch for the given 'wd' and configures the
+// target dirent to stop sending events to this inotify instance.
+func (i *Inotify) RmWatch(wd int32) error {
+	i.mu.Lock()
+
+	// Find the watch we were asked to removed.
+	watch, ok := i.watches[wd]
+	if !ok {
+		i.mu.Unlock()
+		return syserror.EINVAL
+	}
+
+	// Remove the watch from this instance.
+	delete(i.watches, wd)
+
+	// Remove the watch from the watch target.
+	watch.target.Watches.Remove(watch.ID())
+
+	// The watch is now isolated and we can safely drop the instance lock. We
+	// need to do so because watch.destroy() acquires Watch.mu, which cannot be
+	// aquired with Inotify.mu held.
+	i.mu.Unlock()
+
+	// Generate the event for the removal.
+	i.queueEvent(newEvent(watch.wd, "", linux.IN_IGNORED, 0))
+
+	// Remove all pins.
+	watch.destroy()
+
+	return nil
+}
diff --git a/pkg/sentry/fs/inotify_event.go b/pkg/sentry/fs/inotify_event.go
new file mode 100644
index 000000000..217915ba4
--- /dev/null
+++ b/pkg/sentry/fs/inotify_event.go
@@ -0,0 +1,138 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// inotifyEventBaseSize is the base size of linux's struct inotify_event. This
+// must be a power 2 for rounding below.
+const inotifyEventBaseSize = 16
+
+// Event represents a struct inotify_event from linux.
+type Event struct {
+	ilist.Entry
+
+	wd     int32
+	mask   uint32
+	cookie uint32
+
+	// len is computed based on the name field is set automatically by
+	// Event.setName. It should be 0 when no name is set; otherwise it is the
+	// length of the name slice.
+	len uint32
+
+	// The name field has special padding requirements and should only be set by
+	// calling Event.setName.
+	name []byte
+}
+
+func newEvent(wd int32, name string, events, cookie uint32) *Event {
+	e := &Event{
+		wd:     wd,
+		mask:   events,
+		cookie: cookie,
+	}
+	if name != "" {
+		e.setName(name)
+	}
+	return e
+}
+
+// paddedBytes converts a go string to a null-terminated c-string, padded with
+// null bytes to a total size of 'l'. 'l' must be large enough for all the bytes
+// in the 's' plus at least one null byte.
+func paddedBytes(s string, l uint32) []byte {
+	if l < uint32(len(s)+1) {
+		panic("Converting string to byte array results in truncation, this can lead to buffer-overflow due to the missing null-byte!")
+	}
+	b := make([]byte, l)
+	copy(b, s)
+
+	// b was zero-value initialized during make(), so the rest of the slice is
+	// already filled with null bytes.
+
+	return b
+}
+
+// setName sets the optional name for this event.
+func (e *Event) setName(name string) {
+	// We need to pad the name such that the entire event length ends up a
+	// multiple of inotifyEventBaseSize.
+	unpaddedLen := len(name) + 1
+	// Round up to nearest multiple of inotifyEventBaseSize.
+	e.len = uint32((unpaddedLen + inotifyEventBaseSize - 1) & ^(inotifyEventBaseSize - 1))
+	// Make sure we haven't overflowed and wrapped around when rounding.
+	if unpaddedLen > int(e.len) {
+		panic("Overflow when rounding inotify event size, the 'name' field was too big.")
+	}
+	e.name = paddedBytes(name, e.len)
+}
+
+func (e *Event) sizeOf() int {
+	s := inotifyEventBaseSize + int(e.len)
+	if s < inotifyEventBaseSize {
+		panic("overflow")
+	}
+	return s
+}
+
+// CopyTo serializes this event to dst. buf is used as a scratch buffer to
+// construct the output. We use a buffer allocated ahead of time for
+// performance. buf must be at least inotifyEventBaseSize bytes.
+func (e *Event) CopyTo(ctx context.Context, buf []byte, dst usermem.IOSequence) (int64, error) {
+	usermem.ByteOrder.PutUint32(buf[0:], uint32(e.wd))
+	usermem.ByteOrder.PutUint32(buf[4:], e.mask)
+	usermem.ByteOrder.PutUint32(buf[8:], e.cookie)
+	usermem.ByteOrder.PutUint32(buf[12:], e.len)
+
+	writeLen := 0
+
+	n, err := dst.CopyOut(ctx, buf)
+	if err != nil {
+		return 0, err
+	}
+	writeLen += n
+	dst = dst.DropFirst(n)
+
+	if e.len > 0 {
+		n, err = dst.CopyOut(ctx, e.name)
+		if err != nil {
+			return 0, err
+		}
+		writeLen += n
+	}
+
+	// Santiy check.
+	if writeLen != e.sizeOf() {
+		panic(fmt.Sprintf("Serialized unexpected amount of data for an event, expected %v, wrote %v.", e.sizeOf(), writeLen))
+	}
+
+	return int64(writeLen), nil
+}
+
+func (e *Event) equals(other *Event) bool {
+	return e.wd == other.wd &&
+		e.mask == other.mask &&
+		e.cookie == other.cookie &&
+		e.len == other.len &&
+		bytes.Equal(e.name, other.name)
+}
diff --git a/pkg/sentry/fs/inotify_watch.go b/pkg/sentry/fs/inotify_watch.go
new file mode 100644
index 000000000..ff6ec6e3e
--- /dev/null
+++ b/pkg/sentry/fs/inotify_watch.go
@@ -0,0 +1,129 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// Watch represent a particular inotify watch created by inotify_add_watch.
+//
+// While a watch is active, it ensures the target inode is pinned in memory by
+// holding an extra ref on each dirent known (by inotify) to point to the
+// inode. These are known as pins. For a full discussion, see
+// fs/g3doc/inotify.md.
+type Watch struct {
+	// Inotify instance which owns this watch.
+	owner *Inotify
+
+	// Descriptor for this watch. This is unique across an inotify instance.
+	wd int32
+
+	// Events being monitored via this watch.
+	mask uint32
+
+	// The inode being watched. Note that we don't directly hold a reference on
+	// this inode. Instead we hold a reference on the dirent(s) containing the
+	// inode, which we record in pins.
+	target *Inode
+
+	// unpinned indicates whether we have a hard reference on target. This field
+	// may only be modified through atomic ops.
+	unpinned uint32
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// pins is the set of dirents this watch is currently pinning in memory by
+	// holding a reference to them. See Pin()/Unpin().
+	pins map[*Dirent]bool
+}
+
+// ID returns the id of the inotify instance that owns this watch.
+func (w *Watch) ID() uint64 {
+	return w.owner.id
+}
+
+// NotifyParentAfterUnlink indicates whether the parent of the watched object
+// should continue to be be notified of events after the target has been
+// unlinked.
+func (w *Watch) NotifyParentAfterUnlink() bool {
+	return w.mask&linux.IN_EXCL_UNLINK == 0
+}
+
+// isRenameEvent returns true if eventMask describes a rename event.
+func isRenameEvent(eventMask uint32) bool {
+	return eventMask&(linux.IN_MOVED_FROM|linux.IN_MOVED_TO|linux.IN_MOVE_SELF) != 0
+}
+
+// Notify queues a new event on this watch.
+func (w *Watch) Notify(name string, events uint32, cookie uint32) {
+	unmaskableBits := ^uint32(0) &^ linux.IN_ALL_EVENTS
+	effectiveMask := unmaskableBits | w.mask
+	matchedEvents := effectiveMask & events
+
+	if matchedEvents == 0 {
+		// We weren't watching for this event.
+		return
+	}
+
+	w.owner.queueEvent(newEvent(w.wd, name, matchedEvents, cookie))
+}
+
+// Pin acquires a new ref on dirent, which pins the dirent in memory while
+// the watch is active. Calling Pin for a second time on the same dirent for
+// the same watch is a no-op.
+func (w *Watch) Pin(d *Dirent) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if !w.pins[d] {
+		w.pins[d] = true
+		d.IncRef()
+	}
+}
+
+// Unpin drops any extra refs held on dirent due to a previous Pin
+// call. Calling Unpin multiple times for the same dirent, or on a dirent
+// without a corresponding Pin call is a no-op.
+func (w *Watch) Unpin(d *Dirent) {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.pins[d] {
+		delete(w.pins, d)
+		d.DecRef()
+	}
+}
+
+// TargetDestroyed notifies the owner of the watch that the watch target is
+// gone. The owner should release its own references to the watcher upon
+// receiving this notification.
+func (w *Watch) TargetDestroyed() {
+	w.owner.targetDestroyed(w)
+}
+
+// destroy prepares the watch for destruction. It unpins all dirents pinned by
+// this watch. Destroy does not cause any new events to be generated. The caller
+// is responsible for ensuring there are no outstanding references to this
+// watch.
+func (w *Watch) destroy() {
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	for d := range w.pins {
+		d.DecRef()
+	}
+	w.pins = nil
+}
diff --git a/pkg/sentry/fs/lock/BUILD b/pkg/sentry/fs/lock/BUILD
new file mode 100644
index 000000000..c15dde800
--- /dev/null
+++ b/pkg/sentry/fs/lock/BUILD
@@ -0,0 +1,72 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "lock_state",
+    srcs = [
+        "lock.go",
+        "lock_range.go",
+        "lock_set.go",
+    ],
+    out = "lock_state.go",
+    package = "lock",
+)
+
+go_template_instance(
+    name = "lock_range",
+    out = "lock_range.go",
+    package = "lock",
+    prefix = "Lock",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "uint64",
+    },
+)
+
+go_template_instance(
+    name = "lock_set",
+    out = "lock_set.go",
+    consts = {
+        "minDegree": "3",
+    },
+    package = "lock",
+    prefix = "Lock",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "LockRange",
+        "Value": "Lock",
+        "Functions": "lockSetFunctions",
+    },
+)
+
+go_library(
+    name = "lock",
+    srcs = [
+        "lock.go",
+        "lock_range.go",
+        "lock_set.go",
+        "lock_set_functions.go",
+        "lock_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/lock",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/log",
+        "//pkg/state",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "lock_test",
+    size = "small",
+    srcs = [
+        "lock_range_test.go",
+        "lock_test.go",
+    ],
+    embed = [":lock"],
+)
diff --git a/pkg/sentry/fs/lock/lock.go b/pkg/sentry/fs/lock/lock.go
new file mode 100644
index 000000000..24d54c989
--- /dev/null
+++ b/pkg/sentry/fs/lock/lock.go
@@ -0,0 +1,457 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package lock is the API for POSIX-style advisory regional file locks and
+// BSD-style full file locks.
+//
+// Callers needing to enforce these types of locks, like sys_fcntl, can call
+// LockRegion and UnlockRegion on a thread-safe set of Locks.  Locks are
+// specific to a unique file (unique device/inode pair) and for this reason
+// should not be shared between files.
+//
+// A Lock has a set of holders identified by UniqueID.  Normally this is the
+// pid of the thread attempting to acquire the lock.
+//
+// Since these are advisory locks, they do not need to be integrated into
+// Reads/Writes and for this reason there is no way to *check* if a lock is
+// held.  One can only attempt to take a lock or unlock an existing lock.
+//
+// A Lock in a set of Locks is typed: it is either a read lock with any number
+// of readers and no writer, or a write lock with no readers.
+//
+// As expected from POSIX, any attempt to acquire a write lock on a file region
+// when there already exits a write lock held by a different uid will fail. Any
+// attempt to acquire a write lock on a file region when there is more than one
+// reader will fail.  Any attempt to acquire a read lock on a file region when
+// there is already a writer will fail.
+//
+// In special cases, a read lock may be upgraded to a write lock and a write lock
+// can be downgraded to a read lock.  This can only happen if:
+//
+//  * read lock upgrade to write lock: There can be only one reader and the reader
+//    must be the same as the requested write lock holder.
+//
+//  * write lock downgrade to read lock: The writer must be the same as the requested
+//    read lock holder.
+//
+// UnlockRegion always succeeds.  If LockRegion fails the caller should normally
+// interpret this as "try again later".
+package lock
+
+import (
+	"fmt"
+	"math"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// LockType is a type of regional file lock.
+type LockType int
+
+// UniqueID is a unique identifier of the holder of a regional file lock.
+type UniqueID uint64
+
+const (
+	// ReadLock describes a POSIX regional file lock to be taken
+	// read only.  There may be multiple of these locks on a single
+	// file region as long as there is no writer lock on the same
+	// region.
+	ReadLock LockType = iota
+
+	// WriteLock describes a POSIX regional file lock to be taken
+	// write only.  There may be only a single holder of this lock
+	// and no read locks.
+	WriteLock
+)
+
+// LockEOF is the maximal possible end of a regional file lock.
+const LockEOF = math.MaxUint64
+
+// Lock is a regional file lock.  It consists of either a single writer
+// or a set of readers.
+//
+// A Lock may be upgraded from a read lock to a write lock only if there
+// is a single reader and that reader has the same uid as the write lock.
+//
+// A Lock may be downgraded from a write lock to a read lock only if
+// the write lock's uid is the same as the read lock.
+type Lock struct {
+	// Readers are the set of read lock holders identified by UniqueID.
+	// If len(Readers) > 0 then HasWriter must be false.
+	Readers map[UniqueID]bool
+
+	// HasWriter indicates that this is a write lock held by a single
+	// UniqueID.
+	HasWriter bool
+
+	// Writer is only valid if HasWriter is true.  It identifies a
+	// single write lock holder.
+	Writer UniqueID
+}
+
+// Locks is a thread-safe wrapper around a LockSet.
+type Locks struct {
+	// mu protects locks below.
+	mu sync.Mutex `state:"nosave"`
+
+	// locks is the set of region locks currently held on an Inode.
+	locks LockSet
+
+	// blockedQueue is the queue of waiters that are waiting on a lock.
+	blockedQueue waiter.Queue
+}
+
+// Blocker is the interface used for blocking locks. Passing a nil Blocker
+// will be treated as non-blocking.
+type Blocker interface {
+	Block(C chan struct{}) error
+}
+
+const (
+	// EventMaskAll is the mask we will always use for locks, by using the
+	// same mask all the time we can wake up everyone anytime the lock
+	// changes state.
+	EventMaskAll waiter.EventMask = 0xFFFF
+)
+
+// LockRegion attempts to acquire a typed lock for the uid on a region
+// of a file. Returns true if successful in locking the region. If false
+// is returned, the caller should normally interpret this as "try again later" if
+// accquiring the lock in a non-blocking mode or "interrupted" if in a blocking mode.
+// Blocker is the interface used to provide blocking behavior, passing a nil Blocker
+// will result in non-blocking behavior.
+func (l *Locks) LockRegion(uid UniqueID, t LockType, r LockRange, block Blocker) bool {
+	for {
+		l.mu.Lock()
+
+		// Blocking locks must run in a loop because we'll be woken up whenever an unlock event
+		// happens for this lock. We will then attempt to take the lock again and if it fails
+		// continue blocking.
+		res := l.locks.lock(uid, t, r)
+		if !res && block != nil {
+			e, ch := waiter.NewChannelEntry(nil)
+			l.blockedQueue.EventRegister(&e, EventMaskAll)
+			l.mu.Unlock()
+			if err := block.Block(ch); err != nil {
+				// We were interrupted, the caller can translate this to EINTR if applicable.
+				l.blockedQueue.EventUnregister(&e)
+				return false
+			}
+			l.blockedQueue.EventUnregister(&e)
+			continue // Try again now that someone has unlocked.
+		}
+
+		l.mu.Unlock()
+		return res
+	}
+}
+
+// UnlockRegion attempts to release a lock for the uid on a region of a file.
+// This operation is always successful, even if there did not exist a lock on
+// the requested region held by uid in the first place.
+func (l *Locks) UnlockRegion(uid UniqueID, r LockRange) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.locks.unlock(uid, r)
+
+	// Now that we've released the lock, we need to wake up any waiters.
+	l.blockedQueue.Notify(EventMaskAll)
+}
+
+// makeLock returns a new typed Lock that has either uid as its only reader
+// or uid as its only writer.
+func makeLock(uid UniqueID, t LockType) Lock {
+	value := Lock{Readers: make(map[UniqueID]bool)}
+	switch t {
+	case ReadLock:
+		value.Readers[uid] = true
+	case WriteLock:
+		value.HasWriter = true
+		value.Writer = uid
+	default:
+		panic(fmt.Sprintf("makeLock: invalid lock type %d", t))
+	}
+	return value
+}
+
+// isHeld returns true if uid is a holder of Lock.
+func (l Lock) isHeld(uid UniqueID) bool {
+	if l.HasWriter && l.Writer == uid {
+		return true
+	}
+	return l.Readers[uid]
+}
+
+// lock sets uid as a holder of a typed lock on Lock.
+//
+// Preconditions: canLock is true for the range containing this Lock.
+func (l *Lock) lock(uid UniqueID, t LockType) {
+	switch t {
+	case ReadLock:
+		// If we are already a reader, then this is a no-op.
+		if l.Readers[uid] {
+			return
+		}
+		// We cannot downgrade a write lock to a read lock unless the
+		// uid is the same.
+		if l.HasWriter {
+			if l.Writer != uid {
+				panic(fmt.Sprintf("lock: cannot downgrade write lock to read lock for uid %d, writer is %d", uid, l.Writer))
+			}
+			// Ensure that there is only one reader if upgrading.
+			l.Readers = make(map[UniqueID]bool)
+			// Ensure that there is no longer a writer.
+			l.HasWriter = false
+		}
+		l.Readers[uid] = true
+		return
+	case WriteLock:
+		// If we are already the writer, then this is a no-op.
+		if l.HasWriter && l.Writer == uid {
+			return
+		}
+		// We can only upgrade a read lock to a write lock if there
+		// is only one reader and that reader has the same uid as
+		// the write lock.
+		if readers := len(l.Readers); readers > 0 {
+			if readers != 1 {
+				panic(fmt.Sprintf("lock: cannot upgrade read lock to write lock for uid %d, too many readers %v", uid, l.Readers))
+			}
+			if !l.Readers[uid] {
+				panic(fmt.Sprintf("lock: cannot upgrade read lock to write lock for uid %d, conflicting reader %v", uid, l.Readers))
+			}
+		}
+		// Ensure that there is only a writer.
+		l.Readers = make(map[UniqueID]bool)
+		l.HasWriter = true
+		l.Writer = uid
+	default:
+		panic(fmt.Sprintf("lock: invalid lock type %d", t))
+	}
+}
+
+// lockable returns true if check returns true for every Lock in LockRange.
+// Further, check should return true if Lock meets the callers requirements
+// for locking Lock.
+func (l LockSet) lockable(r LockRange, check func(value Lock) bool) bool {
+	// Get our starting point.
+	seg := l.LowerBoundSegment(r.Start)
+	for seg.Ok() && seg.Start() < r.End {
+		// Note that we don't care about overruning the end of the
+		// last segment because if everything checks out we'll just
+		// split the last segment.
+		if !check(seg.Value()) {
+			return false
+		}
+		// Jump to the next segment, ignoring gaps, for the same
+		// reason we ignored the first gap.
+		seg = seg.NextSegment()
+	}
+	// No conflict, we can get a lock for uid over the entire range.
+	return true
+}
+
+// canLock returns true if uid will be able to take a Lock of type t on the
+// entire range specified by LockRange.
+func (l LockSet) canLock(uid UniqueID, t LockType, r LockRange) bool {
+	switch t {
+	case ReadLock:
+		return l.lockable(r, func(value Lock) bool {
+			// If there is no writer, there's no problem adding
+			// another reader.
+			if !value.HasWriter {
+				return true
+			}
+			// If there is a writer, then it must be the same uid
+			// in order to downgrade the lock to a read lock.
+			return value.Writer == uid
+		})
+	case WriteLock:
+		return l.lockable(r, func(value Lock) bool {
+			// If there are only readers.
+			if !value.HasWriter {
+				// Then this uid can only take a write lock if
+				// this is a private upgrade, meaning that the
+				// only reader is uid.
+				return len(value.Readers) == 1 && value.Readers[uid]
+			}
+			// If the uid is already a writer on this region, then
+			// adding a write lock would be a no-op.
+			return value.Writer == uid
+		})
+	default:
+		panic(fmt.Sprintf("canLock: invalid lock type %d", t))
+	}
+}
+
+// lock returns true if uid took a lock of type t on the entire range of LockRange.
+//
+// Preconditions: r.Start <= r.End (will panic otherwise).
+func (l *LockSet) lock(uid UniqueID, t LockType, r LockRange) bool {
+	if r.Start > r.End {
+		panic(fmt.Sprintf("lock: r.Start %d > r.End %d", r.Start, r.End))
+	}
+
+	// Don't attempt to insert anything with a range of 0 and treat this
+	// as a successful no-op.
+	if r.Length() == 0 {
+		return true
+	}
+
+	// Do a first-pass check.  We *could* hold onto the segments we
+	// checked if canLock would return true, but traversing the segment
+	// set should be fast and this keeps things simple.
+	if !l.canLock(uid, t, r) {
+		return false
+	}
+	// Get our starting point.
+	seg, gap := l.Find(r.Start)
+	if gap.Ok() {
+		// Fill in the gap and get the next segment to modify.
+		seg = l.Insert(gap, gap.Range().Intersect(r), makeLock(uid, t)).NextSegment()
+	} else if seg.Start() < r.Start {
+		// Get our first segment to modify.
+		_, seg = l.Split(seg, r.Start)
+	}
+	for seg.Ok() && seg.Start() < r.End {
+		// Split the last one if necessary.
+		if seg.End() > r.End {
+			seg, _ = l.SplitUnchecked(seg, r.End)
+		}
+
+		// Set the lock on the segment.  This is guaranteed to
+		// always be safe, given canLock above.
+		value := seg.ValuePtr()
+		value.lock(uid, t)
+
+		// Fill subsequent gaps.
+		gap = seg.NextGap()
+		if gr := gap.Range().Intersect(r); gr.Length() > 0 {
+			seg = l.Insert(gap, gr, makeLock(uid, t)).NextSegment()
+		} else {
+			seg = gap.NextSegment()
+		}
+	}
+	return true
+}
+
+// unlock is always successful.  If uid has no locks held for the range LockRange,
+// unlock is a no-op.
+//
+// Preconditions: same as lock.
+func (l *LockSet) unlock(uid UniqueID, r LockRange) {
+	if r.Start > r.End {
+		panic(fmt.Sprintf("unlock: r.Start %d > r.End %d", r.Start, r.End))
+	}
+
+	// Same as setlock.
+	if r.Length() == 0 {
+		return
+	}
+
+	// Get our starting point.
+	seg := l.LowerBoundSegment(r.Start)
+	for seg.Ok() && seg.Start() < r.End {
+		// If this segment doesn't have a lock from uid then
+		// there is no need to fragment the set with Isolate (below).
+		// In this case just move on to the next segment.
+		if !seg.Value().isHeld(uid) {
+			seg = seg.NextSegment()
+			continue
+		}
+
+		// Ensure that if we need to unlock a sub-segment that
+		// we don't unlock/remove that entire segment.
+		seg = l.Isolate(seg, r)
+
+		value := seg.Value()
+		var remove bool
+		if value.HasWriter && value.Writer == uid {
+			// If we are unlocking a writer, then since there can
+			// only ever be one writer and no readers, then this
+			// lock should always be removed from the set.
+			remove = true
+		} else if value.Readers[uid] {
+			// If uid is the last reader, then just remove the entire
+			// segment.
+			if len(value.Readers) == 1 {
+				remove = true
+			} else {
+				// Otherwise we need to remove this reader without
+				// affecting any other segment's readers.  To do
+				// this, we need to make a copy of the Readers map
+				// and not add this uid.
+				newValue := Lock{Readers: make(map[UniqueID]bool)}
+				for k, v := range value.Readers {
+					if k != uid {
+						newValue.Readers[k] = v
+					}
+				}
+				seg.SetValue(newValue)
+			}
+		}
+		if remove {
+			seg = l.Remove(seg).NextSegment()
+		} else {
+			seg = seg.NextSegment()
+		}
+	}
+}
+
+// ComputeRange takes a positive file offset and computes the start of a LockRange
+// using start (relative to offset) and the end of the LockRange using length. The
+// values of start and length may be negative but the resulting LockRange must
+// preserve that LockRange.Start < LockRange.End and LockRange.Start > 0.
+func ComputeRange(start, length, offset int64) (LockRange, error) {
+	offset += start
+	// fcntl(2): "l_start can be a negative number provided the offset
+	// does not lie before the start of the file"
+	if offset < 0 {
+		return LockRange{}, syscall.EINVAL
+	}
+
+	// fcntl(2): Specifying 0 for l_len has the  special meaning: lock all
+	// bytes starting at the location specified by l_whence and l_start
+	// through to the end of file, no matter how large the file grows.
+	end := uint64(LockEOF)
+	if length > 0 {
+		// fcntl(2): If l_len is positive, then the range to be locked
+		// covers bytes l_start up to and including l_start+l_len-1.
+		//
+		// Since LockRange.End is exclusive we need not -1 from length..
+		end = uint64(offset + length)
+	} else if length < 0 {
+		// fcntl(2): If l_len is negative, the interval described by
+		// lock covers bytes l_start+l_len up to and including l_start-1.
+		//
+		// Since LockRange.End is exclusive we need not -1 from offset.
+		signedEnd := offset
+		// Add to offset using a negative length (subtract).
+		offset += length
+		if offset < 0 {
+			return LockRange{}, syscall.EINVAL
+		}
+		if signedEnd < offset {
+			return LockRange{}, syscall.EOVERFLOW
+		}
+		// At this point signedEnd cannot be negative,
+		// since we asserted that offset is not negative
+		// and it is not less than offset.
+		end = uint64(signedEnd)
+	}
+	// Offset is guaranteed to be positive at this point.
+	return LockRange{Start: uint64(offset), End: end}, nil
+}
diff --git a/pkg/sentry/fs/lock/lock_range_test.go b/pkg/sentry/fs/lock/lock_range_test.go
new file mode 100644
index 000000000..06a37c701
--- /dev/null
+++ b/pkg/sentry/fs/lock/lock_range_test.go
@@ -0,0 +1,136 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package lock
+
+import (
+	"syscall"
+	"testing"
+)
+
+func TestComputeRange(t *testing.T) {
+	tests := []struct {
+		// Description of test.
+		name string
+
+		// Requested start of the lock range.
+		start int64
+
+		// Requested length of the lock range,
+		// can be negative :(
+		length int64
+
+		// Pre-computed file offset based on whence.
+		// Will be added to start.
+		offset int64
+
+		// Expected error.
+		err error
+
+		// If error is nil, the expected LockRange.
+		LockRange
+	}{
+		{
+			name:      "offset, start, and length all zero",
+			LockRange: LockRange{Start: 0, End: LockEOF},
+		},
+		{
+			name:      "zero offset, zero start, positive length",
+			start:     0,
+			length:    4096,
+			offset:    0,
+			LockRange: LockRange{Start: 0, End: 4096},
+		},
+		{
+			name:   "zero offset, negative start",
+			start:  -4096,
+			offset: 0,
+			err:    syscall.EINVAL,
+		},
+		{
+			name:      "large offset, negative start, positive length",
+			start:     -2048,
+			length:    2048,
+			offset:    4096,
+			LockRange: LockRange{Start: 2048, End: 4096},
+		},
+		{
+			name:      "large offset, negative start, zero length",
+			start:     -2048,
+			length:    0,
+			offset:    4096,
+			LockRange: LockRange{Start: 2048, End: LockEOF},
+		},
+		{
+			name:   "zero offset, zero start, negative length",
+			start:  0,
+			length: -4096,
+			offset: 0,
+			err:    syscall.EINVAL,
+		},
+		{
+			name:      "large offset, zero start, negative length",
+			start:     0,
+			length:    -4096,
+			offset:    4096,
+			LockRange: LockRange{Start: 0, End: 4096},
+		},
+		{
+			name:      "offset, start, and length equal, length is negative",
+			start:     1024,
+			length:    -1024,
+			offset:    1024,
+			LockRange: LockRange{Start: 1024, End: 2048},
+		},
+		{
+			name:      "offset, start, and length equal, start is negative",
+			start:     -1024,
+			length:    1024,
+			offset:    1024,
+			LockRange: LockRange{Start: 0, End: 1024},
+		},
+		{
+			name:      "offset, start, and length equal, offset is negative",
+			start:     1024,
+			length:    1024,
+			offset:    -1024,
+			LockRange: LockRange{Start: 0, End: 1024},
+		},
+		{
+			name:   "offset, start, and length equal, all negative",
+			start:  -1024,
+			length: -1024,
+			offset: -1024,
+			err:    syscall.EINVAL,
+		},
+		{
+			name:      "offset, start, and length equal, all positive",
+			start:     1024,
+			length:    1024,
+			offset:    1024,
+			LockRange: LockRange{Start: 2048, End: 3072},
+		},
+	}
+
+	for _, test := range tests {
+		rng, err := ComputeRange(test.start, test.length, test.offset)
+		if err != test.err {
+			t.Errorf("%s: lockRange(%d, %d, %d) got error %v, want %v", test.name, test.start, test.length, test.offset, err, test.err)
+			continue
+		}
+		if err == nil && rng != test.LockRange {
+			t.Errorf("%s: lockRange(%d, %d, %d) got LockRange %v, want %v", test.name, test.start, test.length, test.offset, rng, test.LockRange)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/lock/lock_set_functions.go b/pkg/sentry/fs/lock/lock_set_functions.go
new file mode 100644
index 000000000..e16f485be
--- /dev/null
+++ b/pkg/sentry/fs/lock/lock_set_functions.go
@@ -0,0 +1,69 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package lock
+
+import (
+	"math"
+)
+
+// LockSet maps a set of Locks into a file.  The key is the file offset.
+
+type lockSetFunctions struct{}
+
+func (lockSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+func (lockSetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+func (lockSetFunctions) ClearValue(l *Lock) {
+	*l = Lock{}
+}
+
+func (lockSetFunctions) Merge(r1 LockRange, val1 Lock, r2 LockRange, val2 Lock) (Lock, bool) {
+	// Merge only if the Readers/Writers are identical.
+	if len(val1.Readers) != len(val2.Readers) {
+		return Lock{}, false
+	}
+	for k := range val1.Readers {
+		if !val2.Readers[k] {
+			return Lock{}, false
+		}
+	}
+	if val1.HasWriter != val2.HasWriter {
+		return Lock{}, false
+	}
+	if val1.HasWriter {
+		if val1.Writer != val2.Writer {
+			return Lock{}, false
+		}
+	}
+	return val1, true
+}
+
+func (lockSetFunctions) Split(r LockRange, val Lock, split uint64) (Lock, Lock) {
+	// Copy the segment so that split segments don't contain map references
+	// to other segments.
+	val0 := Lock{Readers: make(map[UniqueID]bool)}
+	for k, v := range val.Readers {
+		val0.Readers[k] = v
+	}
+	val0.HasWriter = val.HasWriter
+	val0.Writer = val.Writer
+
+	return val, val0
+}
diff --git a/pkg/sentry/fs/lock/lock_test.go b/pkg/sentry/fs/lock/lock_test.go
new file mode 100644
index 000000000..c60f5f7a2
--- /dev/null
+++ b/pkg/sentry/fs/lock/lock_test.go
@@ -0,0 +1,1059 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package lock
+
+import (
+	"reflect"
+	"testing"
+)
+
+type entry struct {
+	Lock
+	LockRange
+}
+
+func equals(e0, e1 []entry) bool {
+	if len(e0) != len(e1) {
+		return false
+	}
+	for i := range e0 {
+		for k := range e0[i].Lock.Readers {
+			if !e1[i].Lock.Readers[k] {
+				return false
+			}
+		}
+		for k := range e1[i].Lock.Readers {
+			if !e0[i].Lock.Readers[k] {
+				return false
+			}
+		}
+		if !reflect.DeepEqual(e0[i].LockRange, e1[i].LockRange) {
+			return false
+		}
+		if e0[i].Lock.HasWriter != e1[i].Lock.HasWriter {
+			return false
+		}
+		if e0[i].Lock.Writer != e1[i].Lock.Writer {
+			return false
+		}
+	}
+	return true
+}
+
+// fill a LockSet with consecutive region locks.  Will panic if
+// LockRanges are not consecutive.
+func fill(entries []entry) LockSet {
+	l := LockSet{}
+	for _, e := range entries {
+		gap := l.FindGap(e.LockRange.Start)
+		if !gap.Ok() {
+			panic("cannot insert into existing segment")
+		}
+		l.Insert(gap, e.LockRange, e.Lock)
+	}
+	return l
+}
+
+func TestCanLockEmpty(t *testing.T) {
+	l := LockSet{}
+
+	// Expect to be able to take any locks given that the set is empty.
+	eof := l.FirstGap().End()
+	r := LockRange{0, eof}
+	if !l.canLock(1, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", ReadLock, r, 1)
+	}
+	if !l.canLock(2, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", ReadLock, r, 2)
+	}
+	if !l.canLock(1, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 1)
+	}
+	if !l.canLock(2, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 2)
+	}
+}
+
+func TestCanLock(t *testing.T) {
+	// + -------------- + ---------- + -------------- + --------- +
+	// | Readers 1 & 2  | Readers 1  | Readers 1 & 3  | Writer 1  |
+	// + -------------  + ---------- + -------------- + --------- +
+	// 0             1024         2048             3072        4096
+	l := fill([]entry{
+		{
+			Lock:      Lock{Readers: map[UniqueID]bool{1: true, 2: true}},
+			LockRange: LockRange{0, 1024},
+		},
+		{
+			Lock:      Lock{Readers: map[UniqueID]bool{1: true}},
+			LockRange: LockRange{1024, 2048},
+		},
+		{
+			Lock:      Lock{Readers: map[UniqueID]bool{1: true, 3: true}},
+			LockRange: LockRange{2048, 3072},
+		},
+		{
+			Lock:      Lock{HasWriter: true, Writer: 1},
+			LockRange: LockRange{3072, 4096},
+		},
+	})
+
+	// Now that we have a mildly interesting layout, try some checks on different
+	// ranges, uids, and lock types.
+	//
+	// Expect to be able to extend the read lock, despite the writer lock, because
+	// the writer has the same uid as the requested read lock.
+	r := LockRange{0, 8192}
+	if !l.canLock(1, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", ReadLock, r, 1)
+	}
+	// Expect to *not* be able to extend the read lock since there is an overlapping
+	// writer region locked by someone other than the uid.
+	if l.canLock(2, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got true, want false", ReadLock, r, 2)
+	}
+	// Expect to be able to extend the read lock if there are only other readers in
+	// the way.
+	r = LockRange{64, 3072}
+	if !l.canLock(2, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", ReadLock, r, 2)
+	}
+	// Expect to be able to set a read lock beyond the range of any existing locks.
+	r = LockRange{4096, 10240}
+	if !l.canLock(2, ReadLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", ReadLock, r, 2)
+	}
+
+	// Expect to not be able to take a write lock with other readers in the way.
+	r = LockRange{0, 8192}
+	if l.canLock(1, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got true, want false", WriteLock, r, 1)
+	}
+	// Expect to be able to extend the write lock for the same uid.
+	r = LockRange{3072, 8192}
+	if !l.canLock(1, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 1)
+	}
+	// Expect to not be able to overlap a write lock for two different uids.
+	if l.canLock(2, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got true, want false", WriteLock, r, 2)
+	}
+	// Expect to be able to set a write lock that is beyond the range of any
+	// existing locks.
+	r = LockRange{8192, 10240}
+	if !l.canLock(2, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 2)
+	}
+	// Expect to be able to upgrade a read lock (any portion of it).
+	r = LockRange{1024, 2048}
+	if !l.canLock(1, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 1)
+	}
+	r = LockRange{1080, 2000}
+	if !l.canLock(1, WriteLock, r) {
+		t.Fatalf("canLock type %d for range %v and uid %d got false, want true", WriteLock, r, 1)
+	}
+}
+
+func TestSetLock(t *testing.T) {
+	tests := []struct {
+		// description of test.
+		name string
+
+		// LockSet entries to pre-fill.
+		before []entry
+
+		// Description of region to lock:
+		//
+		// start is the file offset of the lock.
+		start uint64
+		// end is the end file offset of the lock.
+		end uint64
+		// uid of lock attempter.
+		uid UniqueID
+		// lock type requested.
+		lockType LockType
+
+		// success is true if taking the above
+		// lock should succeed.
+		success bool
+
+		// Expected layout of the set after locking
+		// if success is true.
+		after []entry
+	}{
+		{
+			name:     "set zero length ReadLock on empty set",
+			start:    0,
+			end:      0,
+			uid:      0,
+			lockType: ReadLock,
+			success:  true,
+		},
+		{
+			name:     "set zero length WriteLock on empty set",
+			start:    0,
+			end:      0,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+		},
+		{
+			name:     "set ReadLock on empty set",
+			start:    0,
+			end:      LockEOF,
+			uid:      0,
+			lockType: ReadLock,
+			success:  true,
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+		},
+		{
+			name:     "set WriteLock on empty set",
+			start:    0,
+			end:      LockEOF,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+			// + ----------------------------------------- +
+			// | Writer  0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+		},
+		{
+			name: "set ReadLock on WriteLock same uid",
+			// + ----------------------------------------- +
+			// | Writer 0                                  |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      0,
+			lockType: ReadLock,
+			success:  true,
+			// + ----------- + --------------------------- +
+			// | Readers 0   | Writer 0                    |
+			// + ----------- + --------------------------- +
+			// 0          4096                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, 4096},
+				},
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "set WriteLock on ReadLock same uid",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+			// + ----------- + --------------------------- +
+			// | Writer 0    | Readers 0                   |
+			// + ----------- + --------------------------- +
+			// 0          4096                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "set ReadLock on WriteLock different uid",
+			// + ----------------------------------------- +
+			// | Writer 0                                  |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      1,
+			lockType: ReadLock,
+			success:  false,
+		},
+		{
+			name: "set WriteLock on ReadLock different uid",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      1,
+			lockType: WriteLock,
+			success:  false,
+		},
+		{
+			name: "split ReadLock for overlapping lock at start 0",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      1,
+			lockType: ReadLock,
+			success:  true,
+			// + -------------- + --------------------------- +
+			// | Readers 0 & 1  | Readers 0                   |
+			// + -------------- + --------------------------- +
+			// 0             4096                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "split ReadLock for overlapping lock at non-zero start",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start:    4096,
+			end:      8192,
+			uid:      1,
+			lockType: ReadLock,
+			success:  true,
+			// + ---------- + -------------- + ----------- +
+			// | Readers 0  | Readers 0 & 1  | Readers 0   |
+			// + ---------- + -------------- + ----------- +
+			// 0         4096             8192    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{4096, 8192},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{8192, LockEOF},
+				},
+			},
+		},
+		{
+			name: "fill front gap with ReadLock",
+			// + --------- + ---------------------------- +
+			// | gap       | Readers 0                    |
+			// + --------- + ---------------------------- +
+			// 0        1024                     max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, LockEOF},
+				},
+			},
+			start:    0,
+			end:      8192,
+			uid:      0,
+			lockType: ReadLock,
+			success:  true,
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+		},
+		{
+			name: "fill end gap with ReadLock",
+			// + ---------------------------- +
+			// | Readers 0                    |
+			// + ---------------------------- +
+			// 0                           4096
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, 4096},
+				},
+			},
+			start:    1024,
+			end:      LockEOF,
+			uid:      0,
+			lockType: ReadLock,
+			success:  true,
+			// Note that this is not merged after lock does a Split.  This is
+			// fine because the two locks will still *behave* as one.  In other
+			// words we can fragment any lock all we want and semantically it
+			// makes no difference.
+			//
+			// + ----------- + --------------------------- +
+			// | Readers 0   | Readers 0                   |
+			// + ----------- + --------------------------- +
+			// 0                                  max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, LockEOF},
+				},
+			},
+		},
+		{
+			name: "fill gap with ReadLock and split",
+			// + --------- + ---------------------------- +
+			// | gap       | Readers 0                    |
+			// + --------- + ---------------------------- +
+			// 0        1024                     max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, LockEOF},
+				},
+			},
+			start:    0,
+			end:      4096,
+			uid:      1,
+			lockType: ReadLock,
+			success:  true,
+			// + --------- + ------------- + ------------- +
+			// | Reader 1  | Readers 0 & 1 | Reader 0      |
+			// + ----------+ ------------- + ------------- +
+			// 0        1024            4096      max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{1024, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "upgrade ReadLock to WriteLock for single uid fill gap",
+			// + ------------- + --------- + --- + ------------- +
+			// | Readers 0 & 1 | Readers 0 | gap | Readers 0 & 2 |
+			// + ------------- + --------- + --- + ------------- +
+			// 0            1024        2048  4096      max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, 2048},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start:    1024,
+			end:      4096,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+			// + ------------- + -------- + ------------- +
+			// | Readers 0 & 1 | Writer 0 | Readers 0 & 2 |
+			// + ------------- + -------- + ------------- +
+			// 0            1024       4096      max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{1024, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "upgrade ReadLock to WriteLock for single uid keep gap",
+			// + ------------- + --------- + --- + ------------- +
+			// | Readers 0 & 1 | Readers 0 | gap | Readers 0 & 2 |
+			// + ------------- + --------- + --- + ------------- +
+			// 0            1024        2048  4096      max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, 2048},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start:    1024,
+			end:      3072,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+			// + ------------- + -------- + --- + ------------- +
+			// | Readers 0 & 1 | Writer 0 | gap | Readers 0 & 2 |
+			// + ------------- + -------- + --- + ------------- +
+			// 0            1024       3072  4096      max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{1024, 3072},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "fail to upgrade ReadLock to WriteLock with conflicting Reader",
+			// + ------------- + --------- +
+			// | Readers 0 & 1 | Readers 0 |
+			// + ------------- + --------- +
+			// 0            1024        2048
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, 2048},
+				},
+			},
+			start:    0,
+			end:      2048,
+			uid:      0,
+			lockType: WriteLock,
+			success:  false,
+		},
+		{
+			name: "take WriteLock on whole file if all uids are the same",
+			// + ------------- + --------- + --------- + ---------- +
+			// | Writer 0      | Readers 0 | Readers 0 | Readers 0  |
+			// + ------------- + --------- + --------- + ---------- +
+			// 0            1024        2048        4096   max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{1024, 2048},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{2048, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start:    0,
+			end:      LockEOF,
+			uid:      0,
+			lockType: WriteLock,
+			success:  true,
+			// We do not manually merge locks.  Semantically a fragmented lock
+			// held by the same uid will behave as one lock so it makes no difference.
+			//
+			// + ------------- + ---------------------------- +
+			// | Writer 0      | Writer 0                     |
+			// + ------------- + ---------------------------- +
+			// 0            1024                     max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{1024, LockEOF},
+				},
+			},
+		},
+	}
+
+	for _, test := range tests {
+		l := fill(test.before)
+
+		r := LockRange{Start: test.start, End: test.end}
+		success := l.lock(test.uid, test.lockType, r)
+		var got []entry
+		for seg := l.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+			got = append(got, entry{
+				Lock:      seg.Value(),
+				LockRange: seg.Range(),
+			})
+		}
+
+		if success != test.success {
+			t.Errorf("%s: setlock(%v, %+v, %d, %d) got success %v, want %v", test.name, test.before, r, test.uid, test.lockType, success, test.success)
+			continue
+		}
+
+		if success {
+			if !equals(got, test.after) {
+				t.Errorf("%s: got set %+v, want %+v", test.name, got, test.after)
+			}
+		}
+	}
+}
+
+func TestUnlock(t *testing.T) {
+	tests := []struct {
+		// description of test.
+		name string
+
+		// LockSet entries to pre-fill.
+		before []entry
+
+		// Description of region to unlock:
+		//
+		// start is the file start of the lock.
+		start uint64
+		// end is the end file start of the lock.
+		end uint64
+		// uid of lock holder.
+		uid UniqueID
+
+		// Expected layout of the set after unlocking.
+		after []entry
+	}{
+		{
+			name:  "unlock zero length on empty set",
+			start: 0,
+			end:   0,
+			uid:   0,
+		},
+		{
+			name:  "unlock on empty set (no-op)",
+			start: 0,
+			end:   LockEOF,
+			uid:   0,
+		},
+		{
+			name: "unlock uid not locked (no-op)",
+			// + --------------------------- +
+			// | Readers 1 & 2               |
+			// + --------------------------- +
+			// 0                    max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true, 2: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 1024,
+			end:   4096,
+			uid:   0,
+			// + --------------------------- +
+			// | Readers 1 & 2               |
+			// + --------------------------- +
+			// 0                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true, 2: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock ReadLock over entire file",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 0,
+			end:   LockEOF,
+			uid:   0,
+		},
+		{
+			name: "unlock WriteLock over entire file",
+			// + ----------------------------------------- +
+			// | Writer 0                                  |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 0,
+			end:   LockEOF,
+			uid:   0,
+		},
+		{
+			name: "unlock partial ReadLock (start)",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 0,
+			end:   4096,
+			uid:   0,
+			// + ------ + --------------------------- +
+			// | gap    | Readers 0                   |
+			// +------- + --------------------------- +
+			// 0     4096                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock partial WriteLock (start)",
+			// + ----------------------------------------- +
+			// | Writer  0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 0,
+			end:   4096,
+			uid:   0,
+			// + ------ + --------------------------- +
+			// | gap    | Writer  0                   |
+			// +------- + --------------------------- +
+			// 0     4096                    max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock partial ReadLock (end)",
+			// + ----------------------------------------- +
+			// | Readers 0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 4096,
+			end:   LockEOF,
+			uid:   0,
+			// + --------------------------- +
+			// | Readers 0                   |
+			// +---------------------------- +
+			// 0                          4096
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true}},
+					LockRange: LockRange{0, 4096},
+				},
+			},
+		},
+		{
+			name: "unlock partial WriteLock (end)",
+			// + ----------------------------------------- +
+			// | Writer  0                                 |
+			// + ----------------------------------------- +
+			// 0                                  max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 4096,
+			end:   LockEOF,
+			uid:   0,
+			// + --------------------------- +
+			// | Writer  0                   |
+			// +---------------------------- +
+			// 0                          4096
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 4096},
+				},
+			},
+		},
+		{
+			name: "unlock for single uid",
+			// + ------------- + --------- + ------------------- +
+			// | Readers 0 & 1 | Writer 0  | Readers 0 & 1 & 2   |
+			// + ------------- + --------- + ------------------- +
+			// 0            1024        4096            max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{1024, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start: 0,
+			end:   LockEOF,
+			uid:   0,
+			// + --------- + --- + --------------- +
+			// | Readers 1 | gap | Readers 1 & 2   |
+			// + --------- + --- + --------------- +
+			// 0        1024  4096        max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock subsection locked",
+			// + ------------------------------- +
+			// | Readers 0 & 1 & 2               |
+			// + ------------------------------- +
+			// 0                        max uint64
+			before: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true, 2: true}},
+					LockRange: LockRange{0, LockEOF},
+				},
+			},
+			start: 1024,
+			end:   4096,
+			uid:   0,
+			// + ----------------- + ------------- + ----------------- +
+			// | Readers 0 & 1 & 2 | Readers 1 & 2 | Readers 0 & 1 & 2 |
+			// + ----------------- + ------------- + ----------------- +
+			// 0                1024            4096          max uint64
+			after: []entry{
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true, 2: true}},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true, 2: true}},
+					LockRange: LockRange{1024, 4096},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true, 2: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock mid-gap to increase gap",
+			// + --------- + ----- + ------------------- +
+			// | Writer 0  |  gap  | Readers 0 & 1       |
+			// + --------- + ----- + ------------------- +
+			// 0        1024    4096            max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start: 8,
+			end:   2048,
+			uid:   0,
+			// + --------- + ----- + ------------------- +
+			// | Writer 0  |  gap  | Readers 0 & 1       |
+			// + --------- + ----- + ------------------- +
+			// 0           8    4096            max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 8},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+		},
+		{
+			name: "unlock split region on uid mid-gap",
+			// + --------- + ----- + ------------------- +
+			// | Writer 0  |  gap  | Readers 0 & 1       |
+			// + --------- + ----- + ------------------- +
+			// 0        1024    4096            max uint64
+			before: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{4096, LockEOF},
+				},
+			},
+			start: 2048,
+			end:   8192,
+			uid:   0,
+			// + --------- + ----- + --------- + ------------- +
+			// | Writer 0  |  gap  | Readers 1 | Readers 0 & 1 |
+			// + --------- + ----- + --------- + ------------- +
+			// 0       1024     4096        8192      max uint64
+			after: []entry{
+				{
+					Lock:      Lock{HasWriter: true, Writer: 0},
+					LockRange: LockRange{0, 1024},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{1: true}},
+					LockRange: LockRange{4096, 8192},
+				},
+				{
+					Lock:      Lock{Readers: map[UniqueID]bool{0: true, 1: true}},
+					LockRange: LockRange{8192, LockEOF},
+				},
+			},
+		},
+	}
+
+	for _, test := range tests {
+		l := fill(test.before)
+
+		r := LockRange{Start: test.start, End: test.end}
+		l.unlock(test.uid, r)
+		var got []entry
+		for seg := l.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+			got = append(got, entry{
+				Lock:      seg.Value(),
+				LockRange: seg.Range(),
+			})
+		}
+		if !equals(got, test.after) {
+			t.Errorf("%s: got set %+v, want %+v", test.name, got, test.after)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/mock.go b/pkg/sentry/fs/mock.go
new file mode 100644
index 000000000..b3bfa5268
--- /dev/null
+++ b/pkg/sentry/fs/mock.go
@@ -0,0 +1,177 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// MockInodeOperations implements InodeOperations for testing Inodes.
+type MockInodeOperations struct {
+	InodeOperations
+
+	UAttr UnstableAttr
+
+	createCalled          bool
+	createDirectoryCalled bool
+	createLinkCalled      bool
+	renameCalled          bool
+	walkCalled            bool
+}
+
+// NewMockInode returns a mock *Inode using MockInodeOperations.
+func NewMockInode(ctx context.Context, msrc *MountSource, sattr StableAttr) *Inode {
+	return NewInode(NewMockInodeOperations(ctx), msrc, sattr)
+}
+
+// NewMockInodeOperations returns a *MockInodeOperations.
+func NewMockInodeOperations(ctx context.Context) *MockInodeOperations {
+	return &MockInodeOperations{
+		UAttr: WithCurrentTime(ctx, UnstableAttr{
+			Perms: FilePermsFromMode(0777),
+		}),
+	}
+}
+
+// MockMountSourceOps implements fs.MountSourceOperations.
+type MockMountSourceOps struct {
+	MountSourceOperations
+	keep       bool
+	revalidate bool
+}
+
+// NewMockMountSource returns a new *MountSource using MockMountSourceOps.
+func NewMockMountSource(cache *DirentCache) *MountSource {
+	var keep bool
+	if cache != nil {
+		keep = cache.maxSize > 0
+	}
+	return &MountSource{
+		MountSourceOperations: &MockMountSourceOps{keep: keep},
+		fscache:               cache,
+		children:              make(map[*MountSource]struct{}),
+	}
+}
+
+// Revalidate implements fs.MountSourceOperations.Revalidate.
+func (n *MockMountSourceOps) Revalidate(*Dirent) bool {
+	return n.revalidate
+}
+
+// Keep implements fs.MountSourceOperations.Keep.
+func (n *MockMountSourceOps) Keep(dirent *Dirent) bool {
+	return n.keep
+}
+
+// WriteOut implements fs.InodeOperations.WriteOut.
+func (n *MockInodeOperations) WriteOut(context.Context, *Inode) error {
+	return nil
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (n *MockInodeOperations) UnstableAttr(context.Context, *Inode) (UnstableAttr, error) {
+	return n.UAttr, nil
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+func (n *MockInodeOperations) IsVirtual() bool {
+	return false
+}
+
+// Lookup implements fs.InodeOperations.Lookup.
+func (n *MockInodeOperations) Lookup(ctx context.Context, dir *Inode, p string) (*Dirent, error) {
+	n.walkCalled = true
+	return NewDirent(NewInode(&MockInodeOperations{}, dir.MountSource, StableAttr{}), p), nil
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (n *MockInodeOperations) SetPermissions(context.Context, *Inode, FilePermissions) bool {
+	return false
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (*MockInodeOperations) SetOwner(context.Context, *Inode, FileOwner) error {
+	return syserror.EINVAL
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (n *MockInodeOperations) SetTimestamps(context.Context, *Inode, TimeSpec) error {
+	return nil
+}
+
+// Create implements fs.InodeOperations.Create.
+func (n *MockInodeOperations) Create(ctx context.Context, dir *Inode, p string, flags FileFlags, perms FilePermissions) (*File, error) {
+	n.createCalled = true
+	d := NewDirent(NewInode(&MockInodeOperations{}, dir.MountSource, StableAttr{}), p)
+	return &File{Dirent: d}, nil
+}
+
+// CreateLink implements fs.InodeOperations.CreateLink.
+func (n *MockInodeOperations) CreateLink(_ context.Context, dir *Inode, oldname string, newname string) error {
+	n.createLinkCalled = true
+	return nil
+}
+
+// CreateDirectory implements fs.InodeOperations.CreateDirectory.
+func (n *MockInodeOperations) CreateDirectory(context.Context, *Inode, string, FilePermissions) error {
+	n.createDirectoryCalled = true
+	return nil
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (n *MockInodeOperations) Rename(ctx context.Context, oldParent *Inode, oldName string, newParent *Inode, newName string) error {
+	n.renameCalled = true
+	return nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (n *MockInodeOperations) Check(ctx context.Context, inode *Inode, p PermMask) bool {
+	return ContextCanAccessFile(ctx, inode, p)
+}
+
+// Release implements fs.InodeOperations.Release.
+func (n *MockInodeOperations) Release(context.Context) {}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (n *MockInodeOperations) Truncate(ctx context.Context, inode *Inode, size int64) error {
+	return nil
+}
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev.
+func (n *MockInodeOperations) DeprecatedPwritev(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, nil
+}
+
+// DeprecatedReaddir implements fs.InodeOperations.DeprecatedReaddir.
+func (n *MockInodeOperations) DeprecatedReaddir(context.Context, *DirCtx, int) (int, error) {
+	return 0, nil
+}
+
+// Remove implements fs.InodeOperations.Remove.
+func (n *MockInodeOperations) Remove(context.Context, *Inode, string) error {
+	return nil
+}
+
+// RemoveDirectory implements fs.InodeOperations.RemoveDirectory.
+func (n *MockInodeOperations) RemoveDirectory(context.Context, *Inode, string) error {
+	return nil
+}
+
+// Getlink implements fs.InodeOperations.Getlink.
+func (n *MockInodeOperations) Getlink(context.Context, *Inode) (*Dirent, error) {
+	return nil, syserror.ENOLINK
+}
diff --git a/pkg/sentry/fs/mount.go b/pkg/sentry/fs/mount.go
new file mode 100644
index 000000000..a2943b097
--- /dev/null
+++ b/pkg/sentry/fs/mount.go
@@ -0,0 +1,298 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"bytes"
+	"fmt"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+)
+
+// DirentOperations provide file systems greater control over how long a Dirent stays pinned
+// in core. Implementations must not take Dirent.mu.
+type DirentOperations interface {
+	// Revalidate returns true if the Dirent is stale and its InodeOperations needs to be reloaded. Revalidate
+	// will never be called on a Dirent that is mounted.
+	Revalidate(dirent *Dirent) bool
+
+	// Keep returns true if the Dirent should be kept in memory for as long as possible
+	// beyond any active references.
+	Keep(dirent *Dirent) bool
+}
+
+// MountSourceOperations contains filesystem specific operations.
+type MountSourceOperations interface {
+	// TODO: Add:
+	//
+	// StatFS() (Info, error)
+	// BlockSize() int64
+	// FS() Filesystem
+
+	// DirentOperations provide optional extra management of Dirents.
+	DirentOperations
+
+	// Destroy destroys the MountSource.
+	Destroy()
+
+	// Below are MountSourceOperations that do not conform to Linux.
+
+	// ResetInodeMappings clears all mappings of Inodes before SaveInodeMapping
+	// is called.
+	ResetInodeMappings()
+
+	// SaveInodeMappings is called during saving to store, for each reachable
+	// Inode in the mounted filesystem, a mapping of Inode.StableAttr.InodeID
+	// to the Inode's path relative to its mount point. If an Inode is
+	// reachable at more than one path due to hard links, it is unspecified
+	// which path is mapped. Filesystems that do not use this information to
+	// restore inodes can make SaveInodeMappings a no-op.
+	SaveInodeMapping(inode *Inode, path string)
+}
+
+// InodeMappings defines a fmt.Stringer MountSource Inode mappings.
+type InodeMappings map[uint64]string
+
+// String implements fmt.Stringer.String.
+func (i InodeMappings) String() string {
+	var mappingsBuf bytes.Buffer
+	mappingsBuf.WriteString("\n")
+	for ino, name := range i {
+		mappingsBuf.WriteString(fmt.Sprintf("\t%q\t\tinode number %d\n", name, ino))
+	}
+	return mappingsBuf.String()
+}
+
+// MountSource represents a source of file objects.
+//
+// MountSource corresponds to struct super_block in Linux.
+//
+// A mount source may represent a physical device (or a partition of a physical
+// device) or a virtual source of files such as procfs for a specific PID
+// namespace. There should be only one mount source per logical device. E.g.
+// there should be only procfs mount source for a given PID namespace.
+//
+// A mount source represents files as inodes. Every inode belongs to exactly
+// one mount source. Each file object may only be represented using one inode
+// object in a sentry instance.
+//
+// This is an amalgamation of structs super_block, vfsmount, and mount, while
+// MountSourceOperations is akin to struct super_operations.
+//
+// Hence, mount source also contains common mounted file system state, such as
+// mount flags, the root Dirent, and children mounts. For now, this
+// amalgamation implies that a mount source cannot be shared by multiple mounts
+// (e.g. cannot be mounted at different locations).
+//
+// TODO: Move mount-specific information out of MountSource.
+type MountSource struct {
+	refs.AtomicRefCount
+
+	// MountSourceOperations defines filesystem specific behavior.
+	MountSourceOperations
+
+	// Filesystem is the filesystem backing the mount. Can be nil if there
+	// is no filesystem backing the mount.
+	Filesystem Filesystem
+
+	// Flags are the flags that this filesystem was mounted with.
+	Flags MountSourceFlags
+
+	// fscache keeps Dirents pinned beyond application references to them.
+	// It must be flushed before kernel.SaveTo.
+	fscache *DirentCache `state:"nosave"`
+
+	// direntRefs is the sum of references on all Dirents in this MountSource.
+	//
+	// direntRefs is increased when a Dirent in MountSource is IncRef'd, and
+	// decreased when a Dirent in MountSource is DecRef'd.
+	//
+	// To cleanly unmount a MountSource, one must check that no direntRefs are
+	// held anymore. To check, one must hold root.parent.dirMu of the
+	// MountSource's root Dirent before reading direntRefs to prevent further
+	// walks to Dirents in this MountSource.
+	//
+	// direntRefs must be atomically changed.
+	direntRefs uint64
+
+	// mu protects the fields below, which are set by the MountNamespace
+	// during MountSource/Unmount.
+	mu sync.Mutex `state:"nosave"`
+
+	// id is a unique id for this mount.
+	id uint64
+
+	// root is the root Dirent of this mount.
+	root *Dirent
+
+	// parent is the parent MountSource, or nil if this MountSource is the root.
+	parent *MountSource
+
+	// children are the child MountSources of this MountSource.
+	children map[*MountSource]struct{}
+}
+
+// defaultDirentCacheSize is the number of Dirents that the VFS can hold an extra
+// reference on.
+const defaultDirentCacheSize uint64 = 1000
+
+// NewMountSource returns a new MountSource. Filesystem may be nil if there is no
+// filesystem backing the mount.
+func NewMountSource(mops MountSourceOperations, filesystem Filesystem, flags MountSourceFlags) *MountSource {
+	return &MountSource{
+		MountSourceOperations: mops,
+		Flags:      flags,
+		Filesystem: filesystem,
+		fscache:    NewDirentCache(defaultDirentCacheSize),
+		children:   make(map[*MountSource]struct{}),
+	}
+}
+
+// Parent returns the parent mount, or nil if this mount is the root.
+func (msrc *MountSource) Parent() *MountSource {
+	msrc.mu.Lock()
+	defer msrc.mu.Unlock()
+	return msrc.parent
+}
+
+// ID returns the ID of this mount.
+func (msrc *MountSource) ID() uint64 {
+	msrc.mu.Lock()
+	defer msrc.mu.Unlock()
+	return msrc.id
+}
+
+// Children returns the (immediate) children of this MountSource.
+func (msrc *MountSource) Children() []*MountSource {
+	msrc.mu.Lock()
+	defer msrc.mu.Unlock()
+
+	ms := make([]*MountSource, 0, len(msrc.children))
+	for c := range msrc.children {
+		ms = append(ms, c)
+	}
+	return ms
+}
+
+// Submounts returns all mounts that are descendants of this mount.
+func (msrc *MountSource) Submounts() []*MountSource {
+	var ms []*MountSource
+	for _, c := range msrc.Children() {
+		ms = append(ms, c)
+		ms = append(ms, c.Submounts()...)
+	}
+	return ms
+}
+
+// Root returns the root dirent of this mount.
+func (msrc *MountSource) Root() *Dirent {
+	msrc.mu.Lock()
+	defer msrc.mu.Unlock()
+	return msrc.root
+}
+
+// DirentRefs returns the current mount direntRefs.
+func (msrc *MountSource) DirentRefs() uint64 {
+	return atomic.LoadUint64(&msrc.direntRefs)
+}
+
+// IncDirentRefs increases direntRefs.
+func (msrc *MountSource) IncDirentRefs() {
+	atomic.AddUint64(&msrc.direntRefs, 1)
+}
+
+// DecDirentRefs decrements direntRefs.
+func (msrc *MountSource) DecDirentRefs() {
+	if atomic.AddUint64(&msrc.direntRefs, ^uint64(0)) == ^uint64(0) {
+		panic("Decremented zero mount reference direntRefs")
+	}
+}
+
+func (msrc *MountSource) destroy() {
+	if c := msrc.DirentRefs(); c != 0 {
+		panic(fmt.Sprintf("MountSource with non-zero direntRefs is being destroyed: %d", c))
+	}
+	msrc.MountSourceOperations.Destroy()
+}
+
+// DecRef drops a reference on the MountSource.
+func (msrc *MountSource) DecRef() {
+	msrc.DecRefWithDestructor(msrc.destroy)
+}
+
+// FlushDirentRefs drops all references held by the MountSource on Dirents.
+func (msrc *MountSource) FlushDirentRefs() {
+	msrc.fscache.Invalidate()
+}
+
+// NewCachingMountSource returns a generic mount that will cache dirents
+// aggressively. Filesystem may be nil if there is no backing filesystem.
+func NewCachingMountSource(filesystem Filesystem, flags MountSourceFlags) *MountSource {
+	return NewMountSource(&SimpleMountSourceOperations{
+		keep: true,
+	}, filesystem, flags)
+}
+
+// NewNonCachingMountSource returns a generic mount that will never cache dirents.
+// Filesystem may be nil if there is no backing filesystem.
+func NewNonCachingMountSource(filesystem Filesystem, flags MountSourceFlags) *MountSource {
+	return NewMountSource(&SimpleMountSourceOperations{
+		keep: false,
+	}, filesystem, flags)
+}
+
+// SimpleMountSourceOperations implements MountSourceOperations.
+type SimpleMountSourceOperations struct {
+	keep bool
+}
+
+// Revalidate implements MountSourceOperations.Revalidate.
+func (*SimpleMountSourceOperations) Revalidate(*Dirent) bool {
+	return false
+}
+
+// Keep implements MountSourceOperations.Keep.
+func (smo *SimpleMountSourceOperations) Keep(*Dirent) bool {
+	return smo.keep
+}
+
+// ResetInodeMappings implements MountSourceOperations.ResetInodeMappings.
+func (*SimpleMountSourceOperations) ResetInodeMappings() {}
+
+// SaveInodeMapping implements MountSourceOperations.SaveInodeMapping.
+func (*SimpleMountSourceOperations) SaveInodeMapping(*Inode, string) {}
+
+// Destroy implements MountSourceOperations.Destroy.
+func (*SimpleMountSourceOperations) Destroy() {}
+
+// Info defines attributes of a filesystem.
+type Info struct {
+	// Type is the filesystem type magic value.
+	Type uint64
+
+	// TotalBlocks is the total data blocks in the filesystem.
+	TotalBlocks uint64
+
+	// FreeBlocks is the number of free blocks available.
+	FreeBlocks uint64
+
+	// TotalFiles is the total file nodes in the filesystem.
+	TotalFiles uint64
+
+	// FreeFiles is the number of free file nodes.
+	FreeFiles uint64
+}
diff --git a/pkg/sentry/fs/mount_overlay.go b/pkg/sentry/fs/mount_overlay.go
new file mode 100644
index 000000000..16c25e46c
--- /dev/null
+++ b/pkg/sentry/fs/mount_overlay.go
@@ -0,0 +1,95 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/context"
+
+// overlayMountSourceOperations implements MountSourceOperations for an overlay
+// mount point.
+type overlayMountSourceOperations struct {
+	upper *MountSource
+	lower *MountSource
+}
+
+func newOverlayMountSource(upper, lower *MountSource, flags MountSourceFlags) *MountSource {
+	upper.IncRef()
+	lower.IncRef()
+	return NewMountSource(&overlayMountSourceOperations{
+		upper: upper,
+		lower: lower,
+	}, &overlayFilesystem{}, flags)
+}
+
+// Revalidate panics if the upper or lower MountSource require that dirent be
+// revalidated. Otherwise always returns false.
+func (o *overlayMountSourceOperations) Revalidate(dirent *Dirent) bool {
+	if o.upper.Revalidate(dirent) || o.lower.Revalidate(dirent) {
+		panic("an overlay cannot revalidate file objects")
+	}
+	return false
+}
+
+// Keep returns true if either upper or lower MountSource require that the
+// dirent be kept in memory.
+func (o *overlayMountSourceOperations) Keep(dirent *Dirent) bool {
+	return o.upper.Keep(dirent) || o.lower.Keep(dirent)
+}
+
+// ResetInodeMappings propagates the call to both upper and lower MountSource.
+func (o *overlayMountSourceOperations) ResetInodeMappings() {
+	o.upper.ResetInodeMappings()
+	o.lower.ResetInodeMappings()
+}
+
+// SaveInodeMapping propagates the call to both upper and lower MountSource.
+func (o *overlayMountSourceOperations) SaveInodeMapping(inode *Inode, path string) {
+	inode.overlay.copyMu.RLock()
+	defer inode.overlay.copyMu.RUnlock()
+	if inode.overlay.upper != nil {
+		o.upper.SaveInodeMapping(inode.overlay.upper, path)
+	}
+	if inode.overlay.lower != nil {
+		o.lower.SaveInodeMapping(inode.overlay.lower, path)
+	}
+}
+
+// Destroy drops references on the upper and lower MountSource.
+func (o *overlayMountSourceOperations) Destroy() {
+	o.upper.DecRef()
+	o.lower.DecRef()
+}
+
+// type overlayFilesystem is the filesystem for overlay mounts.
+type overlayFilesystem struct{}
+
+// Name implements Filesystem.Name.
+func (ofs *overlayFilesystem) Name() string {
+	return "overlayfs"
+}
+
+// Flags implements Filesystem.Flags.
+func (ofs *overlayFilesystem) Flags() FilesystemFlags {
+	return 0
+}
+
+// AllowUserMount implements Filesystem.AllowUserMount.
+func (ofs *overlayFilesystem) AllowUserMount() bool {
+	return false
+}
+
+// Mount implements Filesystem.Mount.
+func (ofs *overlayFilesystem) Mount(ctx context.Context, device string, flags MountSourceFlags, data string) (*Inode, error) {
+	panic("overlayFilesystem.Mount should not be called!")
+}
diff --git a/pkg/sentry/fs/mount_state.go b/pkg/sentry/fs/mount_state.go
new file mode 100644
index 000000000..f5ed1dd8d
--- /dev/null
+++ b/pkg/sentry/fs/mount_state.go
@@ -0,0 +1,25 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// afterLoad is invoked by stateify.
+//
+// Beyond the cache, this method's existence is required to ensure that this
+// object is not marked "complete" until all dependent objects are also marked
+// "complete". Implementations (e.g. see gofer_state.go) reach into the
+// MountSourceOperations through this object, this is necessary on restore.
+func (msrc *MountSource) afterLoad() {
+	msrc.fscache = NewDirentCache(defaultDirentCacheSize)
+}
diff --git a/pkg/sentry/fs/mount_test.go b/pkg/sentry/fs/mount_test.go
new file mode 100644
index 000000000..3a053c154
--- /dev/null
+++ b/pkg/sentry/fs/mount_test.go
@@ -0,0 +1,216 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+)
+
+// cacheReallyContains iterates through the dirent cache to determine whether
+// it contains the given dirent.
+func cacheReallyContains(cache *DirentCache, d *Dirent) bool {
+	for i := cache.list.Front(); i != nil; i = i.Next() {
+		if i == d {
+			return true
+		}
+	}
+	return false
+}
+
+// TestMountSourceOnlyCachedOnce tests that a Dirent that is mounted over only ends
+// up in a single Dirent Cache. NOTE: Having a dirent in multiple
+// caches causes major consistency issues.
+func TestMountSourceOnlyCachedOnce(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	rootCache := NewDirentCache(100)
+	rootInode := NewMockInode(ctx, NewMockMountSource(rootCache), StableAttr{
+		Type: Directory,
+	})
+	mm, err := NewMountNamespace(ctx, rootInode)
+	if err != nil {
+		t.Fatalf("NewMountNamespace failed: %v", err)
+	}
+	rootDirent := mm.Root()
+	defer rootDirent.DecRef()
+
+	// Get a child of the root which we will mount over.  Note that the
+	// MockInodeOperations causes Walk to always succeed.
+	child, err := rootDirent.Walk(ctx, rootDirent, "child")
+	if err != nil {
+		t.Fatalf("failed to walk to child dirent: %v", err)
+	}
+	child.maybeExtendReference() // Cache.
+
+	// Ensure that the root cache contains the child.
+	if !cacheReallyContains(rootCache, child) {
+		t.Errorf("wanted rootCache to contain child dirent, but it did not")
+	}
+
+	// Create a new cache and inode, and mount it over child.
+	submountCache := NewDirentCache(100)
+	submountInode := NewMockInode(ctx, NewMockMountSource(submountCache), StableAttr{
+		Type: Directory,
+	})
+	if err := mm.Mount(ctx, child, submountInode); err != nil {
+		t.Fatalf("failed to mount over child: %v", err)
+	}
+
+	// Walk to the child again.
+	child2, err := rootDirent.Walk(ctx, rootDirent, "child")
+	if err != nil {
+		t.Fatalf("failed to walk to child dirent: %v", err)
+	}
+
+	// Should have a different Dirent than before.
+	if child == child2 {
+		t.Fatalf("expected %v not equal to %v, but they are the same", child, child2)
+	}
+
+	// Neither of the caches should no contain the child.
+	if cacheReallyContains(rootCache, child) {
+		t.Errorf("wanted rootCache not to contain child dirent, but it did")
+	}
+	if cacheReallyContains(submountCache, child) {
+		t.Errorf("wanted submountCache not to contain child dirent, but it did")
+	}
+}
+
+// Test that mounts have proper parent/child relationships.
+func TestMountSourceParentChildRelationship(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	rootCache := NewDirentCache(100)
+	rootInode := NewMockInode(ctx, NewMockMountSource(rootCache), StableAttr{
+		Type: Directory,
+	})
+	mm, err := NewMountNamespace(ctx, rootInode)
+	if err != nil {
+		t.Fatalf("NewMountNamespace failed: %v", err)
+	}
+	rootDirent := mm.Root()
+	defer rootDirent.DecRef()
+
+	// Add mounts at the following paths:
+	paths := []string{
+		"/foo",
+		"/foo/bar",
+		"/foo/bar/baz",
+		"/foo/qux",
+		"/waldo",
+	}
+
+	for _, p := range paths {
+		d, err := mm.FindLink(ctx, rootDirent, nil, p, 0)
+		if err != nil {
+			t.Fatalf("could not find path %q in mount manager: %v", p, err)
+		}
+		submountInode := NewMockInode(ctx, NewMockMountSource(nil), StableAttr{
+			Type: Directory,
+		})
+		if err := mm.Mount(ctx, d, submountInode); err != nil {
+			t.Fatalf("could not mount at %q: %v", p, err)
+		}
+	}
+
+	// mm root should contain all submounts (and does not include the root
+	// mount).
+	allMountSources := rootDirent.Inode.MountSource.Submounts()
+	if err := mountPathsAre(rootDirent, allMountSources, paths...); err != nil {
+		t.Error(err)
+	}
+
+	// Each mount should have a unique ID.
+	foundIDs := make(map[uint64]struct{})
+	for _, m := range allMountSources {
+		id := m.ID()
+		if _, ok := foundIDs[id]; ok {
+			t.Errorf("got multiple mounts with id %d", id)
+		}
+		foundIDs[id] = struct{}{}
+	}
+
+	// Root mount should have no parent.
+	rootMountSource := mm.root.Inode.MountSource
+	if p := rootMountSource.Parent(); p != nil {
+		t.Errorf("root.Parent got %v wanted nil", p)
+	}
+
+	// Root mount should have 2 children: foo and waldo.
+	rootChildren := rootMountSource.Children()
+	if err := mountPathsAre(rootDirent, rootChildren, "/foo", "/waldo"); err != nil {
+		t.Error(err)
+	}
+	// All root mount children should have root as parent.
+	for _, c := range rootChildren {
+		if p := c.Parent(); p != rootMountSource {
+			t.Errorf("root mount child got parent %+v, wanted root mount", p)
+		}
+	}
+
+	// "foo" mount should have two children: /foo/bar, and /foo/qux.
+	d, err := mm.FindLink(ctx, rootDirent, nil, "/foo", 0)
+	if err != nil {
+		t.Fatalf("could not find path %q in mount manager: %v", "/foo", err)
+	}
+	fooMountSource := d.Inode.MountSource
+	fooMountSourceChildren := fooMountSource.Children()
+	if err := mountPathsAre(rootDirent, fooMountSourceChildren, "/foo/bar", "/foo/qux"); err != nil {
+		t.Error(err)
+	}
+	// Each child should have fooMountSource as parent.
+	for _, c := range fooMountSourceChildren {
+		if p := c.Parent(); p != fooMountSource {
+			t.Errorf("foo mount child got parent %+v, wanted foo mount", p)
+		}
+	}
+	// Submounts of foo are /foo/bar, /foo/qux, and /foo/bar/baz.
+	if err := mountPathsAre(rootDirent, fooMountSource.Submounts(), "/foo/bar", "/foo/qux", "/foo/bar/baz"); err != nil {
+		t.Error(err)
+	}
+
+	// "waldo" mount should have no submounts or children.
+	waldo, err := mm.FindLink(ctx, rootDirent, nil, "/waldo", 0)
+	if err != nil {
+		t.Fatalf("could not find path %q in mount manager: %v", "/waldo", err)
+	}
+	waldoMountSource := waldo.Inode.MountSource
+	if got := len(waldoMountSource.Children()); got != 0 {
+		t.Errorf("waldo got %d children, wanted 0", got)
+	}
+	if got := len(waldoMountSource.Submounts()); got != 0 {
+		t.Errorf("waldo got %d children, wanted 0", got)
+	}
+}
+
+func mountPathsAre(root *Dirent, got []*MountSource, want ...string) error {
+	if len(got) != len(want) {
+		return fmt.Errorf("mount paths have different lengths: got %d want %d", len(got), len(want))
+	}
+	gotPaths := make(map[string]struct{}, len(got))
+	for _, g := range got {
+		n, _ := g.Root().FullName(root)
+		gotPaths[n] = struct{}{}
+	}
+	for _, w := range want {
+		if _, ok := gotPaths[w]; !ok {
+			return fmt.Errorf("no mount with path %q found", w)
+		}
+	}
+	return nil
+}
diff --git a/pkg/sentry/fs/mounts.go b/pkg/sentry/fs/mounts.go
new file mode 100644
index 000000000..1e6b5b70e
--- /dev/null
+++ b/pkg/sentry/fs/mounts.go
@@ -0,0 +1,511 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// DefaultTraversalLimit provides a sensible default traversal limit that may
+// be passed to FindInode and FindLink. You may want to provide other options in
+// individual syscall implementations, but for internal functions this will be
+// sane.
+const DefaultTraversalLimit = 10
+
+// MountNamespace defines a collection of mounts.
+type MountNamespace struct {
+	refs.AtomicRefCount
+
+	// userns is the user namespace associated with this mount namespace.
+	//
+	// All privileged operations on this mount namespace must have
+	// appropriate capabilities in this userns.
+	//
+	// userns is immutable.
+	userns *auth.UserNamespace
+
+	// root is the root directory.
+	root *Dirent
+
+	// mu protects mounts and mountID counter.
+	mu sync.Mutex `state:"nosave"`
+
+	// mounts is a map of the last mounted Dirent -> stack of old Dirents
+	// that were mounted over, with the oldest mounted Dirent first and
+	// more recent mounted Dirents at the end of the slice.
+	//
+	// A reference to all Dirents in mounts (keys and values) must be held
+	// to ensure the Dirents are recoverable when unmounting.
+	mounts map[*Dirent][]*Dirent
+
+	// mountID is the next mount id to assign.
+	mountID uint64
+}
+
+// NewMountNamespace returns a new MountNamespace, with the provided node at the
+// root, and the given cache size. A root must always be provided.
+func NewMountNamespace(ctx context.Context, root *Inode) (*MountNamespace, error) {
+	creds := auth.CredentialsFromContext(ctx)
+
+	root.MountSource.mu.Lock()
+	defer root.MountSource.mu.Unlock()
+
+	// Set the root dirent and id on the root mount.
+	d := NewDirent(root, "/")
+	root.MountSource.root = d
+	root.MountSource.id = 1
+
+	return &MountNamespace{
+		userns:  creds.UserNamespace,
+		root:    d,
+		mounts:  make(map[*Dirent][]*Dirent),
+		mountID: 2,
+	}, nil
+}
+
+// UserNamespace returns the user namespace associated with this mount manager.
+func (mns *MountNamespace) UserNamespace() *auth.UserNamespace {
+	return mns.userns
+}
+
+// Root returns the MountNamespace's root Dirent and increments its reference
+// count.  The caller must call DecRef when finished.
+func (mns *MountNamespace) Root() *Dirent {
+	mns.root.IncRef()
+	return mns.root
+}
+
+// FlushMountSourceRefs flushes extra references held by MountSources for all active mount points;
+// see fs/mount.go:MountSource.FlushDirentRefs.
+func (mns *MountNamespace) FlushMountSourceRefs() {
+	mns.mu.Lock()
+	defer mns.mu.Unlock()
+	mns.flushMountSourceRefsLocked()
+}
+
+func (mns *MountNamespace) flushMountSourceRefsLocked() {
+	// Flush mounts' MountSource references.
+	for current, stack := range mns.mounts {
+		current.Inode.MountSource.FlushDirentRefs()
+		for _, prev := range stack {
+			prev.Inode.MountSource.FlushDirentRefs()
+		}
+	}
+
+	// Flush root's MountSource references.
+	mns.root.Inode.MountSource.FlushDirentRefs()
+}
+
+// destroy drops root and mounts dirent references and closes any original nodes.
+//
+// After destroy is called, the MountNamespace may continue to be referenced (for
+// example via /proc/mounts), but should free all resources and shouldn't have
+// Find* methods called.
+func (mns *MountNamespace) destroy() {
+	mns.mu.Lock()
+	defer mns.mu.Unlock()
+
+	// Flush all mounts' MountSource references to Dirents. This allows for mount
+	// points to be torn down since there should be no remaining references after
+	// this and DecRef below.
+	mns.flushMountSourceRefsLocked()
+
+	// Teardown mounts.
+	for current, mp := range mns.mounts {
+		// Drop the mount reference on all mounted dirents.
+		for _, d := range mp {
+			d.DecRef()
+		}
+		current.DecRef()
+	}
+	mns.mounts = nil
+
+	// Drop reference on the root.
+	mns.root.DecRef()
+
+	// Wait for asynchronous work (queued by dropping Dirent references
+	// above) to complete before destroying this MountNamespace.
+	AsyncBarrier()
+}
+
+// DecRef implements RefCounter.DecRef with destructor mns.destroy.
+func (mns *MountNamespace) DecRef() {
+	mns.DecRefWithDestructor(mns.destroy)
+}
+
+// Freeze freezes the entire mount tree.
+func (mns *MountNamespace) Freeze() {
+	mns.mu.Lock()
+	defer mns.mu.Unlock()
+
+	// We only want to freeze Dirents with active references, not Dirents referenced
+	// by a mount's MountSource.
+	mns.flushMountSourceRefsLocked()
+
+	// Freeze the entire shebang.
+	mns.root.Freeze()
+}
+
+// withMountLocked prevents further walks to `node`, because `node` is about to
+// be a mount point.
+func (mns *MountNamespace) withMountLocked(node *Dirent, fn func() error) error {
+	mns.mu.Lock()
+	defer mns.mu.Unlock()
+
+	renameMu.Lock()
+	defer renameMu.Unlock()
+
+	// Linux allows mounting over the root (?). It comes with a strange set
+	// of semantics. We'll just not do this for now.
+	if node.parent == nil {
+		return syserror.EBUSY
+	}
+
+	// For both mount and unmount, we take this lock so we can swap out the
+	// appropriate child in parent.children.
+	//
+	// For unmount, this also ensures that if `node` is a mount point, the
+	// underlying mount's MountSource.direntRefs cannot increase by preventing
+	// walks to node.
+	node.parent.dirMu.Lock()
+	defer node.parent.dirMu.Unlock()
+
+	node.parent.mu.Lock()
+	defer node.parent.mu.Unlock()
+
+	// We need not take node.dirMu since we have parent.dirMu.
+
+	// We need to take node.mu, so that we can check for deletion.
+	node.mu.Lock()
+	defer node.mu.Unlock()
+
+	return fn()
+}
+
+// Mount mounts a `inode` over the subtree at `node`.
+func (mns *MountNamespace) Mount(ctx context.Context, node *Dirent, inode *Inode) error {
+	return mns.withMountLocked(node, func() error {
+		// replacement already has one reference taken; this is the mount
+		// reference.
+		replacement, err := node.mount(ctx, inode)
+		if err != nil {
+			return err
+		}
+
+		// Set child/parent dirent relationship.
+		parentMountSource := node.Inode.MountSource
+		childMountSource := inode.MountSource
+		parentMountSource.mu.Lock()
+		defer parentMountSource.mu.Unlock()
+		childMountSource.mu.Lock()
+		defer childMountSource.mu.Unlock()
+
+		parentMountSource.children[childMountSource] = struct{}{}
+		childMountSource.parent = parentMountSource
+
+		// Set the mount's root dirent and id.
+		childMountSource.root = replacement
+		childMountSource.id = mns.mountID
+		mns.mountID++
+
+		// Drop node from its dirent cache.
+		node.dropExtendedReference()
+
+		// If node is already a mount point, push node on the stack so it can
+		// be recovered on unmount.
+		if stack, ok := mns.mounts[node]; ok {
+			mns.mounts[replacement] = append(stack, node)
+			delete(mns.mounts, node)
+			return nil
+		}
+
+		// Was not already mounted, just add another mount point.
+		// Take a reference on node so it can be recovered on unmount.
+		node.IncRef()
+		mns.mounts[replacement] = []*Dirent{node}
+		return nil
+	})
+}
+
+// Unmount ensures no references to the MountSource remain and removes `node` from
+// this subtree. The subtree formerly mounted in `node`'s place will be
+// restored. node's MountSource will be destroyed as soon as the last reference to
+// `node` is dropped, as no references to Dirents within will remain.
+//
+// If detachOnly is set, Unmount merely removes `node` from the subtree, but
+// allows existing references to the MountSource remain. E.g. if an open file still
+// refers to Dirents in MountSource, the Unmount will succeed anyway and MountSource will
+// be destroyed at a later time when all references to Dirents within are
+// dropped.
+//
+// The caller must hold a reference to node from walking to it.
+func (mns *MountNamespace) Unmount(ctx context.Context, node *Dirent, detachOnly bool) error {
+	// This takes locks to prevent further walks to Dirents in this mount
+	// under the assumption that `node` is the root of the mount.
+	return mns.withMountLocked(node, func() error {
+		origs, ok := mns.mounts[node]
+		if !ok {
+			// node is not a mount point.
+			return syserror.EINVAL
+		}
+
+		if len(origs) == 0 {
+			panic("cannot unmount initial dirent")
+		}
+
+		if !detachOnly {
+			m := node.Inode.MountSource
+
+			// Lock the parent MountSource first, if it exists. We are
+			// holding mns.Lock, so the parent can not change out
+			// from under us.
+			parent := m.Parent()
+			if parent != nil {
+				parent.mu.Lock()
+				defer parent.mu.Unlock()
+			}
+
+			// Lock the mount that is being unmounted.
+			m.mu.Lock()
+			defer m.mu.Unlock()
+
+			if m.parent != nil {
+				// Sanity check.
+				if _, ok := m.parent.children[m]; !ok {
+					panic(fmt.Sprintf("mount %+v is not a child of parent %+v", m, m.parent))
+				}
+				delete(m.parent.children, m)
+				m.parent = nil
+			}
+
+			// Flush all references on the mounted node.
+			m.FlushDirentRefs()
+
+			// At this point, exactly two references must be held
+			// to mount: one mount reference on node, and one due
+			// to walking to node.
+			//
+			// We must also be guaranteed that no more references
+			// can be taken on mount. This is why withMountLocked
+			// must be held at this point to prevent any walks to
+			// and from node.
+			if refs := m.DirentRefs(); refs < 2 {
+				panic(fmt.Sprintf("have %d refs on unmount, expect 2 or more", refs))
+			} else if refs != 2 {
+				return syserror.EBUSY
+			}
+		}
+
+		original := origs[len(origs)-1]
+		if err := node.unmount(ctx, original); err != nil {
+			return err
+		}
+
+		switch {
+		case len(origs) > 1:
+			mns.mounts[original] = origs[:len(origs)-1]
+		case len(origs) == 1:
+			// Drop mount reference taken at the end of
+			// MountNamespace.Mount.
+			original.DecRef()
+		}
+
+		delete(mns.mounts, node)
+		return nil
+	})
+}
+
+// FindLink returns an Dirent from a given node, which may be a symlink.
+//
+// The root argument is treated as the root directory, and FindLink will not
+// return anything above that. The wd dirent provides the starting directory,
+// and may be nil which indicates the root should be used. You must call DecRef
+// on the resulting Dirent when you are no longer using the object.
+//
+// If wd is nil, then the root will be used as the working directory. If the
+// path is absolute, this has no functional impact.
+//
+// Precondition: root must be non-nil.
+// Precondition: the path must be non-empty.
+func (mns *MountNamespace) FindLink(ctx context.Context, root, wd *Dirent, path string, maxTraversals uint) (*Dirent, error) {
+	if root == nil {
+		panic("MountNamespace.FindInode: root must not be nil")
+	}
+	if len(path) == 0 {
+		panic("MountNamespace.FindInode: path is empty")
+	}
+
+	// Split the path.
+	first, remainder := SplitFirst(path)
+
+	// Where does this walk originate?
+	current := wd
+	if current == nil {
+		current = root
+	}
+	for first == "/" {
+		// Special case: it's possible that we have nothing to walk at
+		// all. This is necessary since we're resplitting the path.
+		if remainder == "" {
+			root.IncRef()
+			return root, nil
+		}
+
+		// Start at the root and advance the path component so that the
+		// walk below can proceed. Note at this point, it handles the
+		// no-op walk case perfectly fine.
+		current = root
+		first, remainder = SplitFirst(remainder)
+	}
+
+	current.IncRef() // Transferred during walk.
+
+	for {
+		// Check that the file is a directory and that we have
+		// permissions to walk.
+		//
+		// Note that we elide this check for the root directory as an
+		// optimization; a non-executable root may still be walked.  A
+		// non-directory root is hopeless.
+		if current != root {
+			if !IsDir(current.Inode.StableAttr) {
+				current.DecRef() // Drop reference from above.
+				return nil, syserror.ENOTDIR
+			}
+			if err := current.Inode.CheckPermission(ctx, PermMask{Execute: true}); err != nil {
+				current.DecRef() // Drop reference from above.
+				return nil, err
+			}
+		}
+
+		// Move to the next level.
+		next, err := current.Walk(ctx, root, first)
+		if err != nil {
+			// Allow failed walks to cache the dirent, because no
+			// children will acquire a reference at the end.
+			current.maybeExtendReference()
+			current.DecRef()
+			return nil, err
+		}
+
+		// Drop old reference.
+		current.DecRef()
+
+		if remainder != "" {
+			// Ensure it's resolved, unless it's the last level.
+			//
+			// See resolve for reference semantics; on err next
+			// will have one dropped.
+			current, err = mns.resolve(ctx, root, next, maxTraversals)
+			if err != nil {
+				return nil, err
+			}
+		} else {
+			// Allow the file system to take an extra reference on the
+			// found child. This will hold a reference on the containing
+			// directory, so the whole tree will be implicitly cached.
+			next.maybeExtendReference()
+			return next, nil
+		}
+
+		// Move to the next element.
+		first, remainder = SplitFirst(remainder)
+	}
+}
+
+// FindInode is identical to FindLink except the return value is resolved.
+//
+//go:nosplit
+func (mns *MountNamespace) FindInode(ctx context.Context, root, wd *Dirent, path string, maxTraversals uint) (*Dirent, error) {
+	d, err := mns.FindLink(ctx, root, wd, path, maxTraversals)
+	if err != nil {
+		return nil, err
+	}
+
+	// See resolve for reference semantics; on err d will have the
+	// reference dropped.
+	return mns.resolve(ctx, root, d, maxTraversals)
+}
+
+// resolve resolves the given link.
+//
+// If successful, a reference is dropped on node and one is acquired on the
+// caller's behalf for the returned dirent.
+//
+// If not successful, a reference is _also_ dropped on the node and an error
+// returned. This is for convenience in using resolve directly as a return
+// value.
+func (mns *MountNamespace) resolve(ctx context.Context, root, node *Dirent, maxTraversals uint) (*Dirent, error) {
+	// Resolve the path.
+	target, err := node.Inode.Getlink(ctx)
+
+	switch err {
+	case nil:
+		// Make sure we didn't exhaust the traversal budget.
+		if maxTraversals == 0 {
+			target.DecRef()
+			return nil, syscall.ELOOP
+		}
+
+		node.DecRef() // Drop the original reference.
+		return target, nil
+
+	case syscall.ENOLINK:
+		// Not a symlink.
+		return node, nil
+
+	case ErrResolveViaReadlink:
+		defer node.DecRef() // See above.
+
+		// First, check if we should traverse.
+		if maxTraversals == 0 {
+			return nil, syscall.ELOOP
+		}
+
+		// Read the target path.
+		targetPath, err := node.Inode.Readlink(ctx)
+		if err != nil {
+			return nil, err
+		}
+
+		// Find the node; we resolve relative to the current symlink's parent.
+		d, err := mns.FindInode(ctx, root, node.parent, targetPath, maxTraversals-1)
+		if err != nil {
+			return nil, err
+		}
+
+		return d, err
+
+	default:
+		node.DecRef() // Drop for err; see above.
+
+		// Propagate the error.
+		return nil, err
+	}
+}
+
+// SyncAll calls Dirent.SyncAll on the root.
+func (mns *MountNamespace) SyncAll(ctx context.Context) {
+	mns.mu.Lock()
+	defer mns.mu.Unlock()
+	mns.root.SyncAll(ctx)
+}
diff --git a/pkg/sentry/fs/mounts_test.go b/pkg/sentry/fs/mounts_test.go
new file mode 100644
index 000000000..8669f3a38
--- /dev/null
+++ b/pkg/sentry/fs/mounts_test.go
@@ -0,0 +1,102 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs_test
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ramfstest "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs/test"
+)
+
+// Creates a new MountNamespace with filesystem:
+// /       (root dir)
+// |-foo   (dir)
+//   |-bar (file)
+func createMountNamespace(ctx context.Context) (*fs.MountNamespace, error) {
+	perms := fs.FilePermsFromMode(0777)
+	m := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+
+	barFile := ramfstest.NewFile(ctx, perms)
+	fooDir := ramfstest.NewDir(ctx, map[string]*fs.Inode{
+		"bar": fs.NewInode(barFile, m, fs.StableAttr{Type: fs.RegularFile}),
+	}, perms)
+	rootDir := ramfstest.NewDir(ctx, map[string]*fs.Inode{
+		"foo": fs.NewInode(fooDir, m, fs.StableAttr{Type: fs.Directory}),
+	}, perms)
+
+	return fs.NewMountNamespace(ctx, fs.NewInode(rootDir, m, fs.StableAttr{Type: fs.Directory}))
+}
+
+func TestFindLink(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm, err := createMountNamespace(ctx)
+	if err != nil {
+		t.Fatalf("createMountNamespace failed: %v", err)
+	}
+
+	root := mm.Root()
+	defer root.DecRef()
+	foo, err := root.Walk(ctx, root, "foo")
+	if err != nil {
+		t.Fatalf("Error walking to foo: %v", err)
+	}
+
+	// Positive cases.
+	for _, tc := range []struct {
+		findPath string
+		wd       *fs.Dirent
+		wantPath string
+	}{
+		{".", root, "/"},
+		{".", foo, "/foo"},
+		{"..", foo, "/"},
+		{"../../..", foo, "/"},
+		{"///foo", foo, "/foo"},
+		{"/foo", foo, "/foo"},
+		{"/foo/bar", foo, "/foo/bar"},
+		{"/foo/.///./bar", foo, "/foo/bar"},
+		{"/foo///bar", foo, "/foo/bar"},
+		{"/foo/../foo/bar", foo, "/foo/bar"},
+		{"foo/bar", root, "/foo/bar"},
+		{"foo////bar", root, "/foo/bar"},
+		{"bar", foo, "/foo/bar"},
+	} {
+		wdPath, _ := tc.wd.FullName(root)
+		if d, err := mm.FindLink(ctx, root, tc.wd, tc.findPath, 0); err != nil {
+			t.Errorf("FindLink(%q, wd=%q) failed: %v", tc.findPath, wdPath, err)
+		} else if got, _ := d.FullName(root); got != tc.wantPath {
+			t.Errorf("FindLink(%q, wd=%q) got dirent %q, want %q", tc.findPath, wdPath, got, tc.wantPath)
+		}
+	}
+
+	// Negative cases.
+	for _, tc := range []struct {
+		findPath string
+		wd       *fs.Dirent
+	}{
+		{"bar", root},
+		{"/bar", root},
+		{"/foo/../../bar", root},
+		{"foo", foo},
+	} {
+		wdPath, _ := tc.wd.FullName(root)
+		if _, err := mm.FindLink(ctx, root, tc.wd, tc.findPath, 0); err == nil {
+			t.Errorf("FindLink(%q, wd=%q) did not return error", tc.findPath, wdPath)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/offset.go b/pkg/sentry/fs/offset.go
new file mode 100644
index 000000000..7cc8398e6
--- /dev/null
+++ b/pkg/sentry/fs/offset.go
@@ -0,0 +1,65 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// OffsetPageEnd returns the file offset rounded up to the nearest
+// page boundary. OffsetPageEnd panics if rounding up causes overflow,
+// which shouldn't be possible given that offset is an int64.
+func OffsetPageEnd(offset int64) uint64 {
+	end, ok := usermem.Addr(offset).RoundUp()
+	if !ok {
+		panic("impossible overflow")
+	}
+	return uint64(end)
+}
+
+// ReadEndOffset returns an exclusive end offset for a read operation
+// so that the read does not overflow an int64 nor size.
+//
+// Parameters:
+// - offset: the starting offset of the read.
+// - length: the number of bytes to read.
+// - size:   the size of the file.
+//
+// Postconditions: The returned offset is >= offset.
+func ReadEndOffset(offset int64, length int64, size int64) int64 {
+	if offset >= size {
+		return offset
+	}
+	end := offset + length
+	// Don't overflow.
+	if end < offset || end > size {
+		end = size
+	}
+	return end
+}
+
+// WriteEndOffset returns an exclusive end offset for a write operation
+// so that the write does not overflow an int64.
+//
+// Parameters:
+// - offset: the starting offset of the write.
+// - length: the number of bytes to write.
+//
+// Postconditions: The returned offset is >= offset.
+func WriteEndOffset(offset int64, length int64) int64 {
+	return ReadEndOffset(offset, length, math.MaxInt64)
+}
diff --git a/pkg/sentry/fs/overlay.go b/pkg/sentry/fs/overlay.go
new file mode 100644
index 000000000..40eed3feb
--- /dev/null
+++ b/pkg/sentry/fs/overlay.go
@@ -0,0 +1,268 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// The virtual filesystem implements an overlay configuration. For a high-level
+// description, see README.md.
+//
+// Note on whiteouts:
+//
+// This implementation does not use the "Docker-style" whiteouts (symlinks with
+// ".wh." prefix). Instead upper filesystem directories support a set of extended
+// attributes to encode whiteouts: "trusted.overlay.whiteout.<filename>". This
+// gives flexibility to persist whiteouts independently of the filesystem layout
+// while additionally preventing name conflicts with files prefixed with ".wh.".
+//
+// Known deficiencies:
+//
+// - The device number of two files under the same overlay mount point may be
+//   different. This can happen if a file is found in the lower filesystem (takes
+//   the lower filesystem device) and another file is created in the upper
+//   filesystem (takes the upper filesystem device). This may appear odd but
+//   should not break applications.
+//
+// - Registered events on files (i.e. for notification of read/write readiness)
+//   are not copied across copy up. This is fine in the common case of files that
+//   do not block. For files that do block, like pipes and sockets, copy up is not
+//   supported.
+//
+// - Hardlinks in a lower filesystem are broken by copy up. For this reason, no
+//   attempt is made to preserve link count across copy up.
+//
+// - The maximum length of an extended attribute name is the same as the maximum
+//   length of a file path in Linux (XATTR_NAME_MAX == NAME_MAX). This means that
+//   whiteout attributes, if set directly on the host, are limited additionally by
+//   the extra whiteout prefix length (file paths must be strictly shorter than
+//   NAME_MAX). This is not a problem for in-memory filesystems which don't enforce
+//   XATTR_NAME_MAX.
+
+const (
+	// XattrOverlayPrefix is the prefix for extended attributes that affect
+	// the behavior of an overlay.
+	XattrOverlayPrefix = "trusted.overlay."
+
+	// XattrOverlayWhiteoutPrefix is the prefix for extended attributes
+	// that indicate that a whiteout exists.
+	XattrOverlayWhiteoutPrefix = XattrOverlayPrefix + "whiteout."
+)
+
+// XattrOverlayWhiteout returns an extended attribute that indicates a
+// whiteout exists for name. It is supported by directories that wish to
+// mask the existence of name.
+func XattrOverlayWhiteout(name string) string {
+	return XattrOverlayWhiteoutPrefix + name
+}
+
+// NewOverlayRoot produces the root of an overlay.
+//
+// Preconditions:
+//
+// - upper and lower must be non-nil.
+// - lower should not expose character devices, pipes, or sockets, because
+//   copying up these types of files is not supported.
+// - upper and lower must not require that file objects be revalidated.
+// - upper and lower must not have dynamic file/directory content.
+func NewOverlayRoot(ctx context.Context, upper *Inode, lower *Inode, flags MountSourceFlags) (*Inode, error) {
+	if !IsDir(upper.StableAttr) {
+		return nil, fmt.Errorf("upper Inode is not a directory")
+	}
+	if !IsDir(lower.StableAttr) {
+		return nil, fmt.Errorf("lower Inode is not a directory")
+	}
+
+	msrc := newOverlayMountSource(upper.MountSource, lower.MountSource, flags)
+	overlay, err := newOverlayEntry(ctx, upper, lower, true)
+	if err != nil {
+		msrc.DecRef()
+		return nil, err
+	}
+
+	return newOverlayInode(ctx, overlay, msrc), nil
+}
+
+// newOverlayInode creates a new Inode for an overlay.
+func newOverlayInode(ctx context.Context, o *overlayEntry, msrc *MountSource) *Inode {
+	var inode *Inode
+	if o.upper != nil {
+		inode = NewInode(nil, msrc, o.upper.StableAttr)
+	} else {
+		inode = NewInode(nil, msrc, o.lower.StableAttr)
+	}
+	inode.overlay = o
+	return inode
+}
+
+// overlayEntry is the overlay metadata of an Inode. It implements Mappable.
+type overlayEntry struct {
+	// lowerExists is true if an Inode exists for this file in the lower
+	// filesystem. If lowerExists is true, then the overlay must create
+	// a whiteout entry when renaming and removing this entry to mask the
+	// lower Inode.
+	//
+	// Note that this is distinct from actually holding onto a non-nil
+	// lower Inode (below). The overlay does not need to keep a lower Inode
+	// around unless it needs to operate on it, but it always needs to know
+	// whether the lower Inode exists to correctly execute a rename or
+	// remove operation.
+	lowerExists bool
+
+	// lower is an Inode from a lower filesystem. Modifications are
+	// never made on this Inode.
+	lower *Inode
+
+	// copyMu serializes copy-up for operations above
+	// mm.MemoryManager.mappingMu in the lock order.
+	copyMu sync.RWMutex `state:"nosave"`
+
+	// mapsMu serializes copy-up for operations between
+	// mm.MemoryManager.mappingMu and mm.MemoryManager.activeMu in the lock
+	// order.
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings tracks memory mappings of this Mappable so they can be removed
+	// from the lower filesystem Mappable and added to the upper filesystem
+	// Mappable when copy up occurs. It is strictly unnecessary after copy-up.
+	//
+	// mappings is protected by mapsMu.
+	mappings memmap.MappingSet
+
+	// dataMu serializes copy-up for operations below mm.MemoryManager.activeMu
+	// in the lock order.
+	dataMu sync.RWMutex `state:"nosave"`
+
+	// upper is an Inode from an upper filesystem. It is non-nil if
+	// the file exists in the upper filesystem. It becomes non-nil
+	// when the Inode that owns this overlayEntry is modified.
+	//
+	// upper is protected by all of copyMu, mapsMu, and dataMu. Holding any of
+	// these locks is sufficient to read upper; holding all three for writing
+	// is required to mutate it.
+	upper *Inode
+}
+
+// newOverlayEntry returns a new overlayEntry.
+func newOverlayEntry(ctx context.Context, upper *Inode, lower *Inode, lowerExists bool) (*overlayEntry, error) {
+	if upper == nil && lower == nil {
+		panic("invalid overlayEntry, needs at least one Inode")
+	}
+	if upper != nil && upper.overlay != nil {
+		panic("nested writable layers are not supported")
+	}
+	// Check for supported lower filesystem types.
+	if lower != nil {
+		switch lower.StableAttr.Type {
+		case RegularFile, Directory, Symlink, Socket:
+		default:
+			// We don't support copying up from character devices,
+			// named pipes, or anything weird (like proc files).
+			log.Warningf("%s not supported in lower filesytem", lower.StableAttr.Type)
+			return nil, syserror.EINVAL
+		}
+	}
+	return &overlayEntry{
+		lowerExists: lowerExists,
+		lower:       lower,
+		upper:       upper,
+	}, nil
+}
+
+func (o *overlayEntry) release() {
+	// We drop a reference on upper and lower file system Inodes
+	// rather than releasing them, because in-memory filesystems
+	// may hold an extra reference to these Inodes so that they
+	// stay in memory.
+	if o.upper != nil {
+		o.upper.DecRef()
+	}
+	if o.lower != nil {
+		o.lower.DecRef()
+	}
+}
+
+// overlayUpperMountSource gives the upper mount of an overlay mount.
+//
+// The caller may not use this MountSource past the lifetime of overlayMountSource and may
+// not call DecRef on it.
+func overlayUpperMountSource(overlayMountSource *MountSource) *MountSource {
+	return overlayMountSource.MountSourceOperations.(*overlayMountSourceOperations).upper
+}
+
+// Preconditions: At least one of o.copyMu, o.mapsMu, or o.dataMu must be locked.
+func (o *overlayEntry) inodeLocked() *Inode {
+	if o.upper != nil {
+		return o.upper
+	}
+	return o.lower
+}
+
+// Preconditions: At least one of o.copyMu, o.mapsMu, or o.dataMu must be locked.
+func (o *overlayEntry) isMappableLocked() bool {
+	return o.inodeLocked().Mappable() != nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (o *overlayEntry) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	o.mapsMu.Lock()
+	defer o.mapsMu.Unlock()
+	if err := o.inodeLocked().Mappable().AddMapping(ctx, ms, ar, offset); err != nil {
+		return err
+	}
+	o.mappings.AddMapping(ms, ar, offset)
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (o *overlayEntry) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+	o.mapsMu.Lock()
+	defer o.mapsMu.Unlock()
+	o.inodeLocked().Mappable().RemoveMapping(ctx, ms, ar, offset)
+	o.mappings.RemoveMapping(ms, ar, offset)
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (o *overlayEntry) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	o.mapsMu.Lock()
+	defer o.mapsMu.Unlock()
+	if err := o.inodeLocked().Mappable().CopyMapping(ctx, ms, srcAR, dstAR, offset); err != nil {
+		return err
+	}
+	o.mappings.AddMapping(ms, dstAR, offset)
+	return nil
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (o *overlayEntry) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	o.dataMu.RLock()
+	defer o.dataMu.RUnlock()
+	return o.inodeLocked().Mappable().Translate(ctx, required, optional, at)
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (o *overlayEntry) InvalidateUnsavable(ctx context.Context) error {
+	o.mapsMu.Lock()
+	defer o.mapsMu.Unlock()
+	return o.inodeLocked().Mappable().InvalidateUnsavable(ctx)
+}
diff --git a/pkg/sentry/fs/path.go b/pkg/sentry/fs/path.go
new file mode 100644
index 000000000..b74f6ed8c
--- /dev/null
+++ b/pkg/sentry/fs/path.go
@@ -0,0 +1,92 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// TrimTrailingSlashes trims any trailing slashes.
+//
+// The returned boolean indicates whether any changes were made.
+//
+//go:nosplit
+func TrimTrailingSlashes(dir string) (trimmed string, changed bool) {
+	// Trim the trailing slash, except for root.
+	for len(dir) > 1 && dir[len(dir)-1] == '/' {
+		dir = dir[:len(dir)-1]
+		changed = true
+	}
+	return dir, changed
+}
+
+// SplitLast splits the given path into a directory and a file.
+//
+// The "absoluteness" of the path is preserved, but dir is always stripped of
+// trailing slashes.
+//
+//go:nosplit
+func SplitLast(path string) (dir, file string) {
+	path, _ = TrimTrailingSlashes(path)
+	if path == "" {
+		return ".", "."
+	} else if path == "/" {
+		return "/", "."
+	}
+
+	var slash int // Last location of slash in path.
+	for slash = len(path) - 1; slash >= 0 && path[slash] != '/'; slash-- {
+	}
+	switch {
+	case slash < 0:
+		return ".", path
+	case slash == 0:
+		// Directory of the form "/foo", or just "/". We need to
+		// preserve the first slash here, since it indicates an
+		// absolute path.
+		return "/", path[1:]
+	default:
+		// Drop the trailing slash.
+		dir, _ = TrimTrailingSlashes(path[:slash])
+		return dir, path[slash+1:]
+	}
+}
+
+// SplitFirst splits the given path into a first directory and the remainder.
+//
+// If remainder is empty, then the path is a single element.
+//
+//go:nosplit
+func SplitFirst(path string) (current, remainder string) {
+	path, _ = TrimTrailingSlashes(path)
+	if path == "" {
+		return ".", ""
+	}
+
+	var slash int // First location of slash in path.
+	for slash = 0; slash < len(path) && path[slash] != '/'; slash++ {
+	}
+	switch {
+	case slash >= len(path):
+		return path, ""
+	case slash == 0:
+		// See above.
+		return "/", path[1:]
+	default:
+		current = path[:slash]
+		remainder = path[slash+1:]
+		// Strip redundant slashes.
+		for len(remainder) > 0 && remainder[0] == '/' {
+			remainder = remainder[1:]
+		}
+		return current, remainder
+	}
+}
diff --git a/pkg/sentry/fs/path_test.go b/pkg/sentry/fs/path_test.go
new file mode 100644
index 000000000..7ab070855
--- /dev/null
+++ b/pkg/sentry/fs/path_test.go
@@ -0,0 +1,211 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"testing"
+)
+
+// TestSplitLast tests variants of path splitting.
+func TestSplitLast(t *testing.T) {
+	cases := []struct {
+		path string
+		dir  string
+		file string
+	}{
+		{path: "/", dir: "/", file: "."},
+		{path: "/.", dir: "/", file: "."},
+		{path: "/./", dir: "/", file: "."},
+		{path: "/./.", dir: "/.", file: "."},
+		{path: "/././", dir: "/.", file: "."},
+		{path: "/./..", dir: "/.", file: ".."},
+		{path: "/./../", dir: "/.", file: ".."},
+		{path: "/..", dir: "/", file: ".."},
+		{path: "/../", dir: "/", file: ".."},
+		{path: "/../.", dir: "/..", file: "."},
+		{path: "/.././", dir: "/..", file: "."},
+		{path: "/../..", dir: "/..", file: ".."},
+		{path: "/../../", dir: "/..", file: ".."},
+
+		{path: "", dir: ".", file: "."},
+		{path: ".", dir: ".", file: "."},
+		{path: "./", dir: ".", file: "."},
+		{path: "./.", dir: ".", file: "."},
+		{path: "././", dir: ".", file: "."},
+		{path: "./..", dir: ".", file: ".."},
+		{path: "./../", dir: ".", file: ".."},
+		{path: "..", dir: ".", file: ".."},
+		{path: "../", dir: ".", file: ".."},
+		{path: "../.", dir: "..", file: "."},
+		{path: ".././", dir: "..", file: "."},
+		{path: "../..", dir: "..", file: ".."},
+		{path: "../../", dir: "..", file: ".."},
+
+		{path: "/foo", dir: "/", file: "foo"},
+		{path: "/foo/", dir: "/", file: "foo"},
+		{path: "/foo/.", dir: "/foo", file: "."},
+		{path: "/foo/./", dir: "/foo", file: "."},
+		{path: "/foo/./.", dir: "/foo/.", file: "."},
+		{path: "/foo/./..", dir: "/foo/.", file: ".."},
+		{path: "/foo/..", dir: "/foo", file: ".."},
+		{path: "/foo/../", dir: "/foo", file: ".."},
+		{path: "/foo/../.", dir: "/foo/..", file: "."},
+		{path: "/foo/../..", dir: "/foo/..", file: ".."},
+
+		{path: "/foo/bar", dir: "/foo", file: "bar"},
+		{path: "/foo/bar/", dir: "/foo", file: "bar"},
+		{path: "/foo/bar/.", dir: "/foo/bar", file: "."},
+		{path: "/foo/bar/./", dir: "/foo/bar", file: "."},
+		{path: "/foo/bar/./.", dir: "/foo/bar/.", file: "."},
+		{path: "/foo/bar/./..", dir: "/foo/bar/.", file: ".."},
+		{path: "/foo/bar/..", dir: "/foo/bar", file: ".."},
+		{path: "/foo/bar/../", dir: "/foo/bar", file: ".."},
+		{path: "/foo/bar/../.", dir: "/foo/bar/..", file: "."},
+		{path: "/foo/bar/../..", dir: "/foo/bar/..", file: ".."},
+
+		{path: "foo", dir: ".", file: "foo"},
+		{path: "foo", dir: ".", file: "foo"},
+		{path: "foo/", dir: ".", file: "foo"},
+		{path: "foo/.", dir: "foo", file: "."},
+		{path: "foo/./", dir: "foo", file: "."},
+		{path: "foo/./.", dir: "foo/.", file: "."},
+		{path: "foo/./..", dir: "foo/.", file: ".."},
+		{path: "foo/..", dir: "foo", file: ".."},
+		{path: "foo/../", dir: "foo", file: ".."},
+		{path: "foo/../.", dir: "foo/..", file: "."},
+		{path: "foo/../..", dir: "foo/..", file: ".."},
+		{path: "foo/", dir: ".", file: "foo"},
+		{path: "foo/.", dir: "foo", file: "."},
+
+		{path: "foo/bar", dir: "foo", file: "bar"},
+		{path: "foo/bar/", dir: "foo", file: "bar"},
+		{path: "foo/bar/.", dir: "foo/bar", file: "."},
+		{path: "foo/bar/./", dir: "foo/bar", file: "."},
+		{path: "foo/bar/./.", dir: "foo/bar/.", file: "."},
+		{path: "foo/bar/./..", dir: "foo/bar/.", file: ".."},
+		{path: "foo/bar/..", dir: "foo/bar", file: ".."},
+		{path: "foo/bar/../", dir: "foo/bar", file: ".."},
+		{path: "foo/bar/../.", dir: "foo/bar/..", file: "."},
+		{path: "foo/bar/../..", dir: "foo/bar/..", file: ".."},
+		{path: "foo/bar/", dir: "foo", file: "bar"},
+		{path: "foo/bar/.", dir: "foo/bar", file: "."},
+	}
+
+	for _, c := range cases {
+		dir, file := SplitLast(c.path)
+		if dir != c.dir || file != c.file {
+			t.Errorf("SplitLast(%q) got (%q, %q), expected (%q, %q)", c.path, dir, file, c.dir, c.file)
+		}
+	}
+}
+
+// TestSplitFirst tests variants of path splitting.
+func TestSplitFirst(t *testing.T) {
+	cases := []struct {
+		path      string
+		first     string
+		remainder string
+	}{
+		{path: "/", first: "/", remainder: ""},
+		{path: "/.", first: "/", remainder: "."},
+		{path: "///.", first: "/", remainder: "//."},
+		{path: "/.///", first: "/", remainder: "."},
+		{path: "/./.", first: "/", remainder: "./."},
+		{path: "/././", first: "/", remainder: "./."},
+		{path: "/./..", first: "/", remainder: "./.."},
+		{path: "/./../", first: "/", remainder: "./.."},
+		{path: "/..", first: "/", remainder: ".."},
+		{path: "/../", first: "/", remainder: ".."},
+		{path: "/../.", first: "/", remainder: "../."},
+		{path: "/.././", first: "/", remainder: "../."},
+		{path: "/../..", first: "/", remainder: "../.."},
+		{path: "/../../", first: "/", remainder: "../.."},
+
+		{path: "", first: ".", remainder: ""},
+		{path: ".", first: ".", remainder: ""},
+		{path: "./", first: ".", remainder: ""},
+		{path: ".///", first: ".", remainder: ""},
+		{path: "./.", first: ".", remainder: "."},
+		{path: "././", first: ".", remainder: "."},
+		{path: "./..", first: ".", remainder: ".."},
+		{path: "./../", first: ".", remainder: ".."},
+		{path: "..", first: "..", remainder: ""},
+		{path: "../", first: "..", remainder: ""},
+		{path: "../.", first: "..", remainder: "."},
+		{path: ".././", first: "..", remainder: "."},
+		{path: "../..", first: "..", remainder: ".."},
+		{path: "../../", first: "..", remainder: ".."},
+
+		{path: "/foo", first: "/", remainder: "foo"},
+		{path: "/foo/", first: "/", remainder: "foo"},
+		{path: "/foo///", first: "/", remainder: "foo"},
+		{path: "/foo/.", first: "/", remainder: "foo/."},
+		{path: "/foo/./", first: "/", remainder: "foo/."},
+		{path: "/foo/./.", first: "/", remainder: "foo/./."},
+		{path: "/foo/./..", first: "/", remainder: "foo/./.."},
+		{path: "/foo/..", first: "/", remainder: "foo/.."},
+		{path: "/foo/../", first: "/", remainder: "foo/.."},
+		{path: "/foo/../.", first: "/", remainder: "foo/../."},
+		{path: "/foo/../..", first: "/", remainder: "foo/../.."},
+
+		{path: "/foo/bar", first: "/", remainder: "foo/bar"},
+		{path: "///foo/bar", first: "/", remainder: "//foo/bar"},
+		{path: "/foo///bar", first: "/", remainder: "foo///bar"},
+		{path: "/foo/bar/.", first: "/", remainder: "foo/bar/."},
+		{path: "/foo/bar/./", first: "/", remainder: "foo/bar/."},
+		{path: "/foo/bar/./.", first: "/", remainder: "foo/bar/./."},
+		{path: "/foo/bar/./..", first: "/", remainder: "foo/bar/./.."},
+		{path: "/foo/bar/..", first: "/", remainder: "foo/bar/.."},
+		{path: "/foo/bar/../", first: "/", remainder: "foo/bar/.."},
+		{path: "/foo/bar/../.", first: "/", remainder: "foo/bar/../."},
+		{path: "/foo/bar/../..", first: "/", remainder: "foo/bar/../.."},
+
+		{path: "foo", first: "foo", remainder: ""},
+		{path: "foo", first: "foo", remainder: ""},
+		{path: "foo/", first: "foo", remainder: ""},
+		{path: "foo///", first: "foo", remainder: ""},
+		{path: "foo/.", first: "foo", remainder: "."},
+		{path: "foo/./", first: "foo", remainder: "."},
+		{path: "foo/./.", first: "foo", remainder: "./."},
+		{path: "foo/./..", first: "foo", remainder: "./.."},
+		{path: "foo/..", first: "foo", remainder: ".."},
+		{path: "foo/../", first: "foo", remainder: ".."},
+		{path: "foo/../.", first: "foo", remainder: "../."},
+		{path: "foo/../..", first: "foo", remainder: "../.."},
+		{path: "foo/", first: "foo", remainder: ""},
+		{path: "foo/.", first: "foo", remainder: "."},
+
+		{path: "foo/bar", first: "foo", remainder: "bar"},
+		{path: "foo///bar", first: "foo", remainder: "bar"},
+		{path: "foo/bar/", first: "foo", remainder: "bar"},
+		{path: "foo/bar/.", first: "foo", remainder: "bar/."},
+		{path: "foo/bar/./", first: "foo", remainder: "bar/."},
+		{path: "foo/bar/./.", first: "foo", remainder: "bar/./."},
+		{path: "foo/bar/./..", first: "foo", remainder: "bar/./.."},
+		{path: "foo/bar/..", first: "foo", remainder: "bar/.."},
+		{path: "foo/bar/../", first: "foo", remainder: "bar/.."},
+		{path: "foo/bar/../.", first: "foo", remainder: "bar/../."},
+		{path: "foo/bar/../..", first: "foo", remainder: "bar/../.."},
+		{path: "foo/bar/", first: "foo", remainder: "bar"},
+		{path: "foo/bar/.", first: "foo", remainder: "bar/."},
+	}
+
+	for _, c := range cases {
+		first, remainder := SplitFirst(c.path)
+		if first != c.first || remainder != c.remainder {
+			t.Errorf("SplitFirst(%q) got (%q, %q), expected (%q, %q)", c.path, first, remainder, c.first, c.remainder)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD
new file mode 100644
index 000000000..18372cfbf
--- /dev/null
+++ b/pkg/sentry/fs/proc/BUILD
@@ -0,0 +1,95 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "proc_state",
+    srcs = [
+        "cpuinfo.go",
+        "exec_args.go",
+        "fds.go",
+        "file.go",
+        "filesystems.go",
+        "fs.go",
+        "loadavg.go",
+        "meminfo.go",
+        "mounts.go",
+        "net.go",
+        "proc.go",
+        "stat.go",
+        "sys.go",
+        "sys_net.go",
+        "task.go",
+        "uid_gid_map.go",
+        "uptime.go",
+        "version.go",
+    ],
+    out = "proc_state.go",
+    package = "proc",
+)
+
+go_library(
+    name = "proc",
+    srcs = [
+        "cpuinfo.go",
+        "exec_args.go",
+        "fds.go",
+        "file.go",
+        "filesystems.go",
+        "fs.go",
+        "loadavg.go",
+        "meminfo.go",
+        "mounts.go",
+        "net.go",
+        "proc.go",
+        "proc_state.go",
+        "stat.go",
+        "sys.go",
+        "sys_net.go",
+        "task.go",
+        "uid_gid_map.go",
+        "uptime.go",
+        "version.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/proc/device",
+        "//pkg/sentry/fs/proc/seqfile",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/mm",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "proc_test",
+    size = "small",
+    srcs = [
+        "net_test.go",
+        "sys_net_test.go",
+    ],
+    embed = [":proc"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/proc/README.md b/pkg/sentry/fs/proc/README.md
new file mode 100644
index 000000000..c510ee63a
--- /dev/null
+++ b/pkg/sentry/fs/proc/README.md
@@ -0,0 +1,317 @@
+This document tracks what is implemented in procfs. Refer to
+Documentation/filesystems/proc.txt in the Linux project for information about
+procfs generally.
+
+**NOTE**: This document is not guaranteed to be up to date. If you find an
+inconsistency, please file a bug.
+
+[TOC]
+## Kernel data
+
+The following files are implemented:
+
+| File /proc/               | Content                                          |
+| :------------------------ | :----------------------------------------------- |
+| [cpuinfo](#cpuinfo)       | Info about the CPU                               |
+| [filesystem](#filesystem) | Supported filesystems                            |
+| [loadavg](#loadavg)       | Load average of last 1, 5 & 15 minutes           |
+| [meminfo](#meminfo)       | Overall memory info                              |
+| [stat](#stat)             | Overall kernel statistics                        |
+| [sys](#sys)               | Change parameters within the kernel              |
+| [uptime](#uptime)         | Wall clock since boot, combined idle time of all |
+:                           : cpus                                             :
+| [version](#version)       | Kernel version                                   |
+
+### cpuinfo
+
+```bash
+$ cat /proc/cpuinfo
+processor   : 0
+vendor_id   : GenuineIntel
+cpu family  : 6
+model       : 45
+model name  : unknown
+stepping    : unknown
+cpu MHz     : 1234.588
+fpu     : yes
+fpu_exception   : yes
+cpuid level : 13
+wp      : yes
+flags       : fpu vme de pse tsc msr pae mce cx8 apic sep mtrr pge mca cmov pat pse36 clflush dts acpi mmx fxsr sse sse2 ss ht tm pbe syscall nx pdpe1gb rdtscp lm pni pclmulqdq dtes64 monitor ds_cpl vmx smx est tm2 ssse3 cx16 xtpr pdcm pcid dca sse4_1 sse4_2 x2apic popcnt tsc_deadline_timer aes xsave avx xsaveopt
+bogomips    : 1234.59
+clflush size    : 64
+cache_alignment : 64
+address sizes   : 46 bits physical, 48 bits virtual
+power management:
+
+...
+```
+
+Notable divergences:
+
+Field name       | Notes
+:--------------- | :---------------------------------------
+model name       | Always unknown
+stepping         | Always unknown
+fpu              | Always yes
+fpu_exception    | Always yes
+wp               | Always yes
+bogomips         | Bogus value (matches cpu MHz)
+clflush size     | Always 64
+cache_alignment  | Always 64
+address sizes    | Always 46 bits physical, 48 bits virtual
+power management | Always blank
+
+Otherwise fields are derived from the SentryCPUIDSpec proto config.
+
+### filesystem
+
+```bash
+$ cat /proc/filesystem
+nodev   9p
+nodev   devtmpfs
+nodev   proc
+nodev   ramdiskfs
+nodev   sysfs
+nodev   tmpfs
+```
+
+Notable divergences:
+
+Filesystem | Notes
+:--------- | :--------------------------------------------------------
+ramdiskfs  | No Linux equivalent, see the SentryRamdiskFS proto config
+
+### loadavg
+
+```bash
+$ cat /proc/loadavg
+0.00 0.00 0.00 0/0 0
+```
+
+Column                                | Notes
+:------------------------------------ | :----------
+CPU.IO utilization in last 1 minute   | Always zero
+CPU.IO utilization in last 5 minutes  | Always zero
+CPU.IO utilization in last 10 minutes | Always zero
+Num currently running processes       | Always zero
+Total num processes                   | Always zero
+
+TODO: Populate the columns with accurate statistics.
+### meminfo
+
+```bash
+$ cat /proc/meminfo
+MemTotal:        2097152 kB
+MemFree:         2083540 kB
+MemAvailable:    2083540 kB
+Buffers:               0 kB
+Cached:             4428 kB
+SwapCache:             0 kB
+Active:            10812 kB
+Inactive:           2216 kB
+Active(anon):       8600 kB
+Inactive(anon):        0 kB
+Active(file):       2212 kB
+Inactive(file):     2216 kB
+Unevictable:           0 kB
+Mlocked:               0 kB
+SwapTotal:             0 kB
+SwapFree:              0 kB
+Dirty:                 0 kB
+Writeback:             0 kB
+AnonPages:          8600 kB
+Mapped:             4428 kB
+Shmem:                 0 kB
+
+```
+
+Notable divergences:
+
+Field name        | Notes
+:---------------- | :--------------------------------------------------------
+Buffers           | Always zero, no block devices
+SwapCache         | Always zero, no swap
+Inactive(anon)    | Always zero, see SwapCache
+Unevictable       | Always zero TODO
+Mlocked           | Always zero TODO
+SwapTotal         | Always zero, no swap
+SwapFree          | Always zero, no swap
+Dirty             | Always zero TODO
+Writeback         | Always zero TODO
+MemAvailable      | Uses the same value as MemFree since there is no swap.
+Slab              | Missing
+SReclaimable      | Missing
+SUnreclaim        | Missing
+KernelStack       | Missing
+PageTables        | Missing
+NFS_Unstable      | Missing
+Bounce            | Missing
+WritebackTmp      | Missing
+CommitLimit       | Missing
+Committed_AS      | Missing
+VmallocTotal      | Missing
+VmallocUsed       | Missing
+VmallocChunk      | Missing
+HardwareCorrupted | Missing
+AnonHugePages     | Missing
+ShmemHugePages    | Missing
+ShmemPmdMapped    | Missing
+HugePages_Total   | Missing
+HugePages_Free    | Missing
+HugePages_Rsvd    | Missing
+HugePages_Surp    | Missing
+Hugepagesize      | Missing
+DirectMap4k       | Missing
+DirectMap2M       | Missing
+DirectMap1G       | Missing
+
+See [Memory
+Accounting](pkg/sentry/usage/g3doc/memory-accounting.md)
+for general caveats.
+
+### stat
+
+```bash
+$ cat /proc/stat
+cpu  0 0 0 0 0 0 0 0 0 0
+cpu0 0 0 0 0 0 0 0 0 0 0
+cpu1 0 0 0 0 0 0 0 0 0 0
+cpu2 0 0 0 0 0 0 0 0 0 0
+cpu3 0 0 0 0 0 0 0 0 0 0
+cpu4 0 0 0 0 0 0 0 0 0 0
+cpu5 0 0 0 0 0 0 0 0 0 0
+cpu6 0 0 0 0 0 0 0 0 0 0
+cpu7 0 0 0 0 0 0 0 0 0 0
+intr 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
+ctxt 0
+btime 1504040968
+processes 0
+procs_running 0
+procs_blokkcked 0
+softirq 0 0 0 0 0 0 0 0 0 0 0
+```
+
+All fields except for `btime` are always zero.
+TODO: Populate with accurate fields.
+
+### sys
+
+```bash
+$ ls /proc/sys
+kernel vm
+```
+
+Directory | Notes
+:-------- | :----------------------------
+abi       | Missing
+debug     | Missing
+dev       | Missing
+fs        | Missing
+kernel    | Contains hostname (only)
+net       | Missing
+user      | Missing
+vm        | Contains mmap_min_addr (only)
+
+### uptime
+
+```bash
+$ cat /proc/uptime
+3204.62 0.00
+```
+
+Column                           | Notes
+:------------------------------- | :----------------------------
+Total num seconds system running | Time since procfs was mounted
+Number of seconds idle           | Always zero
+
+### version
+
+```bash
+$ cat /proc/version
+Linux version 3.11.10 #1 SMP Fri Nov 29 10:47:50 PST 2013
+```
+
+## Process-specific data
+
+The following files are implemented:
+
+File /proc/PID      | Content
+:------------------ | :---------------------------------------------------
+[auxv](#auxv)       | Copy of auxiliary vector for the process
+[cmdline](#cmdline) | Command line arguments
+[comm](#comm)       | Command name associated with the process
+[exe](#exe)         | Symlink to the process's executable
+[fd](#fd)           | Directory containing links to open file descriptors
+[fdinfo](#fdinfo)   | Information associated with open file descriptors
+[gid_map](#gid_map) | Mappings for group IDs inside the user namespace
+[io](#io)           | IO statistics
+[maps](#maps)       | Memory mappings (anon, executables, library files)
+[ns](#ns)           | Directory containing info about supported namespaces
+[stat](#stat)       | Process statistics
+[status](#status)   | Process status in human readable format
+[task](#task)       | Directory containing info about running threads
+[uid_map](#uid_map) | Mappings for user IDs inside the user namespace
+
+### auxv
+
+TODO
+
+### cmdline
+
+TODO
+
+### comm
+
+TODO
+
+### exe
+
+TODO
+
+### fd
+
+TODO
+
+### fdinfo
+
+TODO
+
+### gid_map
+
+TODO
+
+### io
+
+Only has data for rchar, wchar, syscr, and syscw.
+
+TODO: add more detail.
+
+### maps
+
+TODO
+
+### ns
+
+TODO
+
+### stat
+
+Only has data for pid, comm, state, ppid, utime, stime, cutime, cstime,
+num_threads, and exit_signal.
+
+TODO: add more detail.
+
+### status
+
+Statically created, most of the fields have no data.
+
+TODO: add more detail.
+
+### task
+
+TODO
+
+### uid_map
+
+TODO
diff --git a/pkg/sentry/fs/proc/cpuinfo.go b/pkg/sentry/fs/proc/cpuinfo.go
new file mode 100644
index 000000000..f80aaa5b1
--- /dev/null
+++ b/pkg/sentry/fs/proc/cpuinfo.go
@@ -0,0 +1,64 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// cpuinfo is a file describing the CPU capabilities.
+//
+// Presently cpuinfo never changes, so it doesn't need to be a SeqFile.
+type cpuinfo struct {
+	ramfs.Entry
+
+	// k is the system kernel.
+	k *kernel.Kernel
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (c *cpuinfo) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	features := c.k.FeatureSet()
+	if features == nil {
+		// Kernel is always initialized with a FeatureSet.
+		panic("cpuinfo read with nil FeatureSet")
+	}
+
+	contents := make([]byte, 0, 1024)
+	for i, max := uint(0), c.k.ApplicationCores(); i < max; i++ {
+		contents = append(contents, []byte(features.CPUInfo(i))...)
+	}
+	if offset >= int64(len(contents)) {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOut(ctx, contents[offset:])
+	return int64(n), err
+}
+
+func (p *proc) newCPUInfo(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	f := &cpuinfo{
+		k: p.k,
+	}
+	f.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0444))
+
+	return newFile(f, msrc, fs.SpecialFile, nil)
+}
diff --git a/pkg/sentry/fs/proc/device/BUILD b/pkg/sentry/fs/proc/device/BUILD
new file mode 100644
index 000000000..b62062bd7
--- /dev/null
+++ b/pkg/sentry/fs/proc/device/BUILD
@@ -0,0 +1,11 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "device",
+    srcs = ["device.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device",
+    visibility = ["//pkg/sentry:internal"],
+    deps = ["//pkg/sentry/device"],
+)
diff --git a/pkg/sentry/fs/proc/device/device.go b/pkg/sentry/fs/proc/device/device.go
new file mode 100644
index 000000000..6194afe88
--- /dev/null
+++ b/pkg/sentry/fs/proc/device/device.go
@@ -0,0 +1,23 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package device contains the proc device to avoid dependency loops.
+package device
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+)
+
+// ProcDevice is the kernel proc device.
+var ProcDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/proc/exec_args.go b/pkg/sentry/fs/proc/exec_args.go
new file mode 100644
index 000000000..0e1523bf1
--- /dev/null
+++ b/pkg/sentry/fs/proc/exec_args.go
@@ -0,0 +1,129 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// execArgType enumerates the types of exec arguments that are exposed through
+// proc.
+type execArgType int
+
+const (
+	cmdlineExecArg execArgType = iota
+	environExecArg
+)
+
+// execArgFile is a file containing the exec args (either cmdline or environ)
+// for a given task.
+type execArgFile struct {
+	ramfs.Entry
+
+	// arg is the type of exec argument this file contains.
+	arg execArgType
+
+	// t is the Task to read the exec arg line from.
+	t *kernel.Task
+}
+
+// newExecArgFile creates a file containing the exec args of the given type.
+func newExecArgFile(t *kernel.Task, msrc *fs.MountSource, arg execArgType) *fs.Inode {
+	if arg != cmdlineExecArg && arg != environExecArg {
+		panic(fmt.Sprintf("unknown exec arg type %v", arg))
+	}
+	f := &execArgFile{
+		arg: arg,
+		t:   t,
+	}
+	f.InitEntry(t, fs.RootOwner, fs.FilePermsFromMode(0444))
+	return newFile(f, msrc, fs.SpecialFile, t)
+}
+
+// DeprecatedPreadv reads the exec arg from the process's address space..
+func (f *execArgFile) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	// N.B. Linux 4.2 eliminates the arbitrary one page limit.
+	if offset > usermem.PageSize {
+		return 0, io.EOF
+	}
+	dst = dst.TakeFirst64(usermem.PageSize - offset)
+
+	m, err := getTaskMM(f.t)
+	if err != nil {
+		return 0, err
+	}
+	defer m.DecUsers(ctx)
+
+	// Figure out the bounds of the exec arg we are trying to read.
+	var execArgStart, execArgEnd usermem.Addr
+	switch f.arg {
+	case cmdlineExecArg:
+		execArgStart, execArgEnd = m.ArgvStart(), m.ArgvEnd()
+	case environExecArg:
+		execArgStart, execArgEnd = m.EnvvStart(), m.EnvvEnd()
+	default:
+		panic(fmt.Sprintf("unknown exec arg type %v", f.arg))
+	}
+	if execArgStart == 0 || execArgEnd == 0 {
+		// Don't attempt to read before the start/end are set up.
+		return 0, io.EOF
+	}
+
+	start, ok := execArgStart.AddLength(uint64(offset))
+	if !ok {
+		return 0, io.EOF
+	}
+	if start >= execArgEnd {
+		return 0, io.EOF
+	}
+
+	length := int(execArgEnd - start)
+	if dstlen := dst.NumBytes(); int64(length) > dstlen {
+		length = int(dstlen)
+	}
+
+	buf := make([]byte, length)
+	// N.B. Technically this should be usermem.IOOpts.IgnorePermissions = true
+	// until Linux 4.9 (272ddc8b3735 "proc: don't use FOLL_FORCE for reading
+	// cmdline and environment").
+	copyN, copyErr := m.CopyIn(ctx, start, buf, usermem.IOOpts{})
+	if copyN == 0 {
+		// Nothing to copy.
+		return 0, copyErr
+	}
+	buf = buf[:copyN]
+
+	// TODO: On Linux, if the NUL byte at the end of the
+	// argument vector has been overwritten, it continues reading the
+	// environment vector as part of the argument vector.
+
+	n, dstErr := dst.CopyOut(ctx, buf)
+	if dstErr != nil {
+		return int64(n), dstErr
+	}
+	return int64(n), copyErr
+}
diff --git a/pkg/sentry/fs/proc/fds.go b/pkg/sentry/fs/proc/fds.go
new file mode 100644
index 000000000..2eca9ac31
--- /dev/null
+++ b/pkg/sentry/fs/proc/fds.go
@@ -0,0 +1,258 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+	"sort"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// walkDescriptors finds the descriptor (file-flag pair) for the fd identified
+// by p, and calls the toInodeOperations callback with that descriptor.  This is a helper
+// method for implementing fs.InodeOperations.Lookup.
+func walkDescriptors(t *kernel.Task, p string, toInode func(*fs.File, kernel.FDFlags) *fs.Inode) (*fs.Inode, error) {
+	n, err := strconv.ParseUint(p, 10, 64)
+	if err != nil {
+		// Not found.
+		return nil, syserror.ENOENT
+	}
+
+	var file *fs.File
+	var flags kernel.FDFlags
+	t.WithMuLocked(func(t *kernel.Task) {
+		if fdm := t.FDMap(); fdm != nil {
+			file, flags = fdm.GetDescriptor(kdefs.FD(n))
+		}
+	})
+	if file == nil {
+		return nil, syserror.ENOENT
+	}
+	return toInode(file, flags), nil
+}
+
+// readDescriptors reads fds in the task starting at offset, and calls the
+// toDentAttr callback for each to get a DentAttr, which it then emits. This is
+// a helper for implementing fs.InodeOperations.Readdir.
+func readDescriptors(t *kernel.Task, c *fs.DirCtx, offset int, toDentAttr func(int) fs.DentAttr) (int, error) {
+	var fds kernel.FDs
+	t.WithMuLocked(func(t *kernel.Task) {
+		if fdm := t.FDMap(); fdm != nil {
+			fds = fdm.GetFDs()
+		}
+	})
+
+	fdInts := make([]int, 0, len(fds))
+	for _, fd := range fds {
+		fdInts = append(fdInts, int(fd))
+	}
+
+	// Find the fd to start at.
+	idx := sort.SearchInts(fdInts, offset)
+	if idx == len(fdInts) {
+		return offset, nil
+	}
+	fdInts = fdInts[idx:]
+
+	var fd int
+	for _, fd = range fdInts {
+		name := strconv.FormatUint(uint64(fd), 10)
+		if err := c.DirEmit(name, toDentAttr(fd)); err != nil {
+			// Returned offset is the next fd to serialize.
+			return fd, err
+		}
+	}
+	// We serialized them all.  Next offset should be higher than last
+	// serialized fd.
+	return fd + 1, nil
+}
+
+// fd is a single file in /proc/TID/fd/.
+type fd struct {
+	ramfs.Symlink
+	*fs.File
+}
+
+// newFD returns a new fd based on an existing file.
+//
+// This inherits one reference to the file.
+func newFd(t *kernel.Task, f *fs.File, msrc *fs.MountSource) *fs.Inode {
+	fd := &fd{File: f}
+	// RootOwner by default, is overridden in UnstableAttr()
+	fd.InitSymlink(t, fs.RootOwner, "")
+	return newFile(fd, msrc, fs.Symlink, t)
+}
+
+// GetFile returns the fs.File backing this fd.  The dirent and flags
+// arguments are ignored.
+func (f *fd) GetFile(context.Context, *fs.Dirent, fs.FileFlags) (*fs.File, error) {
+	// Take a reference on the fs.File.
+	f.File.IncRef()
+	return f.File, nil
+}
+
+// Readlink returns the current target.
+func (f *fd) Readlink(ctx context.Context, _ *fs.Inode) (string, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+	n, _ := f.Dirent.FullName(root)
+	return n, nil
+}
+
+// Getlink implements fs.InodeOperations.Getlink.
+func (f *fd) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	f.Dirent.IncRef()
+	return f.Dirent, nil
+}
+
+// Truncate is ignored.
+func (f *fd) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
+
+// Close releases the reference on the file.
+func (f *fd) Close() error {
+	f.DecRef()
+	return nil
+}
+
+// fdDir implements /proc/TID/fd.
+type fdDir struct {
+	ramfs.Dir
+
+	// We hold a reference on the task's fdmap but only keep an indirect
+	// task pointer to avoid Dirent loading circularity caused by fdmap's
+	// potential back pointers into the dirent tree.
+	t *kernel.Task
+}
+
+// newFdDir creates a new fdDir.
+func newFdDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	f := &fdDir{t: t}
+	f.InitDir(t, nil, fs.RootOwner, fs.FilePermissions{User: fs.PermMask{Read: true, Execute: true}})
+	return newFile(f, msrc, fs.SpecialDirectory, t)
+}
+
+// Check implements InodeOperations.Check.
+//
+// This is to match Linux, which uses a special permission handler to guarantee
+// that a process can still access /proc/self/fd after it has executed
+// setuid. See fs/proc/fd.c:proc_fd_permission.
+func (f *fdDir) Check(ctx context.Context, inode *fs.Inode, req fs.PermMask) bool {
+	if fs.ContextCanAccessFile(ctx, inode, req) {
+		return true
+	}
+	if t := kernel.TaskFromContext(ctx); t != nil {
+		// Allow access if the task trying to access it is in the
+		// thread group corresponding to this directory.
+		//
+		// N.B. Technically, in Linux 3.11, this compares what would be
+		// the equivalent of task pointers. However, this was fixed
+		// later in 54708d2858e7 ("proc: actually make
+		// proc_fd_permission() thread-friendly").
+		if f.t.ThreadGroup() == t.ThreadGroup() {
+			return true
+		}
+	}
+	return false
+}
+
+// Lookup loads an Inode in /proc/TID/fd into a Dirent.
+func (f *fdDir) Lookup(ctx context.Context, dir *fs.Inode, p string) (*fs.Dirent, error) {
+	n, err := walkDescriptors(f.t, p, func(file *fs.File, _ kernel.FDFlags) *fs.Inode {
+		return newFd(f.t, file, dir.MountSource)
+	})
+	if err != nil {
+		return nil, err
+	}
+	return fs.NewDirent(n, p), nil
+}
+
+// DeprecatedReaddir lists fds in /proc/TID/fd.
+func (f *fdDir) DeprecatedReaddir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	return readDescriptors(f.t, dirCtx, offset, func(fd int) fs.DentAttr {
+		return fs.GenericDentAttr(fs.Symlink, device.ProcDevice)
+	})
+}
+
+// fdInfo is a single file in /proc/TID/fdinfo/.
+type fdInfo struct {
+	ramfs.File
+
+	flags kernel.FDFlags
+}
+
+// newFdInfo returns a new fdInfo based on an existing file.
+func newFdInfo(t *kernel.Task, _ *fs.File, flags kernel.FDFlags, msrc *fs.MountSource) *fs.Inode {
+	fdi := &fdInfo{flags: flags}
+	fdi.InitFile(t, fs.RootOwner, fs.FilePermissions{User: fs.PermMask{Read: true}})
+	// TODO: Get pos, locks, and other data.  For now we only
+	// have flags.
+	// See https://www.kernel.org/doc/Documentation/filesystems/proc.txt
+	fdi.Append([]byte(fmt.Sprintf("flags: %08o\n", flags)))
+	return newFile(fdi, msrc, fs.SpecialFile, t)
+}
+
+// DeprecatedPwritev implements fs.HandleOperations.DeprecatedPwritev.
+func (*fdInfo) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	return 0, ramfs.ErrInvalidOp
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (*fdInfo) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	return ramfs.ErrInvalidOp
+}
+
+// fdInfoDir implements /proc/TID/fdinfo.  It embeds an fdDir, but overrides
+// Lookup and Readdir.
+type fdInfoDir struct {
+	ramfs.Dir
+
+	t *kernel.Task
+}
+
+// newFdInfoDir creates a new fdInfoDir.
+func newFdInfoDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	fdid := &fdInfoDir{t: t}
+	fdid.InitDir(t, nil, fs.RootOwner, fs.FilePermsFromMode(0500))
+	return newFile(fdid, msrc, fs.SpecialDirectory, t)
+}
+
+// Lookup loads an fd in /proc/TID/fdinfo into a Dirent.
+func (fdid *fdInfoDir) Lookup(ctx context.Context, dir *fs.Inode, p string) (*fs.Dirent, error) {
+	n, err := walkDescriptors(fdid.t, p, func(file *fs.File, flags kernel.FDFlags) *fs.Inode {
+		return newFdInfo(fdid.t, file, flags, dir.MountSource)
+	})
+	if err != nil {
+		return nil, err
+	}
+	return fs.NewDirent(n, p), nil
+}
+
+// DeprecatedReaddir lists fds in /proc/TID/fdinfo.
+func (fdid *fdInfoDir) DeprecatedReaddir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	return readDescriptors(fdid.t, dirCtx, offset, func(fd int) fs.DentAttr {
+		return fs.GenericDentAttr(fs.RegularFile, device.ProcDevice)
+	})
+}
diff --git a/pkg/sentry/fs/proc/file.go b/pkg/sentry/fs/proc/file.go
new file mode 100644
index 000000000..9a433cdf8
--- /dev/null
+++ b/pkg/sentry/fs/proc/file.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type file struct {
+	fs.InodeOperations
+
+	// nodeType is the file type of this file.
+	nodeType fs.InodeType
+
+	// t is the associated kernel task that owns this file.
+	t *kernel.Task
+}
+
+func newFile(node fs.InodeOperations, msrc *fs.MountSource, nodeType fs.InodeType, t *kernel.Task) *fs.Inode {
+	iops := &file{node, nodeType, t}
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      nodeType,
+	}
+	return fs.NewInode(iops, msrc, sattr)
+}
+
+// UnstableAttr returns all attributes of this file.
+func (f *file) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	uattr, err := f.InodeOperations.UnstableAttr(ctx, inode)
+	if err != nil {
+		return fs.UnstableAttr{}, err
+	}
+	if f.t != nil {
+		uattr.Owner = fs.FileOwnerFromContext(f.t)
+	}
+	return uattr, nil
+}
diff --git a/pkg/sentry/fs/proc/filesystems.go b/pkg/sentry/fs/proc/filesystems.go
new file mode 100644
index 000000000..fe4de18ba
--- /dev/null
+++ b/pkg/sentry/fs/proc/filesystems.go
@@ -0,0 +1,55 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+)
+
+// filesystemsData backs /proc/filesystems.
+type filesystemsData struct{}
+
+// NeedsUpdate returns true on the first generation. The set of registered file
+// systems doesn't change so there's no need to generate SeqData more than once.
+func (*filesystemsData) NeedsUpdate(generation int64) bool {
+	return generation == 0
+}
+
+// ReadSeqFileData returns data for the SeqFile reader.
+// SeqData, the current generation and where in the file the handle corresponds to.
+func (*filesystemsData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	// We don't ever expect to see a non-nil SeqHandle.
+	if h != nil {
+		return nil, 0
+	}
+
+	// Generate the file contents.
+	var buf bytes.Buffer
+	for _, sys := range fs.GetFilesystems() {
+		nodev := "nodev"
+		if sys.Flags()&fs.FilesystemRequiresDev != 0 {
+			nodev = ""
+		}
+		// Matches the format of fs/filesystems.c:filesystems_proc_show.
+		fmt.Fprintf(&buf, "%s\t%s\n", nodev, sys.Name())
+	}
+
+	// Return the SeqData and advance the generation counter.
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*filesystemsData)(nil)}}, 1
+}
diff --git a/pkg/sentry/fs/proc/fs.go b/pkg/sentry/fs/proc/fs.go
new file mode 100644
index 000000000..072d00beb
--- /dev/null
+++ b/pkg/sentry/fs/proc/fs.go
@@ -0,0 +1,69 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// filesystem is a procfs.
+type filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&filesystem{})
+}
+
+// FilesystemName is the name underwhich the filesystem is registered.
+// Name matches fs/proc/root.c:proc_fs_type.name.
+const FilesystemName = "proc"
+
+// Name is the name of the file system.
+func (*filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount allows users to mount(2) this file system.
+func (*filesystem) AllowUserMount() bool {
+	return true
+}
+
+// Flags returns that there is nothing special about this file system.
+//
+// In Linux, proc returns FS_USERNS_VISIBLE | FS_USERNS_MOUNT, see fs/proc/root.c.
+func (*filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns the root of a procfs that can be positioned in the vfs.
+func (f *filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// device is always ignored.
+
+	// Parse generic comma-separated key=value options, this file system expects them.
+	options := fs.GenericMountSourceOptions(data)
+
+	// Proc options parsing checks for either a gid= or hidepid= and barfs on
+	// anything else, see fs/proc/root.c:proc_parse_options. Since we don't know
+	// what to do with gid= or hidepid=, we blow up if we get any options.
+	if len(options) > 0 {
+		return nil, fmt.Errorf("unsupported mount options: %v", options)
+	}
+
+	// Construct the procfs root. Since procfs files are all virtual, we
+	// never want them cached.
+	return New(ctx, fs.NewNonCachingMountSource(f, flags))
+}
diff --git a/pkg/sentry/fs/proc/loadavg.go b/pkg/sentry/fs/proc/loadavg.go
new file mode 100644
index 000000000..694cde656
--- /dev/null
+++ b/pkg/sentry/fs/proc/loadavg.go
@@ -0,0 +1,51 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+)
+
+// loadavgData backs /proc/loadavg.
+type loadavgData struct{}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*loadavgData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+func (d *loadavgData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+
+	// TODO: Include real data in fields.
+	// Column 1-3: CPU and IO utilization of the last 1, 5, and 10 minute periods.
+	// Column 4-5: currently running processes and the total number of processes.
+	// Column 6: the last process ID used.
+	fmt.Fprintf(&buf, "%.2f %.2f %.2f %d/%d %d\n", 0.00, 0.00, 0.00, 0, 0, 0)
+
+	return []seqfile.SeqData{
+		{
+			Buf:    buf.Bytes(),
+			Handle: (*loadavgData)(nil),
+		},
+	}, 0
+}
diff --git a/pkg/sentry/fs/proc/meminfo.go b/pkg/sentry/fs/proc/meminfo.go
new file mode 100644
index 000000000..489f796e5
--- /dev/null
+++ b/pkg/sentry/fs/proc/meminfo.go
@@ -0,0 +1,82 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// meminfoData backs /proc/meminfo.
+type meminfoData struct {
+	// k is the owning Kernel.
+	k *kernel.Kernel
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*meminfoData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (d *meminfoData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	mem := d.k.Platform.Memory()
+	mem.UpdateUsage()
+	snapshot, totalUsage := usage.MemoryAccounting.Copy()
+	totalSize := usage.TotalMemory(mem.TotalSize(), totalUsage)
+	anon := snapshot.Anonymous + snapshot.Tmpfs
+	file := snapshot.PageCache + snapshot.Mapped
+	// We don't actually have active/inactive LRUs, so just make up numbers.
+	activeFile := (file / 2) &^ (usermem.PageSize - 1)
+	inactiveFile := file - activeFile
+
+	var buf bytes.Buffer
+	fmt.Fprintf(&buf, "MemTotal:       %8d kB\n", totalSize/1024)
+	memFree := (totalSize - totalUsage) / 1024
+	// We use MemFree as MemAvailable because we don't swap.
+	// TODO: When reclaim is implemented the value of MemAvailable
+	// should change.
+	fmt.Fprintf(&buf, "MemFree:        %8d kB\n", memFree)
+	fmt.Fprintf(&buf, "MemAvailable:   %8d kB\n", memFree)
+	fmt.Fprintf(&buf, "Buffers:               0 kB\n") // memory usage by block devices
+	fmt.Fprintf(&buf, "Cached:         %8d kB\n", (file+snapshot.Tmpfs)/1024)
+	// Emulate a system with no swap, which disables inactivation of anon pages.
+	fmt.Fprintf(&buf, "SwapCache:             0 kB\n")
+	fmt.Fprintf(&buf, "Active:         %8d kB\n", (anon+activeFile)/1024)
+	fmt.Fprintf(&buf, "Inactive:       %8d kB\n", inactiveFile/1024)
+	fmt.Fprintf(&buf, "Active(anon):   %8d kB\n", anon/1024)
+	fmt.Fprintf(&buf, "Inactive(anon):        0 kB\n")
+	fmt.Fprintf(&buf, "Active(file):   %8d kB\n", activeFile/1024)
+	fmt.Fprintf(&buf, "Inactive(file): %8d kB\n", inactiveFile/1024)
+	fmt.Fprintf(&buf, "Unevictable:           0 kB\n") // TODO
+	fmt.Fprintf(&buf, "Mlocked:               0 kB\n") // TODO
+	fmt.Fprintf(&buf, "SwapTotal:             0 kB\n")
+	fmt.Fprintf(&buf, "SwapFree:              0 kB\n")
+	fmt.Fprintf(&buf, "Dirty:                 0 kB\n")
+	fmt.Fprintf(&buf, "Writeback:             0 kB\n")
+	fmt.Fprintf(&buf, "AnonPages:      %8d kB\n", anon/1024)
+	fmt.Fprintf(&buf, "Mapped:         %8d kB\n", file/1024) // doesn't count mapped tmpfs, which we don't know
+	fmt.Fprintf(&buf, "Shmem:          %8d kB\n", snapshot.Tmpfs/1024)
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*meminfoData)(nil)}}, 0
+}
diff --git a/pkg/sentry/fs/proc/mounts.go b/pkg/sentry/fs/proc/mounts.go
new file mode 100644
index 000000000..76092567d
--- /dev/null
+++ b/pkg/sentry/fs/proc/mounts.go
@@ -0,0 +1,176 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// forEachMountSource runs f for the process root mount and  each mount that is a
+// descendant of the root.
+func forEachMountSource(t *kernel.Task, fn func(string, *fs.MountSource)) {
+	// All mount points must be relative to the rootDir, and mounts outside
+	// will be excluded.
+	rootDir := t.FSContext().RootDirectory()
+	defer rootDir.DecRef()
+
+	if rootDir.Inode == nil {
+		panic(fmt.Sprintf("root dirent has nil inode: %+v", rootDir))
+	}
+	if rootDir.Inode.MountSource == nil {
+		panic(fmt.Sprintf("root dirent has nil mount: %+v", rootDir))
+	}
+
+	ms := append(rootDir.Inode.MountSource.Submounts(), rootDir.Inode.MountSource)
+	sort.Slice(ms, func(i, j int) bool {
+		return ms[i].ID() < ms[j].ID()
+	})
+	for _, m := range ms {
+		mountPath, desc := m.Root().FullName(rootDir)
+		if !desc {
+			// MountSources that are not descendants of the chroot jail are ignored.
+			continue
+		}
+
+		fn(mountPath, m)
+	}
+}
+
+// mountInfoFile is used to implement /proc/[pid]/mountinfo.
+type mountInfoFile struct {
+	t *kernel.Task
+}
+
+// NeedsUpdate implements SeqSource.NeedsUpdate.
+func (mif *mountInfoFile) NeedsUpdate(_ int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements SeqSource.ReadSeqFileData.
+func (mif *mountInfoFile) ReadSeqFileData(handle seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if handle != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+	forEachMountSource(mif.t, func(mountPath string, m *fs.MountSource) {
+		// Format:
+		// 36 35 98:0 /mnt1 /mnt2 rw,noatime master:1 - ext3 /dev/root rw,errors=continue
+		// (1)(2)(3)   (4)   (5)      (6)      (7)   (8) (9)   (10)         (11)
+
+		// (1) MountSource ID.
+		fmt.Fprintf(&buf, "%d ", m.ID())
+
+		// (2)  Parent ID (or this ID if there is no parent).
+		pID := m.ID()
+		if p := m.Parent(); p != nil {
+			pID = p.ID()
+		}
+		fmt.Fprintf(&buf, "%d ", pID)
+
+		// (3) Major:Minor device ID. We don't have a superblock, so we
+		// just use the root inode device number.
+		sa := m.Root().Inode.StableAttr
+		fmt.Fprintf(&buf, "%d:%d ", sa.DeviceFileMajor, sa.DeviceFileMinor)
+
+		// (4) Root: the pathname of the directory in the filesystem
+		// which forms the root of this mount.
+		//
+		// NOTE: This will always be "/" until we implement
+		// bind mounts.
+		fmt.Fprintf(&buf, "/ ")
+
+		// (5) Mount point (relative to process root).
+		fmt.Fprintf(&buf, "%s ", mountPath)
+
+		// (6) Mount options.
+		opts := "rw"
+		if m.Flags.ReadOnly {
+			opts = "ro"
+		}
+		if m.Flags.NoAtime {
+			opts += ",noatime"
+		}
+		fmt.Fprintf(&buf, "%s ", opts)
+
+		// (7) Optional fields: zero or more fields of the form "tag[:value]".
+		// (8) Separator: the end of the optional fields is marked by a single hyphen.
+		fmt.Fprintf(&buf, "- ")
+
+		// (9) Filesystem type.
+		name := "none"
+		if m.Filesystem != nil {
+			name = m.Filesystem.Name()
+		}
+		fmt.Fprintf(&buf, "%s ", name)
+
+		// (10) Mount source: filesystem-specific information or "none".
+		fmt.Fprintf(&buf, "none ")
+
+		// (11) Superblock options. Only "ro/rw" is supported for now,
+		// and is the same as the filesystem option.
+		fmt.Fprintf(&buf, "%s\n", opts)
+	})
+
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*mountInfoFile)(nil)}}, 0
+}
+
+// mountsFile is used to implement /proc/[pid]/mountinfo.
+type mountsFile struct {
+	t *kernel.Task
+}
+
+// NeedsUpdate implements SeqSource.NeedsUpdate.
+func (mf *mountsFile) NeedsUpdate(_ int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements SeqSource.ReadSeqFileData.
+func (mf *mountsFile) ReadSeqFileData(handle seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if handle != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+	forEachMountSource(mf.t, func(mountPath string, m *fs.MountSource) {
+		// Format (tab-separated):
+		// <special device or remote filesystem> <mount point> <filesystem type> <mount options> <needs dump> <fsck order>
+		//
+		// We use the filesystem name as the first field, since there
+		// is no real block device we can point to, and we also should
+		// not expose anything about the remote filesystem.
+		//
+		// Only ro/rw option is supported for now.
+		//
+		// The "needs dump"and fsck flags are always 0, which is allowed.
+		opts := "rw"
+		if m.Flags.ReadOnly {
+			opts = "ro"
+		}
+		name := "none"
+		if m.Filesystem != nil {
+			name = m.Filesystem.Name()
+		}
+		fmt.Fprintf(&buf, "%s\t%s\t%s\t%s\t%d\t%d\n", "none", mountPath, name, opts, 0, 0)
+	})
+
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*mountsFile)(nil)}}, 0
+}
diff --git a/pkg/sentry/fs/proc/net.go b/pkg/sentry/fs/proc/net.go
new file mode 100644
index 000000000..6e464857a
--- /dev/null
+++ b/pkg/sentry/fs/proc/net.go
@@ -0,0 +1,151 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+)
+
+// newNet creates a new proc net entry.
+func (p *proc) newNetDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	if s := p.k.NetworkStack(); s != nil && s.SupportsIPv6() {
+		d.AddChild(ctx, "dev", seqfile.NewSeqFileInode(ctx, &netDev{s: s}, msrc))
+		d.AddChild(ctx, "if_inet6", seqfile.NewSeqFileInode(ctx, &ifinet6{s: s}, msrc))
+	}
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
+
+// ifinet6 implements seqfile.SeqSource for /proc/net/if_inet6.
+type ifinet6 struct {
+	s inet.Stack `state:"nosave"` // S/R-FIXME
+}
+
+func (n *ifinet6) contents() []string {
+	var lines []string
+	nics := n.s.Interfaces()
+	for id, naddrs := range n.s.InterfaceAddrs() {
+		nic, ok := nics[id]
+		if !ok {
+			// NIC was added after NICNames was called. We'll just
+			// ignore it.
+			continue
+		}
+
+		for _, a := range naddrs {
+			// IPv6 only.
+			if a.Family != linux.AF_INET6 {
+				continue
+			}
+
+			// Fields:
+			// IPv6 address displayed in 32 hexadecimal chars without colons
+			// Netlink device number (interface index) in hexadecimal (use nic id)
+			// Prefix length in hexadecimal
+			// Scope value (use 0)
+			// Interface flags
+			// Device name
+			lines = append(lines, fmt.Sprintf("%032x %02x %02x %02x %02x %8s\n", a.Addr, id, a.PrefixLen, 0, a.Flags, nic.Name))
+		}
+	}
+	return lines
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*ifinet6) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (n *ifinet6) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	var data []seqfile.SeqData
+	for _, l := range n.contents() {
+		data = append(data, seqfile.SeqData{Buf: []byte(l), Handle: (*ifinet6)(nil)})
+	}
+
+	return data, 0
+}
+
+// netDev implements seqfile.SeqSource for /proc/net/dev.
+type netDev struct {
+	s inet.Stack `state:"nosave"` // S/R-FIXME
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (n *netDev) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData. See Linux's
+// net/core/net-procfs.c:dev_seq_show.
+func (n *netDev) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	interfaces := n.s.Interfaces()
+	contents := make([]string, 2, 2+len(interfaces))
+	// Add the table header. From net/core/net-procfs.c:dev_seq_show.
+	contents[0] = "Inter-|   Receive                                                |  Transmit\n"
+	contents[1] = " face |bytes    packets errs drop fifo frame compressed multicast|bytes    packets errs drop fifo colls carrier compressed\n"
+
+	for _, i := range interfaces {
+		// TODO: Collect stats from each inet.Stack
+		// implementation (hostinet, epsocket, and rpcinet).
+
+		// Implements the same format as
+		// net/core/net-procfs.c:dev_seq_printf_stats.
+		l := fmt.Sprintf("%6s: %7d %7d %4d %4d %4d %5d %10d %9d %8d %7d %4d %4d %4d %5d %7d %10d\n",
+			i.Name,
+			// Received
+			0, // bytes
+			0, // packets
+			0, // errors
+			0, // dropped
+			0, // fifo
+			0, // frame
+			0, // compressed
+			0, // multicast
+			// Transmitted
+			0, // bytes
+			0, // packets
+			0, // errors
+			0, // dropped
+			0, // fifo
+			0, // frame
+			0, // compressed
+			0) // multicast
+		contents = append(contents, l)
+	}
+
+	var data []seqfile.SeqData
+	for _, l := range contents {
+		data = append(data, seqfile.SeqData{Buf: []byte(l), Handle: (*ifinet6)(nil)})
+	}
+
+	return data, 0
+}
diff --git a/pkg/sentry/fs/proc/net_test.go b/pkg/sentry/fs/proc/net_test.go
new file mode 100644
index 000000000..a31a20494
--- /dev/null
+++ b/pkg/sentry/fs/proc/net_test.go
@@ -0,0 +1,74 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+)
+
+func newIPv6TestStack() *inet.TestStack {
+	s := inet.NewTestStack()
+	s.SupportsIPv6Flag = true
+	return s
+}
+
+func TestIfinet6NoAddresses(t *testing.T) {
+	n := &ifinet6{s: newIPv6TestStack()}
+	if got := n.contents(); got != nil {
+		t.Errorf("Got n.contents() = %v, want = %v", got, nil)
+	}
+}
+
+func TestIfinet6(t *testing.T) {
+	s := newIPv6TestStack()
+	s.InterfacesMap[1] = inet.Interface{Name: "eth0"}
+	s.InterfaceAddrsMap[1] = []inet.InterfaceAddr{
+		{
+			Family:    linux.AF_INET6,
+			PrefixLen: 128,
+			Addr:      []byte("\x00\x01\x02\x03\x04\x05\x06\x07\x08\x09\x0a\x0b\x0c\x0d\x0e\x0f"),
+		},
+	}
+	s.InterfacesMap[2] = inet.Interface{Name: "eth1"}
+	s.InterfaceAddrsMap[2] = []inet.InterfaceAddr{
+		{
+			Family:    linux.AF_INET6,
+			PrefixLen: 128,
+			Addr:      []byte("\x10\x11\x12\x13\x14\x15\x16\x17\x18\x19\x1a\x1b\x1c\x1d\x1e\x1f"),
+		},
+	}
+	want := map[string]struct{}{
+		"000102030405060708090a0b0c0d0e0f 01 80 00 00     eth0\n": {},
+		"101112131415161718191a1b1c1d1e1f 02 80 00 00     eth1\n": {},
+	}
+
+	n := &ifinet6{s: s}
+	contents := n.contents()
+	if len(contents) != len(want) {
+		t.Errorf("Got len(n.contents()) = %d, want = %d", len(contents), len(want))
+	}
+	got := map[string]struct{}{}
+	for _, l := range contents {
+		got[l] = struct{}{}
+	}
+
+	if !reflect.DeepEqual(got, want) {
+		t.Errorf("Got n.contents() = %v, want = %v", got, want)
+	}
+}
diff --git a/pkg/sentry/fs/proc/proc.go b/pkg/sentry/fs/proc/proc.go
new file mode 100644
index 000000000..459eb7e62
--- /dev/null
+++ b/pkg/sentry/fs/proc/proc.go
@@ -0,0 +1,182 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package proc implements a partial in-memory file system for profs.
+package proc
+
+import (
+	"fmt"
+	"sort"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// proc is a root proc node.
+type proc struct {
+	ramfs.Dir
+
+	// k is the Kernel containing this proc node.
+	k *kernel.Kernel
+
+	// pidns is the PID namespace of the task that mounted the proc filesystem
+	// that this node represents.
+	pidns *kernel.PIDNamespace
+}
+
+// New returns the root node of a partial simple procfs.
+func New(ctx context.Context, msrc *fs.MountSource) (*fs.Inode, error) {
+	k := kernel.KernelFromContext(ctx)
+	if k == nil {
+		return nil, fmt.Errorf("procfs requires a kernel")
+	}
+	pidns := kernel.PIDNamespaceFromContext(ctx)
+	if pidns == nil {
+		return nil, fmt.Errorf("procfs requires a PID namespace")
+	}
+
+	p := &proc{k: k, pidns: pidns}
+	p.InitDir(ctx, map[string]*fs.Inode{
+		// Note that these are just the static members. There are
+		// dynamic members populated in Readdir and Lookup below.
+		"filesystems": seqfile.NewSeqFileInode(ctx, &filesystemsData{}, msrc),
+		"loadavg":     seqfile.NewSeqFileInode(ctx, &loadavgData{}, msrc),
+		"meminfo":     seqfile.NewSeqFileInode(ctx, &meminfoData{k}, msrc),
+		"mounts":      newMountsSymlink(ctx, msrc),
+		"stat":        seqfile.NewSeqFileInode(ctx, &statData{k}, msrc),
+		"version":     seqfile.NewSeqFileInode(ctx, &versionData{k}, msrc),
+	}, fs.RootOwner, fs.FilePermsFromMode(0555))
+
+	p.AddChild(ctx, "cpuinfo", p.newCPUInfo(ctx, msrc))
+	p.AddChild(ctx, "uptime", p.newUptime(ctx, msrc))
+
+	return newFile(p, msrc, fs.SpecialDirectory, nil), nil
+}
+
+// self is a magical link.
+type self struct {
+	ramfs.Symlink
+
+	pidns *kernel.PIDNamespace
+}
+
+// newSelf returns a new "self" node.
+func (p *proc) newSelf(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	s := &self{pidns: p.pidns}
+	s.InitSymlink(ctx, fs.RootOwner, "")
+	return newFile(s, msrc, fs.Symlink, nil)
+}
+
+// Readlink implements fs.InodeOperations.Readlink.
+func (s *self) Readlink(ctx context.Context, inode *fs.Inode) (string, error) {
+	if t := kernel.TaskFromContext(ctx); t != nil {
+		tgid := s.pidns.IDOfThreadGroup(t.ThreadGroup())
+		if tgid == 0 {
+			return "", ramfs.ErrNotFound
+		}
+		return strconv.FormatUint(uint64(tgid), 10), nil
+	}
+
+	// Who is reading this link?
+	return "", ramfs.ErrInvalidOp
+}
+
+// Lookup loads an Inode at name into a Dirent.
+func (p *proc) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) {
+	// Is it one of the static ones?
+	dirent, walkErr := p.Dir.Lookup(ctx, dir, name)
+	if walkErr == nil {
+		return dirent, nil
+	}
+
+	// Is it a dynamic element?
+	nfs := map[string]func() *fs.Inode{
+		"net":  func() *fs.Inode { return p.newNetDir(ctx, dir.MountSource) },
+		"self": func() *fs.Inode { return p.newSelf(ctx, dir.MountSource) },
+		"sys":  func() *fs.Inode { return p.newSysDir(ctx, dir.MountSource) },
+	}
+	if nf, ok := nfs[name]; ok {
+		return fs.NewDirent(nf(), name), nil
+	}
+
+	// Try to lookup a corresponding task.
+	tid, err := strconv.ParseUint(name, 10, 64)
+	if err != nil {
+		// Ignore the parse error and return the original.
+		return nil, walkErr
+	}
+
+	// Grab the other task.
+	otherTask := p.pidns.TaskWithID(kernel.ThreadID(tid))
+	if otherTask == nil {
+		// Per above.
+		return nil, walkErr
+	}
+
+	// Wrap it in a taskDir.
+	td := newTaskDir(otherTask, dir.MountSource, p.pidns, true)
+	return fs.NewDirent(td, name), nil
+}
+
+// Readdir synthesizes proc contents.
+func (p *proc) DeprecatedReaddir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	// Serialize normal contents.
+	_, err := p.Dir.DeprecatedReaddir(ctx, dirCtx, offset)
+	if err != nil {
+		return offset, err
+	}
+
+	m := make(map[string]fs.DentAttr)
+	var names []string
+
+	// Add special files.
+	m["sys"] = fs.GenericDentAttr(fs.SpecialFile, device.ProcDevice)
+	names = append(names, "sys")
+
+	// Collect tasks.
+	// Per linux we only include it in directory listings if it's the leader.
+	// But for whatever crazy reason, you can still walk to the given node.
+	for _, tg := range p.pidns.ThreadGroups() {
+		if leader := tg.Leader(); leader != nil {
+			name := strconv.FormatUint(uint64(tg.ID()), 10)
+			m[name] = fs.GenericDentAttr(fs.SpecialDirectory, device.ProcDevice)
+			names = append(names, name)
+		}
+	}
+
+	if offset >= len(m) {
+		return offset, nil
+	}
+	sort.Strings(names)
+	names = names[offset:]
+	for _, name := range names {
+		if err := dirCtx.DirEmit(name, m[name]); err != nil {
+			return offset, err
+		}
+		offset++
+	}
+	return offset, err
+}
+
+// newMountsSymlink returns a symlink to "self/mounts"
+func newMountsSymlink(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	s := &ramfs.Symlink{}
+	s.InitSymlink(ctx, fs.RootOwner, "self/mounts")
+	return newFile(s, msrc, fs.Symlink, nil)
+}
diff --git a/pkg/sentry/fs/proc/seqfile/BUILD b/pkg/sentry/fs/proc/seqfile/BUILD
new file mode 100644
index 000000000..48dd25e5b
--- /dev/null
+++ b/pkg/sentry/fs/proc/seqfile/BUILD
@@ -0,0 +1,55 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "seqfile_state",
+    srcs = [
+        "seqfile.go",
+    ],
+    out = "seqfile_state.go",
+    package = "seqfile",
+)
+
+go_library(
+    name = "seqfile",
+    srcs = [
+        "seqfile.go",
+        "seqfile_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/proc/device",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+    ],
+)
+
+go_stateify(
+    name = "seqfile_test_state",
+    srcs = ["seqfile_test.go"],
+    out = "seqfile_test_state.go",
+    package = "seqfile",
+)
+
+go_test(
+    name = "seqfile_test",
+    size = "small",
+    srcs = [
+        "seqfile_test.go",
+        "seqfile_test_state.go",
+    ],
+    embed = [":seqfile"],
+    deps = [
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/ramfs/test",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/proc/seqfile/seqfile.go b/pkg/sentry/fs/proc/seqfile/seqfile.go
new file mode 100644
index 000000000..e37a85869
--- /dev/null
+++ b/pkg/sentry/fs/proc/seqfile/seqfile.go
@@ -0,0 +1,232 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package seqfile
+
+import (
+	"io"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// SeqHandle is a helper handle to seek in the file.
+type SeqHandle interface{}
+
+// SeqData holds the data for one unit in the file.
+type SeqData struct {
+	// The data to be returned to the user.
+	Buf []byte
+
+	// A seek handle used to find the next valid unit in ReadSeqFiledata.
+	Handle SeqHandle
+}
+
+// SeqSource is a data source for a SeqFile file.
+type SeqSource interface {
+	// NeedsUpdate returns true if the consumer of SeqData should call
+	// ReadSeqFileData again. Generation is the generation returned by
+	// ReadSeqFile or 0.
+	NeedsUpdate(generation int64) bool
+
+	// Returns a slice of SeqData ordered by unit and the current
+	// generation. The first entry in the slice is greater than the handle.
+	// If handle is nil then all known records are returned. Generation
+	// must always be greater than 0.
+	ReadSeqFileData(handle SeqHandle) ([]SeqData, int64)
+}
+
+// SeqGenerationCounter is a counter to keep track if the SeqSource should be
+// updated. SeqGenerationCounter is not thread-safe and should be protected
+// with a mutex.
+type SeqGenerationCounter struct {
+	// The generation that the SeqData is at.
+	generation int64
+}
+
+// SetGeneration sets the generation to the new value, be careful to not set it
+// to a value less than current.
+func (s *SeqGenerationCounter) SetGeneration(generation int64) {
+	s.generation = generation
+}
+
+// Update increments the current generation.
+func (s *SeqGenerationCounter) Update() {
+	s.generation++
+}
+
+// Generation returns the current generation counter.
+func (s *SeqGenerationCounter) Generation() int64 {
+	return s.generation
+}
+
+// IsCurrent returns whether the given generation is current or not.
+func (s *SeqGenerationCounter) IsCurrent(generation int64) bool {
+	return s.Generation() == generation
+}
+
+// SeqFile is used to provide dynamic files that can be ordered by record.
+type SeqFile struct {
+	ramfs.Entry
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	SeqSource
+
+	source     []SeqData
+	generation int64
+	lastRead   int64
+}
+
+// NewSeqFile returns a seqfile suitable for use by external consumers.
+func NewSeqFile(ctx context.Context, source SeqSource) *SeqFile {
+	s := &SeqFile{SeqSource: source}
+	s.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0444))
+	return s
+}
+
+// NewSeqFileInode returns an Inode with SeqFile InodeOperations.
+func NewSeqFileInode(ctx context.Context, source SeqSource, msrc *fs.MountSource) *fs.Inode {
+	iops := NewSeqFile(ctx, source)
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.SpecialFile,
+	}
+	return fs.NewInode(iops, msrc, sattr)
+}
+
+// UnstableAttr returns unstable attributes of the SeqFile.
+func (s *SeqFile) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	uattr, _ := s.Entry.UnstableAttr(ctx, inode)
+	uattr.ModificationTime = ktime.NowFromContext(ctx)
+	return uattr, nil
+}
+
+// findIndexAndOffset finds the unit that corresponds to a certain offset.
+// Returns the unit and the offset within the unit. If there are not enough
+// units len(data) and leftover offset is returned.
+func findIndexAndOffset(data []SeqData, offset int64) (int, int64) {
+	for i, buf := range data {
+		l := int64(len(buf.Buf))
+		if offset < l {
+			return i, offset
+		}
+		offset -= l
+	}
+	return len(data), offset
+}
+
+// DeprecatedPreadv reads from the file at the given offset.
+func (s *SeqFile) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.Entry.NotifyAccess(ctx)
+	defer func() { s.lastRead = offset }()
+
+	updated := false
+
+	// Try to find where we should start reading this file.
+	i, recordOffset := findIndexAndOffset(s.source, offset)
+	if i == len(s.source) {
+		// Ok, we're at EOF. Let's first check to see if there might be
+		// more data available to us. If there is more data, add it to
+		// the end and try reading again.
+		if !s.SeqSource.NeedsUpdate(s.generation) {
+			return 0, io.EOF
+		}
+		oldLen := len(s.source)
+		s.updateSourceLocked(len(s.source))
+		updated = true
+		// We know that we had consumed everything up until this point
+		// so we search in the new slice instead of starting over.
+		i, recordOffset = findIndexAndOffset(s.source[oldLen:], recordOffset)
+		i += oldLen
+		// i is at most the length of the slice which is
+		// len(s.source) - oldLen. So at most i will be equal to
+		// len(s.source).
+		if i == len(s.source) {
+			return 0, io.EOF
+		}
+	}
+
+	var done int64
+	// We're reading parts of a record, finish reading the current object
+	// before continuing on to the next. We don't refresh our data source
+	// before this record is completed.
+	if recordOffset != 0 {
+		n, err := dst.CopyOut(ctx, s.source[i].Buf[recordOffset:])
+		done += int64(n)
+		dst = dst.DropFirst(n)
+		if dst.NumBytes() == 0 || err != nil {
+			return done, err
+		}
+		i++
+	}
+
+	// Next/New unit, update the source file if necessary. Make an extra
+	// check to see if we've seeked backwards and if so always update our
+	// data source.
+	if !updated && (s.SeqSource.NeedsUpdate(s.generation) || s.lastRead > offset) {
+		s.updateSourceLocked(i)
+		// recordOffset is 0 here and we won't update records behind the
+		// current one so recordOffset is still 0 even though source
+		// just got updated. Just read the next record.
+	}
+
+	// Finish by reading all the available data.
+	for _, buf := range s.source[i:] {
+		n, err := dst.CopyOut(ctx, buf.Buf)
+		done += int64(n)
+		dst = dst.DropFirst(n)
+		if dst.NumBytes() == 0 || err != nil {
+			return done, err
+		}
+	}
+
+	// If the file shrank (entries not yet read were removed above)
+	// while we tried to read we can end up with nothing read.
+	if done == 0 && dst.NumBytes() != 0 {
+		return 0, io.EOF
+	}
+	return done, nil
+}
+
+// updateSourceLocked requires that s.mu is held.
+func (s *SeqFile) updateSourceLocked(record int) {
+	var h SeqHandle
+	if record == 0 {
+		h = nil
+	} else {
+		h = s.source[record-1].Handle
+	}
+	// Save what we have previously read.
+	s.source = s.source[:record]
+	var newSource []SeqData
+	newSource, s.generation = s.SeqSource.ReadSeqFileData(h)
+	s.source = append(s.source, newSource...)
+}
+
+// DeprecatedPwritev is always denied.
+func (*SeqFile) DeprecatedPwritev(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, ramfs.ErrDenied
+}
diff --git a/pkg/sentry/fs/proc/seqfile/seqfile_test.go b/pkg/sentry/fs/proc/seqfile/seqfile_test.go
new file mode 100644
index 000000000..0bf39ad82
--- /dev/null
+++ b/pkg/sentry/fs/proc/seqfile/seqfile_test.go
@@ -0,0 +1,272 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package seqfile
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	ramfstest "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs/test"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type seqTest struct {
+	actual []SeqData
+	update bool
+}
+
+func (s *seqTest) Init() {
+	var sq []SeqData
+	// Create some SeqData.
+	for i := 0; i < 10; i++ {
+		var b []byte
+		for j := 0; j < 10; j++ {
+			b = append(b, byte(i))
+		}
+		sq = append(sq, SeqData{
+			Buf:    b,
+			Handle: &testHandle{i: i},
+		})
+	}
+	s.actual = sq
+}
+
+// NeedsUpdate reports whether we need to update the data we've previously read.
+func (s *seqTest) NeedsUpdate(int64) bool {
+	return s.update
+}
+
+// ReadSeqFiledata returns a slice of SeqData which contains elements
+// greater than the handle.
+func (s *seqTest) ReadSeqFileData(handle SeqHandle) ([]SeqData, int64) {
+	if handle == nil {
+		return s.actual, 0
+	}
+	h := *handle.(*testHandle)
+	var ret []SeqData
+	for _, b := range s.actual {
+		// We want the next one.
+		h2 := *b.Handle.(*testHandle)
+		if h2.i > h.i {
+			ret = append(ret, b)
+		}
+	}
+	return ret, 0
+}
+
+// Flatten a slice of slices into one slice.
+func flatten(buf ...[]byte) []byte {
+	var flat []byte
+	for _, b := range buf {
+		flat = append(flat, b...)
+	}
+	return flat
+}
+
+type testHandle struct {
+	i int
+}
+
+type testTable struct {
+	offset         int64
+	readBufferSize int
+	expectedData   []byte
+	expectedError  error
+}
+
+func runTableTests(ctx context.Context, table []testTable, n fs.InodeOperations) error {
+	for _, tt := range table {
+		data := make([]byte, tt.readBufferSize)
+		resultLen, err := n.DeprecatedPreadv(ctx, usermem.BytesIOSequence(data), tt.offset)
+		if err != tt.expectedError {
+			return fmt.Errorf("t.Preadv(len: %v, offset: %v) (error) => %v expected %v", tt.readBufferSize, tt.offset, err, tt.expectedError)
+		}
+		expectedLen := int64(len(tt.expectedData))
+		if resultLen != expectedLen {
+			// We make this just an error so we wall through and print the data below.
+			return fmt.Errorf("t.Preadv(len: %v, offset: %v) (size) => %v expected %v", tt.readBufferSize, tt.offset, resultLen, expectedLen)
+		}
+		if !bytes.Equal(data[:expectedLen], tt.expectedData) {
+			return fmt.Errorf("t.Preadv(len: %v, offset: %v) (data) => %v expected %v", tt.readBufferSize, tt.offset, data[:expectedLen], tt.expectedData)
+		}
+	}
+	return nil
+}
+
+func TestSeqFile(t *testing.T) {
+	testSource := &seqTest{}
+	testSource.Init()
+
+	// Create a file that can be R/W.
+	m := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+	ctx := contexttest.Context(t)
+	contents := map[string]*fs.Inode{
+		"foo": NewSeqFileInode(ctx, testSource, m),
+	}
+	root := ramfstest.NewDir(ctx, contents, fs.FilePermsFromMode(0777))
+
+	// How about opening it?
+	inode := fs.NewInode(root, m, fs.StableAttr{Type: fs.Directory})
+	dirent2, err := root.Lookup(ctx, inode, "foo")
+	if err != nil {
+		t.Fatalf("failed to walk to foo for n2: %v", err)
+	}
+	n2 := dirent2.Inode.InodeOperations
+
+	// Writing?
+	if _, err := n2.DeprecatedPwritev(nil, usermem.BytesIOSequence([]byte("test")), 0); err == nil {
+		t.Fatalf("managed to write to n2: %v", err)
+	}
+
+	// How about reading?
+	dirent3, err := root.Lookup(ctx, inode, "foo")
+	if err != nil {
+		t.Fatalf("failed to walk to foo: %v", err)
+	}
+	n3 := dirent3.Inode.InodeOperations
+
+	if n2 != n3 {
+		t.Error("got n2 != n3, want same")
+	}
+
+	testSource.update = true
+
+	table := []testTable{
+		// Read past the end.
+		{100, 4, []byte{}, io.EOF},
+		{110, 4, []byte{}, io.EOF},
+		{200, 4, []byte{}, io.EOF},
+		// Read a truncated first line.
+		{0, 4, testSource.actual[0].Buf[:4], nil},
+		// Read the whole first line.
+		{0, 10, testSource.actual[0].Buf, nil},
+		// Read the whole first line + 5 bytes of second line.
+		{0, 15, flatten(testSource.actual[0].Buf, testSource.actual[1].Buf[:5]), nil},
+		// First 4 bytes of the second line.
+		{10, 4, testSource.actual[1].Buf[:4], nil},
+		// Read the two first lines.
+		{0, 20, flatten(testSource.actual[0].Buf, testSource.actual[1].Buf), nil},
+		// Read three lines.
+		{0, 30, flatten(testSource.actual[0].Buf, testSource.actual[1].Buf, testSource.actual[2].Buf), nil},
+		// Read everything, but use a bigger buffer than necessary.
+		{0, 150, flatten(testSource.actual[0].Buf, testSource.actual[1].Buf, testSource.actual[2].Buf, testSource.actual[3].Buf, testSource.actual[4].Buf, testSource.actual[5].Buf, testSource.actual[6].Buf, testSource.actual[7].Buf, testSource.actual[8].Buf, testSource.actual[9].Buf), nil},
+		// Read the last 3 bytes.
+		{97, 10, testSource.actual[9].Buf[7:], nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed with testSource.update = %v : %v", testSource.update, err)
+	}
+
+	// Disable updates and do it again.
+	testSource.update = false
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed with testSource.update = %v: %v", testSource.update, err)
+	}
+}
+
+// Test that we behave correctly when the file is updated.
+func TestSeqFileFileUpdated(t *testing.T) {
+	testSource := &seqTest{}
+	testSource.Init()
+	testSource.update = true
+
+	// Create a file that can be R/W.
+	m := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+	ctx := contexttest.Context(t)
+	contents := map[string]*fs.Inode{
+		"foo": NewSeqFileInode(ctx, testSource, m),
+	}
+	root := ramfstest.NewDir(ctx, contents, fs.FilePermsFromMode(0777))
+
+	// How about opening it?
+	inode := fs.NewInode(root, m, fs.StableAttr{Type: fs.Directory})
+	dirent2, err := root.Lookup(ctx, inode, "foo")
+	if err != nil {
+		t.Fatalf("failed to walk to foo for n2: %v", err)
+	}
+	n2 := dirent2.Inode.InodeOperations
+
+	table := []testTable{
+		{0, 16, flatten(testSource.actual[0].Buf, testSource.actual[1].Buf[:6]), nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed: %v", err)
+	}
+	// Delete the first entry.
+	cut := testSource.actual[0].Buf
+	testSource.actual = testSource.actual[1:]
+
+	table = []testTable{
+		// Try reading buffer 0 with an offset. This will not delete the old data.
+		{1, 5, cut[1:6], nil},
+		// Reset our file by reading at offset 0.
+		{0, 10, testSource.actual[0].Buf, nil},
+		{16, 14, flatten(testSource.actual[1].Buf[6:], testSource.actual[2].Buf), nil},
+		// Read the same data a second time.
+		{16, 14, flatten(testSource.actual[1].Buf[6:], testSource.actual[2].Buf), nil},
+		// Read the following two lines.
+		{30, 20, flatten(testSource.actual[3].Buf, testSource.actual[4].Buf), nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed after removing first entry: %v", err)
+	}
+
+	// Add a new duplicate line in the middle (6666...)
+	after := testSource.actual[5:]
+	testSource.actual = testSource.actual[:4]
+	// Note the list must be sorted.
+	testSource.actual = append(testSource.actual, after[0])
+	testSource.actual = append(testSource.actual, after...)
+
+	table = []testTable{
+		{50, 20, flatten(testSource.actual[4].Buf, testSource.actual[5].Buf), nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed after adding middle entry: %v", err)
+	}
+	// This will be used in a later test.
+	oldTestData := testSource.actual
+
+	// Delete everything.
+	testSource.actual = testSource.actual[:0]
+	table = []testTable{
+		{20, 20, []byte{}, io.EOF},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed after removing all entries: %v", err)
+	}
+	// Restore some of the data.
+	testSource.actual = oldTestData[:1]
+	table = []testTable{
+		{6, 20, testSource.actual[0].Buf[6:], nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed after adding first entry back: %v", err)
+	}
+
+	// Re-extend the data
+	testSource.actual = oldTestData
+	table = []testTable{
+		{30, 20, flatten(testSource.actual[3].Buf, testSource.actual[4].Buf), nil},
+	}
+	if err := runTableTests(ctx, table, n2); err != nil {
+		t.Errorf("runTableTest failed after extending testSource: %v", err)
+	}
+}
diff --git a/pkg/sentry/fs/proc/stat.go b/pkg/sentry/fs/proc/stat.go
new file mode 100644
index 000000000..dee836a05
--- /dev/null
+++ b/pkg/sentry/fs/proc/stat.go
@@ -0,0 +1,139 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// statData backs /proc/stat.
+type statData struct {
+	// k is the owning Kernel.
+	k *kernel.Kernel
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*statData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// cpuStats contains the breakdown of CPU time for /proc/stat.
+type cpuStats struct {
+	// user is time spent in userspace tasks with non-positive niceness.
+	user uint64
+
+	// nice is time spent in userspace tasks with positive niceness.
+	nice uint64
+
+	// system is time spent in non-interrupt kernel context.
+	system uint64
+
+	// idle is time spent idle.
+	idle uint64
+
+	// ioWait is time spent waiting for IO.
+	ioWait uint64
+
+	// irq is time spent in interrupt context.
+	irq uint64
+
+	// softirq is time spent in software interrupt context.
+	softirq uint64
+
+	// steal is involuntary wait time.
+	steal uint64
+
+	// guest is time spent in guests with non-positive niceness.
+	guest uint64
+
+	// guestNice is time spent in guests with positive niceness.
+	guestNice uint64
+}
+
+// String implements fmt.Stringer.
+func (c cpuStats) String() string {
+	return fmt.Sprintf("%d %d %d %d %d %d %d %d %d %d", c.user, c.nice, c.system, c.idle, c.ioWait, c.irq, c.softirq, c.steal, c.guest, c.guestNice)
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (s *statData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+
+	// TODO: We currently export only zero CPU stats. We could
+	// at least provide some aggregate stats.
+	var cpu cpuStats
+	fmt.Fprintf(&buf, "cpu  %s\n", cpu)
+
+	for c, max := uint(0), s.k.ApplicationCores(); c < max; c++ {
+		fmt.Fprintf(&buf, "cpu%d %s\n", c, cpu)
+	}
+
+	// The total number of interrupts is dependent on the CPUs and PCI
+	// devices on the system. See arch_probe_nr_irqs.
+	//
+	// Since we don't report real interrupt stats, just choose an arbitrary
+	// value from a representative VM.
+	const numInterrupts = 256
+
+	// The Kernel doesn't handle real interrupts, so report all zeroes.
+	// TODO: We could count page faults as #PF.
+	fmt.Fprintf(&buf, "intr 0") // total
+	for i := 0; i < numInterrupts; i++ {
+		fmt.Fprintf(&buf, " 0")
+	}
+	fmt.Fprintf(&buf, "\n")
+
+	// Total number of context switches.
+	// TODO: Count this.
+	fmt.Fprintf(&buf, "ctxt 0\n")
+
+	// CLOCK_REALTIME timestamp from boot, in seconds.
+	fmt.Fprintf(&buf, "btime %d\n", s.k.Timekeeper().BootTime().Seconds())
+
+	// Total number of clones.
+	// TODO: Count this.
+	fmt.Fprintf(&buf, "processes 0\n")
+
+	// Number of runnable tasks.
+	// TODO: Count this.
+	fmt.Fprintf(&buf, "procs_running 0\n")
+
+	// Number of tasks waiting on IO.
+	// TODO: Count this.
+	fmt.Fprintf(&buf, "procs_blocked 0\n")
+
+	// Number of each softirq handled.
+	fmt.Fprintf(&buf, "softirq 0") // total
+	for i := 0; i < linux.NumSoftIRQ; i++ {
+		fmt.Fprintf(&buf, " 0")
+	}
+	fmt.Fprintf(&buf, "\n")
+
+	return []seqfile.SeqData{
+		{
+			Buf:    buf.Bytes(),
+			Handle: (*statData)(nil),
+		},
+	}, 0
+}
diff --git a/pkg/sentry/fs/proc/sys.go b/pkg/sentry/fs/proc/sys.go
new file mode 100644
index 000000000..4323f3650
--- /dev/null
+++ b/pkg/sentry/fs/proc/sys.go
@@ -0,0 +1,117 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// hostname is a file containing the system hostname.
+type hostname struct {
+	ramfs.Entry
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (hostname) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	utsns := kernel.UTSNamespaceFromContext(ctx)
+	contents := []byte(utsns.HostName() + "\n")
+
+	if offset >= int64(len(contents)) {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOut(ctx, contents[offset:])
+	return int64(n), err
+}
+
+func (p *proc) newHostname(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	h := &hostname{}
+	h.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0444))
+	return newFile(h, msrc, fs.SpecialFile, nil)
+}
+
+// mmapMinAddrData backs /proc/sys/vm/mmap_min_addr.
+type mmapMinAddrData struct {
+	k *kernel.Kernel
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*mmapMinAddrData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (d *mmapMinAddrData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+	return []seqfile.SeqData{
+		{
+			Buf:    []byte(fmt.Sprintf("%d\n", d.k.Platform.MinUserAddress())),
+			Handle: (*mmapMinAddrData)(nil),
+		},
+	}, 0
+}
+
+type overcommitMemory struct{}
+
+func (*overcommitMemory) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.
+func (*overcommitMemory) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+	return []seqfile.SeqData{
+		{
+			Buf:    []byte("0\n"),
+			Handle: (*overcommitMemory)(nil),
+		},
+	}, 0
+}
+
+func (p *proc) newKernelDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	d.AddChild(ctx, "hostname", p.newHostname(ctx, msrc))
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
+
+func (p *proc) newVMDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	d.AddChild(ctx, "mmap_min_addr", seqfile.NewSeqFileInode(ctx, &mmapMinAddrData{p.k}, msrc))
+	d.AddChild(ctx, "overcommit_memory", seqfile.NewSeqFileInode(ctx, &overcommitMemory{}, msrc))
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
+
+func (p *proc) newSysDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	d.AddChild(ctx, "kernel", p.newKernelDir(ctx, msrc))
+	d.AddChild(ctx, "vm", p.newVMDir(ctx, msrc))
+	d.AddChild(ctx, "net", p.newSysNetDir(ctx, msrc))
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go
new file mode 100644
index 000000000..db44c95cb
--- /dev/null
+++ b/pkg/sentry/fs/proc/sys_net.go
@@ -0,0 +1,188 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type tcpMemDir int
+
+const (
+	tcpRMem tcpMemDir = iota
+	tcpWMem
+)
+
+type tcpMem struct {
+	ramfs.Entry
+	s    inet.Stack
+	size inet.TCPBufferSize
+	dir  tcpMemDir
+}
+
+func newTCPMem(s inet.Stack, size inet.TCPBufferSize, dir tcpMemDir) *tcpMem {
+	return &tcpMem{s: s, size: size, dir: dir}
+}
+
+func newTCPMemInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack, size inet.TCPBufferSize, dir tcpMemDir) *fs.Inode {
+	tm := newTCPMem(s, size, dir)
+	tm.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0644))
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.SpecialFile,
+	}
+	return fs.NewInode(tm, msrc, sattr)
+}
+
+// DeprecatedPreadv implements fs.InodeOperations.DeprecatedPreadv.
+func (m *tcpMem) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset != 0 {
+		return 0, io.EOF
+	}
+	s := fmt.Sprintf("%d\t%d\t%d\n", m.size.Min, m.size.Default, m.size.Max)
+	n, err := dst.CopyOut(ctx, []byte(s))
+	return int64(n), err
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (*tcpMem) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev.
+func (m *tcpMem) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+	src = src.TakeFirst(usermem.PageSize - 1)
+
+	buf := []int32{int32(m.size.Min), int32(m.size.Default), int32(m.size.Max)}
+	n, cperr := usermem.CopyInt32StringsInVec(ctx, src.IO, src.Addrs, buf, src.Opts)
+	size := inet.TCPBufferSize{
+		Min:     int(buf[0]),
+		Default: int(buf[1]),
+		Max:     int(buf[2]),
+	}
+	var err error
+	switch m.dir {
+	case tcpRMem:
+		err = m.s.SetTCPReceiveBufferSize(size)
+	case tcpWMem:
+		err = m.s.SetTCPSendBufferSize(size)
+	default:
+		panic(fmt.Sprintf("unknown tcpMem.dir: %v", m.dir))
+	}
+	if err != nil {
+		return n, err
+	}
+	return n, cperr
+}
+
+type tcpSack struct {
+	ramfs.Entry
+	s inet.Stack `state:"nosave"` // S/R-FIXME
+}
+
+func newTCPSackInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode {
+	ts := &tcpSack{s: s}
+	ts.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0644))
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.SpecialFile,
+	}
+	return fs.NewInode(ts, msrc, sattr)
+}
+
+func (s *tcpSack) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset != 0 {
+		return 0, io.EOF
+	}
+
+	sack, err := s.s.TCPSACKEnabled()
+	if err != nil {
+		return 0, err
+	}
+
+	val := "0\n"
+	if sack {
+		// Technically, this is not quite compatible with Linux. Linux
+		// stores these as an integer, so if you write "2" into
+		// tcp_sack, you should get 2 back. Tough luck.
+		val = "1\n"
+	}
+	n, err := dst.CopyOut(ctx, []byte(val))
+	return int64(n), err
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (*tcpSack) Truncate(context.Context, *fs.Inode, int64) error {
+	return nil
+}
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev.
+func (s *tcpSack) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+	src = src.TakeFirst(usermem.PageSize - 1)
+
+	var v int32
+	n, err := usermem.CopyInt32StringInVec(ctx, src.IO, src.Addrs, &v, src.Opts)
+	if err != nil {
+		return n, err
+	}
+	return n, s.s.SetTCPSACKEnabled(v != 0)
+}
+
+func newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+
+	// Add tcp_rmem.
+	if rs, err := s.TCPReceiveBufferSize(); err == nil {
+		d.AddChild(ctx, "tcp_rmem", newTCPMemInode(ctx, msrc, s, rs, tcpRMem))
+	}
+
+	// Add tcp_wmem.
+	if ss, err := s.TCPSendBufferSize(); err == nil {
+		d.AddChild(ctx, "tcp_wmem", newTCPMemInode(ctx, msrc, s, ss, tcpWMem))
+	}
+
+	// Add tcp_sack.
+	d.AddChild(ctx, "tcp_sack", newTCPSackInode(ctx, msrc, s))
+
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
+
+func (p *proc) newSysNetDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(ctx, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	if s := p.k.NetworkStack(); s != nil {
+		d.AddChild(ctx, "ipv4", newSysNetIPv4Dir(ctx, msrc, s))
+	}
+	return newFile(d, msrc, fs.SpecialDirectory, nil)
+}
diff --git a/pkg/sentry/fs/proc/sys_net_test.go b/pkg/sentry/fs/proc/sys_net_test.go
new file mode 100644
index 000000000..7ba392346
--- /dev/null
+++ b/pkg/sentry/fs/proc/sys_net_test.go
@@ -0,0 +1,121 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func TestQuerySendBufferSize(t *testing.T) {
+	ctx := context.Background()
+	s := inet.NewTestStack()
+	s.TCPSendBufSize = inet.TCPBufferSize{100, 200, 300}
+	tm := newTCPMem(s, s.TCPSendBufSize, tcpWMem)
+
+	buf := make([]byte, 100)
+	dst := usermem.BytesIOSequence(buf)
+	n, err := tm.DeprecatedPreadv(ctx, dst, 0)
+	if err != nil {
+		t.Fatalf("DeprecatedPreadv failed: %v", err)
+	}
+
+	if got, want := string(buf[:n]), "100\t200\t300\n"; got != want {
+		t.Fatalf("Bad string: got %v, want %v", got, want)
+	}
+}
+
+func TestQueryRecvBufferSize(t *testing.T) {
+	ctx := context.Background()
+	s := inet.NewTestStack()
+	s.TCPRecvBufSize = inet.TCPBufferSize{100, 200, 300}
+	tm := newTCPMem(s, s.TCPRecvBufSize, tcpRMem)
+
+	buf := make([]byte, 100)
+	dst := usermem.BytesIOSequence(buf)
+	n, err := tm.DeprecatedPreadv(ctx, dst, 0)
+	if err != nil {
+		t.Fatalf("DeprecatedPreadv failed: %v", err)
+	}
+
+	if got, want := string(buf[:n]), "100\t200\t300\n"; got != want {
+		t.Fatalf("Bad string: got %v, want %v", got, want)
+	}
+}
+
+var cases = []struct {
+	str     string
+	initial inet.TCPBufferSize
+	final   inet.TCPBufferSize
+}{
+	{
+		str:     "",
+		initial: inet.TCPBufferSize{1, 2, 3},
+		final:   inet.TCPBufferSize{1, 2, 3},
+	},
+	{
+		str:     "100\n",
+		initial: inet.TCPBufferSize{1, 100, 200},
+		final:   inet.TCPBufferSize{100, 100, 200},
+	},
+	{
+		str:     "100 200 300\n",
+		initial: inet.TCPBufferSize{1, 2, 3},
+		final:   inet.TCPBufferSize{100, 200, 300},
+	},
+}
+
+func TestConfigureSendBufferSize(t *testing.T) {
+	ctx := context.Background()
+	s := inet.NewTestStack()
+	for _, c := range cases {
+		s.TCPSendBufSize = c.initial
+		tm := newTCPMem(s, c.initial, tcpWMem)
+
+		// Write the values.
+		src := usermem.BytesIOSequence([]byte(c.str))
+		if n, err := tm.DeprecatedPwritev(ctx, src, 0); n != int64(len(c.str)) || err != nil {
+			t.Errorf("DeprecatedPwritev, case = %q: got (%d, %v), wanted (%d, nil)", c.str, n, err, len(c.str))
+		}
+
+		// Read the values from the stack and check them.
+		if s.TCPSendBufSize != c.final {
+			t.Errorf("TCPSendBufferSize, case = %q: got %v, wanted %v", c.str, s.TCPSendBufSize, c.final)
+		}
+	}
+}
+
+func TestConfigureRecvBufferSize(t *testing.T) {
+	ctx := context.Background()
+	s := inet.NewTestStack()
+	for _, c := range cases {
+		s.TCPRecvBufSize = c.initial
+		tm := newTCPMem(s, c.initial, tcpRMem)
+
+		// Write the values.
+		src := usermem.BytesIOSequence([]byte(c.str))
+		if n, err := tm.DeprecatedPwritev(ctx, src, 0); n != int64(len(c.str)) || err != nil {
+			t.Errorf("DeprecatedPwritev, case = %q: got (%d, %v), wanted (%d, nil)", c.str, n, err, len(c.str))
+		}
+
+		// Read the values from the stack and check them.
+		if s.TCPRecvBufSize != c.final {
+			t.Errorf("TCPRecvBufferSize, case = %q: got %v, wanted %v", c.str, s.TCPRecvBufSize, c.final)
+		}
+	}
+}
diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go
new file mode 100644
index 000000000..3e9a1e50e
--- /dev/null
+++ b/pkg/sentry/fs/proc/task.go
@@ -0,0 +1,567 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"sort"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// getTaskMM returns t's MemoryManager. If getTaskMM succeeds, the MemoryManager's
+// users count is incremented, and must be decremented by the caller when it is
+// no longer in use.
+func getTaskMM(t *kernel.Task) (*mm.MemoryManager, error) {
+	if t.ExitState() == kernel.TaskExitDead {
+		return nil, syserror.ESRCH
+	}
+	var m *mm.MemoryManager
+	t.WithMuLocked(func(t *kernel.Task) {
+		m = t.MemoryManager()
+	})
+	if m == nil || !m.IncUsers() {
+		return nil, io.EOF
+	}
+	return m, nil
+}
+
+// taskDir represents a task-level directory.
+type taskDir struct {
+	ramfs.Dir
+
+	// t is the associated kernel task that owns this file.
+	t *kernel.Task
+}
+
+// newTaskDir creates a new proc task entry.
+func newTaskDir(t *kernel.Task, msrc *fs.MountSource, pidns *kernel.PIDNamespace, showSubtasks bool) *fs.Inode {
+	d := &taskDir{t: t}
+	// TODO: Set EUID/EGID based on dumpability.
+	d.InitDir(t, map[string]*fs.Inode{
+		"auxv":    newAuxvec(t, msrc),
+		"cmdline": newExecArgFile(t, msrc, cmdlineExecArg),
+		"comm":    newComm(t, msrc),
+		"environ": newExecArgFile(t, msrc, environExecArg),
+		"exe":     newExe(t, msrc),
+		"fd":      newFdDir(t, msrc),
+		"fdinfo":  newFdInfoDir(t, msrc),
+		"gid_map": newGIDMap(t, msrc),
+		// TODO: This is incorrect for /proc/[pid]/task/[tid]/io, i.e. if
+		// showSubtasks is false:
+		// http://lxr.free-electrons.com/source/fs/proc/base.c?v=3.11#L2980
+		"io":        newIO(t, msrc),
+		"maps":      newMaps(t, msrc),
+		"mountinfo": seqfile.NewSeqFileInode(t, &mountInfoFile{t: t}, msrc),
+		"mounts":    seqfile.NewSeqFileInode(t, &mountsFile{t: t}, msrc),
+		"ns":        newNamespaceDir(t, msrc),
+		"stat":      newTaskStat(t, msrc, showSubtasks, pidns),
+		"status":    newStatus(t, msrc, pidns),
+		"uid_map":   newUIDMap(t, msrc),
+	}, fs.RootOwner, fs.FilePermsFromMode(0555))
+	if showSubtasks {
+		d.AddChild(t, "task", newSubtasks(t, msrc, pidns))
+	}
+	return newFile(d, msrc, fs.SpecialDirectory, t)
+}
+
+// subtasks represents a /proc/TID/task directory.
+type subtasks struct {
+	ramfs.Dir
+
+	t *kernel.Task
+
+	pidns *kernel.PIDNamespace
+}
+
+func newSubtasks(t *kernel.Task, msrc *fs.MountSource, pidns *kernel.PIDNamespace) *fs.Inode {
+	s := &subtasks{t: t, pidns: pidns}
+	s.InitDir(t, nil, fs.RootOwner, fs.FilePermsFromMode(0555))
+	return newFile(s, msrc, fs.SpecialDirectory, t)
+}
+
+// UnstableAttr returns unstable attributes of the subtasks.
+func (s *subtasks) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	uattr, err := s.Dir.UnstableAttr(ctx, inode)
+	if err != nil {
+		return fs.UnstableAttr{}, err
+	}
+	// We can't rely on ramfs' implementation because the task directories are
+	// generated dynamically.
+	uattr.Links = uint64(2 + s.t.ThreadGroup().Count())
+	return uattr, nil
+}
+
+// Lookup loads an Inode in a task's subtask directory into a Dirent.
+func (s *subtasks) Lookup(ctx context.Context, dir *fs.Inode, p string) (*fs.Dirent, error) {
+	tid, err := strconv.ParseUint(p, 10, 32)
+	if err != nil {
+		return nil, syserror.ENOENT
+	}
+
+	task := s.pidns.TaskWithID(kernel.ThreadID(tid))
+	if task == nil {
+		return nil, syserror.ENOENT
+	}
+	if task.ThreadGroup() != s.t.ThreadGroup() {
+		return nil, syserror.ENOENT
+	}
+
+	td := newTaskDir(task, dir.MountSource, s.pidns, false)
+	return fs.NewDirent(td, p), nil
+}
+
+// DeprecatedReaddir lists a task's subtask directory.
+func (s *subtasks) DeprecatedReaddir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	tasks := s.t.ThreadGroup().MemberIDs(s.pidns)
+	taskInts := make([]int, 0, len(tasks))
+	for _, tid := range tasks {
+		taskInts = append(taskInts, int(tid))
+	}
+
+	// Find the task to start at.
+	idx := sort.SearchInts(taskInts, offset)
+	if idx == len(taskInts) {
+		return offset, nil
+	}
+	taskInts = taskInts[idx:]
+
+	var tid int
+	for _, tid = range taskInts {
+		name := strconv.FormatUint(uint64(tid), 10)
+		attr := fs.GenericDentAttr(fs.SpecialDirectory, device.ProcDevice)
+		if err := dirCtx.DirEmit(name, attr); err != nil {
+			// Returned offset is next tid to serialize.
+			return tid, err
+		}
+	}
+	// We serialized them all.  Next offset should be higher than last
+	// serialized tid.
+	return tid + 1, nil
+}
+
+// exe is an fs.InodeOperations symlink for the /proc/PID/exe file.
+type exe struct {
+	ramfs.Symlink
+
+	t *kernel.Task
+}
+
+func newExe(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	exeSymlink := &exe{t: t}
+	exeSymlink.InitSymlink(t, fs.RootOwner, "")
+	return newFile(exeSymlink, msrc, fs.Symlink, t)
+}
+
+func (e *exe) executable() (d *fs.Dirent, err error) {
+	e.t.WithMuLocked(func(t *kernel.Task) {
+		mm := t.MemoryManager()
+		if mm == nil {
+			// TODO: Check shouldn't allow Readlink once the
+			// Task is zombied.
+			err = syserror.EACCES
+			return
+		}
+
+		// The MemoryManager may be destroyed, in which case
+		// MemoryManager.destroy will simply set the executable to nil
+		// (with locks held).
+		d = mm.Executable()
+		if d == nil {
+			err = syserror.ENOENT
+		}
+	})
+	return
+}
+
+// Readlink implements fs.InodeOperations.
+func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) {
+	if !kernel.ContextCanTrace(ctx, e.t, false) {
+		return "", syserror.EACCES
+	}
+
+	// Pull out the executable for /proc/TID/exe.
+	exec, err := e.executable()
+	if err != nil {
+		return "", err
+	}
+	defer exec.DecRef()
+
+	root := fs.RootFromContext(ctx)
+	if root == nil {
+		// This doesn't correspond to anything in Linux because the vfs is
+		// global there.
+		return "", syserror.EINVAL
+	}
+	defer root.DecRef()
+	n, _ := exec.FullName(root)
+	return n, nil
+}
+
+// namespaceFile represents a file in the namespacefs, such as the files in
+// /proc/<pid>/ns.
+type namespaceFile struct {
+	ramfs.Symlink
+
+	t *kernel.Task
+}
+
+func newNamespaceFile(t *kernel.Task, msrc *fs.MountSource, name string) *fs.Inode {
+	n := &namespaceFile{t: t}
+	n.InitSymlink(t, fs.RootOwner, "")
+
+	// TODO: Namespace symlinks should contain the namespace name and the
+	// inode number for the namespace instance, so for example user:[123456]. We
+	// currently fake the inode number by sticking the symlink inode in its
+	// place.
+	n.Target = fmt.Sprintf("%s:[%d]", name, device.ProcDevice.NextIno())
+
+	return newFile(n, msrc, fs.Symlink, t)
+}
+
+// Getlink implements fs.InodeOperations.Getlink.
+func (n *namespaceFile) Getlink(ctx context.Context, inode *fs.Inode) (*fs.Dirent, error) {
+	if !kernel.ContextCanTrace(ctx, n.t, false) {
+		return nil, syserror.EACCES
+	}
+
+	// Create a new regular file to fake the namespace file.
+	node := &ramfs.Entry{}
+	node.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0777))
+	sattr := fs.StableAttr{
+		DeviceID:  device.ProcDevice.DeviceID(),
+		InodeID:   device.ProcDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.RegularFile,
+	}
+	return fs.NewDirent(fs.NewInode(node, inode.MountSource, sattr), n.Symlink.Target), nil
+}
+
+func newNamespaceDir(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	d := &ramfs.Dir{}
+	d.InitDir(t, map[string]*fs.Inode{
+		"net":  newNamespaceFile(t, msrc, "net"),
+		"pid":  newNamespaceFile(t, msrc, "pid"),
+		"user": newNamespaceFile(t, msrc, "user"),
+	}, fs.RootOwner, fs.FilePermsFromMode(0511))
+	return newFile(d, msrc, fs.SpecialDirectory, t)
+}
+
+// mapsData implements seqfile.SeqSource for /proc/[pid]/maps.
+type mapsData struct {
+	t *kernel.Task
+}
+
+func newMaps(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	return newFile(seqfile.NewSeqFile(t, &mapsData{t}), msrc, fs.SpecialFile, t)
+}
+
+func (md *mapsData) mm() *mm.MemoryManager {
+	var tmm *mm.MemoryManager
+	md.t.WithMuLocked(func(t *kernel.Task) {
+		if mm := t.MemoryManager(); mm != nil {
+			// No additional reference is taken on mm here. This is safe
+			// because MemoryManager.destroy is required to leave the
+			// MemoryManager in a state where it's still usable as a SeqSource.
+			tmm = mm
+		}
+	})
+	return tmm
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (md *mapsData) NeedsUpdate(generation int64) bool {
+	if mm := md.mm(); mm != nil {
+		return mm.NeedsUpdate(generation)
+	}
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (md *mapsData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if mm := md.mm(); mm != nil {
+		return mm.ReadSeqFileData(md.t.AsyncContext(), h)
+	}
+	return []seqfile.SeqData{}, 0
+}
+
+type taskStatData struct {
+	t *kernel.Task
+
+	// If tgstats is true, accumulate fault stats (not implemented) and CPU
+	// time across all tasks in t's thread group.
+	tgstats bool
+
+	// pidns is the PID namespace associated with the proc filesystem that
+	// includes the file using this statData.
+	pidns *kernel.PIDNamespace
+}
+
+func newTaskStat(t *kernel.Task, msrc *fs.MountSource, showSubtasks bool, pidns *kernel.PIDNamespace) *fs.Inode {
+	return newFile(seqfile.NewSeqFile(t, &taskStatData{t, showSubtasks /* tgstats */, pidns}), msrc, fs.SpecialFile, t)
+}
+
+// NeedsUpdate returns whether the generation is old or not.
+func (s *taskStatData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData returns data for the SeqFile reader.
+// SeqData, the current generation and where in the file the handle corresponds to.
+func (s *taskStatData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+
+	fmt.Fprintf(&buf, "%d ", s.pidns.IDOfTask(s.t))
+	fmt.Fprintf(&buf, "(%s) ", s.t.Name())
+	fmt.Fprintf(&buf, "%c ", s.t.StateStatus()[0])
+	ppid := kernel.ThreadID(0)
+	if parent := s.t.Parent(); parent != nil {
+		ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup())
+	}
+	fmt.Fprintf(&buf, "%d ", ppid)
+	fmt.Fprintf(&buf, "%d ", s.pidns.IDOfProcessGroup(s.t.ThreadGroup().ProcessGroup()))
+	fmt.Fprintf(&buf, "%d ", s.pidns.IDOfSession(s.t.ThreadGroup().Session()))
+	fmt.Fprintf(&buf, "0 0 " /* tty_nr tpgid */)
+	fmt.Fprintf(&buf, "0 " /* flags */)
+	fmt.Fprintf(&buf, "0 0 0 0 " /* minflt cminflt majflt cmajflt */)
+	var cputime usage.CPUStats
+	if s.tgstats {
+		cputime = s.t.ThreadGroup().CPUStats()
+	} else {
+		cputime = s.t.CPUStats()
+	}
+	fmt.Fprintf(&buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime))
+	cputime = s.t.ThreadGroup().JoinedChildCPUStats()
+	fmt.Fprintf(&buf, "%d %d ", linux.ClockTFromDuration(cputime.UserTime), linux.ClockTFromDuration(cputime.SysTime))
+	fmt.Fprintf(&buf, "%d %d ", s.t.Priority(), s.t.Niceness())
+	fmt.Fprintf(&buf, "%d ", s.t.ThreadGroup().Count())
+	fmt.Fprintf(&buf, "0 0 " /* itrealvalue starttime */)
+	var vss, rss uint64
+	s.t.WithMuLocked(func(t *kernel.Task) {
+		if mm := t.MemoryManager(); mm != nil {
+			vss = mm.VirtualMemorySize()
+			rss = mm.ResidentSetSize()
+		}
+	})
+	fmt.Fprintf(&buf, "%d %d ", vss, rss/usermem.PageSize)
+	fmt.Fprintf(&buf, "0 0 0 0 0 0 " /* rsslim startcode endcode startstack kstkesp kstkeip */)
+	fmt.Fprintf(&buf, "0 0 0 0 0 " /* signal blocked sigignore sigcatch wchan */)
+	fmt.Fprintf(&buf, "0 0 " /* nswap cnswap */)
+	terminationSignal := linux.Signal(0)
+	if s.t == s.t.ThreadGroup().Leader() {
+		terminationSignal = s.t.ThreadGroup().TerminationSignal()
+	}
+	fmt.Fprintf(&buf, "%d ", terminationSignal)
+	fmt.Fprintf(&buf, "0 0 0 " /* processor rt_priority policy */)
+	fmt.Fprintf(&buf, "0 0 0 " /* delayacct_blkio_ticks guest_time cguest_time */)
+	fmt.Fprintf(&buf, "0 0 0 0 0 0 0 " /* start_data end_data start_brk arg_start arg_end env_start env_end */)
+	fmt.Fprintf(&buf, "0\n" /* exit_code */)
+
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*taskStatData)(nil)}}, 0
+}
+
+// statusData implements seqfile.SeqSource for /proc/[pid]/status.
+type statusData struct {
+	t     *kernel.Task
+	pidns *kernel.PIDNamespace
+}
+
+func newStatus(t *kernel.Task, msrc *fs.MountSource, pidns *kernel.PIDNamespace) *fs.Inode {
+	return newFile(seqfile.NewSeqFile(t, &statusData{t, pidns}), msrc, fs.SpecialFile, t)
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (s *statusData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (s *statusData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	var buf bytes.Buffer
+	fmt.Fprintf(&buf, "Name:\t%s\n", s.t.Name())
+	fmt.Fprintf(&buf, "State:\t%s\n", s.t.StateStatus())
+	fmt.Fprintf(&buf, "Tgid:\t%d\n", s.pidns.IDOfThreadGroup(s.t.ThreadGroup()))
+	fmt.Fprintf(&buf, "Pid:\t%d\n", s.pidns.IDOfTask(s.t))
+	ppid := kernel.ThreadID(0)
+	if parent := s.t.Parent(); parent != nil {
+		ppid = s.pidns.IDOfThreadGroup(parent.ThreadGroup())
+	}
+	fmt.Fprintf(&buf, "PPid:\t%d\n", ppid)
+	tpid := kernel.ThreadID(0)
+	if tracer := s.t.Tracer(); tracer != nil {
+		tpid = s.pidns.IDOfTask(tracer)
+	}
+	fmt.Fprintf(&buf, "TracerPid:\t%d\n", tpid)
+	var fds int
+	var vss, rss uint64
+	s.t.WithMuLocked(func(t *kernel.Task) {
+		if fdm := t.FDMap(); fdm != nil {
+			fds = fdm.Size()
+		}
+		if mm := t.MemoryManager(); mm != nil {
+			vss = mm.VirtualMemorySize()
+			rss = mm.ResidentSetSize()
+		}
+	})
+	fmt.Fprintf(&buf, "FDSize:\t%d\n", fds)
+	fmt.Fprintf(&buf, "VmSize:\t%d kB\n", vss>>10)
+	fmt.Fprintf(&buf, "VmRSS:\t%d kB\n", rss>>10)
+	fmt.Fprintf(&buf, "Threads:\t%d\n", s.t.ThreadGroup().Count())
+	creds := s.t.Credentials()
+	fmt.Fprintf(&buf, "CapInh:\t%016x\n", creds.InheritableCaps)
+	fmt.Fprintf(&buf, "CapPrm:\t%016x\n", creds.PermittedCaps)
+	fmt.Fprintf(&buf, "CapEff:\t%016x\n", creds.EffectiveCaps)
+	fmt.Fprintf(&buf, "CapBnd:\t%016x\n", creds.BoundingCaps)
+	fmt.Fprintf(&buf, "Seccomp:\t%d\n", s.t.SeccompMode())
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*statusData)(nil)}}, 0
+}
+
+// ioUsage is the /proc/<pid>/io and /proc/<pid>/task/<tid>/io data provider.
+type ioUsage interface {
+	// IOUsage returns the io usage data.
+	IOUsage() *usage.IO
+}
+
+type ioData struct {
+	ioUsage
+}
+
+func newIO(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	return newFile(seqfile.NewSeqFile(t, &ioData{t.ThreadGroup()}), msrc, fs.SpecialFile, t)
+}
+
+// NeedsUpdate returns whether the generation is old or not.
+func (i *ioData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData returns data for the SeqFile reader.
+// SeqData, the current generation and where in the file the handle corresponds to.
+func (i *ioData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	io := usage.IO{}
+	io.Accumulate(i.IOUsage())
+
+	var buf bytes.Buffer
+	fmt.Fprintf(&buf, "char: %d\n", io.CharsRead)
+	fmt.Fprintf(&buf, "wchar: %d\n", io.CharsWritten)
+	fmt.Fprintf(&buf, "syscr: %d\n", io.ReadSyscalls)
+	fmt.Fprintf(&buf, "syscw: %d\n", io.WriteSyscalls)
+	fmt.Fprintf(&buf, "read_bytes: %d\n", io.BytesRead)
+	fmt.Fprintf(&buf, "write_bytes: %d\n", io.BytesWritten)
+	fmt.Fprintf(&buf, "cancelled_write_bytes: %d\n", io.BytesWriteCancelled)
+
+	return []seqfile.SeqData{{Buf: buf.Bytes(), Handle: (*ioData)(nil)}}, 0
+}
+
+// comm is a file containing the command name for a task.
+//
+// On Linux, /proc/[pid]/comm is writable, and writing to the comm file changes
+// the thread name. We don't implement this yet as there are no known users of
+// this feature.
+type comm struct {
+	ramfs.Entry
+
+	t *kernel.Task
+}
+
+// newComm returns a new comm file.
+func newComm(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	c := &comm{t: t}
+	c.InitEntry(t, fs.RootOwner, fs.FilePermsFromMode(0444))
+	return newFile(c, msrc, fs.SpecialFile, t)
+}
+
+// DeprecatedPreadv reads the current command name.
+func (c *comm) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	buf := []byte(c.t.Name() + "\n")
+	if offset >= int64(len(buf)) {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOut(ctx, buf[offset:])
+	return int64(n), err
+}
+
+// auxvec is a file containing the auxiliary vector for a task.
+type auxvec struct {
+	ramfs.Entry
+
+	t *kernel.Task
+}
+
+// newAuxvec returns a new auxvec file.
+func newAuxvec(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	a := &auxvec{t: t}
+	a.InitEntry(t, fs.RootOwner, fs.FilePermsFromMode(0400))
+	return newFile(a, msrc, fs.SpecialFile, t)
+}
+
+// DeprecatedPreadv reads the current auxiliary vector.
+func (a *auxvec) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	m, err := getTaskMM(a.t)
+	if err != nil {
+		return 0, err
+	}
+	defer m.DecUsers(ctx)
+	auxv := m.Auxv()
+
+	// Space for buffer with AT_NULL (0) terminator at the end.
+	size := (len(auxv) + 1) * 16
+	if offset >= int64(size) {
+		return 0, io.EOF
+	}
+
+	buf := make([]byte, size)
+	for i, e := range auxv {
+		usermem.ByteOrder.PutUint64(buf[16*i:], e.Key)
+		usermem.ByteOrder.PutUint64(buf[16*i+8:], uint64(e.Value))
+	}
+
+	n, err := dst.CopyOut(ctx, buf[offset:])
+	return int64(n), err
+}
diff --git a/pkg/sentry/fs/proc/uid_gid_map.go b/pkg/sentry/fs/proc/uid_gid_map.go
new file mode 100644
index 000000000..a2a070bdd
--- /dev/null
+++ b/pkg/sentry/fs/proc/uid_gid_map.go
@@ -0,0 +1,152 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// An idMapSeqSource is a seqfile.SeqSource that returns UID or GID mappings
+// from a task's user namespace.
+type idMapSeqSource struct {
+	t    *kernel.Task
+	gids bool
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (imss *idMapSeqSource) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (imss *idMapSeqSource) ReadSeqFileData(handle seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	var start int
+	if handle != nil {
+		start = handle.(*idMapSeqHandle).value
+	}
+	var entries []auth.IDMapEntry
+	if imss.gids {
+		entries = imss.t.UserNamespace().GIDMap()
+	} else {
+		entries = imss.t.UserNamespace().UIDMap()
+	}
+	var data []seqfile.SeqData
+	i := 1
+	for _, e := range entries {
+		if i > start {
+			data = append(data, seqfile.SeqData{
+				Buf:    idMapLineFromEntry(e),
+				Handle: &idMapSeqHandle{i},
+			})
+		}
+		i++
+	}
+	return data, 0
+}
+
+// TODO: Fix issue requiring idMapSeqHandle wrapping an int.
+type idMapSeqHandle struct {
+	value int
+}
+
+type idMapSeqFile struct {
+	seqfile.SeqFile
+}
+
+// newUIDMap returns a new uid_map file.
+func newUIDMap(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	return newIDMap(t, msrc, false /* gids */)
+}
+
+// newGIDMap returns a new gid_map file.
+func newGIDMap(t *kernel.Task, msrc *fs.MountSource) *fs.Inode {
+	return newIDMap(t, msrc, true /* gids */)
+}
+
+func newIDMap(t *kernel.Task, msrc *fs.MountSource, gids bool) *fs.Inode {
+	imsf := &idMapSeqFile{seqfile.SeqFile{SeqSource: &idMapSeqSource{
+		t:    t,
+		gids: gids,
+	}}}
+	imsf.InitEntry(t, fs.RootOwner, fs.FilePermsFromMode(0644))
+	return newFile(imsf, msrc, fs.SpecialFile, t)
+}
+
+func (imsf *idMapSeqFile) source() *idMapSeqSource {
+	return imsf.SeqFile.SeqSource.(*idMapSeqSource)
+}
+
+// "There is an (arbitrary) limit on the number of lines in the file. As at
+// Linux 3.18, the limit is five lines." - user_namespaces(7)
+const maxIDMapLines = 5
+
+// DeprecatedPwritev implements fs.InodeOperations.DeprecatedPwritev.
+func (imsf *idMapSeqFile) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	// "In addition, the number of bytes written to the file must be less than
+	// the system page size, and the write must be performed at the start of
+	// the file ..." - user_namespaces(7)
+	srclen := src.NumBytes()
+	if srclen >= usermem.PageSize || offset != 0 {
+		return 0, syserror.EINVAL
+	}
+	b := make([]byte, srclen)
+	if _, err := src.CopyIn(ctx, b); err != nil {
+		return 0, err
+	}
+	lines := bytes.SplitN(bytes.TrimSpace(b), []byte("\n"), maxIDMapLines+1)
+	if len(lines) > maxIDMapLines {
+		return 0, syserror.EINVAL
+	}
+	entries := make([]auth.IDMapEntry, len(lines))
+	for i, l := range lines {
+		e, err := idMapEntryFromLine(string(l))
+		if err != nil {
+			return 0, syserror.EINVAL
+		}
+		entries[i] = e
+	}
+	t := imsf.source().t
+	var err error
+	if imsf.source().gids {
+		err = t.UserNamespace().SetGIDMap(ctx, entries)
+	} else {
+		err = t.UserNamespace().SetUIDMap(ctx, entries)
+	}
+	if err != nil {
+		return 0, err
+	}
+	return int64(len(b)), nil
+}
+
+func idMapLineFromEntry(e auth.IDMapEntry) []byte {
+	var b bytes.Buffer
+	fmt.Fprintf(&b, "%10d %10d %10d\n", e.FirstID, e.FirstParentID, e.Length)
+	return b.Bytes()
+}
+
+func idMapEntryFromLine(line string) (auth.IDMapEntry, error) {
+	var e auth.IDMapEntry
+	_, err := fmt.Sscan(line, &e.FirstID, &e.FirstParentID, &e.Length)
+	return e, err
+}
diff --git a/pkg/sentry/fs/proc/uptime.go b/pkg/sentry/fs/proc/uptime.go
new file mode 100644
index 000000000..4679d5821
--- /dev/null
+++ b/pkg/sentry/fs/proc/uptime.go
@@ -0,0 +1,61 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// uptime is a file containing the system uptime.
+type uptime struct {
+	ramfs.Entry
+
+	// The "start time" of the sandbox.
+	startTime ktime.Time
+}
+
+// newUptime returns a new uptime file.
+func (p *proc) newUptime(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	u := &uptime{
+		startTime: ktime.NowFromContext(ctx),
+	}
+	u.InitEntry(ctx, fs.RootOwner, fs.FilePermsFromMode(0444))
+	return newFile(u, msrc, fs.SpecialFile, nil)
+}
+
+// DeprecatedPreadv reads the current uptime.
+func (u *uptime) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+
+	now := ktime.NowFromContext(ctx)
+	// Pretend that we've spent zero time sleeping (second number).
+	s := []byte(fmt.Sprintf("%.2f 0.00\n", now.Sub(u.startTime).Seconds()))
+	if offset >= int64(len(s)) {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOut(ctx, s[offset:])
+	return int64(n), err
+}
diff --git a/pkg/sentry/fs/proc/version.go b/pkg/sentry/fs/proc/version.go
new file mode 100644
index 000000000..df3040d37
--- /dev/null
+++ b/pkg/sentry/fs/proc/version.go
@@ -0,0 +1,75 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package proc
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// versionData backs /proc/version.
+type versionData struct {
+	// k is the owning Kernel.
+	k *kernel.Kernel
+}
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (*versionData) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData implements seqfile.SeqSource.ReadSeqFileData.
+func (v *versionData) ReadSeqFileData(h seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	if h != nil {
+		return nil, 0
+	}
+
+	init := v.k.GlobalInit()
+	if init == nil {
+		// Attempted to read before the init Task is created. This can
+		// only occur during startup, which should never need to read
+		// this file.
+		panic("Attempted to read version before initial Task is available")
+	}
+
+	// /proc/version takes the form:
+	//
+	// "SYSNAME version RELEASE (COMPILE_USER@COMPILE_HOST)
+	// (COMPILER_VERSION) VERSION"
+	//
+	// where:
+	// - SYSNAME, RELEASE, and VERSION are the same as returned by
+	// sys_utsname
+	// - COMPILE_USER is the user that build the kernel
+	// - COMPILE_HOST is the hostname of the machine on which the kernel
+	// was built
+	// - COMPILER_VERSION is the version reported by the building compiler
+	//
+	// Since we don't really want to expose build information to
+	// applications, those fields are omitted.
+	//
+	// FIXME: Using Version from the init task SyscallTable
+	// disregards the different version a task may have (e.g., in a uts
+	// namespace).
+	ver := init.Leader().SyscallTable().Version
+	return []seqfile.SeqData{
+		{
+			Buf:    []byte(fmt.Sprintf("%s version %s %s\n", ver.Sysname, ver.Release, ver.Version)),
+			Handle: (*versionData)(nil),
+		},
+	}, 0
+}
diff --git a/pkg/sentry/fs/ramfs/BUILD b/pkg/sentry/fs/ramfs/BUILD
new file mode 100644
index 000000000..663a1aeb9
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/BUILD
@@ -0,0 +1,62 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "ramfs_state",
+    srcs = [
+        "dir.go",
+        "file.go",
+        "ramfs.go",
+        "socket.go",
+        "symlink.go",
+    ],
+    out = "ramfs_state.go",
+    package = "ramfs",
+)
+
+go_library(
+    name = "ramfs",
+    srcs = [
+        "dir.go",
+        "file.go",
+        "ramfs.go",
+        "ramfs_state.go",
+        "socket.go",
+        "symlink.go",
+        "tree.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/amutex",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/secio",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "ramfs_test",
+    size = "small",
+    srcs = ["tree_test.go"],
+    embed = [":ramfs"],
+    deps = [
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+    ],
+)
diff --git a/pkg/sentry/fs/ramfs/dir.go b/pkg/sentry/fs/ramfs/dir.go
new file mode 100644
index 000000000..bf4cd8dfd
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/dir.go
@@ -0,0 +1,364 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// CreateOps represents operations to create different file types.
+type CreateOps struct {
+	// NewDir creates a new directory.
+	NewDir func(ctx context.Context, dir *fs.Inode, perms fs.FilePermissions) (*fs.Inode, error)
+
+	// NewFile creates a new file.
+	NewFile func(ctx context.Context, dir *fs.Inode, perms fs.FilePermissions) (*fs.Inode, error)
+
+	// NewSymlink creates a new symlink with permissions 0777.
+	NewSymlink func(ctx context.Context, dir *fs.Inode, target string) (*fs.Inode, error)
+
+	// NewBoundEndpoint creates a new socket.
+	NewBoundEndpoint func(ctx context.Context, dir *fs.Inode, ep unix.BoundEndpoint, perms fs.FilePermissions) (*fs.Inode, error)
+
+	// NewFifo creates a new fifo.
+	NewFifo func(ctx context.Context, dir *fs.Inode, perm fs.FilePermissions) (*fs.Inode, error)
+}
+
+// Dir represents a single directory in the filesystem.
+type Dir struct {
+	Entry
+
+	// CreateOps may be provided.
+	//
+	// These may only be modified during initialization (while the application
+	// is not running). No sychronization is performed when accessing these
+	// operations during syscalls.
+	*CreateOps `state:"nosave"`
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// children are inodes that are in this directory.  A reference is held
+	// on each inode while it is in the map.
+	children map[string]*fs.Inode
+
+	// dentryMap is a sortedDentryMap containing entries for all children.
+	// Its entries ar kept up-to-date with d.children.
+	dentryMap *fs.SortedDentryMap
+}
+
+// InitDir initializes a directory.
+func (d *Dir) InitDir(ctx context.Context, contents map[string]*fs.Inode, owner fs.FileOwner, perms fs.FilePermissions) {
+	d.InitEntry(ctx, owner, perms)
+	if contents == nil {
+		contents = make(map[string]*fs.Inode)
+	}
+	d.children = contents
+	// Build the entries map ourselves, rather than calling addChildLocked,
+	// because it will be faster.
+	entries := make(map[string]fs.DentAttr, len(contents))
+	for name, inode := range contents {
+		entries[name] = fs.DentAttr{
+			Type:    inode.StableAttr.Type,
+			InodeID: inode.StableAttr.InodeID,
+		}
+	}
+	d.dentryMap = fs.NewSortedDentryMap(entries)
+
+	// Directories have an extra link, corresponding to '.'.
+	d.AddLink()
+}
+
+// addChildLocked add the child inode, inheriting its reference.
+func (d *Dir) addChildLocked(name string, inode *fs.Inode) {
+	d.children[name] = inode
+	d.dentryMap.Add(name, fs.DentAttr{
+		Type:    inode.StableAttr.Type,
+		InodeID: inode.StableAttr.InodeID,
+	})
+
+	// If the child is a directory, increment this dir's link count,
+	// corresponding to '..' from the subdirectory.
+	if fs.IsDir(inode.StableAttr) {
+		d.AddLink()
+	}
+
+	// Given we're now adding this inode to the directory we must also
+	// increase its link count. Similiarly we decremented it in removeChildLocked.
+	inode.AddLink()
+}
+
+// AddChild adds a child to this dir.
+func (d *Dir) AddChild(ctx context.Context, name string, inode *fs.Inode) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	d.addChildLocked(name, inode)
+}
+
+// FindChild returns (child, true) if the directory contains name.
+func (d *Dir) FindChild(name string) (*fs.Inode, bool) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	child, ok := d.children[name]
+	return child, ok
+}
+
+// removeChildLocked attempts to remove an entry from this directory.
+// This Entry's mutex must be held. It returns the removed Inode.
+func (d *Dir) removeChildLocked(ctx context.Context, name string) (*fs.Inode, error) {
+	inode, ok := d.children[name]
+	if !ok {
+		return nil, ErrNotFound
+	}
+
+	delete(d.children, name)
+	d.dentryMap.Remove(name)
+	d.Entry.NotifyModification(ctx)
+
+	// If the child was a subdirectory, then we must decrement this dir's
+	// link count which was the child's ".." directory entry.
+	if fs.IsDir(inode.StableAttr) {
+		d.DropLink()
+	}
+
+	// Update ctime.
+	inode.NotifyStatusChange(ctx)
+
+	// Given we're now removing this inode to the directory we must also
+	// decrease its link count. Similiarly it is increased in addChildLocked.
+	inode.DropLink()
+
+	return inode, nil
+}
+
+// RemoveEntry attempts to remove an entry from this directory.
+func (d *Dir) RemoveEntry(ctx context.Context, name string) error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	inode, err := d.removeChildLocked(ctx, name)
+	if err != nil {
+		return err
+	}
+
+	// Remove our reference on the inode.
+	inode.DecRef()
+	return nil
+}
+
+// Remove removes the named non-directory.
+func (d *Dir) Remove(ctx context.Context, dir *fs.Inode, name string) error {
+	return d.RemoveEntry(ctx, name)
+}
+
+// RemoveDirectory removes the named directory.
+func (d *Dir) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	n, err := d.walkLocked(ctx, name)
+	if err != nil {
+		return err
+	}
+	dirCtx := &fs.DirCtx{}
+	if _, err := n.HandleOps().DeprecatedReaddir(ctx, dirCtx, 0); err != nil {
+		return err
+	}
+	if len(dirCtx.DentAttrs()) > 0 {
+		return ErrNotEmpty
+	}
+	inode, err := d.removeChildLocked(ctx, name)
+	if err != nil {
+		return err
+	}
+
+	// Remove our reference on the inode.
+	inode.DecRef()
+
+	return err
+}
+
+// Lookup loads an inode at p into a Dirent.
+func (d *Dir) Lookup(ctx context.Context, dir *fs.Inode, p string) (*fs.Dirent, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	inode, err := d.walkLocked(ctx, p)
+	if err != nil {
+		return nil, err
+	}
+
+	// Take a reference on the inode before returning it.  This reference
+	// is owned by the dirent we are about to create.
+	inode.IncRef()
+	return fs.NewDirent(inode, p), nil
+}
+
+// walkLocked must be called with this Entry's mutex held.
+func (d *Dir) walkLocked(ctx context.Context, p string) (*fs.Inode, error) {
+	d.Entry.NotifyAccess(ctx)
+
+	// Lookup a child node.
+	if inode, ok := d.children[p]; ok {
+		return inode, nil
+	}
+
+	// fs.InodeOperations.Lookup returns syserror.ENOENT if p
+	// does not exist.
+	return nil, syserror.ENOENT
+}
+
+// createInodeOperationsCommon creates a new child node at this dir by calling
+// makeInodeOperations. It is the common logic for creating a new child.
+func (d *Dir) createInodeOperationsCommon(ctx context.Context, name string, makeInodeOperations func() (*fs.Inode, error)) (*fs.Inode, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	if _, ok := d.children[name]; ok {
+		return nil, syscall.EEXIST
+	}
+
+	inode, err := makeInodeOperations()
+	if err != nil {
+		return nil, err
+	}
+
+	d.addChildLocked(name, inode)
+	d.Entry.NotifyModification(ctx)
+
+	return inode, nil
+}
+
+// Create creates a new Inode with the given name and returns its File.
+func (d *Dir) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perms fs.FilePermissions) (*fs.File, error) {
+	if d.CreateOps == nil || d.CreateOps.NewFile == nil {
+		return nil, ErrDenied
+	}
+
+	inode, err := d.createInodeOperationsCommon(ctx, name, func() (*fs.Inode, error) {
+		return d.NewFile(ctx, dir, perms)
+	})
+	if err != nil {
+		return nil, err
+	}
+
+	// Take an extra ref on inode, which will be owned by the dirent.
+	inode.IncRef()
+
+	// Create the Dirent and corresponding file.
+	created := fs.NewDirent(inode, name)
+	defer created.DecRef()
+	return created.Inode.GetFile(ctx, created, flags)
+}
+
+// CreateLink returns a new link.
+func (d *Dir) CreateLink(ctx context.Context, dir *fs.Inode, oldname, newname string) error {
+	if d.CreateOps == nil || d.CreateOps.NewSymlink == nil {
+		return ErrDenied
+	}
+	_, err := d.createInodeOperationsCommon(ctx, newname, func() (*fs.Inode, error) {
+		return d.NewSymlink(ctx, dir, oldname)
+	})
+	return err
+}
+
+// CreateHardLink creates a new hard link.
+func (d *Dir) CreateHardLink(ctx context.Context, dir *fs.Inode, target *fs.Inode, name string) error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	// Take an extra reference on the inode and add it to our children.
+	target.IncRef()
+
+	// The link count will be incremented in addChildLocked.
+	d.addChildLocked(name, target)
+	d.Entry.NotifyModification(ctx)
+
+	// Update ctime.
+	target.NotifyStatusChange(ctx)
+
+	return nil
+}
+
+// CreateDirectory returns a new subdirectory.
+func (d *Dir) CreateDirectory(ctx context.Context, dir *fs.Inode, name string, perms fs.FilePermissions) error {
+	if d.CreateOps == nil || d.CreateOps.NewDir == nil {
+		return ErrDenied
+	}
+	_, err := d.createInodeOperationsCommon(ctx, name, func() (*fs.Inode, error) {
+		return d.NewDir(ctx, dir, perms)
+	})
+	// TODO: Support updating status times, as those should be
+	// updated by links.
+	return err
+}
+
+// Bind implements fs.InodeOperations.Bind.
+func (d *Dir) Bind(ctx context.Context, dir *fs.Inode, name string, ep unix.BoundEndpoint, perms fs.FilePermissions) error {
+	if d.CreateOps == nil || d.CreateOps.NewBoundEndpoint == nil {
+		return ErrDenied
+	}
+	_, err := d.createInodeOperationsCommon(ctx, name, func() (*fs.Inode, error) {
+		return d.NewBoundEndpoint(ctx, dir, ep, perms)
+	})
+	if err == syscall.EEXIST {
+		return syscall.EADDRINUSE
+	}
+	return err
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo.
+func (d *Dir) CreateFifo(ctx context.Context, dir *fs.Inode, name string, perms fs.FilePermissions) error {
+	if d.CreateOps == nil || d.CreateOps.NewFifo == nil {
+		return ErrDenied
+	}
+	_, err := d.createInodeOperationsCommon(ctx, name, func() (*fs.Inode, error) {
+		return d.NewFifo(ctx, dir, perms)
+	})
+	return err
+}
+
+func (d *Dir) readdirLocked(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	// Serialize the entries in dentryMap.
+	n, err := fs.GenericReaddir(dirCtx, d.dentryMap)
+
+	// Touch the access time.
+	d.Entry.NotifyAccess(ctx)
+
+	return offset + n, err
+}
+
+// DeprecatedReaddir emits the entries contained in this directory.
+func (d *Dir) DeprecatedReaddir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	return d.readdirLocked(ctx, dirCtx, offset)
+}
+
+// DeprecatedPreadv always returns ErrIsDirectory
+func (*Dir) DeprecatedPreadv(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, ErrIsDirectory
+}
+
+// DeprecatedPwritev always returns ErrIsDirectory
+func (*Dir) DeprecatedPwritev(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, ErrIsDirectory
+}
diff --git a/pkg/sentry/fs/ramfs/file.go b/pkg/sentry/fs/ramfs/file.go
new file mode 100644
index 000000000..e8363c3e2
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/file.go
@@ -0,0 +1,148 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"io"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/secio"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// File represents a unique file.  It uses a simple byte slice as storage, and
+// thus should only be used for small files.
+//
+// A File is not mappable.
+type File struct {
+	Entry
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// data tracks backing data for the file.
+	data []byte
+}
+
+// InitFile initializes a file.
+func (f *File) InitFile(ctx context.Context, owner fs.FileOwner, perms fs.FilePermissions) {
+	f.InitEntry(ctx, owner, perms)
+}
+
+// UnstableAttr returns unstable attributes of this ramfs file.
+func (f *File) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	uattr, _ := f.Entry.UnstableAttr(ctx, inode)
+	uattr.Size = int64(len(f.data))
+	uattr.Usage = f.usageLocked()
+
+	return uattr, nil
+}
+
+// usageLocked returns the disk usage. Caller must hold f.mu.
+func (f *File) usageLocked() int64 {
+	return int64(len(f.data))
+}
+
+// Append appends the given data. This is for internal use.
+func (f *File) Append(data []byte) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.data = append(f.data, data...)
+}
+
+// Truncate truncates this node.
+func (f *File) Truncate(ctx context.Context, inode *fs.Inode, l int64) error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if l < int64(len(f.data)) {
+		// Remove excess bytes.
+		f.data = f.data[:l]
+		return nil
+	} else if l > int64(len(f.data)) {
+		// Create a new slice with size l, and copy f.data into it.
+		d := make([]byte, l)
+		copy(d, f.data)
+		f.data = d
+	}
+	f.Entry.NotifyModification(ctx)
+	return nil
+}
+
+// ReadAt implements io.ReaderAt.
+func (f *File) ReadAt(data []byte, offset int64) (int, error) {
+	if offset < 0 {
+		return 0, ErrInvalidOp
+	}
+	if offset >= int64(len(f.data)) {
+		return 0, io.EOF
+	}
+	n := copy(data, f.data[offset:])
+	// Did we read past the end?
+	if offset+int64(len(data)) >= int64(len(f.data)) {
+		return n, io.EOF
+	}
+	return n, nil
+}
+
+// DeprecatedPreadv reads into a collection of slices from a given offset.
+func (f *File) DeprecatedPreadv(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if offset >= int64(len(f.data)) {
+		return 0, io.EOF
+	}
+	n, err := dst.CopyOut(ctx, f.data[offset:])
+	if n > 0 {
+		f.Entry.NotifyAccess(ctx)
+	}
+	return int64(n), err
+}
+
+// WriteAt implements io.WriterAt.
+func (f *File) WriteAt(data []byte, offset int64) (int, error) {
+	if offset < 0 {
+		return 0, ErrInvalidOp
+	}
+	newLen := offset + int64(len(data))
+	if newLen < 0 {
+		// Overflow.
+		return 0, syserror.EINVAL
+	}
+	if newLen > int64(len(f.data)) {
+		// Copy f.data into new slice with expanded length.
+		d := make([]byte, newLen)
+		copy(d, f.data)
+		f.data = d
+	}
+	return copy(f.data[offset:], data), nil
+}
+
+// DeprecatedPwritev writes from a collection of slices at a given offset.
+func (f *File) DeprecatedPwritev(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	n, err := src.CopyInTo(ctx, safemem.FromIOWriter{secio.NewOffsetWriter(f, offset)})
+	if n > 0 {
+		f.Entry.NotifyModification(ctx)
+	}
+	return n, err
+}
diff --git a/pkg/sentry/fs/ramfs/ramfs.go b/pkg/sentry/fs/ramfs/ramfs.go
new file mode 100644
index 000000000..04f2d38de
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/ramfs.go
@@ -0,0 +1,433 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package ramfs implements an in-memory file system that can be associated with
+// any device.
+package ramfs
+
+import (
+	"errors"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+var (
+	// ErrInvalidOp indicates the operation is not valid.
+	ErrInvalidOp = errors.New("invalid operation")
+
+	// ErrDenied indicates the operation was denid.
+	ErrDenied = errors.New("operation denied")
+
+	// ErrNotFound indicates that a node was not found on a walk.
+	ErrNotFound = errors.New("node not found")
+
+	// ErrCrossDevice indicates a cross-device link or rename.
+	ErrCrossDevice = errors.New("can't link across filesystems")
+
+	// ErrIsDirectory indicates that the operation failed because
+	// the node is a directory.
+	ErrIsDirectory = errors.New("is a directory")
+
+	// ErrNotDirectory indicates that the operation failed because
+	// the node is a not directory.
+	ErrNotDirectory = errors.New("not a directory")
+
+	// ErrNotEmpty indicates that the operation failed because the
+	// directory is not empty.
+	ErrNotEmpty = errors.New("directory not empty")
+)
+
+// Entry represents common internal state for file and directory nodes.
+// This may be used by other packages to easily create ramfs files.
+type Entry struct {
+	waiter.AlwaysReady    `state:"nosave"`
+	fsutil.NoMappable     `state:"nosave"`
+	fsutil.NoopWriteOut   `state:"nosave"`
+	fsutil.InodeNotSocket `state:"nosave"`
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// unstable is unstable attributes.
+	unstable fs.UnstableAttr
+
+	// xattrs are the extended attributes of the Entry.
+	xattrs map[string][]byte
+}
+
+// InitEntry initializes an entry.
+func (e *Entry) InitEntry(ctx context.Context, owner fs.FileOwner, p fs.FilePermissions) {
+	e.InitEntryWithAttr(ctx, fs.WithCurrentTime(ctx, fs.UnstableAttr{
+		Owner: owner,
+		Perms: p,
+		// Always start unlinked.
+		Links: 0,
+	}))
+}
+
+// InitEntryWithAttr initializes an entry with a complete set of attributes.
+func (e *Entry) InitEntryWithAttr(ctx context.Context, uattr fs.UnstableAttr) {
+	e.unstable = uattr
+	e.xattrs = make(map[string][]byte)
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (e *Entry) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.unstable, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (*Entry) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// Getxattr implements fs.InodeOperations.Getxattr.
+func (e *Entry) Getxattr(inode *fs.Inode, name string) ([]byte, error) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if value, ok := e.xattrs[name]; ok {
+		return value, nil
+	}
+	return nil, syserror.ENOATTR
+}
+
+// Setxattr implements fs.InodeOperations.Setxattr.
+func (e *Entry) Setxattr(inode *fs.Inode, name string, value []byte) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.xattrs[name] = value
+	return nil
+}
+
+// Listxattr implements fs.InodeOperations.Listxattr.
+func (e *Entry) Listxattr(inode *fs.Inode) (map[string]struct{}, error) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	names := make(map[string]struct{}, len(e.xattrs))
+	for name := range e.xattrs {
+		names[name] = struct{}{}
+	}
+	return names, nil
+}
+
+// GetFile returns a fs.File backed by the dirent argument and flags.
+func (*Entry) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return fsutil.NewHandle(ctx, d, flags, d.Inode.HandleOps()), nil
+}
+
+// SetPermissions always sets the permissions.
+func (e *Entry) SetPermissions(ctx context.Context, inode *fs.Inode, p fs.FilePermissions) bool {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.unstable.Perms = p
+	e.unstable.StatusChangeTime = ktime.NowFromContext(ctx)
+	return true
+}
+
+// SetOwner always sets ownership.
+func (e *Entry) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	if owner.UID.Ok() {
+		e.unstable.Owner.UID = owner.UID
+	}
+	if owner.GID.Ok() {
+		e.unstable.Owner.GID = owner.GID
+	}
+	return nil
+}
+
+// SetTimestamps sets the timestamps.
+func (e *Entry) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	now := ktime.NowFromContext(ctx)
+	if !ts.ATimeOmit {
+		if ts.ATimeSetSystemTime {
+			e.unstable.AccessTime = now
+		} else {
+			e.unstable.AccessTime = ts.ATime
+		}
+	}
+	if !ts.MTimeOmit {
+		if ts.MTimeSetSystemTime {
+			e.unstable.ModificationTime = now
+		} else {
+			e.unstable.ModificationTime = ts.MTime
+		}
+	}
+	e.unstable.StatusChangeTime = now
+	return nil
+}
+
+// NotifyStatusChange updates the status change time (ctime).
+func (e *Entry) NotifyStatusChange(ctx context.Context) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.unstable.StatusChangeTime = ktime.NowFromContext(ctx)
+}
+
+// StatusChangeTime returns the last status change time for this node.
+func (e *Entry) StatusChangeTime() ktime.Time {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.unstable.StatusChangeTime
+}
+
+// NotifyModification updates the modification time and the status change time.
+func (e *Entry) NotifyModification(ctx context.Context) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	now := ktime.NowFromContext(ctx)
+	e.unstable.ModificationTime = now
+	e.unstable.StatusChangeTime = now
+}
+
+// ModificationTime returns the last modification time for this node.
+func (e *Entry) ModificationTime() ktime.Time {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.unstable.ModificationTime
+}
+
+// NotifyAccess updates the access time.
+func (e *Entry) NotifyAccess(ctx context.Context) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	now := ktime.NowFromContext(ctx)
+	e.unstable.AccessTime = now
+}
+
+// AccessTime returns the last access time for this node.
+func (e *Entry) AccessTime() ktime.Time {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.unstable.AccessTime
+}
+
+// Permissions returns permissions on this entry.
+func (e *Entry) Permissions() fs.FilePermissions {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	return e.unstable.Perms
+}
+
+// Lookup is not supported by default.
+func (*Entry) Lookup(context.Context, *fs.Inode, string) (*fs.Dirent, error) {
+	return nil, ErrInvalidOp
+}
+
+// Create is not supported by default.
+func (*Entry) Create(context.Context, *fs.Inode, string, fs.FileFlags, fs.FilePermissions) (*fs.File, error) {
+	return nil, ErrInvalidOp
+}
+
+// CreateLink is not supported by default.
+func (*Entry) CreateLink(context.Context, *fs.Inode, string, string) error {
+	return ErrInvalidOp
+}
+
+// CreateHardLink is not supported by default.
+func (*Entry) CreateHardLink(context.Context, *fs.Inode, *fs.Inode, string) error {
+	return ErrInvalidOp
+}
+
+// IsVirtual returns true.
+func (*Entry) IsVirtual() bool {
+	return true
+}
+
+// CreateDirectory is not supported by default.
+func (*Entry) CreateDirectory(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return ErrInvalidOp
+}
+
+// Bind is not supported by default.
+func (*Entry) Bind(context.Context, *fs.Inode, string, unix.BoundEndpoint, fs.FilePermissions) error {
+	return ErrInvalidOp
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo. CreateFifo is not supported by
+// default.
+func (*Entry) CreateFifo(context.Context, *fs.Inode, string, fs.FilePermissions) error {
+	return ErrInvalidOp
+}
+
+// Remove is not supported by default.
+func (*Entry) Remove(context.Context, *fs.Inode, string) error {
+	return ErrInvalidOp
+}
+
+// RemoveDirectory is not supported by default.
+func (*Entry) RemoveDirectory(context.Context, *fs.Inode, string) error {
+	return ErrInvalidOp
+}
+
+// StatFS always returns ENOSYS.
+func (*Entry) StatFS(context.Context) (fs.Info, error) {
+	return fs.Info{}, syscall.ENOSYS
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (e *Entry) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return Rename(ctx, oldParent.InodeOperations, oldName, newParent.InodeOperations, newName)
+}
+
+// Rename renames from a *ramfs.Dir to another *ramfs.Dir.
+func Rename(ctx context.Context, oldParent fs.InodeOperations, oldName string, newParent fs.InodeOperations, newName string) error {
+	op, ok := oldParent.(*Dir)
+	if !ok {
+		return ErrCrossDevice
+	}
+	np, ok := newParent.(*Dir)
+	if !ok {
+		return ErrCrossDevice
+	}
+
+	np.mu.Lock()
+	defer np.mu.Unlock()
+
+	// Check whether the ramfs entry to be replaced is a non-empty directory.
+	if replaced, ok := np.children[newName]; ok {
+		if fs.IsDir(replaced.StableAttr) {
+			// FIXME: simplify by pinning children of ramfs-backed directories
+			// in the Dirent tree: this allows us to generalize ramfs operations without
+			// relying on an implementation of Readdir (which may do anything, like require
+			// that the file be open ... which would be reasonable).
+			dirCtx := &fs.DirCtx{}
+			_, err := replaced.HandleOps().DeprecatedReaddir(ctx, dirCtx, 0)
+			if err != nil {
+				return err
+			}
+			attrs := dirCtx.DentAttrs()
+
+			// ramfs-backed directories should not contain "." and "..", but we do this
+			// just in case.
+			delete(attrs, ".")
+			delete(attrs, "..")
+
+			// If the directory to be replaced is not empty, reject the rename.
+			if len(attrs) != 0 {
+				return ErrNotEmpty
+			}
+		}
+	}
+
+	// Be careful, we may have already grabbed this mutex above.
+	if op != np {
+		op.mu.Lock()
+		defer op.mu.Unlock()
+	}
+
+	// Do the swap.
+	n := op.children[oldName]
+	op.removeChildLocked(ctx, oldName)
+	np.addChildLocked(newName, n)
+
+	// Update ctime.
+	n.NotifyStatusChange(ctx)
+
+	return nil
+}
+
+// Truncate is not supported by default.
+func (*Entry) Truncate(context.Context, *fs.Inode, int64) error {
+	return ErrInvalidOp
+}
+
+// Readlink always returns ENOLINK.
+func (*Entry) Readlink(context.Context, *fs.Inode) (string, error) {
+	return "", syscall.ENOLINK
+}
+
+// Getlink always returns ENOLINK.
+func (*Entry) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	return nil, syscall.ENOLINK
+}
+
+// Release is a no-op.
+func (e *Entry) Release(context.Context) {}
+
+// AddLink implements InodeOperationss.AddLink.
+func (e *Entry) AddLink() {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.unstable.Links++
+}
+
+// DropLink implements InodeOperationss.DropLink.
+func (e *Entry) DropLink() {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+	e.unstable.Links--
+}
+
+// DeprecatedReaddir is not supported by default.
+func (*Entry) DeprecatedReaddir(context.Context, *fs.DirCtx, int) (int, error) {
+	return 0, ErrNotDirectory
+}
+
+// DeprecatedPreadv always returns ErrInvalidOp.
+func (*Entry) DeprecatedPreadv(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, ErrInvalidOp
+}
+
+// DeprecatedPwritev always returns ErrInvalidOp.
+func (*Entry) DeprecatedPwritev(context.Context, usermem.IOSequence, int64) (int64, error) {
+	return 0, ErrInvalidOp
+}
+
+// DeprecatedFsync is a noop.
+func (*Entry) DeprecatedFsync() error {
+	// Ignore, this is in memory.
+	return nil
+}
+
+// DeprecatedFlush always returns nil.
+func (*Entry) DeprecatedFlush() error {
+	return nil
+}
+
+// DeprecatedMappable implements fs.InodeOperations.DeprecatedMappable.
+func (*Entry) DeprecatedMappable(context.Context, *fs.Inode) (memmap.Mappable, bool) {
+	return nil, false
+}
+
+func init() {
+	// Register ramfs errors.
+	syserror.AddErrorTranslation(ErrInvalidOp, syscall.EINVAL)
+	syserror.AddErrorTranslation(ErrDenied, syscall.EACCES)
+	syserror.AddErrorTranslation(ErrNotFound, syscall.ENOENT)
+	syserror.AddErrorTranslation(ErrCrossDevice, syscall.EXDEV)
+	syserror.AddErrorTranslation(ErrIsDirectory, syscall.EISDIR)
+	syserror.AddErrorTranslation(ErrNotDirectory, syscall.ENOTDIR)
+	syserror.AddErrorTranslation(ErrNotEmpty, syscall.ENOTEMPTY)
+}
diff --git a/pkg/sentry/fs/ramfs/socket.go b/pkg/sentry/fs/ramfs/socket.go
new file mode 100644
index 000000000..b0c79325f
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/socket.go
@@ -0,0 +1,42 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// Socket represents a socket.
+type Socket struct {
+	Entry
+
+	// ep is the bound endpoint.
+	ep unix.BoundEndpoint
+}
+
+// InitSocket initializes a socket.
+func (s *Socket) InitSocket(ctx context.Context, ep unix.BoundEndpoint, owner fs.FileOwner, perms fs.FilePermissions) {
+	s.InitEntry(ctx, owner, perms)
+	s.ep = ep
+}
+
+// BoundEndpoint returns the socket data.
+func (s *Socket) BoundEndpoint(*fs.Inode, string) unix.BoundEndpoint {
+	// ramfs only supports stored sentry internal sockets. Only gofer sockets
+	// care about the path argument.
+	return s.ep
+}
diff --git a/pkg/sentry/fs/ramfs/symlink.go b/pkg/sentry/fs/ramfs/symlink.go
new file mode 100644
index 000000000..9bbf78619
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/symlink.go
@@ -0,0 +1,72 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// Symlink represents a symlink.
+type Symlink struct {
+	Entry
+
+	mu sync.Mutex `state:"nosave"`
+
+	// Target is the symlink target.
+	Target string
+}
+
+// InitSymlink initializes a symlink, pointing to the given target.
+// A symlink is assumed to always have permissions 0777.
+func (s *Symlink) InitSymlink(ctx context.Context, owner fs.FileOwner, target string) {
+	s.InitEntry(ctx, owner, fs.FilePermsFromMode(0777))
+	s.Target = target
+}
+
+// UnstableAttr returns all attributes of this ramfs symlink.
+func (s *Symlink) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	uattr, _ := s.Entry.UnstableAttr(ctx, inode)
+	uattr.Size = int64(len(s.Target))
+	uattr.Usage = uattr.Size
+	return uattr, nil
+}
+
+// Check implements InodeOperations.Check.
+func (s *Symlink) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions on a symlink is always rejected.
+func (s *Symlink) SetPermissions(context.Context, *fs.Inode, fs.FilePermissions) bool {
+	return false
+}
+
+// Readlink reads the symlink value.
+func (s *Symlink) Readlink(ctx context.Context, _ *fs.Inode) (string, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	s.Entry.NotifyAccess(ctx)
+	return s.Target, nil
+}
+
+// Getlink returns ErrResolveViaReadlink, falling back to walking to the result
+// of Readlink().
+func (*Symlink) Getlink(context.Context, *fs.Inode) (*fs.Dirent, error) {
+	return nil, fs.ErrResolveViaReadlink
+}
diff --git a/pkg/sentry/fs/ramfs/test/BUILD b/pkg/sentry/fs/ramfs/test/BUILD
new file mode 100644
index 000000000..074b0f5ad
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/test/BUILD
@@ -0,0 +1,31 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "test_state",
+    srcs = [
+        "test.go",
+    ],
+    out = "test_state.go",
+    package = "test",
+)
+
+go_library(
+    name = "test",
+    testonly = 1,
+    srcs = [
+        "test.go",
+        "test_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs/test",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/state",
+    ],
+)
diff --git a/pkg/sentry/fs/ramfs/test/test.go b/pkg/sentry/fs/ramfs/test/test.go
new file mode 100644
index 000000000..fb669558f
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/test/test.go
@@ -0,0 +1,46 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package test provides a simple ramfs-based filesystem for use in testing.
+package test
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+)
+
+// Dir is a simple ramfs.Dir that supports save/restore as-is.
+type Dir struct {
+	ramfs.Dir
+}
+
+// NewDir returns a simple ramfs directory with the passed contents.
+func NewDir(ctx context.Context, contents map[string]*fs.Inode, perms fs.FilePermissions) *Dir {
+	d := &Dir{}
+	d.InitDir(ctx, contents, fs.RootOwner, perms)
+	return d
+}
+
+// File is a simple ramfs.File that supports save/restore as-is.
+type File struct {
+	ramfs.File
+}
+
+// NewFile returns a simple ramfs File.
+func NewFile(ctx context.Context, perms fs.FilePermissions) *File {
+	f := &File{}
+	f.InitFile(ctx, fs.RootOwner, perms)
+	return f
+}
diff --git a/pkg/sentry/fs/ramfs/tree.go b/pkg/sentry/fs/ramfs/tree.go
new file mode 100644
index 000000000..1fb335f74
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/tree.go
@@ -0,0 +1,71 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"fmt"
+	"path"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// MakeDirectoryTree constructs a ramfs tree of all directories containing
+// subdirs. Each element of subdir must be a clean path, and cannot be empty or
+// "/".
+func MakeDirectoryTree(ctx context.Context, msrc *fs.MountSource, subdirs []string) (*fs.Inode, error) {
+	root := emptyDir(ctx, msrc)
+	for _, subdir := range subdirs {
+		if path.Clean(subdir) != subdir {
+			return nil, fmt.Errorf("cannot add subdir at an unclean path: %q", subdir)
+		}
+		if subdir == "" || subdir == "/" {
+			return nil, fmt.Errorf("cannot add subdir at %q", subdir)
+		}
+		makeSubdir(ctx, msrc, root.InodeOperations.(*Dir), subdir)
+	}
+	return root, nil
+}
+
+// makeSubdir installs into root each component of subdir. The final component is
+// a *ramfs.Dir.
+func makeSubdir(ctx context.Context, msrc *fs.MountSource, root *Dir, subdir string) {
+	for _, c := range strings.Split(subdir, "/") {
+		if len(c) == 0 {
+			continue
+		}
+		child, ok := root.FindChild(c)
+		if !ok {
+			child = emptyDir(ctx, msrc)
+			root.AddChild(ctx, c, child)
+		}
+		root = child.InodeOperations.(*Dir)
+	}
+}
+
+// emptyDir returns an empty *ramfs.Dir that is traversable but not writable.
+func emptyDir(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	dir := &Dir{}
+	dir.InitDir(ctx, make(map[string]*fs.Inode), fs.RootOwner, fs.FilePermsFromMode(0555))
+	return fs.NewInode(dir, msrc, fs.StableAttr{
+		DeviceID:  anon.PseudoDevice.DeviceID(),
+		InodeID:   anon.PseudoDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Directory,
+	})
+}
diff --git a/pkg/sentry/fs/ramfs/tree_test.go b/pkg/sentry/fs/ramfs/tree_test.go
new file mode 100644
index 000000000..68e2929d5
--- /dev/null
+++ b/pkg/sentry/fs/ramfs/tree_test.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ramfs
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+func TestMakeDirectoryTree(t *testing.T) {
+	mount := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+
+	for _, test := range []struct {
+		name    string
+		subdirs []string
+	}{
+		{
+			name: "abs paths",
+			subdirs: []string{
+				"/tmp",
+				"/tmp/a/b",
+				"/tmp/a/c/d",
+				"/tmp/c",
+				"/proc",
+				"/dev/a/b",
+				"/tmp",
+			},
+		},
+		{
+			name: "rel paths",
+			subdirs: []string{
+				"tmp",
+				"tmp/a/b",
+				"tmp/a/c/d",
+				"tmp/c",
+				"proc",
+				"dev/a/b",
+				"tmp",
+			},
+		},
+	} {
+		ctx := contexttest.Context(t)
+		tree, err := MakeDirectoryTree(ctx, mount, test.subdirs)
+		if err != nil {
+			t.Errorf("%s: failed to make ramfs tree, got error %v, want nil", test.name, err)
+			continue
+		}
+
+		// Expect to be able to find each of the paths.
+		mm, err := fs.NewMountNamespace(ctx, tree)
+		if err != nil {
+			t.Errorf("%s: failed to create mount manager: %v", test.name, err)
+			continue
+		}
+		root := mm.Root()
+		defer mm.DecRef()
+
+		for _, p := range test.subdirs {
+			if _, err := mm.FindInode(ctx, root, nil, p, 0); err != nil {
+				t.Errorf("%s: failed to find node %s: %v", test.name, p, err)
+				break
+			}
+		}
+	}
+}
diff --git a/pkg/sentry/fs/restore.go b/pkg/sentry/fs/restore.go
new file mode 100644
index 000000000..b4ac85a27
--- /dev/null
+++ b/pkg/sentry/fs/restore.go
@@ -0,0 +1,75 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"sync"
+)
+
+// RestoreEnvironment is the restore environment for file systems. It consists
+// of things that change across save and restore and therefore cannot be saved
+// in the object graph.
+type RestoreEnvironment struct {
+	// MountSources maps Filesystem.Name() to mount arguments.
+	MountSources map[string][]MountArgs
+
+	// ValidateFileSize indicates file size should not change across S/R.
+	ValidateFileSize bool
+
+	// ValidateFileTimestamp indicates file modification timestamp should
+	// not change across S/R.
+	ValidateFileTimestamp bool
+}
+
+// MountArgs holds arguments to Mount.
+type MountArgs struct {
+	// Dev corresponds to the devname argumnent of Mount.
+	Dev string
+
+	// Flags corresponds to the flags argument of Mount.
+	Flags MountSourceFlags
+
+	// Data corresponds to the data argument of Mount.
+	Data string
+}
+
+// restoreEnv holds the fs package global RestoreEnvironment.
+var restoreEnv = struct {
+	mu  sync.Mutex
+	env RestoreEnvironment
+	set bool
+}{}
+
+// SetRestoreEnvironment sets the RestoreEnvironment. Must be called before
+// state.Load and only once.
+func SetRestoreEnvironment(r RestoreEnvironment) {
+	restoreEnv.mu.Lock()
+	defer restoreEnv.mu.Unlock()
+	if restoreEnv.set {
+		panic("RestoreEnvironment may only be set once")
+	}
+	restoreEnv.env = r
+	restoreEnv.set = true
+}
+
+// CurrentRestoreEnvironment returns the current, read-only RestoreEnvironment.
+// If no RestoreEnvironment was ever set, returns (_, false).
+func CurrentRestoreEnvironment() (RestoreEnvironment, bool) {
+	restoreEnv.mu.Lock()
+	defer restoreEnv.mu.Unlock()
+	e := restoreEnv.env
+	set := restoreEnv.set
+	return e, set
+}
diff --git a/pkg/sentry/fs/save.go b/pkg/sentry/fs/save.go
new file mode 100644
index 000000000..bf2a85143
--- /dev/null
+++ b/pkg/sentry/fs/save.go
@@ -0,0 +1,77 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+)
+
+// SaveInodeMappings saves a mapping of path -> inode ID for every
+// user-reachable Dirent.
+//
+// The entire kernel must be frozen to call this, and filesystem state must not
+// change between SaveInodeMappings and state.Save, otherwise the saved state
+// of any MountSource may be incoherent.
+func SaveInodeMappings() {
+	mountsSeen := make(map[*MountSource]struct{})
+	for dirent := range allDirents.dirents {
+		if _, ok := mountsSeen[dirent.Inode.MountSource]; !ok {
+			dirent.Inode.MountSource.ResetInodeMappings()
+			mountsSeen[dirent.Inode.MountSource] = struct{}{}
+		}
+	}
+
+	for dirent := range allDirents.dirents {
+		if dirent.Inode != nil {
+			// We cannot trust the root provided in the mount due
+			// to the overlay. We can trust the overlay to delegate
+			// SaveInodeMappings to the right underlying
+			// filesystems, though.
+			root := dirent
+			for !root.mounted && root.parent != nil {
+				root = root.parent
+			}
+
+			// Add the mapping.
+			n, reachable := dirent.FullName(root)
+			if !reachable {
+				// Something has gone seriously wrong if we can't reach our root.
+				panic(fmt.Sprintf("Unreachable root on dirent file %s", n))
+			}
+			dirent.Inode.MountSource.SaveInodeMapping(dirent.Inode, n)
+		}
+	}
+}
+
+// SaveFileFsyncError converts an fs.File.Fsync error to an error that
+// indicates that the fs.File was not synced sufficiently to be saved.
+func SaveFileFsyncError(err error) error {
+	switch err {
+	case nil:
+		// We succeeded, everything is great.
+		return nil
+	case syscall.EBADF, syscall.EINVAL, syscall.EROFS, syscall.ENOSYS, syscall.EPERM:
+		// These errors mean that the underlying node might not be syncable,
+		// which we expect to be reported as such even from the gofer.
+		log.Infof("failed to sync during save: %v", err)
+		return nil
+	default:
+		// We failed in some way that indicates potential data loss.
+		return fmt.Errorf("failed to sync: %v, data loss may occur", err)
+	}
+}
diff --git a/pkg/sentry/fs/seek.go b/pkg/sentry/fs/seek.go
new file mode 100644
index 000000000..1268726c2
--- /dev/null
+++ b/pkg/sentry/fs/seek.go
@@ -0,0 +1,43 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// SeekWhence determines seek direction.
+type SeekWhence int
+
+const (
+	// SeekSet sets the absolute offset.
+	SeekSet SeekWhence = iota
+
+	// SeekCurrent sets relative to the current position.
+	SeekCurrent
+
+	// SeekEnd sets relative to the end of the file.
+	SeekEnd
+)
+
+// String returns a human readable string for whence.
+func (s SeekWhence) String() string {
+	switch s {
+	case SeekSet:
+		return "Set"
+	case SeekCurrent:
+		return "Current"
+	case SeekEnd:
+		return "End"
+	default:
+		return "Unknown"
+	}
+}
diff --git a/pkg/sentry/fs/sync.go b/pkg/sentry/fs/sync.go
new file mode 100644
index 000000000..9738a8f22
--- /dev/null
+++ b/pkg/sentry/fs/sync.go
@@ -0,0 +1,43 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package fs
+
+// SyncType enumerates ways in which a File can be synced.
+type SyncType int
+
+const (
+	// SyncAll indicates that modified in-memory metadata and data should
+	// be written to backing storage. SyncAll implies SyncBackingStorage.
+	SyncAll SyncType = iota
+
+	// SyncData indicates that along with modified in-memory data, only
+	// metadata needed to access that data needs to be written.
+	//
+	// For example, changes to access time or modification time do not
+	// need to be written because they are not necessary for a data read
+	// to be handled correctly, unlike the file size.
+	//
+	// The aim of SyncData is to reduce disk activity for applications
+	// that do not require all metadata to be synchronized with the disk,
+	// see fdatasync(2). File systems that implement SyncData as SyncAll
+	// do not support this optimization.
+	//
+	// SyncData implies SyncBackingStorage.
+	SyncData
+
+	// SyncBackingStorage indicates that in-flight write operations to
+	// backing storage should be flushed.
+	SyncBackingStorage
+)
diff --git a/pkg/sentry/fs/sys/BUILD b/pkg/sentry/fs/sys/BUILD
new file mode 100644
index 000000000..0ae2cbac8
--- /dev/null
+++ b/pkg/sentry/fs/sys/BUILD
@@ -0,0 +1,34 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "sys_state",
+    srcs = [
+        "fs.go",
+        "sys.go",
+    ],
+    out = "sys_state.go",
+    package = "sys",
+)
+
+go_library(
+    name = "sys",
+    srcs = [
+        "device.go",
+        "fs.go",
+        "sys.go",
+        "sys_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/sys",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+    ],
+)
diff --git a/pkg/sentry/fs/sys/device.go b/pkg/sentry/fs/sys/device.go
new file mode 100644
index 000000000..54e414d1b
--- /dev/null
+++ b/pkg/sentry/fs/sys/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sys
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// sysfsDevice is the sysfs virtual device.
+var sysfsDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/sys/fs.go b/pkg/sentry/fs/sys/fs.go
new file mode 100644
index 000000000..f25f648c3
--- /dev/null
+++ b/pkg/sentry/fs/sys/fs.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sys
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// filesystem is a sysfs.
+type filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&filesystem{})
+}
+
+// FilesystemName is the name underwhich the filesystem is registered.
+// Name matches fs/sysfs/mount.c:sysfs_fs_type.name.
+const FilesystemName = "sysfs"
+
+// Name is the name of the file system.
+func (*filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount allows users to mount(2) this file system.
+func (*filesystem) AllowUserMount() bool {
+	return true
+}
+
+// Flags returns that there is nothing special about this file system.
+//
+// In Linux, sysfs returns FS_USERNS_VISIBLE | FS_USERNS_MOUNT, see fs/sysfs/mount.c.
+func (*filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns a sysfs root which can be positioned in the vfs.
+func (f *filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// device is always ignored.
+	// sysfs ignores data, see fs/sysfs/mount.c:sysfs_mount.
+
+	return New(ctx, fs.NewNonCachingMountSource(f, flags)), nil
+}
diff --git a/pkg/sentry/fs/sys/sys.go b/pkg/sentry/fs/sys/sys.go
new file mode 100644
index 000000000..ccf56f644
--- /dev/null
+++ b/pkg/sentry/fs/sys/sys.go
@@ -0,0 +1,57 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package sys implements a sysfs filesystem.
+package sys
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type Dir struct {
+	ramfs.Dir
+}
+
+func newDir(ctx context.Context, msrc *fs.MountSource, contents map[string]*fs.Inode) *fs.Inode {
+	d := &Dir{}
+	d.InitDir(ctx, contents, fs.RootOwner, fs.FilePermsFromMode(0555))
+	return fs.NewInode(d, msrc, fs.StableAttr{
+		DeviceID:  sysfsDevice.DeviceID(),
+		InodeID:   sysfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.SpecialDirectory,
+	})
+}
+
+// New returns the root node of a partial simple sysfs.
+func New(ctx context.Context, msrc *fs.MountSource) *fs.Inode {
+	return newDir(ctx, msrc, map[string]*fs.Inode{
+		// Add a basic set of top-level directories. In Linux, these
+		// are dynamically added depending on the KConfig. Here we just
+		// add the most common ones.
+		"block":    newDir(ctx, msrc, nil),
+		"bus":      newDir(ctx, msrc, nil),
+		"class":    newDir(ctx, msrc, nil),
+		"dev":      newDir(ctx, msrc, nil),
+		"devices":  newDir(ctx, msrc, nil),
+		"firmware": newDir(ctx, msrc, nil),
+		"fs":       newDir(ctx, msrc, nil),
+		"kernel":   newDir(ctx, msrc, nil),
+		"module":   newDir(ctx, msrc, nil),
+		"power":    newDir(ctx, msrc, nil),
+	})
+}
diff --git a/pkg/sentry/fs/timerfd/BUILD b/pkg/sentry/fs/timerfd/BUILD
new file mode 100644
index 000000000..7fddc29f4
--- /dev/null
+++ b/pkg/sentry/fs/timerfd/BUILD
@@ -0,0 +1,35 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "timerfd_state",
+    srcs = [
+        "timerfd.go",
+    ],
+    out = "timerfd_state.go",
+    package = "timerfd",
+)
+
+go_library(
+    name = "timerfd",
+    srcs = [
+        "timerfd.go",
+        "timerfd_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/timerfd",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/refs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/fs/timerfd/timerfd.go b/pkg/sentry/fs/timerfd/timerfd.go
new file mode 100644
index 000000000..ae58f6fd7
--- /dev/null
+++ b/pkg/sentry/fs/timerfd/timerfd.go
@@ -0,0 +1,144 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package timerfd implements the semantics of Linux timerfd objects as
+// described by timerfd_create(2).
+package timerfd
+
+import (
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// TimerOperations implements fs.FileOperations for timerfds.
+type TimerOperations struct {
+	fsutil.ZeroSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	fsutil.NoIoctl       `state:"nosave"`
+
+	events waiter.Queue `state:"nosave"`
+	timer  *ktime.Timer
+
+	// val is the number of timer expirations since the last successful call to
+	// Readv, Preadv, or SetTime. val is accessed using atomic memory
+	// operations.
+	val uint64
+}
+
+// NewFile returns a timerfd File that receives time from c.
+func NewFile(ctx context.Context, c ktime.Clock) *fs.File {
+	dirent := fs.NewDirent(anon.NewInode(ctx), "anon_inode:[timerfd]")
+	tops := &TimerOperations{}
+	tops.timer = ktime.NewTimer(c, tops)
+	// Timerfds reject writes, but the Write flag must be set in order to
+	// ensure that our Writev/Pwritev methods actually get called to return
+	// the correct errors.
+	return fs.NewFile(ctx, dirent, fs.FileFlags{Read: true, Write: true}, tops)
+}
+
+// Release implements fs.FileOperations.Release.
+func (t *TimerOperations) Release() {
+	t.timer.Destroy()
+}
+
+// PauseTimer pauses the associated Timer.
+func (t *TimerOperations) PauseTimer() {
+	t.timer.Pause()
+}
+
+// ResumeTimer resumes the associated Timer.
+func (t *TimerOperations) ResumeTimer() {
+	t.timer.Resume()
+}
+
+// Clock returns the associated Timer's Clock.
+func (t *TimerOperations) Clock() ktime.Clock {
+	return t.timer.Clock()
+}
+
+// GetTime returns the associated Timer's setting and the time at which it was
+// observed.
+func (t *TimerOperations) GetTime() (ktime.Time, ktime.Setting) {
+	return t.timer.Get()
+}
+
+// SetTime atomically changes the associated Timer's setting, resets the number
+// of expirations to 0, and returns the previous setting and the time at which
+// it was observed.
+func (t *TimerOperations) SetTime(s ktime.Setting) (ktime.Time, ktime.Setting) {
+	return t.timer.SwapAnd(s, func() { atomic.StoreUint64(&t.val, 0) })
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (t *TimerOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	var ready waiter.EventMask
+	if atomic.LoadUint64(&t.val) != 0 {
+		ready |= waiter.EventIn
+	}
+	return ready
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (t *TimerOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	t.events.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (t *TimerOperations) EventUnregister(e *waiter.Entry) {
+	t.events.EventUnregister(e)
+}
+
+// Read implements fs.FileOperations.Read.
+func (t *TimerOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	const sizeofUint64 = 8
+	if dst.NumBytes() < sizeofUint64 {
+		return 0, syserror.EINVAL
+	}
+	if val := atomic.SwapUint64(&t.val, 0); val != 0 {
+		var buf [sizeofUint64]byte
+		usermem.ByteOrder.PutUint64(buf[:], val)
+		if _, err := dst.CopyOut(ctx, buf[:]); err != nil {
+			// Linux does not undo consuming the number of expirations even if
+			// writing to userspace fails.
+			return 0, err
+		}
+		return sizeofUint64, nil
+	}
+	return 0, syserror.ErrWouldBlock
+}
+
+// Write implements fs.FileOperations.Write.
+func (t *TimerOperations) Write(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EINVAL
+}
+
+// Notify implements ktime.TimerListener.Notify.
+func (t *TimerOperations) Notify(exp uint64) {
+	atomic.AddUint64(&t.val, exp)
+	t.events.Notify(waiter.EventIn)
+}
+
+// Destroy implements ktime.TimerListener.Destroy.
+func (t *TimerOperations) Destroy() {}
diff --git a/pkg/sentry/fs/tmpfs/BUILD b/pkg/sentry/fs/tmpfs/BUILD
new file mode 100644
index 000000000..be4e695d3
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/BUILD
@@ -0,0 +1,64 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "tmpfs_state",
+    srcs = [
+        "file_regular.go",
+        "fs.go",
+        "inode_file.go",
+        "tmpfs.go",
+    ],
+    out = "tmpfs_state.go",
+    package = "tmpfs",
+)
+
+go_library(
+    name = "tmpfs",
+    srcs = [
+        "device.go",
+        "file_regular.go",
+        "fs.go",
+        "inode_file.go",
+        "tmpfs.go",
+        "tmpfs_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/tmpfs",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/ramfs",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/pipe",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "tmpfs_test",
+    size = "small",
+    srcs = ["file_test.go"],
+    embed = [":tmpfs"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/tmpfs/device.go b/pkg/sentry/fs/tmpfs/device.go
new file mode 100644
index 000000000..e588b3440
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// tmpfsDevice is the kernel tmpfs device.
+var tmpfsDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/fs/tmpfs/file_regular.go b/pkg/sentry/fs/tmpfs/file_regular.go
new file mode 100644
index 000000000..9811d90bc
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/file_regular.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// regularFileOperations implements fs.FileOperations for a regular
+// tmpfs file.
+type regularFileOperations struct {
+	waiter.AlwaysReady   `state:"nosave"`
+	fsutil.NoopRelease   `state:"nosave"`
+	fsutil.GenericSeek   `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoopFsync     `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoIoctl       `state:"nosave"`
+
+	// iops is the InodeOperations of a regular tmpfs file. It is
+	// guaranteed to be the same as file.Dirent.Inode.InodeOperations,
+	// see operations that take fs.File below.
+	iops *fileInodeOperations
+}
+
+// Read implements fs.FileOperations.Read.
+func (r *regularFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	return r.iops.read(ctx, dst, offset)
+}
+
+// Write implements fs.FileOperations.Write.
+func (r *regularFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	return r.iops.write(ctx, src, offset)
+}
+
+// ConfigureMMap implements fs.FileOperations.ConfigureMMap.
+func (r *regularFileOperations) ConfigureMMap(ctx context.Context, file *fs.File, opts *memmap.MMapOpts) error {
+	return fsutil.GenericConfigureMMap(file, r.iops, opts)
+}
diff --git a/pkg/sentry/fs/tmpfs/file_test.go b/pkg/sentry/fs/tmpfs/file_test.go
new file mode 100644
index 000000000..f064eb1ac
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/file_test.go
@@ -0,0 +1,73 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+import (
+	"bytes"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func newFileInode(ctx context.Context) *fs.Inode {
+	m := fs.NewCachingMountSource(&Filesystem{}, fs.MountSourceFlags{})
+	iops := NewInMemoryFile(ctx, usage.Tmpfs, fs.WithCurrentTime(ctx, fs.UnstableAttr{}), platform.FromContext(ctx))
+	return fs.NewInode(iops, m, fs.StableAttr{
+		DeviceID:  tmpfsDevice.DeviceID(),
+		InodeID:   tmpfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.RegularFile,
+	})
+}
+
+func newFile(ctx context.Context) *fs.File {
+	inode := newFileInode(ctx)
+	f, _ := inode.GetFile(ctx, fs.NewDirent(inode, "stub"), fs.FileFlags{Read: true, Write: true})
+	return f
+}
+
+// Allocate once, write twice.
+func TestGrow(t *testing.T) {
+	ctx := contexttest.Context(t)
+	f := newFile(ctx)
+	defer f.DecRef()
+
+	abuf := bytes.Repeat([]byte{'a'}, 68)
+	n, err := f.Pwritev(ctx, usermem.BytesIOSequence(abuf), 0)
+	if n != int64(len(abuf)) || err != nil {
+		t.Fatalf("DeprecatedPwritev got (%d, %v) want (%d, nil)", n, err, len(abuf))
+	}
+
+	bbuf := bytes.Repeat([]byte{'b'}, 856)
+	n, err = f.Pwritev(ctx, usermem.BytesIOSequence(bbuf), 68)
+	if n != int64(len(bbuf)) || err != nil {
+		t.Fatalf("DeprecatedPwritev got (%d, %v) want (%d, nil)", n, err, len(bbuf))
+	}
+
+	rbuf := make([]byte, len(abuf)+len(bbuf))
+	n, err = f.Preadv(ctx, usermem.BytesIOSequence(rbuf), 0)
+	if n != int64(len(rbuf)) || err != nil {
+		t.Fatalf("DeprecatedPreadv got (%d, %v) want (%d, nil)", n, err, len(rbuf))
+	}
+
+	if want := append(abuf, bbuf...); !bytes.Equal(rbuf, want) {
+		t.Fatalf("Read %v, want %v", rbuf, want)
+	}
+}
diff --git a/pkg/sentry/fs/tmpfs/fs.go b/pkg/sentry/fs/tmpfs/fs.go
new file mode 100644
index 000000000..639a19b0d
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/fs.go
@@ -0,0 +1,131 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+import (
+	"fmt"
+	"regexp"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+)
+
+const (
+	// Set initial permissions for the root directory.
+	modeKey = "mode"
+
+	// UID for the root directory.
+	rootUIDKey = "uid"
+
+	// GID for the root directory.
+	rootGIDKey = "gid"
+
+	// TODO: support a tmpfs size limit.
+	// size = "size"
+
+	// default permissions are read/write/execute.
+	defaultMode = 0777
+)
+
+// modeRegexp is the expected format of the mode option.
+var modeRegexp = regexp.MustCompile("0[0-7][0-7][0-7]")
+
+// Filesystem is a tmpfs.
+type Filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&Filesystem{})
+}
+
+// FilesystemName is the name underwhich the filesystem is registered.
+// Name matches mm/shmem.c:shmem_fs_type.name.
+const FilesystemName = "tmpfs"
+
+// Name is the name of the file system.
+func (*Filesystem) Name() string {
+	return FilesystemName
+}
+
+// AllowUserMount allows users to mount(2) this file system.
+func (*Filesystem) AllowUserMount() bool {
+	return true
+}
+
+// Flags returns that there is nothing special about this file system.
+//
+// In Linux, tmpfs returns FS_USERNS_MOUNT, see mm/shmem.c.
+func (*Filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// Mount returns a tmpfs root that can be positioned in the vfs.
+func (f *Filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// device is always ignored.
+
+	// Parse generic comma-separated key=value options, this file system expects them.
+	options := fs.GenericMountSourceOptions(data)
+
+	// Parse the root directory permissions.
+	perms := fs.FilePermsFromMode(defaultMode)
+	if m, ok := options[modeKey]; ok {
+		if !modeRegexp.MatchString(m) {
+			return nil, fmt.Errorf("unsupported mode value: 'mode=%s'", m)
+		}
+		// It's basically impossible that we error out at this point,
+		// maybe we should panic.
+		i, err := strconv.ParseUint(m, 8, 32)
+		if err != nil {
+			return nil, fmt.Errorf("mode value not parsable 'mode=%s': %v", m, err)
+		}
+		perms = fs.FilePermsFromMode(linux.FileMode(i))
+		delete(options, modeKey)
+	}
+
+	creds := auth.CredentialsFromContext(ctx)
+	owner := fs.FileOwnerFromContext(ctx)
+	if uidstr, ok := options[rootUIDKey]; ok {
+		uid, err := strconv.ParseInt(uidstr, 10, 32)
+		if err != nil {
+			return nil, fmt.Errorf("uid value not parsable 'uid=%d': %v", uid, err)
+		}
+		owner.UID = creds.UserNamespace.MapToKUID(auth.UID(uid))
+		delete(options, rootUIDKey)
+	}
+
+	if gidstr, ok := options[rootGIDKey]; ok {
+		gid, err := strconv.ParseInt(gidstr, 10, 32)
+		if err != nil {
+			return nil, fmt.Errorf("gid value not parsable 'gid=%d': %v", gid, err)
+		}
+		owner.GID = creds.UserNamespace.MapToKGID(auth.GID(gid))
+		delete(options, rootGIDKey)
+	}
+
+	// Fail if the caller passed us more options than we can parse. They may be
+	// expecting us to set something we can't set.
+	if len(options) > 0 {
+		return nil, fmt.Errorf("unsupported mount options: %v", options)
+	}
+
+	// Construct a mount which will cache dirents.
+	msrc := fs.NewCachingMountSource(f, flags)
+
+	// Construct the tmpfs root.
+	return NewDir(ctx, nil, owner, perms, msrc, platform.FromContext(ctx)), nil
+}
diff --git a/pkg/sentry/fs/tmpfs/inode_file.go b/pkg/sentry/fs/tmpfs/inode_file.go
new file mode 100644
index 000000000..66bc934ae
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/inode_file.go
@@ -0,0 +1,492 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tmpfs
+
+import (
+	"io"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// fileInodeOperations implements fs.InodeOperations for a regular tmpfs file.
+// These files are backed by FrameRegions allocated from a platform.Memory,
+// and may be directly mapped.
+//
+// The tmpfs file memory is backed by FrameRegions, each of which is reference
+// counted. frames maintains a single reference on each of the FrameRegions.
+// Since these contain the contents of the file, the reference may only be
+// decremented once this file is both deleted and all handles to the file have
+// been closed.
+//
+// Mappable users may also call IncRefOn/DecRefOn, generally to indicate that
+// they plan to use MapInto to map the file into an AddressSpace. These calls
+// include an InvalidatorRegion associated with that reference. When the
+// referenced portion of the file is removed (with Truncate), the associated
+// InvalidatorRegion is invalidated.
+type fileInodeOperations struct {
+	fsutil.DeprecatedFileOperations `state:"nosave"`
+	fsutil.InodeNotDirectory        `state:"nosave"`
+	fsutil.InodeNotSocket           `state:"nosave"`
+	fsutil.InodeNotSymlink          `state:"nosave"`
+	fsutil.NoopWriteOut             `state:"nosave"`
+
+	// platform is used to allocate memory that stores the file's contents.
+	platform platform.Platform
+
+	// memUsage is the default memory usage that will be reported by this file.
+	memUsage usage.MemoryKind
+
+	attrMu sync.Mutex `state:"nosave"`
+
+	// attr contains the unstable metadata for the file.
+	//
+	// attr is protected by attrMu. attr.Unstable.Size is protected by both
+	// attrMu and dataMu; reading it requires locking either mutex, while
+	// mutating it requires locking both.
+	attr fsutil.InMemoryAttributes
+
+	mapsMu sync.Mutex `state:"nosave"`
+
+	// mappings tracks mappings of the file into memmap.MappingSpaces.
+	//
+	// mappings is protected by mapsMu.
+	mappings memmap.MappingSet
+
+	dataMu sync.RWMutex `state:"nosave"`
+
+	// data maps offsets into the file to offsets into platform.Memory() that
+	// store the file's data.
+	//
+	// data is protected by dataMu.
+	data fsutil.FileRangeSet
+}
+
+// NewInMemoryFile returns a new file backed by p.Memory().
+func NewInMemoryFile(ctx context.Context, usage usage.MemoryKind, uattr fs.UnstableAttr, p platform.Platform) fs.InodeOperations {
+	return &fileInodeOperations{
+		attr: fsutil.InMemoryAttributes{
+			Unstable: uattr,
+		},
+		platform: p,
+		memUsage: usage,
+	}
+}
+
+// Release implements fs.InodeOperations.Release.
+func (f *fileInodeOperations) Release(context.Context) {
+	f.dataMu.Lock()
+	defer f.dataMu.Unlock()
+	f.data.DropAll(f.platform.Memory())
+}
+
+// Mappable implements fs.InodeOperations.Mappable.
+func (f *fileInodeOperations) Mappable(*fs.Inode) memmap.Mappable {
+	return f
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (*fileInodeOperations) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return rename(ctx, oldParent, oldName, newParent, newName)
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (f *fileInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	flags.Pread = true
+	flags.Pwrite = true
+	return fs.NewFile(ctx, d, flags, &regularFileOperations{iops: f}), nil
+}
+
+// UnstableAttr returns unstable attributes of this tmpfs file.
+func (f *fileInodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	f.dataMu.RLock()
+	defer f.dataMu.RUnlock()
+	attr := f.attr.Unstable
+	attr.Usage = int64(f.data.Span())
+	return attr, nil
+}
+
+// Getxattr implements fs.InodeOperations.Getxattr.
+func (f *fileInodeOperations) Getxattr(inode *fs.Inode, name string) ([]byte, error) {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.Getxattr(name)
+}
+
+// Setxattr implements fs.InodeOperations.Setxattr.
+func (f *fileInodeOperations) Setxattr(inode *fs.Inode, name string, value []byte) error {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.Setxattr(name, value)
+}
+
+// Listxattr implements fs.InodeOperations.Listxattr.
+func (f *fileInodeOperations) Listxattr(inode *fs.Inode) (map[string]struct{}, error) {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.Listxattr()
+}
+
+// Check implements fs.InodeOperations.Check.
+func (f *fileInodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (f *fileInodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, p fs.FilePermissions) bool {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.SetPermissions(ctx, p)
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (f *fileInodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.SetTimestamps(ctx, ts)
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (f *fileInodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	return f.attr.SetOwner(ctx, owner)
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (f *fileInodeOperations) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+
+	f.dataMu.Lock()
+	oldSize := f.attr.Unstable.Size
+	if oldSize != size {
+		f.attr.Unstable.Size = size
+		f.attr.TouchModificationTime(ctx)
+	}
+	f.dataMu.Unlock()
+
+	// Nothing left to do unless shrinking the file.
+	if oldSize <= size {
+		return nil
+	}
+
+	oldpgend := fs.OffsetPageEnd(oldSize)
+	newpgend := fs.OffsetPageEnd(size)
+
+	// Invalidate past translations of truncated pages.
+	if newpgend != oldpgend {
+		f.mapsMu.Lock()
+		f.mappings.Invalidate(memmap.MappableRange{newpgend, oldpgend}, memmap.InvalidateOpts{
+			// Compare Linux's mm/shmem.c:shmem_setattr() =>
+			// mm/memory.c:unmap_mapping_range(evencows=1).
+			InvalidatePrivate: true,
+		})
+		f.mapsMu.Unlock()
+	}
+
+	// We are now guaranteed that there are no translations of truncated pages,
+	// and can remove them.
+	f.dataMu.Lock()
+	defer f.dataMu.Unlock()
+	f.data.Truncate(uint64(size), f.platform.Memory())
+
+	return nil
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+func (f *fileInodeOperations) AddLink() {
+	f.attrMu.Lock()
+	f.attr.Unstable.Links++
+	f.attrMu.Unlock()
+}
+
+// DropLink implements fs.InodeOperations.DropLink.
+func (f *fileInodeOperations) DropLink() {
+	f.attrMu.Lock()
+	f.attr.Unstable.Links--
+	f.attrMu.Unlock()
+}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+func (f *fileInodeOperations) NotifyStatusChange(ctx context.Context) {
+	f.attrMu.Lock()
+	f.attr.TouchStatusChangeTime(ctx)
+	f.attrMu.Unlock()
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+func (*fileInodeOperations) IsVirtual() bool {
+	return true
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+func (*fileInodeOperations) StatFS(context.Context) (fs.Info, error) {
+	return fsInfo, nil
+}
+
+func (f *fileInodeOperations) read(ctx context.Context, dst usermem.IOSequence, offset int64) (int64, error) {
+	// Zero length reads for tmpfs are no-ops.
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Have we reached EOF? We check for this again in
+	// fileReadWriter.ReadToBlocks to avoid holding f.attrMu (which would
+	// serialize reads) or f.dataMu (which would violate lock ordering), but
+	// check here first (before calling into MM) since reading at EOF is
+	// common: getting a return value of 0 from a read syscall is the only way
+	// to detect EOF.
+	//
+	// TODO: Separate out f.attr.Size and use atomics instead of
+	// f.dataMu.
+	f.dataMu.RLock()
+	size := f.attr.Unstable.Size
+	f.dataMu.RUnlock()
+	if offset >= size {
+		return 0, io.EOF
+	}
+
+	n, err := dst.CopyOutFrom(ctx, &fileReadWriter{f, offset})
+	// Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed().
+	f.attrMu.Lock()
+	f.attr.TouchAccessTime(ctx)
+	f.attrMu.Unlock()
+	return n, err
+}
+
+func (f *fileInodeOperations) write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
+	// Zero length writes for tmpfs are no-ops.
+	if src.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	f.attrMu.Lock()
+	defer f.attrMu.Unlock()
+	// Compare Linux's mm/filemap.c:__generic_file_write_iter() => file_update_time().
+	f.attr.TouchModificationTime(ctx)
+	return src.CopyInTo(ctx, &fileReadWriter{f, offset})
+}
+
+type fileReadWriter struct {
+	f      *fileInodeOperations
+	offset int64
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (rw *fileReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	rw.f.dataMu.RLock()
+	defer rw.f.dataMu.RUnlock()
+
+	// Compute the range to read.
+	if rw.offset >= rw.f.attr.Unstable.Size {
+		return 0, io.EOF
+	}
+	end := fs.ReadEndOffset(rw.offset, int64(dsts.NumBytes()), rw.f.attr.Unstable.Size)
+	if end == rw.offset { // dsts.NumBytes() == 0?
+		return 0, nil
+	}
+
+	mem := rw.f.platform.Memory()
+	var done uint64
+	seg, gap := rw.f.data.Find(uint64(rw.offset))
+	for rw.offset < end {
+		mr := memmap.MappableRange{uint64(rw.offset), uint64(end)}
+		switch {
+		case seg.Ok():
+			// Get internal mappings.
+			ims, err := mem.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Read)
+			if err != nil {
+				return done, err
+			}
+
+			// Copy from internal mappings.
+			n, err := safemem.CopySeq(dsts, ims)
+			done += n
+			rw.offset += int64(n)
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Tmpfs holes are zero-filled.
+			gapmr := gap.Range().Intersect(mr)
+			dst := dsts.TakeFirst64(gapmr.Length())
+			n, err := safemem.ZeroSeq(dst)
+			done += n
+			rw.offset += int64(n)
+			dsts = dsts.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = gap.NextSegment(), fsutil.FileRangeGapIterator{}
+
+		default:
+			break
+		}
+	}
+	return done, nil
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+func (rw *fileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	rw.f.dataMu.Lock()
+	defer rw.f.dataMu.Unlock()
+
+	// Compute the range to write.
+	end := fs.WriteEndOffset(rw.offset, int64(srcs.NumBytes()))
+	if end == rw.offset { // srcs.NumBytes() == 0?
+		return 0, nil
+	}
+
+	defer func() {
+		// If the write ends beyond the file's previous size, it causes the
+		// file to grow.
+		if rw.offset > rw.f.attr.Unstable.Size {
+			rw.f.attr.Unstable.Size = rw.offset
+		}
+	}()
+
+	mem := rw.f.platform.Memory()
+	// Page-aligned mr for when we need to allocate memory. RoundUp can't
+	// overflow since end is an int64.
+	pgstartaddr := usermem.Addr(rw.offset).RoundDown()
+	pgendaddr, _ := usermem.Addr(end).RoundUp()
+	pgMR := memmap.MappableRange{uint64(pgstartaddr), uint64(pgendaddr)}
+
+	var done uint64
+	seg, gap := rw.f.data.Find(uint64(rw.offset))
+	for rw.offset < end {
+		mr := memmap.MappableRange{uint64(rw.offset), uint64(end)}
+		switch {
+		case seg.Ok():
+			// Get internal mappings.
+			ims, err := mem.MapInternal(seg.FileRangeOf(seg.Range().Intersect(mr)), usermem.Write)
+			if err != nil {
+				return done, err
+			}
+
+			// Copy to internal mappings.
+			n, err := safemem.CopySeq(ims, srcs)
+			done += n
+			rw.offset += int64(n)
+			srcs = srcs.DropFirst64(n)
+			if err != nil {
+				return done, err
+			}
+
+			// Continue.
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok():
+			// Allocate memory for the write.
+			gapMR := gap.Range().Intersect(pgMR)
+			fr, err := mem.Allocate(gapMR.Length(), rw.f.memUsage)
+			if err != nil {
+				return done, err
+			}
+
+			// Write to that memory as usual.
+			seg, gap = rw.f.data.Insert(gap, gapMR, fr.Start), fsutil.FileRangeGapIterator{}
+
+		default:
+			break
+		}
+	}
+	return done, nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (f *fileInodeOperations) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	f.mapsMu.Lock()
+	defer f.mapsMu.Unlock()
+	f.mappings.AddMapping(ms, ar, offset)
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (f *fileInodeOperations) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+	f.mapsMu.Lock()
+	defer f.mapsMu.Unlock()
+	f.mappings.RemoveMapping(ms, ar, offset)
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (f *fileInodeOperations) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	return f.AddMapping(ctx, ms, dstAR, offset)
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (f *fileInodeOperations) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	f.dataMu.Lock()
+	defer f.dataMu.Unlock()
+
+	// Constrain translations to f.attr.Unstable.Size (rounded up) to prevent
+	// translation to pages that may be concurrently truncated.
+	pgend := fs.OffsetPageEnd(f.attr.Unstable.Size)
+	var buserr error
+	if required.End > pgend {
+		buserr = &memmap.BusError{io.EOF}
+		required.End = pgend
+	}
+	if optional.End > pgend {
+		optional.End = pgend
+	}
+
+	mem := f.platform.Memory()
+	cerr := f.data.Fill(ctx, required, optional, mem, f.memUsage, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) {
+		// Newly-allocated pages are zeroed, so we don't need to do anything.
+		return dsts.NumBytes(), nil
+	})
+
+	var ts []memmap.Translation
+	var translatedEnd uint64
+	for seg := f.data.FindSegment(required.Start); seg.Ok() && seg.Start() < required.End; seg, _ = seg.NextNonEmpty() {
+		segMR := seg.Range().Intersect(optional)
+		ts = append(ts, memmap.Translation{
+			Source: segMR,
+			File:   mem,
+			Offset: seg.FileRangeOf(segMR).Start,
+		})
+		translatedEnd = segMR.End
+	}
+
+	// Don't return the error returned by f.data.Fill if it occurred outside of
+	// required.
+	if translatedEnd < required.End && cerr != nil {
+		return ts, cerr
+	}
+	return ts, buserr
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (f *fileInodeOperations) InvalidateUnsavable(ctx context.Context) error {
+	return nil
+}
diff --git a/pkg/sentry/fs/tmpfs/tmpfs.go b/pkg/sentry/fs/tmpfs/tmpfs.go
new file mode 100644
index 000000000..1cc7ae491
--- /dev/null
+++ b/pkg/sentry/fs/tmpfs/tmpfs.go
@@ -0,0 +1,204 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package tmpfs is a filesystem implementation backed by memory.
+package tmpfs
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/ramfs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/pipe"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+var fsInfo = fs.Info{
+	Type: linux.TMPFS_MAGIC,
+
+	// TODO: allow configuring a tmpfs size and enforce it.
+	TotalBlocks: 0,
+	FreeBlocks:  0,
+}
+
+// rename implements fs.InodeOperations.Rename for tmpfs nodes.
+func rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	op, ok := oldParent.InodeOperations.(*Dir)
+	if !ok {
+		return ramfs.ErrCrossDevice
+	}
+	np, ok := newParent.InodeOperations.(*Dir)
+	if !ok {
+		return ramfs.ErrCrossDevice
+	}
+	return ramfs.Rename(ctx, &op.Dir, oldName, &np.Dir, newName)
+}
+
+// Dir is a directory.
+type Dir struct {
+	ramfs.Dir
+
+	// platform is used to allocate storage for tmpfs Files.
+	platform platform.Platform
+}
+
+// NewDir returns a new directory.
+func NewDir(ctx context.Context, contents map[string]*fs.Inode, owner fs.FileOwner, perms fs.FilePermissions, msrc *fs.MountSource, platform platform.Platform) *fs.Inode {
+	d := &Dir{platform: platform}
+	d.InitDir(ctx, contents, owner, perms)
+
+	// Manually set the CreateOps.
+	d.CreateOps = d.newCreateOps()
+
+	return fs.NewInode(d, msrc, fs.StableAttr{
+		DeviceID:  tmpfsDevice.DeviceID(),
+		InodeID:   tmpfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Directory,
+	})
+}
+
+// afterLoad is invoked by stateify.
+func (d *Dir) afterLoad() {
+	// Per NewDir, manually set the CreateOps.
+	d.Dir.CreateOps = d.newCreateOps()
+}
+
+// newCreateOps builds the custom CreateOps for this Dir.
+func (d *Dir) newCreateOps() *ramfs.CreateOps {
+	return &ramfs.CreateOps{
+		NewDir: func(ctx context.Context, dir *fs.Inode, perms fs.FilePermissions) (*fs.Inode, error) {
+			return NewDir(ctx, nil, fs.FileOwnerFromContext(ctx), perms, dir.MountSource, d.platform), nil
+		},
+		NewFile: func(ctx context.Context, dir *fs.Inode, perms fs.FilePermissions) (*fs.Inode, error) {
+			uattr := fs.WithCurrentTime(ctx, fs.UnstableAttr{
+				Owner: fs.FileOwnerFromContext(ctx),
+				Perms: perms,
+				// Always start unlinked.
+				Links: 0,
+			})
+			iops := NewInMemoryFile(ctx, usage.Tmpfs, uattr, d.platform)
+			return fs.NewInode(iops, dir.MountSource, fs.StableAttr{
+				DeviceID:  tmpfsDevice.DeviceID(),
+				InodeID:   tmpfsDevice.NextIno(),
+				BlockSize: usermem.PageSize,
+				Type:      fs.RegularFile,
+			}), nil
+		},
+		NewSymlink: func(ctx context.Context, dir *fs.Inode, target string) (*fs.Inode, error) {
+			return NewSymlink(ctx, target, fs.FileOwnerFromContext(ctx), dir.MountSource), nil
+		},
+		NewBoundEndpoint: func(ctx context.Context, dir *fs.Inode, socket unix.BoundEndpoint, perms fs.FilePermissions) (*fs.Inode, error) {
+			return NewSocket(ctx, socket, fs.FileOwnerFromContext(ctx), perms, dir.MountSource), nil
+		},
+		NewFifo: func(ctx context.Context, dir *fs.Inode, perms fs.FilePermissions) (*fs.Inode, error) {
+			return NewFifo(ctx, fs.FileOwnerFromContext(ctx), perms, dir.MountSource), nil
+		},
+	}
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (d *Dir) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return rename(ctx, oldParent, oldName, newParent, newName)
+}
+
+// StatFS implments fs.InodeOperations.StatFS.
+func (*Dir) StatFS(context.Context) (fs.Info, error) {
+	return fsInfo, nil
+}
+
+// Symlink is a symlink.
+type Symlink struct {
+	ramfs.Symlink
+}
+
+// NewSymlink returns a new symlink with the provided permissions.
+func NewSymlink(ctx context.Context, target string, owner fs.FileOwner, msrc *fs.MountSource) *fs.Inode {
+	s := &Symlink{}
+	s.InitSymlink(ctx, owner, target)
+	return fs.NewInode(s, msrc, fs.StableAttr{
+		DeviceID:  tmpfsDevice.DeviceID(),
+		InodeID:   tmpfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Symlink,
+	})
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (s *Symlink) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return rename(ctx, oldParent, oldName, newParent, newName)
+}
+
+// StatFS returns the tmpfs info.
+func (s *Symlink) StatFS(context.Context) (fs.Info, error) {
+	return fsInfo, nil
+}
+
+// Socket is a socket.
+type Socket struct {
+	ramfs.Socket
+}
+
+// NewSocket returns a new socket with the provided permissions.
+func NewSocket(ctx context.Context, socket unix.BoundEndpoint, owner fs.FileOwner, perms fs.FilePermissions, msrc *fs.MountSource) *fs.Inode {
+	s := &Socket{}
+	s.InitSocket(ctx, socket, owner, perms)
+	return fs.NewInode(s, msrc, fs.StableAttr{
+		DeviceID:  tmpfsDevice.DeviceID(),
+		InodeID:   tmpfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Socket,
+	})
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (s *Socket) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return rename(ctx, oldParent, oldName, newParent, newName)
+}
+
+// StatFS returns the tmpfs info.
+func (s *Socket) StatFS(context.Context) (fs.Info, error) {
+	return fsInfo, nil
+}
+
+// Fifo is a tmpfs named pipe.
+type Fifo struct {
+	ramfs.Entry
+}
+
+// NewFifo creates a new named pipe.
+func NewFifo(ctx context.Context, owner fs.FileOwner, perms fs.FilePermissions, msrc *fs.MountSource) *fs.Inode {
+	f := &Fifo{}
+	f.InitEntry(ctx, owner, perms)
+	iops := pipe.NewInodeOperations(f, pipe.NewPipe(ctx, true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize))
+	return fs.NewInode(iops, msrc, fs.StableAttr{
+		DeviceID:  tmpfsDevice.DeviceID(),
+		InodeID:   tmpfsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Pipe,
+	})
+}
+
+// Rename implements fs.InodeOperations.Rename.
+func (f *Fifo) Rename(ctx context.Context, oldParent *fs.Inode, oldName string, newParent *fs.Inode, newName string) error {
+	return rename(ctx, oldParent, oldName, newParent, newName)
+}
+
+// StatFS returns the tmpfs info.
+func (*Fifo) StatFS(context.Context) (fs.Info, error) {
+	return fsInfo, nil
+}
diff --git a/pkg/sentry/fs/tty/BUILD b/pkg/sentry/fs/tty/BUILD
new file mode 100644
index 000000000..90b350410
--- /dev/null
+++ b/pkg/sentry/fs/tty/BUILD
@@ -0,0 +1,63 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "tty_state",
+    srcs = [
+        "dir.go",
+        "fs.go",
+        "inode.go",
+        "line_discipline.go",
+        "master.go",
+        "slave.go",
+        "terminal.go",
+    ],
+    out = "tty_state.go",
+    package = "tty",
+)
+
+go_library(
+    name = "tty",
+    srcs = [
+        "dir.go",
+        "fs.go",
+        "inode.go",
+        "line_discipline.go",
+        "master.go",
+        "slave.go",
+        "terminal.go",
+        "tty_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/fs/tty",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "tty_test",
+    size = "small",
+    srcs = ["tty_test.go"],
+    embed = [":tty"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/fs/tty/dir.go b/pkg/sentry/fs/tty/dir.go
new file mode 100644
index 000000000..2c5b2aed6
--- /dev/null
+++ b/pkg/sentry/fs/tty/dir.go
@@ -0,0 +1,398 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package tty provide pseudoterminals via a devpts filesystem.
+package tty
+
+import (
+	"fmt"
+	"math"
+	"strconv"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// dirInodeOperations is the root of a devpts mount.
+//
+// This indirectly manages all terminals within the mount.
+//
+// New Terminals are created by masterInodeOperations.GetFile, which registers
+// the slave Inode in the this directory for discovery via Lookup/Readdir. The
+// slave inode is unregistered when the master file is Released, as the slave
+// is no longer discoverable at that point.
+//
+// References on the underlying Terminal are held by masterFileOperations and
+// slaveInodeOperations.
+//
+// masterInodeOperations and slaveInodeOperations hold a pointer to
+// dirInodeOperations, which is reference counted by the refcount their
+// corresponding Dirents hold on their parent (this directory).
+//
+// dirInodeOperations implements fs.InodeOperations.
+type dirInodeOperations struct {
+	fsutil.DeprecatedFileOperations
+	fsutil.InodeNotSocket
+	fsutil.InodeNotRenameable
+	fsutil.InodeNotSymlink
+	fsutil.InodeNoExtendedAttributes
+	fsutil.NoMappable
+	fsutil.NoopWriteOut
+
+	// msrc is the super block this directory is on.
+	//
+	// TODO: Plumb this through instead of storing it here.
+	msrc *fs.MountSource
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// attr contains the UnstableAttrs.
+	attr fsutil.InMemoryAttributes
+
+	// master is the master PTY inode.
+	master *fs.Inode
+
+	// slaves contains the slave inodes reachable from the directory.
+	//
+	// A new slave is added by allocateTerminal and is removed by
+	// masterFileOperations.Release.
+	//
+	// A reference is held on every slave in the map.
+	slaves map[uint32]*fs.Inode
+
+	// dentryMap is a SortedDentryMap used to implement Readdir containing
+	// the master and all entries in slaves.
+	dentryMap *fs.SortedDentryMap
+
+	// next is the next pty index to use.
+	//
+	// TODO: reuse indices when ptys are closed.
+	next uint32
+}
+
+var _ fs.InodeOperations = (*dirInodeOperations)(nil)
+
+// newDir creates a new dir with a ptmx file and no terminals.
+func newDir(ctx context.Context, m *fs.MountSource) *fs.Inode {
+	d := &dirInodeOperations{
+		attr: fsutil.InMemoryAttributes{
+			Unstable: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+				Owner: fs.RootOwner,
+				Perms: fs.FilePermsFromMode(0555),
+			}),
+		},
+		msrc:      m,
+		slaves:    make(map[uint32]*fs.Inode),
+		dentryMap: fs.NewSortedDentryMap(nil),
+	}
+	// Linux devpts uses a default mode of 0000 for ptmx which can be
+	// changed with the ptmxmode mount option. However, that default is not
+	// useful here (since we'd *always* need the mount option, so it is
+	// accessible by default).
+	d.master = newMasterInode(ctx, d, fs.RootOwner, fs.FilePermsFromMode(0666))
+	d.dentryMap.Add("ptmx", fs.DentAttr{
+		Type:    d.master.StableAttr.Type,
+		InodeID: d.master.StableAttr.InodeID,
+	})
+
+	return fs.NewInode(d, m, fs.StableAttr{
+		DeviceID: ptsDevice.DeviceID(),
+		// N.B. Linux always uses inode id 1 for the directory. See
+		// fs/devpts/inode.c:devpts_fill_super.
+		//
+		// TODO: Since ptsDevice must be shared between
+		// different mounts, we must not assign fixed numbers.
+		InodeID:   ptsDevice.NextIno(),
+		BlockSize: usermem.PageSize,
+		Type:      fs.Directory,
+	})
+}
+
+// Release implements fs.InodeOperations.Release.
+func (d *dirInodeOperations) Release(ctx context.Context) {
+	d.master.DecRef()
+	if len(d.slaves) != 0 {
+		panic(fmt.Sprintf("devpts directory still contains active terminals: %+v", d))
+	}
+}
+
+// Lookup implements fs.InodeOperations.Lookup.
+func (d *dirInodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name string) (*fs.Dirent, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	// Master?
+	if name == "ptmx" {
+		d.master.IncRef()
+		return fs.NewDirent(d.master, name), nil
+	}
+
+	// Slave number?
+	n, err := strconv.ParseUint(name, 10, 32)
+	if err != nil {
+		// Not found.
+		return nil, syserror.ENOENT
+	}
+
+	s, ok := d.slaves[uint32(n)]
+	if !ok {
+		return nil, syserror.ENOENT
+	}
+
+	s.IncRef()
+	return fs.NewDirent(s, name), nil
+}
+
+// Create implements fs.InodeOperations.Create.
+//
+// Creation is never allowed.
+func (d *dirInodeOperations) Create(ctx context.Context, dir *fs.Inode, name string, flags fs.FileFlags, perm fs.FilePermissions) (*fs.File, error) {
+	return nil, syserror.EACCES
+}
+
+// CreateDirectory implements fs.InodeOperations.CreateDirectory.
+//
+// Creation is never allowed.
+func (d *dirInodeOperations) CreateDirectory(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error {
+	return syserror.EACCES
+}
+
+// CreateLink implements fs.InodeOperations.CreateLink.
+//
+// Creation is never allowed.
+func (d *dirInodeOperations) CreateLink(ctx context.Context, dir *fs.Inode, oldname, newname string) error {
+	return syserror.EACCES
+}
+
+// CreateHardLink implements fs.InodeOperations.CreateHardLink.
+//
+// Creation is never allowed.
+func (d *dirInodeOperations) CreateHardLink(ctx context.Context, dir *fs.Inode, target *fs.Inode, name string) error {
+	return syserror.EACCES
+}
+
+// CreateFifo implements fs.InodeOperations.CreateFifo.
+//
+// Creation is never allowed.
+func (d *dirInodeOperations) CreateFifo(ctx context.Context, dir *fs.Inode, name string, perm fs.FilePermissions) error {
+	return syserror.EACCES
+}
+
+// Remove implements fs.InodeOperations.Remove.
+//
+// Removal is never allowed.
+func (d *dirInodeOperations) Remove(ctx context.Context, dir *fs.Inode, name string) error {
+	return syserror.EPERM
+}
+
+// RemoveDirectory implements fs.InodeOperations.RemoveDirectory.
+//
+// Removal is never allowed.
+func (d *dirInodeOperations) RemoveDirectory(ctx context.Context, dir *fs.Inode, name string) error {
+	return syserror.EPERM
+}
+
+// Bind implements fs.InodeOperations.Bind.
+func (d *dirInodeOperations) Bind(ctx context.Context, dir *fs.Inode, name string, data unix.BoundEndpoint, perm fs.FilePermissions) error {
+	return syserror.EPERM
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+func (d *dirInodeOperations) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return fs.NewFile(ctx, dirent, flags, &dirFileOperations{di: d}), nil
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (d *dirInodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	return d.attr.Unstable, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (d *dirInodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions.
+func (d *dirInodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, p fs.FilePermissions) bool {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	return d.attr.SetPermissions(ctx, p)
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (d *dirInodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	return d.attr.SetOwner(ctx, owner)
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (d *dirInodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+	return d.attr.SetTimestamps(ctx, ts)
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (d *dirInodeOperations) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	return syserror.EINVAL
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+func (d *dirInodeOperations) AddLink() {}
+
+// DropLink implements fs.InodeOperations.DropLink.
+func (d *dirInodeOperations) DropLink() {}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+func (d *dirInodeOperations) NotifyStatusChange(ctx context.Context) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	d.attr.TouchStatusChangeTime(ctx)
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+func (d *dirInodeOperations) IsVirtual() bool {
+	return true
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+func (d *dirInodeOperations) StatFS(ctx context.Context) (fs.Info, error) {
+	return fs.Info{
+		Type: linux.DEVPTS_SUPER_MAGIC,
+	}, nil
+}
+
+// allocateTerminal creates a new Terminal and installs a pts node for it.
+//
+// The caller must call DecRef when done with the returned Terminal.
+func (d *dirInodeOperations) allocateTerminal(ctx context.Context) (*Terminal, error) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	n := d.next
+	if n == math.MaxUint32 {
+		return nil, syserror.ENOMEM
+	}
+
+	if _, ok := d.slaves[n]; ok {
+		panic(fmt.Sprintf("pty index collision; index %d already exists", n))
+	}
+
+	t := newTerminal(ctx, d, n)
+	d.next++
+
+	// The reference returned by newTerminal is returned to the caller.
+	// Take another for the slave inode.
+	t.IncRef()
+
+	// Create a pts node. The owner is based on the context that opens
+	// ptmx.
+	creds := auth.CredentialsFromContext(ctx)
+	uid, gid := creds.EffectiveKUID, creds.EffectiveKGID
+	slave := newSlaveInode(ctx, d, t, fs.FileOwner{uid, gid}, fs.FilePermsFromMode(0666))
+
+	d.slaves[n] = slave
+	d.dentryMap.Add(strconv.FormatUint(uint64(n), 10), fs.DentAttr{
+		Type:    slave.StableAttr.Type,
+		InodeID: slave.StableAttr.InodeID,
+	})
+
+	return t, nil
+}
+
+// masterClose is called when the master end of t is closed.
+func (d *dirInodeOperations) masterClose(t *Terminal) {
+	d.mu.Lock()
+	defer d.mu.Unlock()
+
+	// The slave end disappears from the directory when the master end is
+	// closed, even if the slave end is open elsewhere.
+	//
+	// N.B. since we're using a backdoor method to remove a directory entry
+	// we won't properly fire inotify events like Linux would.
+	s, ok := d.slaves[t.n]
+	if !ok {
+		panic(fmt.Sprintf("Terminal %+v doesn't exist in %+v?", t, d))
+	}
+
+	s.DecRef()
+	delete(d.slaves, t.n)
+	d.dentryMap.Remove(strconv.FormatUint(uint64(t.n), 10))
+}
+
+// dirFileOperations are the fs.FileOperations for the directory.
+//
+// This is nearly identical to fsutil.DirFileOperations, except that it takes
+// df.di.mu in IterateDir.
+type dirFileOperations struct {
+	waiter.AlwaysReady `state:"nosave"`
+	fsutil.NoopRelease `state:"nosave"`
+	fsutil.GenericSeek `state:"nosave"`
+	fsutil.NoFsync     `state:"nosave"`
+	fsutil.NoopFlush   `state:"nosave"`
+	fsutil.NoMMap      `state:"nosave"`
+	fsutil.NoIoctl     `state:"nosave"`
+
+	// di is the inode operations.
+	di *dirInodeOperations
+
+	// dirCursor contains the name of the last directory entry that was
+	// serialized.
+	dirCursor string
+}
+
+var _ fs.FileOperations = (*dirFileOperations)(nil)
+
+// IterateDir implements DirIterator.IterateDir.
+func (df *dirFileOperations) IterateDir(ctx context.Context, dirCtx *fs.DirCtx, offset int) (int, error) {
+	df.di.mu.Lock()
+	defer df.di.mu.Unlock()
+
+	n, err := fs.GenericReaddir(dirCtx, df.di.dentryMap)
+	return offset + n, err
+}
+
+// Readdir implements FileOperations.Readdir.
+func (df *dirFileOperations) Readdir(ctx context.Context, file *fs.File, serializer fs.DentrySerializer) (int64, error) {
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+	dirCtx := &fs.DirCtx{
+		Serializer: serializer,
+		DirCursor:  &df.dirCursor,
+	}
+	return fs.DirentReaddir(ctx, file.Dirent, df, root, dirCtx, file.Offset())
+}
+
+// Read implements FileOperations.Read
+func (df *dirFileOperations) Read(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EISDIR
+}
+
+// Write implements FileOperations.Write.
+func (df *dirFileOperations) Write(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syserror.EISDIR
+}
diff --git a/pkg/sentry/fs/tty/fs.go b/pkg/sentry/fs/tty/fs.go
new file mode 100644
index 000000000..f5e7a3162
--- /dev/null
+++ b/pkg/sentry/fs/tty/fs.go
@@ -0,0 +1,95 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// ptsDevice is the pseudo-filesystem device.
+var ptsDevice = device.NewAnonDevice()
+
+// filesystem is a devpts filesystem.
+//
+// This devpts is always in the new "multi-instance" mode. i.e., it contains a
+// ptmx device tied to this mount.
+type filesystem struct{}
+
+func init() {
+	fs.RegisterFilesystem(&filesystem{})
+}
+
+// Name matches drivers/devpts/indoe.c:devpts_fs_type.name.
+func (*filesystem) Name() string {
+	return "devpts"
+}
+
+// AllowUserMount allows users to mount(2) this file system.
+func (*filesystem) AllowUserMount() bool {
+	// TODO: Users may mount this once the terminals are in a
+	// usable state.
+	return false
+}
+
+// Flags returns that there is nothing special about this file system.
+func (*filesystem) Flags() fs.FilesystemFlags {
+	return 0
+}
+
+// MountSource returns a devpts root that can be positioned in the vfs.
+func (f *filesystem) Mount(ctx context.Context, device string, flags fs.MountSourceFlags, data string) (*fs.Inode, error) {
+	// device is always ignored.
+
+	// No options are supported.
+	if data != "" {
+		return nil, syserror.EINVAL
+	}
+
+	return newDir(ctx, fs.NewMountSource(&superOperations{}, f, flags)), nil
+}
+
+// superOperations implements fs.MountSourceOperations, preventing caching.
+type superOperations struct{}
+
+// Revalidate implements fs.DirentOperations.Revalidate.
+//
+// It always returns true, forcing a Lookup for all entries.
+//
+// Slave entries are dropped from dir when their master is closed, so an
+// existing slave Dirent in the tree is not sufficient to guarantee that it
+// still exists on the filesystem.
+func (superOperations) Revalidate(*fs.Dirent) bool {
+	return true
+}
+
+// Keep implements fs.DirentOperations.Keep.
+//
+// Keep returns false because Revalidate would force a lookup on cached entries
+// anyways.
+func (superOperations) Keep(*fs.Dirent) bool {
+	return false
+}
+
+// ResetInodeMappings implements MountSourceOperations.ResetInodeMappings.
+func (superOperations) ResetInodeMappings() {}
+
+// SaveInodeMapping implements MountSourceOperations.SaveInodeMapping.
+func (superOperations) SaveInodeMapping(*fs.Inode, string) {}
+
+// Destroy implements MountSourceOperations.Destroy.
+func (superOperations) Destroy() {}
diff --git a/pkg/sentry/fs/tty/inode.go b/pkg/sentry/fs/tty/inode.go
new file mode 100644
index 000000000..04b9a7727
--- /dev/null
+++ b/pkg/sentry/fs/tty/inode.go
@@ -0,0 +1,143 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// inodeOperations are the base fs.InodeOperations for master and slave Inodes.
+//
+// inodeOperations does not implement:
+//
+// * fs.InodeOperations.Release
+// * fs.InodeOperations.GetFile
+type inodeOperations struct {
+	fsutil.DeprecatedFileOperations  `state:"nosave"`
+	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNotDirectory         `state:"nosave"`
+	fsutil.InodeNotRenameable        `state:"nosave"`
+	fsutil.InodeNotSocket            `state:"nosave"`
+	fsutil.InodeNotSymlink           `state:"nosave"`
+	fsutil.NoMappable                `state:"nosave"`
+	fsutil.NoopWriteOut              `state:"nosave"`
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// uattr is the inode's UnstableAttr.
+	uattr fs.UnstableAttr
+}
+
+// UnstableAttr implements fs.InodeOperations.UnstableAttr.
+func (i *inodeOperations) UnstableAttr(ctx context.Context, inode *fs.Inode) (fs.UnstableAttr, error) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	return i.uattr, nil
+}
+
+// Check implements fs.InodeOperations.Check.
+func (i *inodeOperations) Check(ctx context.Context, inode *fs.Inode, p fs.PermMask) bool {
+	return fs.ContextCanAccessFile(ctx, inode, p)
+}
+
+// SetPermissions implements fs.InodeOperations.SetPermissions
+func (i *inodeOperations) SetPermissions(ctx context.Context, inode *fs.Inode, p fs.FilePermissions) bool {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	i.uattr.Perms = p
+	i.uattr.StatusChangeTime = ktime.NowFromContext(ctx)
+	return true
+}
+
+// SetOwner implements fs.InodeOperations.SetOwner.
+func (i *inodeOperations) SetOwner(ctx context.Context, inode *fs.Inode, owner fs.FileOwner) error {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	if owner.UID.Ok() {
+		i.uattr.Owner.UID = owner.UID
+	}
+	if owner.GID.Ok() {
+		i.uattr.Owner.GID = owner.GID
+	}
+	return nil
+}
+
+// SetTimestamps implements fs.InodeOperations.SetTimestamps.
+func (i *inodeOperations) SetTimestamps(ctx context.Context, inode *fs.Inode, ts fs.TimeSpec) error {
+	if ts.ATimeOmit && ts.MTimeOmit {
+		return nil
+	}
+
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
+	now := ktime.NowFromContext(ctx)
+	if !ts.ATimeOmit {
+		if ts.ATime.IsZero() {
+			i.uattr.AccessTime = now
+		} else {
+			i.uattr.AccessTime = ts.ATime
+		}
+	}
+	if !ts.MTimeOmit {
+		if ts.MTime.IsZero() {
+			i.uattr.ModificationTime = now
+		} else {
+			i.uattr.ModificationTime = ts.MTime
+		}
+	}
+	i.uattr.StatusChangeTime = now
+	return nil
+}
+
+// Truncate implements fs.InodeOperations.Truncate.
+func (i *inodeOperations) Truncate(ctx context.Context, inode *fs.Inode, size int64) error {
+	return syserror.EINVAL
+}
+
+// AddLink implements fs.InodeOperations.AddLink.
+func (i *inodeOperations) AddLink() {
+}
+
+// DropLink implements fs.InodeOperations.DropLink.
+func (i *inodeOperations) DropLink() {
+}
+
+// NotifyStatusChange implements fs.InodeOperations.NotifyStatusChange.
+func (i *inodeOperations) NotifyStatusChange(ctx context.Context) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+	i.uattr.StatusChangeTime = ktime.NowFromContext(ctx)
+}
+
+// IsVirtual implements fs.InodeOperations.IsVirtual.
+func (i *inodeOperations) IsVirtual() bool {
+	return true
+}
+
+// StatFS implements fs.InodeOperations.StatFS.
+func (i *inodeOperations) StatFS(ctx context.Context) (fs.Info, error) {
+	return fs.Info{
+		Type: linux.DEVPTS_SUPER_MAGIC,
+	}, nil
+}
diff --git a/pkg/sentry/fs/tty/line_discipline.go b/pkg/sentry/fs/tty/line_discipline.go
new file mode 100644
index 000000000..fde4e7941
--- /dev/null
+++ b/pkg/sentry/fs/tty/line_discipline.go
@@ -0,0 +1,342 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"bytes"
+	"sync"
+	"unicode/utf8"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+const (
+	spacesPerTab = 8
+)
+
+// lineDiscipline dictates how input and output are handled between the
+// pseudoterminal (pty) master and slave. It can be configured to alter I/O,
+// modify control characters (e.g. Ctrl-C for SIGINT), etc. The following man
+// pages are good resources for how to affect the line discipline:
+//
+//   * termios(3)
+//   * tty_ioctl(4)
+//
+// This file corresponds most closely to drivers/tty/n_tty.c.
+//
+// lineDiscipline has a simple structure but supports a multitude of options
+// (see the above man pages). It consists of two queues of bytes: one from the
+// terminal master to slave (the input queue) and one from slave to master (the
+// output queue). When bytes are written to one end of the pty, the line
+// discipline reads the bytes, modifies them or takes special action if
+// required, and enqueues them to be read by the other end of the pty:
+//
+//       input from terminal    +-------------+   input to process (e.g. bash)
+//    +------------------------>| input queue |---------------------------+
+//    |                         +-------------+                           |
+//    |                                                                   |
+//    |                                                                   v
+// masterFD                                                            slaveFD
+//    ^                                                                   |
+//    |                                                                   |
+//    |   output to terminal   +--------------+    output from process    |
+//    +------------------------| output queue |<--------------------------+
+//                             +--------------+
+//
+// Lock order:
+//  inMu
+//    outMu
+//      termiosMu
+type lineDiscipline struct {
+	// inMu protects inQueue.
+	inMu sync.Mutex `state:"nosave"`
+
+	// inQueue is the input queue of the terminal.
+	inQueue queue
+
+	// outMu protects outQueue.
+	outMu sync.Mutex `state:"nosave"`
+
+	// outQueue is the output queue of the terminal.
+	outQueue queue
+
+	// termiosMu protects termios.
+	termiosMu sync.Mutex `state:"nosave"`
+
+	// termios is the terminal configuration used by the lineDiscipline.
+	termios linux.KernelTermios
+
+	// column is the location in a row of the cursor. This is important for
+	// handling certain special characters like backspace.
+	column int
+}
+
+// getTermios gets the linux.Termios for the tty.
+func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	l.termiosMu.Lock()
+	defer l.termiosMu.Unlock()
+	// We must copy a Termios struct, not KernelTermios.
+	t := l.termios.ToTermios()
+	_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), t, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	return 0, err
+}
+
+// setTermios sets a linux.Termios for the tty.
+func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	l.termiosMu.Lock()
+	defer l.termiosMu.Unlock()
+	// We must copy a Termios struct, not KernelTermios.
+	var t linux.Termios
+	_, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &t, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	l.termios.FromTermios(t)
+	return 0, err
+}
+
+func (l *lineDiscipline) masterReadiness() waiter.EventMask {
+	l.inMu.Lock()
+	defer l.inMu.Unlock()
+	l.outMu.Lock()
+	defer l.outMu.Unlock()
+	return l.inQueue.writeReadiness() | l.outQueue.readReadiness()
+}
+
+func (l *lineDiscipline) slaveReadiness() waiter.EventMask {
+	l.inMu.Lock()
+	defer l.inMu.Unlock()
+	l.outMu.Lock()
+	defer l.outMu.Unlock()
+	return l.outQueue.writeReadiness() | l.inQueue.readReadiness()
+}
+
+// queue represents one of the input or output queues between a pty master and
+// slave.
+type queue struct {
+	waiter.Queue `state:"nosave"`
+	buf          bytes.Buffer `state:".([]byte)"`
+}
+
+// saveBuf is invoked by stateify.
+func (q *queue) saveBuf() []byte {
+	return append([]byte(nil), q.buf.Bytes()...)
+}
+
+// loadBuf is invoked by stateify.
+func (q *queue) loadBuf(b []byte) {
+	q.buf.Write(b)
+}
+
+// readReadiness returns whether q is ready to be read from.
+//
+// Preconditions: q's mutex must be held.
+func (q *queue) readReadiness() waiter.EventMask {
+	ready := waiter.EventMask(0)
+	if q.buf.Len() > 0 {
+		ready |= waiter.EventIn
+	}
+	return ready
+}
+
+// writeReadiness returns whether q is ready to be written to.
+func (q *queue) writeReadiness() waiter.EventMask {
+	return waiter.EventOut
+}
+
+func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	l.inMu.Lock()
+	defer l.inMu.Unlock()
+	return l.queueRead(ctx, dst, &l.inQueue)
+}
+
+func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	l.inMu.Lock()
+	defer l.inMu.Unlock()
+	return l.queueWrite(ctx, src, &l.inQueue, false)
+}
+
+func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	l.outMu.Lock()
+	defer l.outMu.Unlock()
+	return l.queueRead(ctx, dst, &l.outQueue)
+}
+
+func (l *lineDiscipline) outputQueueWrite(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	l.outMu.Lock()
+	defer l.outMu.Unlock()
+	return l.queueWrite(ctx, src, &l.outQueue, true)
+}
+
+// queueRead reads from q to userspace.
+//
+// Preconditions: q's lock must be held.
+func (l *lineDiscipline) queueRead(ctx context.Context, dst usermem.IOSequence, q *queue) (int64, error) {
+	// Copy bytes out to user-space. queueRead doesn't have to do any
+	// processing or other extra work -- that's all taken care of when
+	// writing to a queue.
+	n, err := q.buf.WriteTo(dst.Writer(ctx))
+
+	// If state changed, notify any waiters. If nothing was available to
+	// read, let the caller know we could block.
+	if n > 0 {
+		q.Notify(waiter.EventOut)
+	} else if err == nil {
+		return 0, syserror.ErrWouldBlock
+	}
+	return int64(n), err
+}
+
+// queueWrite writes to q from userspace. `output` is whether the queue being
+// written to should be subject to output processing (i.e. whether it is the
+// output queue).
+//
+// Precondition: q's lock must be held.
+func (l *lineDiscipline) queueWrite(ctx context.Context, src usermem.IOSequence, q *queue, output bool) (int64, error) {
+	// TODO: Use CopyInTo/safemem to avoid extra copying.
+	// Get the bytes to write from user-space.
+	b := make([]byte, src.NumBytes())
+	n, err := src.CopyIn(ctx, b)
+	if err != nil {
+		return 0, err
+	}
+	b = b[:n]
+
+	// If state changed, notify any waiters. If we were unable to write
+	// anything, let the caller know we could block.
+	if n > 0 {
+		q.Notify(waiter.EventIn)
+	} else {
+		return 0, syserror.ErrWouldBlock
+	}
+
+	// Optionally perform line discipline transformations depending on
+	// whether we're writing to the input queue or output queue.
+	var buf *bytes.Buffer
+	l.termiosMu.Lock()
+	if output {
+		buf = l.transformOutput(b)
+	} else {
+		buf = l.transformInput(b)
+	}
+	l.termiosMu.Unlock()
+
+	// Enqueue buf at the end of the queue.
+	buf.WriteTo(&q.buf)
+	return int64(n), err
+}
+
+// transformOutput does ouput processing for one end of the pty. See
+// drivers/tty/n_tty.c:do_output_char for an analagous kernel function.
+//
+// Precondition: l.termiosMu must be held.
+func (l *lineDiscipline) transformOutput(buf []byte) *bytes.Buffer {
+	if !l.termios.OEnabled(linux.OPOST) {
+		return bytes.NewBuffer(buf)
+	}
+
+	var ret bytes.Buffer
+	for len(buf) > 0 {
+		c := l.removeRune(&buf)
+		switch c {
+		case '\n':
+			if l.termios.OEnabled(linux.ONLRET) {
+				l.column = 0
+			}
+			if l.termios.OEnabled(linux.ONLCR) {
+				ret.Write([]byte{'\r', '\n'})
+				continue
+			}
+		case '\r':
+			if l.termios.OEnabled(linux.ONOCR) && l.column == 0 {
+				continue
+			}
+			if l.termios.OEnabled(linux.OCRNL) {
+				c = '\n'
+				if l.termios.OEnabled(linux.ONLRET) {
+					l.column = 0
+				}
+				break
+			}
+			l.column = 0
+		case '\t':
+			spaces := spacesPerTab - l.column%spacesPerTab
+			if l.termios.OutputFlags&linux.TABDLY == linux.XTABS {
+				l.column += spaces
+				ret.Write(bytes.Repeat([]byte{' '}, 8))
+				continue
+			}
+			l.column += spaces
+		case '\b':
+			if l.column > 0 {
+				l.column--
+			}
+		default:
+			l.column++
+		}
+		ret.WriteRune(c)
+	}
+	return &ret
+}
+
+// transformInput does input processing for one end of the pty. Characters
+// read are transformed according to flags set in the termios struct. See
+// drivers/tty/n_tty.c:n_tty_receive_char_special for an analogous kernel
+// function.
+//
+// Precondition: l.termiosMu must be held.
+func (l *lineDiscipline) transformInput(buf []byte) *bytes.Buffer {
+	var ret bytes.Buffer
+	for len(buf) > 0 {
+		c := l.removeRune(&buf)
+		switch c {
+		case '\r':
+			if l.termios.IEnabled(linux.IGNCR) {
+				continue
+			}
+			if l.termios.IEnabled(linux.ICRNL) {
+				c = '\n'
+			}
+		case '\n':
+			if l.termios.IEnabled(linux.INLCR) {
+				c = '\r'
+			}
+		}
+		ret.WriteRune(c)
+	}
+	return &ret
+}
+
+// removeRune removes and returns the first rune from the byte array. The
+// buffer's length is updated accordingly.
+func (l *lineDiscipline) removeRune(b *[]byte) rune {
+	var c rune
+	var size int
+	// If UTF-8 support is enabled, runes might be multiple bytes.
+	if l.termios.IEnabled(linux.IUTF8) {
+		c, size = utf8.DecodeRune(*b)
+	} else {
+		c = rune((*b)[0])
+		size = 1
+	}
+	*b = (*b)[size:]
+	return c
+}
diff --git a/pkg/sentry/fs/tty/master.go b/pkg/sentry/fs/tty/master.go
new file mode 100644
index 000000000..3c47ee517
--- /dev/null
+++ b/pkg/sentry/fs/tty/master.go
@@ -0,0 +1,173 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// masterInodeOperations are the fs.InodeOperations for the master end of the
+// Terminal (ptmx file).
+type masterInodeOperations struct {
+	inodeOperations
+
+	// d is the containing dir.
+	d *dirInodeOperations
+}
+
+var _ fs.InodeOperations = (*masterInodeOperations)(nil)
+
+// newMasterInode creates an Inode for the master end of a terminal.
+func newMasterInode(ctx context.Context, d *dirInodeOperations, owner fs.FileOwner, p fs.FilePermissions) *fs.Inode {
+	iops := &masterInodeOperations{
+		inodeOperations: inodeOperations{
+			uattr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+				Owner: owner,
+				Perms: p,
+				Links: 1,
+				// Size and Blocks are always 0.
+			}),
+		},
+		d: d,
+	}
+
+	return fs.NewInode(iops, d.msrc, fs.StableAttr{
+		DeviceID: ptsDevice.DeviceID(),
+		// N.B. Linux always uses inode id 2 for ptmx. See
+		// fs/devpts/inode.c:mknod_ptmx.
+		//
+		// TODO: Since ptsDevice must be shared between
+		// different mounts, we must not assign fixed numbers.
+		InodeID: ptsDevice.NextIno(),
+		Type:    fs.CharacterDevice,
+		// See fs/devpts/inode.c:devpts_fill_super.
+		BlockSize: 1024,
+		// The PTY master effectively has two different major/minor
+		// device numbers.
+		//
+		// This one is returned by stat for both opened and unopened
+		// instances of this inode.
+		//
+		// When the inode is opened (GetFile), a new device number is
+		// allocated based on major UNIX98_PTY_MASTER_MAJOR and the tty
+		// index as minor number. However, this device number is only
+		// accessible via ioctl(TIOCGDEV) and /proc/TID/stat.
+		DeviceFileMajor: linux.TTYAUX_MAJOR,
+		DeviceFileMinor: linux.PTMX_MINOR,
+	})
+}
+
+// Release implements fs.InodeOperations.Release.
+func (mi *masterInodeOperations) Release(ctx context.Context) {
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+//
+// It allocates a new terminal.
+func (mi *masterInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	t, err := mi.d.allocateTerminal(ctx)
+	if err != nil {
+		return nil, err
+	}
+
+	return fs.NewFile(ctx, d, flags, &masterFileOperations{
+		d: mi.d,
+		t: t,
+	}), nil
+}
+
+// masterFileOperations are the fs.FileOperations for the master end of a terminal.
+type masterFileOperations struct {
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+
+	// d is the containing dir.
+	d *dirInodeOperations
+
+	// t is the connected Terminal.
+	t *Terminal
+}
+
+var _ fs.FileOperations = (*masterFileOperations)(nil)
+
+// Release implements fs.FileOperations.Release.
+func (mf *masterFileOperations) Release() {
+	mf.d.masterClose(mf.t)
+	mf.t.DecRef()
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (mf *masterFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	mf.t.ld.inQueue.EventRegister(e, mask)
+	mf.t.ld.outQueue.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (mf *masterFileOperations) EventUnregister(e *waiter.Entry) {
+	mf.t.ld.inQueue.EventUnregister(e)
+	mf.t.ld.outQueue.EventUnregister(e)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (mf *masterFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return mf.t.ld.masterReadiness()
+}
+
+// Read implements fs.FileOperations.Read.
+func (mf *masterFileOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	return mf.t.ld.outputQueueRead(ctx, dst)
+}
+
+// Write implements fs.FileOperations.Write.
+func (mf *masterFileOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	return mf.t.ld.inputQueueWrite(ctx, src)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (mf *masterFileOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch args[1].Uint() {
+	case linux.TCGETS:
+		// N.B. TCGETS on the master actually returns the configuration
+		// of the slave end.
+		return mf.t.ld.getTermios(ctx, io, args)
+	case linux.TCSETS:
+		// N.B. TCSETS on the master actually affects the configuration
+		// of the slave end.
+		return mf.t.ld.setTermios(ctx, io, args)
+	case linux.TCSETSW:
+		// TODO: This should drain the output queue first.
+		return mf.t.ld.setTermios(ctx, io, args)
+	case linux.TIOCGPTN:
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mf.t.n), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	case linux.TIOCSPTLCK:
+		// TODO: Implement pty locking. For now just pretend we do.
+		return 0, nil
+	default:
+		return 0, syserror.ENOTTY
+	}
+}
diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/slave.go
new file mode 100644
index 000000000..9178071a4
--- /dev/null
+++ b/pkg/sentry/fs/tty/slave.go
@@ -0,0 +1,151 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// slaveInodeOperations are the fs.InodeOperations for the slave end of the
+// Terminal (pts file).
+type slaveInodeOperations struct {
+	inodeOperations
+
+	// d is the containing dir.
+	d *dirInodeOperations
+
+	// t is the connected Terminal.
+	t *Terminal
+}
+
+var _ fs.InodeOperations = (*slaveInodeOperations)(nil)
+
+// newSlaveInode creates an fs.Inode for the slave end of a terminal.
+//
+// newSlaveInode takes ownership of t.
+func newSlaveInode(ctx context.Context, d *dirInodeOperations, t *Terminal, owner fs.FileOwner, p fs.FilePermissions) *fs.Inode {
+	iops := &slaveInodeOperations{
+		inodeOperations: inodeOperations{
+			uattr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+				Owner: owner,
+				Perms: p,
+				Links: 1,
+				// Size and Blocks are always 0.
+			}),
+		},
+		d: d,
+		t: t,
+	}
+
+	return fs.NewInode(iops, d.msrc, fs.StableAttr{
+		DeviceID: ptsDevice.DeviceID(),
+		// N.B. Linux always uses inode id = tty index + 3. See
+		// fs/devpts/inode.c:devpts_pty_new.
+		//
+		// TODO: Since ptsDevice must be shared between
+		// different mounts, we must not assign fixed numbers.
+		InodeID: ptsDevice.NextIno(),
+		Type:    fs.CharacterDevice,
+		// See fs/devpts/inode.c:devpts_fill_super.
+		BlockSize:       1024,
+		DeviceFileMajor: linux.UNIX98_PTY_SLAVE_MAJOR,
+		DeviceFileMinor: t.n,
+	})
+}
+
+// Release implements fs.InodeOperations.Release.
+func (si *slaveInodeOperations) Release(ctx context.Context) {
+	si.t.DecRef()
+}
+
+// GetFile implements fs.InodeOperations.GetFile.
+//
+// This may race with destruction of the terminal. If the terminal is gone, it
+// returns ENOENT.
+func (si *slaveInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	return fs.NewFile(ctx, d, flags, &slaveFileOperations{si: si}), nil
+}
+
+// slaveFileOperations are the fs.FileOperations for the slave end of a terminal.
+type slaveFileOperations struct {
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+
+	// si is the inode operations.
+	si *slaveInodeOperations
+}
+
+var _ fs.FileOperations = (*slaveFileOperations)(nil)
+
+// Release implements fs.FileOperations.Release.
+func (sf *slaveFileOperations) Release() {
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (sf *slaveFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	sf.si.t.ld.outQueue.EventRegister(e, mask)
+	sf.si.t.ld.inQueue.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (sf *slaveFileOperations) EventUnregister(e *waiter.Entry) {
+	sf.si.t.ld.outQueue.EventUnregister(e)
+	sf.si.t.ld.inQueue.EventUnregister(e)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (sf *slaveFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return sf.si.t.ld.slaveReadiness()
+}
+
+// Read implements fs.FileOperations.Read.
+func (sf *slaveFileOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	return sf.si.t.ld.inputQueueRead(ctx, dst)
+}
+
+// Write implements fs.FileOperations.Write.
+func (sf *slaveFileOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	return sf.si.t.ld.outputQueueWrite(ctx, src)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (sf *slaveFileOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch args[1].Uint() {
+	case linux.TCGETS:
+		return sf.si.t.ld.getTermios(ctx, io, args)
+	case linux.TCSETS:
+		return sf.si.t.ld.setTermios(ctx, io, args)
+	case linux.TCSETSW:
+		// TODO: This should drain the output queue first.
+		return sf.si.t.ld.setTermios(ctx, io, args)
+	case linux.TIOCGPTN:
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sf.si.t.n), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	default:
+		return 0, syserror.ENOTTY
+	}
+}
diff --git a/pkg/sentry/fs/tty/terminal.go b/pkg/sentry/fs/tty/terminal.go
new file mode 100644
index 000000000..6ae713a32
--- /dev/null
+++ b/pkg/sentry/fs/tty/terminal.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// Terminal is a pseudoterminal.
+type Terminal struct {
+	refs.AtomicRefCount
+
+	// n is the terminal index.
+	n uint32
+
+	// d is the containing directory.
+	d *dirInodeOperations
+
+	// ld is the line discipline of the terminal.
+	ld lineDiscipline
+}
+
+func newTerminal(ctx context.Context, d *dirInodeOperations, n uint32) *Terminal {
+	termios := linux.DefaultSlaveTermios
+	return &Terminal{
+		d:  d,
+		n:  n,
+		ld: lineDiscipline{termios: termios},
+	}
+}
diff --git a/pkg/sentry/fs/tty/tty_test.go b/pkg/sentry/fs/tty/tty_test.go
new file mode 100644
index 000000000..0c7560ed7
--- /dev/null
+++ b/pkg/sentry/fs/tty/tty_test.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package tty
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func TestSimpleMasterToSlave(t *testing.T) {
+	ld := lineDiscipline{termios: linux.DefaultSlaveTermios}
+	ctx := contexttest.Context(t)
+	inBytes := []byte("hello, tty\n")
+	src := usermem.BytesIOSequence(inBytes)
+	outBytes := make([]byte, 32)
+	dst := usermem.BytesIOSequence(outBytes)
+
+	// Write to the input queue.
+	nw, err := ld.inputQueueWrite(ctx, src)
+	if err != nil {
+		t.Fatalf("error writing to input queue: %v", err)
+	}
+	if nw != int64(len(inBytes)) {
+		t.Fatalf("wrote wrong length: got %d, want %d", nw, len(inBytes))
+	}
+
+	// Read from the input queue.
+	nr, err := ld.inputQueueRead(ctx, dst)
+	if err != nil {
+		t.Fatalf("error reading from input queue: %v", err)
+	}
+	if nr != int64(len(inBytes)) {
+		t.Fatalf("read wrong length: got %d, want %d", nr, len(inBytes))
+	}
+
+	outStr := string(outBytes[:nr])
+	inStr := string(inBytes)
+	if outStr != inStr {
+		t.Fatalf("written and read strings do not match: got %q, want %q", outStr, inStr)
+	}
+}
diff --git a/pkg/sentry/hostcpu/BUILD b/pkg/sentry/hostcpu/BUILD
new file mode 100644
index 000000000..9457618d8
--- /dev/null
+++ b/pkg/sentry/hostcpu/BUILD
@@ -0,0 +1,20 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "hostcpu",
+    srcs = [
+        "getcpu_amd64.s",
+        "hostcpu.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/hostcpu",
+    visibility = ["//:sandbox"],
+)
+
+go_test(
+    name = "hostcpu_test",
+    size = "small",
+    srcs = ["hostcpu_test.go"],
+    embed = [":hostcpu"],
+)
diff --git a/pkg/sentry/hostcpu/getcpu_amd64.s b/pkg/sentry/hostcpu/getcpu_amd64.s
new file mode 100644
index 000000000..7f6247d81
--- /dev/null
+++ b/pkg/sentry/hostcpu/getcpu_amd64.s
@@ -0,0 +1,24 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// func GetCPU() (cpu uint32)
+TEXT ·GetCPU(SB), NOSPLIT, $0-4
+	BYTE $0x0f; BYTE $0x01; BYTE $0xf9; // RDTSCP
+	// On Linux, the bottom 12 bits of IA32_TSC_AUX are CPU and the upper 20
+	// are node. See arch/x86/entry/vdso/vma.c:vgetcpu_cpu_init().
+	ANDL	$0xfff, CX
+	MOVL	CX, cpu+0(FP)
+	RET
diff --git a/pkg/sentry/hostcpu/hostcpu.go b/pkg/sentry/hostcpu/hostcpu.go
new file mode 100644
index 000000000..fa46499ad
--- /dev/null
+++ b/pkg/sentry/hostcpu/hostcpu.go
@@ -0,0 +1,67 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package hostcpu provides utilities for working with CPU information provided
+// by a host Linux kernel.
+package hostcpu
+
+import (
+	"fmt"
+	"io/ioutil"
+	"strconv"
+	"strings"
+	"unicode"
+)
+
+// GetCPU returns the caller's current CPU number, without using the Linux VDSO
+// (which is not available to the sentry) or the getcpu(2) system call (which
+// is relatively slow).
+func GetCPU() uint32
+
+// MaxPossibleCPU returns the highest possible CPU number, which is guaranteed
+// not to change for the lifetime of the host kernel.
+func MaxPossibleCPU() (uint32, error) {
+	const path = "/sys/devices/system/cpu/possible"
+	data, err := ioutil.ReadFile(path)
+	if err != nil {
+		return 0, err
+	}
+	str := string(data)
+	// Linux: drivers/base/cpu.c:show_cpus_attr() =>
+	// include/linux/cpumask.h:cpumask_print_to_pagebuf() =>
+	// lib/bitmap.c:bitmap_print_to_pagebuf()
+	i, err := maxValueInLinuxBitmap(str)
+	if err != nil {
+		return 0, fmt.Errorf("invalid %s (%q): %v", path, str, err)
+	}
+	return uint32(i), nil
+}
+
+// maxValueInLinuxBitmap returns the maximum value specified in str, which is a
+// string emitted by Linux's lib/bitmap.c:bitmap_print_to_pagebuf(list=true).
+func maxValueInLinuxBitmap(str string) (uint64, error) {
+	str = strings.TrimSpace(str)
+	// Find the last decimal number in str.
+	idx := strings.LastIndexFunc(str, func(c rune) bool {
+		return !unicode.IsDigit(c)
+	})
+	if idx != -1 {
+		str = str[idx+1:]
+	}
+	i, err := strconv.ParseUint(str, 10, 64)
+	if err != nil {
+		return 0, err
+	}
+	return i, nil
+}
diff --git a/pkg/sentry/hostcpu/hostcpu_test.go b/pkg/sentry/hostcpu/hostcpu_test.go
new file mode 100644
index 000000000..a82e1a271
--- /dev/null
+++ b/pkg/sentry/hostcpu/hostcpu_test.go
@@ -0,0 +1,52 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostcpu
+
+import (
+	"fmt"
+	"testing"
+)
+
+func TestMaxValueInLinuxBitmap(t *testing.T) {
+	for _, test := range []struct {
+		str string
+		max uint64
+	}{
+		{"0", 0},
+		{"0\n", 0},
+		{"0,2", 2},
+		{"0-63", 63},
+		{"0-3,8-11", 11},
+	} {
+		t.Run(fmt.Sprintf("%q", test.str), func(t *testing.T) {
+			max, err := maxValueInLinuxBitmap(test.str)
+			if err != nil || max != test.max {
+				t.Errorf("maxValueInLinuxBitmap: got (%d, %v), wanted (%d, nil)", max, err, test.max)
+			}
+		})
+	}
+}
+
+func TestMaxValueInLinuxBitmapErrors(t *testing.T) {
+	for _, str := range []string{"", "\n"} {
+		t.Run(fmt.Sprintf("%q", str), func(t *testing.T) {
+			max, err := maxValueInLinuxBitmap(str)
+			if err == nil {
+				t.Errorf("maxValueInLinuxBitmap: got (%d, nil), wanted (_, error)", max)
+			}
+			t.Log(err)
+		})
+	}
+}
diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD
new file mode 100644
index 000000000..207cdb692
--- /dev/null
+++ b/pkg/sentry/inet/BUILD
@@ -0,0 +1,28 @@
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+package(
+    default_visibility = ["//:sandbox"],
+    licenses = ["notice"],  # Apache 2.0
+)
+
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "inet_state",
+    srcs = ["inet.go"],
+    out = "inet_state.go",
+    package = "inet",
+)
+
+go_library(
+    name = "inet",
+    srcs = [
+        "inet.go",
+        "inet_state.go",
+        "test_stack.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/inet",
+    deps = [
+        "//pkg/state",
+    ],
+)
diff --git a/pkg/sentry/inet/inet.go b/pkg/sentry/inet/inet.go
new file mode 100644
index 000000000..e4b326993
--- /dev/null
+++ b/pkg/sentry/inet/inet.go
@@ -0,0 +1,99 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package inet defines semantics for IP stacks.
+package inet
+
+// Stack represents a TCP/IP stack.
+type Stack interface {
+	// Interfaces returns all network interfaces as a mapping from interface
+	// indexes to interface properties. Interface indices are strictly positive
+	// integers.
+	Interfaces() map[int32]Interface
+
+	// InterfaceAddrs returns all network interface addresses as a mapping from
+	// interface indexes to a slice of associated interface address properties.
+	InterfaceAddrs() map[int32][]InterfaceAddr
+
+	// SupportsIPv6 returns true if the stack supports IPv6 connectivity.
+	SupportsIPv6() bool
+
+	// TCPReceiveBufferSize returns TCP receive buffer size settings.
+	TCPReceiveBufferSize() (TCPBufferSize, error)
+
+	// SetTCPReceiveBufferSize attempts to change TCP receive buffer size
+	// settings.
+	SetTCPReceiveBufferSize(size TCPBufferSize) error
+
+	// TCPSendBufferSize returns TCP send buffer size settings.
+	TCPSendBufferSize() (TCPBufferSize, error)
+
+	// SetTCPSendBufferSize attempts to change TCP send buffer size settings.
+	SetTCPSendBufferSize(size TCPBufferSize) error
+
+	// TCPSACKEnabled returns true if RFC 2018 TCP Selective Acknowledgements
+	// are enabled.
+	TCPSACKEnabled() (bool, error)
+
+	// SetTCPSACKEnabled attempts to change TCP selective acknowledgement
+	// settings.
+	SetTCPSACKEnabled(enabled bool) error
+}
+
+// Interface contains information about a network interface.
+type Interface struct {
+	// Keep these fields sorted in the order they appear in rtnetlink(7).
+
+	// DeviceType is the device type, a Linux ARPHRD_* constant.
+	DeviceType uint16
+
+	// Flags is the device flags; see netdevice(7), under "Ioctls",
+	// "SIOCGIFFLAGS, SIOCSIFFLAGS".
+	Flags uint32
+
+	// Name is the device name.
+	Name string
+
+	// Addr is the hardware device address.
+	Addr []byte
+}
+
+// InterfaceAddr contains information about a network interface address.
+type InterfaceAddr struct {
+	// Keep these fields sorted in the order they appear in rtnetlink(7).
+
+	// Family is the address family, a Linux AF_* constant.
+	Family uint8
+
+	// PrefixLen is the address prefix length.
+	PrefixLen uint8
+
+	// Flags is the address flags.
+	Flags uint8
+
+	// Addr is the actual address.
+	Addr []byte
+}
+
+// TCPBufferSize contains settings controlling TCP buffer sizing.
+type TCPBufferSize struct {
+	// Min is the minimum size.
+	Min int
+
+	// Default is the default size.
+	Default int
+
+	// Max is the maximum size.
+	Max int
+}
diff --git a/pkg/sentry/inet/test_stack.go b/pkg/sentry/inet/test_stack.go
new file mode 100644
index 000000000..bc10926ee
--- /dev/null
+++ b/pkg/sentry/inet/test_stack.go
@@ -0,0 +1,83 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package inet
+
+// TestStack is a dummy implementation of Stack for tests.
+type TestStack struct {
+	InterfacesMap     map[int32]Interface
+	InterfaceAddrsMap map[int32][]InterfaceAddr
+	SupportsIPv6Flag  bool
+	TCPRecvBufSize    TCPBufferSize
+	TCPSendBufSize    TCPBufferSize
+	TCPSACKFlag       bool
+}
+
+// NewTestStack returns a TestStack with no network interfaces. The value of
+// all other options is unspecified; tests that rely on specific values must
+// set them explicitly.
+func NewTestStack() *TestStack {
+	return &TestStack{
+		InterfacesMap:     make(map[int32]Interface),
+		InterfaceAddrsMap: make(map[int32][]InterfaceAddr),
+	}
+}
+
+// Interfaces implements Stack.Interfaces.
+func (s *TestStack) Interfaces() map[int32]Interface {
+	return s.InterfacesMap
+}
+
+// InterfaceAddrs implements Stack.InterfaceAddrs.
+func (s *TestStack) InterfaceAddrs() map[int32][]InterfaceAddr {
+	return s.InterfaceAddrsMap
+}
+
+// SupportsIPv6 implements Stack.SupportsIPv6.
+func (s *TestStack) SupportsIPv6() bool {
+	return s.SupportsIPv6Flag
+}
+
+// TCPReceiveBufferSize implements Stack.TCPReceiveBufferSize.
+func (s *TestStack) TCPReceiveBufferSize() (TCPBufferSize, error) {
+	return s.TCPRecvBufSize, nil
+}
+
+// SetTCPReceiveBufferSize implements Stack.SetTCPReceiveBufferSize.
+func (s *TestStack) SetTCPReceiveBufferSize(size TCPBufferSize) error {
+	s.TCPRecvBufSize = size
+	return nil
+}
+
+// TCPSendBufferSize implements Stack.TCPSendBufferSize.
+func (s *TestStack) TCPSendBufferSize() (TCPBufferSize, error) {
+	return s.TCPSendBufSize, nil
+}
+
+// SetTCPSendBufferSize implements Stack.SetTCPSendBufferSize.
+func (s *TestStack) SetTCPSendBufferSize(size TCPBufferSize) error {
+	s.TCPSendBufSize = size
+	return nil
+}
+
+// TCPSACKEnabled implements Stack.TCPSACKEnabled.
+func (s *TestStack) TCPSACKEnabled() (bool, error) {
+	return s.TCPSACKFlag, nil
+}
+
+// SetTCPSACKEnabled implements Stack.SetTCPSACKEnabled.
+func (s *TestStack) SetTCPSACKEnabled(enabled bool) error {
+	s.TCPSACKFlag = enabled
+	return nil
+}
diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD
new file mode 100644
index 000000000..62794cff5
--- /dev/null
+++ b/pkg/sentry/kernel/BUILD
@@ -0,0 +1,234 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "kernel_state",
+    srcs = [
+        "abstract_socket_namespace.go",
+        "fd_map.go",
+        "fs_context.go",
+        "ipc_namespace.go",
+        "kernel.go",
+        "pending_signals.go",
+        "pending_signals_list.go",
+        "process_group_list.go",
+        "ptrace.go",
+        "rseq.go",
+        "session_list.go",
+        "sessions.go",
+        "signal.go",
+        "signal_handlers.go",
+        "syscalls.go",
+        "syscalls_state.go",
+        "syslog.go",
+        "task.go",
+        "task_clone.go",
+        "task_context.go",
+        "task_exec.go",
+        "task_exit.go",
+        "task_list.go",
+        "task_resources.go",
+        "task_run.go",
+        "task_sched.go",
+        "task_signals.go",
+        "task_start.go",
+        "task_syscall.go",
+        "thread_group.go",
+        "threads.go",
+        "timekeeper.go",
+        "timekeeper_state.go",
+        "timer.go",
+        "uts_namespace.go",
+        "vdso.go",
+        "version.go",
+    ],
+    out = "kernel_state.go",
+    imports = ["gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"],
+    package = "kernel",
+)
+
+go_template_instance(
+    name = "pending_signals_list",
+    out = "pending_signals_list.go",
+    package = "kernel",
+    prefix = "pendingSignal",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*pendingSignal",
+    },
+)
+
+go_template_instance(
+    name = "process_group_list",
+    out = "process_group_list.go",
+    package = "kernel",
+    prefix = "processGroup",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*ProcessGroup",
+    },
+)
+
+go_template_instance(
+    name = "seqatomic_taskgoroutineschedinfo",
+    out = "seqatomic_taskgoroutineschedinfo.go",
+    package = "kernel",
+    suffix = "TaskGoroutineSchedInfo",
+    template = "//pkg/sync:generic_seqatomic",
+    types = {
+        "Value": "TaskGoroutineSchedInfo",
+    },
+)
+
+go_template_instance(
+    name = "session_list",
+    out = "session_list.go",
+    package = "kernel",
+    prefix = "session",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*Session",
+    },
+)
+
+go_template_instance(
+    name = "task_list",
+    out = "task_list.go",
+    package = "kernel",
+    prefix = "task",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*Task",
+    },
+)
+
+go_library(
+    name = "kernel",
+    srcs = [
+        "abstract_socket_namespace.go",
+        "context.go",
+        "fd_map.go",
+        "fs_context.go",
+        "ipc_namespace.go",
+        "kernel.go",
+        "kernel_state.go",
+        "pending_signals.go",
+        "pending_signals_list.go",
+        "process_group_list.go",
+        "ptrace.go",
+        "rseq.go",
+        "seccomp.go",
+        "seqatomic_taskgoroutineschedinfo.go",
+        "session_list.go",
+        "sessions.go",
+        "signal.go",
+        "signal_handlers.go",
+        "syscalls.go",
+        "syscalls_state.go",
+        "syslog.go",
+        "task.go",
+        "task_acct.go",
+        "task_block.go",
+        "task_clone.go",
+        "task_context.go",
+        "task_exec.go",
+        "task_exit.go",
+        "task_identity.go",
+        "task_list.go",
+        "task_log.go",
+        "task_net.go",
+        "task_resources.go",
+        "task_run.go",
+        "task_sched.go",
+        "task_signals.go",
+        "task_start.go",
+        "task_stop.go",
+        "task_syscall.go",
+        "task_usermem.go",
+        "thread_group.go",
+        "threads.go",
+        "timekeeper.go",
+        "timekeeper_state.go",
+        "timer.go",
+        "uts_namespace.go",
+        "vdso.go",
+        "version.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel",
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/abi",
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/binary",
+        "//pkg/bits",
+        "//pkg/bpf",
+        "//pkg/cpuid",
+        "//pkg/eventchannel",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/secio",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/fs/timerfd",
+        "//pkg/sentry/hostcpu",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/epoll",
+        "//pkg/sentry/kernel/futex",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/sched",
+        "//pkg/sentry/kernel/semaphore",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/loader",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/mm",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket/netlink/port",
+        "//pkg/sentry/time",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/state/statefile",
+        "//pkg/sync",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/stack",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "kernel_test",
+    size = "small",
+    srcs = [
+        "fd_map_test.go",
+        "table_test.go",
+        "task_test.go",
+        "timekeeper_test.go",
+    ],
+    embed = [":kernel"],
+    deps = [
+        "//pkg/abi",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs/filetest",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/sched",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/time",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/kernel/README.md b/pkg/sentry/kernel/README.md
new file mode 100644
index 000000000..3306780d6
--- /dev/null
+++ b/pkg/sentry/kernel/README.md
@@ -0,0 +1,106 @@
+This package contains:
+
+- A (partial) emulation of the "core Linux kernel", which governs task
+  execution and scheduling, system call dispatch, and signal handling. See
+  below for details.
+
+- The top-level interface for the sentry's Linux kernel emulation in general,
+  used by the `main` function of all versions of the sentry. This interface
+  revolves around the `Env` type (defined in `kernel.go`).
+
+# Background
+
+In Linux, each schedulable context is referred to interchangeably as a "task" or
+"thread". Tasks can be divided into userspace and kernel tasks. In the sentry,
+scheduling is managed by the Go runtime, so each schedulable context is a
+goroutine; only "userspace" (application) contexts are referred to as tasks, and
+represented by Task objects. (From this point forward, "task" refers to the
+sentry's notion of a task unless otherwise specified.)
+
+At a high level, Linux application threads can be thought of as repeating a "run
+loop":
+
+- Some amount of application code is executed in userspace.
+
+- A trap (explicit syscall invocation, hardware interrupt or exception, etc.)
+  causes control flow to switch to the kernel.
+
+- Some amount of kernel code is executed in kernelspace, e.g. to handle the
+  cause of the trap.
+
+- The kernel "returns from the trap" into application code.
+
+Analogously, each task in the sentry is associated with a *task goroutine* that
+executes that task's run loop (`Task.run` in `task_run.go`). However, the
+sentry's task run loop differs in structure in order to support saving execution
+state to, and resuming execution from, checkpoints.
+
+While in kernelspace, a Linux thread can be descheduled (cease execution) in a
+variety of ways:
+
+- It can yield or be preempted, becoming temporarily descheduled but still
+  runnable. At present, the sentry delegates scheduling of runnable threads to
+  the Go runtime.
+
+- It can exit, becoming permanently descheduled. The sentry's equivalent is
+  returning from `Task.run`, terminating the task goroutine.
+
+- It can enter interruptible sleep, a state in which it can be woken by a
+  caller-defined wakeup or the receipt of a signal. In the sentry, interruptible
+  sleep (which is ambiguously referred to as *blocking*) is implemented by
+  making all events that can end blocking (including signal notifications)
+  communicated via Go channels and using `select` to multiplex wakeup sources;
+  see `task_block.go`.
+
+- It can enter uninterruptible sleep, a state in which it can only be woken by a
+  caller-defined wakeup. Killable sleep is a closely related variant in which
+  the task can also be woken by SIGKILL. (These definitions also include Linux's
+  "group-stopped" (`TASK_STOPPED`) and "ptrace-stopped" (`TASK_TRACED`) states.)
+
+To maximize compatibility with Linux, sentry checkpointing appears as a spurious
+signal-delivery interrupt on all tasks; interrupted system calls return `EINTR`
+or are automatically restarted as usual. However, these semantics require that
+uninterruptible and killable sleeps do not appear to be interrupted. In other
+words, the state of the task, including its progress through the interrupted
+operation, must be preserved by checkpointing. For many such sleeps, the wakeup
+condition is application-controlled, making it infeasible to wait for the sleep
+to end before checkpointing. Instead, we must support checkpointing progress
+through sleeping operations.
+
+# Implementation
+
+We break the task's control flow graph into *states*, delimited by:
+
+1. Points where uninterruptible and killable sleeps may occur. For example,
+there exists a state boundary between signal dequeueing and signal delivery
+because there may be an intervening ptrace signal-delivery-stop.
+
+2. Points where sleep-induced branches may "rejoin" normal execution. For
+example, the syscall exit state exists because it can be reached immediately
+following a synchronous syscall, or after a task that is sleeping in `execve()`
+or `vfork()` resumes execution.
+
+3. Points containing large branches. This is strictly for organizational
+purposes. For example, the state that processes interrupt-signaled conditions is
+kept separate from the main "app" state to reduce the size of the latter.
+
+4. `SyscallReinvoke`, which does not correspond to anything in Linux, and exists
+solely to serve the autosave feature.
+
+![dot -Tsvg -Goverlap=false -orun_states.svg run_states.dot](g3doc/run_states.dot "Task control flow graph")
+
+States before which a stop may occur are represented as implementations of the
+`taskRunState` interface named `run(state)`, allowing them to be saved and
+restored. States that cannot be immediately preceded by a stop are simply `Task`
+methods named `do(state)`.
+
+Conditions that can require task goroutines to cease execution for unknown
+lengths of time are called *stops*. Stops are divided into *internal stops*,
+which are stops whose start and end conditions are implemented within the
+sentry, and *external stops*, which are stops whose start and end conditions are
+not known to the sentry. Hence all uninterruptible and killable sleeps are
+internal stops, and the existence of a pending checkpoint operation is an
+external stop. Internal stops are reified into instances of the `TaskStop` type,
+while external stops are merely counted. The task run loop alternates between
+checking for stops and advancing the task's state. This allows checkpointing to
+hold tasks in a stopped state while waiting for all tasks in the system to stop.
diff --git a/pkg/sentry/kernel/abstract_socket_namespace.go b/pkg/sentry/kernel/abstract_socket_namespace.go
new file mode 100644
index 000000000..014c4a3bf
--- /dev/null
+++ b/pkg/sentry/kernel/abstract_socket_namespace.go
@@ -0,0 +1,108 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+type abstractEndpoint struct {
+	ep   unix.BoundEndpoint
+	wr   *refs.WeakRef
+	name string
+	ns   *AbstractSocketNamespace
+}
+
+// WeakRefGone implements refs.WeakRefUser.WeakRefGone.
+func (e *abstractEndpoint) WeakRefGone() {
+	e.ns.mu.Lock()
+	if e.ns.endpoints[e.name].ep == e.ep {
+		delete(e.ns.endpoints, e.name)
+	}
+	e.ns.mu.Unlock()
+}
+
+// AbstractSocketNamespace is used to implement the Linux abstract socket functionality.
+type AbstractSocketNamespace struct {
+	mu sync.Mutex `state:"nosave"`
+
+	// Keeps mapping from name to endpoint.
+	endpoints map[string]abstractEndpoint
+}
+
+// NewAbstractSocketNamespace returns a new AbstractSocketNamespace.
+func NewAbstractSocketNamespace() *AbstractSocketNamespace {
+	return &AbstractSocketNamespace{
+		endpoints: make(map[string]abstractEndpoint),
+	}
+}
+
+// A boundEndpoint wraps a unix.BoundEndpoint to maintain a reference on its
+// backing object.
+type boundEndpoint struct {
+	unix.BoundEndpoint
+	rc refs.RefCounter
+}
+
+// Release implements unix.BoundEndpoint.Release.
+func (e *boundEndpoint) Release() {
+	e.rc.DecRef()
+	e.BoundEndpoint.Release()
+}
+
+// BoundEndpoint retrieves the endpoint bound to the given name. The return
+// value is nil if no endpoint was bound.
+func (a *AbstractSocketNamespace) BoundEndpoint(name string) unix.BoundEndpoint {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+
+	ep, ok := a.endpoints[name]
+	if !ok {
+		return nil
+	}
+
+	rc := ep.wr.Get()
+	if rc == nil {
+		delete(a.endpoints, name)
+		return nil
+	}
+
+	return &boundEndpoint{ep.ep, rc}
+}
+
+// Bind binds the given socket.
+//
+// When the last reference managed by rc is dropped, ep may be removed from the
+// namespace.
+func (a *AbstractSocketNamespace) Bind(name string, ep unix.BoundEndpoint, rc refs.RefCounter) error {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+
+	if ep, ok := a.endpoints[name]; ok {
+		if rc := ep.wr.Get(); rc != nil {
+			rc.DecRef()
+			return syscall.EADDRINUSE
+		}
+	}
+
+	ae := abstractEndpoint{ep: ep, name: name, ns: a}
+	ae.wr = refs.NewWeakRef(rc, &ae)
+	a.endpoints[name] = ae
+	return nil
+}
diff --git a/pkg/sentry/kernel/auth/BUILD b/pkg/sentry/kernel/auth/BUILD
new file mode 100644
index 000000000..7f0680b88
--- /dev/null
+++ b/pkg/sentry/kernel/auth/BUILD
@@ -0,0 +1,73 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "auth_state",
+    srcs = [
+        "credentials.go",
+        "id.go",
+        "id_map_range.go",
+        "id_map_set.go",
+        "user_namespace.go",
+    ],
+    out = "auth_state.go",
+    package = "auth",
+)
+
+go_template_instance(
+    name = "id_map_range",
+    out = "id_map_range.go",
+    package = "auth",
+    prefix = "idMap",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "uint32",
+    },
+)
+
+go_template_instance(
+    name = "id_map_set",
+    out = "id_map_set.go",
+    consts = {
+        "minDegree": "3",
+    },
+    package = "auth",
+    prefix = "idMap",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint32",
+        "Range": "idMapRange",
+        "Value": "uint32",
+        "Functions": "idMapFunctions",
+    },
+)
+
+go_library(
+    name = "auth",
+    srcs = [
+        "auth.go",
+        "auth_state.go",
+        "capability_set.go",
+        "context.go",
+        "credentials.go",
+        "id.go",
+        "id_map.go",
+        "id_map_functions.go",
+        "id_map_range.go",
+        "id_map_set.go",
+        "user_namespace.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/bits",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/kernel/auth/auth.go b/pkg/sentry/kernel/auth/auth.go
new file mode 100644
index 000000000..c49a6b852
--- /dev/null
+++ b/pkg/sentry/kernel/auth/auth.go
@@ -0,0 +1,22 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package auth implements an access control model that is a subset of Linux's.
+//
+// The auth package supports two kinds of access controls: user/group IDs and
+// capabilities. Each resource in the security model is associated with a user
+// namespace; "privileged" operations check that the operator's credentials
+// have the required user/group IDs or capabilities within the user namespace
+// of accessed resources.
+package auth
diff --git a/pkg/sentry/kernel/auth/capability_set.go b/pkg/sentry/kernel/auth/capability_set.go
new file mode 100644
index 000000000..5b8164c49
--- /dev/null
+++ b/pkg/sentry/kernel/auth/capability_set.go
@@ -0,0 +1,61 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+)
+
+// A CapabilitySet is a set of capabilities implemented as a bitset. The zero
+// value of CapabilitySet is a set containing no capabilities.
+type CapabilitySet uint64
+
+// AllCapabilities is a CapabilitySet containing all valid capabilities.
+var AllCapabilities = CapabilitySetOf(linux.MaxCapability+1) - 1
+
+// CapabilitySetOf returns a CapabilitySet containing only the given
+// capability.
+func CapabilitySetOf(cp linux.Capability) CapabilitySet {
+	return CapabilitySet(bits.MaskOf64(int(cp)))
+}
+
+// CapabilitySetOfMany returns a CapabilitySet containing the given capabilities.
+func CapabilitySetOfMany(cps []linux.Capability) CapabilitySet {
+	var cs uint64
+	for _, cp := range cps {
+		cs |= bits.MaskOf64(int(cp))
+	}
+	return CapabilitySet(cs)
+}
+
+// TaskCapabilities represents all the capability sets for a task. Each of these
+// sets is explained in greater detail in capabilities(7).
+type TaskCapabilities struct {
+	// Permitted is a limiting superset for the effective capabilities that
+	// the thread may assume.
+	PermittedCaps CapabilitySet
+	// Inheritable is a set of capabilities preserved across an execve(2).
+	InheritableCaps CapabilitySet
+	// Effective is the set of capabilities used by the kernel to perform
+	// permission checks for the thread.
+	EffectiveCaps CapabilitySet
+	// Bounding is a limiting superset for the capabilities that a thread
+	// can add to its inheritable set using capset(2).
+	BoundingCaps CapabilitySet
+	// Ambient is a set of capabilities that are preserved across an
+	// execve(2) of a program that is not privileged.
+	AmbientCaps CapabilitySet
+}
diff --git a/pkg/sentry/kernel/auth/context.go b/pkg/sentry/kernel/auth/context.go
new file mode 100644
index 000000000..914589b28
--- /dev/null
+++ b/pkg/sentry/kernel/auth/context.go
@@ -0,0 +1,36 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the auth package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxCredentials is a Context.Value key for Credentials.
+	CtxCredentials contextID = iota
+)
+
+// CredentialsFromContext returns a copy of the Credentials used by ctx, or a
+// set of Credentials with no capabilities if ctx does not have Credentials.
+func CredentialsFromContext(ctx context.Context) *Credentials {
+	if v := ctx.Value(CtxCredentials); v != nil {
+		return v.(*Credentials)
+	}
+	return NewAnonymousCredentials()
+}
diff --git a/pkg/sentry/kernel/auth/credentials.go b/pkg/sentry/kernel/auth/credentials.go
new file mode 100644
index 000000000..b832b28fe
--- /dev/null
+++ b/pkg/sentry/kernel/auth/credentials.go
@@ -0,0 +1,227 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Credentials contains information required to authorize privileged operations
+// in a user namespace.
+type Credentials struct {
+	// Real/effective/saved user/group IDs in the root user namespace. None of
+	// these should ever be NoID.
+	RealKUID      KUID
+	EffectiveKUID KUID
+	SavedKUID     KUID
+	RealKGID      KGID
+	EffectiveKGID KGID
+	SavedKGID     KGID
+
+	// Filesystem user/group IDs are not implemented. "... setfsuid() is
+	// nowadays unneeded and should be avoided in new applications (likewise
+	// for setfsgid(2))." - setfsuid(2)
+
+	// Supplementary groups used by set/getgroups.
+	//
+	// ExtraKGIDs slices are immutable, allowing multiple Credentials with the
+	// same ExtraKGIDs to share the same slice.
+	ExtraKGIDs []KGID
+
+	// The capability sets applicable to this set of credentials.
+	PermittedCaps   CapabilitySet
+	InheritableCaps CapabilitySet
+	EffectiveCaps   CapabilitySet
+	BoundingCaps    CapabilitySet
+	// Ambient capabilities are not introduced until Linux 4.3.
+
+	// KeepCaps is the flag for PR_SET_KEEPCAPS which allow capabilities to be
+	// maintained after a switch from root user to non-root user via setuid().
+	KeepCaps bool
+
+	// The user namespace associated with the owner of the credentials.
+	UserNamespace *UserNamespace
+}
+
+// NewAnonymousCredentials returns a set of credentials with no capabilities in
+// any user namespace.
+func NewAnonymousCredentials() *Credentials {
+	// Create a new root user namespace. Since the new namespace's owner is
+	// KUID 0 and the returned credentials have non-zero KUID/KGID, the
+	// returned credentials do not have any capabilities in the new namespace.
+	// Since the new namespace is not part of any existing user namespace
+	// hierarchy, the returned credentials do not have any capabilities in any
+	// other namespace.
+	return &Credentials{
+		RealKUID:      NobodyKUID,
+		EffectiveKUID: NobodyKUID,
+		SavedKUID:     NobodyKUID,
+		RealKGID:      NobodyKGID,
+		EffectiveKGID: NobodyKGID,
+		SavedKGID:     NobodyKGID,
+		UserNamespace: NewRootUserNamespace(),
+	}
+}
+
+// NewRootCredentials returns a set of credentials with KUID and KGID 0 (i.e.
+// global root) in user namespace ns.
+func NewRootCredentials(ns *UserNamespace) *Credentials {
+	// I can't find documentation for this anywhere, but it's correct for the
+	// inheritable capability set to be initially empty (the capabilities test
+	// checks for this property).
+	return &Credentials{
+		RealKUID:      RootKUID,
+		EffectiveKUID: RootKUID,
+		SavedKUID:     RootKUID,
+		RealKGID:      RootKGID,
+		EffectiveKGID: RootKGID,
+		SavedKGID:     RootKGID,
+		PermittedCaps: AllCapabilities,
+		EffectiveCaps: AllCapabilities,
+		BoundingCaps:  AllCapabilities,
+		UserNamespace: ns,
+	}
+}
+
+// NewUserCredentials returns a set of credentials based on the given UID, GIDs,
+// and capabilities in a given namespace. If all arguments are their zero
+// values, this returns the same credentials as NewRootCredentials.
+func NewUserCredentials(kuid KUID, kgid KGID, extraKGIDs []KGID, capabilities *TaskCapabilities, ns *UserNamespace) *Credentials {
+	creds := NewRootCredentials(ns)
+
+	// Set the UID.
+	uid := kuid
+	creds.RealKUID = uid
+	creds.EffectiveKUID = uid
+	creds.SavedKUID = uid
+
+	// Set GID.
+	gid := kgid
+	creds.RealKGID = gid
+	creds.EffectiveKGID = gid
+	creds.SavedKGID = gid
+
+	// Set additional GIDs.
+	creds.ExtraKGIDs = append(creds.ExtraKGIDs, extraKGIDs...)
+
+	// Set capabilities. If capabilities aren't specified, we default to
+	// all capabilities.
+	if capabilities != nil {
+		creds.PermittedCaps = capabilities.PermittedCaps
+		creds.EffectiveCaps = capabilities.EffectiveCaps
+		creds.BoundingCaps = capabilities.BoundingCaps
+		creds.InheritableCaps = capabilities.InheritableCaps
+		// // TODO: Support ambient capabilities.
+	} else {
+		// If no capabilities are specified, grant the same capabilites
+		// that NewRootCredentials does.
+		creds.PermittedCaps = AllCapabilities
+		creds.EffectiveCaps = AllCapabilities
+		creds.BoundingCaps = AllCapabilities
+	}
+
+	return creds
+}
+
+// Fork generates an identical copy of a set of credentials.
+func (c *Credentials) Fork() *Credentials {
+	nc := new(Credentials)
+	*nc = *c // Copy-by-value; this is legal for all fields.
+	return nc
+}
+
+// InGroup returns true if c is in group kgid. Compare Linux's
+// kernel/groups.c:in_group_p().
+func (c *Credentials) InGroup(kgid KGID) bool {
+	if c.EffectiveKGID == kgid {
+		return true
+	}
+	for _, extraKGID := range c.ExtraKGIDs {
+		if extraKGID == kgid {
+			return true
+		}
+	}
+	return false
+}
+
+// HasCapabilityIn returns true if c has capability cp in ns.
+func (c *Credentials) HasCapabilityIn(cp linux.Capability, ns *UserNamespace) bool {
+	for {
+		// "1. A process has a capability inside a user namespace if it is a member
+		// of that namespace and it has the capability in its effective capability
+		// set." - user_namespaces(7)
+		if c.UserNamespace == ns {
+			return CapabilitySetOf(cp)&c.EffectiveCaps != 0
+		}
+		// "3. ... A process that resides in the parent of the user namespace and
+		// whose effective user ID matches the owner of the namespace has all
+		// capabilities in the namespace."
+		if c.UserNamespace == ns.parent && c.EffectiveKUID == ns.owner {
+			return true
+		}
+		// "2. If a process has a capability in a user namespace, then it has that
+		// capability in all child (and further removed descendant) namespaces as
+		// well."
+		if ns.parent == nil {
+			return false
+		}
+		ns = ns.parent
+	}
+}
+
+// HasCapability returns true if c has capability cp in its user namespace.
+func (c *Credentials) HasCapability(cp linux.Capability) bool {
+	return c.HasCapabilityIn(cp, c.UserNamespace)
+}
+
+// UseUID checks that c can use uid in its user namespace, then translates it
+// to the root user namespace.
+//
+// The checks UseUID does are common, but you should verify that it's doing
+// exactly what you want.
+func (c *Credentials) UseUID(uid UID) (KUID, error) {
+	// uid must be mapped.
+	kuid := c.UserNamespace.MapToKUID(uid)
+	if !kuid.Ok() {
+		return NoID, syserror.EINVAL
+	}
+	// If c has CAP_SETUID, then it can use any UID in its user namespace.
+	if c.HasCapability(linux.CAP_SETUID) {
+		return kuid, nil
+	}
+	// Otherwise, c must already have the UID as its real, effective, or saved
+	// set-user-ID.
+	if kuid == c.RealKUID || kuid == c.EffectiveKUID || kuid == c.SavedKUID {
+		return kuid, nil
+	}
+	return NoID, syserror.EPERM
+}
+
+// UseGID checks that c can use gid in its user namespace, then translates it
+// to the root user namespace.
+func (c *Credentials) UseGID(gid GID) (KGID, error) {
+	kgid := c.UserNamespace.MapToKGID(gid)
+	if !kgid.Ok() {
+		return NoID, syserror.EINVAL
+	}
+	if c.HasCapability(linux.CAP_SETGID) {
+		return kgid, nil
+	}
+	if kgid == c.RealKGID || kgid == c.EffectiveKGID || kgid == c.SavedKGID {
+		return kgid, nil
+	}
+	return NoID, syserror.EPERM
+}
diff --git a/pkg/sentry/kernel/auth/id.go b/pkg/sentry/kernel/auth/id.go
new file mode 100644
index 000000000..37522b018
--- /dev/null
+++ b/pkg/sentry/kernel/auth/id.go
@@ -0,0 +1,121 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"math"
+)
+
+// UID is a user ID in an unspecified user namespace.
+type UID uint32
+
+// GID is a group ID in an unspecified user namespace.
+type GID uint32
+
+// In the root user namespace, user/group IDs have a 1-to-1 relationship with
+// the users/groups they represent. In other user namespaces, this is not the
+// case; for example, two different unmapped users may both "have" the overflow
+// UID. This means that it is generally only valid to compare user and group
+// IDs in the root user namespace. We assign distinct types, KUID/KGID, to such
+// IDs to emphasize this distinction. ("k" is for "key", as in "unique key".
+// Linux also uses the prefix "k", but I think they mean "kernel".)
+
+// KUID is a user ID in the root user namespace.
+type KUID uint32
+
+// KGID is a group ID in the root user namespace.
+type KGID uint32
+
+const (
+	// NoID is uint32(-1). -1 is consistently used as a special value, in Linux
+	// and by extension in the auth package, to mean "no ID":
+	//
+	// - ID mapping returns -1 if the ID is not mapped.
+	//
+	// - Most set*id() syscalls accept -1 to mean "do not change this ID".
+	NoID = math.MaxUint32
+
+	// OverflowUID is the default value of /proc/sys/kernel/overflowuid. The
+	// "overflow UID" is usually [1] used when translating a user ID between
+	// namespaces fails because the ID is not mapped. (We don't implement this
+	// file, so the overflow UID is constant.)
+	//
+	// [1] "There is one notable case where unmapped user and group IDs are not
+	// converted to the corresponding overflow ID value. When viewing a uid_map
+	// or gid_map file in which there is no mapping for the second field, that
+	// field is displayed as 4294967295 (-1 as an unsigned integer);" -
+	// user_namespaces(7)
+	OverflowUID = UID(65534)
+	OverflowGID = GID(65534)
+
+	// NobodyKUID is the user ID usually reserved for the least privileged user
+	// "nobody".
+	NobodyKUID = KUID(65534)
+	NobodyKGID = KGID(65534)
+
+	// RootKUID is the user ID usually used for the most privileged user "root".
+	RootKUID = KUID(0)
+	RootKGID = KGID(0)
+	RootUID  = UID(0)
+	RootGID  = GID(0)
+)
+
+// Ok returns true if uid is not -1.
+func (uid UID) Ok() bool {
+	return uid != NoID
+}
+
+// Ok returns true if gid is not -1.
+func (gid GID) Ok() bool {
+	return gid != NoID
+}
+
+// Ok returns true if kuid is not -1.
+func (kuid KUID) Ok() bool {
+	return kuid != NoID
+}
+
+// Ok returns true if kgid is not -1.
+func (kgid KGID) Ok() bool {
+	return kgid != NoID
+}
+
+// OrOverflow returns uid if it is valid and the overflow UID otherwise.
+func (uid UID) OrOverflow() UID {
+	if uid.Ok() {
+		return uid
+	}
+	return OverflowUID
+}
+
+// OrOverflow returns gid if it is valid and the overflow GID otherwise.
+func (gid GID) OrOverflow() GID {
+	if gid.Ok() {
+		return gid
+	}
+	return OverflowGID
+}
+
+// In translates kuid into user namespace ns. If kuid is not mapped in ns, In
+// returns NoID.
+func (kuid KUID) In(ns *UserNamespace) UID {
+	return ns.MapFromKUID(kuid)
+}
+
+// In translates kgid into user namespace ns. If kgid is not mapped in ns, In
+// returns NoID.
+func (kgid KGID) In(ns *UserNamespace) GID {
+	return ns.MapFromKGID(kgid)
+}
diff --git a/pkg/sentry/kernel/auth/id_map.go b/pkg/sentry/kernel/auth/id_map.go
new file mode 100644
index 000000000..6adb33530
--- /dev/null
+++ b/pkg/sentry/kernel/auth/id_map.go
@@ -0,0 +1,283 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// MapFromKUID translates kuid, a UID in the root namespace, to a UID in ns.
+func (ns *UserNamespace) MapFromKUID(kuid KUID) UID {
+	if ns.parent == nil {
+		return UID(kuid)
+	}
+	return UID(ns.mapID(&ns.uidMapFromParent, uint32(ns.parent.MapFromKUID(kuid))))
+}
+
+// MapFromKGID translates kgid, a GID in the root namespace, to a GID in ns.
+func (ns *UserNamespace) MapFromKGID(kgid KGID) GID {
+	if ns.parent == nil {
+		return GID(kgid)
+	}
+	return GID(ns.mapID(&ns.gidMapFromParent, uint32(ns.parent.MapFromKGID(kgid))))
+}
+
+// MapToKUID translates uid, a UID in ns, to a UID in the root namespace.
+func (ns *UserNamespace) MapToKUID(uid UID) KUID {
+	if ns.parent == nil {
+		return KUID(uid)
+	}
+	return ns.parent.MapToKUID(UID(ns.mapID(&ns.uidMapToParent, uint32(uid))))
+}
+
+// MapToKGID translates gid, a GID in ns, to a GID in the root namespace.
+func (ns *UserNamespace) MapToKGID(gid GID) KGID {
+	if ns.parent == nil {
+		return KGID(gid)
+	}
+	return ns.parent.MapToKGID(GID(ns.mapID(&ns.gidMapToParent, uint32(gid))))
+}
+
+func (ns *UserNamespace) mapID(m *idMapSet, id uint32) uint32 {
+	if id == NoID {
+		return NoID
+	}
+	ns.mu.Lock()
+	defer ns.mu.Unlock()
+	if it := m.FindSegment(id); it.Ok() {
+		return it.Value() + (id - it.Start())
+	}
+	return NoID
+}
+
+// allIDsMapped returns true if all IDs in the range [start, end) are mapped in
+// m.
+//
+// Preconditions: end >= start.
+func (ns *UserNamespace) allIDsMapped(m *idMapSet, start, end uint32) bool {
+	ns.mu.Lock()
+	defer ns.mu.Unlock()
+	return m.SpanRange(idMapRange{start, end}) == end-start
+}
+
+// An IDMapEntry represents a mapping from a range of contiguous IDs in a user
+// namespace to an equally-sized range of contiguous IDs in the namespace's
+// parent.
+type IDMapEntry struct {
+	// FirstID is the first ID in the range in the namespace.
+	FirstID uint32
+
+	// FirstParentID is the first ID in the range in the parent namespace.
+	FirstParentID uint32
+
+	// Length is the number of IDs in the range.
+	Length uint32
+}
+
+// SetUIDMap instructs ns to translate UIDs as specified by entries.
+//
+// Note: SetUIDMap does not place an upper bound on the number of entries, but
+// Linux does. This restriction is implemented in SetUIDMap's caller, the
+// implementation of /proc/[pid]/uid_map.
+func (ns *UserNamespace) SetUIDMap(ctx context.Context, entries []IDMapEntry) error {
+	c := CredentialsFromContext(ctx)
+
+	ns.mu.Lock()
+	defer ns.mu.Unlock()
+	// "After the creation of a new user namespace, the uid_map file of *one*
+	// of the processes in the namespace may be written to *once* to define the
+	// mapping of user IDs in the new user namespace. An attempt to write more
+	// than once to a uid_map file in a user namespace fails with the error
+	// EPERM. Similar rules apply for gid_map files." - user_namespaces(7)
+	if !ns.uidMapFromParent.IsEmpty() {
+		return syserror.EPERM
+	}
+	// "At least one line must be written to the file."
+	if len(entries) == 0 {
+		return syserror.EINVAL
+	}
+	// """
+	// In order for a process to write to the /proc/[pid]/uid_map
+	// (/proc/[pid]/gid_map) file, all of the following requirements must be
+	// met:
+	//
+	// 1. The writing process must have the CAP_SETUID (CAP_SETGID) capability
+	// in the user namespace of the process pid.
+	// """
+	if !c.HasCapabilityIn(linux.CAP_SETUID, ns) {
+		return syserror.EPERM
+	}
+	// "2. The writing process must either be in the user namespace of the process
+	// pid or be in the parent user namespace of the process pid."
+	if c.UserNamespace != ns && c.UserNamespace != ns.parent {
+		return syserror.EPERM
+	}
+	// """
+	// 3. (see trySetUIDMap)
+	//
+	// 4. One of the following two cases applies:
+	//
+	// * Either the writing process has the CAP_SETUID (CAP_SETGID) capability
+	// in the parent user namespace.
+	// """
+	if !c.HasCapabilityIn(linux.CAP_SETUID, ns.parent) {
+		// """
+		// * Or otherwise all of the following restrictions apply:
+		//
+		//   + The data written to uid_map (gid_map) must consist of a single line
+		//   that maps the writing process' effective user ID (group ID) in the
+		//   parent user namespace to a user ID (group ID) in the user namespace.
+		// """
+		if len(entries) != 1 || ns.parent.MapToKUID(UID(entries[0].FirstParentID)) != c.EffectiveKUID || entries[0].Length != 1 {
+			return syserror.EPERM
+		}
+		// """
+		//   + The writing process must have the same effective user ID as the
+		//   process that created the user namespace.
+		// """
+		if c.EffectiveKUID != ns.owner {
+			return syserror.EPERM
+		}
+	}
+	// trySetUIDMap leaves data in maps if it fails.
+	if err := ns.trySetUIDMap(entries); err != nil {
+		ns.uidMapFromParent.RemoveAll()
+		ns.uidMapToParent.RemoveAll()
+		return err
+	}
+	return nil
+}
+
+func (ns *UserNamespace) trySetUIDMap(entries []IDMapEntry) error {
+	for _, e := range entries {
+		// Determine upper bounds and check for overflow. This implicitly
+		// checks for NoID.
+		lastID := e.FirstID + e.Length
+		if lastID <= e.FirstID {
+			return syserror.EINVAL
+		}
+		lastParentID := e.FirstParentID + e.Length
+		if lastParentID <= e.FirstParentID {
+			return syserror.EINVAL
+		}
+		// "3. The mapped user IDs (group IDs) must in turn have a mapping in
+		// the parent user namespace."
+		// Only the root namespace has a nil parent, and root is assigned
+		// mappings when it's created, so SetUIDMap would have returned EPERM
+		// without reaching this point if ns is root.
+		if !ns.parent.allIDsMapped(&ns.parent.uidMapToParent, e.FirstParentID, lastParentID) {
+			return syserror.EPERM
+		}
+		// If either of these Adds fail, we have an overlapping range.
+		if !ns.uidMapFromParent.Add(idMapRange{e.FirstParentID, lastParentID}, e.FirstID) {
+			return syserror.EINVAL
+		}
+		if !ns.uidMapToParent.Add(idMapRange{e.FirstID, lastID}, e.FirstParentID) {
+			return syserror.EINVAL
+		}
+	}
+	return nil
+}
+
+// SetGIDMap instructs ns to translate GIDs as specified by entries.
+func (ns *UserNamespace) SetGIDMap(ctx context.Context, entries []IDMapEntry) error {
+	c := CredentialsFromContext(ctx)
+
+	ns.mu.Lock()
+	defer ns.mu.Unlock()
+	if !ns.gidMapFromParent.IsEmpty() {
+		return syserror.EPERM
+	}
+	if len(entries) == 0 {
+		return syserror.EINVAL
+	}
+	if !c.HasCapabilityIn(linux.CAP_SETGID, ns) {
+		return syserror.EPERM
+	}
+	if c.UserNamespace != ns && c.UserNamespace != ns.parent {
+		return syserror.EPERM
+	}
+	if !c.HasCapabilityIn(linux.CAP_SETGID, ns.parent) {
+		if len(entries) != 1 || ns.parent.MapToKGID(GID(entries[0].FirstParentID)) != c.EffectiveKGID || entries[0].Length != 1 {
+			return syserror.EPERM
+		}
+		// It's correct for this to still be UID.
+		if c.EffectiveKUID != ns.owner {
+			return syserror.EPERM
+		}
+		// "In the case of gid_map, use of the setgroups(2) system call must
+		// first be denied by writing "deny" to the /proc/[pid]/setgroups file
+		// (see below) before writing to gid_map." (This file isn't implemented
+		// in the version of Linux we're emulating; see comment in
+		// UserNamespace.)
+	}
+	if err := ns.trySetGIDMap(entries); err != nil {
+		ns.gidMapFromParent.RemoveAll()
+		ns.gidMapToParent.RemoveAll()
+		return err
+	}
+	return nil
+}
+
+func (ns *UserNamespace) trySetGIDMap(entries []IDMapEntry) error {
+	for _, e := range entries {
+		lastID := e.FirstID + e.Length
+		if lastID <= e.FirstID {
+			return syserror.EINVAL
+		}
+		lastParentID := e.FirstParentID + e.Length
+		if lastParentID <= e.FirstParentID {
+			return syserror.EINVAL
+		}
+		if !ns.parent.allIDsMapped(&ns.parent.gidMapToParent, e.FirstParentID, lastParentID) {
+			return syserror.EPERM
+		}
+		if !ns.gidMapFromParent.Add(idMapRange{e.FirstParentID, lastParentID}, e.FirstID) {
+			return syserror.EINVAL
+		}
+		if !ns.gidMapToParent.Add(idMapRange{e.FirstID, lastID}, e.FirstParentID) {
+			return syserror.EINVAL
+		}
+	}
+	return nil
+}
+
+// UIDMap returns the user ID mappings configured for ns. If no mappings
+// have been configured, UIDMap returns nil.
+func (ns *UserNamespace) UIDMap() []IDMapEntry {
+	return ns.getIDMap(&ns.uidMapToParent)
+}
+
+// GIDMap returns the group ID mappings configured for ns. If no mappings
+// have been configured, GIDMap returns nil.
+func (ns *UserNamespace) GIDMap() []IDMapEntry {
+	return ns.getIDMap(&ns.gidMapToParent)
+}
+
+func (ns *UserNamespace) getIDMap(m *idMapSet) []IDMapEntry {
+	ns.mu.Lock()
+	defer ns.mu.Unlock()
+	var entries []IDMapEntry
+	for it := m.FirstSegment(); it.Ok(); it = it.NextSegment() {
+		entries = append(entries, IDMapEntry{
+			FirstID:       it.Start(),
+			FirstParentID: it.Value(),
+			Length:        it.Range().Length(),
+		})
+	}
+	return entries
+}
diff --git a/pkg/sentry/kernel/auth/id_map_functions.go b/pkg/sentry/kernel/auth/id_map_functions.go
new file mode 100644
index 000000000..889291d96
--- /dev/null
+++ b/pkg/sentry/kernel/auth/id_map_functions.go
@@ -0,0 +1,45 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+// idMapFunctions "implements" generic interface segment.Functions for
+// idMapSet. An idMapSet maps non-overlapping ranges of contiguous IDs in one
+// user namespace to non-overlapping ranges of contiguous IDs in another user
+// namespace. Each such ID mapping is implemented as a range-to-value mapping
+// in the set such that [range.Start(), range.End()) => [value, value +
+// range.Length()).
+type idMapFunctions struct{}
+
+func (idMapFunctions) MinKey() uint32 {
+	return 0
+}
+
+func (idMapFunctions) MaxKey() uint32 {
+	return NoID
+}
+
+func (idMapFunctions) ClearValue(*uint32) {}
+
+func (idMapFunctions) Merge(r1 idMapRange, val1 uint32, r2 idMapRange, val2 uint32) (uint32, bool) {
+	// Mapped ranges have to be contiguous.
+	if val1+r1.Length() != val2 {
+		return 0, false
+	}
+	return val1, true
+}
+
+func (idMapFunctions) Split(r idMapRange, val uint32, split uint32) (uint32, uint32) {
+	return val, val + (split - r.Start)
+}
diff --git a/pkg/sentry/kernel/auth/user_namespace.go b/pkg/sentry/kernel/auth/user_namespace.go
new file mode 100644
index 000000000..0980aeadf
--- /dev/null
+++ b/pkg/sentry/kernel/auth/user_namespace.go
@@ -0,0 +1,130 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package auth
+
+import (
+	"math"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// A UserNamespace represents a user namespace. See user_namespaces(7) for
+// details.
+type UserNamespace struct {
+	// parent is this namespace's parent. If this is the root namespace, parent
+	// is nil. The parent pointer is immutable.
+	parent *UserNamespace
+
+	// owner is the effective UID of the namespace's creator in the root
+	// namespace. owner is immutable.
+	owner KUID
+
+	// mu protects the following fields.
+	//
+	// If mu will be locked in multiple UserNamespaces, it must be locked in
+	// descendant namespaces before ancestors.
+	mu sync.Mutex `state:"nosave"`
+
+	// Mappings of user/group IDs between this namespace and its parent.
+	//
+	// All ID maps, once set, cannot be changed. This means that successful
+	// UID/GID translations cannot be racy.
+	uidMapFromParent idMapSet
+	uidMapToParent   idMapSet
+	gidMapFromParent idMapSet
+	gidMapToParent   idMapSet
+
+	// TODO: Consider supporting disabling setgroups(2), which "was
+	// added in Linux 3.19, but was backported to many earlier stable kernel
+	// series, because it addresses a security issue" - user_namespaces(7). (It
+	// was not backported to 3.11.10, which we are currently imitating.)
+}
+
+// NewRootUserNamespace returns a UserNamespace that is appropriate for a
+// system's root user namespace.
+func NewRootUserNamespace() *UserNamespace {
+	var ns UserNamespace
+	// """
+	// The initial user namespace has no parent namespace, but, for
+	// consistency, the kernel provides dummy user and group ID mapping files
+	// for this namespace. Looking at the uid_map file (gid_map is the same)
+	// from a shell in the initial namespace shows:
+	//
+	// $ cat /proc/$$/uid_map
+	// 0          0 4294967295
+	// """ - user_namespaces(7)
+	for _, m := range []*idMapSet{
+		&ns.uidMapFromParent,
+		&ns.uidMapToParent,
+		&ns.gidMapFromParent,
+		&ns.gidMapToParent,
+	} {
+		if !m.Add(idMapRange{0, math.MaxUint32}, 0) {
+			panic("Failed to insert into empty ID map")
+		}
+	}
+	return &ns
+}
+
+// Root returns the root of the user namespace tree containing ns.
+func (ns *UserNamespace) Root() *UserNamespace {
+	for ns.parent != nil {
+		ns = ns.parent
+	}
+	return ns
+}
+
+// "The kernel imposes (since version 3.11) a limit of 32 nested levels of user
+// namespaces." - user_namespaces(7)
+const maxUserNamespaceDepth = 32
+
+func (ns *UserNamespace) depth() int {
+	var i int
+	for ns != nil {
+		i++
+		ns = ns.parent
+	}
+	return i
+}
+
+// NewChildUserNamespace returns a new user namespace created by a caller with
+// credentials c.
+func (c *Credentials) NewChildUserNamespace() (*UserNamespace, error) {
+	if c.UserNamespace.depth() >= maxUserNamespaceDepth {
+		// "... Calls to unshare(2) or clone(2) that would cause this limit to
+		// be exceeded fail with the error EUSERS." - user_namespaces(7)
+		return nil, syserror.EUSERS
+	}
+	// "EPERM: CLONE_NEWUSER was specified in flags, but either the effective
+	// user ID or the effective group ID of the caller does not have a mapping
+	// in the parent namespace (see user_namespaces(7))." - clone(2)
+	// "CLONE_NEWUSER requires that the user ID and group ID of the calling
+	// process are mapped to user IDs and group IDs in the user namespace of
+	// the calling process at the time of the call." - unshare(2)
+	if !c.EffectiveKUID.In(c.UserNamespace).Ok() {
+		return nil, syserror.EPERM
+	}
+	if !c.EffectiveKGID.In(c.UserNamespace).Ok() {
+		return nil, syserror.EPERM
+	}
+	return &UserNamespace{
+		parent: c.UserNamespace,
+		owner:  c.EffectiveKUID,
+		// "When a user namespace is created, it starts without a mapping of
+		// user IDs (group IDs) to the parent user namespace." -
+		// user_namespaces(7)
+	}, nil
+}
diff --git a/pkg/sentry/kernel/context.go b/pkg/sentry/kernel/context.go
new file mode 100644
index 000000000..261ca6f7a
--- /dev/null
+++ b/pkg/sentry/kernel/context.go
@@ -0,0 +1,135 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the kernel package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxCanTrace is a Context.Value key for a function with the same
+	// signature and semantics as kernel.Task.CanTrace.
+	CtxCanTrace contextID = iota
+
+	// CtxKernel is a Context.Value key for a Kernel.
+	CtxKernel
+
+	// CtxPIDNamespace is a Context.Value key for a PIDNamespace.
+	CtxPIDNamespace
+
+	// CtxTask is a Context.Value key for a Task.
+	CtxTask
+
+	// CtxUTSNamespace is a Context.Value key for a UTSNamespace.
+	CtxUTSNamespace
+
+	// CtxIPCNamespace is a Context.Value key for a IPCNamespace.
+	CtxIPCNamespace
+)
+
+// ContextCanTrace returns true if ctx is permitted to trace t, in the same sense
+// as kernel.Task.CanTrace.
+func ContextCanTrace(ctx context.Context, t *Task, attach bool) bool {
+	if v := ctx.Value(CtxCanTrace); v != nil {
+		return v.(func(*Task, bool) bool)(t, attach)
+	}
+	return false
+}
+
+// KernelFromContext returns the Kernel in which ctx is executing, or nil if
+// there is no such Kernel.
+func KernelFromContext(ctx context.Context) *Kernel {
+	if v := ctx.Value(CtxKernel); v != nil {
+		return v.(*Kernel)
+	}
+	return nil
+}
+
+// PIDNamespaceFromContext returns the PID namespace in which ctx is executing,
+// or nil if there is no such PID namespace.
+func PIDNamespaceFromContext(ctx context.Context) *PIDNamespace {
+	if v := ctx.Value(CtxPIDNamespace); v != nil {
+		return v.(*PIDNamespace)
+	}
+	return nil
+}
+
+// UTSNamespaceFromContext returns the UTS namespace in which ctx is executing,
+// or nil if there is no such UTS namespace.
+func UTSNamespaceFromContext(ctx context.Context) *UTSNamespace {
+	if v := ctx.Value(CtxUTSNamespace); v != nil {
+		return v.(*UTSNamespace)
+	}
+	return nil
+}
+
+// IPCNamespaceFromContext returns the IPC namespace in which ctx is executing,
+// or nil if there is no such IPC namespace.
+func IPCNamespaceFromContext(ctx context.Context) *IPCNamespace {
+	if v := ctx.Value(CtxIPCNamespace); v != nil {
+		return v.(*IPCNamespace)
+	}
+	return nil
+}
+
+// TaskFromContext returns the Task associated with ctx, or nil if there is no
+// such Task.
+func TaskFromContext(ctx context.Context) *Task {
+	if v := ctx.Value(CtxTask); v != nil {
+		return v.(*Task)
+	}
+	return nil
+}
+
+// AsyncContext returns a context.Context that may be used by goroutines that
+// do work on behalf of t and therefore share its contextual values, but are
+// not t's task goroutine (e.g. asynchronous I/O).
+func (t *Task) AsyncContext() context.Context {
+	return taskAsyncContext{t: t}
+}
+
+type taskAsyncContext struct {
+	context.NoopSleeper
+	t *Task
+}
+
+// Debugf implements log.Logger.Debugf.
+func (ctx taskAsyncContext) Debugf(format string, v ...interface{}) {
+	ctx.t.Debugf(format, v...)
+}
+
+// Infof implements log.Logger.Infof.
+func (ctx taskAsyncContext) Infof(format string, v ...interface{}) {
+	ctx.t.Infof(format, v...)
+}
+
+// Warningf implements log.Logger.Warningf.
+func (ctx taskAsyncContext) Warningf(format string, v ...interface{}) {
+	ctx.t.Warningf(format, v...)
+}
+
+// IsLogging implements log.Logger.IsLogging.
+func (ctx taskAsyncContext) IsLogging(level log.Level) bool {
+	return ctx.t.IsLogging(level)
+}
+
+// Value implements context.Context.Value.
+func (ctx taskAsyncContext) Value(key interface{}) interface{} {
+	return ctx.t.Value(key)
+}
diff --git a/pkg/sentry/kernel/epoll/BUILD b/pkg/sentry/kernel/epoll/BUILD
new file mode 100644
index 000000000..04651d961
--- /dev/null
+++ b/pkg/sentry/kernel/epoll/BUILD
@@ -0,0 +1,52 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "epoll_autogen_state",
+    srcs = [
+        "epoll.go",
+        "epoll_state.go",
+    ],
+    out = "epoll_autogen_state.go",
+    package = "epoll",
+)
+
+go_library(
+    name = "epoll",
+    srcs = [
+        "epoll.go",
+        "epoll_autogen_state.go",
+        "epoll_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/epoll",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/ilist",
+        "//pkg/refs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "epoll_test",
+    size = "small",
+    srcs = [
+        "epoll_test.go",
+    ],
+    embed = [":epoll"],
+    deps = [
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs/filetest",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/kernel/epoll/epoll.go b/pkg/sentry/kernel/epoll/epoll.go
new file mode 100644
index 000000000..b572fcd7e
--- /dev/null
+++ b/pkg/sentry/kernel/epoll/epoll.go
@@ -0,0 +1,466 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package epoll provides an implementation of Linux's IO event notification
+// facility. See epoll(7) for more details.
+package epoll
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Event describes the event mask that was observed and the user data to be
+// returned when one of the events occurs. It has this format to match the linux
+// format to avoid extra copying/allocation when writing events to userspace.
+type Event struct {
+	// Events is the event mask containing the set of events that have been
+	// observed on an entry.
+	Events uint32
+
+	// Data is an opaque 64-bit value provided by the caller when adding the
+	// entry, and returned to the caller when the entry reports an event.
+	Data [2]int32
+}
+
+// EntryFlags is a bitmask that holds an entry's flags.
+type EntryFlags int
+
+// Valid entry flags.
+const (
+	OneShot EntryFlags = 1 << iota
+	EdgeTriggered
+)
+
+// FileIdentifier identifies a file. We cannot use just the FD because it could
+// potentially be reassigned. We also cannot use just the file pointer because
+// it is possible to have multiple entries for the same file object as long as
+// they are created with different FDs (i.e., the FDs point to the same file).
+type FileIdentifier struct {
+	File *fs.File
+	Fd   kdefs.FD
+}
+
+// pollEntry holds all the state associated with an event poll entry, that is,
+// a file being observed by an event poll object.
+type pollEntry struct {
+	ilist.Entry
+	file     *refs.WeakRef  `state:"manual"`
+	id       FileIdentifier `state:"wait"`
+	userData [2]int32
+	waiter   waiter.Entry `state:"manual"`
+	mask     waiter.EventMask
+	flags    EntryFlags
+
+	epoll *EventPoll
+
+	// We cannot save the current list pointer as it points into EventPoll
+	// struct, while state framework currently does not support such
+	// in-struct pointers. Instead, EventPoll will properly set this field
+	// in its loading logic.
+	curList *ilist.List `state:"nosave"`
+}
+
+// WeakRefGone implements refs.WeakRefUser.WeakRefGone.
+// weakReferenceGone is called when the file in the weak reference is destroyed.
+// The poll entry is removed in response to this.
+func (p *pollEntry) WeakRefGone() {
+	p.epoll.RemoveEntry(p.id)
+}
+
+// EventPoll holds all the state associated with an event poll object, that is,
+// collection of files to observe and their current state.
+type EventPoll struct {
+	fsutil.PipeSeek      `state:"zerovalue"`
+	fsutil.NotDirReaddir `state:"zerovalue"`
+	fsutil.NoFsync       `state:"zerovalue"`
+	fsutil.NoopFlush     `state:"zerovalue"`
+	fsutil.NoMMap        `state:"zerovalue"`
+	fsutil.NoIoctl       `state:"zerovalue"`
+
+	// Wait queue is used to notify interested parties when the event poll
+	// object itself becomes readable or writable.
+	waiter.Queue
+
+	// files is the map of all the files currently being observed, it is
+	// protected by mu.
+	mu    sync.Mutex `state:"nosave"`
+	files map[FileIdentifier]*pollEntry
+
+	// listsMu protects manipulation of the lists below. It needs to be a
+	// different lock to avoid circular lock acquisition order involving
+	// the wait queue mutexes and mu. The full order is mu, observed file
+	// wait queue mutex, then listsMu; this allows listsMu to be acquired
+	// when readyCallback is called.
+	//
+	// An entry is always in one of the following lists:
+	//	readyList -- when there's a chance that it's ready to have
+	//		events delivered to epoll waiters. Given that being
+	//		ready is a transient state, the Readiness() and
+	//		readEvents() functions always call the entry's file
+	//		Readiness() function to confirm it's ready.
+	//	waitingList -- when there's no chance that the entry is ready,
+	//		so it's waiting for the readyCallback to be called
+	//		on it before it gets moved to the readyList.
+	//	disabledList -- when the entry is disabled. This happens when
+	//		a one-shot entry gets delivered via readEvents().
+	listsMu      sync.Mutex `state:"nosave"`
+	readyList    ilist.List
+	waitingList  ilist.List
+	disabledList ilist.List
+}
+
+// cycleMu is used to serialize all the cycle checks. This is only used when
+// an event poll file is added as an entry to another event poll. Such checks
+// are serialized to avoid lock acquisition order inversion: if a thread is
+// adding A to B, and another thread is adding B to A, each would acquire A's
+// and B's mutexes in reverse order, and could cause deadlocks. Having this
+// lock prevents this by allowing only one check at a time to happen.
+//
+// We do the cycle check to prevent callers from introducing potentially
+// infinite recursions. If a caller were to add A to B and then B to A, for
+// event poll A to know if it's readable, it would need to check event poll B,
+// which in turn would need event poll A and so on indefinitely.
+var cycleMu sync.Mutex
+
+// NewEventPoll allocates and initializes a new event poll object.
+func NewEventPoll(ctx context.Context) *fs.File {
+	// name matches fs/eventpoll.c:epoll_create1.
+	dirent := fs.NewDirent(anon.NewInode(ctx), fmt.Sprintf("anon_inode:[eventpoll]"))
+	return fs.NewFile(ctx, dirent, fs.FileFlags{}, &EventPoll{
+		files: make(map[FileIdentifier]*pollEntry),
+	})
+}
+
+// Release implements fs.FileOperations.Release.
+func (e *EventPoll) Release() {
+	// We need to take the lock now because files may be attempting to
+	// remove entries in parallel if they get destroyed.
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	// Go through all entries and clean up.
+	for _, entry := range e.files {
+		entry.id.File.EventUnregister(&entry.waiter)
+		entry.file.Drop()
+	}
+}
+
+// Read implements fs.FileOperations.Read.
+func (*EventPoll) Read(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syscall.ENOSYS
+}
+
+// Write implements fs.FileOperations.Write.
+func (*EventPoll) Write(context.Context, *fs.File, usermem.IOSequence, int64) (int64, error) {
+	return 0, syscall.ENOSYS
+}
+
+// eventsAvailable determines if 'e' has events available for delivery.
+func (e *EventPoll) eventsAvailable() bool {
+	e.listsMu.Lock()
+
+	for it := e.readyList.Front(); it != nil; {
+		entry := it.(*pollEntry)
+		it = it.Next()
+
+		// If the entry is ready, we know 'e' has at least one entry
+		// ready for delivery.
+		ready := entry.id.File.Readiness(entry.mask)
+		if ready != 0 {
+			e.listsMu.Unlock()
+			return true
+		}
+
+		// Entry is not ready, so move it to waiting list.
+		e.readyList.Remove(entry)
+		e.waitingList.PushBack(entry)
+		entry.curList = &e.waitingList
+	}
+
+	e.listsMu.Unlock()
+
+	return false
+}
+
+// Readiness determines if the event poll object is currently readable (i.e.,
+// if there are pending events for delivery).
+func (e *EventPoll) Readiness(mask waiter.EventMask) waiter.EventMask {
+	ready := waiter.EventMask(0)
+
+	if (mask&waiter.EventIn) != 0 && e.eventsAvailable() {
+		ready |= waiter.EventIn
+	}
+
+	return ready
+}
+
+// ReadEvents returns up to max available events.
+func (e *EventPoll) ReadEvents(max int) []Event {
+	var local ilist.List
+	var ret []Event
+
+	e.listsMu.Lock()
+
+	// Go through all entries we believe may be ready.
+	for it := e.readyList.Front(); it != nil && len(ret) < max; {
+		entry := it.(*pollEntry)
+		it = it.Next()
+
+		// Check the entry's readiness. It it's not really ready, we
+		// just put it back in the waiting list and move on to the next
+		// entry.
+		ready := entry.id.File.Readiness(entry.mask) & entry.mask
+		if ready == 0 {
+			e.readyList.Remove(entry)
+			e.waitingList.PushBack(entry)
+			entry.curList = &e.waitingList
+
+			continue
+		}
+
+		// Add event to the array that will be returned to caller.
+		ret = append(ret, Event{
+			Events: uint32(ready),
+			Data:   entry.userData,
+		})
+
+		// The entry is consumed, so we must move it to the disabled
+		// list in case it's one-shot, or back to the wait list if it's
+		// edge-triggered. If it's neither, we leave it in the ready
+		// list so that its readiness can be checked the next time
+		// around; however, we must move it to the end of the list so
+		// that other events can be delivered as well.
+		e.readyList.Remove(entry)
+		if entry.flags&OneShot != 0 {
+			e.disabledList.PushBack(entry)
+			entry.curList = &e.disabledList
+		} else if entry.flags&EdgeTriggered != 0 {
+			e.waitingList.PushBack(entry)
+			entry.curList = &e.waitingList
+		} else {
+			local.PushBack(entry)
+		}
+	}
+
+	e.readyList.PushBackList(&local)
+
+	e.listsMu.Unlock()
+
+	return ret
+}
+
+// readyCallback is called when one of the files we're polling becomes ready. It
+// moves said file to the readyList if it's currently in the waiting list.
+type readyCallback struct{}
+
+// Callback implements waiter.EntryCallback.Callback.
+func (*readyCallback) Callback(w *waiter.Entry) {
+	entry := w.Context.(*pollEntry)
+	e := entry.epoll
+
+	e.listsMu.Lock()
+
+	if entry.curList == &e.waitingList {
+		e.waitingList.Remove(entry)
+		e.readyList.PushBack(entry)
+		entry.curList = &e.readyList
+
+		e.Notify(waiter.EventIn)
+	}
+
+	e.listsMu.Unlock()
+}
+
+// initEntryReadiness initializes the entry's state with regards to its
+// readiness by placing it in the appropriate list and registering for
+// notifications.
+func (e *EventPoll) initEntryReadiness(entry *pollEntry) {
+	// A new entry starts off in the waiting list.
+	e.listsMu.Lock()
+	e.waitingList.PushBack(entry)
+	entry.curList = &e.waitingList
+	e.listsMu.Unlock()
+
+	// Register for event notifications.
+	f := entry.id.File
+	f.EventRegister(&entry.waiter, entry.mask)
+
+	// Check if the file happens to already be in a ready state.
+	ready := f.Readiness(entry.mask) & entry.mask
+	if ready != 0 {
+		(*readyCallback).Callback(nil, &entry.waiter)
+	}
+}
+
+// observes checks if event poll object e is directly or indirectly observing
+// event poll object ep. It uses a bounded recursive depth-first search.
+func (e *EventPoll) observes(ep *EventPoll, depthLeft int) bool {
+	// If we reached the maximum depth, we'll consider that we found it
+	// because we don't want to allow chains that are too long.
+	if depthLeft <= 0 {
+		return true
+	}
+
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	// Go through each observed file and check if it is or observes ep.
+	for id := range e.files {
+		f, ok := id.File.FileOperations.(*EventPoll)
+		if !ok {
+			continue
+		}
+
+		if f == ep || f.observes(ep, depthLeft-1) {
+			return true
+		}
+	}
+
+	return false
+}
+
+// AddEntry adds a new file to the collection of files observed by e.
+func (e *EventPoll) AddEntry(id FileIdentifier, flags EntryFlags, mask waiter.EventMask, data [2]int32) error {
+	// Acquire cycle check lock if another event poll is being added.
+	ep, ok := id.File.FileOperations.(*EventPoll)
+	if ok {
+		cycleMu.Lock()
+		defer cycleMu.Unlock()
+	}
+
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	// Fail if the file already has an entry.
+	if _, ok := e.files[id]; ok {
+		return syscall.EEXIST
+	}
+
+	// Check if a cycle would be created. We use 4 as the limit because
+	// that's the value used by linux and we want to emulate it.
+	if ep != nil {
+		if e == ep {
+			return syscall.EINVAL
+		}
+
+		if ep.observes(e, 4) {
+			return syscall.ELOOP
+		}
+	}
+
+	// Create new entry and add it to map.
+	//
+	// N.B. Even though we are creating a weak reference here, we know it
+	//      won't trigger a callback because we hold a reference to the file
+	//      throughout the execution of this function.
+	entry := &pollEntry{
+		id:       id,
+		userData: data,
+		epoll:    e,
+		flags:    flags,
+		waiter:   waiter.Entry{Callback: &readyCallback{}},
+		mask:     mask,
+	}
+	entry.waiter.Context = entry
+	e.files[id] = entry
+	entry.file = refs.NewWeakRef(id.File, entry)
+
+	// Initialize the readiness state of the new entry.
+	e.initEntryReadiness(entry)
+
+	return nil
+}
+
+// UpdateEntry updates the flags, mask and user data associated with a file that
+// is already part of the collection of observed files.
+func (e *EventPoll) UpdateEntry(id FileIdentifier, flags EntryFlags, mask waiter.EventMask, data [2]int32) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	// Fail if the file doesn't have an entry.
+	entry, ok := e.files[id]
+	if !ok {
+		return syscall.ENOENT
+	}
+
+	// Unregister the old mask and remove entry from the list it's in, so
+	// readyCallback is guaranteed to not be called on this entry anymore.
+	entry.id.File.EventUnregister(&entry.waiter)
+
+	// Remove entry from whatever list it's in. This ensure that no other
+	// threads have access to this entry as the only way left to find it
+	// is via e.files, but we hold e.mu, which prevents that.
+	e.listsMu.Lock()
+	entry.curList.Remove(entry)
+	e.listsMu.Unlock()
+
+	// Initialize new readiness state.
+	entry.flags = flags
+	entry.mask = mask
+	entry.userData = data
+	e.initEntryReadiness(entry)
+
+	return nil
+}
+
+// RemoveEntry a files from the collection of observed files.
+func (e *EventPoll) RemoveEntry(id FileIdentifier) error {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	// Fail if the file doesn't have an entry.
+	entry, ok := e.files[id]
+	if !ok {
+		return syscall.ENOENT
+	}
+
+	// Unregister from file first so that no concurrent attempts will be
+	// made to manipulate the file.
+	entry.id.File.EventUnregister(&entry.waiter)
+
+	// Remove from the current list.
+	e.listsMu.Lock()
+	entry.curList.Remove(entry)
+	entry.curList = nil
+	e.listsMu.Unlock()
+
+	// Remove file from map, and drop weak reference.
+	delete(e.files, id)
+	entry.file.Drop()
+
+	return nil
+}
+
+// UnregisterEpollWaiters removes the epoll waiter objects from the waiting
+// queues. This is different from Release() as the file is not dereferenced.
+func (e *EventPoll) UnregisterEpollWaiters() {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	for _, entry := range e.files {
+		entry.id.File.EventUnregister(&entry.waiter)
+	}
+}
diff --git a/pkg/sentry/kernel/epoll/epoll_state.go b/pkg/sentry/kernel/epoll/epoll_state.go
new file mode 100644
index 000000000..dabb32f49
--- /dev/null
+++ b/pkg/sentry/kernel/epoll/epoll_state.go
@@ -0,0 +1,51 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epoll
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// afterLoad is invoked by stateify.
+func (p *pollEntry) afterLoad() {
+	p.waiter = waiter.Entry{Callback: &readyCallback{}}
+	p.waiter.Context = p
+	p.file = refs.NewWeakRef(p.id.File, p)
+	p.id.File.EventRegister(&p.waiter, p.mask)
+}
+
+// afterLoad is invoked by stateify.
+func (e *EventPoll) afterLoad() {
+	e.listsMu.Lock()
+	defer e.listsMu.Unlock()
+
+	for _, ls := range []*ilist.List{&e.waitingList, &e.readyList, &e.disabledList} {
+		for it := ls.Front(); it != nil; it = it.Next() {
+			it.(*pollEntry).curList = ls
+		}
+	}
+
+	for it := e.waitingList.Front(); it != nil; it = it.Next() {
+		p := it.(*pollEntry)
+		if p.id.File.Readiness(p.mask) != 0 {
+			e.waitingList.Remove(p)
+			e.readyList.PushBack(p)
+			p.curList = &e.readyList
+			e.Notify(waiter.EventIn)
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/epoll/epoll_test.go b/pkg/sentry/kernel/epoll/epoll_test.go
new file mode 100644
index 000000000..bc869fc13
--- /dev/null
+++ b/pkg/sentry/kernel/epoll/epoll_test.go
@@ -0,0 +1,54 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epoll
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/filetest"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+func TestFileDestroyed(t *testing.T) {
+	f := filetest.NewTestFile(t)
+	id := FileIdentifier{f, 12}
+
+	efile := NewEventPoll(contexttest.Context(t))
+	e := efile.FileOperations.(*EventPoll)
+	if err := e.AddEntry(id, 0, waiter.EventIn, [2]int32{}); err != nil {
+		t.Fatalf("addEntry failed: %v", err)
+	}
+
+	// Check that we get an event reported twice in a row.
+	evt := e.ReadEvents(1)
+	if len(evt) != 1 {
+		t.Fatalf("Unexpected number of ready events: want %v, got %v", 1, len(evt))
+	}
+
+	evt = e.ReadEvents(1)
+	if len(evt) != 1 {
+		t.Fatalf("Unexpected number of ready events: want %v, got %v", 1, len(evt))
+	}
+
+	// Destroy the file. Check that we get no more events.
+	f.DecRef()
+
+	evt = e.ReadEvents(1)
+	if len(evt) != 0 {
+		t.Fatalf("Unexpected number of ready events: want %v, got %v", 0, len(evt))
+	}
+
+}
diff --git a/pkg/sentry/kernel/eventfd/BUILD b/pkg/sentry/kernel/eventfd/BUILD
new file mode 100644
index 000000000..2d5a3c693
--- /dev/null
+++ b/pkg/sentry/kernel/eventfd/BUILD
@@ -0,0 +1,46 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "eventfd_state",
+    srcs = [
+        "eventfd.go",
+    ],
+    out = "eventfd_state.go",
+    package = "eventfd",
+)
+
+go_library(
+    name = "eventfd",
+    srcs = [
+        "eventfd.go",
+        "eventfd_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/eventfd",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/refs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "eventfd_test",
+    size = "small",
+    srcs = ["eventfd_test.go"],
+    embed = [":eventfd"],
+    deps = [
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/usermem",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/kernel/eventfd/eventfd.go b/pkg/sentry/kernel/eventfd/eventfd.go
new file mode 100644
index 000000000..c9333719e
--- /dev/null
+++ b/pkg/sentry/kernel/eventfd/eventfd.go
@@ -0,0 +1,172 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package eventfd provides an implementation of Linux's file-based event
+// notification.
+package eventfd
+
+import (
+	"math"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// EventOperations represents an event with the semantics of Linux's file-based event
+// notification (eventfd).
+type EventOperations struct {
+	fsutil.NoopRelease   `state:"nosave"`
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	fsutil.NoIoctl       `state:"nosave"`
+
+	// Mutex that protects accesses to the fields of this event.
+	mu sync.Mutex `state:"nosave"`
+
+	// Queue is used to notify interested parties when the event object
+	// becomes readable or writable.
+	waiter.Queue `state:"nosave"`
+
+	// val is the current value of the event counter.
+	val uint64
+
+	// semMode specifies whether the event is in "semaphore" mode.
+	semMode bool
+}
+
+// New creates a new event object with the supplied initial value and mode.
+func New(ctx context.Context, initVal uint64, semMode bool) *fs.File {
+	// name matches fs/eventfd.c:eventfd_file_create.
+	dirent := fs.NewDirent(anon.NewInode(ctx), "anon_inode:[eventfd]")
+	return fs.NewFile(ctx, dirent, fs.FileFlags{Read: true, Write: true}, &EventOperations{
+		val:     initVal,
+		semMode: semMode,
+	})
+}
+
+// Read implements fs.FileOperations.Read.
+func (e *EventOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() < 8 {
+		return 0, syscall.EINVAL
+	}
+	if err := e.read(ctx, dst); err != nil {
+		return 0, err
+	}
+	return 8, nil
+}
+
+// Write implements fs.FileOperations.Write.
+func (e *EventOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	if src.NumBytes() < 8 {
+		return 0, syscall.EINVAL
+	}
+	if err := e.write(ctx, src); err != nil {
+		return 0, err
+	}
+	return 8, nil
+}
+
+func (e *EventOperations) read(ctx context.Context, dst usermem.IOSequence) error {
+	e.mu.Lock()
+
+	// We can't complete the read if the value is currently zero.
+	if e.val == 0 {
+		e.mu.Unlock()
+		return syserror.ErrWouldBlock
+	}
+
+	// Update the value based on the mode the event is operating in.
+	var val uint64
+	if e.semMode {
+		val = 1
+		// Consistent with Linux, this is done even if writing to memory fails.
+		e.val--
+	} else {
+		val = e.val
+		e.val = 0
+	}
+
+	e.mu.Unlock()
+
+	// Notify writers. We do this even if we were already writable because
+	// it is possible that a writer is waiting to write the maximum value
+	// to the event.
+	e.Notify(waiter.EventOut)
+
+	var buf [8]byte
+	usermem.ByteOrder.PutUint64(buf[:], val)
+	_, err := dst.CopyOut(ctx, buf[:])
+	return err
+}
+
+func (e *EventOperations) write(ctx context.Context, src usermem.IOSequence) error {
+	var buf [8]byte
+	if _, err := src.CopyIn(ctx, buf[:]); err != nil {
+		return err
+	}
+	val := usermem.ByteOrder.Uint64(buf[:])
+
+	return e.Signal(val)
+}
+
+// Signal is an internal function to signal the event fd.
+func (e *EventOperations) Signal(val uint64) error {
+	if val == math.MaxUint64 {
+		return syscall.EINVAL
+	}
+
+	e.mu.Lock()
+
+	// We only allow writes that won't cause the value to go over the max
+	// uint64 minus 1.
+	if val > math.MaxUint64-1-e.val {
+		e.mu.Unlock()
+		return syserror.ErrWouldBlock
+	}
+
+	e.val += val
+	e.mu.Unlock()
+
+	// Always trigger a notification.
+	e.Notify(waiter.EventIn)
+
+	return nil
+}
+
+// Readiness returns the ready events for the event fd.
+func (e *EventOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	ready := waiter.EventMask(0)
+
+	e.mu.Lock()
+	if e.val > 0 {
+		ready |= waiter.EventIn
+	}
+
+	if e.val < math.MaxUint64-1 {
+		ready |= waiter.EventOut
+	}
+	e.mu.Unlock()
+
+	return mask & ready
+}
diff --git a/pkg/sentry/kernel/eventfd/eventfd_test.go b/pkg/sentry/kernel/eventfd/eventfd_test.go
new file mode 100644
index 000000000..71326b62f
--- /dev/null
+++ b/pkg/sentry/kernel/eventfd/eventfd_test.go
@@ -0,0 +1,78 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package eventfd
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+func TestEventfd(t *testing.T) {
+	initVals := []uint64{
+		0,
+		// Using a non-zero initial value verifies that writing to an
+		// eventfd signals when the eventfd's counter was already
+		// non-zero.
+		343,
+	}
+
+	for _, initVal := range initVals {
+		ctx := contexttest.Context(t)
+
+		// Make a new event that is writable.
+		event := New(ctx, initVal, false)
+
+		// Register a callback for a write event.
+		w, ch := waiter.NewChannelEntry(nil)
+		event.EventRegister(&w, waiter.EventIn)
+		defer event.EventUnregister(&w)
+
+		data := []byte("00000124")
+		// Create and submit a write request.
+		n, err := event.Writev(ctx, usermem.BytesIOSequence(data))
+		if err != nil {
+			t.Fatal(err)
+		}
+		if n != 8 {
+			t.Errorf("eventfd.write wrote %d bytes, not full int64", n)
+		}
+
+		// Check if the callback fired due to the write event.
+		select {
+		case <-ch:
+		default:
+			t.Errorf("Didn't get notified of EventIn after write")
+		}
+	}
+}
+
+func TestEventfdStat(t *testing.T) {
+	ctx := contexttest.Context(t)
+
+	// Make a new event that is writable.
+	event := New(ctx, 0, false)
+
+	// Create and submit an stat request.
+	uattr, err := event.Dirent.Inode.UnstableAttr(ctx)
+	if err != nil {
+		t.Fatalf("eventfd stat request failed: %v", err)
+	}
+	if uattr.Size != 0 {
+		t.Fatal("EventFD size should be 0")
+	}
+}
diff --git a/pkg/sentry/kernel/fd_map.go b/pkg/sentry/kernel/fd_map.go
new file mode 100644
index 000000000..ef73125fd
--- /dev/null
+++ b/pkg/sentry/kernel/fd_map.go
@@ -0,0 +1,340 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"bytes"
+	"fmt"
+	"sort"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/lock"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+)
+
+// FDs is an ordering of FD's that can be made stable.
+type FDs []kdefs.FD
+
+func (f FDs) Len() int {
+	return len(f)
+}
+
+func (f FDs) Swap(i, j int) {
+	f[i], f[j] = f[j], f[i]
+}
+
+func (f FDs) Less(i, j int) bool {
+	return f[i] < f[j]
+}
+
+// FDFlags define flags for an individual descriptor.
+type FDFlags struct {
+	// CloseOnExec indicates the descriptor should be closed on exec.
+	CloseOnExec bool
+}
+
+// descriptor holds the details about a file descriptor, namely a pointer the
+// file itself and the descriptor flags.
+type descriptor struct {
+	file  *fs.File
+	flags FDFlags
+}
+
+// FDMap is used to manage File references and flags.
+type FDMap struct {
+	refs.AtomicRefCount
+	k     *Kernel
+	files map[kdefs.FD]descriptor
+	mu    sync.RWMutex `state:"nosave"`
+	uid   uint64
+}
+
+// ID returns a unique identifier for this FDMap.
+func (f *FDMap) ID() uint64 {
+	return f.uid
+}
+
+// NewFDMap allocates a new FDMap that may be used by tasks in k.
+func (k *Kernel) NewFDMap() *FDMap {
+	return &FDMap{
+		k:     k,
+		files: make(map[kdefs.FD]descriptor),
+		uid:   atomic.AddUint64(&k.fdMapUids, 1),
+	}
+}
+
+// destroy removes all of the file descriptors from the map.
+func (f *FDMap) destroy() {
+	f.RemoveIf(func(*fs.File, FDFlags) bool {
+		return true
+	})
+}
+
+// DecRef implements RefCounter.DecRef with destructor f.destroy.
+func (f *FDMap) DecRef() {
+	f.DecRefWithDestructor(f.destroy)
+}
+
+// Size returns the number of file descriptor slots currently allocated.
+func (f *FDMap) Size() int {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+
+	return len(f.files)
+}
+
+// String is a stringer for FDMap.
+func (f *FDMap) String() string {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+
+	var b bytes.Buffer
+	for k, v := range f.files {
+		n, _ := v.file.Dirent.FullName(nil /* root */)
+		b.WriteString(fmt.Sprintf("\tfd:%d => name %s\n", k, n))
+	}
+	return b.String()
+}
+
+// NewFDFrom allocates a new FD guaranteed to be the lowest number available
+// greater than or equal to from. This property is important as Unix programs
+// tend to count on this allocation order.
+func (f *FDMap) NewFDFrom(fd kdefs.FD, file *fs.File, flags FDFlags, limitSet *limits.LimitSet) (kdefs.FD, error) {
+	if fd < 0 {
+		// Don't accept negative FDs.
+		return 0, syscall.EINVAL
+	}
+
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	// Finds the lowest fd not in the handles map.
+	lim := limitSet.Get(limits.NumberOfFiles)
+	for i := fd; lim.Cur == limits.Infinity || i < kdefs.FD(lim.Cur); i++ {
+		if _, ok := f.files[i]; !ok {
+			file.IncRef()
+			f.files[i] = descriptor{file, flags}
+			return i, nil
+		}
+	}
+
+	return -1, syscall.EMFILE
+}
+
+// NewFDAt sets the file reference for the given FD. If there is an
+// active reference for that FD, the ref count for that existing reference
+// is decremented.
+func (f *FDMap) NewFDAt(fd kdefs.FD, file *fs.File, flags FDFlags, limitSet *limits.LimitSet) error {
+	if fd < 0 {
+		// Don't accept negative FDs.
+		return syscall.EBADF
+	}
+
+	// In this one case we do not do a defer of the Unlock.  The
+	// reason is that we must have done all the work needed for
+	// discarding any old open file before we return to the
+	// caller. In other words, the DecRef(), below, must have
+	// completed by the time we return to the caller to ensure
+	// side effects are, in fact, effected. A classic example is
+	// dup2(fd1, fd2); if fd2 was already open, it must be closed,
+	// and we don't want to resume the caller until it is; we have
+	// to block on the DecRef(). Hence we can not just do a 'go
+	// oldfile.DecRef()', since there would be no guarantee that
+	// it would be done before we the caller resumed. Since we
+	// must wait for the DecRef() to finish, and that could take
+	// time, it's best to first call f.muUnlock beore so we are
+	// not blocking other uses of this FDMap on the DecRef() call.
+	f.mu.Lock()
+	oldDesc, oldExists := f.files[fd]
+	lim := limitSet.Get(limits.NumberOfFiles).Cur
+	// if we're closing one then the effective limit is one
+	// more than the actual limit.
+	if oldExists && lim != limits.Infinity {
+		lim++
+	}
+	if lim != limits.Infinity && fd >= kdefs.FD(lim) {
+		f.mu.Unlock()
+		return syscall.EMFILE
+	}
+
+	file.IncRef()
+	f.files[fd] = descriptor{file, flags}
+	f.mu.Unlock()
+
+	if oldExists {
+		oldDesc.file.DecRef()
+	}
+	return nil
+}
+
+// SetFlags sets the flags for the given file descriptor, if it is valid.
+func (f *FDMap) SetFlags(fd kdefs.FD, flags FDFlags) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	desc, ok := f.files[fd]
+	if !ok {
+		return
+	}
+
+	f.files[fd] = descriptor{desc.file, flags}
+}
+
+// GetDescriptor returns a reference to the file and the flags for the FD. It
+// bumps its reference count as well. It returns nil if there is no File
+// for the FD, i.e. if the FD is invalid. The caller must use DecRef
+// when they are done.
+func (f *FDMap) GetDescriptor(fd kdefs.FD) (*fs.File, FDFlags) {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+
+	if desc, ok := f.files[fd]; ok {
+		desc.file.IncRef()
+		return desc.file, desc.flags
+	}
+	return nil, FDFlags{}
+}
+
+// GetFile returns a reference to the File for the FD and bumps
+// its reference count as well. It returns nil if there is no File
+// for the FD, i.e. if the FD is invalid. The caller must use DecRef
+// when they are done.
+func (f *FDMap) GetFile(fd kdefs.FD) *fs.File {
+	f.mu.RLock()
+	if desc, ok := f.files[fd]; ok {
+		desc.file.IncRef()
+		f.mu.RUnlock()
+		return desc.file
+	}
+	f.mu.RUnlock()
+	return nil
+}
+
+// fds returns an ordering of FDs.
+func (f *FDMap) fds() FDs {
+	fds := make(FDs, 0, len(f.files))
+	for fd := range f.files {
+		fds = append(fds, fd)
+	}
+	sort.Sort(fds)
+	return fds
+}
+
+// GetFDs returns a list of valid fds.
+func (f *FDMap) GetFDs() FDs {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+	return f.fds()
+}
+
+// GetRefs returns a stable slice of references to all files and bumps the
+// reference count on each.  The caller must use DecRef on each reference when
+// they're done using the slice.
+func (f *FDMap) GetRefs() []*fs.File {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+
+	fds := f.fds()
+	fs := make([]*fs.File, 0, len(fds))
+	for _, fd := range fds {
+		desc := f.files[fd]
+		desc.file.IncRef()
+		fs = append(fs, desc.file)
+	}
+	return fs
+}
+
+// Fork returns an independent FDMap pointing to the same descriptors.
+func (f *FDMap) Fork() *FDMap {
+	f.mu.RLock()
+	defer f.mu.RUnlock()
+
+	clone := f.k.NewFDMap()
+
+	// Grab a extra reference for every file.
+	for fd, desc := range f.files {
+		desc.file.IncRef()
+		clone.files[fd] = desc
+	}
+
+	// That's it!
+	return clone
+}
+
+// unlock releases all file locks held by this FDMap's uid.  Must only be
+// called on a non-nil *fs.File.
+func (f *FDMap) unlock(file *fs.File) {
+	id := lock.UniqueID(f.ID())
+	file.Dirent.Inode.LockCtx.Posix.UnlockRegion(id, lock.LockRange{0, lock.LockEOF})
+}
+
+// inotifyFileClose generates the appropriate inotify events for f being closed.
+func inotifyFileClose(f *fs.File) {
+	var ev uint32
+	d := f.Dirent
+
+	if fs.IsDir(d.Inode.StableAttr) {
+		ev |= linux.IN_ISDIR
+	}
+
+	if f.Flags().Write {
+		ev |= linux.IN_CLOSE_WRITE
+	} else {
+		ev |= linux.IN_CLOSE_NOWRITE
+	}
+
+	d.InotifyEvent(ev, 0)
+}
+
+// Remove removes an FD from the FDMap, and returns (File, true) if a File
+// one was found. Callers are expected to decrement the reference count on
+// the File. Otherwise returns (nil, false).
+func (f *FDMap) Remove(fd kdefs.FD) (*fs.File, bool) {
+	f.mu.Lock()
+	desc := f.files[fd]
+	delete(f.files, fd)
+	f.mu.Unlock()
+	if desc.file != nil {
+		f.unlock(desc.file)
+		inotifyFileClose(desc.file)
+		return desc.file, true
+	}
+	return nil, false
+}
+
+// RemoveIf removes all FDs where cond is true.
+func (f *FDMap) RemoveIf(cond func(*fs.File, FDFlags) bool) {
+	var removed []*fs.File
+	f.mu.Lock()
+	for fd, desc := range f.files {
+		if desc.file != nil && cond(desc.file, desc.flags) {
+			delete(f.files, fd)
+			removed = append(removed, desc.file)
+		}
+	}
+	f.mu.Unlock()
+
+	for _, file := range removed {
+		f.unlock(file)
+		inotifyFileClose(file)
+		file.DecRef()
+	}
+}
diff --git a/pkg/sentry/kernel/fd_map_test.go b/pkg/sentry/kernel/fd_map_test.go
new file mode 100644
index 000000000..e1ac900e8
--- /dev/null
+++ b/pkg/sentry/kernel/fd_map_test.go
@@ -0,0 +1,134 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/filetest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+)
+
+const (
+	// maxFD is the maximum FD to try to create in the map.
+	// This number of open files has been seen in the wild.
+	maxFD = 2 * 1024
+)
+
+func newTestFDMap() *FDMap {
+	return &FDMap{
+		files: make(map[kdefs.FD]descriptor),
+	}
+}
+
+// TestFDMapMany allocates maxFD FDs, i.e. maxes out the FDMap,
+// until there is no room, then makes sure that NewFDAt works
+// and also that if we remove one and add one that works too.
+func TestFDMapMany(t *testing.T) {
+	file := filetest.NewTestFile(t)
+	limitSet := limits.NewLimitSet()
+	limitSet.Set(limits.NumberOfFiles, limits.Limit{maxFD, maxFD})
+
+	f := newTestFDMap()
+	for i := 0; i < maxFD; i++ {
+		if _, err := f.NewFDFrom(0, file, FDFlags{}, limitSet); err != nil {
+			t.Fatalf("Allocated %v FDs but wanted to allocate %v", i, maxFD)
+		}
+	}
+
+	if _, err := f.NewFDFrom(0, file, FDFlags{}, limitSet); err == nil {
+		t.Fatalf("f.NewFDFrom(0, r) in full map: got nil, wanted error")
+	}
+
+	if err := f.NewFDAt(1, file, FDFlags{}, limitSet); err != nil {
+		t.Fatalf("f.NewFDAt(1, r, FDFlags{}): got %v, wanted nil", err)
+	}
+}
+
+// TestFDMap does a set of simple tests to make sure simple adds,
+// removes, GetRefs, and DecRefs work. The ordering is just weird
+// enough that a table-driven approach seemed clumsy.
+func TestFDMap(t *testing.T) {
+	file := filetest.NewTestFile(t)
+	limitSet := limits.NewLimitSet()
+	limitSet.Set(limits.NumberOfFiles, limits.Limit{1, maxFD})
+
+	f := newTestFDMap()
+	if _, err := f.NewFDFrom(0, file, FDFlags{}, limitSet); err != nil {
+		t.Fatalf("Adding an FD to an empty 1-size map: got %v, want nil", err)
+	}
+
+	if _, err := f.NewFDFrom(0, file, FDFlags{}, limitSet); err == nil {
+		t.Fatalf("Adding an FD to a filled 1-size map: got nil, wanted an error")
+	}
+
+	largeLimit := limits.Limit{maxFD, maxFD}
+	limitSet.Set(limits.NumberOfFiles, largeLimit)
+
+	if fd, err := f.NewFDFrom(0, file, FDFlags{}, limitSet); err != nil {
+		t.Fatalf("Adding an FD to a resized map: got %v, want nil", err)
+	} else if fd != kdefs.FD(1) {
+		t.Fatalf("Added an FD to a resized map: got %v, want 1", fd)
+	}
+
+	if err := f.NewFDAt(1, file, FDFlags{}, limitSet); err != nil {
+		t.Fatalf("Replacing FD 1 via f.NewFDAt(1, r, FDFlags{}): got %v, wanted nil", err)
+	}
+
+	if err := f.NewFDAt(maxFD+1, file, FDFlags{}, limitSet); err == nil {
+		t.Fatalf("Using an FD that was too large via f.NewFDAt(%v, r, FDFlags{}): got nil, wanted an error", maxFD+1)
+	}
+
+	if ref := f.GetFile(1); ref == nil {
+		t.Fatalf("f.GetFile(1): got nil, wanted %v", file)
+	}
+
+	if ref := f.GetFile(2); ref != nil {
+		t.Fatalf("f.GetFile(2): got a %v, wanted nil", ref)
+	}
+
+	ref, ok := f.Remove(1)
+	if !ok {
+		t.Fatalf("f.Remove(1) for an existing FD: failed, want success")
+	}
+	ref.DecRef()
+
+	if ref, ok := f.Remove(1); ok {
+		ref.DecRef()
+		t.Fatalf("r.Remove(1) for a removed FD: got success, want failure")
+	}
+
+}
+
+func TestDescriptorFlags(t *testing.T) {
+	file := filetest.NewTestFile(t)
+	f := newTestFDMap()
+	limitSet := limits.NewLimitSet()
+	limitSet.Set(limits.NumberOfFiles, limits.Limit{maxFD, maxFD})
+
+	if err := f.NewFDAt(2, file, FDFlags{CloseOnExec: true}, limitSet); err != nil {
+		t.Fatalf("f.NewFDAt(2, r, FDFlags{}): got %v, wanted nil", err)
+	}
+
+	newFile, flags := f.GetDescriptor(2)
+	if newFile == nil {
+		t.Fatalf("f.GetFile(2): got a %v, wanted nil", newFile)
+	}
+
+	if !flags.CloseOnExec {
+		t.Fatalf("new File flags %d don't match original %d\n", flags, 0)
+	}
+}
diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go
new file mode 100644
index 000000000..9aa6fa951
--- /dev/null
+++ b/pkg/sentry/kernel/fs_context.go
@@ -0,0 +1,172 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// FSContext contains filesystem context.
+//
+// This includes umask and working directory.
+type FSContext struct {
+	refs.AtomicRefCount
+
+	// mu protects below.
+	mu sync.Mutex `state:"nosave"`
+
+	// root is the filesystem root. Will be nil iff the FSContext has been
+	// destroyed.
+	root *fs.Dirent
+
+	// cwd is the current working directory. Will be nil iff the FSContext
+	// has been destroyed.
+	cwd *fs.Dirent
+
+	// umask is the current file mode creation mask. When a thread using this
+	// context invokes a syscall that creates a file, bits set in umask are
+	// removed from the permissions that the file is created with.
+	umask uint
+}
+
+// newFSContext returns a new filesystem context.
+func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext {
+	root.IncRef()
+	cwd.IncRef()
+	return &FSContext{
+		root:  root,
+		cwd:   cwd,
+		umask: umask,
+	}
+}
+
+// destroy is the destructor for an FSContext.
+//
+// This will call DecRef on both root and cwd Dirents.  If either call to
+// DecRef returns an error, then it will be propigated.  If both calls to
+// DecRef return an error, then the one from root.DecRef will be propigated.
+//
+// Note that there may still be calls to WorkingDirectory() or RootDirectory()
+// (that return nil).  This is because valid references may still be held via
+// proc files or other mechanisms.
+func (f *FSContext) destroy() {
+	f.root.DecRef()
+	f.root = nil
+
+	f.cwd.DecRef()
+	f.cwd = nil
+}
+
+// DecRef implements RefCounter.DecRef with destructor f.destroy.
+func (f *FSContext) DecRef() {
+	f.DecRefWithDestructor(f.destroy)
+}
+
+// Fork forks this FSContext.
+//
+// This is not a valid call after destroy.
+func (f *FSContext) Fork() *FSContext {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.cwd.IncRef()
+	f.root.IncRef()
+	return &FSContext{
+		cwd:   f.cwd,
+		root:  f.root,
+		umask: f.umask,
+	}
+}
+
+// WorkingDirectory returns the current working directory.
+// You should call DecRef on the returned Dirent when finished.
+//
+// This will return nil if called after destroy().
+func (f *FSContext) WorkingDirectory() *fs.Dirent {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if f.cwd != nil {
+		f.cwd.IncRef()
+	}
+	return f.cwd
+}
+
+// SetWorkingDirectory sets the current working directory.
+// This will take an extra reference on the Dirent.
+//
+// This is not a valid call after destroy.
+func (f *FSContext) SetWorkingDirectory(d *fs.Dirent) {
+	if d == nil {
+		panic("FSContext.SetWorkingDirectory called with nil dirent")
+	}
+	if f.cwd == nil {
+		panic(fmt.Sprintf("FSContext.SetWorkingDirectory(%v)) called after destroy", d))
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	old := f.cwd
+	f.cwd = d
+	d.IncRef()
+	old.DecRef()
+}
+
+// RootDirectory returns the current filesystem root.
+// You should call DecRef on the returned Dirent when finished.
+//
+// This will return nil if called after destroy().
+func (f *FSContext) RootDirectory() *fs.Dirent {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.root.IncRef()
+	return f.root
+}
+
+// SetRootDirectory sets the root directory.
+// This will take an extra reference on the Dirent.
+//
+// This is not a valid call after free.
+func (f *FSContext) SetRootDirectory(d *fs.Dirent) {
+	if d == nil {
+		panic("FSContext.SetRootDirectory called with nil dirent")
+	}
+	if f.root == nil {
+		panic(fmt.Sprintf("FSContext.SetRootDirectory(%v)) called after destroy", d))
+	}
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	old := f.root
+	f.root = d
+	d.IncRef()
+	old.DecRef()
+}
+
+// Umask returns the current umask.
+func (f *FSContext) Umask() uint {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.umask
+}
+
+// SwapUmask atomically sets the current umask and returns the old umask.
+func (f *FSContext) SwapUmask(mask uint) uint {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	old := f.umask
+	f.umask = mask
+	return old
+}
diff --git a/pkg/sentry/kernel/futex/BUILD b/pkg/sentry/kernel/futex/BUILD
new file mode 100644
index 000000000..de9897c58
--- /dev/null
+++ b/pkg/sentry/kernel/futex/BUILD
@@ -0,0 +1,48 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_template_instance(
+    name = "waiter_list",
+    out = "waiter_list.go",
+    package = "futex",
+    prefix = "waiter",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*Waiter",
+    },
+)
+
+go_stateify(
+    name = "futex_state",
+    srcs = [
+        "futex.go",
+        "waiter_list.go",
+    ],
+    out = "futex_state.go",
+    package = "futex",
+)
+
+go_library(
+    name = "futex",
+    srcs = [
+        "futex.go",
+        "futex_state.go",
+        "waiter_list.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/futex",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "futex_test",
+    size = "small",
+    srcs = ["futex_test.go"],
+    embed = [":futex"],
+)
diff --git a/pkg/sentry/kernel/futex/futex.go b/pkg/sentry/kernel/futex/futex.go
new file mode 100644
index 000000000..b3ba57a2c
--- /dev/null
+++ b/pkg/sentry/kernel/futex/futex.go
@@ -0,0 +1,405 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package futex provides an implementation of the futex interface as found in
+// the Linux kernel. It allows one to easily transform Wait() calls into waits
+// on a channel, which is useful in a Go-based kernel, for example.
+package futex
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Checker abstracts memory accesses. This is useful because the "addresses"
+// used in this package may not be real addresses (they could be indices of an
+// array, for example), or they could be mapped via some special mechanism.
+//
+// TODO: Replace this with usermem.IO.
+type Checker interface {
+	// Check should validate that given address contains the given value.
+	// If it does not contain the value, syserror.EAGAIN must be returned.
+	// Any other error may be returned, which will be propagated.
+	Check(addr uintptr, val uint32) error
+
+	// Op should atomically perform the operation encoded in op on the data
+	// pointed to by addr, then apply the comparison encoded in op to the
+	// original value at addr, returning the result.
+	// Note that op is an opaque operation whose behaviour is defined
+	// outside of the futex manager.
+	Op(addr uintptr, op uint32) (bool, error)
+}
+
+// Waiter is the struct which gets enqueued into buckets for wake up routines
+// and requeue routines to scan and notify. Once a Waiter has been enqueued by
+// WaitPrepare(), callers may listen on C for wake up events.
+type Waiter struct {
+	// Synchronization:
+	//
+	// - A Waiter that is not enqueued in a bucket is exclusively owned (no
+	// synchronization applies).
+	//
+	// - A Waiter is enqueued in a bucket by calling WaitPrepare(). After this,
+	// waiterEntry, complete, and addr are protected by the bucket.mu ("bucket
+	// lock") of the containing bucket, and bitmask is immutable. complete and
+	// addr are additionally mutated using atomic memory operations, ensuring
+	// that they can be read using atomic memory operations without holding the
+	// bucket lock.
+	//
+	// - A Waiter is only guaranteed to be no longer queued after calling
+	// WaitComplete().
+
+	// waiterEntry links Waiter into bucket.waiters.
+	waiterEntry
+
+	// complete is 1 if the Waiter was removed from its bucket by a wakeup and
+	// 0 otherwise.
+	complete int32
+
+	// C is sent to when the Waiter is woken.
+	C chan struct{}
+
+	// addr is the address being waited on.
+	addr uintptr
+
+	// The bitmask we're waiting on.
+	// This is used the case of a FUTEX_WAKE_BITSET.
+	bitmask uint32
+}
+
+// NewWaiter returns a new unqueued Waiter.
+func NewWaiter() *Waiter {
+	return &Waiter{
+		C: make(chan struct{}, 1),
+	}
+}
+
+// bucket holds a list of waiters for a given address hash.
+type bucket struct {
+	// mu protects waiters and contained Waiter state. See comment in Waiter.
+	mu sync.Mutex `state:"nosave"`
+
+	waiters waiterList `state:"zerovalue"`
+}
+
+// wakeLocked wakes up to n waiters matching the bitmask at the addr for this
+// bucket and returns the number of waiters woken.
+//
+// Preconditions: b.mu must be locked.
+func (b *bucket) wakeLocked(addr uintptr, bitmask uint32, n int) int {
+	done := 0
+	for w := b.waiters.Front(); done < n && w != nil; {
+		if w.addr != addr || w.bitmask&bitmask == 0 {
+			// Not matching.
+			w = w.Next()
+			continue
+		}
+
+		// Remove from the bucket and wake the waiter.
+		woke := w
+		w = w.Next() // Next iteration.
+		b.waiters.Remove(woke)
+		woke.C <- struct{}{}
+
+		// NOTE: The above channel write establishes a write barrier
+		// according to the memory model, so nothing may be ordered
+		// around it. Since we've dequeued w and will never touch it
+		// again, we can safely store 1 to w.complete here and allow
+		// the WaitComplete() to short-circuit grabbing the bucket
+		// lock. If they somehow miss the w.complete, we are still
+		// holding the lock, so we can know that they won't dequeue w,
+		// assume it's free and have the below operation afterwards.
+		atomic.StoreInt32(&woke.complete, 1)
+		done++
+	}
+	return done
+}
+
+// requeueLocked takes n waiters from the bucket and moves them to naddr on the
+// bucket "to".
+//
+// Preconditions: b and to must be locked.
+func (b *bucket) requeueLocked(to *bucket, addr, naddr uintptr, n int) int {
+	done := 0
+	for w := b.waiters.Front(); done < n && w != nil; {
+		if w.addr != addr {
+			// Not matching.
+			w = w.Next()
+			continue
+		}
+
+		requeued := w
+		w = w.Next() // Next iteration.
+		b.waiters.Remove(requeued)
+		atomic.StoreUintptr(&requeued.addr, naddr)
+		to.waiters.PushBack(requeued)
+		done++
+	}
+	return done
+}
+
+const (
+	// bucketCount is the number of buckets per Manager. By having many of
+	// these we reduce contention when concurrent yet unrelated calls are made.
+	bucketCount     = 1 << bucketCountBits
+	bucketCountBits = 10
+)
+
+func checkAddr(addr uintptr) error {
+	// Ensure the address is aligned.
+	// It must be a DWORD boundary.
+	if addr&0x3 != 0 {
+		return syserror.EINVAL
+	}
+
+	return nil
+}
+
+// bucketIndexForAddr returns the index into Manager.buckets for addr.
+func bucketIndexForAddr(addr uintptr) uintptr {
+	// - The bottom 2 bits of addr must be 0, per checkAddr.
+	//
+	// - On amd64, the top 16 bits of addr (bits 48-63) must be equal to bit 47
+	// for a canonical address, and (on all existing platforms) bit 47 must be
+	// 0 for an application address.
+	//
+	// Thus 19 bits of addr are "useless" for hashing, leaving only 45 "useful"
+	// bits. We choose one of the simplest possible hash functions that at
+	// least uses all 45 useful bits in the output, given that bucketCountBits
+	// == 10. This hash function also has the property that it will usually map
+	// adjacent addresses to adjacent buckets, slightly improving memory
+	// locality when an application synchronization structure uses multiple
+	// nearby futexes.
+	//
+	// Note that despite the large number of arithmetic operations in the
+	// function, many components can be computed in parallel, such that the
+	// critical path is 1 bit shift + 3 additions (2 in h1, then h1 + h2). This
+	// is also why h1 and h2 are grouped separately; for "(addr >> 2) + ... +
+	// (addr >> 42)" without any additional grouping, the compiler puts all 4
+	// additions in the critical path.
+	h1 := (addr >> 2) + (addr >> 12) + (addr >> 22)
+	h2 := (addr >> 32) + (addr >> 42)
+	return (h1 + h2) % bucketCount
+}
+
+// Manager holds futex state for a single virtual address space.
+type Manager struct {
+	buckets [bucketCount]bucket
+}
+
+// NewManager returns an initialized futex manager.
+// N.B. we use virtual address to tag futexes, so it only works for private
+// (within a single process) futex.
+func NewManager() *Manager {
+	return &Manager{}
+}
+
+// lockBucket returns a locked bucket for the given addr.
+//
+// Preconditions: checkAddr(addr) == nil.
+func (m *Manager) lockBucket(addr uintptr) *bucket {
+	b := &m.buckets[bucketIndexForAddr(addr)]
+	b.mu.Lock()
+	return b
+}
+
+// lockBuckets returns locked buckets for the given addrs.
+//
+// Preconditions: checkAddr(addr1) == checkAddr(addr2) == nil.
+func (m *Manager) lockBuckets(addr1 uintptr, addr2 uintptr) (*bucket, *bucket) {
+	i1 := bucketIndexForAddr(addr1)
+	i2 := bucketIndexForAddr(addr2)
+	b1 := &m.buckets[i1]
+	b2 := &m.buckets[i2]
+
+	// Ensure that buckets are locked in a consistent order (lowest index
+	// first) to avoid circular locking.
+	switch {
+	case i1 < i2:
+		b1.mu.Lock()
+		b2.mu.Lock()
+	case i2 < i1:
+		b2.mu.Lock()
+		b1.mu.Lock()
+	default:
+		b1.mu.Lock()
+	}
+
+	return b1, b2
+}
+
+// Wake wakes up to n waiters matching the bitmask on the given addr.
+// The number of waiters woken is returned.
+func (m *Manager) Wake(addr uintptr, bitmask uint32, n int) (int, error) {
+	if err := checkAddr(addr); err != nil {
+		return 0, err
+	}
+
+	b := m.lockBucket(addr)
+	// This function is very hot; avoid defer.
+	r := b.wakeLocked(addr, bitmask, n)
+	b.mu.Unlock()
+	return r, nil
+}
+
+func (m *Manager) doRequeue(c Checker, addr uintptr, val uint32, naddr uintptr, nwake int, nreq int) (int, error) {
+	if err := checkAddr(addr); err != nil {
+		return 0, err
+	}
+	if err := checkAddr(naddr); err != nil {
+		return 0, err
+	}
+
+	b1, b2 := m.lockBuckets(addr, naddr)
+	defer b1.mu.Unlock()
+	if b2 != b1 {
+		defer b2.mu.Unlock()
+	}
+
+	// Check our value.
+	// This only applied for RequeueCmp().
+	if c != nil {
+		if err := c.Check(addr, val); err != nil {
+			return 0, err
+		}
+	}
+
+	// Wake the number required.
+	done := b1.wakeLocked(addr, ^uint32(0), nwake)
+
+	// Requeue the number required.
+	b1.requeueLocked(b2, addr, naddr, nreq)
+
+	return done, nil
+}
+
+// Requeue wakes up to nwake waiters on the given addr, and unconditionally
+// requeues up to nreq waiters on naddr.
+func (m *Manager) Requeue(addr uintptr, naddr uintptr, nwake int, nreq int) (int, error) {
+	return m.doRequeue(nil, addr, 0, naddr, nwake, nreq)
+}
+
+// RequeueCmp atomically checks that the addr contains val (via the Checker),
+// wakes up to nwake waiters on addr and then unconditionally requeues nreq
+// waiters on naddr.
+func (m *Manager) RequeueCmp(c Checker, addr uintptr, val uint32, naddr uintptr, nwake int, nreq int) (int, error) {
+	return m.doRequeue(c, addr, val, naddr, nwake, nreq)
+}
+
+// WakeOp atomically applies op to the memory address addr2, wakes up to nwake1
+// waiters unconditionally from addr1, and, based on the original value at addr2
+// and a comparison encoded in op, wakes up to nwake2 waiters from addr2.
+// It returns the total number of waiters woken.
+func (m *Manager) WakeOp(c Checker, addr1 uintptr, addr2 uintptr, nwake1 int, nwake2 int, op uint32) (int, error) {
+	if err := checkAddr(addr1); err != nil {
+		return 0, err
+	}
+	if err := checkAddr(addr2); err != nil {
+		return 0, err
+	}
+
+	b1, b2 := m.lockBuckets(addr1, addr2)
+
+	done := 0
+	cond, err := c.Op(addr2, op)
+	if err == nil {
+		// Wake up up to nwake1 entries from the first bucket.
+		done = b1.wakeLocked(addr1, ^uint32(0), nwake1)
+
+		// Wake up up to nwake2 entries from the second bucket if the
+		// operation yielded true.
+		if cond {
+			done += b2.wakeLocked(addr2, ^uint32(0), nwake2)
+		}
+	}
+
+	b1.mu.Unlock()
+	if b2 != b1 {
+		b2.mu.Unlock()
+	}
+	return done, err
+}
+
+// WaitPrepare atomically checks that addr contains val (via the Checker), then
+// enqueues w to be woken by a send to w.C. If WaitPrepare returns nil, the
+// Waiter must be subsequently removed by calling WaitComplete, whether or not
+// a wakeup is received on w.C.
+func (m *Manager) WaitPrepare(w *Waiter, c Checker, addr uintptr, val uint32, bitmask uint32) error {
+	if err := checkAddr(addr); err != nil {
+		return err
+	}
+
+	// Prepare the Waiter before taking the bucket lock.
+	w.complete = 0
+	select {
+	case <-w.C:
+	default:
+	}
+	w.addr = addr
+	w.bitmask = bitmask
+
+	b := m.lockBucket(addr)
+	// This function is very hot; avoid defer.
+
+	// Perform our atomic check.
+	if err := c.Check(addr, val); err != nil {
+		b.mu.Unlock()
+		return err
+	}
+
+	// Add the waiter to the bucket.
+	b.waiters.PushBack(w)
+
+	b.mu.Unlock()
+	return nil
+}
+
+// WaitComplete must be called when a Waiter previously added by WaitPrepare is
+// no longer eligible to be woken.
+func (m *Manager) WaitComplete(w *Waiter) {
+	// Can we short-circuit acquiring the lock?
+	// This is the happy path where a notification
+	// was received and we don't need to dequeue this
+	// waiter from any list (or take any locks).
+	if atomic.LoadInt32(&w.complete) != 0 {
+		return
+	}
+
+	// Take the bucket lock. Note that without holding the bucket lock, the
+	// waiter is not guaranteed to stay in that bucket, so after we take the
+	// bucket lock, we must ensure that the bucket hasn't changed: if it
+	// happens to have changed, we release the old bucket lock and try again
+	// with the new bucket; if it hasn't changed, we know it won't change now
+	// because we hold the lock.
+	var b *bucket
+	for {
+		addr := atomic.LoadUintptr(&w.addr)
+		b = m.lockBucket(addr)
+		// We still have to use an atomic load here, because if w was racily
+		// requeued then w.addr is not protected by b.mu.
+		if addr == atomic.LoadUintptr(&w.addr) {
+			break
+		}
+		b.mu.Unlock()
+	}
+
+	// Remove waiter from the bucket. w.complete can only be stored with b.mu
+	// locked, so this load doesn't need to use sync/atomic.
+	if w.complete == 0 {
+		b.waiters.Remove(w)
+	}
+	b.mu.Unlock()
+}
diff --git a/pkg/sentry/kernel/futex/futex_test.go b/pkg/sentry/kernel/futex/futex_test.go
new file mode 100644
index 000000000..7b81358ec
--- /dev/null
+++ b/pkg/sentry/kernel/futex/futex_test.go
@@ -0,0 +1,500 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package futex
+
+import (
+	"math"
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"testing"
+	"unsafe"
+)
+
+const (
+	testMutexSize            = 4
+	testMutexLocked   uint32 = 1
+	testMutexUnlocked uint32 = 0
+)
+
+// testData implements the Checker interface, and allows us to
+// treat the address passed for futex operations as an index in
+// a byte slice for testing simplicity.
+type testData []byte
+
+func newTestData(size uint) testData {
+	return make([]byte, size)
+}
+
+func (t testData) Check(addr uintptr, val uint32) error {
+	if val != atomic.LoadUint32((*uint32)(unsafe.Pointer(&t[addr]))) {
+		return syscall.EAGAIN
+	}
+	return nil
+}
+
+func (t testData) Op(addr uintptr, val uint32) (bool, error) {
+	return val == 0, nil
+}
+
+// testMutex ties together a testData slice, an address, and a
+// futex manager in order to implement the sync.Locker interface.
+// Beyond being used as a Locker, this is a simple mechanism for
+// changing the underlying values for simpler tests.
+type testMutex struct {
+	a uintptr
+	d testData
+	m *Manager
+}
+
+func newTestMutex(addr uintptr, d testData, m *Manager) *testMutex {
+	return &testMutex{a: addr, d: d, m: m}
+}
+
+// Lock acquires the testMutex.
+// This may wait for it to be available via the futex manager.
+func (t *testMutex) Lock() {
+	for {
+		// Attempt to grab the lock.
+		if atomic.CompareAndSwapUint32(
+			((*uint32)(unsafe.Pointer(&t.d[t.a]))),
+			testMutexUnlocked,
+			testMutexLocked) {
+			// Lock held.
+			return
+		}
+
+		// Wait for it to be "not locked".
+		w := NewWaiter()
+		err := t.m.WaitPrepare(w, t.d, t.a, testMutexLocked, ^uint32(0))
+		if err == syscall.EAGAIN {
+			continue
+		}
+		if err != nil {
+			// Should never happen.
+			panic("WaitPrepare returned unexpected error: " + err.Error())
+		}
+		<-w.C
+		t.m.WaitComplete(w)
+	}
+}
+
+// Unlock releases the testMutex.
+// This will notify any waiters via the futex manager.
+func (t *testMutex) Unlock() {
+	// Unlock.
+	atomic.StoreUint32(((*uint32)(unsafe.Pointer(&t.d[t.a]))), testMutexUnlocked)
+
+	// Notify all waiters.
+	t.m.Wake(t.a, ^uint32(0), math.MaxInt32)
+}
+
+func TestFutexWake(t *testing.T) {
+	m := NewManager()
+	d := newTestData(testMutexSize)
+
+	// Wait for it to be locked.
+	// (This won't trigger the wake in testMutex)
+	w := NewWaiter()
+	m.WaitPrepare(w, d, 0, testMutexUnlocked, ^uint32(0))
+
+	// Wake the single thread.
+	if _, err := m.Wake(0, ^uint32(0), 1); err != nil {
+		t.Error("wake error:", err)
+	}
+
+	<-w.C
+	m.WaitComplete(w)
+}
+
+func TestFutexWakeBitmask(t *testing.T) {
+	m := NewManager()
+	d := newTestData(testMutexSize)
+
+	// Wait for it to be locked.
+	// (This won't trigger the wake in testMutex)
+	w := NewWaiter()
+	m.WaitPrepare(w, d, 0, testMutexUnlocked, 0x0000ffff)
+
+	// Wake the single thread, not using the bitmask.
+	if _, err := m.Wake(0, 0xffff0000, 1); err != nil {
+		t.Error("wake non-matching bitmask error:", err)
+	}
+
+	select {
+	case <-w.C:
+		t.Error("w is alive?")
+	default:
+	}
+
+	// Now use a matching bitmask.
+	if _, err := m.Wake(0, 0x00000001, 1); err != nil {
+		t.Error("wake matching bitmask error:", err)
+	}
+
+	<-w.C
+	m.WaitComplete(w)
+}
+
+func TestFutexWakeTwo(t *testing.T) {
+	m := NewManager()
+	d := newTestData(testMutexSize)
+
+	// Wait for it to be locked.
+	// (This won't trigger the wake in testMutex)
+	w1 := NewWaiter()
+	w2 := NewWaiter()
+	w3 := NewWaiter()
+	m.WaitPrepare(w1, d, 0, testMutexUnlocked, ^uint32(0))
+	m.WaitPrepare(w2, d, 0, testMutexUnlocked, ^uint32(0))
+	m.WaitPrepare(w3, d, 0, testMutexUnlocked, ^uint32(0))
+
+	// Wake exactly two threads.
+	if _, err := m.Wake(0, ^uint32(0), 2); err != nil {
+		t.Error("wake error:", err)
+	}
+
+	// Ensure exactly two are alive.
+	// We don't get guarantees about exactly which two,
+	// (although we expect them to be w1 and w2).
+	awake := 0
+	for {
+		select {
+		case <-w1.C:
+			awake++
+		case <-w2.C:
+			awake++
+		case <-w3.C:
+			awake++
+		default:
+			if awake != 2 {
+				t.Error("awake != 2?")
+			}
+
+			// Success.
+			return
+		}
+	}
+}
+
+func TestFutexWakeUnrelated(t *testing.T) {
+	m := NewManager()
+	d := newTestData(2 * testMutexSize)
+
+	// Wait for it to be locked.
+	w1 := NewWaiter()
+	w2 := NewWaiter()
+	m.WaitPrepare(w1, d, 0*testMutexSize, testMutexUnlocked, ^uint32(0))
+	m.WaitPrepare(w2, d, 1*testMutexSize, testMutexUnlocked, ^uint32(0))
+
+	// Wake only the second one.
+	if _, err := m.Wake(1*testMutexSize, ^uint32(0), 2); err != nil {
+		t.Error("wake error:", err)
+	}
+
+	// Ensure only r2 is alive.
+	select {
+	case <-w1.C:
+		t.Error("w1 is alive?")
+	default:
+	}
+	<-w2.C
+}
+
+// This function was shamelessly stolen from mutex_test.go.
+func HammerMutex(l sync.Locker, loops int, cdone chan bool) {
+	for i := 0; i < loops; i++ {
+		l.Lock()
+		runtime.Gosched()
+		l.Unlock()
+	}
+	cdone <- true
+}
+
+func TestFutexStress(t *testing.T) {
+	m := NewManager()
+	d := newTestData(testMutexSize)
+	tm := newTestMutex(0*testMutexSize, d, m)
+	c := make(chan bool)
+
+	for i := 0; i < 10; i++ {
+		go HammerMutex(tm, 1000, c)
+	}
+
+	for i := 0; i < 10; i++ {
+		<-c
+	}
+}
+
+func TestWakeOpEmpty(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 0)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 0 {
+		t.Fatalf("Invalid number of wakes: want 0, got %d", n)
+	}
+}
+
+func TestWakeOpFirstNonEmpty(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add two waiters on address 0.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	// Wake up all waiters on address 0.
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 0)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 2 {
+		t.Fatalf("Invalid number of wakes: want 2, got %d", n)
+	}
+}
+
+func TestWakeOpSecondNonEmpty(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add two waiters on address 4.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	// Wake up all waiters on address 4.
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 0)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 2 {
+		t.Fatalf("Invalid number of wakes: want 2, got %d", n)
+	}
+}
+
+func TestWakeOpSecondNonEmptyFailingOp(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add two waiters on address 4.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	// Wake up all waiters on address 4.
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 1)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 0 {
+		t.Fatalf("Invalid number of wakes: want 0, got %d", n)
+	}
+}
+
+func TestWakeOpAllNonEmpty(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add two waiters on address 0.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	// Add two waiters on address 4.
+	w3 := NewWaiter()
+	if err := m.WaitPrepare(w3, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w3)
+
+	w4 := NewWaiter()
+	if err := m.WaitPrepare(w4, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w4)
+
+	// Wake up all waiters on both addresses.
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 0)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 4 {
+		t.Fatalf("Invalid number of wakes: want 4, got %d", n)
+	}
+}
+
+func TestWakeOpAllNonEmptyFailingOp(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add two waiters on address 0.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	// Add two waiters on address 4.
+	w3 := NewWaiter()
+	if err := m.WaitPrepare(w3, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w3)
+
+	w4 := NewWaiter()
+	if err := m.WaitPrepare(w4, d, 4, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w4)
+
+	// Wake up all waiters on both addresses.
+	n, err := m.WakeOp(d, 0, 4, 10, 10, 1)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 2 {
+		t.Fatalf("Invalid number of wakes: want 2, got %d", n)
+	}
+}
+
+func TestWakeOpSameAddress(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add four waiters on address 0.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	w3 := NewWaiter()
+	if err := m.WaitPrepare(w3, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w3)
+
+	w4 := NewWaiter()
+	if err := m.WaitPrepare(w4, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w4)
+
+	// Use the same address, with one at most one waiter from each.
+	n, err := m.WakeOp(d, 0, 0, 1, 1, 0)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 2 {
+		t.Fatalf("Invalid number of wakes: want 2, got %d", n)
+	}
+}
+
+func TestWakeOpSameAddressFailingOp(t *testing.T) {
+	m := NewManager()
+	d := newTestData(8)
+
+	// Add four waiters on address 0.
+	w1 := NewWaiter()
+	if err := m.WaitPrepare(w1, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w1)
+
+	w2 := NewWaiter()
+	if err := m.WaitPrepare(w2, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w2)
+
+	w3 := NewWaiter()
+	if err := m.WaitPrepare(w3, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w3)
+
+	w4 := NewWaiter()
+	if err := m.WaitPrepare(w4, d, 0, 0, ^uint32(0)); err != nil {
+		t.Fatalf("WaitPrepare failed: %v", err)
+	}
+	defer m.WaitComplete(w4)
+
+	// Use the same address, with one at most one waiter from each.
+	n, err := m.WakeOp(d, 0, 0, 1, 1, 1)
+	if err != nil {
+		t.Fatalf("WakeOp failed: %v", err)
+	}
+
+	if n != 1 {
+		t.Fatalf("Invalid number of wakes: want 1, got %d", n)
+	}
+}
diff --git a/pkg/sentry/kernel/g3doc/run_states.dot b/pkg/sentry/kernel/g3doc/run_states.dot
new file mode 100644
index 000000000..7861fe1f5
--- /dev/null
+++ b/pkg/sentry/kernel/g3doc/run_states.dot
@@ -0,0 +1,99 @@
+digraph {
+  subgraph {
+    App;
+  }
+  subgraph {
+    Interrupt;
+    InterruptAfterSignalDeliveryStop;
+  }
+  subgraph {
+    Syscall;
+    SyscallAfterPtraceEventSeccomp;
+    SyscallEnter;
+    SyscallAfterSyscallEnterStop;
+    SyscallAfterSysemuStop;
+    SyscallInvoke;
+    SyscallAfterPtraceEventClone;
+    SyscallAfterExecStop;
+    SyscallAfterVforkStop;
+    SyscallReinvoke;
+    SyscallExit;
+  }
+  subgraph {
+    Vsyscall;
+    VsyscallAfterPtraceEventSeccomp;
+    VsyscallInvoke;
+  }
+  subgraph {
+    Exit;
+    ExitMain; // leave thread group, release resources, reparent children, kill PID namespace and wait if TGID 1
+    ExitNotify; // signal parent/tracer, become waitable
+    ExitDone; // represented by t.runState == nil
+  }
+
+  // Task exit
+  Exit -> ExitMain;
+  ExitMain -> ExitNotify;
+  ExitNotify -> ExitDone;
+
+  // Execution of untrusted application code
+  App -> App;
+
+  // Interrupts (usually signal delivery)
+  App -> Interrupt;
+  Interrupt -> Interrupt; // if other interrupt conditions may still apply
+  Interrupt -> Exit; // if killed
+
+  // Syscalls
+  App -> Syscall;
+  Syscall -> SyscallEnter;
+  SyscallEnter -> SyscallInvoke;
+  SyscallInvoke -> SyscallExit;
+  SyscallExit -> App;
+
+  // exit, exit_group
+  SyscallInvoke -> Exit;
+
+  // execve
+  SyscallInvoke -> SyscallAfterExecStop;
+  SyscallAfterExecStop -> SyscallExit;
+  SyscallAfterExecStop -> App; // fatal signal pending
+
+  // vfork
+  SyscallInvoke -> SyscallAfterVforkStop;
+  SyscallAfterVforkStop -> SyscallExit;
+
+  // Vsyscalls
+  App -> Vsyscall;
+  Vsyscall -> VsyscallInvoke;
+  Vsyscall -> App; // fault while reading return address from stack
+  VsyscallInvoke -> App;
+
+  // ptrace-specific branches
+  Interrupt -> InterruptAfterSignalDeliveryStop;
+  InterruptAfterSignalDeliveryStop -> Interrupt;
+  SyscallEnter -> SyscallAfterSyscallEnterStop;
+  SyscallAfterSyscallEnterStop -> SyscallInvoke;
+  SyscallAfterSyscallEnterStop -> SyscallExit; // skipped by tracer
+  SyscallAfterSyscallEnterStop -> App; // fatal signal pending
+  SyscallEnter -> SyscallAfterSysemuStop;
+  SyscallAfterSysemuStop -> SyscallExit;
+  SyscallAfterSysemuStop -> App; // fatal signal pending
+  SyscallInvoke -> SyscallAfterPtraceEventClone;
+  SyscallAfterPtraceEventClone -> SyscallExit;
+  SyscallAfterPtraceEventClone -> SyscallAfterVforkStop;
+
+  // seccomp
+  Syscall -> App; // SECCOMP_RET_TRAP, SECCOMP_RET_ERRNO, SECCOMP_RET_KILL, SECCOMP_RET_TRACE without tracer
+  Syscall -> SyscallAfterPtraceEventSeccomp; // SECCOMP_RET_TRACE
+  SyscallAfterPtraceEventSeccomp -> SyscallEnter;
+  SyscallAfterPtraceEventSeccomp -> SyscallExit; // skipped by tracer
+  SyscallAfterPtraceEventSeccomp -> App; // fatal signal pending
+  Vsyscall -> VsyscallAfterPtraceEventSeccomp;
+  VsyscallAfterPtraceEventSeccomp -> VsyscallInvoke;
+  VsyscallAfterPtraceEventSeccomp -> App;
+
+  // Autosave
+  SyscallInvoke -> SyscallReinvoke;
+  SyscallReinvoke -> SyscallInvoke;
+}
diff --git a/pkg/sentry/kernel/ipc_namespace.go b/pkg/sentry/kernel/ipc_namespace.go
new file mode 100644
index 000000000..78737f58f
--- /dev/null
+++ b/pkg/sentry/kernel/ipc_namespace.go
@@ -0,0 +1,43 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/semaphore"
+)
+
+// IPCNamespace represents an IPC namespace.
+type IPCNamespace struct {
+	semaphores *semaphore.Registry
+}
+
+// NewIPCNamespace creates a new IPC namespace.
+func NewIPCNamespace() *IPCNamespace {
+	return &IPCNamespace{
+		semaphores: semaphore.NewRegistry(),
+	}
+}
+
+// SemaphoreRegistry returns the semanphore set registry for this namespace.
+func (i *IPCNamespace) SemaphoreRegistry() *semaphore.Registry {
+	return i.semaphores
+}
+
+// IPCNamespace returns the task's IPC namespace.
+func (t *Task) IPCNamespace() *IPCNamespace {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.ipcns
+}
diff --git a/pkg/sentry/kernel/kdefs/BUILD b/pkg/sentry/kernel/kdefs/BUILD
new file mode 100644
index 000000000..b6c00042a
--- /dev/null
+++ b/pkg/sentry/kernel/kdefs/BUILD
@@ -0,0 +1,10 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "kdefs",
+    srcs = ["kdefs.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs",
+    visibility = ["//:sandbox"],
+)
diff --git a/pkg/sentry/kernel/kdefs/kdefs.go b/pkg/sentry/kernel/kdefs/kdefs.go
new file mode 100644
index 000000000..bbb476544
--- /dev/null
+++ b/pkg/sentry/kernel/kdefs/kdefs.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package kdefs defines common kernel definitions.
+//
+package kdefs
+
+// FD is a File Descriptor.
+type FD int32
diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go
new file mode 100644
index 000000000..0932965e0
--- /dev/null
+++ b/pkg/sentry/kernel/kernel.go
@@ -0,0 +1,957 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package kernel provides an emulation of the Linux kernel.
+//
+// See README.md for a detailed overview.
+//
+// Lock order (outermost locks must be taken first):
+//
+// Kernel.extMu
+//   TaskSet.mu
+//     SignalHandlers.mu
+//       Task.mu
+//
+// Locking SignalHandlers.mu in multiple SignalHandlers requires locking
+// TaskSet.mu exclusively first. Locking Task.mu in multiple Tasks at the same
+// time requires locking all of their signal mutexes first.
+package kernel
+
+import (
+	"fmt"
+	"io"
+	"path/filepath"
+	"sync"
+	"sync/atomic"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/timerfd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/hostcpu"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/epoll"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/loader"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink/port"
+	sentrytime "gvisor.googlesource.com/gvisor/pkg/sentry/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/state"
+)
+
+// Kernel represents an emulated Linux kernel. It must be initialized by calling
+// Init() or LoadFrom().
+type Kernel struct {
+	// extMu serializes external changes to the Kernel with calls to
+	// Kernel.SaveTo. (Kernel.SaveTo requires that the state of the Kernel
+	// remains frozen for the duration of the call; it requires that the Kernel
+	// is paused as a precondition, which ensures that none of the tasks
+	// running within the Kernel can affect its state, but extMu is required to
+	// ensure that concurrent users of the Kernel *outside* the Kernel's
+	// control cannot affect its state by calling e.g.
+	// Kernel.SendExternalSignal.)
+	extMu sync.Mutex `state:"nosave"`
+
+	// started is true if Start has been called. Unless otherwise specified,
+	// all Kernel fields become immutable once started becomes true.
+	started bool `state:"nosave"`
+
+	// All of the following fields are immutable unless otherwise specified.
+
+	// Platform is the platform that is used to execute tasks in the
+	// created Kernel. It is embedded so that Kernel can directly serve as
+	// Platform in mm logic and also serve as platform.MemoryProvider in
+	// filemem S/R logic.
+	platform.Platform `state:"nosave"`
+
+	// See InitKernelArgs for the meaning of these fields.
+	featureSet        *cpuid.FeatureSet
+	timekeeper        *Timekeeper
+	tasks             *TaskSet
+	rootUserNamespace *auth.UserNamespace
+	networkStack      inet.Stack `state:"nosave"`
+	applicationCores  uint
+	useHostCores      bool
+	extraAuxv         []arch.AuxEntry
+	vdso              *loader.VDSO
+	rootUTSNamespace  *UTSNamespace
+	rootIPCNamespace  *IPCNamespace
+
+	// mounts holds the state of the virtual filesystem. mounts is initially
+	// nil, and must be set by calling Kernel.SetRootMountNamespace before
+	// Kernel.CreateProcess can succeed.
+	mounts *fs.MountNamespace
+
+	// globalInit is the thread group whose leader has ID 1 in the root PID
+	// namespace. globalInit is stored separately so that it is accessible even
+	// after all tasks in the thread group have exited, such that ID 1 is no
+	// longer mapped.
+	//
+	// globalInit is mutable until it is assigned by the first successful call
+	// to CreateProcess, and is protected by extMu.
+	globalInit *ThreadGroup
+
+	// realtimeClock is a ktime.Clock based on timekeeper's Realtime.
+	realtimeClock *timekeeperClock
+
+	// monotonicClock is a ktime.Clock based on timekeeper's Monotonic.
+	monotonicClock *timekeeperClock
+
+	// syslog is the kernel log.
+	syslog syslog
+
+	// cpuClock is incremented every linux.ClockTick. cpuClock is used to
+	// measure task CPU usage, since sampling monotonicClock twice on every
+	// syscall turns out to be unreasonably expensive. This is similar to how
+	// Linux does task CPU accounting on x86 (CONFIG_IRQ_TIME_ACCOUNTING),
+	// although Linux also uses scheduler timing information to improve
+	// resolution (kernel/sched/cputime.c:cputime_adjust()), which we can't do
+	// since "preeemptive" scheduling is managed by the Go runtime, which
+	// doesn't provide this information.
+	//
+	// cpuClock is mutable, and is accessed using atomic memory operations.
+	cpuClock uint64
+
+	// cpuClockTicker increments cpuClock.
+	cpuClockTicker *ktime.Timer `state:"nosave"`
+
+	// fdMapUids is an ever-increasing counter for generating FDMap uids.
+	//
+	// fdMapUids is mutable, and is accessed using atomic memory operations.
+	fdMapUids uint64
+
+	// uniqueID is used to generate unique identifiers.
+	//
+	// uniqueID is mutable, and is accessed using atomic memory operations.
+	uniqueID uint64
+
+	// nextInotifyCookie is a monotonically increasing counter used for
+	// generating unique inotify event cookies.
+	//
+	// nextInotifyCookie is mutable, and is accesed using atomic memory
+	// operations.
+	nextInotifyCookie uint32
+
+	// netlinkPorts manages allocation of netlink socket port IDs.
+	netlinkPorts *port.Manager
+
+	// exitErr is the error causing the sandbox to exit, if any. It is
+	// protected by extMu.
+	exitErr error
+}
+
+// InitKernelArgs holds arguments to Init.
+type InitKernelArgs struct {
+	// FeatureSet is the emulated CPU feature set.
+	FeatureSet *cpuid.FeatureSet
+
+	// Timekeeper manages time for all tasks in the system.
+	Timekeeper *Timekeeper
+
+	// RootUserNamespace is the root user namespace.
+	RootUserNamespace *auth.UserNamespace
+
+	// NetworkStack is the TCP/IP network stack. NetworkStack may be nil.
+	NetworkStack inet.Stack
+
+	// ApplicationCores is the number of logical CPUs visible to sandboxed
+	// applications. The set of logical CPU IDs is [0, ApplicationCores); thus
+	// ApplicationCores is analogous to Linux's nr_cpu_ids, the index of the
+	// most significant bit in cpu_possible_mask + 1.
+	ApplicationCores uint
+
+	// If UseHostCores is true, Task.CPU() returns the task goroutine's CPU
+	// instead of a virtualized CPU number, and Task.CopyToCPUMask() is a
+	// no-op. If ApplicationCores is less than hostcpu.MaxPossibleCPU(), it
+	// will be overridden.
+	UseHostCores bool
+
+	// ExtraAuxv contains additional auxiliary vector entries that are added to
+	// each process by the ELF loader.
+	ExtraAuxv []arch.AuxEntry
+
+	// Vdso holds the VDSO and its parameter page.
+	Vdso *loader.VDSO
+
+	// RootUTSNamespace is the root UTS namepsace.
+	RootUTSNamespace *UTSNamespace
+
+	// RootIPCNamespace is the root IPC namepsace.
+	RootIPCNamespace *IPCNamespace
+}
+
+// Init initialize the Kernel with no tasks.
+//
+// Callers must manually set Kernel.Platform before caling Init.
+func (k *Kernel) Init(args InitKernelArgs) error {
+	if args.FeatureSet == nil {
+		return fmt.Errorf("FeatureSet is nil")
+	}
+	if args.Timekeeper == nil {
+		return fmt.Errorf("Timekeeper is nil")
+	}
+	if args.RootUserNamespace == nil {
+		return fmt.Errorf("RootUserNamespace is nil")
+	}
+	if args.ApplicationCores == 0 {
+		return fmt.Errorf("ApplicationCores is 0")
+	}
+
+	k.featureSet = args.FeatureSet
+	k.timekeeper = args.Timekeeper
+	k.tasks = newTaskSet()
+	k.rootUserNamespace = args.RootUserNamespace
+	k.rootUTSNamespace = args.RootUTSNamespace
+	k.rootIPCNamespace = args.RootIPCNamespace
+	k.networkStack = args.NetworkStack
+	k.applicationCores = args.ApplicationCores
+	if args.UseHostCores {
+		k.useHostCores = true
+		maxCPU, err := hostcpu.MaxPossibleCPU()
+		if err != nil {
+			return fmt.Errorf("Failed to get maximum CPU number: %v", err)
+		}
+		minAppCores := uint(maxCPU) + 1
+		if k.applicationCores < minAppCores {
+			log.Infof("UseHostCores enabled: increasing ApplicationCores from %d to %d", k.applicationCores, minAppCores)
+			k.applicationCores = minAppCores
+		}
+	}
+	k.extraAuxv = args.ExtraAuxv
+	k.vdso = args.Vdso
+	k.realtimeClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Realtime}
+	k.monotonicClock = &timekeeperClock{tk: args.Timekeeper, c: sentrytime.Monotonic}
+	k.netlinkPorts = port.New()
+
+	return nil
+}
+
+// SaveTo saves the state of k to w.
+//
+// Preconditions: The kernel must be paused throughout the call to SaveTo.
+func (k *Kernel) SaveTo(w io.Writer) error {
+	saveStart := time.Now()
+	ctx := k.SupervisorContext()
+
+	// Do not allow other Kernel methods to affect it while it's being saved.
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+
+	// Stop time.
+	k.pauseTimeLocked()
+	defer k.resumeTimeLocked()
+
+	// Flush write operations on open files so data reaches backing storage.
+	if err := k.tasks.flushWritesToFiles(ctx); err != nil {
+		return err
+	}
+
+	// Remove all epoll waiter objects from underlying wait queues.
+	// NOTE: for programs to resume execution in future snapshot scenarios,
+	// we will need to re-establish these waiter objects after saving.
+	k.tasks.unregisterEpollWaiters()
+
+	// Clear the dirent cache before saving because Dirents must be Loaded in a
+	// particular order (parents before children), and Loading dirents from a cache
+	// breaks that order.
+	k.mounts.FlushMountSourceRefs()
+
+	// Ensure that all pending asynchronous work is complete:
+	//   - inode and mount release
+	//   - asynchronuous IO
+	fs.AsyncBarrier()
+
+	// Once all fs work has completed (flushed references have all been released),
+	// reset mount mappings. This allows individual mounts to save how inodes map
+	// to filesystem resources. Without this, fs.Inodes cannot be restored.
+	fs.SaveInodeMappings()
+
+	// Discard unsavable mappings, such as those for host file descriptors.
+	// This must be done after waiting for "asynchronous fs work", which
+	// includes async I/O that may touch application memory.
+	if err := k.invalidateUnsavableMappings(ctx); err != nil {
+		return fmt.Errorf("failed to invalidate unsavable mappings: %v", err)
+	}
+
+	// Save the kernel state.
+	kernelStart := time.Now()
+	var stats state.Stats
+	if err := state.Save(w, k, &stats); err != nil {
+		return err
+	}
+	log.Infof("Kernel save stats: %s", &stats)
+	log.Infof("Kernel save took [%s].", time.Since(kernelStart))
+
+	// Save the memory state.
+	//
+	// FIXME: In the future, this should not be dispatched via
+	// an abstract memory type. This should be dispatched to a single
+	// memory implementation that belongs to the kernel. (There is
+	// currently a single implementation anyways, it just needs to be
+	// "unabstracted" and reparented appropriately.)
+	memoryStart := time.Now()
+	if err := k.Platform.Memory().SaveTo(w); err != nil {
+		return err
+	}
+	log.Infof("Memory save took [%s].", time.Since(memoryStart))
+
+	log.Infof("Overall save took [%s].", time.Since(saveStart))
+
+	return nil
+}
+
+func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error {
+	ts.mu.RLock()
+	defer ts.mu.RUnlock()
+	for t := range ts.Root.tids {
+		if fdmap := t.FDMap(); fdmap != nil {
+			for _, desc := range fdmap.files {
+				if flags := desc.file.Flags(); !flags.Write {
+					continue
+				}
+				if sattr := desc.file.Dirent.Inode.StableAttr; !fs.IsFile(sattr) && !fs.IsDir(sattr) {
+					continue
+				}
+				// Here we need all metadata synced.
+				syncErr := desc.file.Fsync(ctx, 0, fs.FileMaxOffset, fs.SyncAll)
+				if err := fs.SaveFileFsyncError(syncErr); err != nil {
+					name, _ := desc.file.Dirent.FullName(nil /* root */)
+					return fmt.Errorf("%q was not sufficiently synced: %v", name, err)
+				}
+			}
+		}
+	}
+	return nil
+}
+
+// Preconditions: The kernel must be paused.
+func (k *Kernel) invalidateUnsavableMappings(ctx context.Context) error {
+	invalidated := make(map[*mm.MemoryManager]struct{})
+	k.tasks.mu.RLock()
+	defer k.tasks.mu.RUnlock()
+	for t := range k.tasks.Root.tids {
+		// We can skip locking Task.mu here since the kernel is paused.
+		if mm := t.tc.MemoryManager; mm != nil {
+			if _, ok := invalidated[mm]; !ok {
+				if err := mm.InvalidateUnsavable(ctx); err != nil {
+					return err
+				}
+				invalidated[mm] = struct{}{}
+			}
+		}
+		// I really wish we just had a sync.Map of all MMs...
+		if r, ok := t.runState.(*runSyscallAfterExecStop); ok {
+			if err := r.tc.MemoryManager.InvalidateUnsavable(ctx); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+func (ts *TaskSet) unregisterEpollWaiters() {
+	ts.mu.RLock()
+	defer ts.mu.RUnlock()
+	for t := range ts.Root.tids {
+		if fdmap := t.FDMap(); fdmap != nil {
+			for _, desc := range fdmap.files {
+				if desc.file != nil {
+					if e, ok := desc.file.FileOperations.(*epoll.EventPoll); ok {
+						e.UnregisterEpollWaiters()
+					}
+				}
+			}
+		}
+	}
+}
+
+// LoadFrom returns a new Kernel loaded from args.
+func (k *Kernel) LoadFrom(r io.Reader, p platform.Platform, net inet.Stack) error {
+	loadStart := time.Now()
+	if p == nil {
+		return fmt.Errorf("Platform is nil")
+	}
+
+	k.Platform = p
+	k.networkStack = net
+
+	initAppCores := k.applicationCores
+
+	// Load the kernel state.
+	kernelStart := time.Now()
+	var stats state.Stats
+	if err := state.Load(r, k, &stats); err != nil {
+		return err
+	}
+	log.Infof("Kernel load stats: %s", &stats)
+	log.Infof("Kernel load took [%s].", time.Since(kernelStart))
+
+	// Load the memory state.
+	//
+	// See the note in SaveTo.
+	memoryStart := time.Now()
+	if err := k.Platform.Memory().LoadFrom(r); err != nil {
+		return err
+	}
+	log.Infof("Memory load took [%s].", time.Since(memoryStart))
+
+	// Ensure that all pending asynchronous work is complete:
+	//   - namedpipe opening
+	//   - inode file opening
+	fs.AsyncBarrier()
+
+	log.Infof("Overall load took [%s]", time.Since(loadStart))
+
+	// Applications may size per-cpu structures based on k.applicationCores, so
+	// it can't change across save/restore. When we are virtualizing CPU
+	// numbers, this isn't a problem. However, when we are exposing host CPU
+	// assignments, we can't tolerate an increase in the number of host CPUs,
+	// which could result in getcpu(2) returning CPUs that applications expect
+	// not to exist.
+	if k.useHostCores && initAppCores > k.applicationCores {
+		return fmt.Errorf("UseHostCores enabled: can't increase ApplicationCores from %d to %d after restore", k.applicationCores, initAppCores)
+	}
+
+	return nil
+}
+
+// Destroy releases resources owned by k.
+//
+// Preconditions: There must be no task goroutines running in k.
+func (k *Kernel) Destroy() {
+	if k.mounts != nil {
+		k.mounts.DecRef()
+		k.mounts = nil
+	}
+}
+
+// UniqueID returns a unique identifier.
+func (k *Kernel) UniqueID() uint64 {
+	id := atomic.AddUint64(&k.uniqueID, 1)
+	if id == 0 {
+		panic("unique identifier generator wrapped around")
+	}
+	return id
+}
+
+// CreateProcessArgs holds arguments to kernel.CreateProcess.
+type CreateProcessArgs struct {
+	// Filename is the filename to load.
+	//
+	// If this is provided as "", then the file will be guessed via Argv[0].
+	Filename string
+
+	// Argvv is a list of arguments.
+	Argv []string
+
+	// Envv is a list of environment variables.
+	Envv []string
+
+	// WorkingDirectory is the initial working directory.
+	//
+	// This defaults to the root if empty.
+	WorkingDirectory string
+
+	// Credentials is the initial credentials.
+	Credentials *auth.Credentials
+
+	// FDMap is the initial set of file descriptors. If CreateProcess succeeds,
+	// it takes a reference on FDMap.
+	FDMap *FDMap
+
+	// Umask is the initial umask.
+	Umask uint
+
+	// Limits is the initial resource limits.
+	Limits *limits.LimitSet
+
+	// MaxSymlinkTraversals is the maximum number of symlinks to follow
+	// during resolution.
+	MaxSymlinkTraversals uint
+
+	// UTSNamespace is the initial UTS namespace.
+	UTSNamespace *UTSNamespace
+
+	// IPCNamespace is the initial IPC namespace.
+	IPCNamespace *IPCNamespace
+}
+
+// NewContext returns a context.Context that represents the task that will be
+// created by args.NewContext(k).
+func (args *CreateProcessArgs) NewContext(k *Kernel) *createProcessContext {
+	return &createProcessContext{
+		Logger: log.Log(),
+		k:      k,
+		args:   args,
+	}
+}
+
+// createProcessContext is a context.Context that represents the context
+// associated with a task that is being created.
+type createProcessContext struct {
+	context.NoopSleeper
+	log.Logger
+	k    *Kernel
+	args *CreateProcessArgs
+}
+
+// Value implements context.Context.Value.
+func (ctx *createProcessContext) Value(key interface{}) interface{} {
+	switch key {
+	case CtxKernel:
+		return ctx.k
+	case CtxPIDNamespace:
+		// "The new task ... is in the root PID namespace." -
+		// Kernel.CreateProcess
+		return ctx.k.tasks.Root
+	case CtxUTSNamespace:
+		return ctx.args.UTSNamespace
+	case CtxIPCNamespace:
+		return ctx.args.IPCNamespace
+	case auth.CtxCredentials:
+		return ctx.args.Credentials
+	case fs.CtxRoot:
+		if ctx.k.mounts == nil {
+			return nil
+		}
+		return ctx.k.mounts.Root()
+	case ktime.CtxRealtimeClock:
+		return ctx.k.RealtimeClock()
+	case limits.CtxLimits:
+		return ctx.args.Limits
+	case platform.CtxPlatform:
+		return ctx.k
+	case uniqueid.CtxGlobalUniqueID:
+		return ctx.k.UniqueID()
+	case uniqueid.CtxInotifyCookie:
+		return ctx.k.GenerateInotifyCookie()
+	default:
+		return nil
+	}
+}
+
+// CreateProcess creates a new task in a new thread group with the given
+// options. The new task has no parent and is in the root PID namespace.
+//
+// If k.Start() has already been called, the created task will begin running
+// immediately. Otherwise, it will be started when k.Start() is called.
+//
+// CreateProcess has no analogue in Linux; it is used to create the initial
+// application task, as well as processes started by the control server.
+func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, error) {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	log.Infof("EXEC: %v", args.Argv)
+
+	if k.mounts == nil {
+		return nil, fmt.Errorf("no kernel MountNamespace")
+	}
+
+	tg := NewThreadGroup(k.tasks.Root, NewSignalHandlers(), linux.SIGCHLD, args.Limits, k.monotonicClock)
+	ctx := args.NewContext(k)
+
+	// Grab the root directory.
+	root := fs.RootFromContext(ctx)
+	defer root.DecRef()
+
+	// Grab the working directory.
+	wd := root // Default.
+	if args.WorkingDirectory != "" {
+		var err error
+		wd, err = k.mounts.FindInode(ctx, root, nil, args.WorkingDirectory, args.MaxSymlinkTraversals)
+		if err != nil {
+			return nil, fmt.Errorf("failed to find initial working directory %q: %v", args.WorkingDirectory, err)
+		}
+		defer wd.DecRef()
+	}
+
+	if args.Filename == "" {
+		// Was anything provided?
+		if len(args.Argv) == 0 {
+			return nil, fmt.Errorf("no filename or command provided")
+		}
+		if !filepath.IsAbs(args.Argv[0]) {
+			return nil, fmt.Errorf("'%s' is not an absolute path", args.Argv[0])
+		}
+		args.Filename = args.Argv[0]
+	}
+
+	// Create a fresh task context.
+	tc, err := k.LoadTaskImage(ctx, k.mounts, root, wd, args.MaxSymlinkTraversals, args.Filename, args.Argv, args.Envv, k.featureSet)
+	if err != nil {
+		return nil, err
+	}
+	tr := newTaskResources(args.FDMap, newFSContext(root, wd, args.Umask))
+	// NewTask unconditionally takes ownership of tr, so we never have to call
+	// tr.release.
+
+	// Create the task.
+	config := &TaskConfig{
+		Kernel:         k,
+		ThreadGroup:    tg,
+		TaskContext:    tc,
+		TaskResources:  tr,
+		Credentials:    args.Credentials,
+		UTSNamespace:   args.UTSNamespace,
+		IPCNamespace:   args.IPCNamespace,
+		AllowedCPUMask: sched.NewFullCPUSet(k.applicationCores),
+	}
+	t, err := k.tasks.NewTask(config)
+	if err != nil {
+		return nil, err
+	}
+
+	// Success.
+	if k.started {
+		tid := k.tasks.Root.IDOfTask(t)
+		t.Start(tid)
+	} else if k.globalInit == nil {
+		k.globalInit = tg
+	}
+	return tg, nil
+}
+
+// Start starts execution of all tasks in k.
+//
+// Preconditions: Start may be called exactly once.
+func (k *Kernel) Start() error {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+
+	if k.globalInit == nil {
+		return fmt.Errorf("kernel contains no tasks")
+	}
+	if k.started {
+		return fmt.Errorf("kernel already started")
+	}
+
+	k.started = true
+	k.cpuClockTicker = ktime.NewTimer(k.monotonicClock, kernelCPUClockListener{k})
+	k.cpuClockTicker.Swap(ktime.Setting{
+		Enabled: true,
+		Period:  linux.ClockTick,
+	})
+	// If k was created by LoadKernelFrom, timers were stopped during
+	// Kernel.SaveTo and need to be resumed. If k was created by NewKernel,
+	// this is a no-op.
+	k.resumeTimeLocked()
+	// Start task goroutines.
+	k.tasks.mu.RLock()
+	defer k.tasks.mu.RUnlock()
+	for t, tid := range k.tasks.Root.tids {
+		t.Start(tid)
+	}
+	return nil
+}
+
+// pauseTimeLocked pauses all Timers and Timekeeper updates.
+//
+// Preconditions: Any task goroutines running in k must be stopped. k.extMu
+// must be locked.
+func (k *Kernel) pauseTimeLocked() {
+	// k.cpuClockTicker may be nil since Kernel.SaveTo() may be called before
+	// Kernel.Start().
+	if k.cpuClockTicker != nil {
+		k.cpuClockTicker.Pause()
+	}
+
+	// By precondition, nothing else can be interacting with PIDNamespace.tids
+	// or FDMap.files, so we can iterate them without synchronization. (We
+	// can't hold the TaskSet mutex when pausing thread group timers because
+	// thread group timers call ThreadGroup.SendSignal, which takes the TaskSet
+	// mutex, while holding the Timer mutex.)
+	for t := range k.tasks.Root.tids {
+		if t == t.tg.leader {
+			t.tg.tm.pause()
+		}
+		// This means we'll iterate FDMaps shared by multiple tasks repeatedly,
+		// but ktime.Timer.Pause is idempotent so this is harmless.
+		if fdm := t.tr.FDMap; fdm != nil {
+			for _, desc := range fdm.files {
+				if tfd, ok := desc.file.FileOperations.(*timerfd.TimerOperations); ok {
+					tfd.PauseTimer()
+				}
+			}
+		}
+	}
+	k.timekeeper.PauseUpdates()
+}
+
+// resumeTimeLocked resumes all Timers and Timekeeper updates. If
+// pauseTimeLocked has not been previously called, resumeTimeLocked has no
+// effect.
+//
+// Preconditions: Any task goroutines running in k must be stopped. k.extMu
+// must be locked.
+func (k *Kernel) resumeTimeLocked() {
+	if k.cpuClockTicker != nil {
+		k.cpuClockTicker.Resume()
+	}
+
+	k.timekeeper.ResumeUpdates()
+	for t := range k.tasks.Root.tids {
+		if t == t.tg.leader {
+			t.tg.tm.resume()
+		}
+		if fdm := t.tr.FDMap; fdm != nil {
+			for _, desc := range fdm.files {
+				if tfd, ok := desc.file.FileOperations.(*timerfd.TimerOperations); ok {
+					tfd.ResumeTimer()
+				}
+			}
+		}
+	}
+}
+
+// WaitExited blocks until all tasks in k have exited.
+func (k *Kernel) WaitExited() {
+	k.tasks.liveGoroutines.Wait()
+}
+
+// Kill requests that all tasks in k immediately exit as if group exiting with
+// status es. Kill does not wait for tasks to exit.
+func (k *Kernel) Kill(es ExitStatus) {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	k.tasks.Kill(es)
+}
+
+// Pause requests that all tasks in k temporarily stop executing, and blocks
+// until all tasks in k have stopped. Multiple calls to Pause nest and require
+// an equal number of calls to Unpause to resume execution.
+func (k *Kernel) Pause() {
+	k.extMu.Lock()
+	k.tasks.BeginExternalStop()
+	k.extMu.Unlock()
+	k.tasks.runningGoroutines.Wait()
+}
+
+// Unpause ends the effect of a previous call to Pause. If Unpause is called
+// without a matching preceding call to Pause, Unpause may panic.
+func (k *Kernel) Unpause() {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	k.tasks.EndExternalStop()
+}
+
+// SendExternalSignal injects a signal into the kernel.
+//
+// context is used only for debugging to describe how the signal was received.
+//
+// Returns false if signal could not be sent because the Kernel is not fully
+// initialized yet.
+func (k *Kernel) SendExternalSignal(info *arch.SignalInfo, context string) bool {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	return k.sendExternalSignal(info, context)
+}
+
+// FeatureSet returns the FeatureSet.
+func (k *Kernel) FeatureSet() *cpuid.FeatureSet {
+	return k.featureSet
+}
+
+// Timekeeper returns the Timekeeper.
+func (k *Kernel) Timekeeper() *Timekeeper {
+	return k.timekeeper
+}
+
+// TaskSet returns the TaskSet.
+func (k *Kernel) TaskSet() *TaskSet {
+	return k.tasks
+}
+
+// RootUserNamespace returns the root UserNamespace.
+func (k *Kernel) RootUserNamespace() *auth.UserNamespace {
+	return k.rootUserNamespace
+}
+
+// RootUTSNamespace returns the root UTSNamespace.
+func (k *Kernel) RootUTSNamespace() *UTSNamespace {
+	return k.rootUTSNamespace
+}
+
+// RootIPCNamespace returns the root IPCNamespace.
+func (k *Kernel) RootIPCNamespace() *IPCNamespace {
+	return k.rootIPCNamespace
+}
+
+// RootMountNamespace returns the MountNamespace.
+func (k *Kernel) RootMountNamespace() *fs.MountNamespace {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	return k.mounts
+}
+
+// SetRootMountNamespace sets the MountNamespace.
+func (k *Kernel) SetRootMountNamespace(mounts *fs.MountNamespace) {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	k.mounts = mounts
+}
+
+// NetworkStack returns the network stack. NetworkStack may return nil if no
+// network stack is available.
+func (k *Kernel) NetworkStack() inet.Stack {
+	return k.networkStack
+}
+
+// GlobalInit returns the thread group with ID 1 in the root PID namespace, or
+// nil if no such thread group exists. GlobalInit may return a thread group
+// containing no tasks if the thread group has already exited.
+func (k *Kernel) GlobalInit() *ThreadGroup {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	return k.globalInit
+}
+
+// ApplicationCores returns the number of CPUs visible to sandboxed
+// applications.
+func (k *Kernel) ApplicationCores() uint {
+	return k.applicationCores
+}
+
+// RealtimeClock returns the application CLOCK_REALTIME clock.
+func (k *Kernel) RealtimeClock() ktime.Clock {
+	return k.realtimeClock
+}
+
+// MonotonicClock returns the application CLOCK_MONOTONIC clock.
+func (k *Kernel) MonotonicClock() ktime.Clock {
+	return k.monotonicClock
+}
+
+// CPUClockNow returns the current value of k.cpuClock.
+func (k *Kernel) CPUClockNow() uint64 {
+	return atomic.LoadUint64(&k.cpuClock)
+}
+
+// Syslog returns the syslog.
+func (k *Kernel) Syslog() *syslog {
+	return &k.syslog
+}
+
+// GenerateInotifyCookie generates a unique inotify event cookie.
+//
+// Returned values may overlap with previously returned values if the value
+// space is exhausted. 0 is not a valid cookie value, all other values
+// representable in a uint32 are allowed.
+func (k *Kernel) GenerateInotifyCookie() uint32 {
+	id := atomic.AddUint32(&k.nextInotifyCookie, 1)
+	// Wrap-around is explicitly allowed for inotify event cookies.
+	if id == 0 {
+		id = atomic.AddUint32(&k.nextInotifyCookie, 1)
+	}
+	return id
+}
+
+// NetlinkPorts returns the netlink port manager.
+func (k *Kernel) NetlinkPorts() *port.Manager {
+	return k.netlinkPorts
+}
+
+// ExitError returns the sandbox error that caused the kernel to exit.
+func (k *Kernel) ExitError() error {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	return k.exitErr
+}
+
+// SetExitError sets the sandbox error that caused the kernel to exit, if one is
+// not already set.
+func (k *Kernel) SetExitError(err error) {
+	k.extMu.Lock()
+	defer k.extMu.Unlock()
+	if k.exitErr == nil {
+		k.exitErr = err
+	}
+}
+
+// SupervisorContext returns a Context with maximum privileges in k. It should
+// only be used by goroutines outside the control of the emulated kernel
+// defined by e.
+//
+// Callers are responsible for ensuring that the returned Context is not used
+// concurrently with changes to the Kernel.
+func (k *Kernel) SupervisorContext() context.Context {
+	return supervisorContext{
+		Logger: log.Log(),
+		k:      k,
+	}
+}
+
+type supervisorContext struct {
+	context.NoopSleeper
+	log.Logger
+	k *Kernel
+}
+
+// Value implements context.Context.
+func (ctx supervisorContext) Value(key interface{}) interface{} {
+	switch key {
+	case CtxCanTrace:
+		// The supervisor context can trace anything. (None of
+		// supervisorContext's users are expected to invoke ptrace, but ptrace
+		// permissions are required for certain file accesses.)
+		return func(*Task, bool) bool { return true }
+	case CtxKernel:
+		return ctx.k
+	case CtxPIDNamespace:
+		return ctx.k.tasks.Root
+	case CtxUTSNamespace:
+		return ctx.k.rootUTSNamespace
+	case CtxIPCNamespace:
+		return ctx.k.rootIPCNamespace
+	case auth.CtxCredentials:
+		// The supervisor context is global root.
+		return auth.NewRootCredentials(ctx.k.rootUserNamespace)
+	case fs.CtxRoot:
+		return ctx.k.mounts.Root()
+	case ktime.CtxRealtimeClock:
+		return ctx.k.RealtimeClock()
+	case limits.CtxLimits:
+		// No limits apply.
+		return limits.NewLimitSet()
+	case platform.CtxPlatform:
+		return ctx.k
+	case uniqueid.CtxGlobalUniqueID:
+		return ctx.k.UniqueID()
+	case uniqueid.CtxInotifyCookie:
+		return ctx.k.GenerateInotifyCookie()
+	default:
+		return nil
+	}
+}
+
+type kernelCPUClockListener struct {
+	k *Kernel
+}
+
+// Notify implements ktime.TimerListener.Notify.
+func (l kernelCPUClockListener) Notify(exp uint64) {
+	atomic.AddUint64(&l.k.cpuClock, exp)
+}
+
+// Destroy implements ktime.TimerListener.Destroy.
+func (l kernelCPUClockListener) Destroy() {
+}
diff --git a/pkg/sentry/kernel/memevent/BUILD b/pkg/sentry/kernel/memevent/BUILD
new file mode 100644
index 000000000..c7779e1d5
--- /dev/null
+++ b/pkg/sentry/kernel/memevent/BUILD
@@ -0,0 +1,31 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "memevent",
+    srcs = ["memory_events.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/memevent",
+    visibility = ["//:sandbox"],
+    deps = [
+        ":memory_events_go_proto",
+        "//pkg/eventchannel",
+        "//pkg/log",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/usage",
+    ],
+)
+
+proto_library(
+    name = "memory_events_proto",
+    srcs = ["memory_events.proto"],
+    visibility = ["//visibility:public"],
+)
+
+go_proto_library(
+    name = "memory_events_go_proto",
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/memevent/memory_events_go_proto",
+    proto = ":memory_events_proto",
+    visibility = ["//visibility:public"],
+)
diff --git a/pkg/sentry/kernel/memevent/memory_events.go b/pkg/sentry/kernel/memevent/memory_events.go
new file mode 100644
index 000000000..ecc9151de
--- /dev/null
+++ b/pkg/sentry/kernel/memevent/memory_events.go
@@ -0,0 +1,98 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package memevent implements the memory usage events controller, which
+// periodically emits events via the eventchannel.
+package memevent
+
+import (
+	"sync"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/eventchannel"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/memevent/memory_events_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+)
+
+// MemoryEvents describes the configuration for the global memory event emitter.
+type MemoryEvents struct {
+	k *kernel.Kernel
+
+	// The period is how often to emit an event. The memory events goroutine
+	// will ensure a minimum of one event is emitted per this period, regardless
+	// how of much memory usage has changed.
+	period time.Duration
+
+	// Writing to this channel indicates the memory goroutine should stop.
+	stop chan struct{}
+
+	// done is used to signal when the memory event goroutine has exited.
+	done sync.WaitGroup
+}
+
+// New creates a new MemoryEvents.
+func New(k *kernel.Kernel, period time.Duration) *MemoryEvents {
+	return &MemoryEvents{
+		k:      k,
+		period: period,
+		stop:   make(chan struct{}),
+	}
+}
+
+// Stop stops the memory usage events emitter goroutine. Stop must not be called
+// concurrently with Start and may only be called once.
+func (m *MemoryEvents) Stop() {
+	close(m.stop)
+	m.done.Wait()
+}
+
+// Start starts the memory usage events emitter goroutine. Start must not be
+// called concurrently with Stop and may only be called once.
+func (m *MemoryEvents) Start() {
+	if m.period == 0 {
+		return
+	}
+	go m.run() // S/R-SAFE: doesn't interact with saved state.
+}
+
+func (m *MemoryEvents) run() {
+	m.done.Add(1)
+
+	ticker := time.NewTicker(m.period)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-m.stop:
+			m.done.Done()
+			return
+		case <-ticker.C:
+			m.emit()
+		}
+	}
+}
+
+func (m *MemoryEvents) emit() {
+	totalPlatform, err := m.k.Platform.Memory().TotalUsage()
+	if err != nil {
+		log.Warningf("Failed to fetch memory usage for memory events: %v", err)
+		return
+	}
+	snapshot, _ := usage.MemoryAccounting.Copy()
+	total := totalPlatform + snapshot.Mapped
+
+	eventchannel.Emit(&pb.MemoryUsageEvent{Total: total})
+}
diff --git a/pkg/sentry/kernel/memevent/memory_events.proto b/pkg/sentry/kernel/memevent/memory_events.proto
new file mode 100644
index 000000000..e6e0bd628
--- /dev/null
+++ b/pkg/sentry/kernel/memevent/memory_events.proto
@@ -0,0 +1,25 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package gvisor;
+
+// MemoryUsageEvent describes the memory usage of the sandbox at a single
+// instant in time. These messages are emitted periodically on the eventchannel.
+message MemoryUsageEvent {
+  // The total memory usage of the sandboxed application in bytes, calculated
+  // using the 'fast' method.
+  uint64 total = 1;
+}
diff --git a/pkg/sentry/kernel/pending_signals.go b/pkg/sentry/kernel/pending_signals.go
new file mode 100644
index 000000000..d8701f47a
--- /dev/null
+++ b/pkg/sentry/kernel/pending_signals.go
@@ -0,0 +1,126 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+)
+
+const (
+	// stdSignalCap is the maximum number of instances of a given standard
+	// signal that may be pending. ("[If] multiple instances of a standard
+	// signal are delivered while that signal is currently blocked, then only
+	// one instance is queued.") - signal(7)
+	stdSignalCap = 1
+
+	// rtSignalCap is the maximum number of instances of a given realtime
+	// signal that may be pending.
+	//
+	// TODO: In Linux, the minimum signal queue size is
+	// RLIMIT_SIGPENDING, which is by default max_threads/2.
+	rtSignalCap = 32
+)
+
+// pendingSignals holds a collection of pending signals. The zero value of
+// pendingSignals is a valid empty collection. pendingSignals is thread-unsafe;
+// users must provide synchronization.
+type pendingSignals struct {
+	// signals contains all pending signals.
+	//
+	// Note that signals is zero-indexed, but signal 1 is the first valid
+	// signal, so signals[0] contains signals with signo 1 etc. This offset is
+	// usually handled by using Signal.index().
+	signals [linux.SignalMaximum]pendingSignalQueue
+
+	// Bit i of pendingSet is set iff there is at least one signal with signo
+	// i+1 pending.
+	pendingSet linux.SignalSet
+}
+
+// pendingSignalQueue holds a pendingSignalList for a single signal number.
+type pendingSignalQueue struct {
+	pendingSignalList
+	length int
+}
+
+type pendingSignal struct {
+	// pendingSignalEntry links into a pendingSignalList.
+	pendingSignalEntry
+	*arch.SignalInfo
+}
+
+// enqueue enqueues the given signal. enqueue returns true on success and false
+// on failure (if the given signal's queue is full).
+//
+// Preconditions: info represents a valid signal.
+func (p *pendingSignals) enqueue(info *arch.SignalInfo) bool {
+	sig := linux.Signal(info.Signo)
+	q := &p.signals[sig.Index()]
+	if sig.IsStandard() {
+		if q.length >= stdSignalCap {
+			return false
+		}
+	} else if q.length >= rtSignalCap {
+		return false
+	}
+	q.pendingSignalList.PushBack(&pendingSignal{SignalInfo: info})
+	q.length++
+	p.pendingSet |= linux.SignalSetOf(sig)
+	return true
+}
+
+// dequeue dequeues and returns any pending signal not masked by mask. If no
+// unmasked signals are pending, dequeue returns nil.
+func (p *pendingSignals) dequeue(mask linux.SignalSet) *arch.SignalInfo {
+	// "Real-time signals are delivered in a guaranteed order. Multiple
+	// real-time signals of the same type are delivered in the order they were
+	// sent. If different real-time signals are sent to a process, they are
+	// delivered starting with the lowest-numbered signal. (I.e., low-numbered
+	// signals have highest priority.) By contrast, if multiple standard
+	// signals are pending for a process, the order in which they are delivered
+	// is unspecified. If both standard and real-time signals are pending for a
+	// process, POSIX leaves it unspecified which is delivered first. Linux,
+	// like many other implementations, gives priority to standard signals in
+	// this case." - signal(7)
+	lowestPendingUnblockedBit := bits.TrailingZeros64(uint64(p.pendingSet &^ mask))
+	if lowestPendingUnblockedBit >= linux.SignalMaximum {
+		return nil
+	}
+	return p.dequeueSpecific(linux.Signal(lowestPendingUnblockedBit + 1))
+}
+
+func (p *pendingSignals) dequeueSpecific(sig linux.Signal) *arch.SignalInfo {
+	q := &p.signals[sig.Index()]
+	ps := q.pendingSignalList.Front()
+	if ps == nil {
+		return nil
+	}
+	q.pendingSignalList.Remove(ps)
+	q.length--
+	if q.length == 0 {
+		p.pendingSet &^= linux.SignalSetOf(sig)
+	}
+	return ps.SignalInfo
+}
+
+// discardSpecific causes all pending signals with number sig to be discarded.
+func (p *pendingSignals) discardSpecific(sig linux.Signal) {
+	q := &p.signals[sig.Index()]
+	q.pendingSignalList.Reset()
+	q.length = 0
+	p.pendingSet &^= linux.SignalSetOf(sig)
+}
diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD
new file mode 100644
index 000000000..ca9825f9d
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/BUILD
@@ -0,0 +1,68 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "pipe_state",
+    srcs = [
+        "buffers.go",
+        "node.go",
+        "pipe.go",
+        "reader.go",
+        "reader_writer.go",
+        "writer.go",
+    ],
+    out = "pipe_state.go",
+    package = "pipe",
+)
+
+go_library(
+    name = "pipe",
+    srcs = [
+        "buffers.go",
+        "device.go",
+        "node.go",
+        "pipe.go",
+        "pipe_state.go",
+        "reader.go",
+        "reader_writer.go",
+        "writer.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/pipe",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/amutex",
+        "//pkg/ilist",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
+
+go_test(
+    name = "pipe_test",
+    size = "small",
+    srcs = [
+        "node_test.go",
+        "pipe_test.go",
+    ],
+    embed = [":pipe"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/usermem",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/kernel/pipe/buffers.go b/pkg/sentry/kernel/pipe/buffers.go
new file mode 100644
index 000000000..f300537c5
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/buffers.go
@@ -0,0 +1,50 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+)
+
+// Buffer encapsulates a queueable byte buffer that can
+// easily be truncated.  It is designed only for use with pipes.
+type Buffer struct {
+	ilist.Entry
+	data []byte
+}
+
+// newBuffer initializes a Buffer.
+func newBuffer(buf []byte) *Buffer {
+	return &Buffer{data: buf}
+}
+
+// bytes returns the bytes contained in the buffer.
+func (b *Buffer) bytes() []byte {
+	return b.data
+}
+
+// size returns the number of bytes contained in the buffer.
+func (b *Buffer) size() int {
+	return len(b.data)
+}
+
+// truncate removes the first n bytes from the buffer.
+func (b *Buffer) truncate(n int) int {
+	if n > len(b.data) {
+		panic("Trying to truncate past end of array.")
+	}
+	b.data = b.data[n:]
+	return len(b.data)
+}
diff --git a/pkg/sentry/kernel/pipe/device.go b/pkg/sentry/kernel/pipe/device.go
new file mode 100644
index 000000000..8d383577a
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// pipeDevice is used for all pipe files.
+var pipeDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/kernel/pipe/node.go b/pkg/sentry/kernel/pipe/node.go
new file mode 100644
index 000000000..5b47427ef
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/node.go
@@ -0,0 +1,175 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/amutex"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// inodeOperations wraps fs.InodeOperations operations with common pipe opening semantics.
+type inodeOperations struct {
+	fs.InodeOperations
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// p is the underlying Pipe object representing this fifo.
+	p *Pipe
+
+	// Channels for synchronizing the creation of new readers and writers of
+	// this fifo. See waitFor and newHandleLocked.
+	//
+	// These are not saved/restored because all waiters are unblocked on save,
+	// and either automatically restart (via ERESTARTSYS) or return EINTR on
+	// resume. On restarts via ERESTARTSYS, the appropriate channel will be
+	// recreated.
+	rWakeup chan struct{} `state:"nosave"`
+	wWakeup chan struct{} `state:"nosave"`
+}
+
+// NewInodeOperations creates a new pipe fs.InodeOperations.
+func NewInodeOperations(base fs.InodeOperations, p *Pipe) fs.InodeOperations {
+	return &inodeOperations{
+		InodeOperations: base,
+		p:               p,
+	}
+}
+
+// GetFile implements fs.InodeOperations.GetFile. Named pipes have special blocking
+// semantics during open:
+//
+// "Normally, opening the FIFO blocks until the other end is opened also. A
+// process can open a FIFO in nonblocking mode. In this case, opening for
+// read-only will succeed even if no-one has opened on the write side yet,
+// opening for write-only will fail with ENXIO (no such device or address)
+// unless the other end has already been opened. Under Linux, opening a FIFO
+// for read and write will succeed both in blocking and nonblocking mode. POSIX
+// leaves this behavior undefined. This can be used to open a FIFO for writing
+// while there are no readers available." - fifo(7)
+func (i *inodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) {
+	i.mu.Lock()
+	defer i.mu.Unlock()
+
+	switch {
+	case flags.Read && !flags.Write: // O_RDONLY.
+		r := i.p.ROpen(ctx)
+		i.newHandleLocked(&i.rWakeup)
+
+		if i.p.isNamed && !flags.NonBlocking && !i.p.HasWriters() {
+			if !i.waitFor(&i.wWakeup, ctx) {
+				r.DecRef()
+				return nil, syserror.ErrInterrupted
+			}
+		}
+
+		// By now, either we're doing a nonblocking open or we have a writer. On
+		// a nonblocking read-only open, the open succeeds even if no-one has
+		// opened the write side yet.
+		return r, nil
+
+	case flags.Write && !flags.Read: // O_WRONLY.
+		w := i.p.WOpen(ctx)
+		i.newHandleLocked(&i.wWakeup)
+
+		if i.p.isNamed && !i.p.HasReaders() {
+			// On a nonblocking, write-only open, the open fails with ENXIO if the
+			// read side isn't open yet.
+			if flags.NonBlocking {
+				w.DecRef()
+				return nil, syserror.ENXIO
+			}
+
+			if !i.waitFor(&i.rWakeup, ctx) {
+				w.DecRef()
+				return nil, syserror.ErrInterrupted
+			}
+		}
+		return w, nil
+
+	case flags.Read && flags.Write: // O_RDWR.
+		// Pipes opened for read-write always succeeds without blocking.
+		rw := i.p.RWOpen(ctx)
+		i.newHandleLocked(&i.rWakeup)
+		i.newHandleLocked(&i.wWakeup)
+		return rw, nil
+
+	default:
+		return nil, syserror.EINVAL
+	}
+}
+
+// waitFor blocks until the underlying pipe has at least one reader/writer is
+// announced via 'wakeupChan', or until 'sleeper' is cancelled. Any call to this
+// function will block for either readers or writers, depending on where
+// 'wakeupChan' points.
+//
+// f.mu must be held by the caller. waitFor returns with f.mu held, but it will
+// drop f.mu before blocking for any reader/writers.
+func (i *inodeOperations) waitFor(wakeupChan *chan struct{}, sleeper amutex.Sleeper) bool {
+	// Ideally this function would simply use a condition variable. However, the
+	// wait needs to be interruptible via 'sleeper', so we must sychronize via a
+	// channel. The synchronization below relies on the fact that closing a
+	// channel unblocks all receives on the channel.
+
+	// Does an appropriate wakeup channel already exist? If not, create a new
+	// one. This is all done under f.mu to avoid races.
+	if *wakeupChan == nil {
+		*wakeupChan = make(chan struct{})
+	}
+
+	// Grab a local reference to the wakeup channel since it may disappear as
+	// soon as we drop f.mu.
+	wakeup := *wakeupChan
+
+	// Drop the lock and prepare to sleep.
+	i.mu.Unlock()
+	cancel := sleeper.SleepStart()
+
+	// Wait for either a new reader/write to be signalled via 'wakeup', or
+	// for the sleep to be cancelled.
+	select {
+	case <-wakeup:
+		sleeper.SleepFinish(true)
+	case <-cancel:
+		sleeper.SleepFinish(false)
+	}
+
+	// Take the lock and check if we were woken. If we were woken and
+	// interrupted, the former takes priority.
+	i.mu.Lock()
+	select {
+	case <-wakeup:
+		return true
+	default:
+		return false
+	}
+}
+
+// newHandleLocked signals a new pipe reader or writer depending on where
+// 'wakeupChan' points. This unblocks any corresponding reader or writer
+// waiting for the other end of the channel to be opened, see Fifo.waitFor.
+//
+// i.mu must be held.
+func (*inodeOperations) newHandleLocked(wakeupChan *chan struct{}) {
+	if *wakeupChan != nil {
+		close(*wakeupChan)
+		*wakeupChan = nil
+	}
+}
diff --git a/pkg/sentry/kernel/pipe/node_test.go b/pkg/sentry/kernel/pipe/node_test.go
new file mode 100644
index 000000000..cc1ebf4f6
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/node_test.go
@@ -0,0 +1,308 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"testing"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+type sleeper struct {
+	context.Context
+	ch chan struct{}
+}
+
+func newSleeperContext(t *testing.T) context.Context {
+	return &sleeper{
+		Context: contexttest.Context(t),
+		ch:      make(chan struct{}),
+	}
+}
+
+func (s *sleeper) SleepStart() <-chan struct{} {
+	return s.ch
+}
+
+func (s *sleeper) SleepFinish(bool) {
+}
+
+func (s *sleeper) Cancel() {
+	s.ch <- struct{}{}
+}
+
+type openResult struct {
+	*fs.File
+	error
+}
+
+func testOpenOrDie(ctx context.Context, t *testing.T, n fs.InodeOperations, flags fs.FileFlags, doneChan chan<- struct{}) (*fs.File, error) {
+	file, err := n.GetFile(ctx, nil, flags)
+	if err != nil {
+		t.Fatalf("open with flags %+v failed: %v", flags, err)
+	}
+	if doneChan != nil {
+		doneChan <- struct{}{}
+	}
+	return file, err
+}
+
+func testOpen(ctx context.Context, t *testing.T, n fs.InodeOperations, flags fs.FileFlags, resChan chan<- openResult) (*fs.File, error) {
+	file, err := n.GetFile(ctx, nil, flags)
+	if resChan != nil {
+		resChan <- openResult{file, err}
+	}
+	return file, err
+}
+
+func newNamedPipe(t *testing.T) *Pipe {
+	return NewPipe(contexttest.Context(t), true, DefaultPipeSize, usermem.PageSize)
+}
+
+func newAnonPipe(t *testing.T) *Pipe {
+	return NewPipe(contexttest.Context(t), false, DefaultPipeSize, usermem.PageSize)
+}
+
+// assertRecvBlocks ensures that a recv attempt on c blocks for at least
+// blockDuration. This is useful for checking that a goroutine that is supposed
+// to be executing a blocking operation is actually blocking.
+func assertRecvBlocks(t *testing.T, c <-chan struct{}, blockDuration time.Duration, failMsg string) {
+	select {
+	case <-c:
+		t.Fatalf(failMsg)
+	case <-time.After(blockDuration):
+		// Ok, blocked for the required duration.
+	}
+}
+
+func TestReadOpenBlocksForWriteOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone)
+
+	// Verify that the open for read is blocking.
+	assertRecvBlocks(t, rDone, time.Millisecond*100,
+		"open for read not blocking with no writers")
+
+	wDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+
+	<-wDone
+	<-rDone
+}
+
+func TestWriteOpenBlocksForReadOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	wDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+
+	// Verify that the open for write is blocking
+	assertRecvBlocks(t, wDone, time.Millisecond*100,
+		"open for write not blocking with no readers")
+
+	rDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone)
+
+	<-rDone
+	<-wDone
+}
+
+func TestMultipleWriteOpenDoesntCountAsReadOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rDone1 := make(chan struct{})
+	rDone2 := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone1)
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone2)
+
+	assertRecvBlocks(t, rDone1, time.Millisecond*100,
+		"open for read didn't block with no writers")
+	assertRecvBlocks(t, rDone2, time.Millisecond*100,
+		"open for read didn't block with no writers")
+
+	wDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+
+	<-wDone
+	<-rDone2
+	<-rDone1
+}
+
+func TestClosedReaderBlocksWriteOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rFile, _ := testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true, NonBlocking: true}, nil)
+	rFile.DecRef()
+
+	wDone := make(chan struct{})
+	// This open for write should block because the reader is now gone.
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+	assertRecvBlocks(t, wDone, time.Millisecond*100,
+		"open for write didn't block with no concurrent readers")
+
+	// Open for read again. This should unblock the open for write.
+	rDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone)
+
+	<-rDone
+	<-wDone
+}
+
+func TestReadWriteOpenNeverBlocks(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rwDone := make(chan struct{})
+	// Open for read-write never wait for a reader or writer, even if the
+	// nonblocking flag is not set.
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true, Write: true, NonBlocking: false}, rwDone)
+	<-rwDone
+}
+
+func TestReadWriteOpenUnblocksReadOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone)
+
+	rwDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true, Write: true}, rwDone)
+
+	<-rwDone
+	<-rDone
+}
+
+func TestReadWriteOpenUnblocksWriteOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	wDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+
+	rwDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true, Write: true}, rwDone)
+
+	<-rwDone
+	<-wDone
+}
+
+func TestBlockedOpenIsCancellable(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	done := make(chan openResult)
+	go testOpen(ctx, t, f, fs.FileFlags{Read: true}, done)
+	select {
+	case <-done:
+		t.Fatalf("open for read didn't block with no writers")
+	case <-time.After(time.Millisecond * 100):
+		// Ok.
+	}
+
+	ctx.(*sleeper).Cancel()
+	// If the cancel on the sleeper didn't work, the open for read would never
+	// return.
+	res := <-done
+	if res.error != syserror.ErrInterrupted {
+		t.Fatalf("Cancellation didn't cause GetFile to return fs.ErrInterrupted, got %v.",
+			res.error)
+	}
+}
+
+func TestNonblockingReadOpenNoWriters(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Read: true, NonBlocking: true}, nil); err != nil {
+		t.Fatalf("Nonblocking open for read failed with error %v.", err)
+	}
+}
+
+func TestNonblockingWriteOpenNoReaders(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Write: true, NonBlocking: true}, nil); err != syserror.ENXIO {
+		t.Fatalf("Nonblocking open for write failed unexpected error %v.", err)
+	}
+}
+
+func TestNonBlockingReadOpenWithWriter(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	wDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Write: true}, wDone)
+
+	// Open for write blocks since there are no readers yet.
+	assertRecvBlocks(t, wDone, time.Millisecond*100,
+		"Open for write didn't block with no reader.")
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Read: true, NonBlocking: true}, nil); err != nil {
+		t.Fatalf("Nonblocking open for read failed with error %v.", err)
+	}
+
+	// Open for write should now be unblocked.
+	<-wDone
+}
+
+func TestNonBlockingWriteOpenWithReader(t *testing.T) {
+	f := NewInodeOperations(nil, newNamedPipe(t))
+	ctx := newSleeperContext(t)
+
+	rDone := make(chan struct{})
+	go testOpenOrDie(ctx, t, f, fs.FileFlags{Read: true}, rDone)
+
+	// Open for write blocked, since no reader yet.
+	assertRecvBlocks(t, rDone, time.Millisecond*100,
+		"Open for reader didn't block with no writer.")
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Write: true, NonBlocking: true}, nil); err != nil {
+		t.Fatalf("Nonblocking open for write failed with error %v.", err)
+	}
+
+	// Open for write should now be unblocked.
+	<-rDone
+}
+
+func TestAnonReadOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newAnonPipe(t))
+	ctx := newSleeperContext(t)
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Read: true}, nil); err != nil {
+		t.Fatalf("open anon pipe for read failed: %v", err)
+	}
+}
+
+func TestAnonWriteOpen(t *testing.T) {
+	f := NewInodeOperations(nil, newAnonPipe(t))
+	ctx := newSleeperContext(t)
+
+	if _, err := testOpen(ctx, t, f, fs.FileFlags{Write: true}, nil); err != nil {
+		t.Fatalf("open anon pipe for write failed: %v", err)
+	}
+}
diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go
new file mode 100644
index 000000000..1656c6ff3
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/pipe.go
@@ -0,0 +1,335 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package pipe provides an in-memory implementation of a unidirectional
+// pipe.
+//
+// The goal of this pipe is to emulate the pipe syscall in all of its
+// edge cases and guarantees of atomic IO.
+package pipe
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/ilist"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// DefaultPipeSize is the system-wide default size of a pipe in bytes.
+const DefaultPipeSize = 65536
+
+// Pipe is an encapsulation of a platform-independent pipe.
+// It manages a buffered byte queue shared between a reader/writer
+// pair.
+type Pipe struct {
+	waiter.Queue `state:"nosave"`
+
+	// Whether this is a named or anonymous pipe.
+	isNamed bool
+
+	// The dirent backing this pipe. Shared by all readers and writers.
+	dirent *fs.Dirent
+
+	// The buffered byte queue.
+	data ilist.List
+
+	// Max size of the pipe in bytes.  When this max has been reached,
+	// writers will get EWOULDBLOCK.
+	max int
+
+	// Current size of the pipe in bytes.
+	size int
+
+	// Max number of bytes the pipe can guarantee to read or write
+	// atomically.
+	atomicIOBytes int
+
+	// The number of active readers for this pipe. Load/store atomically.
+	readers int32
+
+	// The number of active writes for this pipe. Load/store atomically.
+	writers int32
+
+	// This flag indicates if this pipe ever had a writer. Note that this does
+	// not necessarily indicate there is *currently* a writer, just that there
+	// has been a writer at some point since the pipe was created.
+	//
+	// Protected by mu.
+	hadWriter bool
+
+	// Lock protecting all pipe internal state.
+	mu sync.Mutex `state:"nosave"`
+}
+
+// NewPipe initializes and returns a pipe. A pipe created by this function is
+// persistent, and will remain valid even without any open fds to it. Named
+// pipes for mknod(2) are created via this function. Note that the
+// implementation of blocking semantics for opening the read and write ends of a
+// named pipe are left to filesystems.
+func NewPipe(ctx context.Context, isNamed bool, sizeBytes, atomicIOBytes int) *Pipe {
+	p := &Pipe{
+		isNamed:       isNamed,
+		max:           sizeBytes,
+		atomicIOBytes: atomicIOBytes,
+	}
+
+	// Build the fs.Dirent of this pipe, shared by all fs.Files associated
+	// with this pipe.
+	ino := pipeDevice.NextIno()
+	base := fsutil.NewSimpleInodeOperations(fsutil.InodeSimpleAttributes{
+		FSType: linux.PIPEFS_MAGIC,
+		UAttr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: fs.FileOwnerFromContext(ctx),
+			Perms: fs.FilePermissions{
+				User: fs.PermMask{Read: true, Write: true},
+			},
+			Links: 1,
+		}),
+	})
+	sattr := fs.StableAttr{
+		Type:      fs.Pipe,
+		DeviceID:  pipeDevice.DeviceID(),
+		InodeID:   ino,
+		BlockSize: int64(atomicIOBytes),
+	}
+	// There is no real filesystem backing this pipe, so we pass in a nil
+	// Filesystem.
+	sb := fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{})
+	p.dirent = fs.NewDirent(fs.NewInode(NewInodeOperations(base, p), sb, sattr), fmt.Sprintf("pipe:[%d]", ino))
+
+	return p
+}
+
+// NewConnectedPipe initializes a pipe and returns a pair of objects (which
+// implement kio.File) representing the read and write ends of the pipe. A pipe
+// created by this function becomes invalid as soon as either the read or write
+// end is closed, and errors on subsequent operations on either end. Pipes
+// for pipe(2) and pipe2(2) are generally created this way.
+func NewConnectedPipe(ctx context.Context, sizeBytes int, atomicIOBytes int) (*fs.File, *fs.File) {
+	p := NewPipe(ctx, false /* isNamed */, sizeBytes, atomicIOBytes)
+	return p.ROpen(ctx), p.WOpen(ctx)
+}
+
+// ROpen opens the pipe for reading.
+func (p *Pipe) ROpen(ctx context.Context) *fs.File {
+	p.rOpen()
+	return fs.NewFile(ctx, p.dirent, fs.FileFlags{Read: true}, &Reader{
+		ReaderWriter: ReaderWriter{Pipe: p},
+	})
+}
+
+// WOpen opens the pipe for writing.
+func (p *Pipe) WOpen(ctx context.Context) *fs.File {
+	p.wOpen()
+	return fs.NewFile(ctx, p.dirent, fs.FileFlags{Write: true}, &Writer{
+		ReaderWriter: ReaderWriter{Pipe: p},
+	})
+}
+
+// RWOpen opens the pipe for both reading and writing.
+func (p *Pipe) RWOpen(ctx context.Context) *fs.File {
+	p.rOpen()
+	p.wOpen()
+	return fs.NewFile(ctx, p.dirent, fs.FileFlags{Read: true, Write: true}, &ReaderWriter{
+		Pipe: p,
+	})
+}
+
+// read reads data from the pipe into dst and returns the number of bytes
+// read, or returns ErrWouldBlock if the pipe is empty.
+func (p *Pipe) read(ctx context.Context, dst usermem.IOSequence) (int64, error) {
+	if !p.HasReaders() {
+		return 0, syscall.EBADF
+	}
+
+	// Don't block for a zero-length read even if the pipe is empty.
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	// If there is nothing to read at the moment but there is a writer, tell the
+	// caller to block.
+	if p.size == 0 {
+		if !p.HasWriters() {
+			// There are no writers, return EOF.
+			return 0, nil
+		}
+		return 0, syserror.ErrWouldBlock
+	}
+	var n int64
+	for b := p.data.Front(); b != nil; b = p.data.Front() {
+		buffer := b.(*Buffer)
+		n0, err := dst.CopyOut(ctx, buffer.bytes())
+		n += int64(n0)
+		p.size -= n0
+		if buffer.truncate(n0) == 0 {
+			p.data.Remove(b)
+		}
+		dst = dst.DropFirst(n0)
+		if dst.NumBytes() == 0 || err != nil {
+			return n, err
+		}
+	}
+	return n, nil
+}
+
+// write writes data from sv into the pipe and returns the number of bytes
+// written. If no bytes are written because the pipe is full (or has less than
+// atomicIOBytes free capacity), write returns ErrWouldBlock.
+func (p *Pipe) write(ctx context.Context, src usermem.IOSequence) (int64, error) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+
+	if !p.HasWriters() {
+		return 0, syscall.EBADF
+	}
+	if !p.HasReaders() {
+		return 0, syscall.EPIPE
+	}
+
+	// POSIX requires that a write smaller than atomicIOBytes (PIPE_BUF) be
+	// atomic, but requires no atomicity for writes larger than this. However,
+	// Linux appears to provide stronger semantics than this in practice:
+	// unmerged writes are done one PAGE_SIZE buffer at a time, so for larger
+	// writes, the writing of each PIPE_BUF-sized chunk is atomic. We implement
+	// this by writing at most atomicIOBytes at a time if we can't service the
+	// write in its entirety.
+	canWrite := src.NumBytes()
+	if canWrite > int64(p.max-p.size) {
+		if p.max-p.size >= p.atomicIOBytes {
+			canWrite = int64(p.atomicIOBytes)
+		} else {
+			return 0, syserror.ErrWouldBlock
+		}
+	}
+
+	// Copy data from user memory into a pipe-owned buffer.
+	buf := make([]byte, canWrite)
+	n, err := src.CopyIn(ctx, buf)
+	if n > 0 {
+		p.data.PushBack(newBuffer(buf[:n]))
+		p.size += n
+	}
+	if int64(n) < src.NumBytes() && err == nil {
+		// Partial write due to full pipe.
+		err = syserror.ErrWouldBlock
+	}
+	return int64(n), err
+}
+
+// rOpen signals a new reader of the pipe.
+func (p *Pipe) rOpen() {
+	atomic.AddInt32(&p.readers, 1)
+}
+
+// wOpen signals a new writer of the pipe.
+func (p *Pipe) wOpen() {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	p.hadWriter = true
+	atomic.AddInt32(&p.writers, 1)
+}
+
+// rClose signals that a reader has closed their end of the pipe.
+func (p *Pipe) rClose() {
+	newReaders := atomic.AddInt32(&p.readers, -1)
+	if newReaders < 0 {
+		panic(fmt.Sprintf("Refcounting bug, pipe has negative readers: %v", newReaders))
+	}
+}
+
+// wClose signals that a writer has closed their end of the pipe.
+func (p *Pipe) wClose() {
+	newWriters := atomic.AddInt32(&p.writers, -1)
+	if newWriters < 0 {
+		panic(fmt.Sprintf("Refcounting bug, pipe has negative writers: %v.", newWriters))
+	}
+}
+
+// HasReaders returns whether the pipe has any active readers.
+func (p *Pipe) HasReaders() bool {
+	return atomic.LoadInt32(&p.readers) > 0
+}
+
+// HasWriters returns whether the pipe has any active writers.
+func (p *Pipe) HasWriters() bool {
+	return atomic.LoadInt32(&p.writers) > 0
+}
+
+func (p *Pipe) rReadinessLocked() waiter.EventMask {
+	ready := waiter.EventMask(0)
+	if p.HasReaders() && p.data.Front() != nil {
+		ready |= waiter.EventIn
+	}
+	if !p.HasWriters() && p.hadWriter {
+		// POLLHUP must be supressed until the pipe has had at least one writer
+		// at some point. Otherwise a reader thread may poll and immediately get
+		// a POLLHUP before the writer ever opens the pipe, which the reader may
+		// interpret as the writer opening then closing the pipe.
+		ready |= waiter.EventHUp
+	}
+	return ready
+}
+
+// rReadiness returns a mask that states whether the read end of the pipe is
+// ready for reading.
+func (p *Pipe) rReadiness() waiter.EventMask {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.rReadinessLocked()
+}
+
+func (p *Pipe) wReadinessLocked() waiter.EventMask {
+	ready := waiter.EventMask(0)
+	if p.HasWriters() && p.size < p.max {
+		ready |= waiter.EventOut
+	}
+	if !p.HasReaders() {
+		ready |= waiter.EventErr
+	}
+	return ready
+}
+
+// wReadiness returns a mask that states whether the write end of the pipe
+// is ready for writing.
+func (p *Pipe) wReadiness() waiter.EventMask {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.wReadinessLocked()
+}
+
+// rwReadiness returns a mask that states whether a read-write handle to the
+// pipe is ready for IO.
+func (p *Pipe) rwReadiness() waiter.EventMask {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.rReadinessLocked() | p.wReadinessLocked()
+}
+
+func (p *Pipe) queuedSize() int {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	return p.size
+}
diff --git a/pkg/sentry/kernel/pipe/pipe_test.go b/pkg/sentry/kernel/pipe/pipe_test.go
new file mode 100644
index 000000000..49ef8c8ac
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/pipe_test.go
@@ -0,0 +1,138 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"bytes"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+func TestPipeRW(t *testing.T) {
+	ctx := contexttest.Context(t)
+	r, w := NewConnectedPipe(ctx, 65536, 4096)
+	defer r.DecRef()
+	defer w.DecRef()
+
+	msg := []byte("here's some bytes")
+	wantN := int64(len(msg))
+	n, err := w.Writev(ctx, usermem.BytesIOSequence(msg))
+	if n != wantN || err != nil {
+		t.Fatalf("Writev: got (%d, %v), wanted (%d, nil)", n, err, wantN)
+	}
+
+	buf := make([]byte, len(msg))
+	n, err = r.Readv(ctx, usermem.BytesIOSequence(buf))
+	if n != wantN || err != nil || !bytes.Equal(buf, msg) {
+		t.Fatalf("Readv: got (%d, %v) %q, wanted (%d, nil) %q", n, err, buf, wantN, msg)
+	}
+}
+
+func TestPipeReadBlock(t *testing.T) {
+	ctx := contexttest.Context(t)
+	r, w := NewConnectedPipe(ctx, 65536, 4096)
+	defer r.DecRef()
+	defer w.DecRef()
+
+	n, err := r.Readv(ctx, usermem.BytesIOSequence(make([]byte, 1)))
+	if n != 0 || err != syserror.ErrWouldBlock {
+		t.Fatalf("Readv: got (%d, %v), wanted (0, %v)", n, err, syserror.ErrWouldBlock)
+	}
+}
+
+func TestPipeWriteBlock(t *testing.T) {
+	const atomicIOBytes = 2
+
+	ctx := contexttest.Context(t)
+	r, w := NewConnectedPipe(ctx, 10, atomicIOBytes)
+	defer r.DecRef()
+	defer w.DecRef()
+
+	msg := []byte("here's some bytes")
+	n, err := w.Writev(ctx, usermem.BytesIOSequence(msg))
+	if wantN, wantErr := int64(atomicIOBytes), syserror.ErrWouldBlock; n != wantN || err != wantErr {
+		t.Fatalf("Writev: got (%d, %v), wanted (%d, %v)", n, err, wantN, wantErr)
+	}
+}
+
+func TestPipeWriteUntilEnd(t *testing.T) {
+	const atomicIOBytes = 2
+
+	ctx := contexttest.Context(t)
+	r, w := NewConnectedPipe(ctx, atomicIOBytes, atomicIOBytes)
+	defer r.DecRef()
+	defer w.DecRef()
+
+	msg := []byte("here's some bytes")
+
+	wDone := make(chan struct{}, 0)
+	rDone := make(chan struct{}, 0)
+	defer func() {
+		// Signal the reader to stop and wait until it does so.
+		close(wDone)
+		<-rDone
+	}()
+
+	go func() {
+		defer close(rDone)
+		// Read from r until done is closed.
+		ctx := contexttest.Context(t)
+		buf := make([]byte, len(msg)+1)
+		dst := usermem.BytesIOSequence(buf)
+		e, ch := waiter.NewChannelEntry(nil)
+		r.EventRegister(&e, waiter.EventIn)
+		defer r.EventUnregister(&e)
+		for {
+			n, err := r.Readv(ctx, dst)
+			dst = dst.DropFirst64(n)
+			if err == syserror.ErrWouldBlock {
+				select {
+				case <-ch:
+					continue
+				case <-wDone:
+					// We expect to have 1 byte left in dst since len(buf) ==
+					// len(msg)+1.
+					if dst.NumBytes() != 1 || !bytes.Equal(buf[:len(msg)], msg) {
+						t.Errorf("Reader: got %q (%d bytes remaining), wanted %q", buf, dst.NumBytes(), msg)
+					}
+					return
+				}
+			}
+			if err != nil {
+				t.Fatalf("Readv: got unexpected error %v", err)
+			}
+		}
+	}()
+
+	src := usermem.BytesIOSequence(msg)
+	e, ch := waiter.NewChannelEntry(nil)
+	w.EventRegister(&e, waiter.EventOut)
+	defer w.EventUnregister(&e)
+	for src.NumBytes() != 0 {
+		n, err := w.Writev(ctx, src)
+		src = src.DropFirst64(n)
+		if err == syserror.ErrWouldBlock {
+			<-ch
+			continue
+		}
+		if err != nil {
+			t.Fatalf("Writev: got (%d, %v)", n, err)
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/pipe/reader.go b/pkg/sentry/kernel/pipe/reader.go
new file mode 100644
index 000000000..40d5e4943
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/reader.go
@@ -0,0 +1,37 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Reader satisfies the fs.FileOperations interface for read-only pipes.
+// Reader should be used with !fs.FileFlags.Write to reject writes.
+type Reader struct {
+	ReaderWriter
+}
+
+// Release implements fs.FileOperations.Release.
+func (r *Reader) Release() {
+	r.Pipe.rClose()
+	// Wake up writers.
+	r.Pipe.Notify(waiter.EventOut)
+}
+
+// Readiness returns the ready events in the underlying pipe.
+func (r *Reader) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return r.Pipe.rReadiness() & mask
+}
diff --git a/pkg/sentry/kernel/pipe/reader_writer.go b/pkg/sentry/kernel/pipe/reader_writer.go
new file mode 100644
index 000000000..dc642a3a6
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/reader_writer.go
@@ -0,0 +1,91 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"fmt"
+	"math"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// ReaderWriter satisfies the FileOperations interface and services both
+// read and write requests. This should only be used directly for named pipes.
+// pipe(2) and pipe2(2) only support unidirectional pipes and should use
+// either pipe.Reader or pipe.Writer.
+type ReaderWriter struct {
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	*Pipe
+}
+
+// Release implements fs.FileOperations.Release.
+func (rw *ReaderWriter) Release() {
+	rw.Pipe.rClose()
+	rw.Pipe.wClose()
+	// Wake up readers and writers.
+	rw.Pipe.Notify(waiter.EventIn | waiter.EventOut)
+}
+
+// Read implements fs.FileOperations.Read.
+func (rw *ReaderWriter) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	n, err := rw.Pipe.read(ctx, dst)
+	if n > 0 {
+		rw.Pipe.Notify(waiter.EventOut)
+	}
+	return n, err
+}
+
+// Write implements fs.FileOperations.Write.
+func (rw *ReaderWriter) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	n, err := rw.Pipe.write(ctx, src)
+	if n > 0 {
+		rw.Pipe.Notify(waiter.EventIn)
+	}
+	return n, err
+}
+
+// Readiness returns the ready events in the underlying pipe.
+func (rw *ReaderWriter) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return rw.Pipe.rwReadiness() & mask
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (rw *ReaderWriter) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// Switch on ioctl request.
+	switch int(args[1].Int()) {
+	case syscall.TIOCINQ:
+		v := rw.queuedSize()
+		if v > math.MaxInt32 {
+			panic(fmt.Sprintf("Impossibly large pipe queued size: %d", v))
+		}
+		// Copy result to user-space.
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	default:
+		return 0, syscall.ENOTTY
+	}
+}
diff --git a/pkg/sentry/kernel/pipe/writer.go b/pkg/sentry/kernel/pipe/writer.go
new file mode 100644
index 000000000..fd13008ac
--- /dev/null
+++ b/pkg/sentry/kernel/pipe/writer.go
@@ -0,0 +1,37 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pipe
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Writer satisfies the fs.FileOperations interface for write-only pipes.
+// Writer should be used with !fs.FileFlags.Read to reject reads.
+type Writer struct {
+	ReaderWriter
+}
+
+// Release implements fs.FileOperations.Release.
+func (w *Writer) Release() {
+	w.Pipe.wClose()
+	// Wake up readers.
+	w.Pipe.Notify(waiter.EventHUp)
+}
+
+// Readiness returns the ready events in the underlying pipe.
+func (w *Writer) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return w.Pipe.wReadiness() & mask
+}
diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go
new file mode 100644
index 000000000..20b1c4cd4
--- /dev/null
+++ b/pkg/sentry/kernel/ptrace.go
@@ -0,0 +1,1054 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// ptrace constants from Linux's include/uapi/linux/ptrace.h.
+const (
+	_PTRACE_EVENT_SECCOMP  = 7
+	PTRACE_SEIZE           = 0x4206
+	PTRACE_INTERRUPT       = 0x4207
+	PTRACE_LISTEN          = 0x4208
+	PTRACE_PEEKSIGINFO     = 0x4209
+	PTRACE_GETSIGMASK      = 0x420a
+	PTRACE_SETSIGMASK      = 0x420b
+	_PTRACE_O_EXITKILL     = 1 << 20
+	_PTRACE_O_TRACESECCOMP = 1 << _PTRACE_EVENT_SECCOMP
+)
+
+// ptraceOptions are the subset of options controlling a task's ptrace behavior
+// that are set by ptrace(PTRACE_SETOPTIONS).
+type ptraceOptions struct {
+	// ExitKill is true if the tracee should be sent SIGKILL when the tracer
+	// exits.
+	ExitKill bool
+
+	// If SysGood is true, set bit 7 in the signal number for
+	// syscall-entry-stop and syscall-exit-stop traps delivered to this task's
+	// tracer.
+	SysGood bool
+
+	// TraceClone is true if the tracer wants to receive PTRACE_EVENT_CLONE
+	// events.
+	TraceClone bool
+
+	// TraceExec is true if the tracer wants to receive PTRACE_EVENT_EXEC
+	// events.
+	TraceExec bool
+
+	// TraceExit is true if the tracer wants to receive PTRACE_EVENT_EXIT
+	// events.
+	TraceExit bool
+
+	// TraceFork is true if the tracer wants to receive PTRACE_EVENT_FORK
+	// events.
+	TraceFork bool
+
+	// TraceSeccomp is true if the tracer wants to receive PTRACE_EVENT_SECCOMP
+	// events.
+	TraceSeccomp bool
+
+	// TraceVfork is true if the tracer wants to receive PTRACE_EVENT_VFORK
+	// events.
+	TraceVfork bool
+
+	// TraceVforkDone is true if the tracer wants to receive
+	// PTRACE_EVENT_VFORK_DONE events.
+	TraceVforkDone bool
+}
+
+// ptraceSyscallMode controls the behavior of a ptraced task at syscall entry
+// and exit.
+type ptraceSyscallMode int
+
+const (
+	// ptraceSyscallNone indicates that the task has never ptrace-stopped, or
+	// that it was resumed from its last ptrace-stop by PTRACE_CONT or
+	// PTRACE_DETACH. The task's syscalls will not be intercepted.
+	ptraceSyscallNone ptraceSyscallMode = iota
+
+	// ptraceSyscallIntercept indicates that the task was resumed from its last
+	// ptrace-stop by PTRACE_SYSCALL. The next time the task enters or exits a
+	// syscall, a ptrace-stop will occur.
+	ptraceSyscallIntercept
+
+	// ptraceSyscallEmu indicates that the task was resumed from its last
+	// ptrace-stop by PTRACE_SYSEMU or PTRACE_SYSEMU_SINGLESTEP. The next time
+	// the task enters a syscall, the syscall will be skipped, and a
+	// ptrace-stop will occur.
+	ptraceSyscallEmu
+)
+
+// CanTrace checks that t is permitted to access target's state, as defined by
+// ptrace(2), subsection "Ptrace access mode checking". If attach is true, it
+// checks for access mode PTRACE_MODE_ATTACH; otherwise, it checks for access
+// mode PTRACE_MODE_READ.
+func (t *Task) CanTrace(target *Task, attach bool) bool {
+	// "1. If the calling thread and the target thread are in the same thread
+	// group, access is always allowed." - ptrace(2)
+	//
+	// Note: Strictly speaking, prior to 73af963f9f30 ("__ptrace_may_access()
+	// should not deny sub-threads", first released in Linux 3.12), the rule
+	// only applies if t and target are the same task. But, as that commit
+	// message puts it, "[any] security check is pointless when the tasks share
+	// the same ->mm."
+	if t.tg == target.tg {
+		return true
+	}
+
+	// """
+	// 2. If the access mode specifies PTRACE_MODE_FSCREDS (ED: snipped,
+	// doesn't exist until Linux 4.5).
+	//
+	// Otherwise, the access mode specifies PTRACE_MODE_REALCREDS, so use the
+	// caller's real UID and GID for the checks in the next step. (Most APIs
+	// that check the caller's UID and GID use the effective IDs. For
+	// historical reasons, the PTRACE_MODE_REALCREDS check uses the real IDs
+	// instead.)
+	//
+	// 3. Deny access if neither of the following is true:
+	//
+	// - The real, effective, and saved-set user IDs of the target match the
+	// caller's user ID, *and* the real, effective, and saved-set group IDs of
+	// the target match the caller's group ID.
+	//
+	// - The caller has the CAP_SYS_PTRACE capability in the user namespace of
+	// the target.
+	//
+	// 4. Deny access if the target process "dumpable" attribute has a value
+	// other than 1 (SUID_DUMP_USER; see the discussion of PR_SET_DUMPABLE in
+	// prctl(2)), and the caller does not have the CAP_SYS_PTRACE capability in
+	// the user namespace of the target process.
+	//
+	// 5. The kernel LSM security_ptrace_access_check() interface is invoked to
+	// see if ptrace access is permitted. The results depend on the LSM(s). The
+	// implementation of this interface in the commoncap LSM performs the
+	// following steps:
+	//
+	// a) If the access mode includes PTRACE_MODE_FSCREDS, then use the
+	// caller's effective capability set; otherwise (the access mode specifies
+	// PTRACE_MODE_REALCREDS, so) use the caller's permitted capability set.
+	//
+	// b) Deny access if neither of the following is true:
+	//
+	// - The caller and the target process are in the same user namespace, and
+	// the caller's capabilities are a proper superset of the target process's
+	// permitted capabilities.
+	//
+	// - The caller has the CAP_SYS_PTRACE capability in the target process's
+	// user namespace.
+	//
+	// Note that the commoncap LSM does not distinguish between
+	// PTRACE_MODE_READ and PTRACE_MODE_ATTACH. (ED: From earlier in this
+	// section: "the commoncap LSM ... is always invoked".)
+	// """
+	callerCreds := t.Credentials()
+	targetCreds := target.Credentials()
+	if callerCreds.HasCapabilityIn(linux.CAP_SYS_PTRACE, targetCreds.UserNamespace) {
+		return true
+	}
+	if cuid := callerCreds.RealKUID; cuid != targetCreds.RealKUID || cuid != targetCreds.EffectiveKUID || cuid != targetCreds.SavedKUID {
+		return false
+	}
+	if cgid := callerCreds.RealKGID; cgid != targetCreds.RealKGID || cgid != targetCreds.EffectiveKGID || cgid != targetCreds.SavedKGID {
+		return false
+	}
+	// TODO: dumpability check
+	if callerCreds.UserNamespace != targetCreds.UserNamespace {
+		return false
+	}
+	if targetCreds.PermittedCaps&^callerCreds.PermittedCaps != 0 {
+		return false
+	}
+	// TODO: Yama LSM
+	return true
+}
+
+// Tracer returns t's ptrace Tracer.
+func (t *Task) Tracer() *Task {
+	return t.ptraceTracer.Load().(*Task)
+}
+
+// hasTracer returns true if t has a ptrace tracer attached.
+func (t *Task) hasTracer() bool {
+	// This isn't just inlined into callers so that if Task.Tracer() turns out
+	// to be too expensive because of e.g. interface conversion, we can switch
+	// to having a separate atomic flag more easily.
+	return t.Tracer() != nil
+}
+
+// ptraceStop is a TaskStop placed on tasks in a ptrace-stop.
+type ptraceStop struct {
+	// If frozen is true, the stopped task's tracer is currently operating on
+	// it, so Task.Kill should not remove the stop.
+	frozen bool
+}
+
+// Killable implements TaskStop.Killable.
+func (s *ptraceStop) Killable() bool {
+	return !s.frozen
+}
+
+// beginPtraceStopLocked initiates an unfrozen ptrace-stop on t. If t has been
+// killed, the stop is skipped, and beginPtraceStopLocked returns false.
+//
+// beginPtraceStopLocked does not signal t's tracer or wake it if it is
+// waiting.
+//
+// Preconditions: The TaskSet mutex must be locked. The caller must be running
+// on the task goroutine.
+func (t *Task) beginPtraceStopLocked() bool {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	// This is analogous to Linux's kernel/signal.c:ptrace_stop() => ... =>
+	// kernel/sched/core.c:__schedule() => signal_pending_state() check, which
+	// is what prevents tasks from entering ptrace-stops after being killed.
+	// Note that if t was SIGKILLed and beingPtraceStopLocked is being called
+	// for PTRACE_EVENT_EXIT, the task will have dequeued the signal before
+	// entering the exit path, so t.killable() will no longer return true. This
+	// is consistent with Linux: "Bugs: ... A SIGKILL signal may still cause a
+	// PTRACE_EVENT_EXIT stop before actual signal death. This may be changed
+	// in the future; SIGKILL is meant to always immediately kill tasks even
+	// under ptrace. Last confirmed on Linux 3.13." - ptrace(2)
+	if t.killedLocked() {
+		return false
+	}
+	t.beginInternalStopLocked(&ptraceStop{})
+	return true
+}
+
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) ptraceTrapLocked(code int32) {
+	t.ptraceCode = code
+	t.ptraceSiginfo = &arch.SignalInfo{
+		Signo: int32(linux.SIGTRAP),
+		Code:  code,
+	}
+	t.ptraceSiginfo.SetPid(int32(t.tg.pidns.tids[t]))
+	t.ptraceSiginfo.SetUid(int32(t.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()))
+	if t.beginPtraceStopLocked() {
+		tracer := t.Tracer()
+		tracer.signalStop(t, arch.CLD_TRAPPED, int32(linux.SIGTRAP))
+		tracer.tg.eventQueue.Notify(EventTraceeStop)
+	}
+}
+
+// ptraceFreeze checks if t is in a ptraceStop. If so, it freezes the
+// ptraceStop, temporarily preventing it from being removed by a concurrent
+// Task.Kill, and returns true. Otherwise it returns false.
+//
+// Preconditions: The TaskSet mutex must be locked. The caller must be running
+// on the task goroutine of t's tracer.
+func (t *Task) ptraceFreeze() bool {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if t.stop == nil {
+		return false
+	}
+	s, ok := t.stop.(*ptraceStop)
+	if !ok {
+		return false
+	}
+	s.frozen = true
+	return true
+}
+
+// ptraceUnfreeze ends the effect of a previous successful call to
+// ptraceFreeze.
+//
+// Preconditions: t must be in a frozen ptraceStop.
+func (t *Task) ptraceUnfreeze() {
+	// t.tg.signalHandlers is stable because t is in a frozen ptrace-stop,
+	// preventing its thread group from completing execve.
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	// Do this even if the task has been killed to ensure a panic if t.stop is
+	// nil or not a ptraceStop.
+	t.stop.(*ptraceStop).frozen = false
+	if t.killedLocked() {
+		t.endInternalStopLocked()
+	}
+}
+
+// ptraceUnstop implements ptrace request PTRACE_CONT, PTRACE_SYSCALL,
+// PTRACE_SINGLESTEP, PTRACE_SYSEMU, or PTRACE_SYSEMU_SINGLESTEP depending on
+// mode and singlestep.
+//
+// Preconditions: t must be in a frozen ptrace stop.
+//
+// Postconditions: If ptraceUnstop returns nil, t will no longer be in a ptrace
+// stop.
+func (t *Task) ptraceUnstop(mode ptraceSyscallMode, singlestep bool, sig linux.Signal) error {
+	if sig != 0 && !sig.IsValid() {
+		return syserror.EIO
+	}
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	t.ptraceCode = int32(sig)
+	t.ptraceSyscallMode = mode
+	t.ptraceSinglestep = singlestep
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.endInternalStopLocked()
+	return nil
+}
+
+func (t *Task) ptraceTraceme() error {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	if t.hasTracer() {
+		return syserror.EPERM
+	}
+	if t.parent == nil {
+		// In Linux, only init can not have a parent, and init is assumed never
+		// to invoke PTRACE_TRACEME. In the sentry, TGID 1 is an arbitrary user
+		// application that may invoke PTRACE_TRACEME; having no parent can
+		// also occur if all tasks in the parent thread group have exited, and
+		// failed to find a living thread group to reparent to. The former case
+		// is treated as if TGID 1 has an exited parent in an invisible
+		// ancestor PID namespace that is an owner of the root user namespace
+		// (and consequently has CAP_SYS_PTRACE), and the latter case is a
+		// special form of the exited parent case below. In either case,
+		// returning nil here is correct.
+		return nil
+	}
+	if !t.parent.CanTrace(t, true) {
+		return syserror.EPERM
+	}
+	if t.parent.exitState != TaskExitNone {
+		// Fail silently, as if we were successfully attached but then
+		// immediately detached. This is consistent with Linux.
+		return nil
+	}
+	t.ptraceTracer.Store(t.parent)
+	t.parent.ptraceTracees[t] = struct{}{}
+	return nil
+}
+
+// ptraceAttach implements ptrace(PTRACE_ATTACH, target). t is the caller.
+func (t *Task) ptraceAttach(target *Task) error {
+	if t.tg == target.tg {
+		return syserror.EPERM
+	}
+	if !t.CanTrace(target, true) {
+		return syserror.EPERM
+	}
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	if target.hasTracer() {
+		return syserror.EPERM
+	}
+	// Attaching to zombies and dead tasks is not permitted; the exit
+	// notification logic relies on this. Linux allows attaching to PF_EXITING
+	// tasks, though.
+	if target.exitState >= TaskExitZombie {
+		return syserror.EPERM
+	}
+	target.ptraceTracer.Store(t)
+	t.ptraceTracees[target] = struct{}{}
+	target.tg.signalHandlers.mu.Lock()
+	target.sendSignalLocked(&arch.SignalInfo{
+		Signo: int32(linux.SIGSTOP),
+		Code:  arch.SignalInfoUser,
+	}, false /* group */)
+	// Undocumented Linux feature: If the tracee is already group-stopped (and
+	// consequently will not report the SIGSTOP just sent), force it to leave
+	// and re-enter the stop so that it will switch to a ptrace-stop.
+	if target.stop == (*groupStop)(nil) {
+		target.groupStopRequired = true
+		target.endInternalStopLocked()
+	}
+	target.tg.signalHandlers.mu.Unlock()
+	return nil
+}
+
+// ptraceDetach implements ptrace(PTRACE_DETACH, target, 0, sig). t is the
+// caller.
+//
+// Preconditions: target must be a tracee of t in a frozen ptrace stop.
+//
+// Postconditions: If ptraceDetach returns nil, target will no longer be in a
+// ptrace stop.
+func (t *Task) ptraceDetach(target *Task, sig linux.Signal) error {
+	if sig != 0 && !sig.IsValid() {
+		return syserror.EIO
+	}
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	target.ptraceCode = int32(sig)
+	target.forgetTracerLocked()
+	delete(t.ptraceTracees, target)
+	return nil
+}
+
+// exitPtrace is called in the exit path to detach all of t's tracees.
+func (t *Task) exitPtrace() {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	for target := range t.ptraceTracees {
+		if target.ptraceOpts.ExitKill {
+			target.tg.signalHandlers.mu.Lock()
+			target.sendSignalLocked(&arch.SignalInfo{
+				Signo: int32(linux.SIGKILL),
+			}, false /* group */)
+			target.tg.signalHandlers.mu.Unlock()
+		}
+		// Leave ptraceCode unchanged so that if the task is ptrace-stopped, it
+		// observes the ptraceCode it set before it entered the stop. I believe
+		// this is consistent with Linux.
+		target.forgetTracerLocked()
+	}
+	// "nil maps cannot be saved"
+	t.ptraceTracees = make(map[*Task]struct{})
+}
+
+// forgetTracerLocked detaches t's tracer and ensures that t is no longer
+// ptrace-stopped.
+//
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) forgetTracerLocked() {
+	t.ptraceOpts = ptraceOptions{}
+	t.ptraceSyscallMode = ptraceSyscallNone
+	t.ptraceSinglestep = false
+	t.ptraceTracer.Store((*Task)(nil))
+	if t.exitTracerNotified && !t.exitTracerAcked {
+		t.exitTracerAcked = true
+		t.exitNotifyLocked(true)
+	}
+	// If t is ptrace-stopped, but its thread group is in a group stop and t is
+	// eligible to participate, make it do so. This is essentially the reverse
+	// of the special case in ptraceAttach, which converts a group stop to a
+	// ptrace stop. ("Handling of restart from group-stop is currently buggy,
+	// but the "as planned" behavior is to leave tracee stopped and waiting for
+	// SIGCONT." - ptrace(2))
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if t.stop == nil {
+		return
+	}
+	if _, ok := t.stop.(*ptraceStop); ok {
+		if t.exitState < TaskExitInitiated && t.tg.groupStopPhase >= groupStopInitiated {
+			t.groupStopRequired = true
+		}
+		t.endInternalStopLocked()
+	}
+}
+
+// ptraceSignalLocked is called after signal dequeueing to check if t should
+// enter ptrace signal-delivery-stop.
+//
+// Preconditions: The signal mutex must be locked. The caller must be running
+// on the task goroutine.
+func (t *Task) ptraceSignalLocked(info *arch.SignalInfo) bool {
+	if linux.Signal(info.Signo) == linux.SIGKILL {
+		return false
+	}
+	if !t.hasTracer() {
+		return false
+	}
+	// The tracer might change this signal into a stop signal, in which case
+	// any SIGCONT received after the signal was originally dequeued should
+	// cancel it. This is consistent with Linux.
+	if t.tg.groupStopPhase == groupStopNone {
+		t.tg.groupStopPhase = groupStopDequeued
+	}
+	// Can't lock the TaskSet mutex while holding a signal mutex.
+	t.tg.signalHandlers.mu.Unlock()
+	defer t.tg.signalHandlers.mu.Lock()
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	tracer := t.Tracer()
+	if tracer == nil {
+		return false
+	}
+	t.ptraceCode = info.Signo
+	t.ptraceSiginfo = info
+	t.Debugf("Entering signal-delivery-stop for signal %d", info.Signo)
+	if t.beginPtraceStopLocked() {
+		tracer.signalStop(t, arch.CLD_TRAPPED, info.Signo)
+		tracer.tg.eventQueue.Notify(EventTraceeStop)
+	}
+	return true
+}
+
+// ptraceSeccomp is called when a seccomp-bpf filter returns action
+// SECCOMP_RET_TRACE to check if t should enter PTRACE_EVENT_SECCOMP stop. data
+// is the lower 16 bits of the filter's return value.
+func (t *Task) ptraceSeccomp(data uint16) bool {
+	if !t.hasTracer() {
+		return false
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	if !t.ptraceOpts.TraceSeccomp {
+		return false
+	}
+	t.Debugf("Entering PTRACE_EVENT_SECCOMP stop")
+	t.ptraceEventLocked(_PTRACE_EVENT_SECCOMP, uint64(data))
+	return true
+}
+
+// ptraceSyscallEnter is called immediately before entering a syscall to check
+// if t should enter ptrace syscall-enter-stop.
+func (t *Task) ptraceSyscallEnter() (taskRunState, bool) {
+	if !t.hasTracer() {
+		return nil, false
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	switch t.ptraceSyscallMode {
+	case ptraceSyscallNone:
+		return nil, false
+	case ptraceSyscallIntercept:
+		t.Debugf("Entering syscall-enter-stop from PTRACE_SYSCALL")
+		t.ptraceSyscallStopLocked()
+		return (*runSyscallAfterSyscallEnterStop)(nil), true
+	case ptraceSyscallEmu:
+		t.Debugf("Entering syscall-enter-stop from PTRACE_SYSEMU")
+		t.ptraceSyscallStopLocked()
+		return (*runSyscallAfterSysemuStop)(nil), true
+	}
+	panic(fmt.Sprintf("Unknown ptraceSyscallMode: %v", t.ptraceSyscallMode))
+}
+
+// ptraceSyscallExit is called immediately after leaving a syscall to check if
+// t should enter ptrace syscall-exit-stop.
+func (t *Task) ptraceSyscallExit() {
+	if !t.hasTracer() {
+		return
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	if t.ptraceSyscallMode != ptraceSyscallIntercept {
+		return
+	}
+	t.Debugf("Entering syscall-exit-stop")
+	t.ptraceSyscallStopLocked()
+}
+
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) ptraceSyscallStopLocked() {
+	code := int32(linux.SIGTRAP)
+	if t.ptraceOpts.SysGood {
+		code |= 0x80
+	}
+	t.ptraceTrapLocked(code)
+}
+
+type ptraceCloneKind int32
+
+const (
+	// ptraceCloneKindClone represents a call to Task.Clone where
+	// TerminationSignal is not SIGCHLD and Vfork is false.
+	ptraceCloneKindClone ptraceCloneKind = iota
+
+	// ptraceCloneKindFork represents a call to Task.Clone where
+	// TerminationSignal is SIGCHLD and Vfork is false.
+	ptraceCloneKindFork
+
+	// ptraceCloneKindVfork represents a call to Task.Clone where Vfork is
+	// true.
+	ptraceCloneKindVfork
+)
+
+// ptraceClone is called at the end of a clone or fork syscall to check if t
+// should enter PTRACE_EVENT_CLONE, PTRACE_EVENT_FORK, or PTRACE_EVENT_VFORK
+// stop. child is the new task.
+func (t *Task) ptraceClone(kind ptraceCloneKind, child *Task, opts *CloneOptions) bool {
+	if !t.hasTracer() {
+		return false
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	event := false
+	if !opts.Untraced {
+		switch kind {
+		case ptraceCloneKindClone:
+			if t.ptraceOpts.TraceClone {
+				t.Debugf("Entering PTRACE_EVENT_CLONE stop")
+				t.ptraceEventLocked(syscall.PTRACE_EVENT_CLONE, uint64(t.tg.pidns.tids[child]))
+				event = true
+			}
+		case ptraceCloneKindFork:
+			if t.ptraceOpts.TraceFork {
+				t.Debugf("Entering PTRACE_EVENT_FORK stop")
+				t.ptraceEventLocked(syscall.PTRACE_EVENT_FORK, uint64(t.tg.pidns.tids[child]))
+				event = true
+			}
+		case ptraceCloneKindVfork:
+			if t.ptraceOpts.TraceVfork {
+				t.Debugf("Entering PTRACE_EVENT_VFORK stop")
+				t.ptraceEventLocked(syscall.PTRACE_EVENT_VFORK, uint64(t.tg.pidns.tids[child]))
+				event = true
+			}
+		default:
+			panic(fmt.Sprintf("Unknown ptraceCloneKind: %v", kind))
+		}
+	}
+	// "If the PTRACE_O_TRACEFORK, PTRACE_O_TRACEVFORK, or PTRACE_O_TRACECLONE
+	// options are in effect, then children created by, respectively, vfork(2)
+	// or clone(2) with the CLONE_VFORK flag, fork(2) or clone(2) with the exit
+	// signal set to SIGCHLD, and other kinds of clone(2), are automatically
+	// attached to the same tracer which traced their parent. SIGSTOP is
+	// delivered to the children, causing them to enter signal-delivery-stop
+	// after they exit the system call which created them." - ptrace(2)
+	//
+	// clone(2)'s documentation of CLONE_UNTRACED and CLONE_PTRACE is
+	// confusingly wrong; see kernel/fork.c:_do_fork() => copy_process() =>
+	// include/linux/ptrace.h:ptrace_init_task().
+	if event || opts.InheritTracer {
+		tracer := t.Tracer()
+		if tracer != nil {
+			child.ptraceTracer.Store(tracer)
+			tracer.ptraceTracees[child] = struct{}{}
+			// "Flags are inherited by new tracees created and "auto-attached"
+			// via active PTRACE_O_TRACEFORK, PTRACE_O_TRACEVFORK, or
+			// PTRACE_O_TRACECLONE options."
+			child.ptraceOpts = t.ptraceOpts
+			child.tg.signalHandlers.mu.Lock()
+			// If the child is PT_SEIZED (currently not possible in the sentry
+			// because PTRACE_SEIZE is unimplemented, but for future
+			// reference), Linux just sets JOBCTL_TRAP_STOP instead, so the
+			// child skips signal-delivery-stop and goes directly to
+			// group-stop.
+			//
+			// The child will self-t.interrupt() when its task goroutine starts
+			// running, so we don't have to.
+			child.pendingSignals.enqueue(&arch.SignalInfo{
+				Signo: int32(linux.SIGSTOP),
+			})
+			child.tg.signalHandlers.mu.Unlock()
+		}
+	}
+	return event
+}
+
+// ptraceVforkDone is called after the end of a vfork stop to check if t should
+// enter PTRACE_EVENT_VFORK_DONE stop. child is the new task's thread ID in t's
+// PID namespace.
+func (t *Task) ptraceVforkDone(child ThreadID) bool {
+	if !t.hasTracer() {
+		return false
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	if !t.ptraceOpts.TraceVforkDone {
+		return false
+	}
+	t.Debugf("Entering PTRACE_EVENT_VFORK_DONE stop")
+	t.ptraceEventLocked(syscall.PTRACE_EVENT_VFORK_DONE, uint64(child))
+	return true
+}
+
+// ptraceExec is called at the end of an execve syscall to check if t should
+// enter PTRACE_EVENT_EXEC stop. oldTID is t's thread ID, in its *tracer's* PID
+// namespace, prior to the execve. (If t did not have a tracer at the time
+// oldTID was read, oldTID may be 0. This is consistent with Linux.)
+func (t *Task) ptraceExec(oldTID ThreadID) {
+	if !t.hasTracer() {
+		return
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	// Recheck with the TaskSet mutex locked. Most ptrace points don't need to
+	// do this because detaching resets ptrace options, but PTRACE_EVENT_EXEC
+	// is special because both TraceExec and !TraceExec do something if a
+	// tracer is attached.
+	if !t.hasTracer() {
+		return
+	}
+	if t.ptraceOpts.TraceExec {
+		t.Debugf("Entering PTRACE_EVENT_EXEC stop")
+		t.ptraceEventLocked(syscall.PTRACE_EVENT_EXEC, uint64(oldTID))
+		return
+	}
+	// "If the PTRACE_O_TRACEEXEC option is not in effect for the execing
+	// tracee, and if the tracee was PTRACE_ATTACHed rather that [sic]
+	// PTRACE_SEIZEd, the kernel delivers an extra SIGTRAP to the tracee after
+	// execve(2) returns. This is an ordinary signal (similar to one which can
+	// be generated by `kill -TRAP`, not a special kind of ptrace-stop.
+	// Employing PTRACE_GETSIGINFO for this signal returns si_code set to 0
+	// (SI_USER). This signal may be blocked by signal mask, and thus may be
+	// delivered (much) later." - ptrace(2)
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.sendSignalLocked(&arch.SignalInfo{
+		Signo: int32(linux.SIGTRAP),
+		Code:  arch.SignalInfoUser,
+	}, false /* group */)
+}
+
+// ptraceExit is called early in the task exit path to check if t should enter
+// PTRACE_EVENT_EXIT stop.
+func (t *Task) ptraceExit() {
+	if !t.hasTracer() {
+		return
+	}
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	if !t.ptraceOpts.TraceExit {
+		return
+	}
+	t.tg.signalHandlers.mu.Lock()
+	status := t.exitStatus.Status()
+	t.tg.signalHandlers.mu.Unlock()
+	t.Debugf("Entering PTRACE_EVENT_EXIT stop")
+	t.ptraceEventLocked(syscall.PTRACE_EVENT_EXIT, uint64(status))
+}
+
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) ptraceEventLocked(event int32, msg uint64) {
+	t.ptraceEventMsg = msg
+	// """
+	// PTRACE_EVENT stops are observed by the tracer as waitpid(2) returning
+	// with WIFSTOPPED(status), and WSTOPSIG(status) returns SIGTRAP. An
+	// additional bit is set in the higher byte of the status word: the value
+	// status>>8 will be
+	//
+	//   (SIGTRAP | PTRACE_EVENT_foo << 8).
+	//
+	// ...
+	//
+	// """ - ptrace(2)
+	t.ptraceTrapLocked(int32(linux.SIGTRAP) | (event << 8))
+}
+
+// ptraceKill implements ptrace(PTRACE_KILL, target). t is the caller.
+func (t *Task) ptraceKill(target *Task) error {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	if target.Tracer() != t {
+		return syserror.ESRCH
+	}
+	target.tg.signalHandlers.mu.Lock()
+	defer target.tg.signalHandlers.mu.Unlock()
+	// "This operation is deprecated; do not use it! Instead, send a SIGKILL
+	// directly using kill(2) or tgkill(2). The problem with PTRACE_KILL is
+	// that it requires the tracee to be in signal-delivery-stop, otherwise it
+	// may not work (i.e., may complete successfully but won't kill the
+	// tracee)." - ptrace(2)
+	if target.stop == nil {
+		return nil
+	}
+	if _, ok := target.stop.(*ptraceStop); !ok {
+		return nil
+	}
+	target.ptraceCode = int32(linux.SIGKILL)
+	target.endInternalStopLocked()
+	return nil
+}
+
+// Ptrace implements the ptrace system call.
+func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error {
+	// PTRACE_TRACEME ignores all other arguments.
+	if req == syscall.PTRACE_TRACEME {
+		return t.ptraceTraceme()
+	}
+	// All other ptrace requests operate on a current or future tracee
+	// specified by pid.
+	target := t.tg.pidns.TaskWithID(pid)
+	if target == nil {
+		return syserror.ESRCH
+	}
+
+	// PTRACE_ATTACH (and PTRACE_SEIZE, which is unimplemented) do not require
+	// that target is not already a tracee.
+	if req == syscall.PTRACE_ATTACH {
+		return t.ptraceAttach(target)
+	}
+	// PTRACE_KILL (and PTRACE_INTERRUPT, which is unimplemented) require that
+	// the target is a tracee, but does not require that it is ptrace-stopped.
+	if req == syscall.PTRACE_KILL {
+		return t.ptraceKill(target)
+	}
+	// All other ptrace requests require that the target is a ptrace-stopped
+	// tracee, and freeze the ptrace-stop so the tracee can be operated on.
+	t.tg.pidns.owner.mu.RLock()
+	if target.Tracer() != t {
+		t.tg.pidns.owner.mu.RUnlock()
+		return syserror.ESRCH
+	}
+	if !target.ptraceFreeze() {
+		t.tg.pidns.owner.mu.RUnlock()
+		// "Most ptrace commands (all except PTRACE_ATTACH, PTRACE_SEIZE,
+		// PTRACE_TRACEME, PTRACE_INTERRUPT, and PTRACE_KILL) require the
+		// tracee to be in a ptrace-stop, otherwise they fail with ESRCH." -
+		// ptrace(2)
+		return syserror.ESRCH
+	}
+	t.tg.pidns.owner.mu.RUnlock()
+	// Even if the target has a ptrace-stop active, the tracee's task goroutine
+	// may not yet have reached Task.doStop; wait for it to do so. This is safe
+	// because there's no way for target to initiate a ptrace-stop and then
+	// block (by calling Task.block) before entering it.
+	//
+	// Caveat: If tasks were just restored, the tracee's first call to
+	// Task.Activate (in Task.run) occurs before its first call to Task.doStop,
+	// which may block if the tracer's address space is active.
+	t.UninterruptibleSleepStart(true)
+	target.waitGoroutineStoppedOrExited()
+	t.UninterruptibleSleepFinish(true)
+
+	// Resuming commands end the ptrace stop, but only if successful.
+	switch req {
+	case syscall.PTRACE_DETACH:
+		if err := t.ptraceDetach(target, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	case syscall.PTRACE_CONT:
+		if err := target.ptraceUnstop(ptraceSyscallNone, false, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	case syscall.PTRACE_SYSCALL:
+		if err := target.ptraceUnstop(ptraceSyscallIntercept, false, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	case syscall.PTRACE_SINGLESTEP:
+		if err := target.ptraceUnstop(ptraceSyscallNone, true, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	case syscall.PTRACE_SYSEMU:
+		if err := target.ptraceUnstop(ptraceSyscallEmu, false, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	case syscall.PTRACE_SYSEMU_SINGLESTEP:
+		if err := target.ptraceUnstop(ptraceSyscallEmu, true, linux.Signal(data)); err != nil {
+			target.ptraceUnfreeze()
+			return err
+		}
+		return nil
+	}
+	// All other ptrace requests expect us to unfreeze the stop.
+	defer target.ptraceUnfreeze()
+
+	switch req {
+	case syscall.PTRACE_PEEKTEXT, syscall.PTRACE_PEEKDATA:
+		// "At the system call level, the PTRACE_PEEKTEXT, PTRACE_PEEKDATA, and
+		// PTRACE_PEEKUSER requests have a different API: they store the result
+		// at the address specified by the data parameter, and the return value
+		// is the error flag." - ptrace(2)
+		word := t.Arch().Native(0)
+		if _, err := usermem.CopyObjectIn(t, target.MemoryManager(), addr, word, usermem.IOOpts{
+			IgnorePermissions: true,
+		}); err != nil {
+			return err
+		}
+		_, err := t.CopyOut(data, word)
+		return err
+
+	case syscall.PTRACE_POKETEXT, syscall.PTRACE_POKEDATA:
+		_, err := usermem.CopyObjectOut(t, target.MemoryManager(), addr, t.Arch().Native(uintptr(data)), usermem.IOOpts{
+			IgnorePermissions: true,
+		})
+		return err
+
+	case syscall.PTRACE_PEEKUSR: // aka PTRACE_PEEKUSER
+		n, err := target.Arch().PtracePeekUser(uintptr(addr))
+		if err != nil {
+			return err
+		}
+		_, err = t.CopyOut(data, n)
+		return err
+
+	case syscall.PTRACE_POKEUSR: // aka PTRACE_POKEUSER
+		return target.Arch().PtracePokeUser(uintptr(addr), uintptr(data))
+
+	case syscall.PTRACE_GETREGS:
+		// "Copy the tracee's general-purpose ... registers ... to the address
+		// data in the tracer. ... (addr is ignored.) Note that SPARC systems
+		// have the meaning of data and addr reversed ..."
+		_, err := target.Arch().PtraceGetRegs(&usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: data,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		})
+		return err
+
+	case syscall.PTRACE_GETFPREGS:
+		_, err := target.Arch().PtraceGetFPRegs(&usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: data,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		})
+		return err
+
+	case syscall.PTRACE_GETREGSET:
+		// "Read the tracee's registers. addr specifies, in an
+		// architecture-dependent way, the type of registers to be read. ...
+		// data points to a struct iovec, which describes the destination
+		// buffer's location and length. On return, the kernel modifies iov.len
+		// to indicate the actual number of bytes returned." - ptrace(2)
+		ars, err := t.CopyInIovecs(data, 1)
+		if err != nil {
+			return err
+		}
+		ar := ars.Head()
+		n, err := target.Arch().PtraceGetRegSet(uintptr(addr), &usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: ar.Start,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		}, int(ar.Length()))
+		if err != nil {
+			return err
+		}
+		ar.End -= usermem.Addr(n)
+		return t.CopyOutIovecs(data, usermem.AddrRangeSeqOf(ar))
+
+	case syscall.PTRACE_SETREGS:
+		_, err := target.Arch().PtraceSetRegs(&usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: data,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		})
+		return err
+
+	case syscall.PTRACE_SETFPREGS:
+		_, err := target.Arch().PtraceSetFPRegs(&usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: data,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		})
+		return err
+
+	case syscall.PTRACE_SETREGSET:
+		ars, err := t.CopyInIovecs(data, 1)
+		if err != nil {
+			return err
+		}
+		ar := ars.Head()
+		n, err := target.Arch().PtraceSetRegSet(uintptr(addr), &usermem.IOReadWriter{
+			Ctx:  t,
+			IO:   t.MemoryManager(),
+			Addr: ar.Start,
+			Opts: usermem.IOOpts{
+				AddressSpaceActive: true,
+			},
+		}, int(ar.Length()))
+		if err != nil {
+			return err
+		}
+		ar.End -= usermem.Addr(n)
+		return t.CopyOutIovecs(data, usermem.AddrRangeSeqOf(ar))
+
+	case syscall.PTRACE_GETSIGINFO:
+		t.tg.pidns.owner.mu.RLock()
+		defer t.tg.pidns.owner.mu.RUnlock()
+		if target.ptraceSiginfo == nil {
+			return syserror.EINVAL
+		}
+		_, err := t.CopyOut(data, target.ptraceSiginfo)
+		return err
+
+	case syscall.PTRACE_SETSIGINFO:
+		var info arch.SignalInfo
+		if _, err := t.CopyIn(data, &info); err != nil {
+			return err
+		}
+		t.tg.pidns.owner.mu.RLock()
+		defer t.tg.pidns.owner.mu.RUnlock()
+		if target.ptraceSiginfo == nil {
+			return syserror.EINVAL
+		}
+		target.ptraceSiginfo = &info
+		return nil
+
+	case PTRACE_GETSIGMASK:
+		if addr != linux.SignalSetSize {
+			return syserror.EINVAL
+		}
+		target.mu.Lock()
+		defer target.mu.Unlock()
+		_, err := t.CopyOut(data, target.tr.SignalMask)
+		return err
+
+	case PTRACE_SETSIGMASK:
+		if addr != linux.SignalSetSize {
+			return syserror.EINVAL
+		}
+		var mask linux.SignalSet
+		if _, err := t.CopyIn(data, &mask); err != nil {
+			return err
+		}
+		// The target's task goroutine is stopped, so this is safe:
+		target.SetSignalMask(mask &^ UnblockableSignals)
+		return nil
+
+	case syscall.PTRACE_SETOPTIONS:
+		t.tg.pidns.owner.mu.Lock()
+		defer t.tg.pidns.owner.mu.Unlock()
+		validOpts := uintptr(_PTRACE_O_EXITKILL | syscall.PTRACE_O_TRACESYSGOOD | syscall.PTRACE_O_TRACECLONE |
+			syscall.PTRACE_O_TRACEEXEC | syscall.PTRACE_O_TRACEEXIT | syscall.PTRACE_O_TRACEFORK |
+			_PTRACE_O_TRACESECCOMP | syscall.PTRACE_O_TRACEVFORK | syscall.PTRACE_O_TRACEVFORKDONE)
+		if uintptr(data)&^validOpts != 0 {
+			return syserror.EINVAL
+		}
+		target.ptraceOpts = ptraceOptions{
+			ExitKill:       data&_PTRACE_O_EXITKILL != 0,
+			SysGood:        data&syscall.PTRACE_O_TRACESYSGOOD != 0,
+			TraceClone:     data&syscall.PTRACE_O_TRACECLONE != 0,
+			TraceExec:      data&syscall.PTRACE_O_TRACEEXEC != 0,
+			TraceExit:      data&syscall.PTRACE_O_TRACEEXIT != 0,
+			TraceFork:      data&syscall.PTRACE_O_TRACEFORK != 0,
+			TraceSeccomp:   data&_PTRACE_O_TRACESECCOMP != 0,
+			TraceVfork:     data&syscall.PTRACE_O_TRACEVFORK != 0,
+			TraceVforkDone: data&syscall.PTRACE_O_TRACEVFORKDONE != 0,
+		}
+		return nil
+
+	case syscall.PTRACE_GETEVENTMSG:
+		t.tg.pidns.owner.mu.RLock()
+		defer t.tg.pidns.owner.mu.RUnlock()
+		_, err := t.CopyOut(usermem.Addr(data), target.ptraceEventMsg)
+		return err
+
+	default:
+		// PEEKSIGINFO is unimplemented but seems to have no users anywhere.
+		return syserror.EIO
+	}
+}
diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go
new file mode 100644
index 000000000..635372993
--- /dev/null
+++ b/pkg/sentry/kernel/rseq.go
@@ -0,0 +1,118 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/hostcpu"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Restartable sequences, as described in https://lwn.net/Articles/650333/.
+
+// RSEQCriticalRegion describes a restartable sequence critical region.
+type RSEQCriticalRegion struct {
+	// When a task in this thread group has its CPU preempted (as defined by
+	// platform.ErrContextCPUPreempted) or has a signal delivered to an
+	// application handler while its instruction pointer is in CriticalSection,
+	// set the instruction pointer to Restart and application register r10 (on
+	// amd64) to the former instruction pointer.
+	CriticalSection usermem.AddrRange
+	Restart         usermem.Addr
+}
+
+// RSEQAvailable returns true if t supports restartable sequences.
+func (t *Task) RSEQAvailable() bool {
+	return t.k.useHostCores && t.k.Platform.DetectsCPUPreemption()
+}
+
+// RSEQCriticalRegion returns a copy of t's thread group's current restartable
+// sequence.
+func (t *Task) RSEQCriticalRegion() RSEQCriticalRegion {
+	return *t.tg.rscr.Load().(*RSEQCriticalRegion)
+}
+
+// SetRSEQCriticalRegion replaces t's thread group's restartable sequence.
+//
+// Preconditions: t.RSEQAvailable() == true.
+func (t *Task) SetRSEQCriticalRegion(rscr RSEQCriticalRegion) error {
+	// These checks are somewhat more lenient than in Linux, which (bizarrely)
+	// requires rscr.CriticalSection to be non-empty and rscr.Restart to be
+	// outside of rscr.CriticalSection, even if rscr.CriticalSection.Start == 0
+	// (which disables the critical region).
+	if rscr.CriticalSection.Start == 0 {
+		rscr.CriticalSection.End = 0
+		rscr.Restart = 0
+		t.tg.rscr.Store(&rscr)
+		return nil
+	}
+	if rscr.CriticalSection.Start >= rscr.CriticalSection.End {
+		return syserror.EINVAL
+	}
+	if rscr.CriticalSection.Contains(rscr.Restart) {
+		return syserror.EINVAL
+	}
+	// TODO: check that rscr.CriticalSection and rscr.Restart are in
+	// the application address range, for consistency with Linux
+	t.tg.rscr.Store(&rscr)
+	return nil
+}
+
+// RSEQCPUAddr returns the address that RSEQ will keep updated with t's CPU
+// number.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) RSEQCPUAddr() usermem.Addr {
+	return t.rseqCPUAddr
+}
+
+// SetRSEQCPUAddr replaces the address that RSEQ will keep updated with t's CPU
+// number.
+//
+// Preconditions: t.RSEQAvailable() == true. The caller must be running on the
+// task goroutine. t's AddressSpace must be active.
+func (t *Task) SetRSEQCPUAddr(addr usermem.Addr) error {
+	t.rseqCPUAddr = addr
+	if addr != 0 {
+		if err := t.rseqCopyOutCPU(); err != nil {
+			t.rseqCPUAddr = 0
+			t.rseqCPU = -1
+			return syserror.EINVAL // yes, EINVAL, not err or EFAULT
+		}
+	} else {
+		t.rseqCPU = -1
+	}
+	return nil
+}
+
+// Preconditions: The caller must be running on the task goroutine. t's
+// AddressSpace must be active.
+func (t *Task) rseqCopyOutCPU() error {
+	t.rseqCPU = int32(hostcpu.GetCPU())
+	buf := t.CopyScratchBuffer(4)
+	usermem.ByteOrder.PutUint32(buf, uint32(t.rseqCPU))
+	_, err := t.CopyOutBytes(t.rseqCPUAddr, buf)
+	return err
+}
+
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) rseqInterrupt() {
+	rscr := t.tg.rscr.Load().(*RSEQCriticalRegion)
+	if ip := t.Arch().IP(); rscr.CriticalSection.Contains(usermem.Addr(ip)) {
+		t.Debugf("Interrupted RSEQ critical section at %#x; restarting at %#x", ip, rscr.Restart)
+		t.Arch().SetIP(uintptr(rscr.Restart))
+		t.Arch().SetRSEQInterruptedIP(ip)
+	}
+}
diff --git a/pkg/sentry/kernel/sched/BUILD b/pkg/sentry/kernel/sched/BUILD
new file mode 100644
index 000000000..b533c51c4
--- /dev/null
+++ b/pkg/sentry/kernel/sched/BUILD
@@ -0,0 +1,20 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "sched",
+    srcs = [
+        "cpuset.go",
+        "sched.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched",
+    visibility = ["//pkg/sentry:internal"],
+)
+
+go_test(
+    name = "sched_test",
+    size = "small",
+    srcs = ["cpuset_test.go"],
+    embed = [":sched"],
+)
diff --git a/pkg/sentry/kernel/sched/cpuset.go b/pkg/sentry/kernel/sched/cpuset.go
new file mode 100644
index 000000000..0a97603f0
--- /dev/null
+++ b/pkg/sentry/kernel/sched/cpuset.go
@@ -0,0 +1,105 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sched
+
+import "math/bits"
+
+const (
+	bitsPerByte  = 8
+	bytesPerLong = 8 // only for 64-bit architectures
+)
+
+// CPUSet contains a bitmap to record CPU information.
+//
+// Note that this definition is only correct for little-endian architectures,
+// since Linux's cpumask_t uses unsigned long.
+type CPUSet []byte
+
+// CPUSetSize returns the size in bytes of a CPUSet that can contain num cpus.
+func CPUSetSize(num uint) uint {
+	// NOTE: Applications may expect that the size of a CPUSet in
+	// bytes is always a multiple of sizeof(unsigned long), since this is true
+	// in Linux. Thus we always round up.
+	bytes := (num + bitsPerByte - 1) / bitsPerByte
+	longs := (bytes + bytesPerLong - 1) / bytesPerLong
+	return longs * bytesPerLong
+}
+
+// NewCPUSet returns a CPUSet for the given number of CPUs which initially
+// contains no CPUs.
+func NewCPUSet(num uint) CPUSet {
+	return CPUSet(make([]byte, CPUSetSize(num)))
+}
+
+// NewFullCPUSet returns a CPUSet for the given number of CPUs, all of which
+// are present in the set.
+func NewFullCPUSet(num uint) CPUSet {
+	c := NewCPUSet(num)
+	var i uint
+	for ; i < num/bitsPerByte; i++ {
+		c[i] = 0xff
+	}
+	if rem := num % bitsPerByte; rem != 0 {
+		c[i] = (1 << rem) - 1
+	}
+	return c
+}
+
+// Size returns the size of 'c' in bytes.
+func (c CPUSet) Size() uint {
+	return uint(len(c))
+}
+
+// NumCPUs returns how many cpus are set in the CPUSet.
+func (c CPUSet) NumCPUs() uint {
+	var n int
+	for _, b := range c {
+		n += bits.OnesCount8(b)
+	}
+	return uint(n)
+}
+
+// Copy returns a copy of the CPUSet.
+func (c CPUSet) Copy() CPUSet {
+	return append(CPUSet(nil), c...)
+}
+
+// Set sets the bit corresponding to cpu.
+func (c *CPUSet) Set(cpu uint) {
+	(*c)[cpu/bitsPerByte] |= 1 << (cpu % bitsPerByte)
+}
+
+// ClearAbove clears bits corresponding to cpu and all higher cpus.
+func (c *CPUSet) ClearAbove(cpu uint) {
+	i := cpu / bitsPerByte
+	if i >= c.Size() {
+		return
+	}
+	(*c)[i] &^= 0xff << (cpu % bitsPerByte)
+	for i++; i < c.Size(); i++ {
+		(*c)[i] = 0
+	}
+}
+
+// ForEachCPU iterates over the CPUSet and calls fn with the cpu index if
+// it's set.
+func (c CPUSet) ForEachCPU(fn func(uint)) {
+	for i := uint(0); i < c.Size()*bitsPerByte; i++ {
+		bit := uint(1) << (i & (bitsPerByte - 1))
+		if uint(c[i/bitsPerByte])&bit == bit {
+			fn(i)
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/sched/cpuset_test.go b/pkg/sentry/kernel/sched/cpuset_test.go
new file mode 100644
index 000000000..8a6e12958
--- /dev/null
+++ b/pkg/sentry/kernel/sched/cpuset_test.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sched
+
+import (
+	"testing"
+)
+
+func TestNumCPUs(t *testing.T) {
+	for i := uint(0); i < 1024; i++ {
+		c := NewCPUSet(i)
+		for j := uint(0); j < i; j++ {
+			c.Set(j)
+		}
+		n := c.NumCPUs()
+		if n != i {
+			t.Errorf("got wrong number of cpus %d, want %d", n, i)
+		}
+	}
+}
+
+func TestClearAbove(t *testing.T) {
+	const n = 1024
+	c := NewFullCPUSet(n)
+	for i := uint(0); i < n; i++ {
+		cpu := n - i
+		c.ClearAbove(cpu)
+		if got := c.NumCPUs(); got != cpu {
+			t.Errorf("iteration %d: got %d cpus, wanted %d", i, got, cpu)
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/sched/sched.go b/pkg/sentry/kernel/sched/sched.go
new file mode 100644
index 000000000..f1de1da60
--- /dev/null
+++ b/pkg/sentry/kernel/sched/sched.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package sched implements scheduler related features.
+package sched
diff --git a/pkg/sentry/kernel/seccomp.go b/pkg/sentry/kernel/seccomp.go
new file mode 100644
index 000000000..b7c4a507f
--- /dev/null
+++ b/pkg/sentry/kernel/seccomp.go
@@ -0,0 +1,205 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/bpf"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const maxSyscallFilterInstructions = 1 << 15
+
+type seccompResult int
+
+const (
+	// seccompResultDeny indicates that a syscall should not be executed.
+	seccompResultDeny seccompResult = iota
+
+	// seccompResultAllow indicates that a syscall should be executed.
+	seccompResultAllow
+
+	// seccompResultKill indicates that the task should be killed immediately,
+	// with the exit status indicating that the task was killed by SIGSYS.
+	seccompResultKill
+
+	// seccompResultTrace indicates that a ptracer was successfully notified as
+	// a result of a SECCOMP_RET_TRACE.
+	seccompResultTrace
+)
+
+// seccompData is equivalent to struct seccomp_data, which contains the data
+// passed to seccomp-bpf filters.
+type seccompData struct {
+	// nr is the system call number.
+	nr int32
+
+	// arch is an AUDIT_ARCH_* value indicating the system call convention.
+	arch uint32
+
+	// instructionPointer is the value of the instruction pointer at the time
+	// of the system call.
+	instructionPointer uint64
+
+	// args contains the first 6 system call arguments.
+	args [6]uint64
+}
+
+func (d *seccompData) asBPFInput() bpf.Input {
+	return bpf.InputBytes{binary.Marshal(nil, usermem.ByteOrder, d), usermem.ByteOrder}
+}
+
+func seccompSiginfo(t *Task, errno, sysno int32, ip usermem.Addr) *arch.SignalInfo {
+	si := &arch.SignalInfo{
+		Signo: int32(linux.SIGSYS),
+		Errno: errno,
+		Code:  arch.SYS_SECCOMP,
+	}
+	si.SetCallAddr(uint64(ip))
+	si.SetSyscall(sysno)
+	si.SetArch(t.SyscallTable().AuditNumber)
+	return si
+}
+
+// checkSeccompSyscall applies the task's seccomp filters before the execution
+// of syscall sysno at instruction pointer ip. (These parameters must be passed
+// in because vsyscalls do not use the values in t.Arch().)
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) checkSeccompSyscall(sysno int32, args arch.SyscallArguments, ip usermem.Addr) seccompResult {
+	result := t.evaluateSyscallFilters(sysno, args, ip)
+	switch result & linux.SECCOMP_RET_ACTION {
+	case linux.SECCOMP_RET_TRAP:
+		// "Results in the kernel sending a SIGSYS signal to the triggering
+		// task without executing the system call. ... The SECCOMP_RET_DATA
+		// portion of the return value will be passed as si_errno." -
+		// Documentation/prctl/seccomp_filter.txt
+		t.SendSignal(seccompSiginfo(t, int32(result&linux.SECCOMP_RET_DATA), sysno, ip))
+		return seccompResultDeny
+
+	case linux.SECCOMP_RET_ERRNO:
+		// "Results in the lower 16-bits of the return value being passed to
+		// userland as the errno without executing the system call."
+		t.Arch().SetReturn(-uintptr(result & linux.SECCOMP_RET_DATA))
+		return seccompResultDeny
+
+	case linux.SECCOMP_RET_TRACE:
+		// "When returned, this value will cause the kernel to attempt to
+		// notify a ptrace()-based tracer prior to executing the system call.
+		// If there is no tracer present, -ENOSYS is returned to userland and
+		// the system call is not executed."
+		if t.ptraceSeccomp(uint16(result & linux.SECCOMP_RET_DATA)) {
+			return seccompResultTrace
+		}
+		// This useless-looking temporary is needed because Go.
+		tmp := uintptr(syscall.ENOSYS)
+		t.Arch().SetReturn(-tmp)
+		return seccompResultDeny
+
+	case linux.SECCOMP_RET_ALLOW:
+		// "Results in the system call being executed."
+		return seccompResultAllow
+
+	case linux.SECCOMP_RET_KILL:
+		// "Results in the task exiting immediately without executing the
+		// system call. The exit status of the task will be SIGSYS, not
+		// SIGKILL."
+		fallthrough
+	default: // consistent with Linux
+		return seccompResultKill
+	}
+}
+
+func (t *Task) evaluateSyscallFilters(sysno int32, args arch.SyscallArguments, ip usermem.Addr) uint32 {
+	data := seccompData{
+		nr:                 sysno,
+		arch:               t.tc.st.AuditNumber,
+		instructionPointer: uint64(ip),
+	}
+	// data.args is []uint64 and args is []arch.SyscallArgument (uintptr), so
+	// we can't do any slicing tricks or even use copy/append here.
+	for i, arg := range args {
+		if i >= len(data.args) {
+			break
+		}
+		data.args[i] = arg.Uint64()
+	}
+	input := data.asBPFInput()
+
+	ret := uint32(linux.SECCOMP_RET_ALLOW)
+	// "Every filter successfully installed will be evaluated (in reverse
+	// order) for each system call the task makes." - kernel/seccomp.c
+	for i := len(t.syscallFilters) - 1; i >= 0; i-- {
+		thisRet, err := bpf.Exec(t.syscallFilters[i], input)
+		if err != nil {
+			t.Debugf("seccomp-bpf filter %d returned error: %v", i, err)
+			thisRet = linux.SECCOMP_RET_KILL
+		}
+		// "If multiple filters exist, the return value for the evaluation of a
+		// given system call will always use the highest precedent value." -
+		// Documentation/prctl/seccomp_filter.txt
+		//
+		// (Note that this contradicts prctl(2): "If the filters permit prctl()
+		// calls, then additional filters can be added; they are run in order
+		// until the first non-allow result is seen." prctl(2) is incorrect.)
+		//
+		// "The ordering ensures that a min_t() over composed return values
+		// always selects the least permissive choice." -
+		// include/uapi/linux/seccomp.h
+		if (thisRet & linux.SECCOMP_RET_ACTION) < (ret & linux.SECCOMP_RET_ACTION) {
+			ret = thisRet
+		}
+	}
+
+	return ret
+}
+
+// AppendSyscallFilter adds BPF program p as a system call filter.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) AppendSyscallFilter(p bpf.Program) error {
+	// Cap the combined length of all syscall filters (plus a penalty of 4
+	// instructions per filter beyond the first) to
+	// maxSyscallFilterInstructions. (This restriction is inherited from
+	// Linux.)
+	totalLength := p.Length()
+	for _, f := range t.syscallFilters {
+		totalLength += f.Length() + 4
+	}
+	if totalLength > maxSyscallFilterInstructions {
+		return syserror.ENOMEM
+	}
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.syscallFilters = append(t.syscallFilters, p)
+	return nil
+}
+
+// SeccompMode returns a SECCOMP_MODE_* constant indicating the task's current
+// seccomp syscall filtering mode, appropriate for both prctl(PR_GET_SECCOMP)
+// and /proc/[pid]/status.
+func (t *Task) SeccompMode() int {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if len(t.syscallFilters) > 0 {
+		return linux.SECCOMP_MODE_FILTER
+	}
+	return linux.SECCOMP_MODE_NONE
+}
diff --git a/pkg/sentry/kernel/semaphore/BUILD b/pkg/sentry/kernel/semaphore/BUILD
new file mode 100644
index 000000000..1656ad126
--- /dev/null
+++ b/pkg/sentry/kernel/semaphore/BUILD
@@ -0,0 +1,62 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_template_instance(
+    name = "waiter_list",
+    out = "waiter_list.go",
+    package = "semaphore",
+    prefix = "waiter",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*waiter",
+    },
+)
+
+go_stateify(
+    name = "semaphore_state",
+    srcs = [
+        "semaphore.go",
+        "waiter_list.go",
+    ],
+    out = "semaphore_autogen_state.go",
+    package = "semaphore",
+)
+
+go_library(
+    name = "semaphore",
+    srcs = [
+        "semaphore.go",
+        "semaphore_autogen_state.go",
+        "waiter_list.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/semaphore",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/time",
+        "//pkg/state",
+        "//pkg/state/statefile",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "semaphore_test",
+    size = "small",
+    srcs = ["semaphore_test.go"],
+    embed = [":semaphore"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go
new file mode 100644
index 000000000..19ad5d537
--- /dev/null
+++ b/pkg/sentry/kernel/semaphore/semaphore.go
@@ -0,0 +1,473 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package semaphore implements System V semaphores.
+package semaphore
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	valueMax = 32767 // SEMVMX
+
+	// semaphoresMax is "maximum number of semaphores per semaphore ID" (SEMMSL).
+	semaphoresMax = 32000
+
+	// setMax is "system-wide limit on the number of semaphore sets" (SEMMNI).
+	setsMax = 32000
+
+	// semaphoresTotalMax is "system-wide limit on the number of semaphores"
+	// (SEMMNS = SEMMNI*SEMMSL).
+	semaphoresTotalMax = 1024000000
+)
+
+// Registry maintains a set of semaphores that can be found by key or ID.
+type Registry struct {
+	// mu protects all fields below.
+	mu         sync.Mutex `state:"nosave"`
+	semaphores map[int32]*Set
+	lastIDUsed int32
+}
+
+// Set represents a set of semaphores that can be operated atomically.
+type Set struct {
+	// Id is a handle that identifies the set.
+	ID int32
+
+	// key is an user provided key that can be shared between processes.
+	key int32
+
+	// creator is the user that created the set. Immutable.
+	creator fs.FileOwner
+
+	// mu protects all fields below.
+	mu         sync.Mutex `state:"nosave"`
+	owner      fs.FileOwner
+	perms      fs.FilePermissions
+	opTime     ktime.Time
+	changeTime ktime.Time
+	sems       []sem
+
+	// dead is set to true when the set is removed and can't be reached anymore.
+	// All waiters must wake up and fail when set is dead.
+	dead bool
+}
+
+// sem represents a single semanphore from a set.
+type sem struct {
+	value   int16
+	waiters waiterList `state:"zerovalue"`
+}
+
+// waiter represents a caller that is waiting for the semaphore value to
+// become positive or zero.
+type waiter struct {
+	waiterEntry
+
+	// value represents how much resource the waiter needs to wake up.
+	value int16
+	ch    chan struct{}
+}
+
+// NewRegistry creates a new semaphore set registry.
+func NewRegistry() *Registry {
+	return &Registry{semaphores: make(map[int32]*Set)}
+}
+
+// FindOrCreate searches for a semaphore set that matches 'key'. If not found,
+// it may create a new one if requested. If private is true, key is ignored and
+// a new set is always created. If create is false, it fails if a set cannot
+// be found. If exclusive is true, it fails if a set with the same key already
+// exists.
+func (r *Registry) FindOrCreate(ctx context.Context, key, nsems int32, mode linux.FileMode, private, create, exclusive bool) (*Set, error) {
+	if nsems < 0 || nsems > semaphoresMax {
+		return nil, syserror.EINVAL
+	}
+
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	if !private {
+		// Look up an existing semaphore.
+		if set := r.findByKey(key); set != nil {
+			// Check that caller can access semaphore set.
+			creds := auth.CredentialsFromContext(ctx)
+			if !set.checkPerms(creds, fs.PermsFromMode(mode)) {
+				return nil, syserror.EACCES
+			}
+
+			// Validate parameters.
+			if nsems > int32(set.size()) {
+				return nil, syserror.EINVAL
+			}
+			if create && exclusive {
+				return nil, syserror.EEXIST
+			}
+			return set, nil
+		}
+
+		if !create {
+			// Semaphore not found and should not be created.
+			return nil, syserror.ENOENT
+		}
+	}
+
+	// Zero is only valid if an existing set is found.
+	if nsems == 0 {
+		return nil, syserror.EINVAL
+	}
+
+	// Apply system limits.
+	if len(r.semaphores) >= setsMax {
+		return nil, syserror.EINVAL
+	}
+	if r.totalSems() > int(semaphoresTotalMax-nsems) {
+		return nil, syserror.EINVAL
+	}
+
+	// Finally create a new set.
+	owner := fs.FileOwnerFromContext(ctx)
+	perms := fs.FilePermsFromMode(mode)
+	return r.newSet(ctx, key, owner, owner, perms, nsems)
+}
+
+// RemoveID removes set with give 'id' from the registry and marks the set as
+// dead. All waiters will be awakened and fail.
+func (r *Registry) RemoveID(id int32, creds *auth.Credentials) error {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+
+	set := r.semaphores[id]
+	if set == nil {
+		return syserror.EINVAL
+	}
+
+	// "The effective user ID of the calling process must match the creator or
+	// owner of the semaphore set, or the caller must be privileged."
+	if !set.checkCredentials(creds) && !set.checkCapability(creds) {
+		return syserror.EACCES
+	}
+
+	delete(r.semaphores, set.ID)
+	set.destroy()
+	return nil
+}
+
+func (r *Registry) newSet(ctx context.Context, key int32, owner, creator fs.FileOwner, perms fs.FilePermissions, nsems int32) (*Set, error) {
+	set := &Set{
+		key:        key,
+		owner:      owner,
+		creator:    owner,
+		perms:      perms,
+		changeTime: ktime.NowFromContext(ctx),
+		sems:       make([]sem, nsems),
+	}
+
+	// Find the next available ID.
+	for id := r.lastIDUsed + 1; id != r.lastIDUsed; id++ {
+		// Handle wrap around.
+		if id < 0 {
+			id = 0
+			continue
+		}
+		if r.semaphores[id] == nil {
+			r.lastIDUsed = id
+			r.semaphores[id] = set
+			set.ID = id
+			return set, nil
+		}
+	}
+
+	log.Warningf("Semaphore map is full, they must be leaking")
+	return nil, syserror.ENOMEM
+}
+
+// FindByID looks up a set given an ID.
+func (r *Registry) FindByID(id int32) *Set {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.semaphores[id]
+}
+
+func (r *Registry) findByKey(key int32) *Set {
+	for _, v := range r.semaphores {
+		if v.key == key {
+			return v
+		}
+	}
+	return nil
+}
+
+func (r *Registry) totalSems() int {
+	totalSems := 0
+	for _, v := range r.semaphores {
+		totalSems += v.size()
+	}
+	return totalSems
+}
+
+func (s *Set) findSem(num int32) *sem {
+	if num < 0 || int(num) >= s.size() {
+		return nil
+	}
+	return &s.sems[num]
+}
+
+func (s *Set) size() int {
+	return len(s.sems)
+}
+
+// Change changes some fields from the set atomically.
+func (s *Set) Change(ctx context.Context, creds *auth.Credentials, owner fs.FileOwner, perms fs.FilePermissions) error {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// "The effective UID of the calling process must match the owner or creator
+	// of the semaphore set, or the caller must be privileged."
+	if !s.checkCredentials(creds) && !s.checkCapability(creds) {
+		return syserror.EACCES
+	}
+
+	s.owner = owner
+	s.perms = perms
+	s.changeTime = ktime.NowFromContext(ctx)
+	return nil
+}
+
+// SetVal overrides a semaphore value, waking up waiters as needed.
+func (s *Set) SetVal(ctx context.Context, num int32, val int16, creds *auth.Credentials) error {
+	if val < 0 || val > valueMax {
+		return syserror.ERANGE
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// "The calling process must have alter permission on the semaphore set."
+	if !s.checkPerms(creds, fs.PermMask{Write: true}) {
+		return syserror.EACCES
+	}
+
+	sem := s.findSem(num)
+	if sem == nil {
+		return syserror.ERANGE
+	}
+
+	// TODO: Clear undo entries in all processes
+	sem.value = val
+	s.changeTime = ktime.NowFromContext(ctx)
+	sem.wakeWaiters()
+	return nil
+}
+
+// GetVal returns a semaphore value.
+func (s *Set) GetVal(num int32, creds *auth.Credentials) (int16, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// "The calling process must have read permission on the semaphore set."
+	if !s.checkPerms(creds, fs.PermMask{Read: true}) {
+		return 0, syserror.EACCES
+	}
+
+	sem := s.findSem(num)
+	if sem == nil {
+		return 0, syserror.ERANGE
+	}
+	return sem.value, nil
+}
+
+// ExecuteOps attempts to execute a list of operations to the set. It only
+// suceeds when all operations can be applied. No changes are made if it fails.
+//
+// On failure, it may return an error (retries are hopeless) or it may return
+// a channel that can be waited on before attempting again.
+func (s *Set) ExecuteOps(ctx context.Context, ops []linux.Sembuf, creds *auth.Credentials) (chan struct{}, int32, error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Did it race with a removal operation?
+	if s.dead {
+		return nil, 0, syserror.EIDRM
+	}
+
+	// Validate the operations.
+	readOnly := true
+	for _, op := range ops {
+		if s.findSem(int32(op.SemNum)) == nil {
+			return nil, 0, syserror.EFBIG
+		}
+		if op.SemOp != 0 {
+			readOnly = false
+		}
+	}
+
+	if !s.checkPerms(creds, fs.PermMask{Read: readOnly, Write: !readOnly}) {
+		return nil, 0, syserror.EACCES
+	}
+
+	ch, num, err := s.executeOps(ctx, ops)
+	if err != nil {
+		return nil, 0, err
+	}
+	return ch, num, nil
+}
+
+func (s *Set) executeOps(ctx context.Context, ops []linux.Sembuf) (chan struct{}, int32, error) {
+	// Changes to semaphores go to this slice temporarily until they all succeed.
+	tmpVals := make([]int16, len(s.sems))
+	for i := range s.sems {
+		tmpVals[i] = s.sems[i].value
+	}
+
+	for _, op := range ops {
+		sem := &s.sems[op.SemNum]
+		if op.SemOp == 0 {
+			// Handle 'wait for zero' operation.
+			if tmpVals[op.SemNum] != 0 {
+				// Semaphore isn't 0, must wait.
+				if op.SemFlg&linux.IPC_NOWAIT != 0 {
+					return nil, 0, syserror.ErrWouldBlock
+				}
+
+				w := newWaiter(op.SemOp)
+				sem.waiters.PushBack(w)
+				return w.ch, int32(op.SemNum), nil
+			}
+		} else {
+			if op.SemOp < 0 {
+				// Handle 'wait' operation.
+				if -op.SemOp > valueMax {
+					return nil, 0, syserror.ERANGE
+				}
+				if -op.SemOp > tmpVals[op.SemNum] {
+					// Not enough resources, must wait.
+					if op.SemFlg&linux.IPC_NOWAIT != 0 {
+						return nil, 0, syserror.ErrWouldBlock
+					}
+
+					w := newWaiter(op.SemOp)
+					sem.waiters.PushBack(w)
+					return w.ch, int32(op.SemNum), nil
+				}
+			} else {
+				// op.SemOp > 0: Handle 'signal' operation.
+				if tmpVals[op.SemNum] > valueMax-op.SemOp {
+					return nil, 0, syserror.ERANGE
+				}
+			}
+
+			tmpVals[op.SemNum] += op.SemOp
+		}
+	}
+
+	// All operations succeeded, apply them.
+	// TODO: handle undo operations.
+	for i, v := range tmpVals {
+		s.sems[i].value = v
+		s.sems[i].wakeWaiters()
+	}
+	s.opTime = ktime.NowFromContext(ctx)
+	return nil, 0, nil
+}
+
+// AbortWait notifies that a waiter is giving up and will not wait on the
+// channel anymore.
+func (s *Set) AbortWait(num int32, ch chan struct{}) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	sem := &s.sems[num]
+	for w := sem.waiters.Front(); w != nil; w = w.Next() {
+		if w.ch == ch {
+			sem.waiters.Remove(w)
+			return
+		}
+	}
+	// Waiter may not be found in case it raced with wakeWaiters().
+}
+
+func (s *Set) checkCredentials(creds *auth.Credentials) bool {
+	return s.owner.UID == creds.EffectiveKUID ||
+		s.owner.GID == creds.EffectiveKGID ||
+		s.creator.UID == creds.EffectiveKUID ||
+		s.creator.GID == creds.EffectiveKGID
+}
+
+func (s *Set) checkCapability(creds *auth.Credentials) bool {
+	return creds.HasCapability(linux.CAP_IPC_OWNER) && creds.UserNamespace.MapFromKUID(s.owner.UID).Ok()
+}
+
+func (s *Set) checkPerms(creds *auth.Credentials, reqPerms fs.PermMask) bool {
+	// Are we owner, or in group, or other?
+	p := s.perms.Other
+	if s.owner.UID == creds.EffectiveKUID {
+		p = s.perms.User
+	} else if creds.InGroup(s.owner.GID) {
+		p = s.perms.Group
+	}
+
+	// Are permissions satisfied without capability checks?
+	if p.SupersetOf(reqPerms) {
+		return true
+	}
+
+	return s.checkCapability(creds)
+}
+
+func (s *Set) destroy() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// Notify all waiters. Tney will fail on the next attempt to execute
+	// operations and return error.
+	s.dead = true
+	for _, s := range s.sems {
+		for w := s.waiters.Front(); w != nil; w = w.Next() {
+			w.ch <- struct{}{}
+		}
+		s.waiters.Reset()
+	}
+}
+
+// wakeWaiters goes over all waiters and checks which of them can be notified.
+func (s *sem) wakeWaiters() {
+	// Note that this will release all waiters waiting for 0 too.
+	for w := s.waiters.Front(); w != nil; {
+		if s.value < w.value {
+			// Still blocked, skip it.
+			continue
+		}
+		w.ch <- struct{}{}
+		old := w
+		w = w.Next()
+		s.waiters.Remove(old)
+	}
+}
+
+func newWaiter(val int16) *waiter {
+	return &waiter{
+		value: val,
+		ch:    make(chan struct{}, 1),
+	}
+}
diff --git a/pkg/sentry/kernel/semaphore/semaphore_test.go b/pkg/sentry/kernel/semaphore/semaphore_test.go
new file mode 100644
index 000000000..0386586ab
--- /dev/null
+++ b/pkg/sentry/kernel/semaphore/semaphore_test.go
@@ -0,0 +1,172 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package semaphore
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func executeOps(ctx context.Context, t *testing.T, set *Set, ops []linux.Sembuf, block bool) chan struct{} {
+	ch, _, err := set.executeOps(ctx, ops)
+	if err != nil {
+		t.Fatalf("ExecuteOps(ops) failed, err: %v, ops: %+v", err, ops)
+	}
+	if block {
+		if ch == nil {
+			t.Fatalf("ExecuteOps(ops) got: nil, expected: !nil, ops: %+v", ops)
+		}
+		if signalled(ch) {
+			t.Fatalf("ExecuteOps(ops) channel should not have been signalled, ops: %+v", ops)
+		}
+	} else {
+		if ch != nil {
+			t.Fatalf("ExecuteOps(ops) got: %v, expected: nil, ops: %+v", ch, ops)
+		}
+	}
+	return ch
+}
+
+func signalled(ch chan struct{}) bool {
+	select {
+	case <-ch:
+		return true
+	default:
+		return false
+	}
+}
+
+func TestBasic(t *testing.T) {
+	ctx := contexttest.Context(t)
+	set := &Set{ID: 123, sems: make([]sem, 1)}
+	ops := []linux.Sembuf{
+		linux.Sembuf{SemOp: 1},
+	}
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = -1
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = -1
+	ch1 := executeOps(ctx, t, set, ops, true)
+
+	ops[0].SemOp = 1
+	executeOps(ctx, t, set, ops, false)
+	if !signalled(ch1) {
+		t.Fatalf("ExecuteOps(ops) channel should not have been signalled, ops: %+v", ops)
+	}
+}
+
+func TestWaitForZero(t *testing.T) {
+	ctx := contexttest.Context(t)
+	set := &Set{ID: 123, sems: make([]sem, 1)}
+	ops := []linux.Sembuf{
+		linux.Sembuf{SemOp: 0},
+	}
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = -2
+	ch1 := executeOps(ctx, t, set, ops, true)
+
+	ops[0].SemOp = 0
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = 1
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = 0
+	chZero1 := executeOps(ctx, t, set, ops, true)
+
+	ops[0].SemOp = 0
+	chZero2 := executeOps(ctx, t, set, ops, true)
+
+	ops[0].SemOp = 1
+	executeOps(ctx, t, set, ops, false)
+	if !signalled(ch1) {
+		t.Fatalf("ExecuteOps(ops) channel should have been signalled, ops: %+v, set: %+v", ops, set)
+	}
+
+	ops[0].SemOp = -2
+	executeOps(ctx, t, set, ops, false)
+	if !signalled(chZero1) {
+		t.Fatalf("ExecuteOps(ops) channel zero 1 should have been signalled, ops: %+v, set: %+v", ops, set)
+	}
+	if !signalled(chZero2) {
+		t.Fatalf("ExecuteOps(ops) channel zero 2 should have been signalled, ops: %+v, set: %+v", ops, set)
+	}
+}
+
+func TestNoWait(t *testing.T) {
+	ctx := contexttest.Context(t)
+	set := &Set{ID: 123, sems: make([]sem, 1)}
+	ops := []linux.Sembuf{
+		linux.Sembuf{SemOp: 1},
+	}
+	executeOps(ctx, t, set, ops, false)
+
+	ops[0].SemOp = -2
+	ops[0].SemFlg = linux.IPC_NOWAIT
+	if _, _, err := set.executeOps(ctx, ops); err != syserror.ErrWouldBlock {
+		t.Fatalf("ExecuteOps(ops) wrong result, got: %v, expected: %v", err, syserror.ErrWouldBlock)
+	}
+
+	ops[0].SemOp = 0
+	ops[0].SemFlg = linux.IPC_NOWAIT
+	if _, _, err := set.executeOps(ctx, ops); err != syserror.ErrWouldBlock {
+		t.Fatalf("ExecuteOps(ops) wrong result, got: %v, expected: %v", err, syserror.ErrWouldBlock)
+	}
+}
+
+func TestUnregister(t *testing.T) {
+	ctx := contexttest.Context(t)
+	r := NewRegistry()
+	set, err := r.FindOrCreate(ctx, 123, 2, linux.FileMode(0x600), true, true, true)
+	if err != nil {
+		t.Fatalf("FindOrCreate() failed, err: %v", err)
+	}
+	if got := r.FindByID(set.ID); got.ID != set.ID {
+		t.Fatalf("FindById(%d) failed, got: %+v, expected: %+v", set.ID, got, set)
+	}
+
+	ops := []linux.Sembuf{
+		linux.Sembuf{SemOp: -1},
+	}
+	chs := make([]chan struct{}, 0, 5)
+	for i := 0; i < 5; i++ {
+		ch := executeOps(ctx, t, set, ops, true)
+		chs = append(chs, ch)
+	}
+
+	creds := auth.CredentialsFromContext(ctx)
+	if err := r.RemoveID(set.ID, creds); err != nil {
+		t.Fatalf("RemoveID(%d) failed, err: %v", set.ID, err)
+	}
+	if !set.dead {
+		t.Fatalf("set is not dead: %+v", set)
+	}
+	if got := r.FindByID(set.ID); got != nil {
+		t.Fatalf("FindById(%d) failed, got: %+v, expected: nil", set.ID, got)
+	}
+	for i, ch := range chs {
+		if !signalled(ch) {
+			t.Fatalf("channel %d should have been signalled", i)
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go
new file mode 100644
index 000000000..53d8fb844
--- /dev/null
+++ b/pkg/sentry/kernel/sessions.go
@@ -0,0 +1,462 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SessionID is the public identifier.
+type SessionID ThreadID
+
+// ProcessGroupID is the public identifier.
+type ProcessGroupID ThreadID
+
+// Session contains a leader threadgroup and a list of ProcessGroups.
+type Session struct {
+	refs refs.AtomicRefCount
+
+	// leader is the originator of the Session.
+	//
+	// Note that this may no longer be running (and may be reaped), so the
+	// ID is cached upon initial creation. The leader is still required
+	// however, since its PIDNamespace defines the scope of the Session.
+	//
+	// The leader is immutable.
+	leader *ThreadGroup
+
+	// id is the cached identifier in the leader's namespace.
+	//
+	// The id is immutable.
+	id SessionID
+
+	// ProcessGroups is a list of process groups in this Session. This is
+	// protected by TaskSet.mu.
+	processGroups processGroupList
+
+	// sessionEntry is the embed for TaskSet.sessions. This is protected by
+	// TaskSet.mu.
+	sessionEntry
+}
+
+// incRef grabs a reference.
+func (s *Session) incRef() {
+	s.refs.IncRef()
+}
+
+// decRef drops a reference.
+//
+// Precondition: callers must hold TaskSet.mu for writing.
+func (s *Session) decRef() {
+	s.refs.DecRefWithDestructor(func() {
+		// Remove translations from the leader.
+		for ns := s.leader.pidns; ns != nil; ns = ns.parent {
+			id := ns.sids[s]
+			delete(ns.sids, s)
+			delete(ns.sessions, id)
+		}
+
+		// Remove from the list of global Sessions.
+		s.leader.pidns.owner.sessions.Remove(s)
+	})
+}
+
+// ProcessGroup contains an originator threadgroup and a parent Session.
+type ProcessGroup struct {
+	refs refs.AtomicRefCount // not exported.
+
+	// originator is the originator of the group.
+	//
+	// See note re: leader in Session. The same applies here.
+	//
+	// The originator is immutable.
+	originator *ThreadGroup
+
+	// id is the cached identifier in the originator's namespace.
+	//
+	// The id is immutable.
+	id ProcessGroupID
+
+	// Session is the parent Session.
+	//
+	// The session is immutable.
+	session *Session
+
+	// ancestors is the number of thread groups in this process group whose
+	// parent is in a different process group in the same session.
+	//
+	// The name is derived from the fact that process groups where
+	// ancestors is zero are considered "orphans".
+	//
+	// ancestors is protected by TaskSet.mu.
+	ancestors uint32
+
+	// processGroupEntry is the embedded entry for Sessions.groups. This is
+	// protected by TaskSet.mu.
+	processGroupEntry
+}
+
+// incRefWithParent grabs a reference.
+//
+// This function is called when this ProcessGroup is being associated with some
+// new ThreadGroup, tg. parentPG is the ProcessGroup of tg's parent
+// ThreadGroup. If tg is init, then parentPG may be nil.
+//
+// Precondition: callers must hold TaskSet.mu for writing.
+func (pg *ProcessGroup) incRefWithParent(parentPG *ProcessGroup) {
+	// We acquire an "ancestor" reference in the case of a nil parent.
+	// This is because the process being associated is init, and init can
+	// never be orphaned (we count it as always having an ancestor).
+	if pg != parentPG && (parentPG == nil || pg.session == parentPG.session) {
+		pg.ancestors++
+	}
+
+	pg.refs.IncRef()
+}
+
+// decRefWithParent drops a reference.
+//
+// parentPG is per incRefWithParent.
+//
+// Precondition: callers must hold TaskSet.mu for writing.
+func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) {
+	// See incRefWithParent regarding parent == nil.
+	if pg != parentPG && (parentPG == nil || pg.session == parentPG.session) {
+		pg.ancestors--
+	}
+
+	alive := true
+	pg.refs.DecRefWithDestructor(func() {
+		alive = false // don't bother with handleOrphan.
+
+		// Remove translations from the originator.
+		for ns := pg.originator.pidns; ns != nil; ns = ns.parent {
+			id := ns.pgids[pg]
+			delete(ns.pgids, pg)
+			delete(ns.processGroups, id)
+		}
+
+		// Remove the list of process groups.
+		pg.session.processGroups.Remove(pg)
+		pg.session.decRef()
+	})
+	if alive {
+		pg.handleOrphan()
+	}
+}
+
+// parentPG returns the parent process group.
+//
+// Precondition: callers must hold TaskSet.mu.
+func (tg *ThreadGroup) parentPG() *ProcessGroup {
+	if tg.leader.parent != nil {
+		return tg.leader.parent.tg.processGroup
+	}
+	return nil
+}
+
+// handleOrphan checks whether the process group is an orphan and has any
+// stopped jobs. If yes, then appropriate signals are delivered to each thread
+// group within the process group.
+//
+// Precondition: callers must hold TaskSet.mu for writing.
+func (pg *ProcessGroup) handleOrphan() {
+	// Check if this process is an orphan.
+	if pg.ancestors != 0 {
+		return
+	}
+
+	// See if there are any stopped jobs.
+	hasStopped := false
+	pg.originator.pidns.owner.forEachThreadGroupLocked(func(tg *ThreadGroup) {
+		if tg.processGroup != pg {
+			return
+		}
+		tg.signalHandlers.mu.Lock()
+		if tg.groupStopPhase == groupStopComplete {
+			hasStopped = true
+		}
+		tg.signalHandlers.mu.Unlock()
+	})
+	if !hasStopped {
+		return
+	}
+
+	// Deliver appropriate signals to all thread groups.
+	pg.originator.pidns.owner.forEachThreadGroupLocked(func(tg *ThreadGroup) {
+		if tg.processGroup != pg {
+			return
+		}
+		tg.signalHandlers.mu.Lock()
+		tg.leader.sendSignalLocked(sigPriv(linux.SIGHUP), true /* group */)
+		tg.leader.sendSignalLocked(sigPriv(linux.SIGCONT), true /* group */)
+		tg.signalHandlers.mu.Unlock()
+	})
+
+	return
+}
+
+// CreateSession creates a new Session, with the ThreadGroup as the leader.
+//
+// EPERM may be returned if either the given ThreadGroup is already a Session
+// leader, or a ProcessGroup already exists for the ThreadGroup's ID.
+func (tg *ThreadGroup) CreateSession() error {
+	tg.pidns.owner.mu.Lock()
+	defer tg.pidns.owner.mu.Unlock()
+	return tg.createSession()
+}
+
+// createSession creates a new session for a threadgroup.
+//
+// Precondition: callers must hold TaskSet.mu for writing.
+func (tg *ThreadGroup) createSession() error {
+	// Get the ID for this thread in the current namespace.
+	id := tg.pidns.tids[tg.leader]
+
+	// Check if this ThreadGroup already leads a Session, or
+	// if the proposed group is already taken.
+	for s := tg.pidns.owner.sessions.Front(); s != nil; s = s.Next() {
+		if s.leader.pidns != tg.pidns {
+			continue
+		}
+		if s.leader == tg {
+			return syserror.EPERM
+		}
+		if s.id == SessionID(id) {
+			return syserror.EPERM
+		}
+		for pg := s.processGroups.Front(); pg != nil; pg = pg.Next() {
+			if pg.id == ProcessGroupID(id) {
+				return syserror.EPERM
+			}
+		}
+	}
+
+	// Create a new Session, with a single reference.
+	s := &Session{
+		id:     SessionID(id),
+		leader: tg,
+	}
+
+	// Create a new ProcessGroup, belonging to that Session.
+	// This also has a single reference (assigned below).
+	//
+	// Note that since this is a new session and a new process group, there
+	// will be zero ancestors for this process group. (It is an orphan at
+	// this point.)
+	pg := &ProcessGroup{
+		id:         ProcessGroupID(id),
+		originator: tg,
+		session:    s,
+		ancestors:  0,
+	}
+
+	// Tie them and return the result.
+	s.processGroups.PushBack(pg)
+	tg.pidns.owner.sessions.PushBack(s)
+
+	// Leave the current group, and assign the new one.
+	if tg.processGroup != nil {
+		oldParentPG := tg.parentPG()
+		tg.forEachChildThreadGroupLocked(func(childTG *ThreadGroup) {
+			childTG.processGroup.incRefWithParent(pg)
+			childTG.processGroup.decRefWithParent(oldParentPG)
+		})
+		tg.processGroup.decRefWithParent(oldParentPG)
+		tg.processGroup = pg
+	} else {
+		// The current process group may be nil only in the case of an
+		// unparented thread group (i.e. the init process). This would
+		// not normally occur, but we allow it for the convenience of
+		// CreateSession working from that point. There will be no
+		// child processes. We always say that the very first group
+		// created has ancestors (avoids checks elsewhere).
+		//
+		// Note that this mirrors the parent == nil logic in
+		// incRef/decRef/reparent, which counts nil as an ancestor.
+		tg.processGroup = pg
+		tg.processGroup.ancestors++
+	}
+
+	// Ensure a translation is added to all namespaces.
+	for ns := tg.pidns; ns != nil; ns = ns.parent {
+		local := ns.tids[tg.leader]
+		ns.sids[s] = SessionID(local)
+		ns.sessions[SessionID(local)] = s
+		ns.pgids[pg] = ProcessGroupID(local)
+		ns.processGroups[ProcessGroupID(local)] = pg
+	}
+
+	return nil
+}
+
+// CreateProcessGroup creates a new process group.
+//
+// An EPERM error will be returned if the ThreadGroup belongs to a different
+// Session, is a Session leader or the group already exists.
+func (tg *ThreadGroup) CreateProcessGroup() error {
+	tg.pidns.owner.mu.Lock()
+	defer tg.pidns.owner.mu.Unlock()
+
+	// Get the ID for this thread in the current namespace.
+	id := tg.pidns.tids[tg.leader]
+
+	// Per above, check for a Session leader or existing group.
+	for s := tg.pidns.owner.sessions.Front(); s != nil; s = s.Next() {
+		if s.leader.pidns != tg.pidns {
+			continue
+		}
+		if s.leader == tg {
+			return syserror.EPERM
+		}
+		for pg := s.processGroups.Front(); pg != nil; pg = pg.Next() {
+			if pg.id == ProcessGroupID(id) {
+				return syserror.EPERM
+			}
+		}
+	}
+
+	// Create a new ProcessGroup, belonging to the current Session.
+	//
+	// We manually adjust the ancestors if the parent is in the same
+	// session.
+	tg.processGroup.session.incRef()
+	pg := &ProcessGroup{
+		id:         ProcessGroupID(id),
+		originator: tg,
+		session:    tg.processGroup.session,
+	}
+	if tg.leader.parent != nil && tg.leader.parent.tg.processGroup.session == pg.session {
+		pg.ancestors++
+	}
+
+	// Assign the new process group; adjust children.
+	oldParentPG := tg.parentPG()
+	tg.forEachChildThreadGroupLocked(func(childTG *ThreadGroup) {
+		childTG.processGroup.incRefWithParent(pg)
+		childTG.processGroup.decRefWithParent(oldParentPG)
+	})
+	tg.processGroup.decRefWithParent(oldParentPG)
+	tg.processGroup = pg
+
+	// Ensure this translation is added to all namespaces.
+	for ns := tg.pidns; ns != nil; ns = ns.parent {
+		local := ns.tids[tg.leader]
+		ns.pgids[pg] = ProcessGroupID(local)
+		ns.processGroups[ProcessGroupID(local)] = pg
+	}
+
+	return nil
+}
+
+// JoinProcessGroup joins an existing process group.
+//
+// This function will return EACCES if an exec has been performed since fork
+// by the given ThreadGroup, and EPERM if the Sessions are not the same or the
+// group does not exist.
+//
+// If checkExec is set, then the join is not permitted after the process has
+// executed exec at least once.
+func (tg *ThreadGroup) JoinProcessGroup(pidns *PIDNamespace, pgid ProcessGroupID, checkExec bool) error {
+	pidns.owner.mu.Lock()
+	defer pidns.owner.mu.Unlock()
+
+	// Lookup the ProcessGroup.
+	pg := pidns.processGroups[pgid]
+	if pg == nil {
+		return syserror.EPERM
+	}
+
+	// Disallow the join if an execve has performed, per POSIX.
+	if checkExec && tg.execed {
+		return syserror.EACCES
+	}
+
+	// See if it's in the same session as ours.
+	if pg.session != tg.processGroup.session {
+		return syserror.EPERM
+	}
+
+	// Join the group; adjust children.
+	parentPG := tg.parentPG()
+	pg.incRefWithParent(parentPG)
+	tg.forEachChildThreadGroupLocked(func(childTG *ThreadGroup) {
+		childTG.processGroup.incRefWithParent(pg)
+		childTG.processGroup.decRefWithParent(tg.processGroup)
+	})
+	tg.processGroup.decRefWithParent(parentPG)
+	tg.processGroup = pg
+
+	return nil
+}
+
+// Session returns the ThreadGroup's Session.
+//
+// A reference is not taken on the session.
+func (tg *ThreadGroup) Session() *Session {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.processGroup.session
+}
+
+// IDOfSession returns the Session assigned to s in PID namespace ns.
+//
+// If this group isn't visible in this namespace, zero will be returned. It is
+// the callers responsibility to check that before using this function.
+func (pidns *PIDNamespace) IDOfSession(s *Session) SessionID {
+	pidns.owner.mu.RLock()
+	defer pidns.owner.mu.RUnlock()
+	return pidns.sids[s]
+}
+
+// SessionWithID returns the Session with the given ID in the PID namespace ns,
+// or nil if that given ID is not defined in this namespace.
+//
+// A reference is not taken on the session.
+func (pidns *PIDNamespace) SessionWithID(id SessionID) *Session {
+	pidns.owner.mu.RLock()
+	defer pidns.owner.mu.RUnlock()
+	return pidns.sessions[id]
+}
+
+// ProcessGroup returns the ThreadGroup's ProcessGroup.
+//
+// A reference is not taken on the process group.
+func (tg *ThreadGroup) ProcessGroup() *ProcessGroup {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.processGroup
+}
+
+// IDOfProcessGroup returns the process group assigned to pg in PID namespace ns.
+//
+// The same constraints apply as IDOfSession.
+func (pidns *PIDNamespace) IDOfProcessGroup(pg *ProcessGroup) ProcessGroupID {
+	pidns.owner.mu.RLock()
+	defer pidns.owner.mu.RUnlock()
+	return pidns.pgids[pg]
+}
+
+// ProcessGroupWithID returns the ProcessGroup with the given ID in the PID
+// namespace ns, or nil if that given ID is not defined in this namespace.
+//
+// A reference is not taken on the process group.
+func (pidns *PIDNamespace) ProcessGroupWithID(id ProcessGroupID) *ProcessGroup {
+	pidns.owner.mu.RLock()
+	defer pidns.owner.mu.RUnlock()
+	return pidns.processGroups[id]
+}
diff --git a/pkg/sentry/kernel/signal.go b/pkg/sentry/kernel/signal.go
new file mode 100644
index 000000000..8edd05cdf
--- /dev/null
+++ b/pkg/sentry/kernel/signal.go
@@ -0,0 +1,69 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+)
+
+// SignalPanic is used to panic the running threads. It is a signal which
+// cannot be used by the application: it must be caught and ignored by the
+// runtime (in order to catch possible races).
+const SignalPanic = linux.SIGUSR2
+
+// sendExternalSignal is called when an asynchronous signal is sent to the
+// sentry ("in sentry context"). On some platforms, it may also be called when
+// an asynchronous signal is sent to sandboxed application threads ("in
+// application context").
+//
+// context is used only for debugging to differentiate these cases.
+//
+// Returns false if signal could not be sent because the Kernel is not fully
+// initialized yet.
+func (k *Kernel) sendExternalSignal(info *arch.SignalInfo, context string) bool {
+	switch linux.Signal(info.Signo) {
+	case platform.SignalInterrupt:
+		// Assume that a call to platform.Context.Interrupt() misfired.
+		return true
+
+	case SignalPanic:
+		// SignalPanic is also specially handled in sentry setup to ensure that
+		// it causes a panic even after tasks exit, but SignalPanic may also
+		// be sent here if it is received while in app context.
+		panic("Signal-induced panic")
+
+	default:
+		log.Infof("Received external signal %d in %s context", info.Signo, context)
+		if k.globalInit == nil {
+			log.Warningf("Received external signal %d before init created", info.Signo)
+			return false
+		}
+		k.globalInit.SendSignal(info)
+	}
+
+	return true
+}
+
+// sigPriv returns a SignalInfo representing a signal sent by the sentry. (The
+// name reflects its equivalence to Linux's SEND_SIG_PRIV.)
+func sigPriv(sig linux.Signal) *arch.SignalInfo {
+	return &arch.SignalInfo{
+		Signo: int32(sig),
+		Code:  arch.SignalInfoKernel,
+	}
+}
diff --git a/pkg/sentry/kernel/signal_handlers.go b/pkg/sentry/kernel/signal_handlers.go
new file mode 100644
index 000000000..21ba4ee70
--- /dev/null
+++ b/pkg/sentry/kernel/signal_handlers.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+)
+
+// SignalHandlers holds information about signal actions.
+type SignalHandlers struct {
+	// mu protects actions, as well as the signal state of all tasks and thread
+	// groups using this SignalHandlers object. (See comment on
+	// ThreadGroup.signalHandlers.)
+	mu sync.Mutex `state:"nosave"`
+
+	// actions is the action to be taken upon receiving each signal.
+	actions map[linux.Signal]arch.SignalAct
+}
+
+// NewSignalHandlers returns a new SignalHandlers specifying all default
+// actions.
+func NewSignalHandlers() *SignalHandlers {
+	return &SignalHandlers{
+		actions: make(map[linux.Signal]arch.SignalAct),
+	}
+}
+
+// Fork returns a copy of sh for a new thread group.
+func (sh *SignalHandlers) Fork() *SignalHandlers {
+	sh2 := NewSignalHandlers()
+	sh.mu.Lock()
+	defer sh.mu.Unlock()
+	for sig, act := range sh.actions {
+		sh2.actions[sig] = act
+	}
+	return sh2
+}
+
+// CopyForExec returns a copy of sh for a thread group that is undergoing an
+// execve. (See comments in Task.finishExec.)
+func (sh *SignalHandlers) CopyForExec() *SignalHandlers {
+	sh2 := NewSignalHandlers()
+	sh.mu.Lock()
+	defer sh.mu.Unlock()
+	for sig, act := range sh.actions {
+		if act.Handler == arch.SignalActIgnore {
+			sh2.actions[sig] = arch.SignalAct{
+				Handler: arch.SignalActIgnore,
+			}
+		}
+	}
+	return sh2
+}
+
+// dequeueActionLocked returns the SignalAct that should be used to handle sig.
+//
+// Preconditions: sh.mu must be locked.
+func (sh *SignalHandlers) dequeueAction(sig linux.Signal) arch.SignalAct {
+	act := sh.actions[sig]
+	if act.IsResetHandler() {
+		delete(sh.actions, sig)
+	}
+	return act
+}
diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go
new file mode 100644
index 000000000..e20fa3eb6
--- /dev/null
+++ b/pkg/sentry/kernel/syscalls.go
@@ -0,0 +1,305 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// maxSyscallNum is the highest supported syscall number.
+//
+// The types below create fast lookup slices for all syscalls. This maximum
+// serves as a sanity check that we don't allocate huge slices for a very large
+// syscall.
+const maxSyscallNum = 2000
+
+// SyscallFn is a syscall implementation.
+type SyscallFn func(t *Task, args arch.SyscallArguments) (uintptr, *SyscallControl, error)
+
+// MissingFn is a syscall to be called when an implementation is missing.
+type MissingFn func(t *Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error)
+
+// Possible flags for SyscallFlagsTable.enable.
+const (
+	// syscallPresent indicates that this is not a missing syscall.
+	//
+	// This flag is used internally in SyscallFlagsTable.
+	syscallPresent = 1 << iota
+
+	// StraceEnableLog enables syscall log tracing.
+	StraceEnableLog
+
+	// StraceEnableEvent enables syscall event tracing.
+	StraceEnableEvent
+
+	// ExternalBeforeEnable enables the external hook before syscall execution.
+	ExternalBeforeEnable
+
+	// ExternalAfterEnable enables the external hook after syscall execution.
+	ExternalAfterEnable
+)
+
+// StraceEnableBits combines both strace log and event flags.
+const StraceEnableBits = StraceEnableLog | StraceEnableEvent
+
+// SyscallFlagsTable manages a set of enable/disable bit fields on a per-syscall
+// basis.
+type SyscallFlagsTable struct {
+	// mu protects writes to the fields below.
+	//
+	// Atomic loads are always allowed. Atomic stores are allowed only
+	// while mu is held.
+	mu sync.Mutex
+
+	// enable contains the enable bits for each syscall.
+	//
+	// missing syscalls have the same value in enable as missingEnable to
+	// avoid an extra branch in Word.
+	enable []uint32
+
+	// missingEnable contains the enable bits for missing syscalls.
+	missingEnable uint32
+}
+
+// Init initializes the struct, with all syscalls in table set to enable.
+//
+// max is the largest syscall number in table.
+func (e *SyscallFlagsTable) init(table map[uintptr]SyscallFn, max uintptr) {
+	e.enable = make([]uint32, max+1)
+	for num := range table {
+		e.enable[num] = syscallPresent
+	}
+}
+
+// Word returns the enable bitfield for sysno.
+func (e *SyscallFlagsTable) Word(sysno uintptr) uint32 {
+	if sysno < uintptr(len(e.enable)) {
+		return atomic.LoadUint32(&e.enable[sysno])
+	}
+
+	return atomic.LoadUint32(&e.missingEnable)
+}
+
+// Enable sets enable bit bit for all syscalls based on s.
+//
+// Syscalls missing from s are disabled.
+//
+// Syscalls missing from the initial table passed to Init cannot be added as
+// individual syscalls. If present in s they will be ignored.
+//
+// Callers to Word may see either the old or new value while this function
+// is executing.
+func (e *SyscallFlagsTable) Enable(bit uint32, s map[uintptr]bool, missingEnable bool) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	missingVal := atomic.LoadUint32(&e.missingEnable)
+	if missingEnable {
+		missingVal |= bit
+	} else {
+		missingVal &^= bit
+	}
+	atomic.StoreUint32(&e.missingEnable, missingVal)
+
+	for num := range e.enable {
+		val := atomic.LoadUint32(&e.enable[num])
+		if !bits.IsOn32(val, syscallPresent) {
+			// Missing.
+			atomic.StoreUint32(&e.enable[num], missingVal)
+			continue
+		}
+
+		if s[uintptr(num)] {
+			val |= bit
+		} else {
+			val &^= bit
+		}
+		atomic.StoreUint32(&e.enable[num], val)
+	}
+}
+
+// EnableAll sets enable bit bit for all syscalls, present and missing.
+func (e *SyscallFlagsTable) EnableAll(bit uint32) {
+	e.mu.Lock()
+	defer e.mu.Unlock()
+
+	missingVal := atomic.LoadUint32(&e.missingEnable)
+	missingVal |= bit
+	atomic.StoreUint32(&e.missingEnable, missingVal)
+
+	for num := range e.enable {
+		val := atomic.LoadUint32(&e.enable[num])
+		if !bits.IsOn32(val, syscallPresent) {
+			// Missing.
+			atomic.StoreUint32(&e.enable[num], missingVal)
+			continue
+		}
+
+		val |= bit
+		atomic.StoreUint32(&e.enable[num], val)
+	}
+}
+
+// Stracer traces syscall execution.
+type Stracer interface {
+	// SyscallEnter is called on syscall entry.
+	//
+	// The returned private data is passed to SyscallExit.
+	//
+	// TODO: remove kernel imports from the strace package so
+	// that the type can be used directly.
+	SyscallEnter(t *Task, sysno uintptr, args arch.SyscallArguments, flags uint32) interface{}
+
+	// SyscallExit is called on syscall exit.
+	SyscallExit(context interface{}, t *Task, sysno, rval uintptr, err error)
+}
+
+// SyscallTable is a lookup table of system calls. Critically, a SyscallTable
+// is *immutable*. In order to make supporting suspend and resume sane, they
+// must be uniquely registered and may not change during operation.
+type SyscallTable struct {
+	// OS is the operating system that this syscall table implements.
+	OS abi.OS `state:"wait"`
+
+	// Arch is the architecture that this syscall table targets.
+	Arch arch.Arch `state:"wait"`
+
+	// The OS version that this syscall table implements.
+	Version Version `state:"manual"`
+
+	// AuditNumber is a numeric constant that represents the syscall table. If
+	// non-zero, auditNumber must be one of the AUDIT_ARCH_* values defined by
+	// linux/audit.h.
+	AuditNumber uint32 `state:"manual"`
+
+	// Table is the collection of functions.
+	Table map[uintptr]SyscallFn `state:"manual"`
+
+	// lookup is a fixed-size array that holds the syscalls (indexed by
+	// their numbers). It is used for fast look ups.
+	lookup []SyscallFn `state:"manual"`
+
+	// Emulate is a collection of instruction addresses to emulate. The
+	// keys are addresses, and the values are system call numbers.
+	Emulate map[usermem.Addr]uintptr `state:"manual"`
+
+	// The function to call in case of a missing system call.
+	Missing MissingFn `state:"manual"`
+
+	// Stracer traces this syscall table.
+	Stracer Stracer `state:"manual"`
+
+	// External is used to handle an external callback.
+	External func(*Kernel) `state:"manual"`
+
+	// ExternalFilterBefore is called before External is called before the syscall is executed.
+	// External is not called if it returns false.
+	ExternalFilterBefore func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"`
+
+	// ExternalFilterAfter is called before External is called after the syscall is executed.
+	// External is not called if it returns false.
+	ExternalFilterAfter func(*Task, uintptr, arch.SyscallArguments) bool `state:"manual"`
+
+	// FeatureEnable stores the strace and one-shot enable bits.
+	FeatureEnable SyscallFlagsTable `state:"manual"`
+}
+
+// allSyscallTables contains all known tables.
+var allSyscallTables []*SyscallTable
+
+// SyscallTables returns a read-only slice of registered SyscallTables.
+func SyscallTables() []*SyscallTable {
+	return allSyscallTables
+}
+
+// LookupSyscallTable returns the SyscallCall table for the OS/Arch combination.
+func LookupSyscallTable(os abi.OS, a arch.Arch) (*SyscallTable, bool) {
+	for _, s := range allSyscallTables {
+		if s.OS == os && s.Arch == a {
+			return s, true
+		}
+	}
+	return nil, false
+}
+
+// RegisterSyscallTable registers a new syscall table for use by a Kernel.
+func RegisterSyscallTable(s *SyscallTable) {
+	if s.Table == nil {
+		// Ensure non-nil lookup table.
+		s.Table = make(map[uintptr]SyscallFn)
+	}
+	if s.Emulate == nil {
+		// Ensure non-nil emulate table.
+		s.Emulate = make(map[usermem.Addr]uintptr)
+	}
+
+	var max uintptr
+	for num := range s.Table {
+		if num > max {
+			max = num
+		}
+	}
+
+	if max > maxSyscallNum {
+		panic(fmt.Sprintf("SyscallTable %+v contains too large syscall number %d", s, max))
+	}
+
+	s.lookup = make([]SyscallFn, max+1)
+
+	// Initialize the fast-lookup table.
+	for num, fn := range s.Table {
+		s.lookup[num] = fn
+	}
+
+	s.FeatureEnable.init(s.Table, max)
+
+	if _, ok := LookupSyscallTable(s.OS, s.Arch); ok {
+		panic(fmt.Sprintf("Duplicate SyscallTable registered for OS %v Arch %v", s.OS, s.Arch))
+	}
+
+	// Save a reference to this table.
+	//
+	// This is required for a Kernel to find the table and for save/restore
+	// operations below.
+	allSyscallTables = append(allSyscallTables, s)
+}
+
+// Lookup returns the syscall implementation, if one exists.
+func (s *SyscallTable) Lookup(sysno uintptr) SyscallFn {
+	if sysno < uintptr(len(s.lookup)) {
+		return s.lookup[sysno]
+	}
+
+	return nil
+}
+
+// LookupEmulate looks up an emulation syscall number.
+func (s *SyscallTable) LookupEmulate(addr usermem.Addr) (uintptr, bool) {
+	sysno, ok := s.Emulate[addr]
+	return sysno, ok
+}
+
+// mapLookup is similar to Lookup, except that it only uses the syscall table,
+// that is, it skips the fast look array. This is available for benchmarking.
+func (s *SyscallTable) mapLookup(sysno uintptr) SyscallFn {
+	return s.Table[sysno]
+}
diff --git a/pkg/sentry/kernel/syscalls_state.go b/pkg/sentry/kernel/syscalls_state.go
new file mode 100644
index 000000000..826809a70
--- /dev/null
+++ b/pkg/sentry/kernel/syscalls_state.go
@@ -0,0 +1,29 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import "fmt"
+
+// afterLoad is invoked by stateify.
+func (s *SyscallTable) afterLoad() {
+	otherTable, ok := LookupSyscallTable(s.OS, s.Arch)
+	if !ok {
+		// Couldn't find a reference?
+		panic(fmt.Sprintf("syscall table not found for OS %v Arch %v", s.OS, s.Arch))
+	}
+
+	// Copy the table.
+	*s = *otherTable
+}
diff --git a/pkg/sentry/kernel/syslog.go b/pkg/sentry/kernel/syslog.go
new file mode 100644
index 000000000..31541749e
--- /dev/null
+++ b/pkg/sentry/kernel/syslog.go
@@ -0,0 +1,100 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"math/rand"
+	"sync"
+)
+
+// syslog represents a sentry-global kernel log.
+//
+// Currently, it contains only fun messages for a dmesg easter egg.
+type syslog struct {
+	// mu protects the below.
+	mu sync.Mutex `state:"nosave"`
+
+	// msg is the syslog message buffer. It is lazily initialized.
+	msg []byte
+}
+
+// Log returns a copy of the syslog.
+func (s *syslog) Log() []byte {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if s.msg != nil {
+		// Already initialized, just return a copy.
+		o := make([]byte, len(s.msg))
+		copy(o, s.msg)
+		return o
+	}
+
+	// Not initialized, create message.
+	allMessages := []string{
+		"Synthesizing system calls...",
+		"Mounting deweydecimalfs...",
+		"Moving files to filing cabinet...",
+		"Digging up root...",
+		"Constructing home...",
+		"Segmenting fault lines...",
+		"Creating bureaucratic processes...",
+		"Searching for needles in stacks...",
+		"Preparing for the zombie uprising...",
+		"Feeding the init monster...",
+		"Creating cloned children...",
+		"Daemonizing children...",
+		"Waiting for children...",
+		"Gathering forks...",
+		"Committing treasure map to memory...",
+		"Reading process obituaries...",
+		"Searching for socket adapter...",
+		"Creating process schedule...",
+		"Generating random numbers by fair dice roll...",
+		"Rewriting operating system in Javascript...",
+		"Consulting tar man page...",
+		"Forking spaghetti code...",
+		"Checking naughty and nice process list...",
+		"Checking naughty and nice process list...", // Check it up to twice.
+		"Granting licence to kill(2)...",            // British spelling for British movie.
+		"Letting the watchdogs out...",
+	}
+
+	selectMessage := func() string {
+		i := rand.Intn(len(allMessages))
+		m := allMessages[i]
+
+		// Delete the selected message.
+		allMessages[i] = allMessages[len(allMessages)-1]
+		allMessages = allMessages[:len(allMessages)-1]
+
+		return m
+	}
+
+	time := 0.0
+	for i := 0; i < 10; i++ {
+		time += rand.Float64() / 2
+		s.msg = append(s.msg, []byte(fmt.Sprintf("<6>[%11.6f] %s\n", time, selectMessage()))...)
+	}
+
+	time += rand.Float64() / 2
+	s.msg = append(s.msg, []byte(fmt.Sprintf("<6>[%11.6f] Ready!\n", time))...)
+
+	// Return a copy.
+	o := make([]byte, len(s.msg))
+	copy(o, s.msg)
+	return o
+}
diff --git a/pkg/sentry/kernel/table_test.go b/pkg/sentry/kernel/table_test.go
new file mode 100644
index 000000000..71ca75555
--- /dev/null
+++ b/pkg/sentry/kernel/table_test.go
@@ -0,0 +1,108 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+)
+
+const (
+	maxTestSyscall = 1000
+)
+
+func createSyscallTable() *SyscallTable {
+	m := make(map[uintptr]SyscallFn)
+	for i := uintptr(0); i <= maxTestSyscall; i++ {
+		j := i
+		m[i] = func(*Task, arch.SyscallArguments) (uintptr, *SyscallControl, error) {
+			return j, nil, nil
+		}
+	}
+
+	s := &SyscallTable{
+		OS:    abi.Linux,
+		Arch:  arch.AMD64,
+		Table: m,
+	}
+
+	RegisterSyscallTable(s)
+	return s
+}
+
+func TestTable(t *testing.T) {
+	table := createSyscallTable()
+	defer func() {
+		// Cleanup registered tables to keep tests separate.
+		allSyscallTables = []*SyscallTable{}
+	}()
+
+	// Go through all functions and check that they return the right value.
+	for i := uintptr(0); i < maxTestSyscall; i++ {
+		fn := table.Lookup(i)
+		if fn == nil {
+			t.Errorf("Syscall %v is set to nil", i)
+			continue
+		}
+
+		v, _, _ := fn(nil, arch.SyscallArguments{})
+		if v != i {
+			t.Errorf("Wrong return value for syscall %v: expected %v, got %v", i, i, v)
+		}
+	}
+
+	// Check that values outside the range return nil.
+	for i := uintptr(maxTestSyscall + 1); i < maxTestSyscall+100; i++ {
+		fn := table.Lookup(i)
+		if fn != nil {
+			t.Errorf("Syscall %v is not nil: %v", i, fn)
+			continue
+		}
+	}
+}
+
+func BenchmarkTableLookup(b *testing.B) {
+	table := createSyscallTable()
+
+	b.ResetTimer()
+
+	j := uintptr(0)
+	for i := 0; i < b.N; i++ {
+		table.Lookup(j)
+		j = (j + 1) % 310
+	}
+
+	b.StopTimer()
+	// Cleanup registered tables to keep tests separate.
+	allSyscallTables = []*SyscallTable{}
+}
+
+func BenchmarkTableMapLookup(b *testing.B) {
+	table := createSyscallTable()
+
+	b.ResetTimer()
+
+	j := uintptr(0)
+	for i := 0; i < b.N; i++ {
+		table.mapLookup(j)
+		j = (j + 1) % 310
+	}
+
+	b.StopTimer()
+	// Cleanup registered tables to keep tests separate.
+	allSyscallTables = []*SyscallTable{}
+}
diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go
new file mode 100644
index 000000000..3d2e035e9
--- /dev/null
+++ b/pkg/sentry/kernel/task.go
@@ -0,0 +1,606 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bpf"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/futex"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	ssync "gvisor.googlesource.com/gvisor/pkg/sync"
+)
+
+// Task represents a thread of execution in the untrusted app.  It
+// includes registers and any thread-specific state that you would
+// normally expect.
+//
+// Each task is associated with a goroutine, called the task goroutine, that
+// executes code (application code, system calls, etc.) on behalf of that task.
+// See Task.run (task_run.go).
+//
+// All fields that are "owned by the task goroutine" can only be mutated by the
+// task goroutine while it is running. The task goroutine does not require
+// synchronization to read these fields, although it still requires
+// synchronization as described for those fields to mutate them.
+//
+// All fields that are "exclusive to the task goroutine" can only be accessed
+// by the task goroutine while it is running. The task goroutine does not
+// require synchronization to read or write these fields.
+type Task struct {
+	taskNode
+
+	// runState is what the task goroutine is executing if it is not stopped.
+	// If runState is nil, the task goroutine should exit or has exited.
+	// runState is exclusive to the task goroutine.
+	runState taskRunState
+
+	// haveSyscallReturn is true if tc.Arch().Return() represents a value
+	// returned by a syscall (or set by ptrace after a syscall).
+	//
+	// haveSyscallReturn is exclusive to the task goroutine.
+	haveSyscallReturn bool
+
+	// interruptChan is notified whenever the task goroutine is interrupted
+	// (usually by a pending signal). interruptChan is effectively a condition
+	// variable that can be used in select statements.
+	//
+	// interruptChan is not saved; because saving interrupts all tasks,
+	// interruptChan is always notified after restore (see Task.run).
+	interruptChan chan struct{} `state:"nosave"`
+
+	// gosched contains the current scheduling state of the task goroutine.
+	//
+	// gosched is protected by goschedSeq. gosched is owned by the task
+	// goroutine.
+	goschedSeq ssync.SeqCount `state:"nosave"`
+	gosched    TaskGoroutineSchedInfo
+
+	// yieldCount is the number of times the task goroutine has called
+	// Task.InterruptibleSleepStart, Task.UninterruptibleSleepStart, or
+	// Task.Yield(), voluntarily ceasing execution.
+	//
+	// yieldCount is accessed using atomic memory operations. yieldCount is
+	// owned by the task goroutine.
+	yieldCount uint64
+
+	// pendingSignals is the set of pending signals that may be handled only by
+	// this task.
+	//
+	// pendingSignals is protected by (taskNode.)tg.signalHandlers.mu
+	// (hereafter "the signal mutex"); see comment on
+	// ThreadGroup.signalHandlers.
+	pendingSignals pendingSignals
+
+	// If haveSavedSignalMask is true, savedSignalMask is the signal mask that
+	// should be applied after the task has either delivered one signal to a
+	// user handler or is about to resume execution in the untrusted
+	// application.
+	//
+	// Both haveSavedSignalMask and savedSignalMask are exclusive to the task
+	// goroutine.
+	haveSavedSignalMask bool
+	savedSignalMask     linux.SignalSet
+
+	// signalStack is the alternate signal stack used by signal handlers for
+	// which the SA_ONSTACK flag is set.
+	//
+	// signalStack is exclusive to the task goroutine.
+	signalStack arch.SignalStack
+
+	// If groupStopRequired is true, the task should enter a group stop in the
+	// interrupt path. groupStopRequired is not redundant with
+	// tg.groupStopPhase != groupStopNone, because ptrace allows tracers to
+	// resume individual tasks from a group stop without ending the group stop
+	// as a whole.
+	//
+	// groupStopRequired is analogous to JOBCTL_TRAP_STOP in Linux, except that
+	// Linux only uses that flag for ptraced tasks.
+	//
+	// groupStopRequired is protected by the signal mutex.
+	groupStopRequired bool
+
+	// If groupStopAcknowledged is true, the task has already acknowledged that
+	// it is entering the most recent group stop that has been initiated on its
+	// thread group. groupStopAcknowledged is only meaningful if
+	// tg.groupStopPhase == groupStopInitiated.
+	//
+	// groupStopAcknowledged is analogous to !JOBCTL_STOP_CONSUME in Linux.
+	//
+	// groupStopAcknowledged is protected by the signal mutex.
+	groupStopAcknowledged bool
+
+	// If stop is not nil, it is the internally-initiated condition that
+	// currently prevents the task goroutine from running.
+	//
+	// stop is protected by the signal mutex.
+	stop TaskStop
+
+	// stopCount is the number of active external stops (calls to
+	// Task.BeginExternalStop that have not been paired with a call to
+	// Task.EndExternalStop), plus 1 if stop is not nil. Hence stopCount is
+	// non-zero if the task goroutine should stop.
+	//
+	// Mutating stopCount requires both locking the signal mutex and using
+	// atomic memory operations. Reading stopCount requires either locking the
+	// signal mutex or using atomic memory operations. This allows Task.doStop
+	// to require only a single atomic read in the common case where stopCount
+	// is 0.
+	//
+	// stopCount is not saved, because external stops cannot be retained across
+	// a save/restore cycle. (Suppose a sentryctl command issues an external
+	// stop; after a save/restore cycle, the restored sentry has no knowledge
+	// of the pre-save sentryctl command, and the stopped task would remain
+	// stopped forever.)
+	stopCount int32 `state:"nosave"`
+
+	// endStopCond is signaled when stopCount transitions to 0. The combination
+	// of stopCount and endStopCond effectively form a sync.WaitGroup, but
+	// WaitGroup provides no way to read its counter value.
+	//
+	// Invariant: endStopCond.L is the signal mutex. (This is not racy because
+	// sync.Cond.Wait is the only user of sync.Cond.L; only the task goroutine
+	// calls sync.Cond.Wait; and only the task goroutine can change the
+	// identity of the signal mutex, in Task.finishExec.)
+	endStopCond sync.Cond `state:"nosave"`
+
+	// exitStatus is the task's exit status.
+	//
+	// exitStatus is protected by the signal mutex.
+	exitStatus ExitStatus
+
+	// syscallRestartBlock represents a custom restart function to run in
+	// restart_syscall(2) to resume an interrupted syscall.
+	//
+	// syscallRestartBlock is exclusive to the task goroutine.
+	syscallRestartBlock SyscallRestartBlock
+
+	// mu protects some of the following fields.
+	mu sync.Mutex `state:"nosave"`
+
+	// tc and tr form the majority of the task's data.
+	//
+	// tc and tr are protected by mu. tc and tr are owned by the task
+	// goroutine. tr.signalMask is protected by the signal mutex and must be
+	// written using atomic memory operations (such that reading tr.signalMask
+	// is safe if the signal mutex is locked or if atomic memory operations are
+	// used), but is also owned by the task goroutine.
+	tc TaskContext
+	tr TaskResources
+
+	// p provides the mechanism by which the task runs code in userspace. The p
+	// interface object is immutable.
+	p platform.Context `state:"nosave"`
+
+	// k is the Kernel that this task belongs to. The k pointer is immutable.
+	k *Kernel
+
+	// If vforkParent is not nil, it is the task that created this task with
+	// vfork() or clone(CLONE_VFORK), and should have its vforkStop ended when
+	// this TaskContext is released.
+	//
+	// vforkParent is protected by the TaskSet mutex.
+	vforkParent *Task
+
+	// exitState is the task's progress through the exit path.
+	//
+	// exitState is protected by the TaskSet mutex. exitState is owned by the
+	// task goroutine.
+	exitState TaskExitState
+
+	// exitTracerNotified is true if the exit path has either signaled the
+	// task's tracer to indicate the exit, or determined that no such signal is
+	// needed. exitTracerNotified can only be true if exitState is
+	// TaskExitZombie or TaskExitDead.
+	//
+	// exitTracerNotified is protected by the TaskSet mutex.
+	exitTracerNotified bool
+
+	// exitTracerAcked is true if exitTracerNotified is true and either the
+	// task's tracer has acknowledged the exit notification, or the exit path
+	// has determined that no such notification is needed.
+	//
+	// exitTracerAcked is protected by the TaskSet mutex.
+	exitTracerAcked bool
+
+	// exitParentNotified is true if the exit path has either signaled the
+	// task's parent to indicate the exit, or determined that no such signal is
+	// needed. exitParentNotified can only be true if exitState is
+	// TaskExitZombie or TaskExitDead.
+	//
+	// exitParentNotified is protected by the TaskSet mutex.
+	exitParentNotified bool
+
+	// exitParentAcked is true if exitParentNotified is true and either the
+	// task's parent has acknowledged the exit notification, or the exit path
+	// has determined that no such acknowledgment is needed.
+	//
+	// exitParentAcked is protected by the TaskSet mutex.
+	exitParentAcked bool
+
+	// goroutineStopped is a WaitGroup whose counter value is 1 when the task
+	// goroutine is running and 0 when the task goroutine is stopped or has
+	// exited.
+	goroutineStopped sync.WaitGroup `state:"nosave"`
+
+	// ptraceTracer is the task that is ptrace-attached to this one. If
+	// ptraceTracer is nil, this task is not being traced. Note that due to
+	// atomic.Value limitations (atomic.Value.Store(nil) panics), a nil
+	// ptraceTracer is always represented as a typed nil (i.e. (*Task)(nil)).
+	//
+	// ptraceTracer is protected by the TaskSet mutex, and accessed with atomic
+	// operations. This allows paths that wouldn't otherwise lock the TaskSet
+	// mutex, notably the syscall path, to check if ptraceTracer is nil without
+	// additional synchronization.
+	ptraceTracer atomic.Value `state:".(*Task)"`
+
+	// ptraceTracees is the set of tasks that this task is ptrace-attached to.
+	//
+	// ptraceTracees is protected by the TaskSet mutex.
+	ptraceTracees map[*Task]struct{}
+
+	// ptraceOpts contains ptrace options explicitly set by the tracer. If
+	// ptraceTracer is nil, ptraceOpts is expected to be the zero value.
+	//
+	// ptraceOpts is protected by the TaskSet mutex.
+	ptraceOpts ptraceOptions
+
+	// ptraceSyscallMode controls ptrace behavior around syscall entry and
+	// exit.
+	//
+	// ptraceSyscallMode is protected by the TaskSet mutex.
+	ptraceSyscallMode ptraceSyscallMode
+
+	// If ptraceSinglestep is true, the next time the task executes application
+	// code, single-stepping should be enabled. ptraceSinglestep is stored
+	// independently of the architecture-specific trap flag because tracer
+	// detaching (which can happen concurrently with the tracee's execution if
+	// the tracer exits) must disable single-stepping, and the task's
+	// architectural state is implicitly exclusive to the task goroutine (no
+	// synchronization occurs before passing registers to SwitchToApp).
+	//
+	// ptraceSinglestep is analogous to Linux's TIF_SINGLESTEP.
+	//
+	// ptraceSinglestep is protected by the TaskSet mutex.
+	ptraceSinglestep bool
+
+	// If t is ptrace-stopped, ptraceCode is a ptrace-defined value set at the
+	// time that t entered the ptrace stop, reset to 0 when the tracer
+	// acknowledges the stop with a wait*() syscall. Otherwise, it is the
+	// signal number passed to the ptrace operation that ended the last ptrace
+	// stop on this task. In the latter case, the effect of ptraceCode depends
+	// on the nature of the ptrace stop; signal-delivery-stop uses it to
+	// conditionally override ptraceSiginfo, syscall-entry/exit-stops send the
+	// signal to the task after leaving the stop, and PTRACE_EVENT stops and
+	// traced group stops ignore it entirely.
+	//
+	// Linux contextually stores the equivalent of ptraceCode in
+	// task_struct::exit_code.
+	//
+	// ptraceCode is protected by the TaskSet mutex.
+	ptraceCode int32
+
+	// ptraceSiginfo is the value returned to the tracer by
+	// ptrace(PTRACE_GETSIGINFO) and modified by ptrace(PTRACE_SETSIGINFO).
+	// (Despite the name, PTRACE_PEEKSIGINFO is completely unrelated.)
+	// ptraceSiginfo is nil if the task is in a ptraced group-stop (this is
+	// required for PTRACE_GETSIGINFO to return EINVAL during such stops, which
+	// is in turn required to distinguish group stops from other ptrace stops,
+	// per subsection "Group-stop" in ptrace(2)).
+	//
+	// ptraceSiginfo is analogous to Linux's task_struct::last_siginfo.
+	//
+	// ptraceSiginfo is protected by the TaskSet mutex.
+	ptraceSiginfo *arch.SignalInfo
+
+	// ptraceEventMsg is the value set by PTRACE_EVENT stops and returned to
+	// the tracer by ptrace(PTRACE_GETEVENTMSG).
+	//
+	// ptraceEventMsg is protected by the TaskSet mutex.
+	ptraceEventMsg uint64
+
+	// The struct that holds the IO-related usage. The ioUsage pointer is
+	// immutable.
+	ioUsage *usage.IO
+
+	// logPrefix is a string containing the task's thread ID in the root PID
+	// namespace, and is prepended to log messages emitted by Task.Infof etc.
+	logPrefix atomic.Value `state:".(string)"`
+
+	// creds is the task's credentials.
+	//
+	// creds is protected by mu.
+	creds *auth.Credentials
+
+	// utsns is the task's UTS namespace.
+	//
+	// utsns is protected by mu.
+	utsns *UTSNamespace
+
+	// ipcns is the task's IPC namespace.
+	//
+	// ipcns is protected by mu.
+	ipcns *IPCNamespace
+
+	// parentDeathSignal is sent to this task's thread group when its parent exits.
+	//
+	// parentDeathSignal is protected by mu.
+	parentDeathSignal linux.Signal
+
+	// syscallFilters is all seccomp-bpf syscall filters applicable to the
+	// task, in the order in which they were installed.
+	//
+	// syscallFilters is protected by mu. syscallFilters is owned by the task
+	// goroutine.
+	syscallFilters []bpf.Program
+
+	// If cleartid is non-zero, treat it as a pointer to a ThreadID in the
+	// task's virtual address space; when the task exits, set the pointed-to
+	// ThreadID to 0, and wake any futex waiters.
+	//
+	// cleartid is exclusive to the task goroutine.
+	cleartid usermem.Addr
+
+	// This is mostly a fake cpumask just for sched_set/getaffinity as we
+	// don't really control the affinity.
+	//
+	// Invariant: allowedCPUMask.Size() ==
+	// sched.CPUMaskSize(Kernel.applicationCores).
+	//
+	// allowedCPUMask is protected by mu.
+	allowedCPUMask sched.CPUSet
+
+	// cpu is the fake cpu number returned by getcpu(2). cpu is ignored
+	// entirely if Kernel.useHostCores is true.
+	//
+	// cpu is accessed using atomic memory operations.
+	cpu int32
+
+	// This is used to keep track of changes made to a process' priority/niceness.
+	// It is mostly used to provide some reasonable return value from
+	// getpriority(2) after a call to setpriority(2) has been made.
+	// We currently do not actually modify a process' scheduling priority.
+	// NOTE: This represents the userspace view of priority (nice).
+	// This means that the value should be in the range [-20, 19].
+	//
+	// niceness is protected by mu.
+	niceness int
+
+	// This is used to track the numa policy for the current thread. This can be
+	// modified through a set_mempolicy(2) syscall. Since we always report a
+	// single numa node, all policies are no-ops. We only track this information
+	// so that we can return reasonable values if the application calls
+	// get_mempolicy(2) after setting a non-default policy. Note that in the
+	// real syscall, nodemask can be longer than 4 bytes, but we always report a
+	// single node so never need to save more than a single bit.
+	//
+	// numaPolicy and numaNodeMask are protected by mu.
+	numaPolicy   int32
+	numaNodeMask uint32
+
+	// If netns is true, the task is in a non-root network namespace. Network
+	// namespaces aren't currently implemented in full; being in a network
+	// namespace simply prevents the task from observing any network devices
+	// (including loopback) or using abstract socket addresses (see unix(7)).
+	//
+	// netns is protected by mu. netns is owned by the task goroutine.
+	netns bool
+
+	// If rseqPreempted is true, before the next call to p.Switch(), interrupt
+	// RSEQ critical regions as defined by tg.rseq and write the task
+	// goroutine's CPU number to rseqCPUAddr. rseqCPU is the last CPU number
+	// written to rseqCPUAddr.
+	//
+	// If rseqCPUAddr is 0, rseqCPU is -1.
+	//
+	// rseqCPUAddr, rseqCPU, and rseqPreempted are exclusive to the task
+	// goroutine.
+	rseqPreempted bool `state:"nosave"`
+	rseqCPUAddr   usermem.Addr
+	rseqCPU       int32
+
+	// copyScratchBuffer is a buffer available to CopyIn/CopyOut
+	// implementations that require an intermediate buffer to copy data
+	// into/out of. It prevents these buffers from being allocated/zeroed in
+	// each syscall and eventually garbage collected.
+	//
+	// copyScratchBuffer is exclusive to the task goroutine.
+	copyScratchBuffer [copyScratchBufferLen]byte `state:"nosave"`
+
+	// blockingTimer is used for blocking timeouts. blockingTimerChan is the
+	// channel that is sent to when blockingTimer fires.
+	//
+	// blockingTimer is exclusive to the task goroutine.
+	blockingTimer     *ktime.Timer    `state:"nosave"`
+	blockingTimerChan <-chan struct{} `state:"nosave"`
+
+	// futexWaiter is used for futex(FUTEX_WAIT) syscalls.
+	//
+	// futexWaiter is exclusive to the task goroutine.
+	futexWaiter *futex.Waiter `state:"nosave"`
+
+	// startTime is the real time at which the task started. It is set when
+	// a Task is created or invokes execve(2).
+	//
+	// startTime is protected by mu.
+	startTime ktime.Time
+}
+
+func (t *Task) savePtraceTracer() *Task {
+	return t.ptraceTracer.Load().(*Task)
+}
+
+func (t *Task) loadPtraceTracer(tracer *Task) {
+	t.ptraceTracer.Store(tracer)
+}
+
+func (t *Task) saveLogPrefix() string {
+	return t.logPrefix.Load().(string)
+}
+
+func (t *Task) loadLogPrefix(prefix string) {
+	t.logPrefix.Store(prefix)
+}
+
+// afterLoad is invoked by stateify.
+func (t *Task) afterLoad() {
+	t.interruptChan = make(chan struct{}, 1)
+	t.gosched.State = TaskGoroutineNonexistent
+	if t.stop != nil {
+		t.stopCount = 1
+	}
+	t.endStopCond.L = &t.tg.signalHandlers.mu
+	t.p = t.k.Platform.NewContext()
+	t.rseqPreempted = true
+	t.futexWaiter = futex.NewWaiter()
+}
+
+// copyScratchBufferLen is the length of the copyScratchBuffer field of the Task
+// struct.
+const copyScratchBufferLen = 52
+
+// TaskMaybe is the interface for extracting Tasks out of things which may be
+// or contain Task objects.
+type TaskMaybe interface {
+	// ExtractTask returns the Task.
+	ExtractTask() *Task
+}
+
+// CopyScratchBuffer returns a scratch buffer to be used in CopyIn/CopyOut
+// functions. It must only be used within those functions and can only be used
+// by the task goroutine; it exists to improve performance and thus
+// intentionally lacks any synchronization.
+//
+// Callers should pass a constant value as an argument, which will allow the
+// compiler to inline and optimize out the if statement below.
+func (t *Task) CopyScratchBuffer(size int) []byte {
+	if size > copyScratchBufferLen {
+		return make([]byte, size)
+	}
+	return t.copyScratchBuffer[:size]
+}
+
+// FutexWaiter returns the Task's futex.Waiter.
+func (t *Task) FutexWaiter() *futex.Waiter {
+	return t.futexWaiter
+}
+
+// ExtractTask implements TaskMaybe.ExtractTask.
+func (t *Task) ExtractTask() *Task {
+	return t
+}
+
+// TaskContext returns t's TaskContext.
+//
+// Precondition: The caller must be running on the task goroutine, or t.mu must
+// be locked.
+func (t *Task) TaskContext() *TaskContext {
+	return &t.tc
+}
+
+// TaskResources returns t's TaskResources.
+//
+// Precondition: The caller must be running on the task goroutine, or t.mu must
+// be locked.
+func (t *Task) TaskResources() *TaskResources {
+	return &t.tr
+}
+
+// WithMuLocked executes f with t.mu locked.
+func (t *Task) WithMuLocked(f func(*Task)) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	f(t)
+}
+
+// Kernel returns the Kernel containing t.
+func (t *Task) Kernel() *Kernel {
+	return t.k
+}
+
+// Value implements context.Context.Value.
+func (t *Task) Value(key interface{}) interface{} {
+	switch key {
+	case CtxCanTrace:
+		return t.CanTrace
+	case CtxKernel:
+		return t.k
+	case CtxPIDNamespace:
+		return t.tg.pidns
+	case CtxUTSNamespace:
+		return t.utsns
+	case CtxIPCNamespace:
+		return t.ipcns
+	case CtxTask:
+		return t
+	case auth.CtxCredentials:
+		return t.creds
+	case fs.CtxRoot:
+		return t.FSContext().RootDirectory()
+	case ktime.CtxRealtimeClock:
+		return t.k.RealtimeClock()
+	case limits.CtxLimits:
+		return t.tg.limits
+	case platform.CtxPlatform:
+		return t.k
+	case uniqueid.CtxGlobalUniqueID:
+		return t.k.UniqueID()
+	case uniqueid.CtxInotifyCookie:
+		return t.k.GenerateInotifyCookie()
+	default:
+		return nil
+	}
+}
+
+// SetClearTID sets t's cleartid.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) SetClearTID(addr usermem.Addr) {
+	t.cleartid = addr
+}
+
+// SetSyscallRestartBlock sets the restart block for use in
+// restart_syscall(2). After registering a restart block, a syscall should
+// return ERESTART_RESTARTBLOCK to request a restart using the block.
+//
+// Precondition: The caller must be running on the task goroutine.
+func (t *Task) SetSyscallRestartBlock(r SyscallRestartBlock) {
+	t.syscallRestartBlock = r
+}
+
+// SyscallRestartBlock returns the currently registered restart block for use in
+// restart_syscall(2). This function is *not* idempotent and may be called once
+// per syscall. This function must not be called if a restart block has not been
+// registered for the current syscall.
+//
+// Precondition: The caller must be running on the task goroutine.
+func (t *Task) SyscallRestartBlock() SyscallRestartBlock {
+	r := t.syscallRestartBlock
+	// Explicitly set the restart block to nil so that a future syscall can't
+	// accidentally reuse it.
+	t.syscallRestartBlock = nil
+	return r
+}
diff --git a/pkg/sentry/kernel/task_acct.go b/pkg/sentry/kernel/task_acct.go
new file mode 100644
index 000000000..ce12cdb64
--- /dev/null
+++ b/pkg/sentry/kernel/task_acct.go
@@ -0,0 +1,111 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// Accounting, limits, timers.
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+)
+
+// IOUsage returns the io usage of the thread.
+func (t *Task) IOUsage() *usage.IO {
+	return t.ioUsage
+}
+
+// IOUsage returns the total io usage of all dead and live threads in the group.
+func (tg *ThreadGroup) IOUsage() *usage.IO {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+
+	io := *tg.ioUsage
+	// Account for active tasks.
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		io.Accumulate(t.IOUsage())
+	}
+	return &io
+}
+
+// Name returns t's name.
+func (t *Task) Name() string {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.tc.Name
+}
+
+// SetName changes t's name.
+func (t *Task) SetName(name string) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.tc.Name = name
+	t.Debugf("Set thread name to %q", name)
+}
+
+// SetCPUTimer is used by setrlimit(RLIMIT_CPU) to enforce the hard and soft
+// limits on CPU time used by this process.
+func (tg *ThreadGroup) SetCPUTimer(l *limits.Limit) {
+	tg.Timer().applyCPULimits(*l)
+}
+
+// Limits implements context.Context.Limits.
+func (t *Task) Limits() *limits.LimitSet {
+	return t.ThreadGroup().Limits()
+}
+
+// StartTime returns t's start time.
+func (t *Task) StartTime() ktime.Time {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.startTime
+}
+
+// MaxRSS returns the maximum resident set size of the task in bytes. which
+// should be one of RUSAGE_SELF, RUSAGE_CHILDREN, RUSAGE_THREAD, or
+// RUSAGE_BOTH. See getrusage(2) for documentation on the behavior of these
+// flags.
+func (t *Task) MaxRSS(which int32) uint64 {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+
+	switch which {
+	case linux.RUSAGE_SELF, linux.RUSAGE_THREAD:
+		// If there's an active mm we can use its value.
+		if mm := t.MemoryManager(); mm != nil {
+			if mmMaxRSS := mm.MaxResidentSetSize(); mmMaxRSS > t.tg.maxRSS {
+				return mmMaxRSS
+			}
+		}
+		return t.tg.maxRSS
+	case linux.RUSAGE_CHILDREN:
+		return t.tg.childMaxRSS
+	case linux.RUSAGE_BOTH:
+		maxRSS := t.tg.maxRSS
+		if maxRSS < t.tg.childMaxRSS {
+			maxRSS = t.tg.childMaxRSS
+		}
+		if mm := t.MemoryManager(); mm != nil {
+			if mmMaxRSS := mm.MaxResidentSetSize(); mmMaxRSS > maxRSS {
+				return mmMaxRSS
+			}
+		}
+		return maxRSS
+	default:
+		// We'll only get here if which is invalid.
+		return 0
+	}
+}
diff --git a/pkg/sentry/kernel/task_block.go b/pkg/sentry/kernel/task_block.go
new file mode 100644
index 000000000..9fd24f134
--- /dev/null
+++ b/pkg/sentry/kernel/task_block.go
@@ -0,0 +1,207 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"time"
+
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// BlockWithTimeout blocks t until an event is received from C, the application
+// monotonic clock indicates that timeout has elapsed (only if haveTimeout is true),
+// or t is interrupted. It returns:
+//
+// - The remaining timeout, which is guaranteed to be 0 if the timeout expired,
+// and is unspecified if haveTimeout is false.
+//
+// - An error which is nil if an event is received from C, ETIMEDOUT if the timeout
+// expired, and syserror.ErrInterrupted if t is interrupted.
+func (t *Task) BlockWithTimeout(C chan struct{}, haveTimeout bool, timeout time.Duration) (time.Duration, error) {
+	if !haveTimeout {
+		return timeout, t.block(C, nil)
+	}
+
+	start := t.Kernel().MonotonicClock().Now()
+	deadline := start.Add(timeout)
+	err := t.BlockWithDeadline(C, true, deadline)
+
+	// Timeout, explicitly return a remaining duration of 0.
+	if err == syserror.ETIMEDOUT {
+		return 0, err
+	}
+
+	// Compute the remaining timeout. Note that even if block() above didn't
+	// return due to a timeout, we may have used up any of the remaining time
+	// since then. We cap the remaining timeout to 0 to make it easier to
+	// directly use the returned duration.
+	end := t.Kernel().MonotonicClock().Now()
+	remainingTimeout := timeout - end.Sub(start)
+	if remainingTimeout < 0 {
+		remainingTimeout = 0
+	}
+
+	return remainingTimeout, err
+}
+
+// BlockWithDeadline blocks t until an event is received from C, the
+// application monotonic clock indicates a time of deadline (only if
+// haveDeadline is true), or t is interrupted. It returns nil if an event is
+// received from C, ETIMEDOUT if the deadline expired, and
+// syserror.ErrInterrupted if t is interrupted.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) BlockWithDeadline(C chan struct{}, haveDeadline bool, deadline ktime.Time) error {
+	if !haveDeadline {
+		return t.block(C, nil)
+	}
+
+	// Start the timeout timer.
+	t.blockingTimer.Swap(ktime.Setting{
+		Enabled: true,
+		Next:    deadline,
+	})
+
+	err := t.block(C, t.blockingTimerChan)
+
+	// Stop the timeout timer and drain the channel.
+	t.blockingTimer.Swap(ktime.Setting{})
+	select {
+	case <-t.blockingTimerChan:
+	default:
+	}
+
+	return err
+}
+
+// BlockWithTimer blocks t until an event is received from C or tchan, or t is
+// interrupted. It returns nil if an event is received from C, ETIMEDOUT if an
+// event is received from tchan, and syserror.ErrInterrupted if t is
+// interrupted.
+//
+// Most clients should use BlockWithDeadline or BlockWithTimeout instead.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) BlockWithTimer(C chan struct{}, tchan <-chan struct{}) error {
+	return t.block(C, tchan)
+}
+
+// Block blocks t until an event is received from C or t is interrupted. It
+// returns nil if an event is received from C and syserror.ErrInterrupted if t
+// is interrupted.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) Block(C chan struct{}) error {
+	return t.block(C, nil)
+}
+
+// block blocks a task on one of many events.
+// N.B. defer is too expensive to be used here.
+func (t *Task) block(C chan struct{}, timerChan <-chan struct{}) error {
+	// Fast path if the request is already done.
+	select {
+	case <-C:
+		return nil
+	default:
+	}
+
+	// Deactive our address space, we don't need it.
+	interrupt := t.SleepStart()
+
+	select {
+	case <-C:
+		t.SleepFinish(true)
+		return nil
+
+	case <-interrupt:
+		t.SleepFinish(false)
+		// Return the indicated error on interrupt.
+		return syserror.ErrInterrupted
+
+	case <-timerChan:
+		// We've timed out.
+		t.SleepFinish(true)
+		return syserror.ETIMEDOUT
+	}
+}
+
+// SleepStart implements amutex.Sleeper.SleepStart.
+func (t *Task) SleepStart() <-chan struct{} {
+	t.Deactivate()
+	t.accountTaskGoroutineEnter(TaskGoroutineBlockedInterruptible)
+	return t.interruptChan
+}
+
+// SleepFinish implements amutex.Sleeper.SleepFinish.
+func (t *Task) SleepFinish(success bool) {
+	if !success {
+		// The interrupted notification is consumed only at the top-level
+		// (Run). Therefore we attempt to reset the pending notification.
+		// This will also elide our next entry back into the task, so we
+		// will process signals, state changes, etc.
+		t.interruptSelf()
+	}
+	t.accountTaskGoroutineLeave(TaskGoroutineBlockedInterruptible)
+	t.Activate()
+}
+
+// UninterruptibleSleepStart implements context.Context.UninterruptibleSleepStart.
+func (t *Task) UninterruptibleSleepStart(deactivate bool) {
+	if deactivate {
+		t.Deactivate()
+	}
+	t.accountTaskGoroutineEnter(TaskGoroutineBlockedUninterruptible)
+}
+
+// UninterruptibleSleepFinish implements context.Context.UninterruptibleSleepFinish.
+func (t *Task) UninterruptibleSleepFinish(activate bool) {
+	t.accountTaskGoroutineLeave(TaskGoroutineBlockedUninterruptible)
+	if activate {
+		t.Activate()
+	}
+}
+
+// interrupted returns true if interrupt or interruptSelf has been called at
+// least once since the last call to interrupted.
+func (t *Task) interrupted() bool {
+	select {
+	case <-t.interruptChan:
+		return true
+	default:
+		return false
+	}
+}
+
+// interrupt unblocks the task and interrupts it if it's currently running in
+// userspace.
+func (t *Task) interrupt() {
+	t.interruptSelf()
+	t.p.Interrupt()
+}
+
+// interruptSelf is like Interrupt, but can only be called by the task
+// goroutine.
+func (t *Task) interruptSelf() {
+	select {
+	case t.interruptChan <- struct{}{}:
+		t.Debugf("Interrupt queued")
+	default:
+		t.Debugf("Dropping duplicate interrupt")
+	}
+	// platform.Context.Interrupt() is unnecessary since a task goroutine
+	// calling interruptSelf() cannot also be blocked in
+	// platform.Context.Switch().
+}
diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go
new file mode 100644
index 000000000..3a74abdfb
--- /dev/null
+++ b/pkg/sentry/kernel/task_clone.go
@@ -0,0 +1,475 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bpf"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SharingOptions controls what resources are shared by a new task created by
+// Task.Clone, or an existing task affected by Task.Unshare.
+type SharingOptions struct {
+	// If NewAddressSpace is true, the task should have an independent virtual
+	// address space.
+	NewAddressSpace bool
+
+	// If NewSignalHandlers is true, the task should use an independent set of
+	// signal handlers.
+	NewSignalHandlers bool
+
+	// If NewThreadGroup is true, the task should be the leader of its own
+	// thread group. TerminationSignal is the signal that the thread group
+	// will send to its parent when it exits. If NewThreadGroup is false,
+	// TerminationSignal is ignored.
+	NewThreadGroup    bool
+	TerminationSignal linux.Signal
+
+	// If NewPIDNamespace is true:
+	//
+	// - In the context of Task.Clone, the new task should be the init task
+	// (TID 1) in a new PID namespace.
+	//
+	// - In the context of Task.Unshare, the task should create a new PID
+	// namespace, and all subsequent clones of the task should be members of
+	// the new PID namespace.
+	NewPIDNamespace bool
+
+	// If NewUserNamespace is true, the task should have an independent user
+	// namespace.
+	NewUserNamespace bool
+
+	// If NewNetworkNamespace is true, the task should have an independent
+	// network namespace. (Note that network namespaces are not really
+	// implemented; see comment on Task.netns for details.)
+	NewNetworkNamespace bool
+
+	// If NewFiles is true, the task should use an independent file descriptor
+	// table.
+	NewFiles bool
+
+	// If NewFSContext is true, the task should have an independent FSContext.
+	NewFSContext bool
+
+	// If NewUTSNamespace is true, the task should have an independent UTS
+	// namespace.
+	NewUTSNamespace bool
+
+	// If NewIPCNamespace is true, the task should have an independent IPC
+	// namespace.
+	NewIPCNamespace bool
+}
+
+// CloneOptions controls the behavior of Task.Clone.
+type CloneOptions struct {
+	// SharingOptions defines the set of resources that the new task will share
+	// with its parent.
+	SharingOptions
+
+	// Stack is the initial stack pointer of the new task. If Stack is 0, the
+	// new task will start with the same stack pointer as its parent.
+	Stack usermem.Addr
+
+	// If SetTLS is true, set the new task's TLS (thread-local storage)
+	// descriptor to TLS. If SetTLS is false, TLS is ignored.
+	SetTLS bool
+	TLS    usermem.Addr
+
+	// If ChildClearTID is true, when the child exits, 0 is written to the
+	// address ChildTID in the child's memory, and if the write is successful a
+	// futex wake on the same address is performed.
+	//
+	// If ChildSetTID is true, the child's thread ID (in the child's PID
+	// namespace) is written to address ChildTID in the child's memory. (As in
+	// Linux, failed writes are silently ignored.)
+	ChildClearTID bool
+	ChildSetTID   bool
+	ChildTID      usermem.Addr
+
+	// If ParentSetTID is true, the child's thread ID (in the parent's PID
+	// namespace) is written to address ParentTID in the parent's memory. (As
+	// in Linux, failed writes are silently ignored.)
+	//
+	// Older versions of the clone(2) man page state that CLONE_PARENT_SETTID
+	// causes the child's thread ID to be written to ptid in both the parent
+	// and child's memory, but this is a documentation error fixed by
+	// 87ab04792ced ("clone.2: Fix description of CLONE_PARENT_SETTID").
+	ParentSetTID bool
+	ParentTID    usermem.Addr
+
+	// If Vfork is true, place the parent in vforkStop until the cloned task
+	// releases its TaskContext.
+	Vfork bool
+
+	// If Untraced is true, do not report PTRACE_EVENT_CLONE/FORK/VFORK for
+	// this clone(), and do not ptrace-attach the caller's tracer to the new
+	// task. (PTRACE_EVENT_VFORK_DONE will still be reported if appropriate).
+	Untraced bool
+
+	// If InheritTracer is true, ptrace-attach the caller's tracer to the new
+	// task, even if no PTRACE_EVENT_CLONE/FORK/VFORK event would be reported
+	// for it. If both Untraced and InheritTracer are true, no event will be
+	// reported, but tracer inheritance will still occur.
+	InheritTracer bool
+}
+
+// Clone implements the clone(2) syscall and returns the thread ID of the new
+// task in t's PID namespace. Clone may return both a non-zero thread ID and a
+// non-nil error.
+//
+// Preconditions: The caller must be running Task.doSyscallInvoke on the task
+// goroutine.
+func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
+	// Since signal actions may refer to application signal handlers by virtual
+	// address, any set of signal handlers must refer to the same address
+	// space.
+	if !opts.NewSignalHandlers && opts.NewAddressSpace {
+		return 0, nil, syserror.EINVAL
+	}
+	// In order for the behavior of thread-group-directed signals to be sane,
+	// all tasks in a thread group must share signal handlers.
+	if !opts.NewThreadGroup && opts.NewSignalHandlers {
+		return 0, nil, syserror.EINVAL
+	}
+	// All tasks in a thread group must be in the same PID namespace.
+	if !opts.NewThreadGroup && (opts.NewPIDNamespace || t.childPIDNamespace != nil) {
+		return 0, nil, syserror.EINVAL
+	}
+	// The two different ways of specifying a new PID namespace are
+	// incompatible.
+	if opts.NewPIDNamespace && t.childPIDNamespace != nil {
+		return 0, nil, syserror.EINVAL
+	}
+	// Thread groups and FS contexts cannot span user namespaces.
+	if opts.NewUserNamespace && (!opts.NewThreadGroup || !opts.NewFSContext) {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// "If CLONE_NEWUSER is specified along with other CLONE_NEW* flags in a
+	// single clone(2) or unshare(2) call, the user namespace is guaranteed to
+	// be created first, giving the child (clone(2)) or caller (unshare(2))
+	// privileges over the remaining namespaces created by the call." -
+	// user_namespaces(7)
+	creds := t.Credentials()
+	var userns *auth.UserNamespace
+	if opts.NewUserNamespace {
+		var err error
+		// "EPERM (since Linux 3.9): CLONE_NEWUSER was specified in flags and
+		// the caller is in a chroot environment (i.e., the caller's root
+		// directory does not match the root directory of the mount namespace
+		// in which it resides)." - clone(2). Neither chroot(2) nor
+		// user_namespaces(7) document this.
+		if t.IsChrooted() {
+			return 0, nil, syserror.EPERM
+		}
+		userns, err = creds.NewChildUserNamespace()
+		if err != nil {
+			return 0, nil, err
+		}
+	}
+	if (opts.NewPIDNamespace || opts.NewNetworkNamespace || opts.NewUTSNamespace) && !creds.HasCapability(linux.CAP_SYS_ADMIN) {
+		return 0, nil, syserror.EPERM
+	}
+
+	utsns := t.UTSNamespace()
+	if opts.NewUTSNamespace {
+		// Note that this must happen after NewUserNamespace so we get
+		// the new userns if there is one.
+		utsns = t.UTSNamespace().Clone(userns)
+	}
+
+	ipcns := t.IPCNamespace()
+	if opts.NewIPCNamespace {
+		// Note that "If CLONE_NEWIPC is set, then create the process in a new IPC
+		// namespace"
+		ipcns = NewIPCNamespace()
+	}
+
+	tc, err := t.tc.Fork(t, !opts.NewAddressSpace)
+	if err != nil {
+		return 0, nil, err
+	}
+	// clone() returns 0 in the child.
+	tc.Arch.SetReturn(0)
+	if opts.Stack != 0 {
+		tc.Arch.SetStack(uintptr(opts.Stack))
+	}
+	if opts.SetTLS {
+		tc.Arch.StateData().Regs.Fs_base = uint64(opts.TLS)
+	}
+
+	pidns := t.tg.pidns
+	if t.childPIDNamespace != nil {
+		pidns = t.childPIDNamespace
+	} else if opts.NewPIDNamespace {
+		pidns = pidns.NewChild(userns)
+	}
+	tg := t.tg
+	parent := t.parent
+	if opts.NewThreadGroup {
+		sh := t.tg.signalHandlers
+		if opts.NewSignalHandlers {
+			sh = sh.Fork()
+		}
+		tg = NewThreadGroup(pidns, sh, opts.TerminationSignal, tg.limits.GetCopy(), t.k.monotonicClock)
+		parent = t
+	}
+	cfg := &TaskConfig{
+		Kernel:            t.k,
+		Parent:            parent,
+		ThreadGroup:       tg,
+		TaskContext:       tc,
+		TaskResources:     t.tr.Fork(!opts.NewFiles, !opts.NewFSContext),
+		Niceness:          t.Niceness(),
+		Credentials:       creds.Fork(),
+		NetworkNamespaced: t.netns,
+		AllowedCPUMask:    t.CPUMask(),
+		UTSNamespace:      utsns,
+		IPCNamespace:      ipcns,
+	}
+	if opts.NewNetworkNamespace {
+		cfg.NetworkNamespaced = true
+	}
+	nt, err := t.tg.pidns.owner.NewTask(cfg)
+	if err != nil {
+		if opts.NewThreadGroup {
+			tg.release()
+		}
+		return 0, nil, err
+	}
+
+	// "A child process created via fork(2) inherits a copy of its parent's
+	// alternate signal stack settings" - sigaltstack(2).
+	//
+	// However kernel/fork.c:copy_process() adds a limitation to this:
+	// "sigaltstack should be cleared when sharing the same VM".
+	if opts.NewAddressSpace || opts.Vfork {
+		nt.SetSignalStack(t.SignalStack())
+	}
+
+	if userns != nil {
+		if err := nt.SetUserNamespace(userns); err != nil {
+			// This shouldn't be possible: userns was created from nt.creds, so
+			// nt should have CAP_SYS_ADMIN in userns.
+			panic("Task.Clone: SetUserNamespace failed: " + err.Error())
+		}
+	}
+
+	// This has to happen last, because e.g. ptraceClone may send a SIGSTOP to
+	// nt that it must receive before its task goroutine starts running.
+	tid := nt.k.tasks.Root.IDOfTask(nt)
+	defer nt.Start(tid)
+
+	// "If fork/clone and execve are allowed by @prog, any child processes will
+	// be constrained to the same filters and system call ABI as the parent." -
+	// Documentation/prctl/seccomp_filter.txt
+	nt.syscallFilters = append([]bpf.Program(nil), t.syscallFilters...)
+	if opts.Vfork {
+		nt.vforkParent = t
+	}
+
+	if opts.ChildClearTID {
+		nt.SetClearTID(opts.ChildTID)
+	}
+	if opts.ChildSetTID {
+		// Can't use Task.CopyOut, which assumes AddressSpaceActive.
+		usermem.CopyObjectOut(t, nt.MemoryManager(), opts.ChildTID, nt.ThreadID(), usermem.IOOpts{})
+	}
+	ntid := t.tg.pidns.IDOfTask(nt)
+	if opts.ParentSetTID {
+		t.CopyOut(opts.ParentTID, ntid)
+	}
+
+	kind := ptraceCloneKindClone
+	if opts.Vfork {
+		kind = ptraceCloneKindVfork
+	} else if opts.TerminationSignal == linux.SIGCHLD {
+		kind = ptraceCloneKindFork
+	}
+	if t.ptraceClone(kind, nt, opts) {
+		if opts.Vfork {
+			return ntid, &SyscallControl{next: &runSyscallAfterPtraceEventClone{vforkChild: nt, vforkChildTID: ntid}}, nil
+		}
+		return ntid, &SyscallControl{next: &runSyscallAfterPtraceEventClone{}}, nil
+	}
+	if opts.Vfork {
+		t.maybeBeginVforkStop(nt)
+		return ntid, &SyscallControl{next: &runSyscallAfterVforkStop{childTID: ntid}}, nil
+	}
+	return ntid, nil, nil
+}
+
+// maybeBeginVforkStop checks if a previously-started vfork child is still
+// running and has not yet released its MM, such that its parent t should enter
+// a vforkStop.
+//
+// Preconditions: The caller must be running on t's task goroutine.
+func (t *Task) maybeBeginVforkStop(child *Task) {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if t.killedLocked() {
+		child.vforkParent = nil
+		return
+	}
+	if child.vforkParent == t {
+		t.beginInternalStopLocked((*vforkStop)(nil))
+	}
+}
+
+func (t *Task) unstopVforkParent() {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	if p := t.vforkParent; p != nil {
+		p.tg.signalHandlers.mu.Lock()
+		defer p.tg.signalHandlers.mu.Unlock()
+		if _, ok := p.stop.(*vforkStop); ok {
+			p.endInternalStopLocked()
+		}
+		// Parent no longer needs to be unstopped.
+		t.vforkParent = nil
+	}
+}
+
+type runSyscallAfterPtraceEventClone struct {
+	vforkChild *Task
+
+	// If vforkChild is not nil, vforkChildTID is its thread ID in the parent's
+	// PID namespace. vforkChildTID must be stored since the child may exit and
+	// release its TID before the PTRACE_EVENT stop ends.
+	vforkChildTID ThreadID
+}
+
+func (r *runSyscallAfterPtraceEventClone) execute(t *Task) taskRunState {
+	if r.vforkChild != nil {
+		t.maybeBeginVforkStop(r.vforkChild)
+		return &runSyscallAfterVforkStop{r.vforkChildTID}
+	}
+	return (*runSyscallExit)(nil)
+}
+
+type runSyscallAfterVforkStop struct {
+	// childTID has the same meaning as
+	// runSyscallAfterPtraceEventClone.vforkChildTID.
+	childTID ThreadID
+}
+
+func (r *runSyscallAfterVforkStop) execute(t *Task) taskRunState {
+	t.ptraceVforkDone(r.childTID)
+	return (*runSyscallExit)(nil)
+}
+
+// Unshare changes the set of resources t shares with other tasks, as specified
+// by opts.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) Unshare(opts *SharingOptions) error {
+	// In Linux unshare(2), NewThreadGroup implies NewSignalHandlers and
+	// NewSignalHandlers implies NewAddressSpace. All three flags are no-ops if
+	// t is the only task using its MM, which due to clone(2)'s rules imply
+	// that it is also the only task using its signal handlers / in its thread
+	// group, and cause EINVAL to be returned otherwise.
+	//
+	// Since we don't count the number of tasks using each address space or set
+	// of signal handlers, we reject NewSignalHandlers and NewAddressSpace
+	// altogether, and interpret NewThreadGroup as requiring that t be the only
+	// member of its thread group. This seems to be logically coherent, in the
+	// sense that clone(2) allows a task to share signal handlers and address
+	// spaces with tasks in other thread groups.
+	if opts.NewAddressSpace || opts.NewSignalHandlers {
+		return syserror.EINVAL
+	}
+	if opts.NewThreadGroup {
+		t.tg.signalHandlers.mu.Lock()
+		if t.tg.tasksCount != 1 {
+			t.tg.signalHandlers.mu.Unlock()
+			return syserror.EINVAL
+		}
+		t.tg.signalHandlers.mu.Unlock()
+		// This isn't racy because we're the only living task, and therefore
+		// the only task capable of creating new ones, in our thread group.
+	}
+	if opts.NewUserNamespace {
+		if t.IsChrooted() {
+			return syserror.EPERM
+		}
+		// This temporary is needed because Go.
+		creds := t.Credentials()
+		newUserNS, err := creds.NewChildUserNamespace()
+		if err != nil {
+			return err
+		}
+		err = t.SetUserNamespace(newUserNS)
+		if err != nil {
+			return err
+		}
+	}
+	haveCapSysAdmin := t.HasCapability(linux.CAP_SYS_ADMIN)
+	if opts.NewPIDNamespace {
+		if !haveCapSysAdmin {
+			return syserror.EPERM
+		}
+		t.childPIDNamespace = t.tg.pidns.NewChild(t.UserNamespace())
+	}
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if opts.NewNetworkNamespace {
+		if !haveCapSysAdmin {
+			return syserror.EPERM
+		}
+		t.netns = true
+	}
+	if opts.NewUTSNamespace {
+		if !haveCapSysAdmin {
+			return syserror.EPERM
+		}
+		// Note that this must happen after NewUserNamespace, so the
+		// new user namespace is used if there is one.
+		t.utsns = t.utsns.Clone(t.creds.UserNamespace)
+	}
+	if opts.NewIPCNamespace {
+		if !haveCapSysAdmin {
+			return syserror.EPERM
+		}
+		// Note that "If CLONE_NEWIPC is set, then create the process in a new IPC
+		// namespace"
+		t.ipcns = NewIPCNamespace()
+	}
+	if opts.NewFiles {
+		oldFDMap := t.tr.FDMap
+		t.tr.FDMap = oldFDMap.Fork()
+		oldFDMap.DecRef()
+	}
+	if opts.NewFSContext {
+		oldFS := t.tr.FSContext
+		t.tr.FSContext = oldFS.Fork()
+		oldFS.DecRef()
+	}
+	return nil
+}
+
+// vforkStop is a TaskStop imposed on a task that creates a child with
+// CLONE_VFORK or vfork(2), that ends when the child task ceases to use its
+// current MM. (Normally, CLONE_VFORK is used in conjunction with CLONE_VM, so
+// that the child and parent share mappings until the child execve()s into a
+// new process image or exits.)
+type vforkStop struct{}
+
+// StopIgnoresKill implements TaskStop.Killable.
+func (*vforkStop) Killable() bool { return true }
diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go
new file mode 100644
index 000000000..5c563ba08
--- /dev/null
+++ b/pkg/sentry/kernel/task_context.go
@@ -0,0 +1,179 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"errors"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/futex"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/loader"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// ErrNoSyscalls is returned if there is no syscall table.
+var ErrNoSyscalls = errors.New("no syscall table found")
+
+// Auxmap contains miscellaneous data for the task.
+type Auxmap map[string]interface{}
+
+// TaskContext is the subset of a task's data that is provided by the loader.
+type TaskContext struct {
+	// Name is the thread name set by the prctl(PR_SET_NAME) system call.
+	Name string
+
+	// Arch is the architecture-specific context (registers, etc.)
+	Arch arch.Context
+
+	// MemoryManager is the task's address space.
+	MemoryManager *mm.MemoryManager
+
+	// fu implements futexes in the address space.
+	fu *futex.Manager
+
+	// st is the task's syscall table.
+	st *SyscallTable
+}
+
+// release releases all resources held by the TaskContext. release is called by
+// the task when it execs into a new TaskContext or exits.
+func (tc *TaskContext) release() {
+	// Nil out pointers so that if the task is saved after release, it doesn't
+	// follow the pointers to possibly now-invalid objects.
+	if tc.MemoryManager != nil {
+		// TODO
+		tc.MemoryManager.DecUsers(context.Background())
+		tc.MemoryManager = nil
+	}
+	tc.fu = nil
+}
+
+// Fork returns a duplicate of tc. The copied TaskContext always has an
+// independent arch.Context. If shareAddressSpace is true, the copied
+// TaskContext shares an address space with the original; otherwise, the copied
+// TaskContext has an independent address space that is initially a duplicate
+// of the original's.
+func (tc *TaskContext) Fork(ctx context.Context, shareAddressSpace bool) (*TaskContext, error) {
+	newTC := &TaskContext{
+		Arch: tc.Arch.Fork(),
+		st:   tc.st,
+	}
+	if shareAddressSpace {
+		newTC.MemoryManager = tc.MemoryManager
+		if newTC.MemoryManager != nil {
+			if !newTC.MemoryManager.IncUsers() {
+				// Shouldn't be possible since tc.MemoryManager should be a
+				// counted user.
+				panic(fmt.Sprintf("TaskContext.Fork called with userless TaskContext.MemoryManager"))
+			}
+		}
+		newTC.fu = tc.fu
+	} else {
+		newMM, err := tc.MemoryManager.Fork(ctx)
+		if err != nil {
+			return nil, err
+		}
+		newTC.MemoryManager = newMM
+		// TODO: revisit when shmem is supported.
+		newTC.fu = futex.NewManager()
+	}
+	return newTC, nil
+}
+
+// Arch returns t's arch.Context.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) Arch() arch.Context {
+	return t.tc.Arch
+}
+
+// MemoryManager returns t's MemoryManager. MemoryManager does not take an
+// additional reference on the returned MM.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) MemoryManager() *mm.MemoryManager {
+	return t.tc.MemoryManager
+}
+
+// Futex returns t's futex manager.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) Futex() *futex.Manager {
+	return t.tc.fu
+}
+
+// SyscallTable returns t's syscall table.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) SyscallTable() *SyscallTable {
+	return t.tc.st
+}
+
+// Stack returns the userspace stack.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) Stack() *arch.Stack {
+	return &arch.Stack{t.Arch(), t.MemoryManager(), usermem.Addr(t.Arch().Stack())}
+}
+
+// LoadTaskImage loads filename into a new TaskContext.
+//
+// It takes several arguments:
+//  * mounts: MountNamespace to lookup filename in
+//  * root: Root to lookup filename under
+//  * wd: Working directory to lookup filename under
+//  * maxTraversals: maximum number of symlinks to follow
+//  * filename: path to binary to load
+//  * argv: Binary argv
+//  * envv: Binary envv
+//  * fs: Binary FeatureSet
+func (k *Kernel) LoadTaskImage(ctx context.Context, mounts *fs.MountNamespace, root, wd *fs.Dirent, maxTraversals uint, filename string, argv, envv []string, fs *cpuid.FeatureSet) (*TaskContext, error) {
+	// Prepare a new user address space to load into.
+	m := mm.NewMemoryManager(k)
+	defer m.DecUsers(ctx)
+
+	os, ac, name, err := loader.Load(ctx, m, mounts, root, wd, maxTraversals, fs, filename, argv, envv, k.extraAuxv, k.vdso)
+	if err != nil {
+		return nil, err
+	}
+
+	// Lookup our new syscall table.
+	st, ok := LookupSyscallTable(os, ac.Arch())
+	if !ok {
+		// No syscall table found. Yikes.
+		return nil, ErrNoSyscalls
+	}
+
+	if !m.IncUsers() {
+		panic("Failed to increment users count on new MM")
+	}
+	return &TaskContext{
+		Name:          name,
+		Arch:          ac,
+		MemoryManager: m,
+		fu:            futex.NewManager(),
+		st:            st,
+	}, nil
+}
diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go
new file mode 100644
index 000000000..2285847a2
--- /dev/null
+++ b/pkg/sentry/kernel/task_exec.go
@@ -0,0 +1,240 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// This file implements the machinery behind the execve() syscall. In brief, a
+// thread executes an execve() by killing all other threads in its thread
+// group, assuming the leader's identity, and then switching process images.
+//
+// This design is effectively mandated by Linux. From ptrace(2):
+//
+// """
+// execve(2) under ptrace
+//     When one thread in a multithreaded process calls execve(2), the
+//     kernel destroys all other threads in the process, and resets the
+//     thread ID of the execing thread to the thread group ID (process ID).
+//     (Or, to put things another way, when a multithreaded process does an
+//     execve(2), at completion of the call, it appears as though the
+//     execve(2) occurred in the thread group leader, regardless of which
+//     thread did the execve(2).)  This resetting of the thread ID looks
+//     very confusing to tracers:
+//
+//     *  All other threads stop in PTRACE_EVENT_EXIT stop, if the
+//        PTRACE_O_TRACEEXIT option was turned on.  Then all other threads
+//        except the thread group leader report death as if they exited via
+//        _exit(2) with exit code 0.
+//
+//     *  The execing tracee changes its thread ID while it is in the
+//        execve(2).  (Remember, under ptrace, the "pid" returned from
+//        waitpid(2), or fed into ptrace calls, is the tracee's thread ID.)
+//        That is, the tracee's thread ID is reset to be the same as its
+//        process ID, which is the same as the thread group leader's thread
+//        ID.
+//
+//     *  Then a PTRACE_EVENT_EXEC stop happens, if the PTRACE_O_TRACEEXEC
+//        option was turned on.
+//
+//     *  If the thread group leader has reported its PTRACE_EVENT_EXIT stop
+//        by this time, it appears to the tracer that the dead thread leader
+//        "reappears from nowhere".  (Note: the thread group leader does not
+//        report death via WIFEXITED(status) until there is at least one
+//        other live thread.  This eliminates the possibility that the
+//        tracer will see it dying and then reappearing.)  If the thread
+//        group leader was still alive, for the tracer this may look as if
+//        thread group leader returns from a different system call than it
+//        entered, or even "returned from a system call even though it was
+//        not in any system call".  If the thread group leader was not
+//        traced (or was traced by a different tracer), then during
+//        execve(2) it will appear as if it has become a tracee of the
+//        tracer of the execing tracee.
+//
+//     All of the above effects are the artifacts of the thread ID change in
+//     the tracee.
+// """
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// execStop is a TaskStop that a task sets on itself when it wants to execve
+// and is waiting for the other tasks in its thread group to exit first.
+type execStop struct{}
+
+// Killable implements TaskStop.Killable.
+func (*execStop) Killable() bool { return true }
+
+// Execve implements the execve(2) syscall by killing all other tasks in its
+// thread group and switching to newTC. Execve always takes ownership of newTC.
+//
+// Preconditions: The caller must be running Task.doSyscallInvoke on the task
+// goroutine.
+func (t *Task) Execve(newTC *TaskContext) (*SyscallControl, error) {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+
+	if t.tg.exiting || t.tg.execing != nil {
+		// We lost to a racing group-exit, kill, or exec from another thread
+		// and should just exit.
+		newTC.release()
+		return nil, syserror.EINTR
+	}
+
+	// Cancel any racing group stops.
+	t.tg.endGroupStopLocked(false)
+
+	// If the task has any siblings, they have to exit before the exec can
+	// continue.
+	t.tg.execing = t
+	if t.tg.tasks.Front() != t.tg.tasks.Back() {
+		// "[All] other threads except the thread group leader report death as
+		// if they exited via _exit(2) with exit code 0." - ptrace(2)
+		for sibling := t.tg.tasks.Front(); sibling != nil; sibling = sibling.Next() {
+			if t != sibling {
+				sibling.killLocked()
+			}
+		}
+		// The last sibling to exit will wake t.
+		t.beginInternalStopLocked((*execStop)(nil))
+	}
+
+	return &SyscallControl{next: &runSyscallAfterExecStop{newTC}, ignoreReturn: true}, nil
+}
+
+// The runSyscallAfterExecStop state continues execve(2) after all siblings of
+// a thread in the execve syscall have exited.
+type runSyscallAfterExecStop struct {
+	tc *TaskContext
+}
+
+func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState {
+	t.tg.pidns.owner.mu.Lock()
+	t.tg.execing = nil
+	if t.killed() {
+		t.tg.pidns.owner.mu.Unlock()
+		r.tc.release()
+		return (*runInterrupt)(nil)
+	}
+	// We are the thread group leader now. Save our old thread ID for
+	// PTRACE_EVENT_EXEC. This is racy in that if a tracer attaches after this
+	// point it will get a PID of 0, but this is consistent with Linux.
+	oldTID := ThreadID(0)
+	if tracer := t.Tracer(); tracer != nil {
+		oldTID = tracer.tg.pidns.tids[t]
+	}
+	t.promoteLocked()
+	// "During an execve(2), the dispositions of handled signals are reset to
+	// the default; the dispositions of ignored signals are left unchanged. ...
+	// [The] signal mask is preserved across execve(2). ... [The] pending
+	// signal set is preserved across an execve(2)." - signal(7)
+	//
+	// Details:
+	//
+	// - If the thread group is sharing its signal handlers with another thread
+	// group via CLONE_SIGHAND, execve forces the signal handlers to be copied
+	// (see Linux's fs/exec.c:de_thread). We're not reference-counting signal
+	// handlers, so we always make a copy.
+	//
+	// - "Disposition" only means sigaction::sa_handler/sa_sigaction; flags,
+	// restorer (if present), and mask are always reset. (See Linux's
+	// fs/exec.c:setup_new_exec => kernel/signal.c:flush_signal_handlers.)
+	t.tg.signalHandlers = t.tg.signalHandlers.CopyForExec()
+	t.endStopCond.L = &t.tg.signalHandlers.mu
+	// "Any alternate signal stack is not preserved (sigaltstack(2))." - execve(2)
+	t.signalStack = arch.SignalStack{Flags: arch.SignalStackFlagDisable}
+	// "The termination signal is reset to SIGCHLD (see clone(2))."
+	t.tg.terminationSignal = linux.SIGCHLD
+	// execed indicates that the process can no longer join a process group
+	// in some scenarios (namely, the parent call setpgid(2) on the child).
+	// See the JoinProcessGroup function in sessions.go for more context.
+	t.tg.execed = true
+	// Maximum RSS is preserved across execve(2).
+	t.updateRSSLocked()
+	// Restartable sequence state is discarded.
+	t.rseqPreempted = false
+	t.rseqCPUAddr = 0
+	t.rseqCPU = -1
+	t.tg.rscr.Store(&RSEQCriticalRegion{})
+	t.tg.pidns.owner.mu.Unlock()
+
+	// Remove FDs with the CloseOnExec flag set.
+	t.FDMap().RemoveIf(func(file *fs.File, flags FDFlags) bool {
+		return flags.CloseOnExec
+	})
+
+	// Switch to the new process.
+	t.MemoryManager().Deactivate()
+	t.mu.Lock()
+	// Update credentials to reflect the execve. This should precede switching
+	// MMs to ensure that dumpability has been reset first, if needed.
+	t.updateCredsForExecLocked()
+	t.tc.release()
+	t.tc = *r.tc
+	t.mu.Unlock()
+	t.unstopVforkParent()
+	// NOTE: All locks must be dropped prior to calling Activate.
+	t.MemoryManager().Activate()
+
+	t.ptraceExec(oldTID)
+	return (*runSyscallExit)(nil)
+}
+
+// promoteLocked makes t the leader of its thread group. If t is already the
+// thread group leader, promoteLocked is a no-op.
+//
+// Preconditions: All other tasks in t's thread group, including the existing
+// leader (if it is not t), have reached TaskExitZombie. The TaskSet mutex must
+// be locked for writing.
+func (t *Task) promoteLocked() {
+	oldLeader := t.tg.leader
+	if t == oldLeader {
+		return
+	}
+	// Swap the leader's TIDs with the execing task's. The latter will be
+	// released when the old leader is reaped below.
+	for ns := t.tg.pidns; ns != nil; ns = ns.parent {
+		oldTID, leaderTID := ns.tids[t], ns.tids[oldLeader]
+		ns.tids[oldLeader] = oldTID
+		ns.tids[t] = leaderTID
+		ns.tasks[oldTID] = oldLeader
+		ns.tasks[leaderTID] = t
+	}
+
+	// Inherit the old leader's start time.
+	oldStartTime := oldLeader.StartTime()
+	t.mu.Lock()
+	t.startTime = oldStartTime
+	t.mu.Unlock()
+
+	t.tg.leader = t
+	t.Infof("Becoming TID %d (in root PID namespace)", t.tg.pidns.owner.Root.tids[t])
+	t.updateLogPrefixLocked()
+	// Reap the original leader. If it has a tracer, detach it instead of
+	// waiting for it to acknowledge the original leader's death.
+	oldLeader.exitParentNotified = true
+	oldLeader.exitParentAcked = true
+	if tracer := oldLeader.Tracer(); tracer != nil {
+		delete(tracer.ptraceTracees, oldLeader)
+		oldLeader.forgetTracerLocked()
+		// Notify the tracer that it will no longer be receiving these events
+		// from the tracee.
+		tracer.tg.eventQueue.Notify(EventExit | EventTraceeStop | EventGroupContinue)
+	}
+	oldLeader.exitNotifyLocked(false)
+}
diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go
new file mode 100644
index 000000000..3d49ae350
--- /dev/null
+++ b/pkg/sentry/kernel/task_exit.go
@@ -0,0 +1,1139 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// This file implements the task exit cycle:
+//
+// - Tasks are asynchronously requested to exit with Task.Kill.
+//
+// - When able, the task goroutine enters the exit path starting from state
+// runExit.
+//
+// - Other tasks observe completed exits with Task.Wait (which implements the
+// wait*() family of syscalls).
+
+import (
+	"errors"
+	"fmt"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// An ExitStatus is a value communicated from an exiting task or thread group
+// to the party that reaps it.
+type ExitStatus struct {
+	// Code is the numeric value passed to the call to exit or exit_group that
+	// caused the exit. If the exit was not caused by such a call, Code is 0.
+	Code int
+
+	// Signo is the signal that caused the exit. If the exit was not caused by
+	// a signal, Signo is 0.
+	Signo int
+}
+
+// Signaled returns true if the ExitStatus indicates that the exiting task or
+// thread group was killed by a signal.
+func (es ExitStatus) Signaled() bool {
+	return es.Signo != 0
+}
+
+// Status returns the numeric representation of the ExitStatus returned by e.g.
+// the wait4() system call.
+func (es ExitStatus) Status() uint32 {
+	return ((uint32(es.Code) & 0xff) << 8) | (uint32(es.Signo) & 0xff)
+}
+
+// ShellExitCode returns the numeric exit code that Bash would return for an
+// exit status of es.
+func (es ExitStatus) ShellExitCode() int {
+	if es.Signaled() {
+		return 128 + es.Signo
+	}
+	return es.Code
+}
+
+// TaskExitState represents a step in the task exit path.
+//
+// "Exiting" and "exited" are often ambiguous; prefer to name specific states.
+type TaskExitState int
+
+const (
+	// TaskExitNone indicates that the task has not begun exiting.
+	TaskExitNone TaskExitState = iota
+
+	// TaskExitInitiated indicates that the task goroutine has entered the exit
+	// path, and the task is no longer eligible to participate in group stops
+	// or group signal handling. TaskExitInitiated is analogous to Linux's
+	// PF_EXITING.
+	TaskExitInitiated
+
+	// TaskExitZombie indicates that the task has released its resources, and
+	// the task no longer prevents a sibling thread from completing execve.
+	TaskExitZombie
+
+	// TaskExitDead indicates that the task's thread IDs have been released,
+	// and the task no longer prevents its thread group leader from being
+	// reaped. ("Reaping" refers to the transitioning of a task from
+	// TaskExitZombie to TaskExitDead.)
+	TaskExitDead
+)
+
+// String implements fmt.Stringer.
+func (t TaskExitState) String() string {
+	switch t {
+	case TaskExitNone:
+		return "TaskExitNone"
+	case TaskExitInitiated:
+		return "TaskExitInitiated"
+	case TaskExitZombie:
+		return "TaskExitZombie"
+	case TaskExitDead:
+		return "TaskExitDead"
+	default:
+		return strconv.Itoa(int(t))
+	}
+}
+
+// killLocked marks t as killed by enqueueing a SIGKILL, without causing the
+// thread-group-affecting side effects SIGKILL usually has.
+//
+// Preconditions: The signal mutex must be locked.
+func (t *Task) killLocked() {
+	// Clear killable stops.
+	if t.stop != nil && t.stop.Killable() {
+		t.endInternalStopLocked()
+	}
+	t.groupStopRequired = false
+	t.pendingSignals.enqueue(&arch.SignalInfo{
+		Signo: int32(linux.SIGKILL),
+		// Linux just sets SIGKILL in the pending signal bitmask without
+		// enqueueing an actual siginfo, such that
+		// kernel/signal.c:collect_signal() initalizes si_code to SI_USER.
+		Code: arch.SignalInfoUser,
+	})
+	t.interrupt()
+}
+
+// killed returns true if t has a SIGKILL pending. killed is analogous to
+// Linux's fatal_signal_pending().
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) killed() bool {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	return t.killedLocked()
+}
+
+func (t *Task) killedLocked() bool {
+	return t.pendingSignals.pendingSet&linux.SignalSetOf(linux.SIGKILL) != 0
+}
+
+// PrepareExit indicates an exit with status es.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) PrepareExit(es ExitStatus) {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.exitStatus = es
+}
+
+// PrepareGroupExit indicates a group exit with status es to t's thread group.
+//
+// PrepareGroupExit is analogous to Linux's do_group_exit(), except that it
+// does not tail-call do_exit(), except that it *does* set Task.exitStatus.
+// (Linux does not do so until within do_exit(), since it reuses exit_code for
+// ptrace.)
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) PrepareGroupExit(es ExitStatus) {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if t.tg.exiting || t.tg.execing != nil {
+		// Note that if t.tg.exiting is false but t.tg.execing is not nil, i.e.
+		// this "group exit" is being executed by the killed sibling of an
+		// execing task, then Task.Execve never set t.tg.exitStatus, so it's
+		// still the zero value. This is consistent with Linux, both in intent
+		// ("all other threads ... report death as if they exited via _exit(2)
+		// with exit code 0" - ptrace(2), "execve under ptrace") and in
+		// implementation (compare fs/exec.c:de_thread() =>
+		// kernel/signal.c:zap_other_threads() and
+		// kernel/exit.c:do_group_exit() =>
+		// include/linux/sched.h:signal_group_exit()).
+		t.exitStatus = t.tg.exitStatus
+		return
+	}
+	t.tg.exiting = true
+	t.tg.exitStatus = es
+	t.exitStatus = es
+	for sibling := t.tg.tasks.Front(); sibling != nil; sibling = sibling.Next() {
+		if sibling != t {
+			sibling.killLocked()
+		}
+	}
+}
+
+// Kill requests that all tasks in ts exit as if group exiting with status es.
+// Kill does not wait for tasks to exit.
+//
+// Kill has no analogue in Linux; it's provided for save/restore only.
+func (ts *TaskSet) Kill(es ExitStatus) {
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	ts.Root.exiting = true
+	for t := range ts.Root.tids {
+		t.tg.signalHandlers.mu.Lock()
+		if !t.tg.exiting {
+			t.tg.exiting = true
+			t.tg.exitStatus = es
+		}
+		t.killLocked()
+		t.tg.signalHandlers.mu.Unlock()
+	}
+}
+
+// advanceExitStateLocked checks that t's current exit state is oldExit, then
+// sets it to newExit. If t's current exit state is not oldExit,
+// advanceExitStateLocked panics.
+//
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) advanceExitStateLocked(oldExit, newExit TaskExitState) {
+	if t.exitState != oldExit {
+		panic(fmt.Sprintf("Transitioning from exit state %v to %v: unexpected preceding state %v", oldExit, newExit, t.exitState))
+	}
+	t.Debugf("Transitioning from exit state %v to %v", oldExit, newExit)
+	t.exitState = newExit
+}
+
+// runExit is the entry point into the task exit path.
+type runExit struct{}
+
+func (*runExit) execute(t *Task) taskRunState {
+	t.ptraceExit()
+	return (*runExitMain)(nil)
+}
+
+type runExitMain struct{}
+
+func (*runExitMain) execute(t *Task) taskRunState {
+	lastExiter := t.exitThreadGroup()
+
+	// If the task has a cleartid, and the thread group wasn't killed by a
+	// signal, handle that before releasing the MM.
+	if t.cleartid != 0 {
+		t.tg.signalHandlers.mu.Lock()
+		signaled := t.tg.exiting && t.tg.exitStatus.Signaled()
+		t.tg.signalHandlers.mu.Unlock()
+		if !signaled {
+			if _, err := t.CopyOut(t.cleartid, ThreadID(0)); err == nil {
+				t.Futex().Wake(uintptr(t.cleartid), ^uint32(0), 1)
+			}
+			// If the CopyOut fails, there's nothing we can do.
+		}
+	}
+
+	// Deactivate the address space before releasing the MM.
+	t.Deactivate()
+
+	// Update the max resident set size before releasing t.tc.mm.
+	t.tg.pidns.owner.mu.Lock()
+	t.updateRSSLocked()
+	t.tg.pidns.owner.mu.Unlock()
+
+	// Release all of the task's resources.
+	t.mu.Lock()
+	t.tc.release()
+	t.tr.release()
+	t.mu.Unlock()
+	t.unstopVforkParent()
+
+	// If this is the last task to exit from the thread group, release the
+	// thread group's resources.
+	if lastExiter {
+		t.tg.release()
+	}
+
+	// Detach tracees.
+	t.exitPtrace()
+
+	// Reparent the task's children.
+	t.exitChildren()
+
+	// Don't tail-call runExitNotify, as exitChildren may have initiated a stop
+	// to wait for a PID namespace to die.
+	return (*runExitNotify)(nil)
+}
+
+// exitThreadGroup transitions t to TaskExitInitiated, indicating to t's thread
+// group that it is no longer eligible to participate in group activities. It
+// returns true if t is the last task in its thread group to call
+// exitThreadGroup.
+func (t *Task) exitThreadGroup() bool {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	t.tg.signalHandlers.mu.Lock()
+	// Can't defer unlock: see below.
+
+	t.advanceExitStateLocked(TaskExitNone, TaskExitInitiated)
+	t.tg.activeTasks--
+	last := t.tg.activeTasks == 0
+
+	// Ensure that someone will handle the signals we can't.
+	t.setSignalMaskLocked(^linux.SignalSet(0))
+
+	// Check if this task's exit interacts with an initiated group stop.
+	if t.tg.groupStopPhase != groupStopInitiated {
+		t.tg.signalHandlers.mu.Unlock()
+		return last
+	}
+	if t.groupStopAcknowledged {
+		// Un-acknowledge the group stop.
+		t.tg.groupStopCount--
+		t.groupStopAcknowledged = false
+		// If the group stop wasn't complete before, then there is still at
+		// least one other task that hasn't acknowledged the group stop, so
+		// it is still not complete now.
+		t.tg.signalHandlers.mu.Unlock()
+		return last
+	}
+	if t.tg.groupStopCount != t.tg.activeTasks {
+		t.tg.signalHandlers.mu.Unlock()
+		return last
+	}
+	t.Debugf("Completing group stop")
+	t.tg.groupStopPhase = groupStopComplete
+	t.tg.groupStopWaitable = true
+	sig := t.tg.groupStopSignal
+	t.tg.groupContNotify = false
+	t.tg.groupContWaitable = false
+	// signalStop must be called with t's signal mutex unlocked.
+	t.tg.signalHandlers.mu.Unlock()
+	if t.tg.leader.parent != nil {
+		t.tg.leader.parent.signalStop(t, arch.CLD_STOPPED, int32(sig))
+		t.tg.leader.parent.tg.eventQueue.Notify(EventChildGroupStop)
+	}
+	return last
+}
+
+func (t *Task) exitChildren() {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	newParent := t.findReparentTargetLocked()
+	if newParent == nil {
+		// "If the init process of a PID namespace terminates, the kernel
+		// terminates all of the processes in the namespace via a SIGKILL
+		// signal." - pid_namespaces(7)
+		t.Debugf("Init process terminating, killing namespace")
+		t.tg.pidns.exiting = true
+		for other := range t.tg.pidns.tids {
+			if other.tg != t.tg {
+				other.tg.signalHandlers.mu.Lock()
+				other.sendSignalLocked(&arch.SignalInfo{
+					Signo: int32(linux.SIGKILL),
+				}, false /* group */)
+				other.tg.signalHandlers.mu.Unlock()
+			}
+		}
+		// TODO: The init process waits for all processes in the
+		// namespace to exit before completing its own exit
+		// (kernel/pid_namespace.c:zap_pid_ns_processes()). Stop until all
+		// other tasks in the namespace are dead, except possibly for this
+		// thread group's leader (which can't be reaped until this task exits).
+	}
+	// This is correct even if newParent is nil (it ensures that children don't
+	// wait for a parent to reap them.)
+	for c := range t.children {
+		if sig := c.ParentDeathSignal(); sig != 0 {
+			siginfo := &arch.SignalInfo{
+				Signo: int32(sig),
+				Code:  arch.SignalInfoUser,
+			}
+			siginfo.SetPid(int32(c.tg.pidns.tids[t]))
+			siginfo.SetUid(int32(t.Credentials().RealKUID.In(c.UserNamespace()).OrOverflow()))
+			c.tg.signalHandlers.mu.Lock()
+			c.sendSignalLocked(siginfo, true /* group */)
+			c.tg.signalHandlers.mu.Unlock()
+		}
+		c.reparentLocked(newParent)
+		if newParent != nil {
+			newParent.children[c] = struct{}{}
+		}
+	}
+}
+
+// findReparentTargetLocked returns the task to which t's children should be
+// reparented. If no such task exists, findNewParentLocked returns nil.
+//
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) findReparentTargetLocked() *Task {
+	// Reparent to any sibling in the same thread group that hasn't begun
+	// exiting.
+	if t2 := t.tg.anyNonExitingTaskLocked(); t2 != nil {
+		return t2
+	}
+	// "A child process that is orphaned within the namespace will be
+	// reparented to [the init process for the namespace] ..." -
+	// pid_namespaces(7)
+	if init := t.tg.pidns.tasks[InitTID]; init != nil {
+		return init.tg.anyNonExitingTaskLocked()
+	}
+	return nil
+}
+
+func (tg *ThreadGroup) anyNonExitingTaskLocked() *Task {
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		if t.exitState == TaskExitNone {
+			return t
+		}
+	}
+	return nil
+}
+
+// reparentLocked changes t's parent. The new parent may be nil.
+//
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) reparentLocked(parent *Task) {
+	oldParent := t.parent
+	t.parent = parent
+	// If a thread group leader's parent changes, reset the thread group's
+	// termination signal to SIGCHLD and re-check exit notification. (Compare
+	// kernel/exit.c:reparent_leader().)
+	if t != t.tg.leader {
+		return
+	}
+	if oldParent == nil && parent == nil {
+		return
+	}
+	if oldParent != nil && parent != nil && oldParent.tg == parent.tg {
+		return
+	}
+	t.tg.terminationSignal = linux.SIGCHLD
+	if t.exitParentNotified && !t.exitParentAcked {
+		t.exitParentNotified = false
+		t.exitNotifyLocked(false)
+	}
+}
+
+// When a task exits, other tasks in the system, notably the task's parent and
+// ptracer, may want to be notified. The exit notification system ensures that
+// interested tasks receive signals and/or are woken from blocking calls to
+// wait*() syscalls; these notifications must be resolved before exiting tasks
+// can be reaped and disappear from the system.
+//
+// Each task may have a parent task and/or a tracer task. If both a parent and
+// a tracer exist, they may be the same task, different tasks in the same
+// thread group, or tasks in different thread groups. (In the last case, Linux
+// refers to the task as being ptrace-reparented due to an implementation
+// detail; we avoid this terminology to avoid confusion.)
+//
+// A thread group is *empty* if all non-leader tasks in the thread group are
+// dead, and the leader is either a zombie or dead. The exit of a thread group
+// leader is never waitable - by either the parent or tracer - until the thread
+// group is empty.
+//
+// There are a few ways for an exit notification to be resolved:
+//
+// - The exit notification may be acknowledged by a call to Task.Wait with
+// WaitOptions.ConsumeEvent set (e.g. due to a wait4() syscall).
+//
+// - If the notified party is the parent, and the parent thread group is not
+// also the tracer thread group, and the notification signal is SIGCHLD, the
+// parent may explicitly ignore the notification (see quote in exitNotify).
+// Note that it's possible for the notified party to ignore the signal in other
+// cases, but the notification is only resolved under the above conditions.
+// (Actually, there is one exception; see the last paragraph of the "leader,
+// has tracer, tracer thread group is parent thread group" case below.)
+//
+// - If the notified party is the parent, and the parent does not exist, the
+// notification is resolved as if ignored. (This is only possible in the
+// sentry. In Linux, the only task / thread group without a parent is global
+// init, and killing global init causes a kernel panic.)
+//
+// - If the notified party is a tracer, the tracer may detach the traced task.
+// (Zombie tasks cannot be ptrace-attached, so the reverse is not possible.)
+//
+// In addition, if the notified party is the parent, the parent may exit and
+// cause the notifying task to be reparented to another thread group. This does
+// not resolve the notification; instead, the notification must be resent to
+// the new parent.
+//
+// The series of notifications generated for a given task's exit depend on
+// whether it is a thread group leader; whether the task is ptraced; and, if
+// so, whether the tracer thread group is the same as the parent thread group.
+//
+// - Non-leader, no tracer: No notification is generated; the task is reaped
+// immediately.
+//
+// - Non-leader, has tracer: SIGCHLD is sent to the tracer. When the tracer
+// notification is resolved (by waiting or detaching), the task is reaped. (For
+// non-leaders, whether the tracer and parent thread groups are the same is
+// irrelevant.)
+//
+// - Leader, no tracer: The task remains a zombie, with no notification sent,
+// until all other tasks in the thread group are dead. (In Linux terms, this
+// condition is indicated by include/linux/sched.h:thread_group_empty(); tasks
+// are removed from their thread_group list in kernel/exit.c:release_task() =>
+// __exit_signal() => __unhash_process().) Then the thread group's termination
+// signal is sent to the parent. When the parent notification is resolved (by
+// waiting or ignoring), the task is reaped.
+//
+// - Leader, has tracer, tracer thread group is not parent thread group:
+// SIGCHLD is sent to the tracer. When the tracer notification is resolved (by
+// waiting or detaching), and all other tasks in the thread group are dead, the
+// thread group's termination signal is sent to the parent. (Note that the
+// tracer cannot resolve the exit notification by waiting until the thread
+// group is empty.) When the parent notification is resolved, the task is
+// reaped.
+//
+// - Leader, has tracer, tracer thread group is parent thread group:
+//
+// If all other tasks in the thread group are dead, the thread group's
+// termination signal is sent to the parent. At this point, the notification
+// can only be resolved by waiting. If the parent detaches from the task as a
+// tracer, the notification is not resolved, but the notification can now be
+// resolved by waiting or ignoring. When the parent notification is resolved,
+// the task is reaped.
+//
+// If at least one task in the thread group is not dead, SIGCHLD is sent to the
+// parent. At this point, the notification cannot be resolved at all; once the
+// thread group becomes empty, it can be resolved only by waiting. If the
+// parent detaches from the task as a tracer before all remaining tasks die,
+// then exit notification proceeds as in the case where the leader never had a
+// tracer. If the parent detaches from the task as a tracer after all remaining
+// tasks die, the notification is not resolved, but the notification can now be
+// resolved by waiting or ignoring. When the parent notification is resolved,
+// the task is reaped.
+//
+// In both of the above cases, when the parent detaches from the task as a
+// tracer while the thread group is empty, whether or not the parent resolves
+// the notification by ignoring it is based on the parent's SIGCHLD signal
+// action, whether or not the thread group's termination signal is SIGCHLD
+// (Linux: kernel/ptrace.c:__ptrace_detach() => ignoring_children()).
+//
+// There is one final wrinkle: A leader can become a non-leader due to a
+// sibling execve. In this case, the execing thread detaches the leader's
+// tracer (if one exists) and reaps the leader immediately. In Linux, this is
+// in fs/exec.c:de_thread(); in the sentry, this is in Task.promoteLocked().
+
+type runExitNotify struct{}
+
+func (*runExitNotify) execute(t *Task) taskRunState {
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+	t.advanceExitStateLocked(TaskExitInitiated, TaskExitZombie)
+	t.tg.liveTasks--
+	// Check if this completes a sibling's execve.
+	if t.tg.execing != nil && t.tg.liveTasks == 1 {
+		// execing blocks the addition of new tasks to the thread group, so
+		// the sole living task must be the execing one.
+		e := t.tg.execing
+		e.tg.signalHandlers.mu.Lock()
+		if _, ok := e.stop.(*execStop); ok {
+			e.endInternalStopLocked()
+		}
+		e.tg.signalHandlers.mu.Unlock()
+	}
+	t.exitNotifyLocked(false)
+	// The task goroutine will now exit.
+	return nil
+}
+
+// exitNotifyLocked is called after changes to t's state that affect exit
+// notification.
+//
+// If fromPtraceDetach is true, the caller is ptraceDetach or exitPtrace;
+// thanks to Linux's haphazard implementation of this functionality, such cases
+// determine whether parent notifications are ignored based on the parent's
+// handling of SIGCHLD, regardless of what the exited task's thread group's
+// termination signal is.
+//
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) exitNotifyLocked(fromPtraceDetach bool) {
+	if t.exitState != TaskExitZombie {
+		return
+	}
+	if !t.exitTracerNotified {
+		t.exitTracerNotified = true
+		tracer := t.Tracer()
+		if tracer == nil {
+			t.exitTracerAcked = true
+		} else if t != t.tg.leader || t.parent == nil || tracer.tg != t.parent.tg {
+			// Don't set exitParentNotified if t is non-leader, even if the
+			// tracer is in the parent thread group, so that if the parent
+			// detaches the following call to exitNotifyLocked passes through
+			// the !exitParentNotified case below and causes t to be reaped
+			// immediately.
+			//
+			// Tracer notification doesn't care about about
+			// SIG_IGN/SA_NOCLDWAIT.
+			tracer.tg.signalHandlers.mu.Lock()
+			tracer.sendSignalLocked(t.exitNotificationSignal(linux.SIGCHLD, tracer), true /* group */)
+			tracer.tg.signalHandlers.mu.Unlock()
+			// Wake EventTraceeStop waiters as well since this task will never
+			// ptrace-stop again.
+			tracer.tg.eventQueue.Notify(EventExit | EventTraceeStop)
+		} else {
+			// t is a leader and the tracer is in the parent thread group.
+			t.exitParentNotified = true
+			sig := linux.SIGCHLD
+			if t.tg.tasksCount == 1 {
+				sig = t.tg.terminationSignal
+			}
+			// This notification doesn't care about SIG_IGN/SA_NOCLDWAIT either
+			// (in Linux, the check in do_notify_parent() is gated by
+			// !tsk->ptrace.)
+			t.parent.tg.signalHandlers.mu.Lock()
+			t.parent.sendSignalLocked(t.exitNotificationSignal(sig, t.parent), true /* group */)
+			t.parent.tg.signalHandlers.mu.Unlock()
+			// See below for rationale for this event mask.
+			t.parent.tg.eventQueue.Notify(EventExit | EventChildGroupStop | EventGroupContinue)
+		}
+	}
+	if t.exitTracerAcked && !t.exitParentNotified {
+		if t != t.tg.leader {
+			t.exitParentNotified = true
+			t.exitParentAcked = true
+		} else if t.tg.tasksCount == 1 {
+			t.exitParentNotified = true
+			if t.parent == nil {
+				t.exitParentAcked = true
+			} else {
+				// "POSIX.1-2001 specifies that if the disposition of SIGCHLD is
+				// set to SIG_IGN or the SA_NOCLDWAIT flag is set for SIGCHLD (see
+				// sigaction(2)), then children that terminate do not become
+				// zombies and a call to wait() or waitpid() will block until all
+				// children have terminated, and then fail with errno set to
+				// ECHILD. (The original POSIX standard left the behavior of
+				// setting SIGCHLD to SIG_IGN unspecified. Note that even though
+				// the default disposition of SIGCHLD is "ignore", explicitly
+				// setting the disposition to SIG_IGN results in different
+				// treatment of zombie process children.) Linux 2.6 conforms to
+				// this specification." - wait(2)
+				//
+				// Some undocumented Linux-specific details:
+				//
+				// - All of the above is ignored if the termination signal isn't
+				// SIGCHLD.
+				//
+				// - SA_NOCLDWAIT causes the leader to be immediately reaped, but
+				// does not suppress the SIGCHLD.
+				signalParent := t.tg.terminationSignal.IsValid()
+				t.parent.tg.signalHandlers.mu.Lock()
+				if t.tg.terminationSignal == linux.SIGCHLD || fromPtraceDetach {
+					if act, ok := t.parent.tg.signalHandlers.actions[linux.SIGCHLD]; ok {
+						if act.Handler == arch.SignalActIgnore {
+							t.exitParentAcked = true
+							signalParent = false
+						} else if act.Flags&arch.SignalFlagNoCldWait != 0 {
+							t.exitParentAcked = true
+						}
+					}
+				}
+				if signalParent {
+					t.parent.tg.leader.sendSignalLocked(t.exitNotificationSignal(t.tg.terminationSignal, t.parent), true /* group */)
+				}
+				t.parent.tg.signalHandlers.mu.Unlock()
+				// If a task in the parent was waiting for a child group stop
+				// or continue, it needs to be notified of the exit, because
+				// there may be no remaining eligible tasks (so that wait
+				// should return ECHILD).
+				t.parent.tg.eventQueue.Notify(EventExit | EventChildGroupStop | EventGroupContinue)
+			}
+		}
+	}
+	if t.exitTracerAcked && t.exitParentAcked {
+		t.advanceExitStateLocked(TaskExitZombie, TaskExitDead)
+		for ns := t.tg.pidns; ns != nil; ns = ns.parent {
+			tid := ns.tids[t]
+			delete(ns.tasks, tid)
+			delete(ns.tids, t)
+		}
+		t.tg.exitedCPUStats.Accumulate(t.CPUStats())
+		t.tg.ioUsage.Accumulate(t.ioUsage)
+		t.tg.signalHandlers.mu.Lock()
+		t.tg.tasks.Remove(t)
+		if t.tg.lastTimerSignalTask == t {
+			t.tg.lastTimerSignalTask = nil
+		}
+		t.tg.tasksCount--
+		tc := t.tg.tasksCount
+		t.tg.signalHandlers.mu.Unlock()
+		if tc == 1 && t != t.tg.leader {
+			// Our fromPtraceDetach doesn't matter here (in Linux terms, this
+			// is via a call to release_task()).
+			t.tg.leader.exitNotifyLocked(false)
+		} else if tc == 0 {
+			t.tg.processGroup.decRefWithParent(t.tg.parentPG())
+		}
+		if t.parent != nil {
+			delete(t.parent.children, t)
+			t.parent = nil
+		}
+	}
+}
+
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) exitNotificationSignal(sig linux.Signal, receiver *Task) *arch.SignalInfo {
+	info := &arch.SignalInfo{
+		Signo: int32(sig),
+	}
+	info.SetPid(int32(receiver.tg.pidns.tids[t]))
+	info.SetUid(int32(t.Credentials().RealKUID.In(receiver.UserNamespace()).OrOverflow()))
+	if t.exitStatus.Signaled() {
+		info.Code = arch.CLD_KILLED
+		info.SetStatus(int32(t.exitStatus.Signo))
+	} else {
+		info.Code = arch.CLD_EXITED
+		info.SetStatus(int32(t.exitStatus.Code))
+	}
+	// TODO: Set utime, stime.
+	return info
+}
+
+// ExitStatus returns t's exit status, which is only guaranteed to be
+// meaningful if t.ExitState() != TaskExitNone.
+func (t *Task) ExitStatus() ExitStatus {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	return t.exitStatus
+}
+
+// ExitStatus returns the exit status that would be returned by a consuming
+// wait*() on tg.
+func (tg *ThreadGroup) ExitStatus() ExitStatus {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	tg.signalHandlers.mu.Lock()
+	defer tg.signalHandlers.mu.Unlock()
+	if tg.exiting {
+		return tg.exitStatus
+	}
+	return tg.leader.exitStatus
+}
+
+// TerminationSignal returns the thread group's termination signal.
+func (tg *ThreadGroup) TerminationSignal() linux.Signal {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.terminationSignal
+}
+
+// Task events that can be waited for.
+const (
+	// EventExit represents an exit notification generated for a child thread
+	// group leader or a tracee under the conditions specified in the comment
+	// above runExitNotify.
+	EventExit waiter.EventMask = 1 << iota
+
+	// EventChildGroupStop occurs when a child thread group completes a group
+	// stop (i.e. all tasks in the child thread group have entered a stopped
+	// state as a result of a group stop).
+	EventChildGroupStop
+
+	// EventTraceeStop occurs when a task that is ptraced by a task in the
+	// notified thread group enters a ptrace stop (see ptrace(2)).
+	EventTraceeStop
+
+	// EventGroupContinue occurs when a child thread group, or a thread group
+	// whose leader is ptraced by a task in the notified thread group, that had
+	// initiated or completed a group stop leaves the group stop, due to the
+	// child thread group or any task in the child thread group being sent
+	// SIGCONT.
+	EventGroupContinue
+)
+
+// WaitOptions controls the behavior of Task.Wait.
+type WaitOptions struct {
+	// If SpecificTID is non-zero, only events from the task with thread ID
+	// SpecificTID are eligible to be waited for. SpecificTID is resolved in
+	// the PID namespace of the waiter (the method receiver of Task.Wait). If
+	// no such task exists, or that task would not otherwise be eligible to be
+	// waited for by the waiting task, then there are no waitable tasks and
+	// Wait will return ECHILD.
+	SpecificTID ThreadID
+
+	// If SpecificPGID is non-zero, only events from ThreadGroups with a
+	// matching ProcessGroupID are eligible to be waited for. (Same
+	// constraints as SpecificTID apply.)
+	SpecificPGID ProcessGroupID
+
+	// Terminology note: Per waitpid(2), "a clone child is one which delivers
+	// no signal, or a signal other than SIGCHLD to its parent upon
+	// termination." In Linux, termination signal is technically a per-task
+	// property rather than a per-thread-group property. However, clone()
+	// forces no termination signal for tasks created with CLONE_THREAD, and
+	// execve() resets the termination signal to SIGCHLD, so all
+	// non-group-leader threads have no termination signal and are therefore
+	// "clone tasks".
+
+	// If NonCloneTasks is true, events from non-clone tasks are eligible to be
+	// waited for.
+	NonCloneTasks bool
+
+	// If CloneTasks is true, events from clone tasks are eligible to be waited
+	// for.
+	CloneTasks bool
+
+	// Events is a bitwise combination of the events defined above that specify
+	// what events are of interest to the call to Wait.
+	Events waiter.EventMask
+
+	// If ConsumeEvent is true, the Wait should consume the event such that it
+	// cannot be returned by a future Wait. Note that if a task exit is
+	// consumed in this way, in most cases the task will be reaped.
+	ConsumeEvent bool
+
+	// If BlockInterruptErr is not nil, Wait will block until either an event
+	// is available or there are no tasks that could produce a waitable event;
+	// if that blocking is interrupted, Wait returns BlockInterruptErr. If
+	// BlockInterruptErr is nil, Wait will not block.
+	BlockInterruptErr error
+}
+
+// Preconditions: The TaskSet mutex must be locked (for reading or writing).
+func (o *WaitOptions) matchesTask(t *Task, pidns *PIDNamespace) bool {
+	if o.SpecificTID != 0 && o.SpecificTID != pidns.tids[t] {
+		return false
+	}
+	if o.SpecificPGID != 0 && o.SpecificPGID != pidns.pgids[t.tg.processGroup] {
+		return false
+	}
+	if t == t.tg.leader && t.tg.terminationSignal == linux.SIGCHLD {
+		return o.NonCloneTasks
+	}
+	return o.CloneTasks
+}
+
+// ErrNoWaitableEvent is returned by non-blocking Task.Waits (e.g.
+// waitpid(WNOHANG)) that find no waitable events, but determine that waitable
+// events may exist in the future. (In contrast, if a non-blocking or blocking
+// Wait determines that there are no tasks that can produce a waitable event,
+// Task.Wait returns ECHILD.)
+var ErrNoWaitableEvent = errors.New("non-blocking Wait found eligible threads but no waitable events")
+
+// WaitResult contains information about a waited-for event.
+type WaitResult struct {
+	// Task is the task that reported the event.
+	Task *Task
+
+	// TID is the thread ID of Task in the PID namespace of the task that
+	// called Wait (that is, the method receiver of the call to Task.Wait). TID
+	// is provided because consuming exit waits cause the thread ID to be
+	// deallocated.
+	TID ThreadID
+
+	// UID is the real UID of Task in the user namespace of the task that
+	// called Wait.
+	UID auth.UID
+
+	// Event is exactly one of the events defined above.
+	Event waiter.EventMask
+
+	// Status is the numeric status associated with the event.
+	Status uint32
+}
+
+// Wait waits for an event from a thread group that is a child of t's thread
+// group, or a task in such a thread group, or a task that is ptraced by t,
+// subject to the options specified in opts.
+func (t *Task) Wait(opts *WaitOptions) (*WaitResult, error) {
+	if opts.BlockInterruptErr == nil {
+		return t.waitOnce(opts)
+	}
+	w, ch := waiter.NewChannelEntry(nil)
+	t.tg.eventQueue.EventRegister(&w, opts.Events)
+	defer t.tg.eventQueue.EventUnregister(&w)
+	for {
+		wr, err := t.waitOnce(opts)
+		if err != ErrNoWaitableEvent {
+			// This includes err == nil.
+			return wr, err
+		}
+		if err := t.Block(ch); err != nil {
+			return wr, syserror.ConvertIntr(err, opts.BlockInterruptErr)
+		}
+	}
+}
+
+func (t *Task) waitOnce(opts *WaitOptions) (*WaitResult, error) {
+	anyWaitableTasks := false
+
+	t.tg.pidns.owner.mu.Lock()
+	defer t.tg.pidns.owner.mu.Unlock()
+
+	// Without the (unimplemented) __WNOTHREAD flag, a task can wait on the
+	// children and tracees of any task in the same thread group.
+	for parent := t.tg.tasks.Front(); parent != nil; parent = parent.Next() {
+		for child := range parent.children {
+			if !opts.matchesTask(child, parent.tg.pidns) {
+				continue
+			}
+			// Non-leaders don't notify parents on exit and aren't eligible to
+			// be waited on.
+			if opts.Events&EventExit != 0 && child == child.tg.leader && !child.exitParentAcked {
+				anyWaitableTasks = true
+				if wr := t.waitCollectZombieLocked(child, opts, false); wr != nil {
+					return wr, nil
+				}
+			}
+			// Check for group stops and continues. Tasks that have passed
+			// TaskExitInitiated can no longer participate in group stops.
+			if opts.Events&(EventChildGroupStop|EventGroupContinue) == 0 {
+				continue
+			}
+			if child.exitState >= TaskExitInitiated {
+				continue
+			}
+			// If the waiter is in the same thread group as the task's
+			// tracer, do not report its group stops; they will be reported
+			// as ptrace stops instead. This also skips checking for group
+			// continues, but they'll be checked for when scanning tracees
+			// below. (Per kernel/exit.c:wait_consider_task(): "If a
+			// ptracer wants to distinguish the two events for its own
+			// children, it should create a separate process which takes
+			// the role of real parent.")
+			if tracer := child.Tracer(); tracer != nil && tracer.tg == parent.tg {
+				continue
+			}
+			anyWaitableTasks = true
+			if opts.Events&EventChildGroupStop != 0 {
+				if wr := t.waitCollectChildGroupStopLocked(child, opts); wr != nil {
+					return wr, nil
+				}
+			}
+			if opts.Events&EventGroupContinue != 0 {
+				if wr := t.waitCollectGroupContinueLocked(child, opts); wr != nil {
+					return wr, nil
+				}
+			}
+		}
+		for tracee := range parent.ptraceTracees {
+			if !opts.matchesTask(tracee, parent.tg.pidns) {
+				continue
+			}
+			// Non-leaders do notify tracers on exit.
+			if opts.Events&EventExit != 0 && !tracee.exitTracerAcked {
+				anyWaitableTasks = true
+				if wr := t.waitCollectZombieLocked(tracee, opts, true); wr != nil {
+					return wr, nil
+				}
+			}
+			if opts.Events&(EventTraceeStop|EventGroupContinue) == 0 {
+				continue
+			}
+			if tracee.exitState >= TaskExitInitiated {
+				continue
+			}
+			anyWaitableTasks = true
+			if opts.Events&EventTraceeStop != 0 {
+				if wr := t.waitCollectTraceeStopLocked(tracee, opts); wr != nil {
+					return wr, nil
+				}
+			}
+			if opts.Events&EventGroupContinue != 0 {
+				if wr := t.waitCollectGroupContinueLocked(tracee, opts); wr != nil {
+					return wr, nil
+				}
+			}
+		}
+	}
+
+	if anyWaitableTasks {
+		return nil, ErrNoWaitableEvent
+	}
+	return nil, syserror.ECHILD
+}
+
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) waitCollectZombieLocked(target *Task, opts *WaitOptions, asPtracer bool) *WaitResult {
+	if asPtracer && !target.exitTracerNotified {
+		return nil
+	}
+	if !asPtracer && !target.exitParentNotified {
+		return nil
+	}
+	// Zombied thread group leaders are never waitable until their thread group
+	// is otherwise empty. Usually this is caught by the
+	// target.exitParentNotified check above, but if t is both (in the thread
+	// group of) target's tracer and parent, asPtracer may be true.
+	if target == target.tg.leader && target.tg.tasksCount != 1 {
+		return nil
+	}
+	pid := t.tg.pidns.tids[target]
+	uid := target.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()
+	status := target.exitStatus.Status()
+	if !opts.ConsumeEvent {
+		return &WaitResult{
+			Task:   target,
+			TID:    pid,
+			UID:    uid,
+			Event:  EventExit,
+			Status: status,
+		}
+	}
+	// Surprisingly, the exit status reported by a non-consuming wait can
+	// differ from that reported by a consuming wait; the latter will return
+	// the group exit code if one is available.
+	if target.tg.exiting {
+		status = target.tg.exitStatus.Status()
+	}
+	// t may be (in the thread group of) target's parent, tracer, or both. We
+	// don't need to check for !exitTracerAcked because tracees are detached
+	// here, and we don't need to check for !exitParentAcked because zombies
+	// will be reaped here.
+	if tracer := target.Tracer(); tracer != nil && tracer.tg == t.tg && target.exitTracerNotified {
+		target.exitTracerAcked = true
+		target.ptraceTracer.Store((*Task)(nil))
+		delete(t.ptraceTracees, target)
+	}
+	if target.parent != nil && target.parent.tg == t.tg && target.exitParentNotified {
+		target.exitParentAcked = true
+		if target == target.tg.leader {
+			// target.tg.exitedCPUStats doesn't include target.CPUStats() yet,
+			// and won't until after target.exitNotifyLocked() (maybe). Include
+			// target.CPUStats() explicitly. This is consistent with Linux,
+			// which accounts an exited task's cputime to its thread group in
+			// kernel/exit.c:release_task() => __exit_signal(), and uses
+			// thread_group_cputime_adjusted() in wait_task_zombie().
+			t.tg.childCPUStats.Accumulate(target.CPUStats())
+			t.tg.childCPUStats.Accumulate(target.tg.exitedCPUStats)
+			t.tg.childCPUStats.Accumulate(target.tg.childCPUStats)
+			// Update t's child max resident set size. The size will be the maximum
+			// of this thread's size and all its childrens' sizes.
+			if t.tg.childMaxRSS < target.tg.maxRSS {
+				t.tg.childMaxRSS = target.tg.maxRSS
+			}
+			if t.tg.childMaxRSS < target.tg.childMaxRSS {
+				t.tg.childMaxRSS = target.tg.childMaxRSS
+			}
+		}
+	}
+	target.exitNotifyLocked(false)
+	return &WaitResult{
+		Task:   target,
+		TID:    pid,
+		UID:    uid,
+		Event:  EventExit,
+		Status: status,
+	}
+}
+
+// updateRSSLocked updates t.tg.maxRSS.
+//
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) updateRSSLocked() {
+	if mmMaxRSS := t.MemoryManager().MaxResidentSetSize(); t.tg.maxRSS < mmMaxRSS {
+		t.tg.maxRSS = mmMaxRSS
+	}
+}
+
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) waitCollectChildGroupStopLocked(target *Task, opts *WaitOptions) *WaitResult {
+	target.tg.signalHandlers.mu.Lock()
+	defer target.tg.signalHandlers.mu.Unlock()
+	if !target.tg.groupStopWaitable {
+		return nil
+	}
+	pid := t.tg.pidns.tids[target]
+	uid := target.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()
+	sig := target.tg.groupStopSignal
+	if opts.ConsumeEvent {
+		target.tg.groupStopWaitable = false
+	}
+	return &WaitResult{
+		Task:  target,
+		TID:   pid,
+		UID:   uid,
+		Event: EventChildGroupStop,
+		// There is no name for these status constants.
+		Status: (uint32(sig)&0xff)<<8 | 0x7f,
+	}
+}
+
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) waitCollectGroupContinueLocked(target *Task, opts *WaitOptions) *WaitResult {
+	target.tg.signalHandlers.mu.Lock()
+	defer target.tg.signalHandlers.mu.Unlock()
+	if !target.tg.groupContWaitable {
+		return nil
+	}
+	pid := t.tg.pidns.tids[target]
+	uid := target.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()
+	if opts.ConsumeEvent {
+		target.tg.groupContWaitable = false
+	}
+	return &WaitResult{
+		Task:   target,
+		TID:    pid,
+		UID:    uid,
+		Event:  EventGroupContinue,
+		Status: 0xffff,
+	}
+}
+
+// Preconditions: The TaskSet mutex must be locked for writing.
+func (t *Task) waitCollectTraceeStopLocked(target *Task, opts *WaitOptions) *WaitResult {
+	target.tg.signalHandlers.mu.Lock()
+	defer target.tg.signalHandlers.mu.Unlock()
+	if target.stop == nil {
+		return nil
+	}
+	if _, ok := target.stop.(*ptraceStop); !ok {
+		return nil
+	}
+	if target.ptraceCode == 0 {
+		return nil
+	}
+	pid := t.tg.pidns.tids[target]
+	uid := target.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()
+	code := target.ptraceCode
+	if opts.ConsumeEvent {
+		target.ptraceCode = 0
+	}
+	return &WaitResult{
+		Task:   target,
+		TID:    pid,
+		UID:    uid,
+		Event:  EventTraceeStop,
+		Status: uint32(code)<<8 | 0x7f,
+	}
+}
+
+// ExitState returns t's current progress through the exit path.
+func (t *Task) ExitState() TaskExitState {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	return t.exitState
+}
+
+// ParentDeathSignal returns t's parent death signal.
+func (t *Task) ParentDeathSignal() linux.Signal {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.parentDeathSignal
+}
+
+// SetParentDeathSignal sets t's parent death signal.
+func (t *Task) SetParentDeathSignal(sig linux.Signal) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.parentDeathSignal = sig
+}
diff --git a/pkg/sentry/kernel/task_identity.go b/pkg/sentry/kernel/task_identity.go
new file mode 100644
index 000000000..a51fa9d7e
--- /dev/null
+++ b/pkg/sentry/kernel/task_identity.go
@@ -0,0 +1,557 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Credentials returns t's credentials by value.
+func (t *Task) Credentials() auth.Credentials {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return *t.creds // Copy out with lock held.
+}
+
+// UserNamespace returns the user namespace associated with the task.
+func (t *Task) UserNamespace() *auth.UserNamespace {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.creds.UserNamespace
+}
+
+// HasCapabilityIn checks if the task has capability cp in user namespace ns.
+func (t *Task) HasCapabilityIn(cp linux.Capability, ns *auth.UserNamespace) bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.creds.HasCapabilityIn(cp, ns)
+}
+
+// HasCapability checks if the task has capability cp in its user namespace.
+func (t *Task) HasCapability(cp linux.Capability) bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.creds.HasCapability(cp)
+}
+
+// SetUID implements the semantics of setuid(2).
+func (t *Task) SetUID(uid auth.UID) error {
+	// setuid considers -1 to be invalid.
+	if !uid.Ok() {
+		return syserror.EINVAL
+	}
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	kuid := t.creds.UserNamespace.MapToKUID(uid)
+	if !kuid.Ok() {
+		return syserror.EINVAL
+	}
+	// "setuid() sets the effective user ID of the calling process. If the
+	// effective UID of the caller is root (more precisely: if the caller has
+	// the CAP_SETUID capability), the real UID and saved set-user-ID are also
+	// set." - setuid(2)
+	if t.creds.HasCapability(linux.CAP_SETUID) {
+		t.setKUIDsUncheckedLocked(kuid, kuid, kuid)
+		return nil
+	}
+	// "EPERM: The user is not privileged (Linux: does not have the CAP_SETUID
+	// capability) and uid does not match the real UID or saved set-user-ID of
+	// the calling process."
+	if kuid != t.creds.RealKUID && kuid != t.creds.SavedKUID {
+		return syserror.EPERM
+	}
+	t.setKUIDsUncheckedLocked(t.creds.RealKUID, kuid, t.creds.SavedKUID)
+	return nil
+}
+
+// SetREUID implements the semantics of setreuid(2).
+func (t *Task) SetREUID(r, e auth.UID) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	// "Supplying a value of -1 for either the real or effective user ID forces
+	// the system to leave that ID unchanged." - setreuid(2)
+	newR := t.creds.RealKUID
+	if r.Ok() {
+		newR = t.creds.UserNamespace.MapToKUID(r)
+		if !newR.Ok() {
+			return syserror.EINVAL
+		}
+	}
+	newE := t.creds.EffectiveKUID
+	if e.Ok() {
+		newE = t.creds.UserNamespace.MapToKUID(e)
+		if !newE.Ok() {
+			return syserror.EINVAL
+		}
+	}
+	if !t.creds.HasCapability(linux.CAP_SETUID) {
+		// "Unprivileged processes may only set the effective user ID to the
+		// real user ID, the effective user ID, or the saved set-user-ID."
+		if newE != t.creds.RealKUID && newE != t.creds.EffectiveKUID && newE != t.creds.SavedKUID {
+			return syserror.EPERM
+		}
+		// "Unprivileged users may only set the real user ID to the real user
+		// ID or the effective user ID."
+		if newR != t.creds.RealKUID && newR != t.creds.EffectiveKUID {
+			return syserror.EPERM
+		}
+	}
+	// "If the real user ID is set (i.e., ruid is not -1) or the effective user
+	// ID is set to a value not equal to the previous real user ID, the saved
+	// set-user-ID will be set to the new effective user ID."
+	newS := t.creds.SavedKUID
+	if r.Ok() || (e.Ok() && newE != t.creds.EffectiveKUID) {
+		newS = newE
+	}
+	t.setKUIDsUncheckedLocked(newR, newE, newS)
+	return nil
+}
+
+// SetRESUID implements the semantics of the setresuid(2) syscall.
+func (t *Task) SetRESUID(r, e, s auth.UID) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	// "Unprivileged user processes may change the real UID, effective UID, and
+	// saved set-user-ID, each to one of: the current real UID, the current
+	// effective UID or the current saved set-user-ID. Privileged processes (on
+	// Linux, those having the CAP_SETUID capability) may set the real UID,
+	// effective UID, and saved set-user-ID to arbitrary values. If one of the
+	// arguments equals -1, the corresponding value is not changed." -
+	// setresuid(2)
+	var err error
+	newR := t.creds.RealKUID
+	if r.Ok() {
+		newR, err = t.creds.UseUID(r)
+		if err != nil {
+			return err
+		}
+	}
+	newE := t.creds.EffectiveKUID
+	if e.Ok() {
+		newE, err = t.creds.UseUID(e)
+		if err != nil {
+			return err
+		}
+	}
+	newS := t.creds.SavedKUID
+	if s.Ok() {
+		newS, err = t.creds.UseUID(s)
+		if err != nil {
+			return err
+		}
+	}
+	t.setKUIDsUncheckedLocked(newR, newE, newS)
+	return nil
+}
+
+// Preconditions: t.mu must be locked.
+func (t *Task) setKUIDsUncheckedLocked(newR, newE, newS auth.KUID) {
+	root := t.creds.UserNamespace.MapToKUID(auth.RootUID)
+	oldR, oldE, oldS := t.creds.RealKUID, t.creds.EffectiveKUID, t.creds.SavedKUID
+	t.creds.RealKUID, t.creds.EffectiveKUID, t.creds.SavedKUID = newR, newE, newS
+
+	// "1. If one or more of the real, effective or saved set user IDs was
+	// previously 0, and as a result of the UID changes all of these IDs have a
+	// nonzero value, then all capabilities are cleared from the permitted and
+	// effective capability sets." - capabilities(7)
+	if (oldR == root || oldE == root || oldS == root) && (newR != root && newE != root && newS != root) {
+		// prctl(2): "PR_SET_KEEPCAP: Set the state of the calling thread's
+		// "keep capabilities" flag, which determines whether the thread's permitted
+		// capability set is cleared when a change is made to the
+		// thread's user IDs such that the thread's real UID, effective
+		// UID, and saved set-user-ID all become nonzero when at least
+		// one of them previously had the value 0.  By default, the
+		// permitted capability set is cleared when such a change is
+		// made; setting the "keep capabilities" flag prevents it from
+		// being cleared." (A thread's effective capability set is always
+		// cleared when such a credential change is made,
+		// regardless of the setting of the "keep capabilities" flag.)
+		if !t.creds.KeepCaps {
+			t.creds.PermittedCaps = 0
+			t.creds.EffectiveCaps = 0
+		}
+	}
+	// """
+	// 2. If the effective user ID is changed from 0 to nonzero, then all
+	// capabilities are cleared from the effective set.
+	//
+	// 3. If the effective user ID is changed from nonzero to 0, then the
+	// permitted set is copied to the effective set.
+	// """
+	if oldE == root && newE != root {
+		t.creds.EffectiveCaps = 0
+	} else if oldE != root && newE == root {
+		t.creds.EffectiveCaps = t.creds.PermittedCaps
+	}
+	// "4. If the filesystem user ID is changed from 0 to nonzero (see
+	// setfsuid(2)), then the following capabilities are cleared from the
+	// effective set: ..."
+	// (filesystem UIDs aren't implemented, nor are any of the capabilities in
+	// question)
+
+	// Not documented, but compare Linux's kernel/cred.c:commit_creds().
+	if oldE != newE {
+		t.parentDeathSignal = 0
+	}
+}
+
+// SetGID implements the semantics of setgid(2).
+func (t *Task) SetGID(gid auth.GID) error {
+	if !gid.Ok() {
+		return syserror.EINVAL
+	}
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	kgid := t.creds.UserNamespace.MapToKGID(gid)
+	if !kgid.Ok() {
+		return syserror.EINVAL
+	}
+	if t.creds.HasCapability(linux.CAP_SETGID) {
+		t.setKGIDsUncheckedLocked(kgid, kgid, kgid)
+		return nil
+	}
+	if kgid != t.creds.RealKGID && kgid != t.creds.SavedKGID {
+		return syserror.EPERM
+	}
+	t.setKGIDsUncheckedLocked(t.creds.RealKGID, kgid, t.creds.SavedKGID)
+	return nil
+}
+
+// SetREGID implements the semantics of setregid(2).
+func (t *Task) SetREGID(r, e auth.GID) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	newR := t.creds.RealKGID
+	if r.Ok() {
+		newR = t.creds.UserNamespace.MapToKGID(r)
+		if !newR.Ok() {
+			return syserror.EINVAL
+		}
+	}
+	newE := t.creds.EffectiveKGID
+	if e.Ok() {
+		newE = t.creds.UserNamespace.MapToKGID(e)
+		if !newE.Ok() {
+			return syserror.EINVAL
+		}
+	}
+	if !t.creds.HasCapability(linux.CAP_SETGID) {
+		if newE != t.creds.RealKGID && newE != t.creds.EffectiveKGID && newE != t.creds.SavedKGID {
+			return syserror.EPERM
+		}
+		if newR != t.creds.RealKGID && newR != t.creds.EffectiveKGID {
+			return syserror.EPERM
+		}
+	}
+	newS := t.creds.SavedKGID
+	if r.Ok() || (e.Ok() && newE != t.creds.EffectiveKGID) {
+		newS = newE
+	}
+	t.setKGIDsUncheckedLocked(newR, newE, newS)
+	return nil
+}
+
+// SetRESGID implements the semantics of the setresgid(2) syscall.
+func (t *Task) SetRESGID(r, e, s auth.GID) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	var err error
+	newR := t.creds.RealKGID
+	if r.Ok() {
+		newR, err = t.creds.UseGID(r)
+		if err != nil {
+			return err
+		}
+	}
+	newE := t.creds.EffectiveKGID
+	if e.Ok() {
+		newE, err = t.creds.UseGID(e)
+		if err != nil {
+			return err
+		}
+	}
+	newS := t.creds.SavedKGID
+	if s.Ok() {
+		newS, err = t.creds.UseGID(s)
+		if err != nil {
+			return err
+		}
+	}
+	t.setKGIDsUncheckedLocked(newR, newE, newS)
+	return nil
+}
+
+func (t *Task) setKGIDsUncheckedLocked(newR, newE, newS auth.KGID) {
+	oldE := t.creds.EffectiveKGID
+	t.creds.RealKGID, t.creds.EffectiveKGID, t.creds.SavedKGID = newR, newE, newS
+
+	// Not documented, but compare Linux's kernel/cred.c:commit_creds().
+	if oldE != newE {
+		t.parentDeathSignal = 0
+	}
+}
+
+// SetExtraGIDs attempts to change t's supplemental groups. All IDs are
+// interpreted as being in t's user namespace.
+func (t *Task) SetExtraGIDs(gids []auth.GID) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if !t.creds.HasCapability(linux.CAP_SETGID) {
+		return syserror.EPERM
+	}
+	kgids := make([]auth.KGID, len(gids))
+	for i, gid := range gids {
+		kgid := t.creds.UserNamespace.MapToKGID(gid)
+		if !kgid.Ok() {
+			return syserror.EINVAL
+		}
+		kgids[i] = kgid
+	}
+	t.creds.ExtraKGIDs = kgids
+	return nil
+}
+
+// SetCapabilitySets attempts to change t's permitted, inheritable, and
+// effective capability sets.
+func (t *Task) SetCapabilitySets(permitted, inheritable, effective auth.CapabilitySet) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	// "Permitted: This is a limiting superset for the effective capabilities
+	// that the thread may assume." - capabilities(7)
+	if effective & ^permitted != 0 {
+		return syserror.EPERM
+	}
+	// "It is also a limiting superset for the capabilities that may be added
+	// to the inheritable set by a thread that does not have the CAP_SETPCAP
+	// capability in its effective set."
+	if !t.creds.HasCapability(linux.CAP_SETPCAP) && (inheritable & ^(t.creds.InheritableCaps|t.creds.PermittedCaps) != 0) {
+		return syserror.EPERM
+	}
+	// "If a thread drops a capability from its permitted set, it can never
+	// reacquire that capability (unless it execve(2)s ..."
+	if permitted & ^t.creds.PermittedCaps != 0 {
+		return syserror.EPERM
+	}
+	// "... if a capability is not in the bounding set, then a thread can't add
+	// this capability to its inheritable set, even if it was in its permitted
+	// capabilities ..."
+	if inheritable & ^(t.creds.InheritableCaps|t.creds.BoundingCaps) != 0 {
+		return syserror.EPERM
+	}
+	t.creds.PermittedCaps = permitted
+	t.creds.InheritableCaps = inheritable
+	t.creds.EffectiveCaps = effective
+	return nil
+}
+
+// DropBoundingCapability attempts to drop capability cp from t's capability
+// bounding set.
+func (t *Task) DropBoundingCapability(cp linux.Capability) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if !t.creds.HasCapability(linux.CAP_SETPCAP) {
+		return syserror.EPERM
+	}
+	t.creds.BoundingCaps &^= auth.CapabilitySetOf(cp)
+	return nil
+}
+
+// SetUserNamespace attempts to move c into ns.
+func (t *Task) SetUserNamespace(ns *auth.UserNamespace) error {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	// "A process reassociating itself with a user namespace must have the
+	// CAP_SYS_ADMIN capability in the target user namespace." - setns(2)
+	//
+	// If t just created ns, then t.creds is guaranteed to have CAP_SYS_ADMIN
+	// in ns (by rule 3 in auth.Credentials.HasCapability).
+	if !t.creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, ns) {
+		return syserror.EPERM
+	}
+
+	t.creds.UserNamespace = ns
+	// "The child process created by clone(2) with the CLONE_NEWUSER flag
+	// starts out with a complete set of capabilities in the new user
+	// namespace. Likewise, a process that creates a new user namespace using
+	// unshare(2) or joins an existing user namespace using setns(2) gains a
+	// full set of capabilities in that namespace."
+	t.creds.PermittedCaps = auth.AllCapabilities
+	t.creds.InheritableCaps = 0
+	t.creds.EffectiveCaps = auth.AllCapabilities
+	t.creds.BoundingCaps = auth.AllCapabilities
+	// "A call to clone(2), unshare(2), or setns(2) using the CLONE_NEWUSER
+	// flag sets the "securebits" flags (see capabilities(7)) to their default
+	// values (all flags disabled) in the child (for clone(2)) or caller (for
+	// unshare(2), or setns(2)." - user_namespaces(7)
+	t.creds.KeepCaps = false
+
+	return nil
+}
+
+// SetKeepCaps will set the keep capabilities flag PR_SET_KEEPCAPS.
+func (t *Task) SetKeepCaps(k bool) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.creds.KeepCaps = k
+}
+
+// updateCredsForExec updates t.creds to reflect an execve().
+//
+// NOTE: We currently do not implement privileged executables
+// (set-user/group-ID bits and file capabilities). This allows us to make a lot
+// of simplifying assumptions:
+//
+// - We assume the no_new_privs bit (set by prctl(SET_NO_NEW_PRIVS)), which
+// disables the features we don't support anyway, is always set. This
+// drastically simplifies this function.
+//
+// - We don't implement AT_SECURE, because no_new_privs always being set means
+// that the conditions that require AT_SECURE never arise. (Compare Linux's
+// security/commoncap.c:cap_bprm_set_creds() and cap_bprm_secureexec().)
+//
+// - We don't check for CAP_SYS_ADMIN in prctl(PR_SET_SECCOMP), since
+// seccomp-bpf is also allowed if the task has no_new_privs set.
+//
+// - Task.ptraceAttach does not serialize with execve as it does in Linux,
+// since no_new_privs being set has the same effect as the presence of an
+// unprivileged tracer.
+//
+// Preconditions: t.mu must be locked.
+func (t *Task) updateCredsForExecLocked() {
+	// """
+	// During an execve(2), the kernel calculates the new capabilities of
+	// the process using the following algorithm:
+	//
+	//     P'(permitted) = (P(inheritable) & F(inheritable)) |
+	//                     (F(permitted) & cap_bset)
+	//
+	//     P'(effective) = F(effective) ? P'(permitted) : 0
+	//
+	//     P'(inheritable) = P(inheritable)    [i.e., unchanged]
+	//
+	// where:
+	//
+	//     P         denotes the value of a thread capability set before the
+	//               execve(2)
+	//
+	//     P'        denotes the value of a thread capability set after the
+	//               execve(2)
+	//
+	//     F         denotes a file capability set
+	//
+	//     cap_bset  is the value of the capability bounding set
+	//
+	// ...
+	//
+	// In order to provide an all-powerful root using capability sets, during
+	// an execve(2):
+	//
+	// 1. If a set-user-ID-root program is being executed, or the real user ID
+	// of the process is 0 (root) then the file inheritable and permitted sets
+	// are defined to be all ones (i.e. all capabilities enabled).
+	//
+	// 2. If a set-user-ID-root program is being executed, then the file
+	// effective bit is defined to be one (enabled).
+	//
+	// The upshot of the above rules, combined with the capabilities
+	// transformations described above, is that when a process execve(2)s a
+	// set-user-ID-root program, or when a process with an effective UID of 0
+	// execve(2)s a program, it gains all capabilities in its permitted and
+	// effective capability sets, except those masked out by the capability
+	// bounding set.
+	// """ - capabilities(7)
+	// (ambient capability sets omitted)
+	//
+	// As the last paragraph implies, the case of "a set-user-ID root program
+	// is being executed" also includes the case where (namespace) root is
+	// executing a non-set-user-ID program; the actual check is just based on
+	// the effective user ID.
+	var newPermitted auth.CapabilitySet // since F(inheritable) == F(permitted) == 0
+	fileEffective := false
+	root := t.creds.UserNamespace.MapToKUID(auth.RootUID)
+	if t.creds.EffectiveKUID == root || t.creds.RealKUID == root {
+		newPermitted = t.creds.InheritableCaps | t.creds.BoundingCaps
+		if t.creds.EffectiveKUID == root {
+			fileEffective = true
+		}
+	}
+
+	// Now we enter poorly-documented, somewhat confusing territory. (The
+	// accompanying comment in Linux's security/commoncap.c:cap_bprm_set_creds
+	// is not very helpful.) My reading of it is:
+	//
+	// If at least one of the following is true:
+	//
+	// A1. The execing task is ptraced, and the tracer did not have
+	// CAP_SYS_PTRACE in the execing task's user namespace at the time of
+	// PTRACE_ATTACH.
+	//
+	// A2. The execing task shares its FS context with at least one task in
+	// another thread group.
+	//
+	// A3. The execing task has no_new_privs set.
+	//
+	// AND at least one of the following is true:
+	//
+	// B1. The new effective user ID (which may come from set-user-ID, or be the
+	// execing task's existing effective user ID) is not equal to the task's
+	// real UID.
+	//
+	// B2. The new effective group ID (which may come from set-group-ID, or be
+	// the execing task's existing effective group ID) is not equal to the
+	// task's real GID.
+	//
+	// B3. The new permitted capability set contains capabilities not in the
+	// task's permitted capability set.
+	//
+	// Then:
+	//
+	// C1. Limit the new permitted capability set to the task's permitted
+	// capability set.
+	//
+	// C2. If either the task does not have CAP_SETUID in its user namespace, or
+	// the task has no_new_privs set, force the new effective UID and GID to
+	// the task's real UID and GID.
+	//
+	// But since no_new_privs is always set (A3 is always true), this becomes
+	// much simpler. If B1 and B2 are false, C2 is a no-op. If B3 is false, C1
+	// is a no-op. So we can just do C1 and C2 unconditionally.
+	if t.creds.EffectiveKUID != t.creds.RealKUID || t.creds.EffectiveKGID != t.creds.RealKGID {
+		t.creds.EffectiveKUID = t.creds.RealKUID
+		t.creds.EffectiveKGID = t.creds.RealKGID
+		t.parentDeathSignal = 0
+	}
+	// (Saved set-user-ID is always set to the new effective user ID, and saved
+	// set-group-ID is always set to the new effective group ID, regardless of
+	// the above.)
+	t.creds.SavedKUID = t.creds.RealKUID
+	t.creds.SavedKGID = t.creds.RealKGID
+	t.creds.PermittedCaps &= newPermitted
+	if fileEffective {
+		t.creds.EffectiveCaps = t.creds.PermittedCaps
+	} else {
+		t.creds.EffectiveCaps = 0
+	}
+
+	// prctl(2): The "keep capabilities" value will be reset to 0 on subsequent
+	// calls to execve(2).
+	t.creds.KeepCaps = false
+
+	// "The bounding set is inherited at fork(2) from the thread's parent, and
+	// is preserved across an execve(2)". So we're done.
+}
diff --git a/pkg/sentry/kernel/task_log.go b/pkg/sentry/kernel/task_log.go
new file mode 100644
index 000000000..18efacb19
--- /dev/null
+++ b/pkg/sentry/kernel/task_log.go
@@ -0,0 +1,137 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"sort"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// maxStackDebugBytes is the maximum number of user stack bytes that may be
+	// printed by debugDumpStack.
+	maxStackDebugBytes = 1024
+)
+
+// Infof logs an formatted info message by calling log.Infof.
+func (t *Task) Infof(fmt string, v ...interface{}) {
+	if log.IsLogging(log.Info) {
+		log.Infof(t.logPrefix.Load().(string)+fmt, v...)
+	}
+}
+
+// Warningf logs a warning string by calling log.Warningf.
+func (t *Task) Warningf(fmt string, v ...interface{}) {
+	if log.IsLogging(log.Warning) {
+		log.Warningf(t.logPrefix.Load().(string)+fmt, v...)
+	}
+}
+
+// Debugf creates a debug string that includes the task ID.
+func (t *Task) Debugf(fmt string, v ...interface{}) {
+	if log.IsLogging(log.Debug) {
+		log.Debugf(t.logPrefix.Load().(string)+fmt, v...)
+	}
+}
+
+// IsLogging returns true iff this level is being logged.
+func (t *Task) IsLogging(level log.Level) bool {
+	return log.IsLogging(level)
+}
+
+// DebugDumpState logs task state at log level debug.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) DebugDumpState() {
+	t.debugDumpRegisters()
+	t.debugDumpStack()
+	if mm := t.MemoryManager(); mm != nil {
+		t.Debugf("Mappings:\n%s", mm)
+	}
+	t.Debugf("FDMap:\n%s", t.FDMap())
+}
+
+// debugDumpRegisters logs register state at log level debug.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) debugDumpRegisters() {
+	if !t.IsLogging(log.Debug) {
+		return
+	}
+	regmap, err := t.Arch().RegisterMap()
+	if err != nil {
+		t.Debugf("Registers: %v", err)
+	} else {
+		t.Debugf("Registers:")
+		var regs []string
+		for reg := range regmap {
+			regs = append(regs, reg)
+		}
+		sort.Strings(regs)
+		for _, reg := range regs {
+			t.Debugf("%-8s = %016x", reg, regmap[reg])
+		}
+	}
+}
+
+// debugDumpStack logs user stack contents at log level debug.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) debugDumpStack() {
+	if !t.IsLogging(log.Debug) {
+		return
+	}
+	m := t.MemoryManager()
+	if m == nil {
+		t.Debugf("Memory manager for task is gone, skipping application stack dump.")
+		return
+	}
+	t.Debugf("Stack:")
+	start := usermem.Addr(t.Arch().Stack())
+	// Round addr down to a 16-byte boundary.
+	start &= ^usermem.Addr(15)
+	// Print 16 bytes per line, one byte at a time.
+	for offset := uint64(0); offset < maxStackDebugBytes; offset += 16 {
+		addr, ok := start.AddLength(offset)
+		if !ok {
+			break
+		}
+		var data [16]byte
+		n, err := m.CopyIn(t, addr, data[:], usermem.IOOpts{
+			IgnorePermissions: true,
+		})
+		// Print as much of the line as we can, even if an error was
+		// encountered.
+		if n > 0 {
+			t.Debugf("%x: % x", addr, data[:n])
+		}
+		if err != nil {
+			t.Debugf("Error reading stack at address %x: %v", addr+usermem.Addr(n), err)
+			break
+		}
+	}
+}
+
+// updateLogPrefix updates the task's cached log prefix to reflect its
+// current thread ID.
+//
+// Preconditions: The task's owning TaskSet.mu must be locked.
+func (t *Task) updateLogPrefixLocked() {
+	// Use the task's TID in the root PID namespace for logging.
+	t.logPrefix.Store(fmt.Sprintf("[% 4d] ", t.tg.pidns.owner.Root.tids[t]))
+}
diff --git a/pkg/sentry/kernel/task_net.go b/pkg/sentry/kernel/task_net.go
new file mode 100644
index 000000000..4df2e53d3
--- /dev/null
+++ b/pkg/sentry/kernel/task_net.go
@@ -0,0 +1,35 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+)
+
+// IsNetworkNamespaced returns true if t is in a non-root network namespace.
+func (t *Task) IsNetworkNamespaced() bool {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.netns
+}
+
+// NetworkContext returns the network stack used by the task. NetworkContext
+// may return nil if no network stack is available.
+func (t *Task) NetworkContext() inet.Stack {
+	if t.IsNetworkNamespaced() {
+		return nil
+	}
+	return t.k.networkStack
+}
diff --git a/pkg/sentry/kernel/task_resources.go b/pkg/sentry/kernel/task_resources.go
new file mode 100644
index 000000000..e529f0c2d
--- /dev/null
+++ b/pkg/sentry/kernel/task_resources.go
@@ -0,0 +1,126 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+)
+
+// TaskResources is the subset of a task's data provided by its creator that is
+// not provided by the loader.
+type TaskResources struct {
+	// SignalMask is the set of signals whose delivery is currently blocked.
+	//
+	// FIXME: Determine if we also need RealSignalMask
+	SignalMask linux.SignalSet
+
+	// FSContext is the filesystem context.
+	*FSContext
+
+	// FDMap provides access to files to the task.
+	*FDMap
+
+	// Tracks abstract sockets that are in use.
+	AbstractSockets *AbstractSocketNamespace
+}
+
+// newTaskResources returns a new TaskResources, taking an additional reference
+// on fdm.
+func newTaskResources(fdm *FDMap, fc *FSContext) *TaskResources {
+	fdm.IncRef()
+	return &TaskResources{
+		FDMap:           fdm,
+		FSContext:       fc,
+		AbstractSockets: NewAbstractSocketNamespace(),
+	}
+}
+
+// release releases all resources held by the TaskResources. release is called
+// by the task when it exits.
+func (tr *TaskResources) release() {
+	tr.FDMap.DecRef()
+	tr.FDMap = nil
+	tr.FSContext.DecRef()
+	tr.FSContext = nil
+	tr.AbstractSockets = nil
+}
+
+// Fork returns a duplicate of tr.
+//
+// FIXME: Preconditions: When tr is owned by a Task, that task's
+// signal mutex must be locked, or Fork must be called by the task's goroutine.
+func (tr *TaskResources) Fork(shareFiles bool, shareFSContext bool) *TaskResources {
+	var fdmap *FDMap
+	if shareFiles {
+		fdmap = tr.FDMap
+		fdmap.IncRef()
+	} else {
+		fdmap = tr.FDMap.Fork()
+	}
+
+	var fsc *FSContext
+	if shareFSContext {
+		fsc = tr.FSContext
+		fsc.IncRef()
+	} else {
+		fsc = tr.FSContext.Fork()
+	}
+
+	return &TaskResources{
+		SignalMask:      tr.SignalMask,
+		FDMap:           fdmap,
+		FSContext:       fsc,
+		AbstractSockets: tr.AbstractSockets,
+	}
+}
+
+// FDMap returns t's FDMap.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) FDMap() *FDMap {
+	return t.tr.FDMap
+}
+
+// FSContext returns t's FSContext.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) FSContext() *FSContext {
+	return t.tr.FSContext
+}
+
+// MountNamespace returns t's MountNamespace. MountNamespace does not take an additional
+// reference on the returned MountNamespace.
+func (t *Task) MountNamespace() *fs.MountNamespace {
+	return t.k.mounts
+}
+
+// AbstractSockets returns t's AbstractSocketNamespace.
+func (t *Task) AbstractSockets() *AbstractSocketNamespace {
+	return t.tr.AbstractSockets
+}
+
+// IsChrooted returns true if the root directory of t's FSContext is not the
+// root directory of t's MountNamespace.
+//
+// Preconditions: The caller must be running on the task goroutine, or t.mu
+// must be locked.
+func (t *Task) IsChrooted() bool {
+	realRoot := t.k.mounts.Root()
+	defer realRoot.DecRef()
+	return t.tr.FSContext.root != realRoot
+}
diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go
new file mode 100644
index 000000000..94ce5582b
--- /dev/null
+++ b/pkg/sentry/kernel/task_run.go
@@ -0,0 +1,346 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"bytes"
+	"runtime"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// A taskRunState is a reified state in the task state machine. See README.md
+// for details. The canonical list of all run states, as well as transitions
+// between them, is given in run_states.dot.
+//
+// The set of possible states is enumerable and completely defined by the
+// kernel package, so taskRunState would ideally be represented by a
+// discriminated union. However, Go does not support sum types.
+//
+// Hence, as with TaskStop, data-free taskRunStates should be represented as
+// typecast nils to avoid unnecessary allocation.
+type taskRunState interface {
+	// execute executes the code associated with this state over the given task
+	// and returns the following state. If execute returns nil, the task
+	// goroutine should exit.
+	//
+	// It is valid to tail-call a following state's execute to avoid the
+	// overhead of converting the following state to an interface object and
+	// checking for stops, provided that the tail-call cannot recurse.
+	execute(*Task) taskRunState
+}
+
+// run runs the task goroutine.
+//
+// threadID a dummy value set to the task's TID in the root PID namespace to
+// make it visible in stack dumps. A goroutine for a given task can be identified
+// searching for Task.run()'s argument value.
+func (t *Task) run(threadID uintptr) {
+	// Construct t.blockingTimer here. We do this here because we can't
+	// reconstruct t.blockingTimer during restore in Task.afterLoad(), because
+	// kernel.timekeeper.SetClocks() hasn't been called yet.
+	blockingTimerNotifier, blockingTimerChan := ktime.NewChannelNotifier()
+	t.blockingTimer = ktime.NewTimer(t.k.MonotonicClock(), blockingTimerNotifier)
+	defer t.blockingTimer.Destroy()
+	t.blockingTimerChan = blockingTimerChan
+
+	// Activate our address space.
+	t.Activate()
+	// The corresponding t.Deactivate occurs in the exit path
+	// (runExitMain.execute) so that when
+	// Platform.CooperativelySharesAddressSpace() == true, we give up the
+	// AddressSpace before the task goroutine finishes executing.
+
+	// Ensure that thread group timers for execution time reflect that this
+	// task now exists.
+	t.tg.tm.kick()
+
+	// If this is a newly-started task, it should check for participation in
+	// group stops. If this is a task resuming after restore, it was
+	// interrupted by saving. In either case, the task is initially
+	// interrupted.
+	t.interruptSelf()
+
+	for {
+		// Explanation for this ordering:
+		//
+		// - A freshly-started task that is stopped should not do anything
+		// before it enters the stop.
+		//
+		// - If taskRunState.execute returns nil, the task goroutine should
+		// exit without checking for a stop.
+		//
+		// - Task.Start won't start Task.run if t.runState is nil, so this
+		// ordering is safe.
+		t.doStop()
+		t.runState = t.runState.execute(t)
+		if t.runState == nil {
+			t.accountTaskGoroutineEnter(TaskGoroutineNonexistent)
+			t.goroutineStopped.Done()
+			t.tg.liveGoroutines.Done()
+			t.tg.pidns.owner.liveGoroutines.Done()
+			t.tg.pidns.owner.runningGoroutines.Done()
+
+			// Keep argument alive because stack trace for dead variables may not be correct.
+			runtime.KeepAlive(threadID)
+			return
+		}
+	}
+}
+
+// doStop is called by Task.run to block until the task is not stopped.
+func (t *Task) doStop() {
+	if atomic.LoadInt32(&t.stopCount) == 0 {
+		return
+	}
+	t.Deactivate()
+	// NOTE: t.Activate() must be called without any locks held, so
+	// this defer must precede the defer for unlocking the signal mutex.
+	defer t.Activate()
+	t.accountTaskGoroutineEnter(TaskGoroutineStopped)
+	defer t.accountTaskGoroutineLeave(TaskGoroutineStopped)
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.tg.pidns.owner.runningGoroutines.Add(-1)
+	defer t.tg.pidns.owner.runningGoroutines.Add(1)
+	t.goroutineStopped.Add(-1)
+	defer t.goroutineStopped.Add(1)
+	for t.stopCount > 0 {
+		t.endStopCond.Wait()
+	}
+}
+
+// The runApp state checks for interrupts before executing untrusted
+// application code.
+type runApp struct{}
+
+func (*runApp) execute(t *Task) taskRunState {
+	if t.interrupted() {
+		// Checkpointing instructs tasks to stop by sending an interrupt, so we
+		// must check for stops before entering runInterrupt (instead of
+		// tail-calling it).
+		return (*runInterrupt)(nil)
+	}
+
+	// We're about to switch to the application again. If there's still a
+	// unhandled SyscallRestartErrno that wasn't translated to an EINTR,
+	// restart the syscall that was interrupted. If there's a saved signal
+	// mask, restore it. (Note that restoring the saved signal mask may unblock
+	// a pending signal, causing another interruption, but that signal should
+	// not interact with the interrupted syscall.)
+	if t.haveSyscallReturn {
+		if sre, ok := SyscallRestartErrnoFromReturn(t.Arch().Return()); ok {
+			if sre == ERESTART_RESTARTBLOCK {
+				t.Debugf("Restarting syscall %d with restart block after errno %d: not interrupted by handled signal", t.Arch().SyscallNo(), sre)
+				t.Arch().RestartSyscallWithRestartBlock()
+			} else {
+				t.Debugf("Restarting syscall %d after errno %d: not interrupted by handled signal", t.Arch().SyscallNo(), sre)
+				t.Arch().RestartSyscall()
+			}
+		}
+		t.haveSyscallReturn = false
+	}
+	if t.haveSavedSignalMask {
+		t.SetSignalMask(t.savedSignalMask)
+		t.haveSavedSignalMask = false
+		if t.interrupted() {
+			return (*runInterrupt)(nil)
+		}
+	}
+
+	// Apply restartable sequences.
+	if t.rseqPreempted {
+		t.rseqPreempted = false
+		if t.rseqCPUAddr != 0 {
+			if err := t.rseqCopyOutCPU(); err != nil {
+				t.Warningf("Failed to copy CPU to %#x for RSEQ: %v", t.rseqCPUAddr, err)
+				t.forceSignal(linux.SIGSEGV, false)
+				t.SendSignal(sigPriv(linux.SIGSEGV))
+				// Re-enter the task run loop for signal delivery.
+				return (*runApp)(nil)
+			}
+		}
+		t.rseqInterrupt()
+	}
+
+	// Check if we need to enable single-stepping. Tracers expect that the
+	// kernel preserves the value of the single-step flag set by PTRACE_SETREGS
+	// whether or not PTRACE_SINGLESTEP/PTRACE_SYSEMU_SINGLESTEP is used (this
+	// includes our ptrace platform, by the way), so we should only clear the
+	// single-step flag if we're responsible for setting it. (clearSinglestep
+	// is therefore analogous to Linux's TIF_FORCED_TF.)
+	//
+	// Strictly speaking, we should also not clear the single-step flag if we
+	// single-step through an instruction that sets the single-step flag
+	// (arch/x86/kernel/step.c:is_setting_trap_flag()). But nobody sets their
+	// own TF. (Famous last words, I know.)
+	clearSinglestep := false
+	if t.hasTracer() {
+		t.tg.pidns.owner.mu.RLock()
+		if t.ptraceSinglestep {
+			clearSinglestep = !t.Arch().SingleStep()
+			t.Arch().SetSingleStep()
+		}
+		t.tg.pidns.owner.mu.RUnlock()
+	}
+
+	t.accountTaskGoroutineEnter(TaskGoroutineRunningApp)
+	info, at, err := t.p.Switch(t.MemoryManager().AddressSpace(), t.Arch(), t.rseqCPU)
+	t.accountTaskGoroutineLeave(TaskGoroutineRunningApp)
+
+	if clearSinglestep {
+		t.Arch().ClearSingleStep()
+	}
+
+	switch err {
+	case nil:
+		// Handle application system call.
+		return t.doSyscall()
+
+	case platform.ErrContextInterrupt:
+		// Interrupted by platform.Context.Interrupt(). Re-enter the run
+		// loop to figure out why.
+		return (*runApp)(nil)
+
+	case platform.ErrContextSignal:
+		// Looks like a signal has been delivered to us. If it's a synchronous
+		// signal (SEGV, SIGBUS, etc.), it should be sent to the application
+		// thread that received it.
+		sig := linux.Signal(info.Signo)
+
+		// Was it a fault that we should handle internally? If so, this wasn't
+		// an application-generated signal and we should continue execution
+		// normally.
+		if at.Any() {
+			addr := usermem.Addr(info.Addr())
+			err := t.MemoryManager().HandleUserFault(t, addr, at, usermem.Addr(t.Arch().Stack()))
+			if err == nil {
+				// The fault was handled appropriately.
+				// We can resume running the application.
+				return (*runApp)(nil)
+			}
+
+			// Is this a vsyscall that we need emulate?
+			if at.Execute {
+				if sysno, ok := t.tc.st.LookupEmulate(addr); ok {
+					return t.doVsyscall(addr, sysno)
+				}
+			}
+
+			// The JVM will trigger these errors constantly, so don't
+			// spam logs with this error.
+			if err == syserror.EFAULT || err == syserror.EPERM {
+				t.Debugf("Unhandled user fault: addr=%x ip=%x access=%v err=%v", addr, t.Arch().IP(), at, err)
+			} else {
+				t.Warningf("Unhandled user fault: addr=%x ip=%x access=%v err=%v", addr, t.Arch().IP(), at, err)
+			}
+			t.DebugDumpState()
+
+			// Continue to signal handling.
+			//
+			// Convert a BusError error to a SIGBUS from a SIGSEGV. All
+			// other info bits stay the same (address, etc.).
+			if _, ok := err.(*memmap.BusError); ok {
+				sig = linux.SIGBUS
+				info.Signo = int32(linux.SIGBUS)
+			}
+		}
+
+		switch sig {
+		case linux.SIGILL:
+			// N.B. The debug stuff here is arguably
+			// expensive.  Don't fret. This gets called
+			// about 5 times for a typical application, if
+			// that.
+			t.Debugf("SIGILL @ %x", t.Arch().IP())
+
+			// Is this a CPUID instruction?
+			expected := arch.CPUIDInstruction[:]
+			found := make([]byte, len(expected))
+			_, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found)
+			if err == nil && bytes.Equal(expected, found) {
+				// Skip the cpuid instruction.
+				t.Arch().CPUIDEmulate(t)
+				t.Arch().SetIP(t.Arch().IP() + uintptr(len(expected)))
+				break
+			}
+
+			// Treat it like any other synchronous signal.
+			fallthrough
+
+		case linux.SIGSEGV, linux.SIGBUS, linux.SIGFPE, linux.SIGTRAP:
+			// Synchronous signal. Send it to ourselves. Assume the signal is
+			// legitimate and force it (work around the signal being ignored or
+			// blocked) like Linux does. Conveniently, this is even the correct
+			// behavior for SIGTRAP from single-stepping.
+			t.forceSignal(linux.Signal(sig), false /* unconditional */)
+			t.SendSignal(info)
+
+		case platform.SignalInterrupt:
+			// Assume that a call to platform.Context.Interrupt() misfired.
+
+		case linux.SIGPROF:
+			// It's a profiling interrupt: there's not much
+			// we can do. We've already paid a decent cost
+			// by intercepting the signal, at this point we
+			// simply ignore it.
+
+		default:
+			// Asynchronous signal. Let the system deal with it.
+			t.k.sendExternalSignal(info, "application")
+		}
+
+		return (*runApp)(nil)
+
+	case platform.ErrContextCPUPreempted:
+		// Ensure that RSEQ critical sections are interrupted and per-thread
+		// CPU values are updated before the next platform.Context.Switch().
+		t.rseqPreempted = true
+		return (*runApp)(nil)
+
+	default:
+		// What happened? Can't continue.
+		t.Warningf("Unexpected SwitchToApp error: %v", err)
+		t.PrepareExit(ExitStatus{Code: t.ExtractErrno(err, -1)})
+		return (*runExit)(nil)
+	}
+}
+
+// waitGoroutineStoppedOrExited blocks until t's task goroutine stops or exits.
+func (t *Task) waitGoroutineStoppedOrExited() {
+	t.goroutineStopped.Wait()
+}
+
+// WaitExited blocks until all task goroutines in tg have exited.
+//
+// WaitExited does not correspond to anything in Linux; it's provided so that
+// external callers of Kernel.CreateProcess can wait for the created thread
+// group to terminate.
+func (tg *ThreadGroup) WaitExited() {
+	tg.liveGoroutines.Wait()
+}
+
+// Yield yields the processor for the calling task.
+func (t *Task) Yield() {
+	atomic.AddUint64(&t.yieldCount, 1)
+	runtime.Gosched()
+}
diff --git a/pkg/sentry/kernel/task_sched.go b/pkg/sentry/kernel/task_sched.go
new file mode 100644
index 000000000..b50139077
--- /dev/null
+++ b/pkg/sentry/kernel/task_sched.go
@@ -0,0 +1,329 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// CPU scheduling, real and fake.
+
+import (
+	"fmt"
+	"sync/atomic"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/hostcpu"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// TaskGoroutineState is a coarse representation of the current execution
+// status of a kernel.Task goroutine.
+type TaskGoroutineState int
+
+const (
+	// TaskGoroutineNonexistent indicates that the task goroutine has either
+	// not yet been created by Task.Start() or has returned from Task.run().
+	// This must be the zero value for TaskGoroutineState.
+	TaskGoroutineNonexistent TaskGoroutineState = iota
+
+	// TaskGoroutineRunningSys indicates that the task goroutine is executing
+	// sentry code.
+	TaskGoroutineRunningSys
+
+	// TaskGoroutineRunningApp indicates that the task goroutine is executing
+	// application code.
+	TaskGoroutineRunningApp
+
+	// TaskGoroutineBlockedInterruptible indicates that the task goroutine is
+	// blocked in Task.block(), and hence may be woken by Task.interrupt()
+	// (e.g. due to signal delivery).
+	TaskGoroutineBlockedInterruptible
+
+	// TaskGoroutineBlockedUninterruptible indicates that the task goroutine is
+	// stopped outside of Task.block() and Task.doStop(), and hence cannot be
+	// woken by Task.interrupt().
+	TaskGoroutineBlockedUninterruptible
+
+	// TaskGoroutineStopped indicates that the task goroutine is blocked in
+	// Task.doStop(). TaskGoroutineStopped is similar to
+	// TaskGoroutineBlockedUninterruptible, but is a separate state to make it
+	// possible to determine when Task.stop is meaningful.
+	TaskGoroutineStopped
+)
+
+// TaskGoroutineSchedInfo contains task goroutine scheduling state which must
+// be read and updated atomically.
+type TaskGoroutineSchedInfo struct {
+	// Timestamp was the value of Kernel.cpuClock when this
+	// TaskGoroutineSchedInfo was last updated.
+	Timestamp uint64
+
+	// State is the current state of the task goroutine.
+	State TaskGoroutineState
+
+	// UserTicks is the amount of time the task goroutine has spent executing
+	// its associated Task's application code, in units of linux.ClockTick.
+	UserTicks uint64
+
+	// SysTicks is the amount of time the task goroutine has spent executing in
+	// the sentry, in units of linux.ClockTick.
+	SysTicks uint64
+}
+
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) accountTaskGoroutineEnter(state TaskGoroutineState) {
+	now := t.k.CPUClockNow()
+	if t.gosched.State != TaskGoroutineRunningSys {
+		panic(fmt.Sprintf("Task goroutine switching from state %v (expected %v) to %v", t.gosched.State, TaskGoroutineRunningSys, state))
+	}
+	t.goschedSeq.BeginWrite()
+	// This function is very hot; avoid defer.
+	t.gosched.SysTicks += now - t.gosched.Timestamp
+	t.gosched.Timestamp = now
+	t.gosched.State = state
+	t.goschedSeq.EndWrite()
+}
+
+// Preconditions: The caller must be running on the task goroutine, and leaving
+// a state indicated by a previous call to
+// t.accountTaskGoroutineEnter(state).
+func (t *Task) accountTaskGoroutineLeave(state TaskGoroutineState) {
+	now := t.k.CPUClockNow()
+	if t.gosched.State != state {
+		panic(fmt.Sprintf("Task goroutine switching from state %v (expected %v) to %v", t.gosched.State, state, TaskGoroutineRunningSys))
+	}
+	t.goschedSeq.BeginWrite()
+	// This function is very hot; avoid defer.
+	if state == TaskGoroutineRunningApp {
+		t.gosched.UserTicks += now - t.gosched.Timestamp
+	}
+	t.gosched.Timestamp = now
+	t.gosched.State = TaskGoroutineRunningSys
+	t.goschedSeq.EndWrite()
+}
+
+// TaskGoroutineSchedInfo returns a copy of t's task goroutine scheduling info.
+// Most clients should use t.CPUStats() instead.
+func (t *Task) TaskGoroutineSchedInfo() TaskGoroutineSchedInfo {
+	return SeqAtomicLoadTaskGoroutineSchedInfo(&t.goschedSeq, &t.gosched)
+}
+
+// CPUStats returns the CPU usage statistics of t.
+func (t *Task) CPUStats() usage.CPUStats {
+	return t.cpuStatsAt(t.k.CPUClockNow())
+}
+
+// Preconditions: now <= Kernel.CPUClockNow(). (Since Kernel.cpuClock is
+// monotonic, this is satisfied if now is the result of a previous call to
+// Kernel.CPUClockNow().) This requirement exists because otherwise a racing
+// change to t.gosched can cause cpuStatsAt to adjust stats by too much, making
+// the returned stats non-monotonic.
+func (t *Task) cpuStatsAt(now uint64) usage.CPUStats {
+	tsched := t.TaskGoroutineSchedInfo()
+	if tsched.Timestamp < now {
+		// Update stats to reflect execution since the last update to
+		// t.gosched.
+		switch tsched.State {
+		case TaskGoroutineRunningSys:
+			tsched.SysTicks += now - tsched.Timestamp
+		case TaskGoroutineRunningApp:
+			tsched.UserTicks += now - tsched.Timestamp
+		}
+	}
+	return usage.CPUStats{
+		UserTime:          time.Duration(tsched.UserTicks * uint64(linux.ClockTick)),
+		SysTime:           time.Duration(tsched.SysTicks * uint64(linux.ClockTick)),
+		VoluntarySwitches: atomic.LoadUint64(&t.yieldCount),
+	}
+}
+
+// CPUStats returns the combined CPU usage statistics of all past and present
+// threads in tg.
+func (tg *ThreadGroup) CPUStats() usage.CPUStats {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	// Hack to get a pointer to the Kernel.
+	if tg.leader == nil {
+		// Per comment on tg.leader, this is only possible if nothing in the
+		// ThreadGroup has ever executed anyway.
+		return usage.CPUStats{}
+	}
+	now := tg.leader.k.CPUClockNow()
+	stats := tg.exitedCPUStats
+	// Account for active tasks.
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		stats.Accumulate(t.cpuStatsAt(now))
+	}
+	return stats
+}
+
+// JoinedChildCPUStats implements the semantics of RUSAGE_CHILDREN: "Return
+// resource usage statistics for all children of [tg] that have terminated and
+// been waited for. These statistics will include the resources used by
+// grandchildren, and further removed descendants, if all of the intervening
+// descendants waited on their terminated children."
+func (tg *ThreadGroup) JoinedChildCPUStats() usage.CPUStats {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.childCPUStats
+}
+
+// StateStatus returns a string representation of the task's current state,
+// appropriate for /proc/[pid]/status.
+func (t *Task) StateStatus() string {
+	switch s := t.TaskGoroutineSchedInfo().State; s {
+	case TaskGoroutineNonexistent:
+		t.tg.pidns.owner.mu.RLock()
+		defer t.tg.pidns.owner.mu.RUnlock()
+		switch t.exitState {
+		case TaskExitZombie:
+			return "Z (zombie)"
+		case TaskExitDead:
+			return "X (dead)"
+		default:
+			// The task goroutine can't exit before passing through
+			// runExitNotify, so this indicates that the task has been created,
+			// but the task goroutine hasn't yet started. The Linux equivalent
+			// is struct task_struct::state == TASK_NEW
+			// (kernel/fork.c:copy_process() =>
+			// kernel/sched/core.c:sched_fork()), but the TASK_NEW bit is
+			// masked out by TASK_REPORT for /proc/[pid]/status, leaving only
+			// TASK_RUNNING.
+			return "R (running)"
+		}
+	case TaskGoroutineRunningSys, TaskGoroutineRunningApp:
+		return "R (running)"
+	case TaskGoroutineBlockedInterruptible:
+		return "S (sleeping)"
+	case TaskGoroutineStopped:
+		t.tg.signalHandlers.mu.Lock()
+		defer t.tg.signalHandlers.mu.Unlock()
+		switch t.stop.(type) {
+		case *groupStop:
+			return "T (stopped)"
+		case *ptraceStop:
+			return "t (tracing stop)"
+		}
+		fallthrough
+	case TaskGoroutineBlockedUninterruptible:
+		// This is the name Linux uses for TASK_UNINTERRUPTIBLE and
+		// TASK_KILLABLE (= TASK_UNINTERRUPTIBLE | TASK_WAKEKILL):
+		// fs/proc/array.c:task_state_array.
+		return "D (disk sleep)"
+	default:
+		panic(fmt.Sprintf("Invalid TaskGoroutineState: %v", s))
+	}
+}
+
+// CPUMask returns a copy of t's allowed CPU mask.
+func (t *Task) CPUMask() sched.CPUSet {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.allowedCPUMask.Copy()
+}
+
+// SetCPUMask sets t's allowed CPU mask based on mask. It takes ownership of
+// mask.
+//
+// Preconditions: mask.Size() ==
+// sched.CPUSetSize(t.Kernel().ApplicationCores()).
+func (t *Task) SetCPUMask(mask sched.CPUSet) error {
+	if want := sched.CPUSetSize(t.k.applicationCores); mask.Size() != want {
+		panic(fmt.Sprintf("Invalid CPUSet %v (expected %d bytes)", mask, want))
+	}
+
+	// Remove CPUs in mask above Kernel.applicationCores.
+	mask.ClearAbove(t.k.applicationCores)
+
+	// Ensure that at least 1 CPU is still allowed.
+	if mask.NumCPUs() == 0 {
+		return syserror.EINVAL
+	}
+
+	if t.k.useHostCores {
+		// No-op; pretend the mask was immediately changed back.
+		return nil
+	}
+
+	t.tg.pidns.owner.mu.RLock()
+	rootTID := t.tg.pidns.owner.Root.tids[t]
+	t.tg.pidns.owner.mu.RUnlock()
+
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.allowedCPUMask = mask
+	atomic.StoreInt32(&t.cpu, assignCPU(mask, rootTID))
+	return nil
+}
+
+// CPU returns the cpu id for a given task.
+func (t *Task) CPU() int32 {
+	if t.k.useHostCores {
+		return int32(hostcpu.GetCPU())
+	}
+
+	return atomic.LoadInt32(&t.cpu)
+}
+
+// assignCPU returns the virtualized CPU number for the task with global TID
+// tid and allowedCPUMask allowed.
+func assignCPU(allowed sched.CPUSet, tid ThreadID) (cpu int32) {
+	// To pretend that threads are evenly distributed to allowed CPUs, choose n
+	// to be less than the number of CPUs in allowed ...
+	n := int(tid) % int(allowed.NumCPUs())
+	// ... then pick the nth CPU in allowed.
+	allowed.ForEachCPU(func(c uint) {
+		if n--; n == 0 {
+			cpu = int32(c)
+		}
+	})
+	return cpu
+}
+
+// Niceness returns t's niceness.
+func (t *Task) Niceness() int {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.niceness
+}
+
+// Priority returns t's priority.
+func (t *Task) Priority() int {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.niceness + 20
+}
+
+// SetNiceness sets t's niceness to n.
+func (t *Task) SetNiceness(n int) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.niceness = n
+}
+
+// NumaPolicy returns t's current numa policy.
+func (t *Task) NumaPolicy() (policy int32, nodeMask uint32) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.numaPolicy, t.numaNodeMask
+}
+
+// SetNumaPolicy sets t's numa policy.
+func (t *Task) SetNumaPolicy(policy int32, nodeMask uint32) {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.numaPolicy = policy
+	t.numaNodeMask = nodeMask
+}
diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go
new file mode 100644
index 000000000..2340256b0
--- /dev/null
+++ b/pkg/sentry/kernel/task_signals.go
@@ -0,0 +1,1056 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// This file defines the behavior of task signal handling.
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SignalAction is an internal signal action.
+type SignalAction int
+
+// Available signal actions.
+// Note that although we refer the complete set internally,
+// the application is only capable of using the Default and
+// Ignore actions from the system call interface.
+const (
+	SignalActionTerm SignalAction = iota
+	SignalActionCore
+	SignalActionStop
+	SignalActionIgnore
+	SignalActionHandler
+)
+
+// Default signal handler actions. Note that for most signals,
+// (except SIGKILL and SIGSTOP) these can be overridden by the app.
+var defaultActions = map[linux.Signal]SignalAction{
+	// POSIX.1-1990 standard.
+	linux.SIGHUP:  SignalActionTerm,
+	linux.SIGINT:  SignalActionTerm,
+	linux.SIGQUIT: SignalActionCore,
+	linux.SIGILL:  SignalActionCore,
+	linux.SIGABRT: SignalActionCore,
+	linux.SIGFPE:  SignalActionCore,
+	linux.SIGKILL: SignalActionTerm, // but see ThreadGroup.applySignalSideEffects
+	linux.SIGSEGV: SignalActionCore,
+	linux.SIGPIPE: SignalActionTerm,
+	linux.SIGALRM: SignalActionTerm,
+	linux.SIGTERM: SignalActionTerm,
+	linux.SIGUSR1: SignalActionTerm,
+	linux.SIGUSR2: SignalActionTerm,
+	linux.SIGCHLD: SignalActionIgnore,
+	linux.SIGCONT: SignalActionIgnore, // but see ThreadGroup.applySignalSideEffects
+	linux.SIGSTOP: SignalActionStop,
+	linux.SIGTSTP: SignalActionStop,
+	linux.SIGTTIN: SignalActionStop,
+	linux.SIGTTOU: SignalActionStop,
+	// POSIX.1-2001 standard.
+	linux.SIGBUS:    SignalActionCore,
+	linux.SIGPROF:   SignalActionTerm,
+	linux.SIGSYS:    SignalActionCore,
+	linux.SIGTRAP:   SignalActionCore,
+	linux.SIGURG:    SignalActionIgnore,
+	linux.SIGVTALRM: SignalActionTerm,
+	linux.SIGXCPU:   SignalActionCore,
+	linux.SIGXFSZ:   SignalActionCore,
+	// The rest on linux.
+	linux.SIGSTKFLT: SignalActionTerm,
+	linux.SIGIO:     SignalActionTerm,
+	linux.SIGPWR:    SignalActionTerm,
+	linux.SIGWINCH:  SignalActionIgnore,
+}
+
+// computeAction figures out what to do given a signal number
+// and an arch.SignalAct. SIGSTOP always results in a SignalActionStop,
+// and SIGKILL always results in a SignalActionTerm.
+// Signal 0 is always ignored as many programs use it for various internal functions
+// and don't expect it to do anything.
+//
+// In the event the signal is not one of these, act.Handler determines what
+// happens next.
+// If act.Handler is:
+// 0, the default action is taken;
+// 1, the signal is ignored;
+// anything else, the function returns SignalActionHandler.
+func computeAction(sig linux.Signal, act arch.SignalAct) SignalAction {
+	switch sig {
+	case linux.SIGSTOP:
+		return SignalActionStop
+	case linux.SIGKILL:
+		return SignalActionTerm
+	case linux.Signal(0):
+		return SignalActionIgnore
+	}
+
+	switch act.Handler {
+	case arch.SignalActDefault:
+		return defaultActions[sig]
+	case arch.SignalActIgnore:
+		return SignalActionIgnore
+	default:
+		return SignalActionHandler
+	}
+}
+
+// UnblockableSignals contains the set of signals which cannot be blocked.
+var UnblockableSignals = linux.MakeSignalSet(linux.SIGKILL, linux.SIGSTOP)
+
+// StopSignals is the set of signals whose default action is SignalActionStop.
+var StopSignals = linux.MakeSignalSet(linux.SIGSTOP, linux.SIGTSTP, linux.SIGTTIN, linux.SIGTTOU)
+
+// dequeueSignalLocked returns a pending unmasked signal. If there are no
+// pending unmasked signals, dequeueSignalLocked returns nil.
+//
+// Preconditions: t.tg.signalHandlers.mu must be locked.
+func (t *Task) dequeueSignalLocked() *arch.SignalInfo {
+	if info := t.pendingSignals.dequeue(t.tr.SignalMask); info != nil {
+		return info
+	}
+	if info := t.tg.pendingSignals.dequeue(t.tr.SignalMask); info != nil {
+		return info
+	}
+	return nil
+}
+
+// TakeSignal returns a pending signal not blocked by mask. Signal handlers are
+// not affected. If there are no pending signals not blocked by mask,
+// TakeSignal returns a nil SignalInfo.
+func (t *Task) TakeSignal(mask linux.SignalSet) *arch.SignalInfo {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if info := t.pendingSignals.dequeue(mask); info != nil {
+		return info
+	}
+	if info := t.tg.pendingSignals.dequeue(mask); info != nil {
+		return info
+	}
+	return nil
+}
+
+// discardSpecificLocked removes all instances of the given signal from all
+// signal queues in tg.
+//
+// Preconditions: The signal mutex must be locked.
+func (tg *ThreadGroup) discardSpecificLocked(sig linux.Signal) {
+	tg.pendingSignals.discardSpecific(sig)
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		t.pendingSignals.discardSpecific(sig)
+	}
+}
+
+// PendingSignals returns the set of pending signals.
+func (t *Task) PendingSignals() linux.SignalSet {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	return t.pendingSignals.pendingSet | t.tg.pendingSignals.pendingSet
+}
+
+// deliverSignal delivers the given signal and returns the following run state.
+func (t *Task) deliverSignal(info *arch.SignalInfo, act arch.SignalAct) taskRunState {
+	sigact := computeAction(linux.Signal(info.Signo), act)
+
+	if t.haveSyscallReturn {
+		if sre, ok := SyscallRestartErrnoFromReturn(t.Arch().Return()); ok {
+			// Signals that are ignored, cause a thread group stop, or
+			// terminate the thread group do not interact with interrupted
+			// syscalls; in Linux terms, they are never returned to the signal
+			// handling path from get_signal => get_signal_to_deliver. The
+			// behavior of an interrupted syscall is determined by the first
+			// signal that is actually handled (by userspace).
+			if sigact == SignalActionHandler {
+				switch {
+				case sre == ERESTARTNOHAND:
+					fallthrough
+				case sre == ERESTART_RESTARTBLOCK:
+					fallthrough
+				case (sre == ERESTARTSYS && !act.IsRestart()):
+					t.Debugf("Not restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo)
+					t.Arch().SetReturn(uintptr(-t.ExtractErrno(syserror.EINTR, -1)))
+				default:
+					t.Debugf("Restarting syscall %d after errno %d: interrupted by signal %d", t.Arch().SyscallNo(), sre, info.Signo)
+					t.Arch().RestartSyscall()
+				}
+			}
+		}
+	}
+
+	switch sigact {
+	case SignalActionTerm, SignalActionCore:
+		// "Default action is to terminate the process." - signal(7)
+		t.Debugf("Signal %d: terminating thread group", info.Signo)
+		t.PrepareGroupExit(ExitStatus{Signo: int(info.Signo)})
+		return (*runExit)(nil)
+
+	case SignalActionStop:
+		// "Default action is to stop the process."
+		t.initiateGroupStop(info)
+
+	case SignalActionIgnore:
+		// "Default action is to ignore the signal."
+		t.Debugf("Signal %d: ignored", info.Signo)
+
+	case SignalActionHandler:
+		// Try to deliver the signal to the user-configured handler.
+		t.Debugf("Signal %d: delivering to handler", info.Signo)
+		if err := t.deliverSignalToHandler(info, act); err != nil {
+			t.Warningf("Failed to deliver signal %+v to user handler: %v", info, err)
+			// Send a forced SIGSEGV. If the signal that couldn't be delivered
+			// was a SIGSEGV, force the handler to SIG_DFL.
+			t.forceSignal(linux.SIGSEGV, linux.Signal(info.Signo) == linux.SIGSEGV /* unconditional */)
+			t.SendSignal(sigPriv(linux.SIGSEGV))
+		}
+
+	default:
+		panic(fmt.Sprintf("Unknown signal action %+v, %d?", info, computeAction(linux.Signal(info.Signo), act)))
+	}
+	return (*runInterrupt)(nil)
+}
+
+// deliverSignalToHandler changes the task's userspace state to enter the given
+// user-configured handler for the given signal.
+func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) error {
+	// Signal delivery to an application handler interrupts restartable
+	// sequences.
+	t.rseqInterrupt()
+
+	// Are executing on the main stack,
+	// or the provided alternate stack?
+	sp := usermem.Addr(t.Arch().Stack())
+
+	// N.B. This is a *copy* of the alternate stack that the user's signal
+	// handler expects to see in its ucontext (even if it's not in use).
+	alt := t.signalStack
+	if act.IsOnStack() && alt.IsEnabled() {
+		alt.SetOnStack()
+		if !t.OnSignalStack(alt) {
+			sp = usermem.Addr(alt.Top())
+		}
+	}
+
+	// Set up the signal handler. If we have a saved signal mask, the signal
+	// handler should run with the current mask, but sigreturn should restore
+	// the saved one.
+	st := &arch.Stack{t.Arch(), t.MemoryManager(), sp}
+	mask := t.tr.SignalMask
+	if t.haveSavedSignalMask {
+		mask = t.savedSignalMask
+	}
+	if err := t.Arch().SignalSetup(st, &act, info, &alt, mask); err != nil {
+		return err
+	}
+	t.haveSavedSignalMask = false
+
+	// Add our signal mask.
+	newMask := t.tr.SignalMask | act.Mask
+	if !act.IsNoDefer() {
+		newMask |= linux.SignalSetOf(linux.Signal(info.Signo))
+	}
+	t.SetSignalMask(newMask)
+
+	return nil
+}
+
+var ctrlResume = &SyscallControl{ignoreReturn: true}
+
+// SignalReturn implements sigreturn(2) (if rt is false) or rt_sigreturn(2) (if
+// rt is true).
+func (t *Task) SignalReturn(rt bool) (*SyscallControl, error) {
+	st := t.Stack()
+	sigset, err := t.Arch().SignalRestore(st, rt)
+	if err != nil {
+		return nil, err
+	}
+
+	// Restore our signal mask. SIGKILL and SIGSTOP should not be blocked.
+	t.SetSignalMask(sigset &^ UnblockableSignals)
+
+	// TODO: sys_rt_sigreturn also calls restore_altstack from
+	// uc.stack, allowing the signal handler to implicitly mutate the signal
+	// stack.
+
+	return ctrlResume, nil
+}
+
+// SendSignal sends the given signal to t.
+//
+// The following errors may be returned:
+//
+//	syserror.ESRCH - The task has exited.
+//	syserror.EINVAL - The signal is not valid.
+//	syserror.EAGAIN - THe signal is realtime, and cannot be queued.
+//
+func (t *Task) SendSignal(info *arch.SignalInfo) error {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	return t.sendSignalLocked(info, false /* group */)
+}
+
+// SendGroupSignal sends the given signal to t's thread group.
+func (t *Task) SendGroupSignal(info *arch.SignalInfo) error {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	return t.sendSignalLocked(info, true /* group */)
+}
+
+// SendSignal sends the given signal to tg, using tg's leader to determine if
+// the signal is blocked.
+func (tg *ThreadGroup) SendSignal(info *arch.SignalInfo) error {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	tg.signalHandlers.mu.Lock()
+	defer tg.signalHandlers.mu.Unlock()
+	return tg.leader.sendSignalLocked(info, true /* group */)
+}
+
+// Preconditions: The TaskSet mutex must be locked.
+func (t *Task) onCPULocked(includeSys bool) bool {
+	// Task is exiting.
+	if t.exitState != TaskExitNone {
+		return false
+	}
+
+	switch t.TaskGoroutineSchedInfo().State {
+	case TaskGoroutineRunningSys:
+		return includeSys
+	case TaskGoroutineRunningApp:
+		return true
+	default:
+		return false
+	}
+}
+
+// SendTimerSignal mimics the process timer signal delivery behavior in linux:
+// signals are delivered to the thread that triggers the timer expiration (see
+// kernel/time/posix-cpu-timers.c:check_process_timers(). This
+// means
+//   1) the thread is running on cpu at the time.
+//   2) a thread runs more frequently will get more of those signals.
+//
+// We approximate this behavior by selecting a running task in a round-robin
+// fashion. Statistically, a thread running more often should have a higher
+// probability to be selected.
+func (tg *ThreadGroup) SendTimerSignal(info *arch.SignalInfo, includeSys bool) error {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	tg.signalHandlers.mu.Lock()
+	defer tg.signalHandlers.mu.Unlock()
+
+	// Find the next running threads.
+	var t *Task
+	if tg.lastTimerSignalTask == nil {
+		t = tg.tasks.Front()
+	} else {
+		t = tg.lastTimerSignalTask.Next()
+	}
+
+	// Iterate from lastTimerSignalTask.Next() to the last task in the task list.
+	for t != nil {
+		if t.onCPULocked(includeSys) {
+			tg.lastTimerSignalTask = t
+			return t.sendSignalLocked(info, true /* group */)
+		}
+		t = t.Next()
+	}
+
+	// t is nil when we reach here. If lastTimerSignalTask is not nil, iterate
+	// from Front to lastTimerSignalTask.
+	if tg.lastTimerSignalTask != nil {
+		for t := tg.tasks.Front(); t != tg.lastTimerSignalTask.Next(); t = t.Next() {
+			if t.onCPULocked(includeSys) {
+				tg.lastTimerSignalTask = t
+				return t.sendSignalLocked(info, true /* group */)
+			}
+		}
+	}
+
+	// No running threads? Just try the leader.
+	tg.lastTimerSignalTask = tg.leader
+	return tg.leader.sendSignalLocked(info, true /* group */)
+}
+
+func (t *Task) sendSignalLocked(info *arch.SignalInfo, group bool) error {
+	if t.exitState == TaskExitDead {
+		return syserror.ESRCH
+	}
+	sig := linux.Signal(info.Signo)
+	if sig == 0 {
+		return nil
+	}
+	if !sig.IsValid() {
+		return syserror.EINVAL
+	}
+
+	// Signal side effects apply even if the signal is ultimately discarded.
+	t.tg.applySignalSideEffectsLocked(sig)
+
+	// TODO: "Only signals for which the "init" process has established a
+	// signal handler can be sent to the "init" process by other members of the
+	// PID namespace. This restriction applies even to privileged processes,
+	// and prevents other members of the PID namespace from accidentally
+	// killing the "init" process." - pid_namespaces(7). We don't currently do
+	// this for child namespaces, though we should; we also don't do this for
+	// the root namespace (the same restriction applies to global init on
+	// Linux), where whether or not we should is much murkier. In practice,
+	// most sandboxed applications are not prepared to function as an init
+	// process.
+
+	// Unmasked, ignored signals are discarded without being queued, unless
+	// they will be visible to a tracer. Even for group signals, it's the
+	// originally-targeted task's signal mask and tracer that matter; compare
+	// Linux's kernel/signal.c:__send_signal() => prepare_signal() =>
+	// sig_ignored().
+	ignored := computeAction(sig, t.tg.signalHandlers.actions[sig]) == SignalActionIgnore
+	if linux.SignalSetOf(sig)&t.tr.SignalMask == 0 && ignored && !t.hasTracer() {
+		t.Debugf("Discarding ignored signal %d", sig)
+		return nil
+	}
+
+	q := &t.pendingSignals
+	if group {
+		q = &t.tg.pendingSignals
+	}
+	if !q.enqueue(info) {
+		if sig.IsRealtime() {
+			return syserror.EAGAIN
+		}
+		t.Debugf("Discarding duplicate signal %d", sig)
+		return nil
+	}
+
+	// Find a receiver to notify. Note that the task we choose to notify, if
+	// any, may not be the task that actually dequeues and handles the signal;
+	// e.g. a racing signal mask change may cause the notified task to become
+	// ineligible, or a racing sibling task may dequeue the signal first.
+	if t.canReceiveSignalLocked(sig) {
+		t.Debugf("Notified of signal %d", sig)
+		t.interrupt()
+		return nil
+	}
+	if group {
+		if nt := t.tg.findSignalReceiverLocked(sig); nt != nil {
+			nt.Debugf("Notified of group signal %d", sig)
+			nt.interrupt()
+			return nil
+		}
+	}
+	t.Debugf("No task notified of signal %d", sig)
+	return nil
+}
+
+func (tg *ThreadGroup) applySignalSideEffectsLocked(sig linux.Signal) {
+	switch {
+	case linux.SignalSetOf(sig)&StopSignals != 0:
+		// Stop signals cause all prior SIGCONT to be discarded. (This is
+		// despite the fact this has little effect since SIGCONT's most
+		// important effect is applied when the signal is sent in the branch
+		// below, not when the signal is delivered.)
+		tg.discardSpecificLocked(linux.SIGCONT)
+	case sig == linux.SIGCONT:
+		// "The SIGCONT signal has a side effect of waking up (all threads of)
+		// a group-stopped process. This side effect happens before
+		// signal-delivery-stop. The tracer can't suppress this side effect (it
+		// can only suppress signal injection, which only causes the SIGCONT
+		// handler to not be executed in the tracee, if such a handler is
+		// installed." - ptrace(2)
+		tg.endGroupStopLocked(true)
+	case sig == linux.SIGKILL:
+		// "SIGKILL does not generate signal-delivery-stop and therefore the
+		// tracer can't suppress it. SIGKILL kills even within system calls
+		// (syscall-exit-stop is not generated prior to death by SIGKILL)." -
+		// ptrace(2)
+		//
+		// Note that this differs from ThreadGroup.requestExit in that it
+		// ignores tg.execing.
+		if !tg.exiting {
+			tg.exiting = true
+			tg.exitStatus = ExitStatus{Signo: int(linux.SIGKILL)}
+		}
+		for t := tg.tasks.Front(); t != nil; t = t.Next() {
+			t.killLocked()
+		}
+	}
+}
+
+// canReceiveSignalLocked returns true if t should be interrupted to receive
+// the given signal. canReceiveSignalLocked is analogous to Linux's
+// kernel/signal.c:wants_signal(), but see below for divergences.
+//
+// Preconditions: The signal mutex must be locked.
+func (t *Task) canReceiveSignalLocked(sig linux.Signal) bool {
+	// - Do not choose tasks that are blocking the signal.
+	if linux.SignalSetOf(sig)&t.tr.SignalMask != 0 {
+		return false
+	}
+	// - No need to check Task.exitState, as the exit path sets every bit in the
+	// signal mask when it transitions from TaskExitNone to TaskExitInitiated.
+	// - No special case for SIGKILL: SIGKILL already interrupted all tasks in the
+	// task group via applySignalSideEffects => killLocked.
+	// - Do not choose stopped tasks, which cannot handle signals.
+	if t.stop != nil {
+		return false
+	}
+	// - TODO: No special case for when t is also the sending task,
+	// because the identity of the sender is unknown.
+	// - Do not choose tasks that have already been interrupted, as they may be
+	// busy handling another signal.
+	if len(t.interruptChan) != 0 {
+		return false
+	}
+	return true
+}
+
+// findSignalReceiverLocked returns a task in tg that should be interrupted to
+// receive the given signal. If no such task exists, findSignalReceiverLocked
+// returns nil.
+//
+// Linux actually records curr_target to balance the group signal targets.
+//
+// Preconditions: The signal mutex must be locked.
+func (tg *ThreadGroup) findSignalReceiverLocked(sig linux.Signal) *Task {
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		if t.canReceiveSignalLocked(sig) {
+			return t
+		}
+	}
+	return nil
+}
+
+// forceSignal ensures that the task is not ignoring or blocking the given
+// signal. If unconditional is true, forceSignal takes action even if the
+// signal isn't being ignored or blocked.
+func (t *Task) forceSignal(sig linux.Signal, unconditional bool) {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.forceSignalLocked(sig, unconditional)
+}
+
+func (t *Task) forceSignalLocked(sig linux.Signal, unconditional bool) {
+	blocked := linux.SignalSetOf(sig)&t.tr.SignalMask != 0
+	act := t.tg.signalHandlers.actions[sig]
+	ignored := act.Handler == arch.SignalActIgnore
+	if blocked || ignored || unconditional {
+		act.Handler = arch.SignalActDefault
+		t.tg.signalHandlers.actions[sig] = act
+		if blocked {
+			t.setSignalMaskLocked(t.tr.SignalMask &^ linux.SignalSetOf(sig))
+		}
+	}
+}
+
+// SignalMask returns a copy of t's signal mask.
+func (t *Task) SignalMask() linux.SignalSet {
+	return linux.SignalSet(atomic.LoadUint64((*uint64)(&t.tr.SignalMask)))
+}
+
+// SetSignalMask sets t's signal mask.
+//
+// Preconditions: SetSignalMask can only be called by the task goroutine.
+// t.exitState < TaskExitZombie.
+func (t *Task) SetSignalMask(mask linux.SignalSet) {
+	// By precondition, t prevents t.tg from completing an execve and mutating
+	// t.tg.signalHandlers, so we can skip the TaskSet mutex.
+	t.tg.signalHandlers.mu.Lock()
+	t.setSignalMaskLocked(mask)
+	t.tg.signalHandlers.mu.Unlock()
+}
+
+// Preconditions: The signal mutex must be locked.
+func (t *Task) setSignalMaskLocked(mask linux.SignalSet) {
+	oldMask := t.tr.SignalMask
+	atomic.StoreUint64((*uint64)(&t.tr.SignalMask), uint64(mask))
+
+	// If the new mask blocks any signals that were not blocked by the old
+	// mask, and at least one such signal is pending in tg.pendingSignals, and
+	// t has been woken, it could be the case that t was woken to handle that
+	// signal, but will no longer do so as a result of its new signal mask, so
+	// we have to pick a replacement.
+	blocked := mask &^ oldMask
+	blockedGroupPending := blocked & t.tg.pendingSignals.pendingSet
+	if blockedGroupPending != 0 && t.interrupted() {
+		linux.ForEachSignal(blockedGroupPending, func(sig linux.Signal) {
+			if nt := t.tg.findSignalReceiverLocked(sig); nt != nil {
+				nt.interrupt()
+				return
+			}
+		})
+		// We have to re-issue the interrupt consumed by t.interrupted() since
+		// it might have been for a different reason.
+		t.interruptSelf()
+	}
+
+	// Conversely, if the new mask unblocks any signals that were blocked by
+	// the old mask, and at least one such signal is pending, we may now need
+	// to handle that signal.
+	unblocked := oldMask &^ mask
+	unblockedPending := unblocked & (t.pendingSignals.pendingSet | t.tg.pendingSignals.pendingSet)
+	if unblockedPending != 0 {
+		t.interruptSelf()
+	}
+}
+
+// SetSavedSignalMask sets the saved signal mask (see Task.savedSignalMask's
+// comment).
+//
+// Preconditions: SetSavedSignalMask can only be called by the task goroutine.
+func (t *Task) SetSavedSignalMask(mask linux.SignalSet) {
+	t.savedSignalMask = mask
+	t.haveSavedSignalMask = true
+}
+
+// SignalStack returns the task-private signal stack.
+func (t *Task) SignalStack() arch.SignalStack {
+	return t.signalStack
+}
+
+// OnSignalStack returns true if, when the task resumes running, it will run on
+// the task-private signal stack.
+func (t *Task) OnSignalStack(s arch.SignalStack) bool {
+	sp := usermem.Addr(t.Arch().Stack())
+	return usermem.Addr(s.Addr) <= sp && sp < usermem.Addr(s.Addr+s.Size)
+}
+
+// SetSignalStack sets the task-private signal stack and clears the
+// SignalStackFlagDisable, since we have a signal stack.
+func (t *Task) SetSignalStack(alt arch.SignalStack) error {
+	// Mask out irrelevant parts: only disable matters.
+	alt.Flags &= arch.SignalStackFlagDisable
+	t.signalStack = alt
+	return nil
+}
+
+// SetSignalAct atomically sets the thread group's signal action for signal sig
+// to *actptr (if actptr is not nil) and returns the old signal action.
+func (tg *ThreadGroup) SetSignalAct(sig linux.Signal, actptr *arch.SignalAct) (arch.SignalAct, error) {
+	if !sig.IsValid() {
+		return arch.SignalAct{}, syserror.EINVAL
+	}
+
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	sh := tg.signalHandlers
+	sh.mu.Lock()
+	defer sh.mu.Unlock()
+	oldact := sh.actions[sig]
+	if actptr != nil {
+		if sig == linux.SIGKILL || sig == linux.SIGSTOP {
+			return oldact, syserror.EINVAL
+		}
+
+		act := *actptr
+		act.Mask &^= UnblockableSignals
+		sh.actions[sig] = act
+		// From POSIX, by way of Linux:
+		//
+		// "Setting a signal action to SIG_IGN for a signal that is pending
+		// shall cause the pending signal to be discarded, whether or not it is
+		// blocked."
+		//
+		// "Setting a signal action to SIG_DFL for a signal that is pending and
+		// whose default action is to ignore the signal (for example, SIGCHLD),
+		// shall cause the pending signal to be discarded, whether or not it is
+		// blocked."
+		if computeAction(sig, act) == SignalActionIgnore {
+			tg.discardSpecificLocked(sig)
+		}
+	}
+	return oldact, nil
+}
+
+// CopyOutSignalAct converts the given SignalAct into an architecture-specific
+// type and then copies it out to task memory.
+func (t *Task) CopyOutSignalAct(addr usermem.Addr, s *arch.SignalAct) error {
+	n := t.Arch().NewSignalAct()
+	n.SerializeFrom(s)
+	_, err := t.CopyOut(addr, n)
+	return err
+}
+
+// CopyInSignalAct copies an architecture-specific sigaction type from task
+// memory and then converts it into a SignalAct.
+func (t *Task) CopyInSignalAct(addr usermem.Addr) (arch.SignalAct, error) {
+	n := t.Arch().NewSignalAct()
+	var s arch.SignalAct
+	if _, err := t.CopyIn(addr, n); err != nil {
+		return s, err
+	}
+	n.DeserializeTo(&s)
+	return s, nil
+}
+
+// CopyOutSignalStack converts the given SignalStack into an
+// architecture-specific type and then copies it out to task memory.
+func (t *Task) CopyOutSignalStack(addr usermem.Addr, s *arch.SignalStack) error {
+	n := t.Arch().NewSignalStack()
+	n.SerializeFrom(s)
+	_, err := t.CopyOut(addr, n)
+	return err
+}
+
+// CopyInSignalStack copies an architecture-specific stack_t from task memory
+// and then converts it into a SignalStack.
+func (t *Task) CopyInSignalStack(addr usermem.Addr) (arch.SignalStack, error) {
+	n := t.Arch().NewSignalStack()
+	var s arch.SignalStack
+	if _, err := t.CopyIn(addr, n); err != nil {
+		return s, err
+	}
+	n.DeserializeTo(&s)
+	return s, nil
+}
+
+// groupStop is a TaskStop placed on tasks that have received a stop signal
+// (SIGSTOP, SIGTSTP, SIGTTIN, SIGTTOU). (The term "group-stop" originates from
+// the ptrace man page.)
+type groupStop struct{}
+
+// Killable implements TaskStop.Killable.
+func (*groupStop) Killable() bool { return true }
+
+type groupStopPhase int
+
+const (
+	// groupStopNone indicates that a thread group is not in, or attempting to
+	// enter or leave, a group stop.
+	groupStopNone groupStopPhase = iota
+
+	// groupStopDequeued indicates that at least one task in a thread group has
+	// dequeued a stop signal (or dequeued any signal and entered a
+	// signal-delivery-stop as a result, which allows ptrace to change the
+	// signal into a stop signal), but temporarily dropped the signal mutex
+	// without initiating the group stop.
+	//
+	// groupStopDequeued is analogous to JOBCTL_STOP_DEQUEUED in Linux.
+	groupStopDequeued
+
+	// groupStopInitiated indicates that a task in a thread group has initiated
+	// a group stop, but not all tasks in the thread group have acknowledged
+	// entering the group stop.
+	//
+	// groupStopInitiated is represented by JOBCTL_STOP_PENDING &&
+	// !SIGNAL_STOP_STOPPED in Linux.
+	groupStopInitiated
+
+	// groupStopComplete indicates that all tasks in a thread group have
+	// acknowledged entering the group stop, and the last one to do so has
+	// notified the thread group's parent.
+	//
+	// groupStopComplete is represented by JOBCTL_STOP_PENDING &&
+	// SIGNAL_STOP_STOPPED in Linux.
+	groupStopComplete
+)
+
+// initiateGroupStop attempts to initiate a group stop based on a
+// previously-dequeued stop signal.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) initiateGroupStop(info *arch.SignalInfo) {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	if t.tg.groupStopPhase != groupStopDequeued {
+		t.Debugf("Signal %d: not stopping thread group: lost to racing signal", info.Signo)
+		return
+	}
+	if t.tg.exiting {
+		t.Debugf("Signal %d: not stopping thread group: lost to racing group exit", info.Signo)
+		return
+	}
+	if t.tg.execing != nil {
+		t.Debugf("Signal %d: not stopping thread group: lost to racing execve", info.Signo)
+		return
+	}
+	t.Debugf("Signal %d: stopping thread group", info.Signo)
+	t.tg.groupStopPhase = groupStopInitiated
+	t.tg.groupStopSignal = linux.Signal(info.Signo)
+	t.tg.groupStopCount = 0
+	for t2 := t.tg.tasks.Front(); t2 != nil; t2 = t2.Next() {
+		t2.groupStopRequired = true
+		t2.groupStopAcknowledged = false
+		t2.interrupt()
+	}
+}
+
+// endGroupStopLocked ensures that all prior stop signals received by tg are
+// not stopping tg and will not stop tg in the future. If broadcast is true,
+// parent and tracer notification will be scheduled if appropriate.
+//
+// Preconditions: The signal mutex must be locked.
+func (tg *ThreadGroup) endGroupStopLocked(broadcast bool) {
+	// Discard all previously-queued stop signals.
+	linux.ForEachSignal(StopSignals, tg.discardSpecificLocked)
+
+	if tg.groupStopPhase != groupStopNone {
+		tg.leader.Debugf("Ending group stop currently in phase %d", tg.groupStopPhase)
+		if tg.groupStopPhase == groupStopInitiated || tg.groupStopPhase == groupStopComplete {
+			tg.groupStopSignal = 0
+			for t := tg.tasks.Front(); t != nil; t = t.Next() {
+				if _, ok := t.stop.(*groupStop); ok {
+					t.endInternalStopLocked()
+				}
+			}
+			if broadcast {
+				// Instead of notifying the parent here, set groupContNotify so
+				// that one of the continuing tasks does so. (Linux does
+				// something similar.) The reason we do this is to keep locking
+				// sane. In order to send a signal to the parent, we need to
+				// lock its signal mutex, but we're already holding tg's signal
+				// mutex, and the TaskSet mutex must be locked for writing for
+				// us to hold two signal mutexes. Since we don't want to
+				// require this for endGroupStopLocked (which is called from
+				// signal-sending paths), nor do we want to lose atomicity by
+				// releasing the mutexes we're already holding, just let the
+				// continuing thread group deal with it.
+				tg.groupContNotify = true
+				tg.groupContInterrupted = tg.groupStopPhase == groupStopInitiated
+				tg.groupContWaitable = true
+			}
+		}
+		// If groupStopPhase was groupStopDequeued, setting it to groupStopNone
+		// will cause following calls to initiateGroupStop to recognize that
+		// the group stop has been cancelled.
+		tg.groupStopPhase = groupStopNone
+	}
+}
+
+// signalStop sends a signal to t's thread group of a new group stop, group
+// continue, or ptrace stop, if appropriate. code and status are set in the
+// signal sent to tg, if any.
+//
+// Preconditions: The TaskSet mutex must be locked (for reading or writing).
+func (t *Task) signalStop(target *Task, code int32, status int32) {
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	act, ok := t.tg.signalHandlers.actions[linux.SIGCHLD]
+	if !ok || (act.Handler != arch.SignalActIgnore && act.Flags&arch.SignalFlagNoCldStop == 0) {
+		sigchld := &arch.SignalInfo{
+			Signo: int32(linux.SIGCHLD),
+			Code:  code,
+		}
+		sigchld.SetPid(int32(t.tg.pidns.tids[target]))
+		sigchld.SetUid(int32(target.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()))
+		sigchld.SetStatus(status)
+		// TODO: Set utime, stime.
+		t.sendSignalLocked(sigchld, true /* group */)
+	}
+}
+
+// The runInterrupt state handles conditions indicated by interrupts.
+type runInterrupt struct{}
+
+func (*runInterrupt) execute(t *Task) taskRunState {
+	// Interrupts are de-duplicated (if t is interrupted twice before
+	// t.interrupted() is called, t.interrupted() will only return true once),
+	// so early exits from this function must re-enter the runInterrupt state
+	// to check for more interrupt-signaled conditions.
+
+	t.tg.signalHandlers.mu.Lock()
+
+	// Did we just leave a group stop?
+	if t.tg.groupContNotify {
+		t.tg.groupContNotify = false
+		sig := t.tg.groupStopSignal
+		intr := t.tg.groupContInterrupted
+		t.tg.signalHandlers.mu.Unlock()
+		t.tg.pidns.owner.mu.RLock()
+		// For consistency with Linux, if the parent and (thread group
+		// leader's) tracer are in the same thread group, deduplicate
+		// notifications.
+		notifyParent := t.tg.leader.parent != nil
+		if tracer := t.tg.leader.ptraceTracer.Load().(*Task); tracer != nil {
+			if notifyParent && tracer.tg == t.tg.leader.parent.tg {
+				notifyParent = false
+			}
+			// Sending CLD_STOPPED to the tracer doesn't really make any sense;
+			// the thread group leader may have already entered the stop and
+			// notified its tracer accordingly. But it's consistent with
+			// Linux...
+			if intr {
+				tracer.signalStop(t.tg.leader, arch.CLD_STOPPED, int32(sig))
+				if !notifyParent {
+					tracer.tg.eventQueue.Notify(EventGroupContinue | EventTraceeStop | EventChildGroupStop)
+				} else {
+					tracer.tg.eventQueue.Notify(EventGroupContinue | EventTraceeStop)
+				}
+			} else {
+				tracer.signalStop(t.tg.leader, arch.CLD_CONTINUED, int32(sig))
+				tracer.tg.eventQueue.Notify(EventGroupContinue)
+			}
+		}
+		if notifyParent {
+			// If groupContInterrupted, do as Linux does and pretend the group
+			// stop completed just before it ended. The theoretical behavior in
+			// this case would be to send a SIGCHLD indicating the completed
+			// stop, followed by a SIGCHLD indicating the continue. However,
+			// SIGCHLD is a standard signal, so the latter would always be
+			// dropped. Hence sending only the former is equivalent.
+			if intr {
+				t.tg.leader.parent.signalStop(t.tg.leader, arch.CLD_STOPPED, int32(sig))
+				t.tg.leader.parent.tg.eventQueue.Notify(EventGroupContinue | EventChildGroupStop)
+			} else {
+				t.tg.leader.parent.signalStop(t.tg.leader, arch.CLD_CONTINUED, int32(sig))
+				t.tg.leader.parent.tg.eventQueue.Notify(EventGroupContinue)
+			}
+		}
+		t.tg.pidns.owner.mu.RUnlock()
+		return (*runInterrupt)(nil)
+	}
+
+	// Do we need to enter a group stop?
+	if t.groupStopRequired {
+		t.groupStopRequired = false
+		sig := t.tg.groupStopSignal
+		notifyParent := false
+		if !t.groupStopAcknowledged {
+			t.groupStopAcknowledged = true
+			t.tg.groupStopCount++
+			if t.tg.groupStopCount == t.tg.activeTasks {
+				t.Debugf("Completing group stop")
+				notifyParent = true
+				t.tg.groupStopPhase = groupStopComplete
+				t.tg.groupStopWaitable = true
+				t.tg.groupContNotify = false
+				t.tg.groupContWaitable = false
+			}
+		}
+		// Drop the signal mutex so we can take the TaskSet mutex.
+		t.tg.signalHandlers.mu.Unlock()
+
+		t.tg.pidns.owner.mu.RLock()
+		if t.tg.leader.parent == nil {
+			notifyParent = false
+		}
+		if tracer := t.Tracer(); tracer != nil {
+			t.ptraceCode = int32(sig)
+			t.ptraceSiginfo = nil
+			if t.beginPtraceStopLocked() {
+				tracer.signalStop(t, arch.CLD_STOPPED, int32(sig))
+				// For consistency with Linux, if the parent and tracer are in the
+				// same thread group, deduplicate notification signals.
+				if notifyParent && tracer.tg == t.tg.leader.parent.tg {
+					notifyParent = false
+					tracer.tg.eventQueue.Notify(EventChildGroupStop | EventTraceeStop)
+				} else {
+					tracer.tg.eventQueue.Notify(EventTraceeStop)
+				}
+			}
+		} else {
+			t.tg.signalHandlers.mu.Lock()
+			if !t.killedLocked() {
+				t.beginInternalStopLocked((*groupStop)(nil))
+			}
+			t.tg.signalHandlers.mu.Unlock()
+		}
+		if notifyParent {
+			t.tg.leader.parent.signalStop(t.tg.leader, arch.CLD_STOPPED, int32(sig))
+			t.tg.leader.parent.tg.eventQueue.Notify(EventChildGroupStop)
+		}
+		t.tg.pidns.owner.mu.RUnlock()
+
+		return (*runInterrupt)(nil)
+	}
+
+	// Are there signals pending?
+	if info := t.dequeueSignalLocked(); info != nil {
+		if linux.SignalSetOf(linux.Signal(info.Signo))&StopSignals != 0 && t.tg.groupStopPhase == groupStopNone {
+			// Indicate that we've dequeued a stop signal before
+			// unlocking the signal mutex; initiateGroupStop will check
+			// that the phase hasn't changed (or is at least another
+			// "stop signal dequeued" phase) after relocking it.
+			t.tg.groupStopPhase = groupStopDequeued
+		}
+		if t.ptraceSignalLocked(info) {
+			// Dequeueing the signal action must wait until after the
+			// signal-delivery-stop ends since the tracer can change or
+			// suppress the signal.
+			t.tg.signalHandlers.mu.Unlock()
+			return (*runInterruptAfterSignalDeliveryStop)(nil)
+		}
+		act := t.tg.signalHandlers.dequeueAction(linux.Signal(info.Signo))
+		t.tg.signalHandlers.mu.Unlock()
+		return t.deliverSignal(info, act)
+	}
+
+	t.tg.signalHandlers.mu.Unlock()
+	return (*runApp)(nil)
+}
+
+type runInterruptAfterSignalDeliveryStop struct{}
+
+func (*runInterruptAfterSignalDeliveryStop) execute(t *Task) taskRunState {
+	t.tg.pidns.owner.mu.Lock()
+	// Can't defer unlock: deliverSignal must be called without holding TaskSet
+	// mutex.
+	sig := linux.Signal(t.ptraceCode)
+	defer func() {
+		t.ptraceSiginfo = nil
+	}()
+	if !sig.IsValid() {
+		t.tg.pidns.owner.mu.Unlock()
+		return (*runInterrupt)(nil)
+	}
+	info := t.ptraceSiginfo
+	if sig != linux.Signal(info.Signo) {
+		info.Signo = int32(sig)
+		info.Errno = 0
+		info.Code = arch.SignalInfoUser
+		// pid isn't a valid field for all signal numbers, but Linux
+		// doesn't care (kernel/signal.c:ptrace_signal()).
+		//
+		// Linux uses t->parent for the tid and uid here, which is the tracer
+		// if it hasn't detached or the real parent otherwise.
+		parent := t.parent
+		if tracer := t.Tracer(); tracer != nil {
+			parent = tracer
+		}
+		if parent == nil {
+			// Tracer has detached and t was created by Kernel.CreateProcess().
+			// Pretend the parent is in an ancestor PID + user namespace.
+			info.SetPid(0)
+			info.SetUid(int32(auth.OverflowUID))
+		} else {
+			info.SetPid(int32(t.tg.pidns.tids[parent]))
+			info.SetUid(int32(parent.Credentials().RealKUID.In(t.UserNamespace()).OrOverflow()))
+		}
+	}
+	t.tg.signalHandlers.mu.Lock()
+	t.tg.pidns.owner.mu.Unlock()
+	// If the signal is masked, re-queue it.
+	if linux.SignalSetOf(sig)&t.tr.SignalMask != 0 {
+		t.sendSignalLocked(info, false /* group */)
+		t.tg.signalHandlers.mu.Unlock()
+		return (*runInterrupt)(nil)
+	}
+	act := t.tg.signalHandlers.dequeueAction(linux.Signal(info.Signo))
+	t.tg.signalHandlers.mu.Unlock()
+	return t.deliverSignal(info, act)
+}
diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go
new file mode 100644
index 000000000..801cb3395
--- /dev/null
+++ b/pkg/sentry/kernel/task_start.go
@@ -0,0 +1,252 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/futex"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// TaskConfig defines the configuration of a new Task (see below).
+type TaskConfig struct {
+	// Kernel is the owning Kernel.
+	*Kernel
+
+	// Parent is the new task's parent. Parent may be nil.
+	Parent *Task
+
+	// ThreadGroup is the ThreadGroup the new task belongs to.
+	*ThreadGroup
+
+	// TaskContext is the TaskContext of the new task.
+	*TaskContext
+
+	// TaskResources is the TaskResources of the new task.
+	*TaskResources
+
+	// Credentials is the Credentials of the new task.
+	Credentials *auth.Credentials
+
+	// Niceness is the niceness of the new task.
+	Niceness int
+
+	// If NetworkNamespaced is true, the new task should observe a non-root
+	// network namespace.
+	NetworkNamespaced bool
+
+	// AllowedCPUMask contains the cpus that this task can run on.
+	AllowedCPUMask sched.CPUSet
+
+	// UTSNamespace is the UTSNamespace of the new task.
+	UTSNamespace *UTSNamespace
+
+	// IPCNamespace is the IPCNamespace of the new task.
+	IPCNamespace *IPCNamespace
+}
+
+// NewTask creates a new task defined by TaskConfig.
+// Whether or not NewTask is successful, it takes ownership of both TaskContext
+// and TaskResources of the TaskConfig.
+//
+// NewTask does not start the returned task; the caller must call Task.Start.
+func (ts *TaskSet) NewTask(cfg *TaskConfig) (*Task, error) {
+	t, err := ts.newTask(cfg)
+	if err != nil {
+		cfg.TaskContext.release()
+		cfg.TaskResources.release()
+		return nil, err
+	}
+	return t, nil
+}
+
+// newTask is a helper for TaskSet.NewTask that only takes ownership of TaskContext
+// and TaskResources of the TaskConfig if it succeeds.
+func (ts *TaskSet) newTask(cfg *TaskConfig) (*Task, error) {
+	tg := cfg.ThreadGroup
+	tc := cfg.TaskContext
+	t := &Task{
+		taskNode: taskNode{
+			tg:       tg,
+			parent:   cfg.Parent,
+			children: make(map[*Task]struct{}),
+		},
+		runState:       (*runApp)(nil),
+		interruptChan:  make(chan struct{}, 1),
+		signalStack:    arch.SignalStack{Flags: arch.SignalStackFlagDisable},
+		tc:             *tc,
+		tr:             *cfg.TaskResources,
+		p:              cfg.Kernel.Platform.NewContext(),
+		k:              cfg.Kernel,
+		ptraceTracees:  make(map[*Task]struct{}),
+		allowedCPUMask: cfg.AllowedCPUMask.Copy(),
+		ioUsage:        &usage.IO{},
+		creds:          cfg.Credentials,
+		niceness:       cfg.Niceness,
+		netns:          cfg.NetworkNamespaced,
+		utsns:          cfg.UTSNamespace,
+		ipcns:          cfg.IPCNamespace,
+		rseqCPU:        -1,
+		futexWaiter:    futex.NewWaiter(),
+	}
+	t.endStopCond.L = &t.tg.signalHandlers.mu
+	t.ptraceTracer.Store((*Task)(nil))
+	// We don't construct t.blockingTimer until Task.run(); see that function
+	// for justification.
+
+	// Make the new task (and possibly thread group) visible to the rest of
+	// the system atomically.
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	tg.signalHandlers.mu.Lock()
+	defer tg.signalHandlers.mu.Unlock()
+	if tg.exiting || tg.execing != nil {
+		// If the caller is in the same thread group, then what we return
+		// doesn't matter too much since the caller will exit before it returns
+		// to userspace. If the caller isn't in the same thread group, then
+		// we're in uncharted territory and can return whatever we want.
+		return nil, syserror.EINTR
+	}
+	if err := ts.assignTIDsLocked(t); err != nil {
+		return nil, err
+	}
+	// Below this point, newTask is expected not to fail (there is no rollback
+	// of assignTIDsLocked or any of the following).
+
+	// Logging on t's behalf will panic if t.logPrefix hasn't been initialized.
+	// This is the earliest point at which we can do so (since t now has thread
+	// IDs).
+	t.updateLogPrefixLocked()
+
+	if t.parent != nil {
+		t.parent.children[t] = struct{}{}
+	}
+
+	if tg.leader == nil {
+		// New thread group.
+		tg.leader = t
+		if parentPG := tg.parentPG(); parentPG == nil {
+			tg.createSession()
+		} else {
+			// Inherit the process group.
+			parentPG.incRefWithParent(parentPG)
+			tg.processGroup = parentPG
+		}
+	}
+	tg.tasks.PushBack(t)
+	tg.tasksCount++
+	tg.liveTasks++
+	tg.activeTasks++
+
+	// Propagate external TaskSet stops to the new task.
+	t.stopCount = ts.stopCount
+
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	t.cpu = assignCPU(t.allowedCPUMask, ts.Root.tids[t])
+
+	t.startTime = t.k.RealtimeClock().Now()
+
+	return t, nil
+}
+
+// assignTIDsLocked ensures that new task t is visible in all PID namespaces in
+// which it should be visible.
+//
+// Preconditions: ts.mu must be locked for writing.
+func (ts *TaskSet) assignTIDsLocked(t *Task) error {
+	type allocatedTID struct {
+		ns  *PIDNamespace
+		tid ThreadID
+	}
+	var allocatedTIDs []allocatedTID
+	for ns := t.tg.pidns; ns != nil; ns = ns.parent {
+		tid, err := ns.allocateTID()
+		if err != nil {
+			// Failure. Remove the tids we already allocated in descendant
+			// namespaces.
+			for _, a := range allocatedTIDs {
+				delete(a.ns.tasks, a.tid)
+				delete(a.ns.tids, t)
+			}
+			return err
+		}
+		ns.tasks[tid] = t
+		ns.tids[t] = tid
+		allocatedTIDs = append(allocatedTIDs, allocatedTID{ns, tid})
+	}
+	return nil
+}
+
+// allocateTID returns an unused ThreadID from ns.
+//
+// Preconditions: ns.owner.mu must be locked for writing.
+func (ns *PIDNamespace) allocateTID() (ThreadID, error) {
+	if ns.exiting {
+		// "In this case, a subsequent fork(2) into this PID namespace will
+		// fail with the error ENOMEM; it is not possible to create a new
+		// processes [sic] in a PID namespace whose init process has
+		// terminated." - pid_namespaces(7)
+		return 0, syserror.ENOMEM
+	}
+	tid := ns.last
+	for {
+		// Next.
+		tid++
+		if tid > TasksLimit {
+			tid = InitTID + 1
+		}
+
+		// Is it available?
+		_, ok := ns.tasks[tid]
+		if !ok {
+			ns.last = tid
+			return tid, nil
+		}
+
+		// Did we do a full cycle?
+		if tid == ns.last {
+			// No tid available.
+			return 0, syserror.EAGAIN
+		}
+	}
+}
+
+// Start starts the task goroutine. Start must be called exactly once for each
+// task returned by NewTask.
+//
+// 'tid' must be the task's TID in the root PID namespace and it's used for
+// debugging purposes only (set as parameter to Task.run to make it visible
+// in stack dumps).
+func (t *Task) Start(tid ThreadID) {
+	// If the task was restored, it may be "starting" after having already exited.
+	if t.runState == nil {
+		return
+	}
+	t.goroutineStopped.Add(1)
+	t.tg.liveGoroutines.Add(1)
+	t.tg.pidns.owner.liveGoroutines.Add(1)
+	t.tg.pidns.owner.runningGoroutines.Add(1)
+
+	// Task is now running in system mode.
+	t.accountTaskGoroutineLeave(TaskGoroutineNonexistent)
+
+	// Use the task's TID in the root PID namespace to make it visible in stack dumps.
+	go t.run(uintptr(tid)) // S/R-SAFE: synchronizes with saving through stops
+}
diff --git a/pkg/sentry/kernel/task_stop.go b/pkg/sentry/kernel/task_stop.go
new file mode 100644
index 000000000..feaf6cae4
--- /dev/null
+++ b/pkg/sentry/kernel/task_stop.go
@@ -0,0 +1,226 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// This file implements task stops, which represent the equivalent of Linux's
+// uninterruptible sleep states in a way that is compatible with save/restore.
+// Task stops comprise both internal stops (which form part of the task's
+// "normal" control flow) and external stops (which do not); see README.md for
+// details.
+//
+// There are multiple interfaces for interacting with stops because there are
+// multiple cases to consider:
+//
+// - A task goroutine can begin a stop on its associated task (e.g. a
+// vfork() syscall stopping the calling task until the child task releases its
+// MM). In this case, calling Task.interrupt is both unnecessary (the task
+// goroutine obviously cannot be blocked in Task.block or executing application
+// code) and undesirable (as it may spuriously interrupt a in-progress
+// syscall).
+//
+// Beginning internal stops in this case is implemented by
+// Task.beginInternalStop / Task.beginInternalStopLocked. As of this writing,
+// there are no instances of this case that begin external stops, except for
+// autosave; however, autosave terminates the sentry without ending the
+// external stop, so the spurious interrupt is moot.
+//
+// - An arbitrary goroutine can begin a stop on an unrelated task (e.g. all
+// tasks being stopped in preparation for state checkpointing). If the task
+// goroutine may be in Task.block or executing application code, it must be
+// interrupted by Task.interrupt for it to actually enter the stop; since,
+// strictly speaking, we have no way of determining this, we call
+// Task.interrupt unconditionally.
+//
+// Beginning external stops in this case is implemented by
+// Task.BeginExternalStop. As of this writing, there are no instances of this
+// case that begin internal stops.
+//
+// - An arbitrary goroutine can end a stop on an unrelated task (e.g. an
+// exiting task resuming a sibling task that has been blocked in an execve()
+// syscall waiting for other tasks to exit). In this case, Task.endStopCond
+// must be notified to kick the task goroutine out of Task.doStop.
+//
+// Ending internal stops in this case is implemented by
+// Task.endInternalStopLocked. Ending external stops in this case is
+// implemented by Task.EndExternalStop.
+//
+// - Hypothetically, a task goroutine can end an internal stop on its
+// associated task. As of this writing, there are no instances of this case.
+// However, any instances of this case could still use the above functions,
+// since notifying Task.endStopCond would be unnecessary but harmless.
+
+import (
+	"fmt"
+	"sync/atomic"
+)
+
+// A TaskStop is a condition visible to the task control flow graph that
+// prevents a task goroutine from running or exiting, i.e. an internal stop.
+//
+// NOTE: Most TaskStops don't contain any data; they're
+// distinguished by their type. The obvious way to implement such a TaskStop
+// is:
+//
+//     type groupStop struct{}
+//     func (groupStop) Killable() bool { return true }
+//     ...
+//     t.beginInternalStop(groupStop{})
+//
+// However, this doesn't work because the state package can't serialize values,
+// only pointers. Furthermore, the correctness of save/restore depends on the
+// ability to pass a TaskStop to endInternalStop that will compare equal to the
+// TaskStop that was passed to beginInternalStop, even if a save/restore cycle
+// occurred between the two. As a result, the current idiom is to always use a
+// typecast nil for data-free TaskStops:
+//
+//     type groupStop struct{}
+//     func (*groupStop) Killable() bool { return true }
+//     ...
+//     t.beginInternalStop((*groupStop)(nil))
+//
+// This is pretty gross, but the alternatives seem grosser.
+type TaskStop interface {
+	// Killable returns true if Task.Kill should end the stop prematurely.
+	// Killable is analogous to Linux's TASK_WAKEKILL.
+	Killable() bool
+}
+
+// beginInternalStop indicates the start of an internal stop that applies to t.
+//
+// Preconditions: The task must not already be in an internal stop (i.e. t.stop
+// == nil). The caller must be running on the task goroutine.
+func (t *Task) beginInternalStop(s TaskStop) {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.beginInternalStopLocked(s)
+}
+
+// Preconditions: The signal mutex must be locked. All preconditions for
+// Task.beginInternalStop also apply.
+func (t *Task) beginInternalStopLocked(s TaskStop) {
+	if t.stop != nil {
+		panic(fmt.Sprintf("Attempting to enter internal stop %#v when already in internal stop %#v", s, t.stop))
+	}
+	t.Debugf("Entering internal stop %#v", s)
+	t.stop = s
+	t.beginStopLocked()
+}
+
+// endInternalStopLocked indicates the end of an internal stop that applies to
+// t. endInternalStopLocked does not wait for the task to resume.
+//
+// The caller is responsible for ensuring that the internal stop they expect
+// actually applies to t; this requires holding the signal mutex which protects
+// t.stop, which is why there is no endInternalStop that locks the signal mutex
+// for you.
+//
+// Preconditions: The signal mutex must be locked. The task must be in an
+// internal stop (i.e. t.stop != nil).
+func (t *Task) endInternalStopLocked() {
+	if t.stop == nil {
+		panic("Attempting to leave non-existent internal stop")
+	}
+	t.Debugf("Leaving internal stop %#v", t.stop)
+	t.stop = nil
+	t.endStopLocked()
+}
+
+// BeginExternalStop indicates the start of an external stop that applies to t.
+// BeginExternalStop does not wait for t's task goroutine to stop.
+func (t *Task) BeginExternalStop() {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.beginStopLocked()
+	t.interrupt()
+}
+
+// EndExternalStop indicates the end of an external stop started by a previous
+// call to Task.BeginExternalStop. EndExternalStop does not wait for t's task
+// goroutine to resume.
+func (t *Task) EndExternalStop() {
+	t.tg.pidns.owner.mu.RLock()
+	defer t.tg.pidns.owner.mu.RUnlock()
+	t.tg.signalHandlers.mu.Lock()
+	defer t.tg.signalHandlers.mu.Unlock()
+	t.endStopLocked()
+}
+
+// beginStopLocked increments t.stopCount to indicate that a new internal or
+// external stop applies to t.
+//
+// Preconditions: The signal mutex must be locked.
+func (t *Task) beginStopLocked() {
+	if newval := atomic.AddInt32(&t.stopCount, 1); newval <= 0 {
+		// Most likely overflow.
+		panic(fmt.Sprintf("Invalid stopCount: %d", newval))
+	}
+}
+
+// endStopLocked decerements t.stopCount to indicate that an existing internal
+// or external stop no longer applies to t.
+//
+// Preconditions: The signal mutex must be locked.
+func (t *Task) endStopLocked() {
+	if newval := atomic.AddInt32(&t.stopCount, -1); newval < 0 {
+		panic(fmt.Sprintf("Invalid stopCount: %d", newval))
+	} else if newval == 0 {
+		t.endStopCond.Signal()
+	}
+}
+
+// BeginExternalStop indicates the start of an external stop that applies to
+// all current and future tasks in ts. BeginExternalStop does not wait for
+// task goroutines to stop.
+func (ts *TaskSet) BeginExternalStop() {
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	ts.stopCount++
+	if ts.stopCount <= 0 {
+		panic(fmt.Sprintf("Invalid stopCount: %d", ts.stopCount))
+	}
+	if ts.Root == nil {
+		return
+	}
+	for t := range ts.Root.tids {
+		t.tg.signalHandlers.mu.Lock()
+		t.beginStopLocked()
+		t.tg.signalHandlers.mu.Unlock()
+		t.interrupt()
+	}
+}
+
+// EndExternalStop indicates the end of an external stop started by a previous
+// call to TaskSet.BeginExternalStop. EndExternalStop does not wait for task
+// goroutines to resume.
+func (ts *TaskSet) EndExternalStop() {
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	ts.stopCount--
+	if ts.stopCount < 0 {
+		panic(fmt.Sprintf("Invalid stopCount: %d", ts.stopCount))
+	}
+	if ts.Root == nil {
+		return
+	}
+	for t := range ts.Root.tids {
+		t.tg.signalHandlers.mu.Lock()
+		t.endStopLocked()
+		t.tg.signalHandlers.mu.Unlock()
+	}
+}
diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go
new file mode 100644
index 000000000..79f4ff60c
--- /dev/null
+++ b/pkg/sentry/kernel/task_syscall.go
@@ -0,0 +1,434 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"os"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SyscallRestartErrno represents a ERESTART* errno defined in the Linux's kernel
+// include/linux/errno.h. These errnos are never returned to userspace
+// directly, but are used to communicate the expected behavior of an
+// interrupted syscall from the syscall to signal handling.
+type SyscallRestartErrno int
+
+// These numeric values are significant because ptrace syscall exit tracing can
+// observe them.
+//
+// For all of the following errnos, if the syscall is not interrupted by a
+// signal delivered to a user handler, the syscall is restarted.
+const (
+	// ERESTARTSYS is returned by an interrupted syscall to indicate that it
+	// should be converted to EINTR if interrupted by a signal delivered to a
+	// user handler without SA_RESTART set, and restarted otherwise.
+	ERESTARTSYS = SyscallRestartErrno(512)
+
+	// ERESTARTNOINTR is returned by an interrupted syscall to indicate that it
+	// should always be restarted.
+	ERESTARTNOINTR = SyscallRestartErrno(513)
+
+	// ERESTARTNOHAND is returned by an interrupted syscall to indicate that it
+	// should be converted to EINTR if interrupted by a signal delivered to a
+	// user handler, and restarted otherwise.
+	ERESTARTNOHAND = SyscallRestartErrno(514)
+
+	// ERESTART_RESTARTBLOCK is returned by an interrupted syscall to indicate
+	// that it should be restarted using a custom function. The interrupted
+	// syscall must register a custom restart function by calling
+	// Task.SetRestartSyscallFn.
+	ERESTART_RESTARTBLOCK = SyscallRestartErrno(516)
+)
+
+// Error implements error.Error.
+func (e SyscallRestartErrno) Error() string {
+	// Descriptions are borrowed from strace.
+	switch e {
+	case ERESTARTSYS:
+		return "to be restarted if SA_RESTART is set"
+	case ERESTARTNOINTR:
+		return "to be restarted"
+	case ERESTARTNOHAND:
+		return "to be restarted if no handler"
+	case ERESTART_RESTARTBLOCK:
+		return "interrupted by signal"
+	default:
+		return "(unknown interrupt error)"
+	}
+}
+
+// SyscallRestartErrnoFromReturn returns the SyscallRestartErrno represented by
+// rv, the value in a syscall return register.
+func SyscallRestartErrnoFromReturn(rv uintptr) (SyscallRestartErrno, bool) {
+	switch int(rv) {
+	case -int(ERESTARTSYS):
+		return ERESTARTSYS, true
+	case -int(ERESTARTNOINTR):
+		return ERESTARTNOINTR, true
+	case -int(ERESTARTNOHAND):
+		return ERESTARTNOHAND, true
+	case -int(ERESTART_RESTARTBLOCK):
+		return ERESTART_RESTARTBLOCK, true
+	default:
+		return 0, false
+	}
+}
+
+// SyscallRestartBlock represents the restart block for a syscall restartable
+// with a custom function. It encapsulates the state required to restart a
+// syscall across a S/R.
+type SyscallRestartBlock interface {
+	Restart(t *Task) (uintptr, error)
+}
+
+// SyscallControl is returned by syscalls to control the behavior of
+// Task.doSyscallInvoke.
+type SyscallControl struct {
+	// next is the state that the task goroutine should switch to. If next is
+	// nil, the task goroutine should continue to syscall exit as usual.
+	next taskRunState
+
+	// If ignoreReturn is true, Task.doSyscallInvoke should not store any value
+	// in the task's syscall return value register.
+	ignoreReturn bool
+}
+
+var (
+	// CtrlDoExit is returned by the implementations of the exit and exit_group
+	// syscalls to enter the task exit path directly, skipping syscall exit
+	// tracing.
+	CtrlDoExit = &SyscallControl{next: (*runExit)(nil), ignoreReturn: true}
+
+	// ctrlStopAndReinvokeSyscall is returned by syscalls using the external
+	// feature before syscall execution. This causes Task.doSyscallInvoke
+	// to return runSyscallReinvoke, allowing Task.run to check for stops
+	// before immediately re-invoking the syscall (skipping the re-checking
+	// of seccomp filters and ptrace which would confuse userspace
+	// tracing).
+	ctrlStopAndReinvokeSyscall = &SyscallControl{next: (*runSyscallReinvoke)(nil), ignoreReturn: true}
+
+	// ctrlStopBeforeSyscallExit is returned by syscalls that initiate a stop at
+	// their end. This causes Task.doSyscallInvoke to return runSyscallExit, rather
+	// than tail-calling it, allowing stops to be checked before syscall exit.
+	ctrlStopBeforeSyscallExit = &SyscallControl{next: (*runSyscallExit)(nil)}
+)
+
+func (t *Task) invokeExternal() {
+	t.BeginExternalStop()
+	go func() { // S/R-SAFE: External control flow.
+		defer t.EndExternalStop()
+		t.SyscallTable().External(t.Kernel())
+	}()
+}
+
+func (t *Task) executeSyscall(sysno uintptr, args arch.SyscallArguments) (rval uintptr, ctrl *SyscallControl, err error) {
+	s := t.SyscallTable()
+
+	fe := s.FeatureEnable.Word(sysno)
+
+	var straceContext interface{}
+	if bits.IsAnyOn32(fe, StraceEnableBits) {
+		straceContext = s.Stracer.SyscallEnter(t, sysno, args, fe)
+	}
+
+	if bits.IsOn32(fe, ExternalBeforeEnable) && (s.ExternalFilterBefore == nil || s.ExternalFilterBefore(t, sysno, args)) {
+		t.invokeExternal()
+		// Ensure we check for stops, then invoke the syscall again.
+		ctrl = ctrlStopAndReinvokeSyscall
+	} else {
+		fn := s.Lookup(sysno)
+		if fn != nil {
+			// Call our syscall implementation.
+			rval, ctrl, err = fn(t, args)
+		} else {
+			// Use the missing function if not found.
+			rval, err = t.SyscallTable().Missing(t, sysno, args)
+		}
+	}
+
+	if bits.IsOn32(fe, ExternalAfterEnable) && (s.ExternalFilterAfter == nil || s.ExternalFilterAfter(t, sysno, args)) {
+		t.invokeExternal()
+		// Don't reinvoke the syscall.
+	}
+
+	if bits.IsAnyOn32(fe, StraceEnableBits) {
+		s.Stracer.SyscallExit(straceContext, t, sysno, rval, err)
+	}
+
+	return
+}
+
+// doSyscall is the entry point for an invocation of a system call specified by
+// the current state of t's registers.
+//
+// The syscall path is very hot; avoid defer.
+func (t *Task) doSyscall() taskRunState {
+	sysno := t.Arch().SyscallNo()
+	args := t.Arch().SyscallArgs()
+
+	// Tracers expect to see this between when the task traps into the kernel
+	// to perform a syscall and when the syscall is actually invoked.
+	// This useless-looking temporary is needed because Go.
+	tmp := uintptr(syscall.ENOSYS)
+	t.Arch().SetReturn(-tmp)
+
+	// Check seccomp filters. The nil check is for performance (as seccomp use
+	// is rare), not needed for correctness.
+	if t.syscallFilters != nil {
+		switch r := t.checkSeccompSyscall(int32(sysno), args, usermem.Addr(t.Arch().IP())); r {
+		case seccompResultDeny:
+			t.Debugf("Syscall %d: denied by seccomp", sysno)
+			return (*runSyscallExit)(nil)
+		case seccompResultAllow:
+			// ok
+		case seccompResultKill:
+			t.Debugf("Syscall %d: killed by seccomp", sysno)
+			t.PrepareExit(ExitStatus{Signo: int(linux.SIGSYS)})
+			return (*runExit)(nil)
+		case seccompResultTrace:
+			t.Debugf("Syscall %d: stopping for PTRACE_EVENT_SECCOMP", sysno)
+			return (*runSyscallAfterPtraceEventSeccomp)(nil)
+		default:
+			panic(fmt.Sprintf("Unknown seccomp result %d", r))
+		}
+	}
+
+	return t.doSyscallEnter(sysno, args)
+}
+
+type runSyscallAfterPtraceEventSeccomp struct{}
+
+func (*runSyscallAfterPtraceEventSeccomp) execute(t *Task) taskRunState {
+	if t.killed() {
+		// "[S]yscall-exit-stop is not generated prior to death by SIGKILL." -
+		// ptrace(2)
+		return (*runInterrupt)(nil)
+	}
+	sysno := t.Arch().SyscallNo()
+	// "The tracer can skip the system call by changing the syscall number to
+	// -1." - Documentation/prctl/seccomp_filter.txt
+	if sysno == ^uintptr(0) {
+		return (*runSyscallExit)(nil).execute(t)
+	}
+	args := t.Arch().SyscallArgs()
+	return t.doSyscallEnter(sysno, args)
+}
+
+func (t *Task) doSyscallEnter(sysno uintptr, args arch.SyscallArguments) taskRunState {
+	if next, ok := t.ptraceSyscallEnter(); ok {
+		return next
+	}
+	return t.doSyscallInvoke(sysno, args)
+}
+
+type runSyscallAfterSyscallEnterStop struct{}
+
+func (*runSyscallAfterSyscallEnterStop) execute(t *Task) taskRunState {
+	if sig := linux.Signal(t.ptraceCode); sig.IsValid() {
+		t.tg.signalHandlers.mu.Lock()
+		t.sendSignalLocked(sigPriv(sig), false /* group */)
+		t.tg.signalHandlers.mu.Unlock()
+	}
+	if t.killed() {
+		return (*runInterrupt)(nil)
+	}
+	sysno := t.Arch().SyscallNo()
+	if sysno == ^uintptr(0) {
+		return (*runSyscallExit)(nil)
+	}
+	args := t.Arch().SyscallArgs()
+	return t.doSyscallInvoke(sysno, args)
+}
+
+type runSyscallAfterSysemuStop struct{}
+
+func (*runSyscallAfterSysemuStop) execute(t *Task) taskRunState {
+	if sig := linux.Signal(t.ptraceCode); sig.IsValid() {
+		t.tg.signalHandlers.mu.Lock()
+		t.sendSignalLocked(sigPriv(sig), false /* group */)
+		t.tg.signalHandlers.mu.Unlock()
+	}
+	if t.killed() {
+		return (*runInterrupt)(nil)
+	}
+	return (*runSyscallExit)(nil).execute(t)
+}
+
+func (t *Task) doSyscallInvoke(sysno uintptr, args arch.SyscallArguments) taskRunState {
+	rval, ctrl, err := t.executeSyscall(sysno, args)
+
+	if ctrl != nil {
+		if !ctrl.ignoreReturn {
+			t.Arch().SetReturn(rval)
+		}
+		if ctrl.next != nil {
+			return ctrl.next
+		}
+	} else if err != nil {
+		t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno))))
+		t.haveSyscallReturn = true
+	} else {
+		t.Arch().SetReturn(rval)
+	}
+
+	return (*runSyscallExit)(nil).execute(t)
+}
+
+type runSyscallReinvoke struct{}
+
+func (*runSyscallReinvoke) execute(t *Task) taskRunState {
+	if t.killed() {
+		// It's possible that since the last execution, the task has
+		// been forcible killed. Invoking the system call here could
+		// result in an infinite loop if it is again preempted by an
+		// external stop and reinvoked.
+		return (*runInterrupt)(nil)
+	}
+
+	sysno := t.Arch().SyscallNo()
+	args := t.Arch().SyscallArgs()
+	return t.doSyscallInvoke(sysno, args)
+}
+
+type runSyscallExit struct{}
+
+func (*runSyscallExit) execute(t *Task) taskRunState {
+	t.ptraceSyscallExit()
+	return (*runApp)(nil)
+}
+
+// doVsyscall is the entry point for a vsyscall invocation of syscall sysno, as
+// indicated by an execution fault at address addr. doVsyscall returns the
+// task's next run state.
+func (t *Task) doVsyscall(addr usermem.Addr, sysno uintptr) taskRunState {
+	// Grab the caller up front, to make sure there's a sensible stack.
+	caller := t.Arch().Native(uintptr(0))
+	if _, err := t.CopyIn(usermem.Addr(t.Arch().Stack()), caller); err != nil {
+		t.Debugf("vsyscall %d: error reading return address from stack: %v", sysno, err)
+		t.forceSignal(linux.SIGSEGV, false /* unconditional */)
+		t.SendSignal(sigPriv(linux.SIGSEGV))
+		return (*runApp)(nil)
+	}
+
+	// For _vsyscalls_, there is no need to translate System V calling convention
+	// to syscall ABI because they both use RDI, RSI, and RDX for the first three
+	// arguments and none of the vsyscalls uses more than two arguments.
+	args := t.Arch().SyscallArgs()
+	if t.syscallFilters != nil {
+		switch r := t.checkSeccompSyscall(int32(sysno), args, addr); r {
+		case seccompResultDeny:
+			t.Debugf("vsyscall %d, caller %x: denied by seccomp", sysno, t.Arch().Value(caller))
+			return (*runApp)(nil)
+		case seccompResultAllow:
+			// ok
+		case seccompResultTrace:
+			t.Debugf("vsyscall %d, caller %x: stopping for PTRACE_EVENT_SECCOMP", sysno, t.Arch().Value(caller))
+			return &runVsyscallAfterPtraceEventSeccomp{addr, sysno, caller}
+		default:
+			panic(fmt.Sprintf("Unknown seccomp result %d", r))
+		}
+	}
+
+	return t.doVsyscallInvoke(sysno, args, caller)
+}
+
+type runVsyscallAfterPtraceEventSeccomp struct {
+	addr   usermem.Addr
+	sysno  uintptr
+	caller interface{}
+}
+
+func (r *runVsyscallAfterPtraceEventSeccomp) execute(t *Task) taskRunState {
+	if t.killed() {
+		return (*runInterrupt)(nil)
+	}
+	sysno := t.Arch().SyscallNo()
+	// "... the syscall may not be changed to another system call using the
+	// orig_rax register. It may only be changed to -1 order [sic] to skip the
+	// currently emulated call. ... The tracer MUST NOT modify rip or rsp." -
+	// Documentation/prctl/seccomp_filter.txt. On Linux, changing orig_ax or ip
+	// causes do_exit(SIGSYS), and changing sp is ignored.
+	if (sysno != ^uintptr(0) && sysno != r.sysno) || usermem.Addr(t.Arch().IP()) != r.addr {
+		t.PrepareExit(ExitStatus{Signo: int(linux.SIGSYS)})
+		return (*runExit)(nil)
+	}
+	if sysno == ^uintptr(0) {
+		return (*runApp)(nil)
+	}
+	return t.doVsyscallInvoke(sysno, t.Arch().SyscallArgs(), r.caller)
+}
+
+func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, caller interface{}) taskRunState {
+	rval, ctrl, err := t.executeSyscall(sysno, args)
+	if ctrl != nil {
+		t.Debugf("vsyscall %d, caller %x: syscall control: %v", sysno, t.Arch().Value(caller), ctrl)
+		// Set the return value. The stack has already been adjusted.
+		t.Arch().SetReturn(0)
+	} else if err == nil {
+		t.Debugf("vsyscall %d, caller %x: successfully emulated syscall", sysno, t.Arch().Value(caller))
+		// Set the return value. The stack has already been adjusted.
+		t.Arch().SetReturn(uintptr(rval))
+	} else {
+		t.Debugf("vsyscall %d, caller %x: emulated syscall returned error: %v", sysno, t.Arch().Value(caller), err)
+		if err == syserror.EFAULT {
+			t.forceSignal(linux.SIGSEGV, false /* unconditional */)
+			t.SendSignal(sigPriv(linux.SIGSEGV))
+			// A return is not emulated in this case.
+			return (*runApp)(nil)
+		}
+		t.Arch().SetReturn(uintptr(-t.ExtractErrno(err, int(sysno))))
+	}
+	t.Arch().SetIP(t.Arch().Value(caller))
+	t.Arch().SetStack(t.Arch().Stack() + uintptr(t.Arch().Width()))
+	return (*runApp)(nil)
+}
+
+// ExtractErrno extracts an integer error number from the error.
+// The syscall number is purely for context in the error case. Use -1 if
+// syscall number is unknown.
+func (t *Task) ExtractErrno(err error, sysno int) int {
+	switch err := err.(type) {
+	case nil:
+		return 0
+	case syscall.Errno:
+		return int(err)
+	case SyscallRestartErrno:
+		return int(err)
+	case *memmap.BusError:
+		// Bus errors may generate SIGBUS, but for syscalls they still
+		// return EFAULT. See case in task_run.go where the fault is
+		// handled (and the SIGBUS is delivered).
+		return int(syscall.EFAULT)
+	case *os.PathError:
+		return t.ExtractErrno(err.Err, sysno)
+	case *os.LinkError:
+		return t.ExtractErrno(err.Err, sysno)
+	case *os.SyscallError:
+		return t.ExtractErrno(err.Err, sysno)
+	default:
+		if errno, ok := syserror.TranslateError(err); ok {
+			return int(errno)
+		}
+	}
+	panic(fmt.Sprintf("Unknown syscall %d error: %v", sysno, err))
+}
diff --git a/pkg/sentry/kernel/task_test.go b/pkg/sentry/kernel/task_test.go
new file mode 100644
index 000000000..82ef858a1
--- /dev/null
+++ b/pkg/sentry/kernel/task_test.go
@@ -0,0 +1,69 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+)
+
+func TestTaskCPU(t *testing.T) {
+	for _, test := range []struct {
+		mask sched.CPUSet
+		tid  ThreadID
+		cpu  int32
+	}{
+		{
+			mask: []byte{0xff},
+			tid:  1,
+			cpu:  0,
+		},
+		{
+			mask: []byte{0xff},
+			tid:  10,
+			cpu:  1,
+		},
+		{
+			// more than 8 cpus.
+			mask: []byte{0xff, 0xff},
+			tid:  10,
+			cpu:  9,
+		},
+		{
+			// missing the first cpu.
+			mask: []byte{0xfe},
+			tid:  1,
+			cpu:  1,
+		},
+		{
+			mask: []byte{0xfe},
+			tid:  10,
+			cpu:  3,
+		},
+		{
+			// missing the fifth cpu.
+			mask: []byte{0xef},
+			tid:  10,
+			cpu:  2,
+		},
+	} {
+		assigned := assignCPU(test.mask, test.tid)
+		if test.cpu != assigned {
+			t.Errorf("assignCPU(%v, %v) got %v, want %v", test.mask, test.tid, assigned, test.cpu)
+		}
+	}
+
+}
diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go
new file mode 100644
index 000000000..7a62ab674
--- /dev/null
+++ b/pkg/sentry/kernel/task_usermem.go
@@ -0,0 +1,298 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// _MAX_RW_COUNT is the maximum size in bytes of a single read or write.
+// Reads and writes that exceed this size may be silently truncated.
+// (Linux: include/linux/fs.h:MAX_RW_COUNT)
+var _MAX_RW_COUNT = int(usermem.Addr(math.MaxInt32).RoundDown())
+
+// Activate ensures that the task has an active address space.
+func (t *Task) Activate() {
+	if mm := t.MemoryManager(); mm != nil {
+		if err := mm.Activate(); err != nil {
+			panic("unable to activate mm: " + err.Error())
+		}
+	}
+}
+
+// Deactivate relinquishes the task's active address space.
+func (t *Task) Deactivate() {
+	if mm := t.MemoryManager(); mm != nil {
+		if err := mm.Deactivate(); err != nil {
+			panic("unable to deactivate mm: " + err.Error())
+		}
+	}
+}
+
+// CopyIn copies a fixed-size value or slice of fixed-size values in from the
+// task's memory. The copy will fail with syscall.EFAULT if it traverses user
+// memory that is unmapped or not readable by the user.
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyIn(addr usermem.Addr, dst interface{}) (int, error) {
+	return usermem.CopyObjectIn(t, t.MemoryManager(), addr, dst, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+}
+
+// CopyInBytes is a fast version of CopyIn if the caller can serialize the
+// data without reflection and pass in a byte slice.
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyInBytes(addr usermem.Addr, dst []byte) (int, error) {
+	return t.MemoryManager().CopyIn(t, addr, dst, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+}
+
+// CopyOut copies a fixed-size value or slice of fixed-size values out to the
+// task's memory. The copy will fail with syscall.EFAULT if it traverses user
+// memory that is unmapped or not writeable by the user.
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyOut(addr usermem.Addr, src interface{}) (int, error) {
+	return usermem.CopyObjectOut(t, t.MemoryManager(), addr, src, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+}
+
+// CopyOutBytes is a fast version of CopyOut if the caller can serialize the
+// data without reflection and pass in a byte slice.
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyOutBytes(addr usermem.Addr, src []byte) (int, error) {
+	return t.MemoryManager().CopyOut(t, addr, src, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+}
+
+// CopyInString copies a NUL-terminated string of length at most maxlen in from
+// the task's memory. The copy will fail with syscall.EFAULT if it traverses
+// user memory that is unmapped or not readable by the user.
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyInString(addr usermem.Addr, maxlen int) (string, error) {
+	return usermem.CopyStringIn(t, t.MemoryManager(), addr, maxlen, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+}
+
+// CopyInVector copies a NULL-terminated vector of strings from the task's
+// memory. The copy will fail with syscall.EFAULT if it traverses
+// user memory that is unmapped or not readable by the user.
+//
+// maxElemSize is the maximum size of each individual element.
+//
+// maxTotalSize is the maximum total length of all elements plus the total
+// number of elements. For example, the following strings correspond to
+// the following set of sizes:
+//
+//     { "a", "b", "c" } => 6 (3 for lengths, 3 for elements)
+//     { "abc" }         => 4 (3 for length, 1 for elements)
+//
+// This Task's AddressSpace must be active.
+func (t *Task) CopyInVector(addr usermem.Addr, maxElemSize, maxTotalSize int) ([]string, error) {
+	var v []string
+	for {
+		argAddr := t.Arch().Native(0)
+		if _, err := t.CopyIn(addr, argAddr); err != nil {
+			return v, err
+		}
+		if t.Arch().Value(argAddr) == 0 {
+			break
+		}
+		// Each string has a zero terminating byte counted, so copying out a string
+		// requires at least one byte of space. Also, see the calculation below.
+		if maxTotalSize <= 0 {
+			return nil, syserror.ENOMEM
+		}
+		thisMax := maxElemSize
+		if maxTotalSize < thisMax {
+			thisMax = maxTotalSize
+		}
+		arg, err := t.CopyInString(usermem.Addr(t.Arch().Value(argAddr)), thisMax)
+		if err != nil {
+			return v, err
+		}
+		v = append(v, arg)
+		addr += usermem.Addr(t.Arch().Width())
+		maxTotalSize -= len(arg) + 1
+	}
+	return v, nil
+}
+
+// CopyOutIovecs converts src to an array of struct iovecs and copies it to the
+// memory mapped at addr.
+//
+// Preconditions: As for usermem.IO.CopyOut. The caller must be running on the
+// task goroutine. t's AddressSpace must be active.
+func (t *Task) CopyOutIovecs(addr usermem.Addr, src usermem.AddrRangeSeq) error {
+	switch t.Arch().Width() {
+	case 8:
+		const itemLen = 16
+		if _, ok := addr.AddLength(uint64(src.NumRanges()) * itemLen); !ok {
+			return syserror.EFAULT
+		}
+
+		b := t.CopyScratchBuffer(itemLen)
+		for ; !src.IsEmpty(); src = src.Tail() {
+			ar := src.Head()
+			usermem.ByteOrder.PutUint64(b[0:8], uint64(ar.Start))
+			usermem.ByteOrder.PutUint64(b[8:16], uint64(ar.Length()))
+			if _, err := t.CopyOutBytes(addr, b); err != nil {
+				return err
+			}
+			addr += itemLen
+		}
+
+	default:
+		return syserror.ENOSYS
+	}
+
+	return nil
+}
+
+// CopyInIovecs copies an array of numIovecs struct iovecs from the memory
+// mapped at addr, converts them to usermem.AddrRanges, and returns them as a
+// usermem.AddrRangeSeq.
+//
+// CopyInIovecs shares the following properties with Linux's
+// lib/iov_iter.c:import_iovec() => fs/read_write.c:rw_copy_check_uvector():
+//
+// - If the length of any AddrRange would exceed the range of an ssize_t,
+// CopyInIovecs returns EINVAL.
+//
+// - If the length of any AddrRange would cause its end to overflow,
+// CopyInIovecs returns EFAULT.
+//
+// - The combined length of all AddrRanges is limited to _MAX_RW_COUNT. If the
+// combined length of all AddrRanges would otherwise exceed this amount, ranges
+// beyond _MAX_RW_COUNT are silently truncated.
+//
+// Preconditions: As for usermem.IO.CopyIn. The caller must be running on the
+// task goroutine. t's AddressSpace must be active.
+func (t *Task) CopyInIovecs(addr usermem.Addr, numIovecs int) (usermem.AddrRangeSeq, error) {
+	if numIovecs == 0 {
+		return usermem.AddrRangeSeq{}, nil
+	}
+
+	var dst []usermem.AddrRange
+	if numIovecs > 1 {
+		dst = make([]usermem.AddrRange, 0, numIovecs)
+	}
+
+	switch t.Arch().Width() {
+	case 8:
+		const itemLen = 16
+		if _, ok := addr.AddLength(uint64(numIovecs) * itemLen); !ok {
+			return usermem.AddrRangeSeq{}, syserror.EFAULT
+		}
+
+		b := t.CopyScratchBuffer(itemLen)
+		for i := 0; i < numIovecs; i++ {
+			if _, err := t.CopyInBytes(addr, b); err != nil {
+				return usermem.AddrRangeSeq{}, err
+			}
+
+			base := usermem.Addr(usermem.ByteOrder.Uint64(b[0:8]))
+			length := usermem.ByteOrder.Uint64(b[8:16])
+			if length > math.MaxInt64 {
+				return usermem.AddrRangeSeq{}, syserror.EINVAL
+			}
+			ar, ok := base.ToRange(length)
+			if !ok {
+				return usermem.AddrRangeSeq{}, syserror.EFAULT
+			}
+
+			if numIovecs == 1 {
+				// Special case to avoid allocating dst.
+				return usermem.AddrRangeSeqOf(ar).TakeFirst(_MAX_RW_COUNT), nil
+			}
+			dst = append(dst, ar)
+
+			addr += itemLen
+		}
+
+	default:
+		return usermem.AddrRangeSeq{}, syserror.ENOSYS
+	}
+
+	// Truncate to _MAX_RW_COUNT.
+	var total uint64
+	for i := range dst {
+		dstlen := uint64(dst[i].Length())
+		if rem := uint64(_MAX_RW_COUNT) - total; rem < dstlen {
+			dst[i].End -= usermem.Addr(dstlen - rem)
+			dstlen = rem
+		}
+		total += dstlen
+	}
+
+	return usermem.AddrRangeSeqFromSlice(dst), nil
+}
+
+// SingleIOSequence returns a usermem.IOSequence representing [addr,
+// addr+length) in t's address space. If length exceeds _MAX_RW_COUNT, it is
+// silently truncated.
+//
+// SingleIOSequence is analogous to Linux's
+// lib/iov_iter.c:import_single_range(). (Note that the non-vectorized read and
+// write syscalls in Linux do not use import_single_range(), but are still
+// truncated to _MAX_RW_COUNT by fs/read_write.c:rw_verify_area().)
+func (t *Task) SingleIOSequence(addr usermem.Addr, length int, opts usermem.IOOpts) (usermem.IOSequence, error) {
+	if length > _MAX_RW_COUNT {
+		length = _MAX_RW_COUNT
+	}
+	ar, ok := addr.ToRange(uint64(length))
+	if !ok {
+		return usermem.IOSequence{}, syserror.EFAULT
+	}
+	return usermem.IOSequence{
+		IO:    t.MemoryManager(),
+		Addrs: usermem.AddrRangeSeqOf(ar),
+		Opts:  opts,
+	}, nil
+}
+
+// IovecsIOSequence returns a usermem.IOSequence representing the array of
+// iovcnt struct iovecs at addr in t's address space. opts applies to the
+// returned IOSequence, not the reading of the struct iovec array.
+//
+// IovecsIOSequence is analogous to Linux's lib/iov_iter.c:import_iovec().
+//
+// Preconditions: As for Task.CopyInIovecs.
+func (t *Task) IovecsIOSequence(addr usermem.Addr, iovcnt int, opts usermem.IOOpts) (usermem.IOSequence, error) {
+	if iovcnt < 0 || iovcnt > linux.UIO_MAXIOV {
+		return usermem.IOSequence{}, syserror.EINVAL
+	}
+	ars, err := t.CopyInIovecs(addr, iovcnt)
+	if err != nil {
+		return usermem.IOSequence{}, err
+	}
+	return usermem.IOSequence{
+		IO:    t.MemoryManager(),
+		Addrs: ars,
+		Opts:  opts,
+	}, nil
+}
diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go
new file mode 100644
index 000000000..8fffd3446
--- /dev/null
+++ b/pkg/sentry/kernel/thread_group.go
@@ -0,0 +1,269 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+)
+
+// A ThreadGroup is a logical grouping of tasks that has widespread
+// significance to other kernel features (e.g. signal handling). ("Thread
+// groups" are usually called "processes" in userspace documentation.)
+//
+// ThreadGroup is a superset of Linux's struct signal_struct.
+type ThreadGroup struct {
+	threadGroupNode
+
+	// signalHandlers is the set of signal handlers used by every task in this
+	// thread group. (signalHandlers may also be shared with other thread
+	// groups.)
+	//
+	// signalHandlers.mu (hereafter "the signal mutex") protects state related
+	// to signal handling, as well as state that usually needs to be atomic
+	// with signal handling, for all ThreadGroups and Tasks using
+	// signalHandlers. (This is analogous to Linux's use of struct
+	// sighand_struct::siglock.)
+	//
+	// The signalHandlers pointer can only be mutated during an execve
+	// (Task.finishExec). Consequently, when it's possible for a task in the
+	// thread group to be completing an execve, signalHandlers is protected by
+	// the owning TaskSet.mu. Otherwise, it is possible to read the
+	// signalHandlers pointer without synchronization. In particular,
+	// completing an execve requires that all other tasks in the thread group
+	// have exited, so task goroutines do not need the owning TaskSet.mu to
+	// read the signalHandlers pointer of their thread groups.
+	signalHandlers *SignalHandlers
+
+	// pendingSignals is the set of pending signals that may be handled by any
+	// task in this thread group.
+	//
+	// pendingSignals is protected by the signal mutex.
+	pendingSignals pendingSignals
+
+	// lastTimerSignalTask records the last task we deliver a process timer signal to.
+	// Please see SendTimerSignal for more details.
+	//
+	// lastTimerSignalTask is protected by the signal mutex.
+	lastTimerSignalTask *Task
+
+	// groupStopPhase indicates the state of a group stop in progress on the
+	// thread group, if any.
+	//
+	// groupStopPhase is protected by the signal mutex.
+	groupStopPhase groupStopPhase
+
+	// groupStopSignal is the signal that caused a group stop to be initiated.
+	// groupStopSignal is only meaningful if groupStopPhase is
+	// groupStopInitiated or groupStopComplete.
+	//
+	// groupStopSignal is protected by the signal mutex.
+	groupStopSignal linux.Signal
+
+	// groupStopCount is the number of non-exited tasks in the thread group
+	// that have acknowledged an initiated group stop. groupStopCount is only
+	// meaningful if groupStopPhase is groupStopInitiated.
+	//
+	// groupStopCount is protected by the signal mutex.
+	groupStopCount int
+
+	// If groupStopWaitable is true, the thread group is indicating a waitable
+	// group stop event (as defined by EventChildGroupStop).
+	//
+	// Linux represents the analogous state as SIGNAL_STOP_STOPPED being set
+	// and group_exit_code being non-zero.
+	//
+	// groupStopWaitable is protected by the signal mutex.
+	groupStopWaitable bool
+
+	// If groupContNotify is true, then a SIGCONT has recently ended a group
+	// stop on this thread group, and the first task to observe it should
+	// notify its parent.
+	//
+	// groupContNotify is protected by the signal mutex.
+	groupContNotify bool
+
+	// If groupContNotify is true, groupContInterrupted is true iff SIGCONT
+	// ended a group stop in phase groupStopInitiated. If groupContNotify is
+	// false, groupContInterrupted is meaningless.
+	//
+	// Analogues in Linux:
+	//
+	// - groupContNotify && groupContInterrupted is represented by
+	// SIGNAL_CLD_STOPPED.
+	//
+	// - groupContNotify && !groupContInterrupted is represented by
+	// SIGNAL_CLD_CONTINUED.
+	//
+	// - !groupContNotify is represented by neither flag being set.
+	//
+	// groupContInterrupted is protected by the signal mutex.
+	groupContInterrupted bool
+
+	// If groupContWaitable is true, the thread group is indicating a waitable
+	// continue event (as defined by EventGroupContinue).
+	//
+	// groupContWaitable is analogous to Linux's SIGNAL_STOP_CONTINUED.
+	//
+	// groupContWaitable is protected by the signal mutex.
+	groupContWaitable bool
+
+	// exiting is true if all tasks in the ThreadGroup should exit. exiting is
+	// analogous to Linux's SIGNAL_GROUP_EXIT.
+	//
+	// exiting is protected by the signal mutex. exiting can only transition
+	// from false to true.
+	exiting bool
+
+	// exitStatus is the thread group's exit status.
+	//
+	// While exiting is false, exitStatus is protected by the signal mutex.
+	// When exiting becomes true, exitStatus becomes immutable.
+	exitStatus ExitStatus
+
+	// terminationSignal is the signal that this thread group's leader will
+	// send to its parent when it exits.
+	//
+	// terminationSignal is protected by the TaskSet mutex.
+	terminationSignal linux.Signal
+
+	// liveGoroutines is the number of non-exited task goroutines in the thread
+	// group.
+	//
+	// liveGoroutines is not saved; it is reset as task goroutines are
+	// restarted by Task.Start.
+	liveGoroutines sync.WaitGroup `state:"nosave"`
+
+	// tm contains process timers. TimerManager fields are immutable.
+	tm TimerManager
+
+	// exitedCPUStats is the CPU usage for all exited tasks in the thread
+	// group. exitedCPUStats is protected by the TaskSet mutex.
+	exitedCPUStats usage.CPUStats
+
+	// childCPUStats is the CPU usage of all joined descendants of this thread
+	// group. childCPUStats is protected by the TaskSet mutex.
+	childCPUStats usage.CPUStats
+
+	// ioUsage is the I/O usage for all exited tasks in the thread group.
+	// The ioUsage pointer is immutable.
+	ioUsage *usage.IO
+
+	// maxRSS is the historical maximum resident set size of the thread group, updated when:
+	//
+	// - A task in the thread group exits, since after all tasks have
+	// exited the MemoryManager is no longer reachable.
+	//
+	// - The thread group completes an execve, since this changes
+	// MemoryManagers.
+	//
+	// maxRSS is protected by the TaskSet mutex.
+	maxRSS uint64
+
+	// childMaxRSS is the maximum resident set size in bytes of all joined
+	// descendants of this thread group.
+	//
+	// childMaxRSS is protected by the TaskSet mutex.
+	childMaxRSS uint64
+
+	// Resource limits for this ThreadGroup. The limits pointer is immutable.
+	limits *limits.LimitSet
+
+	// processGroup is the processGroup for this thread group.
+	//
+	// processGroup is protected by the TaskSet mutex.
+	processGroup *ProcessGroup
+
+	// execed indicates an exec has occurred since creation. This will be
+	// set by finishExec, and new TheadGroups will have this field cleared.
+	// When execed is set, the processGroup may no longer be changed.
+	//
+	// execed is protected by the TaskSet mutex.
+	execed bool
+
+	// rscr is the thread group's RSEQ critical region.
+	rscr atomic.Value `state:".(*RSEQCriticalRegion)"`
+}
+
+// NewThreadGroup returns a new, empty thread group in PID namespace ns. The
+// thread group leader will send its parent terminationSignal when it exits.
+// The new thread group isn't visible to the system until a task has been
+// created inside of it by a successful call to TaskSet.NewTask.
+func NewThreadGroup(ns *PIDNamespace, sh *SignalHandlers, terminationSignal linux.Signal, limits *limits.LimitSet, monotonicClock *timekeeperClock) *ThreadGroup {
+	tg := &ThreadGroup{
+		threadGroupNode: threadGroupNode{
+			pidns: ns,
+		},
+		signalHandlers:    sh,
+		terminationSignal: terminationSignal,
+		ioUsage:           &usage.IO{},
+		limits:            limits,
+	}
+	tg.tm = newTimerManager(tg, monotonicClock)
+	tg.rscr.Store(&RSEQCriticalRegion{})
+	return tg
+}
+
+// saveRscr is invopked by stateify.
+func (tg *ThreadGroup) saveRscr() *RSEQCriticalRegion {
+	return tg.rscr.Load().(*RSEQCriticalRegion)
+}
+
+// loadRscr is invoked by stateify.
+func (tg *ThreadGroup) loadRscr(rscr *RSEQCriticalRegion) {
+	tg.rscr.Store(rscr)
+}
+
+// SignalHandlers returns the signal handlers used by tg.
+//
+// Preconditions: The caller must provide the synchronization required to read
+// tg.signalHandlers, as described in the field's comment.
+func (tg *ThreadGroup) SignalHandlers() *SignalHandlers {
+	return tg.signalHandlers
+}
+
+// Timer returns tg's timers.
+func (tg *ThreadGroup) Timer() *TimerManager {
+	return &tg.tm
+}
+
+// Limits returns tg's limits.
+func (tg *ThreadGroup) Limits() *limits.LimitSet {
+	return tg.limits
+}
+
+// release releases the thread group's resources.
+func (tg *ThreadGroup) release() {
+	// This must be done without holding the TaskSet mutex since thread group
+	// timers call SendSignal with Timer.mu locked.
+	tg.tm.destroy()
+}
+
+// forEachChildThreadGroupLocked indicates over all child ThreadGroups.
+//
+// Precondition: TaskSet.mu must be held.
+func (tg *ThreadGroup) forEachChildThreadGroupLocked(fn func(*ThreadGroup)) {
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		for child := range t.children {
+			if child == child.tg.leader {
+				fn(child.tg)
+			}
+		}
+	}
+}
diff --git a/pkg/sentry/kernel/threads.go b/pkg/sentry/kernel/threads.go
new file mode 100644
index 000000000..440da9dad
--- /dev/null
+++ b/pkg/sentry/kernel/threads.go
@@ -0,0 +1,443 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// TasksLimit is the maximum number of threads for untrusted application.
+// Linux doesn't really limit this directly, rather it is limited by total
+// memory size, stacks allocated and a global maximum. There's no real reason
+// for us to limit it either, (esp. since threads are backed by go routines),
+// and we would expect to hit resource limits long before hitting this number.
+// However, for correctness, we still check that the user doesn't exceed this
+// number.
+//
+// Note that because of the way futexes are implemented, there *are* in fact
+// serious restrictions on valid thread IDs. They are limited to 2^30 - 1
+// (kernel/fork.c:MAX_THREADS).
+const TasksLimit = (1 << 16)
+
+// ThreadID is a generic thread identifier.
+type ThreadID int32
+
+// String returns a decimal representation of the ThreadID.
+func (tid ThreadID) String() string {
+	return fmt.Sprintf("%d", tid)
+}
+
+// InitTID is the TID given to the first task added to each PID namespace. The
+// thread group led by InitTID is called the namespace's init process. The
+// death of a PID namespace's init process causes all tasks visible in that
+// namespace to be killed.
+const InitTID ThreadID = 1
+
+// A TaskSet comprises all tasks in a system.
+type TaskSet struct {
+	// mu protects all relationships betweens tasks and thread groups in the
+	// TaskSet. (mu is approximately equivalent to Linux's tasklist_lock.)
+	mu sync.RWMutex `state:"nosave"`
+
+	// Root is the root PID namespace, in which all tasks in the TaskSet are
+	// visible. The Root pointer is immutable.
+	Root *PIDNamespace
+
+	// sessions is the set of all sessions.
+	sessions sessionList
+
+	// stopCount is the number of active external stops applicable to all tasks
+	// in the TaskSet (calls to TaskSet.BeginExternalStop that have not been
+	// paired with a call to TaskSet.EndExternalStop). stopCount is protected
+	// by mu.
+	//
+	// stopCount is not saved for the same reason as Task.stopCount; it is
+	// always reset to zero after restore.
+	stopCount int32 `state:"nosave"`
+
+	// liveGoroutines is the number of non-exited task goroutines in the
+	// TaskSet.
+	//
+	// liveGoroutines is not saved; it is reset as task goroutines are
+	// restarted by Task.Start.
+	liveGoroutines sync.WaitGroup `state:"nosave"`
+
+	// runningGoroutines is the number of running task goroutines in the
+	// TaskSet.
+	//
+	// runningGoroutines is not saved; its counter value is required to be zero
+	// at time of save (but note that this is not necessarily the same thing as
+	// sync.WaitGroup's zero value).
+	runningGoroutines sync.WaitGroup `state:"nosave"`
+}
+
+// newTaskSet returns a new, empty TaskSet.
+func newTaskSet() *TaskSet {
+	ts := &TaskSet{}
+	ts.Root = newPIDNamespace(ts, nil /* parent */, auth.NewRootUserNamespace())
+	return ts
+}
+
+// forEachThreadGroupLocked applies f to each thread group in ts.
+//
+// Preconditions: ts.mu must be locked (for reading or writing).
+func (ts *TaskSet) forEachThreadGroupLocked(f func(tg *ThreadGroup)) {
+	for t := range ts.Root.tids {
+		if t == t.tg.leader {
+			f(t.tg)
+		}
+	}
+}
+
+// A PIDNamespace represents a PID namespace, a bimap between thread IDs and
+// tasks. See the pid_namespaces(7) man page for further details.
+//
+// N.B. A task is said to be visible in a PID namespace if the PID namespace
+// contains a thread ID that maps to that task.
+type PIDNamespace struct {
+	// owner is the TaskSet that this PID namespace belongs to. The owner
+	// pointer is immutable.
+	owner *TaskSet
+
+	// parent is the PID namespace of the process that created this one. If
+	// this is the root PID namespace, parent is nil. The parent pointer is
+	// immutable.
+	//
+	// Invariant: All tasks that are visible in this namespace are also visible
+	// in all ancestor namespaces.
+	parent *PIDNamespace
+
+	// userns is the user namespace with which this PID namespace is
+	// associated. Privileged operations on this PID namespace must have
+	// appropriate capabilities in userns. The userns pointer is immutable.
+	userns *auth.UserNamespace
+
+	// The following fields are protected by owner.mu.
+
+	// last is the last ThreadID to be allocated in this namespace.
+	last ThreadID
+
+	// tasks is a mapping from ThreadIDs in this namespace to tasks visible in
+	// the namespace.
+	tasks map[ThreadID]*Task
+
+	// tids is a mapping from tasks visible in this namespace to their
+	// identifiers in this namespace.
+	tids map[*Task]ThreadID
+
+	// sessions is a mapping from SessionIDs in this namespace to sessions
+	// visible in the namespace.
+	sessions map[SessionID]*Session
+
+	// sids is a mapping from sessions visible in this namespace to their
+	// identifiers in this namespace.
+	sids map[*Session]SessionID
+
+	// processGroups is a mapping from ProcessGroupIDs in this namespace to
+	// process groups visible in the namespace.
+	processGroups map[ProcessGroupID]*ProcessGroup
+
+	// pgids is a mapping from process groups visible in this namespace to
+	// their identifiers in this namespace.
+	pgids map[*ProcessGroup]ProcessGroupID
+
+	// exiting indicates that the namespace's init process is exiting or has
+	// exited.
+	exiting bool
+}
+
+func newPIDNamespace(ts *TaskSet, parent *PIDNamespace, userns *auth.UserNamespace) *PIDNamespace {
+	return &PIDNamespace{
+		owner:         ts,
+		parent:        parent,
+		userns:        userns,
+		tasks:         make(map[ThreadID]*Task),
+		tids:          make(map[*Task]ThreadID),
+		sessions:      make(map[SessionID]*Session),
+		sids:          make(map[*Session]SessionID),
+		processGroups: make(map[ProcessGroupID]*ProcessGroup),
+		pgids:         make(map[*ProcessGroup]ProcessGroupID),
+	}
+}
+
+// NewChild returns a new, empty PID namespace that is a child of ns. Authority
+// over the new PID namespace is controlled by userns.
+func (ns *PIDNamespace) NewChild(userns *auth.UserNamespace) *PIDNamespace {
+	return newPIDNamespace(ns.owner, ns, userns)
+}
+
+// TaskWithID returns the task with thread ID tid in PID namespace ns. If no
+// task has that TID, TaskWithID returns nil.
+func (ns *PIDNamespace) TaskWithID(tid ThreadID) *Task {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	return ns.tasks[tid]
+}
+
+// ThreadGroupWithID returns the thread group lead by the task with thread ID
+// tid in PID namespace ns. If no task has that TID, or if the task with that
+// TID is not a thread group leader, ThreadGroupWithID returns nil.
+func (ns *PIDNamespace) ThreadGroupWithID(tid ThreadID) *ThreadGroup {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	t := ns.tasks[tid]
+	if t == nil {
+		return nil
+	}
+	if t != t.tg.leader {
+		return nil
+	}
+	return t.tg
+}
+
+// IDOfTask returns the TID assigned to the given task in PID namespace ns. If
+// the task is not visible in that namespace, IDOfTask returns 0. (This return
+// value is significant in some cases, e.g. getppid() is documented as
+// returning 0 if the caller's parent is in an ancestor namespace and
+// consequently not visible to the caller.) If the task is nil, IDOfTask returns
+// 0.
+func (ns *PIDNamespace) IDOfTask(t *Task) ThreadID {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	return ns.tids[t]
+}
+
+// IDOfThreadGroup returns the TID assigned to tg's leader in PID namespace ns.
+// If the task is not visible in that namespace, IDOfThreadGroup returns 0.
+func (ns *PIDNamespace) IDOfThreadGroup(tg *ThreadGroup) ThreadID {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	return ns.tids[tg.leader]
+}
+
+// Tasks returns a snapshot of the tasks in ns.
+func (ns *PIDNamespace) Tasks() []*Task {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	tasks := make([]*Task, 0, len(ns.tasks))
+	for t := range ns.tids {
+		tasks = append(tasks, t)
+	}
+	return tasks
+}
+
+// ThreadGroups returns a snapshot of the thread groups in ns.
+func (ns *PIDNamespace) ThreadGroups() []*ThreadGroup {
+	ns.owner.mu.RLock()
+	defer ns.owner.mu.RUnlock()
+	var tgs []*ThreadGroup
+	for t := range ns.tids {
+		if t == t.tg.leader {
+			tgs = append(tgs, t.tg)
+		}
+	}
+	return tgs
+}
+
+// UserNamespace returns the user namespace associated with PID namespace ns.
+func (ns *PIDNamespace) UserNamespace() *auth.UserNamespace {
+	return ns.userns
+}
+
+// A threadGroupNode defines the relationship between a thread group and the
+// rest of the system. Conceptually, threadGroupNode is data belonging to the
+// owning TaskSet, as if TaskSet contained a field `nodes
+// map[*ThreadGroup]*threadGroupNode`. However, for practical reasons,
+// threadGroupNode is embedded in the ThreadGroup it represents.
+// (threadGroupNode is an anonymous field in ThreadGroup; this is to expose
+// threadGroupEntry's methods on ThreadGroup to make it implement
+// threadGroupLinker.)
+type threadGroupNode struct {
+	// pidns is the PID namespace containing the thread group and all of its
+	// member tasks. The pidns pointer is immutable.
+	pidns *PIDNamespace
+
+	// eventQueue is notified whenever a event of interest to Task.Wait occurs
+	// in a child of this thread group, or a ptrace tracee of a task in this
+	// thread group. Events are defined in task_exit.go.
+	//
+	// Note that we cannot check and save this wait queue similarly to other
+	// wait queues, as the queue will not be empty by the time of saving, due
+	// to the wait sourced from Exec().
+	eventQueue waiter.Queue `state:"nosave"`
+
+	// leader is the thread group's leader, which is the oldest task in the
+	// thread group; usually the last task in the thread group to call
+	// execve(), or if no such task exists then the first task in the thread
+	// group, which was created by a call to fork() or clone() without
+	// CLONE_THREAD. Once a thread group has been made visible to the rest of
+	// the system by TaskSet.newTask, leader is never nil.
+	//
+	// Note that it's possible for the leader to exit without causing the rest
+	// of the thread group to exit; in such a case, leader will still be valid
+	// and non-nil, but leader will not be in tasks.
+	//
+	// leader is protected by the TaskSet mutex.
+	leader *Task
+
+	// If execing is not nil, it is a task in the thread group that has killed
+	// all other tasks so that it can become the thread group leader and
+	// perform an execve. (execing may already be the thread group leader.)
+	//
+	// execing is analogous to Linux's signal_struct::group_exit_task.
+	//
+	// execing is protected by the TaskSet mutex.
+	execing *Task
+
+	// tasks is all tasks in the thread group that have not yet been reaped.
+	//
+	// tasks is protected by both the TaskSet mutex and the signal mutex:
+	// Mutating tasks requires locking the TaskSet mutex for writing *and*
+	// locking the signal mutex. Reading tasks requires locking the TaskSet
+	// mutex *or* locking the signal mutex.
+	tasks taskList
+
+	// tasksCount is the number of tasks in the thread group that have not yet
+	// been reaped; equivalently, tasksCount is the number of tasks in tasks.
+	//
+	// tasksCount is protected by both the TaskSet mutex and the signal mutex,
+	// as with tasks.
+	tasksCount int
+
+	// liveTasks is the number of tasks in the thread group that have not yet
+	// reached TaskExitZombie.
+	//
+	// liveTasks is protected by the TaskSet mutex (NOT the signal mutex).
+	liveTasks int
+
+	// activeTasks is the number of tasks in the thread group that have not yet
+	// reached TaskExitInitiated.
+	//
+	// activeTasks is protected by both the TaskSet mutex and the signal mutex,
+	// as with tasks.
+	activeTasks int
+}
+
+// PIDNamespace returns the PID namespace containing tg.
+func (tg *ThreadGroup) PIDNamespace() *PIDNamespace {
+	return tg.pidns
+}
+
+// TaskSet returns the TaskSet containing tg.
+func (tg *ThreadGroup) TaskSet() *TaskSet {
+	return tg.pidns.owner
+}
+
+// Leader returns tg's leader.
+func (tg *ThreadGroup) Leader() *Task {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.leader
+}
+
+// Count returns the number of non-exited threads in the group.
+func (tg *ThreadGroup) Count() int {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	var count int
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		count++
+	}
+	return count
+}
+
+// MemberIDs returns a snapshot of the ThreadIDs (in PID namespace pidns) for
+// all tasks in tg.
+func (tg *ThreadGroup) MemberIDs(pidns *PIDNamespace) []ThreadID {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+
+	var tasks []ThreadID
+	for t := tg.tasks.Front(); t != nil; t = t.Next() {
+		if id, ok := pidns.tids[t]; ok {
+			tasks = append(tasks, id)
+		}
+	}
+	return tasks
+}
+
+// ID returns tg's leader's thread ID in its own PID namespace. If tg's leader
+// is dead, ID returns 0.
+func (tg *ThreadGroup) ID() ThreadID {
+	tg.pidns.owner.mu.RLock()
+	defer tg.pidns.owner.mu.RUnlock()
+	return tg.pidns.tids[tg.leader]
+}
+
+// A taskNode defines the relationship between a task and the rest of the
+// system. The comments on threadGroupNode also apply to taskNode.
+type taskNode struct {
+	// tg is the thread group that this task belongs to. The tg pointer is
+	// immutable.
+	tg *ThreadGroup `state:"wait"`
+
+	// taskEntry links into tg.tasks. Note that this means that
+	// Task.Next/Prev/SetNext/SetPrev refer to sibling tasks in the same thread
+	// group. See threadGroupNode.tasks for synchronization info.
+	taskEntry
+
+	// parent is the task's parent. parent may be nil.
+	//
+	// parent is protected by the TaskSet mutex.
+	parent *Task
+
+	// children is this task's children.
+	//
+	// children is protected by the TaskSet mutex.
+	children map[*Task]struct{}
+
+	// If childPIDNamespace is not nil, all new tasks created by this task will
+	// be members of childPIDNamespace rather than this one. (As a corollary,
+	// this task becomes unable to create sibling tasks in the same thread
+	// group.)
+	//
+	// childPIDNamespace is exclusive to the task goroutine.
+	childPIDNamespace *PIDNamespace
+}
+
+// ThreadGroup returns the thread group containing t.
+func (t *Task) ThreadGroup() *ThreadGroup {
+	return t.tg
+}
+
+// PIDNamespace returns the PID namespace containing t.
+func (t *Task) PIDNamespace() *PIDNamespace {
+	return t.tg.pidns
+}
+
+// TaskSet returns the TaskSet containing t.
+func (t *Task) TaskSet() *TaskSet {
+	return t.tg.pidns.owner
+}
+
+// Timekeeper returns the system Timekeeper.
+func (t *Task) Timekeeper() *Timekeeper {
+	return t.k.timekeeper
+}
+
+// Parent returns t's parent.
+func (t *Task) Parent() *Task {
+	return t.parent
+}
+
+// ThreadID returns t's thread ID in its own PID namespace. If the task is
+// dead, ThreadID returns 0.
+func (t *Task) ThreadID() ThreadID {
+	return t.tg.pidns.IDOfTask(t)
+}
diff --git a/pkg/sentry/kernel/time/BUILD b/pkg/sentry/kernel/time/BUILD
new file mode 100644
index 000000000..84f31b2dc
--- /dev/null
+++ b/pkg/sentry/kernel/time/BUILD
@@ -0,0 +1,32 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "time_state",
+    srcs = [
+        "time.go",
+    ],
+    out = "time_state.go",
+    package = "time",
+)
+
+go_library(
+    name = "time",
+    srcs = [
+        "context.go",
+        "time.go",
+        "time_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/kernel/time/context.go b/pkg/sentry/kernel/time/context.go
new file mode 100644
index 000000000..ac4dc01d8
--- /dev/null
+++ b/pkg/sentry/kernel/time/context.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the time package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxRealtimeClock is a Context.Value key for the current real time.
+	CtxRealtimeClock contextID = iota
+)
+
+// RealtimeClockFromContext returns the real time clock associated with context
+// ctx.
+func RealtimeClockFromContext(ctx context.Context) Clock {
+	if v := ctx.Value(CtxRealtimeClock); v != nil {
+		return v.(Clock)
+	}
+	return nil
+}
+
+// NowFromContext returns the current real time associated with context ctx.
+func NowFromContext(ctx context.Context) Time {
+	if clk := RealtimeClockFromContext(ctx); clk != nil {
+		return clk.Now()
+	}
+	panic("encountered context without RealtimeClock")
+}
diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go
new file mode 100644
index 000000000..c223c2f19
--- /dev/null
+++ b/pkg/sentry/kernel/time/time.go
@@ -0,0 +1,649 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package time defines the Timer type, which provides a periodic timer that
+// works by sampling a user-provided clock.
+package time
+
+import (
+	"fmt"
+	"math"
+	"sync"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// Events that may be generated by a Clock.
+const (
+	// ClockEventSet occurs when a Clock undergoes a discontinuous change.
+	ClockEventSet waiter.EventMask = 1 << iota
+
+	// ClockEventRateIncrease occurs when the rate at which a Clock advances
+	// increases significantly, such that values returned by previous calls to
+	// Clock.WallTimeUntil may be too large.
+	ClockEventRateIncrease
+)
+
+// Time represents an instant in time with nanosecond precision.
+//
+// Time may represent time with respect to any clock and may not have any
+// meaning in the real world.
+type Time struct {
+	ns int64
+}
+
+var (
+	// MinTime is the zero time instant, the lowest possible time that can
+	// be represented by Time.
+	MinTime = Time{ns: math.MinInt64}
+
+	// MaxTime is the highest possible time that can be represented by
+	// Time.
+	MaxTime = Time{ns: math.MaxInt64}
+
+	// ZeroTime represents the zero time in an unspecified Clock's domain.
+	ZeroTime = Time{ns: 0}
+)
+
+const (
+	// MinDuration is the minimum duration representable by time.Duration.
+	MinDuration = time.Duration(math.MinInt64)
+
+	// MaxDuration is the maximum duration representable by time.Duration.
+	MaxDuration = time.Duration(math.MaxInt64)
+)
+
+// FromNanoseconds returns a Time representing the point ns nanoseconds after
+// an unspecified Clock's zero time.
+func FromNanoseconds(ns int64) Time {
+	return Time{ns}
+}
+
+// FromSeconds returns a Time representing the point s seconds after an
+// unspecified Clock's zero time.
+func FromSeconds(s int64) Time {
+	if s > math.MaxInt64/time.Second.Nanoseconds() {
+		return MaxTime
+	}
+	return Time{s * 1e9}
+}
+
+// FromUnix converts from Unix seconds and nanoseconds to Time, assuming a real
+// time Unix clock domain.
+func FromUnix(s int64, ns int64) Time {
+	if s > math.MaxInt64/time.Second.Nanoseconds() {
+		return MaxTime
+	}
+	t := s * 1e9
+	if t > math.MaxInt64-ns {
+		return MaxTime
+	}
+	return Time{t + ns}
+}
+
+// FromTimespec converts from Linux Timespec to Time.
+func FromTimespec(ts linux.Timespec) Time {
+	return Time{ts.ToNsecCapped()}
+}
+
+// FromTimeval converts a Linux Timeval to Time.
+func FromTimeval(tv linux.Timeval) Time {
+	return Time{tv.ToNsecCapped()}
+}
+
+// Nanoseconds returns nanoseconds elapsed since the zero time in t's Clock
+// domain. If t represents walltime, this is nanoseconds since the Unix epoch.
+func (t Time) Nanoseconds() int64 {
+	return t.ns
+}
+
+// Seconds returns seconds elapsed since the zero time in t's Clock domain. If
+// t represents walltime, this is seconds since Unix epoch.
+func (t Time) Seconds() int64 {
+	return t.Nanoseconds() / time.Second.Nanoseconds()
+}
+
+// Timespec converts Time to a Linux timespec.
+func (t Time) Timespec() linux.Timespec {
+	return linux.NsecToTimespec(t.Nanoseconds())
+}
+
+// Unix returns the (seconds, nanoseconds) representation of t such that
+// seconds*1e9 + nanoseconds = t.
+func (t Time) Unix() (s int64, ns int64) {
+	s = t.ns / 1e9
+	ns = t.ns % 1e9
+	return
+}
+
+// TimeT converts Time to a Linux time_t.
+func (t Time) TimeT() linux.TimeT {
+	return linux.NsecToTimeT(t.Nanoseconds())
+}
+
+// Timeval converts Time to a Linux timeval.
+func (t Time) Timeval() linux.Timeval {
+	return linux.NsecToTimeval(t.Nanoseconds())
+}
+
+// Add adds the duration of d to t.
+func (t Time) Add(d time.Duration) Time {
+	if t.ns > 0 && d.Nanoseconds() > math.MaxInt64-int64(t.ns) {
+		return MaxTime
+	}
+	if t.ns < 0 && d.Nanoseconds() < math.MinInt64-int64(t.ns) {
+		return MinTime
+	}
+	return Time{int64(t.ns) + d.Nanoseconds()}
+}
+
+// AddTime adds the duration of u to t.
+func (t Time) AddTime(u Time) Time {
+	return t.Add(time.Duration(u.ns))
+}
+
+// Equal reports whether the two times represent the same instant in time.
+func (t Time) Equal(u Time) bool {
+	return t.ns == u.ns
+}
+
+// Before reports whether the instant t is before the instant u.
+func (t Time) Before(u Time) bool {
+	return t.ns < u.ns
+}
+
+// After reports whether the instant t is after the instant u.
+func (t Time) After(u Time) bool {
+	return t.ns > u.ns
+}
+
+// Sub returns the duration of t - u.
+//
+// N.B. This measure may not make sense for every Time returned by ktime.Clock.
+// Callers who need wall time duration can use ktime.Clock.WallTimeUntil to
+// estimate that wall time.
+func (t Time) Sub(u Time) time.Duration {
+	dur := time.Duration(int64(t.ns)-int64(u.ns)) * time.Nanosecond
+	switch {
+	case u.Add(dur).Equal(t):
+		return dur
+	case t.Before(u):
+		return MinDuration
+	default:
+		return MaxDuration
+	}
+}
+
+// IsMin returns whether t represents the lowest possible time instant.
+func (t Time) IsMin() bool {
+	return t == MinTime
+}
+
+// IsZero returns whether t represents the zero time instant in t's Clock domain.
+func (t Time) IsZero() bool {
+	return t == ZeroTime
+}
+
+// String returns the time represented in nanoseconds as a string.
+func (t Time) String() string {
+	return fmt.Sprintf("%dns", t.Nanoseconds())
+}
+
+// A Clock is an abstract time source.
+type Clock interface {
+	// Now returns the current time in nanoseconds according to the Clock.
+	Now() Time
+
+	// WallTimeUntil returns the estimated wall time until Now will return a
+	// value greater than or equal to t, given that a recent call to Now
+	// returned now. If t has already passed, WallTimeUntil may return 0 or a
+	// negative value.
+	//
+	// WallTimeUntil must be abstract to support Clocks that do not represent
+	// wall time (e.g. thread group execution timers). Clocks that represent
+	// wall times may embed the WallRateClock type to obtain an appropriate
+	// trivial implementation of WallTimeUntil.
+	//
+	// WallTimeUntil is used to determine when associated Timers should next
+	// check for expirations. Returning too small a value may result in
+	// spurious Timer goroutine wakeups, while returning too large a value may
+	// result in late expirations. Implementations should usually err on the
+	// side of underestimating.
+	WallTimeUntil(t, now Time) time.Duration
+
+	// Waitable methods may be used to subscribe to Clock events. Waiters will
+	// not be preserved by Save and must be re-established during restore.
+	//
+	// Since Clock events are transient, implementations of
+	// waiter.Waitable.Readiness should return 0.
+	waiter.Waitable
+}
+
+// WallRateClock implements Clock.WallTimeUntil for Clocks that elapse at the
+// same rate as wall time.
+type WallRateClock struct{}
+
+// WallTimeUntil implements Clock.WallTimeUntil.
+func (WallRateClock) WallTimeUntil(t, now Time) time.Duration {
+	return t.Sub(now)
+}
+
+// NoClockEvents implements waiter.Waitable for Clocks that do not generate
+// events.
+type NoClockEvents struct{}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (NoClockEvents) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return 0
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (NoClockEvents) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (NoClockEvents) EventUnregister(e *waiter.Entry) {
+}
+
+// ClockEventsQueue implements waiter.Waitable by wrapping waiter.Queue and
+// defining waiter.Waitable.Readiness as required by Clock.
+type ClockEventsQueue struct {
+	waiter.Queue
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (ClockEventsQueue) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return 0
+}
+
+// A TimerListener receives expirations from a Timer.
+type TimerListener interface {
+	// Notify is called when its associated Timer expires. exp is the number of
+	// expirations.
+	//
+	// Notify is called with the associated Timer's mutex locked, so Notify
+	// must not take any locks that precede Timer.mu in lock order.
+	//
+	// Preconditions: exp > 0.
+	Notify(exp uint64)
+
+	// Destroy is called when the timer is destroyed.
+	Destroy()
+}
+
+// Setting contains user-controlled mutable Timer properties.
+type Setting struct {
+	// Enabled is true if the timer is running.
+	Enabled bool
+
+	// Next is the time in nanoseconds of the next expiration.
+	Next Time
+
+	// Period is the time in nanoseconds between expirations. If Period is
+	// zero, the timer will not automatically restart after expiring.
+	//
+	// Invariant: Period >= 0.
+	Period time.Duration
+}
+
+// SettingFromSpec converts a (value, interval) pair to a Setting based on a
+// reading from c. value is interpreted as a time relative to c.Now().
+func SettingFromSpec(value time.Duration, interval time.Duration, c Clock) (Setting, error) {
+	if value < 0 {
+		return Setting{}, syserror.EINVAL
+	}
+	if value == 0 {
+		return Setting{Period: interval}, nil
+	}
+	return Setting{
+		Enabled: true,
+		Next:    c.Now().Add(value),
+		Period:  interval,
+	}, nil
+}
+
+// SettingFromAbsSpec converts a (value, interval) pair to a Setting based on a
+// reading from c. value is interpreted as an absolute time.
+func SettingFromAbsSpec(value Time, interval time.Duration) (Setting, error) {
+	if value.Before(ZeroTime) {
+		return Setting{}, syserror.EINVAL
+	}
+	if value.IsZero() {
+		return Setting{Period: interval}, nil
+	}
+	return Setting{
+		Enabled: true,
+		Next:    value,
+		Period:  interval,
+	}, nil
+}
+
+// SpecFromSetting converts a timestamp and a Setting to a (relative value,
+// interval) pair, as used by most Linux syscalls that return a struct
+// itimerval or struct itimerspec.
+func SpecFromSetting(now Time, s Setting) (value, period time.Duration) {
+	if !s.Enabled {
+		return 0, s.Period
+	}
+	return s.Next.Sub(now), s.Period
+}
+
+// advancedTo returns an updated Setting and a number of expirations after
+// the associated Clock indicates a time of now.
+//
+// Settings may be created by successive calls to advancedTo with decreasing
+// values of now (i.e. time may appear to go backward). Supporting this is
+// required to support non-monotonic clocks, as well as allowing
+// Timer.clock.Now() to be called without holding Timer.mu.
+func (s Setting) advancedTo(now Time) (Setting, uint64) {
+	if !s.Enabled {
+		return s, 0
+	}
+	if s.Next.After(now) {
+		return s, 0
+	}
+	if s.Period == 0 {
+		s.Enabled = false
+		return s, 1
+	}
+	exp := 1 + uint64(now.Sub(s.Next).Nanoseconds())/uint64(s.Period)
+	s.Next = s.Next.Add(time.Duration(uint64(s.Period) * exp))
+	return s, exp
+}
+
+// Timer is an optionally-periodic timer driven by sampling a user-specified
+// Clock. Timer's semantics support the requirements of Linux's interval timers
+// (setitimer(2), timer_create(2), timerfd_create(2)).
+//
+// Timers should be created using NewTimer and must be cleaned up by calling
+// Timer.Destroy when no longer used.
+type Timer struct {
+	// clock is the time source. clock is immutable.
+	clock Clock
+
+	// listener is notified of expirations. listener is immutable.
+	listener TimerListener
+
+	// mu protects the following mutable fields.
+	mu sync.Mutex `state:"nosave"`
+
+	// setting is the timer setting. setting is protected by mu.
+	setting Setting
+
+	// paused is true if the Timer is paused. paused is protected by mu.
+	paused bool
+
+	// kicker is used to wake the Timer goroutine. The kicker pointer is
+	// immutable, but its state is protected by mu.
+	kicker *time.Timer `state:"nosave"`
+
+	// entry is registered with clock.EventRegister. entry is immutable.
+	//
+	// Per comment in Clock, entry must be re-registered after restore; per
+	// comment in Timer.Load, this is done in Timer.Resume.
+	entry waiter.Entry `state:"nosave"`
+
+	// events is the channel that will be notified whenever entry receives an
+	// event. It is also closed by Timer.Destroy to instruct the Timer
+	// goroutine to exit.
+	events chan struct{} `state:"nosave"`
+}
+
+// timerTickEvents are Clock events that require the Timer goroutine to Tick
+// prematurely.
+const timerTickEvents = ClockEventSet | ClockEventRateIncrease
+
+// NewTimer returns a new Timer that will obtain time from clock and send
+// expirations to listener. The Timer is initially stopped and has no first
+// expiration or period configured.
+func NewTimer(clock Clock, listener TimerListener) *Timer {
+	t := &Timer{
+		clock:    clock,
+		listener: listener,
+	}
+	t.init()
+	return t
+}
+
+// After waits for the duration to elapse according to clock and then sends a
+// notification on the returned channel. The timer is started immediately and
+// will fire exactly once. The second return value is the start time used with
+// the duration.
+//
+// Callers must call Timer.Destroy.
+func After(clock Clock, duration time.Duration) (*Timer, Time, <-chan struct{}) {
+	notifier, tchan := NewChannelNotifier()
+	t := NewTimer(clock, notifier)
+	now := clock.Now()
+
+	t.Swap(Setting{
+		Enabled: true,
+		Period:  0,
+		Next:    now.Add(duration),
+	})
+	return t, now, tchan
+}
+
+// init initializes Timer state that is not preserved across save/restore. If
+// init has already been called, calling it again is a no-op.
+//
+// Preconditions: t.mu must be locked, or the caller must have exclusive access
+// to t.
+func (t *Timer) init() {
+	if t.kicker != nil {
+		return
+	}
+	// If t.kicker is nil, the Timer goroutine can't be running, so we can't
+	// race with it.
+	t.kicker = time.NewTimer(0)
+	t.entry, t.events = waiter.NewChannelEntry(nil)
+	t.clock.EventRegister(&t.entry, timerTickEvents)
+	go t.runGoroutine() // S/R-SAFE: synchronized by t.mu
+}
+
+// Destroy releases resources owned by the Timer. A Destroyed Timer must not be
+// used again; in particular, a Destroyed Timer should not be Saved.
+func (t *Timer) Destroy() {
+	// Stop the Timer, ensuring that the Timer goroutine will not call
+	// t.kicker.Reset, before calling t.kicker.Stop.
+	t.mu.Lock()
+	t.setting.Enabled = false
+	t.mu.Unlock()
+	t.kicker.Stop()
+	// Unregister t.entry, ensuring that the Clock will not send to t.events,
+	// before closing t.events to instruct the Timer goroutine to exit.
+	t.clock.EventUnregister(&t.entry)
+	close(t.events)
+	t.listener.Destroy()
+}
+
+func (t *Timer) runGoroutine() {
+	for {
+		select {
+		case <-t.kicker.C:
+		case _, ok := <-t.events:
+			if !ok {
+				// Channel closed by Destroy.
+				return
+			}
+		}
+		t.Tick()
+	}
+}
+
+// Tick requests that the Timer immediately check for expirations and
+// re-evaluate when it should next check for expirations.
+func (t *Timer) Tick() {
+	now := t.clock.Now()
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if t.paused {
+		return
+	}
+	s, exp := t.setting.advancedTo(now)
+	t.setting = s
+	if exp > 0 {
+		t.listener.Notify(exp)
+	}
+	t.resetKickerLocked(now)
+}
+
+// Pause pauses the Timer, ensuring that it does not generate any further
+// expirations until Resume is called. If the Timer is already paused, Pause
+// has no effect.
+func (t *Timer) Pause() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.paused = true
+	// t.kicker may be nil if we were restored but never resumed.
+	if t.kicker != nil {
+		t.kicker.Stop()
+	}
+}
+
+// Resume ends the effect of Pause. If the Timer is not paused, Resume has no
+// effect.
+func (t *Timer) Resume() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if !t.paused {
+		return
+	}
+	t.paused = false
+
+	// Lazily initialize the Timer. We can't call Timer.init until Timer.Resume
+	// because save/restore will restore Timers before
+	// kernel.Timekeeper.SetClocks() has been called, so if t.clock is backed
+	// by a kernel.Timekeeper then the Timer goroutine will panic if it calls
+	// t.clock.Now().
+	t.init()
+
+	// Kick the Timer goroutine in case it was already initialized, but the
+	// Timer goroutine was sleeping.
+	t.kicker.Reset(0)
+}
+
+// Get returns a snapshot of the Timer's current Setting and the time
+// (according to the Timer's Clock) at which the snapshot was taken.
+//
+// Preconditions: The Timer must not be paused (since its Setting cannot
+// be advanced to the current time while it is paused.)
+func (t *Timer) Get() (Time, Setting) {
+	now := t.clock.Now()
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if t.paused {
+		panic(fmt.Sprintf("Timer.Get called on paused Timer %p", t))
+	}
+	s, exp := t.setting.advancedTo(now)
+	t.setting = s
+	if exp > 0 {
+		t.listener.Notify(exp)
+	}
+	t.resetKickerLocked(now)
+	return now, s
+}
+
+// Swap atomically changes the Timer's Setting and returns the Timer's previous
+// Setting and the time (according to the Timer's Clock) at which the snapshot
+// was taken. Setting s.Enabled to true starts the Timer, while setting
+// s.Enabled to false stops it.
+//
+// Preconditions: The Timer must not be paused.
+func (t *Timer) Swap(s Setting) (Time, Setting) {
+	return t.SwapAnd(s, nil)
+}
+
+// SwapAnd atomically changes the Timer's Setting, calls f if it is not nil,
+// and returns the Timer's previous Setting and the time (according to the
+// Timer's Clock) at which the Setting was changed. Setting s.Enabled to true
+// starts the timer, while setting s.Enabled to false stops it.
+//
+// Preconditions: The Timer must not be paused. f cannot call any Timer methods
+// since it is called with the Timer mutex locked.
+func (t *Timer) SwapAnd(s Setting, f func()) (Time, Setting) {
+	now := t.clock.Now()
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	if t.paused {
+		panic(fmt.Sprintf("Timer.SwapAnd called on paused Timer %p", t))
+	}
+	oldS, oldExp := t.setting.advancedTo(now)
+	if oldExp > 0 {
+		t.listener.Notify(oldExp)
+	}
+	if f != nil {
+		f()
+	}
+	newS, newExp := s.advancedTo(now)
+	t.setting = newS
+	if newExp > 0 {
+		t.listener.Notify(newExp)
+	}
+	t.resetKickerLocked(now)
+	return now, oldS
+}
+
+// Preconditions: t.mu must be locked.
+func (t *Timer) resetKickerLocked(now Time) {
+	if t.setting.Enabled {
+		// Clock.WallTimeUntil may return a negative value. This is fine;
+		// time.when treats negative Durations as 0.
+		t.kicker.Reset(t.clock.WallTimeUntil(t.setting.Next, now))
+	}
+	// We don't call t.kicker.Stop if !t.setting.Enabled because in most cases
+	// resetKickerLocked will be called from the Timer goroutine itself, in
+	// which case t.kicker has already fired and t.kicker.Stop will be an
+	// expensive no-op (time.Timer.Stop => time.stopTimer => runtime.stopTimer
+	// => runtime.deltimer).
+}
+
+// Clock returns the Clock used by t.
+func (t *Timer) Clock() Clock {
+	return t.clock
+}
+
+// ChannelNotifier is a TimerListener that sends a message on an empty struct
+// channel.
+//
+// ChannelNotifier cannot be saved or loaded.
+type ChannelNotifier struct {
+	// tchan must be a buffered channel.
+	tchan chan struct{}
+}
+
+// NewChannelNotifier creates a new channel notifier.
+//
+// If the notifier is used with a timer, Timer.Destroy will close the channel
+// returned here.
+func NewChannelNotifier() (TimerListener, <-chan struct{}) {
+	tchan := make(chan struct{}, 1)
+	return &ChannelNotifier{tchan}, tchan
+}
+
+// Notify implements ktime.TimerListener.Notify.
+func (c *ChannelNotifier) Notify(uint64) {
+	select {
+	case c.tchan <- struct{}{}:
+	default:
+	}
+}
+
+// Destroy implements ktime.TimerListener.Destroy and will close the channel.
+func (c *ChannelNotifier) Destroy() {
+	close(c.tchan)
+}
diff --git a/pkg/sentry/kernel/timekeeper.go b/pkg/sentry/kernel/timekeeper.go
new file mode 100644
index 000000000..3f16c1676
--- /dev/null
+++ b/pkg/sentry/kernel/timekeeper.go
@@ -0,0 +1,270 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	sentrytime "gvisor.googlesource.com/gvisor/pkg/sentry/time"
+)
+
+// Timekeeper manages all of the kernel clocks.
+type Timekeeper struct {
+	// clocks are the clock sources.
+	//
+	// These are not saved directly, as the new machine's clock may behave
+	// differently.
+	//
+	// It is set only once, by SetClocks.
+	clocks sentrytime.Clocks `state:"nosave"`
+
+	// bootTime is the realtime when the system "booted". i.e., when
+	// SetClocks was called in the initial (not restored) run.
+	bootTime ktime.Time
+
+	// monotonicOffset is the offset to apply to the monotonic clock output
+	// from clocks.
+	//
+	// It is set only once, by SetClocks.
+	monotonicOffset int64 `state:"nosave"`
+
+	// restored indicates that this Timekeeper was restored from a state
+	// file.
+	restored bool `state:"nosave"`
+
+	// saveMonotonic is the (offset) value of the monotonic clock at the
+	// time of save.
+	//
+	// It is only valid if restored is true.
+	//
+	// It is only used in SetClocks after restore to compute the new
+	// monotonicOffset.
+	saveMonotonic int64
+
+	// saveRealtime is the value of the realtime clock at the time of save.
+	//
+	// It is only valid if restored is true.
+	//
+	// It is only used in SetClocks after restore to compute the new
+	// monotonicOffset.
+	saveRealtime int64
+
+	// params manages the parameter page.
+	params *VDSOParamPage
+
+	// mu protects destruction with stop and wg.
+	mu sync.Mutex `state:"nosave"`
+
+	// stop is used to tell the update goroutine to exit.
+	stop chan struct{} `state:"nosave"`
+
+	// wg is used to indicate that the update goroutine has exited.
+	wg sync.WaitGroup `state:"nosave"`
+}
+
+// NewTimekeeper returns a Timekeeper that is automatically kept up-to-date.
+// NewTimekeeper does not take ownership of paramPage.
+//
+// SetClocks must be called on the returned Timekeeper before it is usable.
+func NewTimekeeper(platform platform.Platform, paramPage platform.FileRange) (*Timekeeper, error) {
+	return &Timekeeper{
+		params: NewVDSOParamPage(platform, paramPage),
+	}, nil
+}
+
+// SetClocks the backing clock source.
+//
+// SetClocks must be called before the Timekeeper is used, and it may not be
+// called more than once, as changing the clock source without extra correction
+// could cause time discontinuities.
+//
+// It must also be called after Load.
+func (t *Timekeeper) SetClocks(c sentrytime.Clocks) {
+	// Update the params, marking them "not ready", as we may need to
+	// restart calibration on this new machine.
+	if t.restored {
+		if err := t.params.Write(func() vdsoParams {
+			return vdsoParams{}
+		}); err != nil {
+			panic("unable to reset VDSO params: " + err.Error())
+		}
+	}
+
+	if t.clocks != nil {
+		panic("SetClocks called on previously-initialized Timekeeper")
+	}
+
+	t.clocks = c
+
+	// Compute the offset of the monotonic clock from the base Clocks.
+	//
+	// In a fresh (not restored) sentry, monotonic time starts at zero.
+	//
+	// In a restored sentry, monotonic time jumps forward by approximately
+	// the same amount as real time. There are no guarantees here, we are
+	// just making a best-effort attempt to to make it appear that the app
+	// was simply not scheduled for a long period, rather than that the
+	// real time clock was changed.
+	//
+	// If real time went backwards, it remains the same.
+	wantMonotonic := int64(0)
+
+	nowMonotonic, err := t.clocks.GetTime(sentrytime.Monotonic)
+	if err != nil {
+		panic("Unable to get current monotonic time: " + err.Error())
+	}
+
+	nowRealtime, err := t.clocks.GetTime(sentrytime.Realtime)
+	if err != nil {
+		panic("Unable to get current realtime: " + err.Error())
+	}
+
+	if t.restored {
+		wantMonotonic = t.saveMonotonic
+		elapsed := nowRealtime - t.saveRealtime
+		if elapsed > 0 {
+			wantMonotonic += elapsed
+		}
+	}
+
+	t.monotonicOffset = wantMonotonic - nowMonotonic
+
+	if !t.restored {
+		// Hold on to the initial "boot" time.
+		t.bootTime = ktime.FromNanoseconds(nowRealtime)
+	}
+
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.startUpdater()
+}
+
+// startUpdater starts an update goroutine that keeps the clocks updated.
+//
+// mu must be held.
+func (t *Timekeeper) startUpdater() {
+	if t.stop != nil {
+		// Timekeeper already started
+		return
+	}
+	t.stop = make(chan struct{})
+
+	// Keep the clocks up to date.
+	//
+	// Note that the Go runtime uses host CLOCK_MONOTONIC to service the
+	// timer, so it may run at a *slightly* different rate from the
+	// application CLOCK_MONOTONIC. That is fine, as we only need to update
+	// at approximately this rate.
+	timer := time.NewTicker(sentrytime.ApproxUpdateInterval)
+	t.wg.Add(1)
+	go func() { // S/R-SAFE: stopped during save.
+		for {
+			// Start with an update immediately, so the clocks are
+			// ready ASAP.
+
+			// Call Update within a Write block to prevent the VDSO
+			// from using the old params between Update and
+			// Write.
+			if err := t.params.Write(func() vdsoParams {
+				monotonicParams, monotonicOk, realtimeParams, realtimeOk := t.clocks.Update()
+
+				var p vdsoParams
+				if monotonicOk {
+					p.monotonicReady = 1
+					p.monotonicBaseCycles = int64(monotonicParams.BaseCycles)
+					p.monotonicBaseRef = int64(monotonicParams.BaseRef) + t.monotonicOffset
+					p.monotonicFrequency = monotonicParams.Frequency
+				}
+				if realtimeOk {
+					p.realtimeReady = 1
+					p.realtimeBaseCycles = int64(realtimeParams.BaseCycles)
+					p.realtimeBaseRef = int64(realtimeParams.BaseRef)
+					p.realtimeFrequency = realtimeParams.Frequency
+				}
+
+				log.Debugf("Updating VDSO parameters: %+v", p)
+
+				return p
+			}); err != nil {
+				log.Warningf("Unable to update VDSO parameter page: %v", err)
+			}
+
+			select {
+			case <-timer.C:
+			case <-t.stop:
+				t.wg.Done()
+				return
+			}
+		}
+	}()
+}
+
+// stopUpdater stops the update goroutine, blocking until it exits.
+//
+// mu must be held.
+func (t *Timekeeper) stopUpdater() {
+	if t.stop == nil {
+		// Updater not running.
+		return
+	}
+
+	close(t.stop)
+	t.wg.Wait()
+	t.stop = nil
+}
+
+// Destroy destroys the Timekeeper, freeing all associated resources.
+func (t *Timekeeper) Destroy() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+
+	t.stopUpdater()
+}
+
+// PauseUpdates stops clock parameter updates. This should only be used when
+// Tasks are not running and thus cannot access the clock.
+func (t *Timekeeper) PauseUpdates() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.stopUpdater()
+}
+
+// ResumeUpdates restarts clock parameter updates stopped by PauseUpdates.
+func (t *Timekeeper) ResumeUpdates() {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	t.startUpdater()
+}
+
+// GetTime returns the current time in nanoseconds.
+func (t *Timekeeper) GetTime(c sentrytime.ClockID) (int64, error) {
+	if t.clocks == nil {
+		panic("Timekeeper used before initialized with SetClocks")
+	}
+	now, err := t.clocks.GetTime(c)
+	if err == nil && c == sentrytime.Monotonic {
+		now += t.monotonicOffset
+	}
+	return now, err
+}
+
+// BootTime returns the system boot real time.
+func (t *Timekeeper) BootTime() ktime.Time {
+	return t.bootTime
+}
diff --git a/pkg/sentry/kernel/timekeeper_state.go b/pkg/sentry/kernel/timekeeper_state.go
new file mode 100644
index 000000000..aee983ac7
--- /dev/null
+++ b/pkg/sentry/kernel/timekeeper_state.go
@@ -0,0 +1,41 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/time"
+)
+
+// beforeSave is invoked by stateify.
+func (t *Timekeeper) beforeSave() {
+	if t.stop != nil {
+		panic("pauseUpdates must be called before Save")
+	}
+
+	// N.B. we want the *offset* monotonic time.
+	var err error
+	if t.saveMonotonic, err = t.GetTime(time.Monotonic); err != nil {
+		panic("unable to get current monotonic time: " + err.Error())
+	}
+
+	if t.saveRealtime, err = t.GetTime(time.Realtime); err != nil {
+		panic("unable to get current realtime: " + err.Error())
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (t *Timekeeper) afterLoad() {
+	t.restored = true
+}
diff --git a/pkg/sentry/kernel/timekeeper_test.go b/pkg/sentry/kernel/timekeeper_test.go
new file mode 100644
index 000000000..08bacba4f
--- /dev/null
+++ b/pkg/sentry/kernel/timekeeper_test.go
@@ -0,0 +1,156 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	sentrytime "gvisor.googlesource.com/gvisor/pkg/sentry/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// mockClocks is a sentrytime.Clocks that simply returns the times in the
+// struct.
+type mockClocks struct {
+	monotonic int64
+	realtime  int64
+}
+
+// Update implements sentrytime.Clocks.Update. It does nothing.
+func (*mockClocks) Update() (monotonicParams sentrytime.Parameters, monotonicOk bool, realtimeParam sentrytime.Parameters, realtimeOk bool) {
+	return
+}
+
+// Update implements sentrytime.Clocks.GetTime.
+func (c *mockClocks) GetTime(id sentrytime.ClockID) (int64, error) {
+	switch id {
+	case sentrytime.Monotonic:
+		return c.monotonic, nil
+	case sentrytime.Realtime:
+		return c.realtime, nil
+	default:
+		return 0, syserror.EINVAL
+	}
+}
+
+// stateTestClocklessTimekeeper returns a test Timekeeper which has not had
+// SetClocks called.
+func stateTestClocklessTimekeeper(tb testing.TB) *Timekeeper {
+	ctx := contexttest.Context(tb)
+	p := platform.FromContext(ctx)
+	fr, err := p.Memory().Allocate(usermem.PageSize, usage.Anonymous)
+	if err != nil {
+		tb.Fatalf("failed to allocate memory: %v", err)
+	}
+	return &Timekeeper{
+		params: NewVDSOParamPage(p, fr),
+	}
+}
+
+func stateTestTimekeeper(tb testing.TB) *Timekeeper {
+	t := stateTestClocklessTimekeeper(tb)
+	t.SetClocks(sentrytime.NewCalibratedClocks())
+	return t
+}
+
+// TestTimekeeperMonotonicZero tests that monotonic time starts at zero.
+func TestTimekeeperMonotonicZero(t *testing.T) {
+	c := &mockClocks{
+		monotonic: 100000,
+	}
+
+	tk := stateTestClocklessTimekeeper(t)
+	tk.SetClocks(c)
+	defer tk.Destroy()
+
+	now, err := tk.GetTime(sentrytime.Monotonic)
+	if err != nil {
+		t.Errorf("GetTime err got %v want nil", err)
+	}
+	if now != 0 {
+		t.Errorf("GetTime got %d want 0", now)
+	}
+
+	c.monotonic += 10
+
+	now, err = tk.GetTime(sentrytime.Monotonic)
+	if err != nil {
+		t.Errorf("GetTime err got %v want nil", err)
+	}
+	if now != 10 {
+		t.Errorf("GetTime got %d want 10", now)
+	}
+}
+
+// TestTimekeeperMonotonicJumpForward tests that monotonic time jumps forward
+// after restore.
+func TestTimekeeperMonotonicForward(t *testing.T) {
+	c := &mockClocks{
+		monotonic: 900000,
+		realtime:  600000,
+	}
+
+	tk := stateTestClocklessTimekeeper(t)
+	tk.restored = true
+	tk.saveMonotonic = 100000
+	tk.saveRealtime = 400000
+	tk.SetClocks(c)
+	defer tk.Destroy()
+
+	// The monotonic clock should jump ahead by 200000 to 300000.
+	//
+	// The new system monotonic time (900000) is irrelevant to what the app
+	// sees.
+	now, err := tk.GetTime(sentrytime.Monotonic)
+	if err != nil {
+		t.Errorf("GetTime err got %v want nil", err)
+	}
+	if now != 300000 {
+		t.Errorf("GetTime got %d want 300000", now)
+	}
+}
+
+// TestTimekeeperMonotonicJumpBackwards tests that monotonic time does not jump
+// backwards when realtime goes backwards.
+func TestTimekeeperMonotonicJumpBackwards(t *testing.T) {
+	c := &mockClocks{
+		monotonic: 900000,
+		realtime:  400000,
+	}
+
+	tk := stateTestClocklessTimekeeper(t)
+	tk.restored = true
+	tk.saveMonotonic = 100000
+	tk.saveRealtime = 600000
+	tk.SetClocks(c)
+	defer tk.Destroy()
+
+	// The monotonic clock should remain at 100000.
+	//
+	// The new system monotonic time (900000) is irrelevant to what the app
+	// sees and we don't want to jump the monotonic clock backwards like
+	// realtime did.
+	now, err := tk.GetTime(sentrytime.Monotonic)
+	if err != nil {
+		t.Errorf("GetTime err got %v want nil", err)
+	}
+	if now != 100000 {
+		t.Errorf("GetTime got %d want 100000", now)
+	}
+}
diff --git a/pkg/sentry/kernel/timer.go b/pkg/sentry/kernel/timer.go
new file mode 100644
index 000000000..03a3310be
--- /dev/null
+++ b/pkg/sentry/kernel/timer.go
@@ -0,0 +1,282 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	sentrytime "gvisor.googlesource.com/gvisor/pkg/sentry/time"
+)
+
+// timekeeperClock is a ktime.Clock that reads time from a
+// kernel.Timekeeper-managed clock.
+type timekeeperClock struct {
+	tk *Timekeeper
+	c  sentrytime.ClockID
+
+	// Implements ktime.Clock.WallTimeUntil.
+	ktime.WallRateClock `state:"nosave"`
+
+	// Implements waiter.Waitable. (We have no ability to detect
+	// discontinuities from external changes to CLOCK_REALTIME).
+	ktime.NoClockEvents `state:"nosave"`
+}
+
+// Now implements ktime.Clock.Now.
+func (tc *timekeeperClock) Now() ktime.Time {
+	now, err := tc.tk.GetTime(tc.c)
+	if err != nil {
+		panic(fmt.Sprintf("timekeeperClock(ClockID=%v)).Now: %v", tc.c, err))
+	}
+	return ktime.FromNanoseconds(now)
+}
+
+// tgClock is a ktime.Clock that measures the time a thread group has spent
+// executing.
+type tgClock struct {
+	tg *ThreadGroup
+
+	// If includeSys is true, the tgClock includes both time spent executing
+	// application code as well as time spent in the sentry. Otherwise, the
+	// tgClock includes only time spent executing application code.
+	includeSys bool
+
+	// Implements waiter.Waitable.
+	ktime.ClockEventsQueue `state:"nosave"`
+}
+
+// UserCPUClock returns a ktime.Clock that measures the time that a thread
+// group has spent executing.
+func (tg *ThreadGroup) UserCPUClock() ktime.Clock {
+	return tg.tm.virtClock
+}
+
+// CPUClock returns a ktime.Clock that measures the time that a thread group
+// has spent executing, including sentry time.
+func (tg *ThreadGroup) CPUClock() ktime.Clock {
+	return tg.tm.profClock
+}
+
+// Now implements ktime.Clock.Now.
+func (tgc *tgClock) Now() ktime.Time {
+	stats := tgc.tg.CPUStats()
+	if tgc.includeSys {
+		return ktime.FromNanoseconds((stats.UserTime + stats.SysTime).Nanoseconds())
+	}
+	return ktime.FromNanoseconds(stats.UserTime.Nanoseconds())
+}
+
+// WallTimeUntil implements ktime.Clock.WallTimeUntil.
+func (tgc *tgClock) WallTimeUntil(t, now ktime.Time) time.Duration {
+	// The assumption here is that the time spent in this process (not matter
+	// virtual or prof) should not exceed wall time * active tasks, since
+	// Task.exitThreadGroup stops accounting as it transitions to
+	// TaskExitInitiated.
+	tgc.tg.pidns.owner.mu.RLock()
+	n := tgc.tg.activeTasks
+	tgc.tg.pidns.owner.mu.RUnlock()
+	if n == 0 {
+		if t.Before(now) {
+			return 0
+		}
+		// The timer tick raced with thread group exit, after which no more
+		// tasks can enter the thread group. So tgc.Now() will never advance
+		// again. Return a large delay; the timer should be stopped long before
+		// it comes again anyway.
+		return time.Hour
+	}
+	// This is a lower bound on the amount of time that can elapse before an
+	// associated timer expires, so returning this value tends to result in a
+	// sequence of closely-spaced ticks just before timer expiry. To avoid
+	// this, round up to the nearest ClockTick; CPU usage measurements are
+	// limited to this resolution anyway.
+	remaining := time.Duration(int64(t.Sub(now))/int64(n)) * time.Nanosecond
+	return ((remaining + (linux.ClockTick - time.Nanosecond)) / linux.ClockTick) * linux.ClockTick
+}
+
+// taskClock is a ktime.Clock that measures the time that a task has spent
+// executing.
+type taskClock struct {
+	t *Task
+
+	// If includeSys is true, the taskClock includes both time spent executing
+	// application code as well as time spent in the sentry. Otherwise, the
+	// taskClock includes only time spent executing application code.
+	includeSys bool
+
+	// Implements waiter.Waitable. TimeUntil wouldn't change its estimation
+	// based on either of the clock events, so there's no event to be
+	// notified for.
+	ktime.NoClockEvents `state:"nosave"`
+
+	// Implements ktime.Clock.WallTimeUntil.
+	//
+	// As an upper bound, a task's clock cannot advance faster than CPU
+	// time. It would have to execute at a rate of more than 1 task-second
+	// per 1 CPU-second, which isn't possible.
+	ktime.WallRateClock `state:"nosave"`
+}
+
+// UserCPUClock returns a clock measuring the CPU time the task has spent
+// executing application code.
+func (t *Task) UserCPUClock() ktime.Clock {
+	return &taskClock{t: t, includeSys: false}
+}
+
+// CPUClock returns a clock measuring the CPU time the task has spent executing
+// application and "kernel" code.
+func (t *Task) CPUClock() ktime.Clock {
+	return &taskClock{t: t, includeSys: true}
+}
+
+// Now implements ktime.Clock.Now.
+func (tc *taskClock) Now() ktime.Time {
+	stats := tc.t.CPUStats()
+	if tc.includeSys {
+		return ktime.FromNanoseconds((stats.UserTime + stats.SysTime).Nanoseconds())
+	}
+	return ktime.FromNanoseconds(stats.UserTime.Nanoseconds())
+}
+
+// signalNotifier is a ktime.Listener that sends signals to a ThreadGroup.
+type signalNotifier struct {
+	tg         *ThreadGroup
+	signal     linux.Signal
+	realTimer  bool
+	includeSys bool
+}
+
+// Notify implements ktime.TimerListener.Notify.
+func (s *signalNotifier) Notify(exp uint64) {
+	// Since all signals sent using a signalNotifier are standard (not
+	// real-time) signals, we can ignore the number of expirations and send
+	// only a single signal.
+	if s.realTimer {
+		// real timer signal sent to leader. See kernel/time/itimer.c:it_real_fn
+		s.tg.SendSignal(sigPriv(s.signal))
+	} else {
+		s.tg.SendTimerSignal(sigPriv(s.signal), s.includeSys)
+	}
+}
+
+// Destroy implements ktime.TimerListener.Destroy.
+func (s *signalNotifier) Destroy() {}
+
+// TimerManager is a collection of supported process cpu timers.
+type TimerManager struct {
+	// Clocks used to drive thread group execution time timers.
+	virtClock *tgClock
+	profClock *tgClock
+
+	RealTimer      *ktime.Timer
+	VirtualTimer   *ktime.Timer
+	ProfTimer      *ktime.Timer
+	SoftLimitTimer *ktime.Timer
+	HardLimitTimer *ktime.Timer
+}
+
+// newTimerManager returns a new instance of TimerManager.
+func newTimerManager(tg *ThreadGroup, monotonicClock ktime.Clock) TimerManager {
+	virtClock := &tgClock{tg: tg, includeSys: false}
+	profClock := &tgClock{tg: tg, includeSys: true}
+	tm := TimerManager{
+		virtClock: virtClock,
+		profClock: profClock,
+		RealTimer: ktime.NewTimer(monotonicClock, &signalNotifier{
+			tg:         tg,
+			signal:     linux.SIGALRM,
+			realTimer:  true,
+			includeSys: false,
+		}),
+		VirtualTimer: ktime.NewTimer(virtClock, &signalNotifier{
+			tg:         tg,
+			signal:     linux.SIGVTALRM,
+			realTimer:  false,
+			includeSys: false,
+		}),
+		ProfTimer: ktime.NewTimer(profClock, &signalNotifier{
+			tg:         tg,
+			signal:     linux.SIGPROF,
+			realTimer:  false,
+			includeSys: true,
+		}),
+		SoftLimitTimer: ktime.NewTimer(profClock, &signalNotifier{
+			tg:         tg,
+			signal:     linux.SIGXCPU,
+			realTimer:  false,
+			includeSys: true,
+		}),
+		HardLimitTimer: ktime.NewTimer(profClock, &signalNotifier{
+			tg:         tg,
+			signal:     linux.SIGKILL,
+			realTimer:  false,
+			includeSys: true,
+		}),
+	}
+	tm.applyCPULimits(tg.Limits().Get(limits.CPU))
+	return tm
+}
+
+// Save saves this TimerManger.
+
+// destroy destroys all timers.
+func (tm *TimerManager) destroy() {
+	tm.RealTimer.Destroy()
+	tm.VirtualTimer.Destroy()
+	tm.ProfTimer.Destroy()
+	tm.SoftLimitTimer.Destroy()
+	tm.HardLimitTimer.Destroy()
+}
+
+func (tm *TimerManager) applyCPULimits(l limits.Limit) {
+	tm.SoftLimitTimer.Swap(ktime.Setting{
+		Enabled: l.Cur != limits.Infinity,
+		Next:    ktime.FromNanoseconds((time.Duration(l.Cur) * time.Second).Nanoseconds()),
+		Period:  time.Second,
+	})
+	tm.HardLimitTimer.Swap(ktime.Setting{
+		Enabled: l.Max != limits.Infinity,
+		Next:    ktime.FromNanoseconds((time.Duration(l.Max) * time.Second).Nanoseconds()),
+	})
+}
+
+// kick is called when the number of threads in the thread group associated
+// with tm increases.
+func (tm *TimerManager) kick() {
+	tm.virtClock.Notify(ktime.ClockEventRateIncrease)
+	tm.profClock.Notify(ktime.ClockEventRateIncrease)
+}
+
+// pause is to pause the timers and stop timer signal delivery.
+func (tm *TimerManager) pause() {
+	tm.RealTimer.Pause()
+	tm.VirtualTimer.Pause()
+	tm.ProfTimer.Pause()
+	tm.SoftLimitTimer.Pause()
+	tm.HardLimitTimer.Pause()
+}
+
+// resume is to resume the timers and continue timer signal delivery.
+func (tm *TimerManager) resume() {
+	tm.RealTimer.Resume()
+	tm.VirtualTimer.Resume()
+	tm.ProfTimer.Resume()
+	tm.SoftLimitTimer.Resume()
+	tm.HardLimitTimer.Resume()
+}
diff --git a/pkg/sentry/kernel/uts_namespace.go b/pkg/sentry/kernel/uts_namespace.go
new file mode 100644
index 000000000..58e9b4d1b
--- /dev/null
+++ b/pkg/sentry/kernel/uts_namespace.go
@@ -0,0 +1,100 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+)
+
+// UTSNamespace represents a UTS namespace, a holder of two system identifiers:
+// the hostname and domain name.
+type UTSNamespace struct {
+	// mu protects all fields below.
+	mu         sync.Mutex `state:"nosave"`
+	hostName   string
+	domainName string
+
+	// userns is the user namespace associated with the UTSNamespace.
+	// Privileged operations on this UTSNamespace must have appropriate
+	// capabilities in userns.
+	//
+	// userns is immutable.
+	userns *auth.UserNamespace
+}
+
+// NewUTSNamespace creates a new UTS namespace.
+func NewUTSNamespace(hostName, domainName string, userns *auth.UserNamespace) *UTSNamespace {
+	return &UTSNamespace{
+		hostName:   hostName,
+		domainName: domainName,
+		userns:     userns,
+	}
+}
+
+// UTSNamespace returns the task's UTS namespace.
+func (t *Task) UTSNamespace() *UTSNamespace {
+	t.mu.Lock()
+	defer t.mu.Unlock()
+	return t.utsns
+}
+
+// HostName returns the host name of this UTS namespace.
+func (u *UTSNamespace) HostName() string {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return u.hostName
+}
+
+// SetHostName sets the host name of this UTS namespace.
+func (u *UTSNamespace) SetHostName(host string) {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	u.hostName = host
+}
+
+// DomainName returns the domain name of this UTS namespace.
+func (u *UTSNamespace) DomainName() string {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return u.domainName
+}
+
+// SetDomainName sets the domain name of this UTS namespace.
+func (u *UTSNamespace) SetDomainName(domain string) {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	u.domainName = domain
+}
+
+// UserNamespace returns the user namespace associated with this UTS namespace.
+func (u *UTSNamespace) UserNamespace() *auth.UserNamespace {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return u.userns
+}
+
+// Clone makes a copy of this UTS namespace, associating the given user
+// namespace.
+func (u *UTSNamespace) Clone(userns *auth.UserNamespace) *UTSNamespace {
+	u.mu.Lock()
+	defer u.mu.Unlock()
+	return &UTSNamespace{
+		hostName:   u.hostName,
+		domainName: u.domainName,
+		userns:     userns,
+	}
+}
diff --git a/pkg/sentry/kernel/vdso.go b/pkg/sentry/kernel/vdso.go
new file mode 100644
index 000000000..0bacbea49
--- /dev/null
+++ b/pkg/sentry/kernel/vdso.go
@@ -0,0 +1,145 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// vdsoParams are the parameters exposed to the VDSO.
+//
+// They are exposed to the VDSO via a parameter page managed by VDSOParamPage,
+// which also includes a sequence counter.
+type vdsoParams struct {
+	monotonicReady      uint64
+	monotonicBaseCycles int64
+	monotonicBaseRef    int64
+	monotonicFrequency  uint64
+
+	realtimeReady      uint64
+	realtimeBaseCycles int64
+	realtimeBaseRef    int64
+	realtimeFrequency  uint64
+}
+
+// VDSOParamPage manages a VDSO parameter page.
+//
+// Its memory layout looks like:
+//
+// type page struct {
+//	// seq is a sequence counter that protects the fields below.
+//	seq uint64
+//	vdsoParams
+// }
+//
+// Everything in the struct is 8 bytes for easy alignment.
+//
+// It must be kept in sync with params in vdso/vdso_time.cc.
+type VDSOParamPage struct {
+	// The parameter page is fr, allocated from platform.Memory().
+	platform platform.Platform
+	fr       platform.FileRange
+
+	// seq is the current sequence count written to the page.
+	//
+	// A write is in progress if bit 1 of the counter is set.
+	//
+	// Timekeeper's updater goroutine may call Write before equality is
+	// checked in state_test_util tests, causing this field to change across
+	// save / restore.
+	seq uint64
+}
+
+// NewVDSOParamPage returns a VDSOParamPage.
+//
+// Preconditions:
+//
+// * fr is a single page allocated from platform.Memory(). VDSOParamPage does
+//   not take ownership of fr; it must remain allocated for the lifetime of the
+//   VDSOParamPage.
+//
+// * VDSOParamPage must be the only writer to fr.
+//
+// * platform.Memory().MapInternal(fr) must return a single safemem.Block.
+func NewVDSOParamPage(platform platform.Platform, fr platform.FileRange) *VDSOParamPage {
+	return &VDSOParamPage{platform: platform, fr: fr}
+}
+
+// access returns a mapping of the param page.
+func (v *VDSOParamPage) access() (safemem.Block, error) {
+	bs, err := v.platform.Memory().MapInternal(v.fr, usermem.ReadWrite)
+	if err != nil {
+		return safemem.Block{}, err
+	}
+	if bs.NumBlocks() != 1 {
+		panic(fmt.Sprintf("Multiple blocks (%d) in VDSO param BlockSeq", bs.NumBlocks()))
+	}
+	return bs.Head(), nil
+}
+
+// incrementSeq increments the sequence counter in the param page.
+func (v *VDSOParamPage) incrementSeq(paramPage safemem.Block) error {
+	next := v.seq + 1
+	old, err := safemem.SwapUint64(paramPage, next)
+	if err != nil {
+		return err
+	}
+
+	if old != v.seq {
+		return fmt.Errorf("unexpected VDSOParamPage seq value: got %d expected %d. Application may hang or get incorrect time from the VDSO.", old, v.seq)
+	}
+
+	v.seq = next
+	return nil
+}
+
+// Write updates the VDSO parameters.
+//
+// Write starts a write block, calls f to get the new parameters, writes
+// out the new parameters, then ends the write block.
+func (v *VDSOParamPage) Write(f func() vdsoParams) error {
+	paramPage, err := v.access()
+	if err != nil {
+		return err
+	}
+
+	// Write begin.
+	next := v.seq + 1
+	if next%2 != 1 {
+		panic("Out-of-order sequence count")
+	}
+
+	err = v.incrementSeq(paramPage)
+	if err != nil {
+		return err
+	}
+
+	// Get the new params.
+	p := f()
+	buf := binary.Marshal(nil, usermem.ByteOrder, p)
+
+	// Skip the sequence counter.
+	if _, err := safemem.Copy(paramPage.DropFirst(8), safemem.BlockFromSafeSlice(buf)); err != nil {
+		panic(fmt.Sprintf("Unable to get set VDSO parameters: %v", err))
+	}
+
+	// Write end.
+	return v.incrementSeq(paramPage)
+}
diff --git a/pkg/sentry/kernel/version.go b/pkg/sentry/kernel/version.go
new file mode 100644
index 000000000..a9e84673f
--- /dev/null
+++ b/pkg/sentry/kernel/version.go
@@ -0,0 +1,33 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kernel
+
+// Version defines the application-visible system version.
+type Version struct {
+	// Operating system name (e.g. "Linux").
+	Sysname string
+
+	// Operating system release (e.g. "3.11.10-amd64").
+	Release string
+
+	// Operating system version. On Linux this takes the shape
+	// "#VERSION CONFIG_FLAGS TIMESTAMP"
+	// where:
+	// - VERSION is a sequence counter incremented on every successful build
+	// - CONFIG_FLAGS is a space-separated list of major enabled kernel features
+	//   (e.g. "SMP" and "PREEMPT")
+	// - TIMESTAMP is the build timestamp as returned by `date`
+	Version string
+}
diff --git a/pkg/sentry/limits/BUILD b/pkg/sentry/limits/BUILD
new file mode 100644
index 000000000..06c3e72b0
--- /dev/null
+++ b/pkg/sentry/limits/BUILD
@@ -0,0 +1,39 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "limits_state",
+    srcs = [
+        "limits.go",
+    ],
+    out = "limits_state.go",
+    package = "limits",
+)
+
+go_library(
+    name = "limits",
+    srcs = [
+        "context.go",
+        "limits.go",
+        "limits_state.go",
+        "linux.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/limits",
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/state",
+    ],
+)
+
+go_test(
+    name = "limits_test",
+    size = "small",
+    srcs = [
+        "limits_test.go",
+    ],
+    embed = [":limits"],
+)
diff --git a/pkg/sentry/limits/context.go b/pkg/sentry/limits/context.go
new file mode 100644
index 000000000..75e97bf92
--- /dev/null
+++ b/pkg/sentry/limits/context.go
@@ -0,0 +1,35 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package limits
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the limit package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxLimits is a Context.Value key for a LimitSet.
+	CtxLimits contextID = iota
+)
+
+// FromContext returns the limits that apply to ctx.
+func FromContext(ctx context.Context) *LimitSet {
+	if v := ctx.Value(CtxLimits); v != nil {
+		return v.(*LimitSet)
+	}
+	return nil
+}
diff --git a/pkg/sentry/limits/limits.go b/pkg/sentry/limits/limits.go
new file mode 100644
index 000000000..4230ba958
--- /dev/null
+++ b/pkg/sentry/limits/limits.go
@@ -0,0 +1,128 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package limits provides resource limits.
+package limits
+
+import (
+	"sync"
+	"syscall"
+)
+
+// LimitType defines a type of resource limit.
+type LimitType int
+
+// Set of constants defining the different types of resource limits.
+const (
+	CPU LimitType = iota
+	FileSize
+	Data
+	Stack
+	Core
+	Rss
+	ProcessCount
+	NumberOfFiles
+	MemoryPagesLocked
+	AS
+	Locks
+	SignalsPending
+	MessageQueueBytes
+	Nice
+	RealTimePriority
+	Rttime
+)
+
+// Infinity is a constant representing a resource with no limit.
+const Infinity = ^uint64(0)
+
+// Limit specifies a system limit.
+type Limit struct {
+	// Cur specifies the current limit.
+	Cur uint64
+	// Max specifies the maximum settable limit.
+	Max uint64
+}
+
+// LimitSet represents the Limits that correspond to each LimitType.
+type LimitSet struct {
+	mu   sync.Mutex `state:"nosave"`
+	data map[LimitType]Limit
+}
+
+// NewLimitSet creates a new, empty LimitSet.
+func NewLimitSet() *LimitSet {
+	return &LimitSet{
+		data: make(map[LimitType]Limit),
+	}
+}
+
+// GetCopy returns a clone of the LimitSet.
+func (l *LimitSet) GetCopy() *LimitSet {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	copyData := make(map[LimitType]Limit)
+	for k, v := range l.data {
+		copyData[k] = v
+	}
+	return &LimitSet{
+		data: copyData,
+	}
+}
+
+// Get returns the resource limit associated with LimitType t.
+// If no limit is provided, it defaults to an infinite limit.Infinity.
+func (l *LimitSet) Get(t LimitType) Limit {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	s, ok := l.data[t]
+	if !ok {
+		return Limit{Cur: Infinity, Max: Infinity}
+	}
+	return s
+}
+
+// GetCapped returns the current value for the limit, capped as specified.
+func (l *LimitSet) GetCapped(t LimitType, max uint64) uint64 {
+	s := l.Get(t)
+	if s.Cur == Infinity || s.Cur > max {
+		return max
+	}
+	return s.Cur
+}
+
+// SetUnchecked assigns value v to resource of LimitType t.
+func (l *LimitSet) SetUnchecked(t LimitType, v Limit) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	l.data[t] = v
+}
+
+// Set assigns value v to resource of LimitType t and returns the old value.
+func (l *LimitSet) Set(t LimitType, v Limit) (Limit, error) {
+	l.mu.Lock()
+	defer l.mu.Unlock()
+	// If a limit is already set, make sure the new limit doesn't
+	// exceed the previous max limit.
+	if _, ok := l.data[t]; ok {
+		if l.data[t].Max < v.Max {
+			return Limit{}, syscall.EPERM
+		}
+		if v.Cur > v.Max {
+			return Limit{}, syscall.EINVAL
+		}
+	}
+	old := l.data[t]
+	l.data[t] = v
+	return old, nil
+}
diff --git a/pkg/sentry/limits/limits_test.go b/pkg/sentry/limits/limits_test.go
new file mode 100644
index 000000000..dd6f80750
--- /dev/null
+++ b/pkg/sentry/limits/limits_test.go
@@ -0,0 +1,37 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package limits
+
+import (
+	"syscall"
+	"testing"
+)
+
+func TestSet(t *testing.T) {
+	ls := NewLimitSet()
+	ls.Set(1, Limit{Cur: 50, Max: 50})
+	if _, err := ls.Set(1, Limit{Cur: 20, Max: 50}); err != nil {
+		t.Fatalf("Tried to lower Limit to valid new value: got %v, wanted nil", err)
+	}
+	if _, err := ls.Set(1, Limit{Cur: 20, Max: 60}); err != syscall.EPERM {
+		t.Fatalf("Tried to raise limit.Max to invalid higher value: got %v, wanted syscall.EPERM", err)
+	}
+	if _, err := ls.Set(1, Limit{Cur: 60, Max: 50}); err != syscall.EINVAL {
+		t.Fatalf("Tried to raise limit.Cur to invalid higher value: got %v, wanted syscall.EINVAL", err)
+	}
+	if _, err := ls.Set(1, Limit{Cur: 11, Max: 10}); err != syscall.EINVAL {
+		t.Fatalf("Tried to set new limit with Cur > Max: got %v, wanted syscall.EINVAL", err)
+	}
+}
diff --git a/pkg/sentry/limits/linux.go b/pkg/sentry/limits/linux.go
new file mode 100644
index 000000000..8e6a24341
--- /dev/null
+++ b/pkg/sentry/limits/linux.go
@@ -0,0 +1,100 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package limits
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// FromLinuxResource maps linux resources to sentry LimitTypes.
+var FromLinuxResource = map[int]LimitType{
+	linux.RLIMIT_CPU:        CPU,
+	linux.RLIMIT_FSIZE:      FileSize,
+	linux.RLIMIT_DATA:       Data,
+	linux.RLIMIT_STACK:      Stack,
+	linux.RLIMIT_CORE:       Core,
+	linux.RLIMIT_RSS:        Rss,
+	linux.RLIMIT_NPROC:      ProcessCount,
+	linux.RLIMIT_NOFILE:     NumberOfFiles,
+	linux.RLIMIT_MEMLOCK:    MemoryPagesLocked,
+	linux.RLIMIT_AS:         AS,
+	linux.RLIMIT_LOCKS:      Locks,
+	linux.RLIMIT_SIGPENDING: SignalsPending,
+	linux.RLIMIT_MSGQUEUE:   MessageQueueBytes,
+	linux.RLIMIT_NICE:       Nice,
+	linux.RLIMIT_RTPRIO:     RealTimePriority,
+	linux.RLIMIT_RTTIME:     Rttime,
+}
+
+// FromLinux maps linux rlimit values to sentry Limits, being careful to handle
+// infinities.
+func FromLinux(rl uint64) uint64 {
+	if rl == linux.RLimInfinity {
+		return Infinity
+	}
+	return rl
+}
+
+// ToLinux maps sentry Limits to linux rlimit values, being careful to handle
+// infinities.
+func ToLinux(l uint64) uint64 {
+	if l == Infinity {
+		return linux.RLimInfinity
+	}
+	return l
+}
+
+// NewLinuxLimitSet returns a LimitSet whose values match the default rlimits
+// in Linux.
+func NewLinuxLimitSet() (*LimitSet, error) {
+	ls := NewLimitSet()
+	for rlt, rl := range linux.InitRLimits {
+		lt, ok := FromLinuxResource[rlt]
+		if !ok {
+			return nil, fmt.Errorf("unknown rlimit type %v", rlt)
+		}
+		ls.SetUnchecked(lt, Limit{
+			Cur: FromLinux(rl.Cur),
+			Max: FromLinux(rl.Max),
+		})
+	}
+	return ls, nil
+}
+
+// NewLinuxDistroLimitSet returns a new LimitSet whose values are typical
+// for a booted Linux distro.
+//
+// Many Linux init systems adjust the default Linux limits to values more
+// expected by the rest of the userspace. NewLinuxDistroLimitSet returns a
+// LimitSet with sensible defaults for applications that aren't starting
+// their own init system.
+func NewLinuxDistroLimitSet() (*LimitSet, error) {
+	ls, err := NewLinuxLimitSet()
+	if err != nil {
+		return nil, err
+	}
+
+	// Adjust ProcessCount to a lower value because GNU bash allocates 16
+	// bytes per proc and OOMs if this number is set too high. Value was
+	// picked arbitrarily.
+	//
+	// 1,048,576 ought to be enough for anyone.
+	l := ls.Get(ProcessCount)
+	l.Cur = 1 << 20
+	ls.Set(ProcessCount, l)
+	return ls, nil
+}
diff --git a/pkg/sentry/loader/BUILD b/pkg/sentry/loader/BUILD
new file mode 100644
index 000000000..917ec8cc8
--- /dev/null
+++ b/pkg/sentry/loader/BUILD
@@ -0,0 +1,59 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_embed_data", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_embed_data(
+    name = "vdso_bin",
+    src = "//vdso:vdso.so",
+    package = "loader",
+    var = "vdsoBin",
+)
+
+go_stateify(
+    name = "loader_state",
+    srcs = [
+        "vdso.go",
+        "vdso_state.go",
+    ],
+    out = "loader_state.go",
+    package = "loader",
+)
+
+go_library(
+    name = "loader",
+    srcs = [
+        "elf.go",
+        "interpreter.go",
+        "loader.go",
+        "loader_state.go",
+        "vdso.go",
+        "vdso_state.go",
+        ":vdso_bin",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/loader",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi",
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/cpuid",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/mm",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/uniqueid",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go
new file mode 100644
index 000000000..d23dc1096
--- /dev/null
+++ b/pkg/sentry/loader/elf.go
@@ -0,0 +1,637 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package loader
+
+import (
+	"bytes"
+	"debug/elf"
+	"fmt"
+	"io"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	// elfMagic identifies an ELF file.
+	elfMagic = "\x7fELF"
+
+	// maxTotalPhdrSize is the maximum combined size of all program
+	// headers.  Linux limits this to one page.
+	maxTotalPhdrSize = usermem.PageSize
+)
+
+var (
+	// header64Size is the size of elf.Header64.
+	header64Size = int(binary.Size(elf.Header64{}))
+
+	// Prog64Size is the size of elf.Prog64.
+	prog64Size = int(binary.Size(elf.Prog64{}))
+)
+
+func progFlagsAsPerms(f elf.ProgFlag) usermem.AccessType {
+	var p usermem.AccessType
+	if f&elf.PF_R == elf.PF_R {
+		p.Read = true
+	}
+	if f&elf.PF_W == elf.PF_W {
+		p.Write = true
+	}
+	if f&elf.PF_X == elf.PF_X {
+		p.Execute = true
+	}
+	return p
+}
+
+// elfInfo contains the metadata needed to load an ELF binary.
+type elfInfo struct {
+	// os is the target OS of the ELF.
+	os abi.OS
+
+	// arch is the target architecture of the ELF.
+	arch arch.Arch
+
+	// entry is the program entry point.
+	entry usermem.Addr
+
+	// phdrs are the program headers.
+	phdrs []elf.ProgHeader
+
+	// phdrSize is the size of a single program header in the ELF.
+	phdrSize int
+
+	// phdrOff is the offset of the program headers in the file.
+	phdrOff uint64
+
+	// sharedObject is true if the ELF represents a shared object.
+	sharedObject bool
+}
+
+// parseHeader parse the ELF header, verifying that this is a supported ELF
+// file and returning the ELF program headers.
+//
+// This is similar to elf.NewFile, except that it is more strict about what it
+// accepts from the ELF, and it doesn't parse unnecessary parts of the file.
+//
+// ctx may be nil if f does not need it.
+func parseHeader(ctx context.Context, f *fs.File) (elfInfo, error) {
+	// Check ident first; it will tell us the endianness of the rest of the
+	// structs.
+	var ident [elf.EI_NIDENT]byte
+	_, err := readFull(ctx, f, usermem.BytesIOSequence(ident[:]), 0)
+	if err != nil {
+		log.Infof("Error reading ELF ident: %v", err)
+		// The entire ident array always exists.
+		if err == io.EOF || err == io.ErrUnexpectedEOF {
+			err = syserror.ENOEXEC
+		}
+		return elfInfo{}, err
+	}
+
+	// Only some callers pre-check the ELF magic.
+	if !bytes.Equal(ident[:len(elfMagic)], []byte(elfMagic)) {
+		log.Infof("File is not an ELF")
+		return elfInfo{}, syserror.ENOEXEC
+	}
+
+	// We only support 64-bit, little endian binaries
+	if class := elf.Class(ident[elf.EI_CLASS]); class != elf.ELFCLASS64 {
+		log.Infof("Unsupported ELF class: %v", class)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	if endian := elf.Data(ident[elf.EI_DATA]); endian != elf.ELFDATA2LSB {
+		log.Infof("Unsupported ELF endianness: %v", endian)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	byteOrder := binary.LittleEndian
+
+	if version := elf.Version(ident[elf.EI_VERSION]); version != elf.EV_CURRENT {
+		log.Infof("Unsupported ELF version: %v", version)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	// EI_OSABI is ignored by Linux, which is the only OS supported.
+	os := abi.Linux
+
+	var hdr elf.Header64
+	hdrBuf := make([]byte, header64Size)
+	_, err = readFull(ctx, f, usermem.BytesIOSequence(hdrBuf), 0)
+	if err != nil {
+		log.Infof("Error reading ELF header: %v", err)
+		// The entire header always exists.
+		if err == io.EOF || err == io.ErrUnexpectedEOF {
+			err = syserror.ENOEXEC
+		}
+		return elfInfo{}, err
+	}
+	binary.Unmarshal(hdrBuf, byteOrder, &hdr)
+
+	// We only support amd64.
+	if machine := elf.Machine(hdr.Machine); machine != elf.EM_X86_64 {
+		log.Infof("Unsupported ELF machine %d", machine)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	a := arch.AMD64
+
+	var sharedObject bool
+	elfType := elf.Type(hdr.Type)
+	switch elfType {
+	case elf.ET_EXEC:
+		sharedObject = false
+	case elf.ET_DYN:
+		sharedObject = true
+	default:
+		log.Infof("Unsupported ELF type %v", elfType)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+
+	if int(hdr.Phentsize) != prog64Size {
+		log.Infof("Unsupported phdr size %d", hdr.Phentsize)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	totalPhdrSize := prog64Size * int(hdr.Phnum)
+	if totalPhdrSize < prog64Size {
+		log.Warningf("No phdrs or total phdr size overflows: prog64Size: %d phnum: %d", prog64Size, int(hdr.Phnum))
+		return elfInfo{}, syserror.ENOEXEC
+	}
+	if totalPhdrSize > maxTotalPhdrSize {
+		log.Infof("Too many phdrs (%d): total size %d > %d", hdr.Phnum, totalPhdrSize, maxTotalPhdrSize)
+		return elfInfo{}, syserror.ENOEXEC
+	}
+
+	phdrBuf := make([]byte, totalPhdrSize)
+	_, err = readFull(ctx, f, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff))
+	if err != nil {
+		log.Infof("Error reading ELF phdrs: %v", err)
+		// If phdrs were specified, they should all exist.
+		if err == io.EOF || err == io.ErrUnexpectedEOF {
+			err = syserror.ENOEXEC
+		}
+		return elfInfo{}, err
+	}
+
+	phdrs := make([]elf.ProgHeader, hdr.Phnum)
+	for i := range phdrs {
+		var prog64 elf.Prog64
+		binary.Unmarshal(phdrBuf[:prog64Size], byteOrder, &prog64)
+		phdrBuf = phdrBuf[prog64Size:]
+		phdrs[i] = elf.ProgHeader{
+			Type:   elf.ProgType(prog64.Type),
+			Flags:  elf.ProgFlag(prog64.Flags),
+			Off:    prog64.Off,
+			Vaddr:  prog64.Vaddr,
+			Paddr:  prog64.Paddr,
+			Filesz: prog64.Filesz,
+			Memsz:  prog64.Memsz,
+			Align:  prog64.Align,
+		}
+	}
+
+	return elfInfo{
+		os:           os,
+		arch:         a,
+		entry:        usermem.Addr(hdr.Entry),
+		phdrs:        phdrs,
+		phdrOff:      hdr.Phoff,
+		phdrSize:     prog64Size,
+		sharedObject: sharedObject,
+	}, nil
+}
+
+// mapSegment maps a phdr into the Task. offset is the offset to apply to
+// phdr.Vaddr.
+func mapSegment(ctx context.Context, m *mm.MemoryManager, f *fs.File, phdr *elf.ProgHeader, offset usermem.Addr) error {
+	// Alignment of vaddr and offset must match. We'll need to map on the
+	// page boundary.
+	adjust := usermem.Addr(phdr.Vaddr).PageOffset()
+	if adjust != usermem.Addr(phdr.Off).PageOffset() {
+		ctx.Infof("Alignment of vaddr %#x != off %#x", phdr.Vaddr, phdr.Off)
+		return syserror.ENOEXEC
+	}
+
+	addr, ok := offset.AddLength(phdr.Vaddr)
+	if !ok {
+		// If offset != 0 we should have ensured this would fit.
+		ctx.Warningf("Computed segment load address overflows: %#x + %#x", phdr.Vaddr, offset)
+		return syserror.ENOEXEC
+	}
+	addr -= usermem.Addr(adjust)
+
+	fileOffset := phdr.Off - adjust
+	fileSize := phdr.Filesz + adjust
+	if fileSize < phdr.Filesz {
+		ctx.Infof("Computed segment file size overflows: %#x + %#x", phdr.Filesz, adjust)
+		return syserror.ENOEXEC
+	}
+	memSize := phdr.Memsz + adjust
+	if memSize < phdr.Memsz {
+		ctx.Infof("Computed segment mem size overflows: %#x + %#x", phdr.Memsz, adjust)
+		return syserror.ENOEXEC
+	}
+	ms, ok := usermem.Addr(fileSize).RoundUp()
+	if !ok {
+		ctx.Infof("fileSize %#x too large", fileSize)
+		return syserror.ENOEXEC
+	}
+	mapSize := uint64(ms)
+
+	prot := progFlagsAsPerms(phdr.Flags)
+	mopts := memmap.MMapOpts{
+		Length: mapSize,
+		Offset: fileOffset,
+		Addr:   addr,
+		Fixed:  true,
+		// Linux will happily allow conflicting segments to map over
+		// one another.
+		Unmap:    true,
+		Private:  true,
+		Perms:    prot,
+		MaxPerms: usermem.AnyAccess,
+	}
+	if err := f.ConfigureMMap(ctx, &mopts); err != nil {
+		ctx.Infof("File is not memory-mappable: %v", err)
+		return err
+	}
+	if _, err := m.MMap(ctx, mopts); err != nil {
+		ctx.Infof("Error mapping PT_LOAD segment %+v at %#x: %v", phdr, addr, err)
+		return err
+	}
+
+	// We need to clear the end of the last page that exceeds fileSize so
+	// we don't map part of the file beyond fileSize.
+	//
+	// Note that Linux *does not* clear the portion of the first page
+	// before phdr.Off.
+	if mapSize > fileSize {
+		zeroAddr, ok := addr.AddLength(fileSize)
+		if !ok {
+			panic(fmt.Sprintf("successfully mmaped address overflows? %#x + %#x", addr, fileSize))
+		}
+		zeroSize := int64(mapSize - fileSize)
+		if zeroSize < 0 {
+			panic(fmt.Sprintf("zeroSize too big? %#x", uint64(zeroSize)))
+		}
+		if _, err := m.ZeroOut(ctx, zeroAddr, zeroSize, usermem.IOOpts{IgnorePermissions: true}); err != nil {
+			ctx.Warningf("Failed to zero end of page [%#x, %#x): %v", zeroAddr, zeroAddr+usermem.Addr(zeroSize), err)
+			return err
+		}
+	}
+
+	// Allocate more anonymous pages if necessary.
+	if mapSize < memSize {
+		anonAddr, ok := addr.AddLength(mapSize)
+		if !ok {
+			panic(fmt.Sprintf("anonymous memory doesn't fit in pre-sized range? %#x + %#x", addr, mapSize))
+		}
+		anonSize, ok := usermem.Addr(memSize - mapSize).RoundUp()
+		if !ok {
+			ctx.Infof("extra anon pages too large: %#x", memSize-mapSize)
+			return syserror.ENOEXEC
+		}
+
+		if _, err := m.MMap(ctx, memmap.MMapOpts{
+			Length: uint64(anonSize),
+			Addr:   anonAddr,
+			// Fixed without Unmap will fail the mmap if something is
+			// already at addr.
+			Fixed:    true,
+			Private:  true,
+			Perms:    progFlagsAsPerms(phdr.Flags),
+			MaxPerms: usermem.AnyAccess,
+		}); err != nil {
+			ctx.Infof("Error mapping PT_LOAD segment %v anonymous memory: %v", phdr, err)
+			return err
+		}
+	}
+
+	return nil
+}
+
+// loadedELF describes an ELF that has been successfully loaded.
+type loadedELF struct {
+	// os is the target OS of the ELF.
+	os abi.OS
+
+	// arch is the target architecture of the ELF.
+	arch arch.Arch
+
+	// entry is the entry point of the ELF.
+	entry usermem.Addr
+
+	// start is the end of the ELF.
+	start usermem.Addr
+
+	// end is the end of the ELF.
+	end usermem.Addr
+
+	// interpter is the path to the ELF interpreter.
+	interpreter string
+
+	// phdrAddr is the address of the ELF program headers.
+	phdrAddr usermem.Addr
+
+	// phdrSize is the size of a single program header in the ELF.
+	phdrSize int
+
+	// phdrNum is the number of program headers.
+	phdrNum int
+
+	// auxv contains a subset of ELF-specific auxiliary vector entries:
+	// * AT_PHDR
+	// * AT_PHENT
+	// * AT_PHNUM
+	// * AT_BASE
+	// * AT_ENTRY
+	auxv arch.Auxv
+}
+
+// loadParsedELF loads f into mm.
+//
+// info is the parsed elfInfo from the header.
+//
+// It does not load the ELF interpreter, or return any auxv entries.
+//
+// Preconditions:
+//  * f is an ELF file
+func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) {
+	first := true
+	var start, end usermem.Addr
+	var interpreter string
+	for _, phdr := range info.phdrs {
+		switch phdr.Type {
+		case elf.PT_LOAD:
+			vaddr := usermem.Addr(phdr.Vaddr)
+			if first {
+				first = false
+				start = vaddr
+			}
+			if vaddr < end {
+				ctx.Infof("PT_LOAD headers out-of-order. %#x < %#x", vaddr, end)
+				return loadedELF{}, syserror.ENOEXEC
+			}
+			var ok bool
+			end, ok = vaddr.AddLength(phdr.Memsz)
+			if !ok {
+				ctx.Infof("PT_LOAD header size overflows. %#x + %#x", vaddr, phdr.Memsz)
+				return loadedELF{}, syserror.ENOEXEC
+			}
+
+		case elf.PT_INTERP:
+			if phdr.Filesz > syscall.PathMax {
+				ctx.Infof("PT_INTERP path too big: %v", phdr.Filesz)
+				return loadedELF{}, syserror.ENOEXEC
+			}
+
+			path := make([]byte, phdr.Filesz)
+			_, err := readFull(ctx, f, usermem.BytesIOSequence(path), int64(phdr.Off))
+			if err != nil {
+				ctx.Infof("Error reading PT_INTERP path: %v", err)
+				// If an interpreter was specified, it should exist.
+				if err == io.EOF || err == io.ErrUnexpectedEOF {
+					err = syserror.ENOEXEC
+				}
+				return loadedELF{}, syserror.ENOEXEC
+			}
+
+			if path[len(path)-1] != 0 {
+				ctx.Infof("PT_INTERP path not NUL-terminated: %v", path)
+				return loadedELF{}, syserror.ENOEXEC
+			}
+
+			// Strip NUL-terminator from string.
+			interpreter = string(path[:len(path)-1])
+		}
+	}
+
+	// Shared objects don't have fixed load addresses. We need to pick a
+	// base address big enough to fit all segments, so we first create a
+	// mapping for the total size just to find a region that is big enough.
+	//
+	// It is safe to unmap it immediately with racing with another mapping
+	// because we are the only one in control of the MemoryManager.
+	//
+	// Note that the vaddr of the first PT_LOAD segment is ignored when
+	// choosing the load address (even if it is non-zero). The vaddr does
+	// become an offset from that load address.
+	var offset usermem.Addr
+	if info.sharedObject {
+		totalSize := end - start
+		totalSize, ok := totalSize.RoundUp()
+		if !ok {
+			ctx.Infof("ELF PT_LOAD segments too big")
+			return loadedELF{}, syserror.ENOEXEC
+		}
+
+		var err error
+		offset, err = m.MMap(ctx, memmap.MMapOpts{
+			Length:  uint64(totalSize),
+			Addr:    sharedLoadOffset,
+			Private: true,
+		})
+		if err != nil {
+			ctx.Infof("Error allocating address space for shared object: %v", err)
+			return loadedELF{}, err
+		}
+		if err := m.MUnmap(ctx, offset, uint64(totalSize)); err != nil {
+			panic(fmt.Sprintf("Failed to unmap base address: %v", err))
+		}
+
+		start, ok = start.AddLength(uint64(offset))
+		if !ok {
+			panic(fmt.Sprintf("Start %#x + offset %#x overflows?", start, offset))
+		}
+
+		end, ok = end.AddLength(uint64(offset))
+		if !ok {
+			panic(fmt.Sprintf("End %#x + offset %#x overflows?", end, offset))
+		}
+
+		info.entry, ok = info.entry.AddLength(uint64(offset))
+		if !ok {
+			ctx.Infof("Entrypoint %#x + offset %#x overflows? Is the entrypoint within a segment?", info.entry, offset)
+			return loadedELF{}, err
+		}
+	}
+
+	// Map PT_LOAD segments.
+	for _, phdr := range info.phdrs {
+		switch phdr.Type {
+		case elf.PT_LOAD:
+			if phdr.Memsz == 0 {
+				// No need to load segments with size 0, but
+				// they exist in some binaries.
+				continue
+			}
+
+			if err := mapSegment(ctx, m, f, &phdr, offset); err != nil {
+				ctx.Infof("Failed to map PT_LOAD segment: %+v", phdr)
+				return loadedELF{}, err
+			}
+		}
+	}
+
+	// This assumes that the first segment contains the ELF headers. This
+	// may not be true in a malformed ELF, but Linux makes the same
+	// assumption.
+	phdrAddr, ok := start.AddLength(info.phdrOff)
+	if !ok {
+		ctx.Warningf("ELF start address %#x + phdr offset %#x overflows", start, info.phdrOff)
+		phdrAddr = 0
+	}
+
+	return loadedELF{
+		os:          info.os,
+		arch:        info.arch,
+		entry:       info.entry,
+		start:       start,
+		end:         end,
+		interpreter: interpreter,
+		phdrAddr:    phdrAddr,
+		phdrSize:    info.phdrSize,
+		phdrNum:     len(info.phdrs),
+	}, nil
+}
+
+// loadInitialELF loads f into mm.
+//
+// It creates an arch.Context for the ELF and prepares the mm for this arch.
+//
+// It does not load the ELF interpreter, or return any auxv entries.
+//
+// Preconditions:
+//  * f is an ELF file
+//  * f is the first ELF loaded into m
+func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureSet, f *fs.File) (loadedELF, arch.Context, error) {
+	info, err := parseHeader(ctx, f)
+	if err != nil {
+		ctx.Infof("Failed to parse initial ELF: %v", err)
+		return loadedELF{}, nil, err
+	}
+
+	// Create the arch.Context now so we can prepare the mmap layout before
+	// mapping anything.
+	ac := arch.New(info.arch, fs)
+
+	l, err := m.SetMmapLayout(ac, limits.FromContext(ctx))
+	if err != nil {
+		ctx.Warningf("Failed to set mmap layout: %v", err)
+		return loadedELF{}, nil, err
+	}
+
+	// PIELoadAddress tries to move the ELF out of the way of the default
+	// mmap base to ensure that the initial brk has sufficient space to
+	// grow.
+	le, err := loadParsedELF(ctx, m, f, info, ac.PIELoadAddress(l))
+	return le, ac, err
+}
+
+// loadInterpreterELF loads f into mm.
+//
+// The interpreter must be for the same OS/Arch as the initial ELF.
+//
+// It does not return any auxv entries.
+//
+// Preconditions:
+//  * f is an ELF file
+func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f *fs.File, initial loadedELF) (loadedELF, error) {
+	info, err := parseHeader(ctx, f)
+	if err != nil {
+		if err == syserror.ENOEXEC {
+			// Bad interpreter.
+			err = syserror.ELIBBAD
+		}
+		return loadedELF{}, err
+	}
+
+	if info.os != initial.os {
+		ctx.Infof("Initial ELF OS %v and interpreter ELF OS %v differ", initial.os, info.os)
+		return loadedELF{}, syserror.ELIBBAD
+	}
+	if info.arch != initial.arch {
+		ctx.Infof("Initial ELF arch %v and interpreter ELF arch %v differ", initial.arch, info.arch)
+		return loadedELF{}, syserror.ELIBBAD
+	}
+
+	// The interpreter is not given a load offset, as its location does not
+	// affect brk.
+	return loadParsedELF(ctx, m, f, info, 0)
+}
+
+// loadELF loads f into the Task address space.
+//
+// If loadELF returns ErrSwitchFile it should be called again with the returned
+// path and argv.
+//
+// Preconditions:
+//  * f is an ELF file
+func loadELF(ctx context.Context, m *mm.MemoryManager, mounts *fs.MountNamespace, root, wd *fs.Dirent, maxTraversals uint, fs *cpuid.FeatureSet, f *fs.File) (loadedELF, arch.Context, error) {
+	bin, ac, err := loadInitialELF(ctx, m, fs, f)
+	if err != nil {
+		ctx.Infof("Error loading binary: %v", err)
+		return loadedELF{}, nil, err
+	}
+
+	var interp loadedELF
+	if bin.interpreter != "" {
+		d, i, err := openPath(ctx, mounts, root, wd, maxTraversals, bin.interpreter)
+		if err != nil {
+			ctx.Infof("Error opening interpreter %s: %v", bin.interpreter, err)
+			return loadedELF{}, nil, err
+		}
+		defer i.DecRef()
+		// We don't need the Dirent.
+		d.DecRef()
+
+		interp, err = loadInterpreterELF(ctx, m, i, bin)
+		if err != nil {
+			ctx.Infof("Error loading interpreter: %v", err)
+			return loadedELF{}, nil, err
+		}
+
+		if interp.interpreter != "" {
+			// No recursive interpreters!
+			ctx.Infof("Interpreter requires an interpreter")
+			return loadedELF{}, nil, syserror.ENOEXEC
+		}
+	}
+
+	// ELF-specific auxv entries.
+	bin.auxv = arch.Auxv{
+		arch.AuxEntry{linux.AT_PHDR, bin.phdrAddr},
+		arch.AuxEntry{linux.AT_PHENT, usermem.Addr(bin.phdrSize)},
+		arch.AuxEntry{linux.AT_PHNUM, usermem.Addr(bin.phdrNum)},
+		arch.AuxEntry{linux.AT_ENTRY, bin.entry},
+	}
+	if bin.interpreter != "" {
+		bin.auxv = append(bin.auxv, arch.AuxEntry{linux.AT_BASE, interp.start})
+
+		// Start in the interpreter.
+		// N.B. AT_ENTRY above contains the *original* entry point.
+		bin.entry = interp.entry
+	}
+
+	return bin, ac, nil
+}
diff --git a/pkg/sentry/loader/interpreter.go b/pkg/sentry/loader/interpreter.go
new file mode 100644
index 000000000..b8ecbe92f
--- /dev/null
+++ b/pkg/sentry/loader/interpreter.go
@@ -0,0 +1,105 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package loader
+
+import (
+	"bytes"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	// interpreterScriptMagic identifies an interpreter script.
+	interpreterScriptMagic = "#!"
+
+	// interpMaxLineLength is the maximum length for the first line of an
+	// interpreter script.
+	//
+	// From execve(2): "A maximum line length of 127 characters is allowed
+	// for the first line in a #! executable shell script."
+	interpMaxLineLength = 127
+)
+
+// parseInterpreterScript returns the interpreter path and argv.
+func parseInterpreterScript(ctx context.Context, filename string, f *fs.File, argv, envv []string) (newpath string, newargv []string, err error) {
+	line := make([]byte, interpMaxLineLength)
+	n, err := readFull(ctx, f, usermem.BytesIOSequence(line), 0)
+	// Short read is OK.
+	if err != nil && err != io.ErrUnexpectedEOF {
+		if err == io.EOF {
+			err = syserror.ENOEXEC
+		}
+		return "", []string{}, err
+	}
+	line = line[:n]
+
+	if !bytes.Equal(line[:2], []byte(interpreterScriptMagic)) {
+		return "", []string{}, syserror.ENOEXEC
+	}
+	// Ignore #!.
+	line = line[2:]
+
+	// Ignore everything after newline.
+	// Linux silently truncates the remainder of the line if it exceeds
+	// interpMaxLineLength.
+	i := bytes.IndexByte(line, '\n')
+	if i > 0 {
+		line = line[:i]
+	}
+
+	// Skip any whitespace before the interpeter.
+	line = bytes.TrimLeft(line, " \t")
+
+	// Linux only looks for a space or tab delimiting the interpreter and
+	// arg.
+	//
+	// execve(2): "On Linux, the entire string following the interpreter
+	// name is passed as a single argument to the interpreter, and this
+	// string can include white space."
+	interp := line
+	var arg []byte
+	i = bytes.IndexAny(line, " \t")
+	if i >= 0 {
+		interp = line[:i]
+		if i+1 < len(line) {
+			arg = line[i+1:]
+		}
+	}
+
+	// Build the new argument list:
+	//
+	// 1. The interpreter.
+	newargv = append(newargv, string(interp))
+
+	// 2. The optional interpreter argument.
+	if len(arg) > 0 {
+		newargv = append(newargv, string(arg))
+	}
+
+	// 3. The original arguments. The original argv[0] is replaced with the
+	// full script filename.
+	if len(argv) > 0 {
+		argv[0] = filename
+	} else {
+		argv = []string{filename}
+	}
+	newargv = append(newargv, argv...)
+
+	return string(interp), newargv, nil
+}
diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go
new file mode 100644
index 000000000..94c281b72
--- /dev/null
+++ b/pkg/sentry/loader/loader.go
@@ -0,0 +1,277 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package loader loads a binary into a MemoryManager.
+package loader
+
+import (
+	"bytes"
+	"crypto/rand"
+	"io"
+	"path"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// readFull behaves like io.ReadFull for an *fs.File.
+func readFull(ctx context.Context, f *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	var total int64
+	for dst.NumBytes() > 0 {
+		n, err := f.Preadv(ctx, dst, offset+total)
+		total += n
+		if err == io.EOF && total != 0 {
+			return total, io.ErrUnexpectedEOF
+		} else if err != nil {
+			return total, err
+		}
+		dst = dst.DropFirst64(n)
+	}
+	return total, nil
+}
+
+// openPath opens name for loading.
+//
+// openPath returns the fs.Dirent and an *fs.File for name, which is not
+// installed in the Task FDMap. The caller takes ownership of both.
+//
+// name must be a readable, executable, regular file.
+func openPath(ctx context.Context, mm *fs.MountNamespace, root, wd *fs.Dirent, maxTraversals uint, name string) (*fs.Dirent, *fs.File, error) {
+	d, err := mm.FindInode(ctx, root, wd, name, maxTraversals)
+	if err != nil {
+		return nil, nil, err
+	}
+	defer d.DecRef()
+
+	perms := fs.PermMask{
+		// TODO: Linux requires only execute permission,
+		// not read. However, our backing filesystems may prevent us
+		// from reading the file without read permission.
+		//
+		// Additionally, a task with a non-readable executable has
+		// additional constraints on access via ptrace and procfs.
+		Read:    true,
+		Execute: true,
+	}
+	if err := d.Inode.CheckPermission(ctx, perms); err != nil {
+		return nil, nil, err
+	}
+
+	// If they claim it's a directory, then make sure.
+	//
+	// N.B. we reject directories below, but we must first reject
+	// non-directories passed as directories.
+	if len(name) > 0 && name[len(name)-1] == '/' && !fs.IsDir(d.Inode.StableAttr) {
+		return nil, nil, syserror.ENOTDIR
+	}
+
+	// No exec-ing directories, pipes, etc!
+	if !fs.IsRegular(d.Inode.StableAttr) {
+		ctx.Infof("Error regularing %s: %v", name, d.Inode.StableAttr)
+		return nil, nil, syserror.EACCES
+	}
+
+	// Create a new file.
+	file, err := d.Inode.GetFile(ctx, d, fs.FileFlags{Read: true})
+	if err != nil {
+		return nil, nil, err
+	}
+
+	// We must be able to read at arbitrary offsets.
+	if !file.Flags().Pread {
+		file.DecRef()
+		ctx.Infof("%s cannot be read at an offset: %+v", name, file.Flags())
+		return nil, nil, syserror.EACCES
+	}
+
+	// Grab a reference for the caller.
+	d.IncRef()
+	return d, file, nil
+}
+
+// allocStack allocates and maps a stack in to any available part of the address space.
+func allocStack(ctx context.Context, m *mm.MemoryManager, a arch.Context) (*arch.Stack, error) {
+	ar, err := m.MapStack(ctx)
+	if err != nil {
+		return nil, err
+	}
+	return &arch.Stack{a, m, ar.End}, nil
+}
+
+const (
+	// maxLoaderAttempts is the maximum number of attempts to try to load
+	// an interpreter scripts, to prevent loops. 6 (inital + 5 changes) is
+	// what the Linux kernel allows (fs/exec.c:search_binary_handler).
+	maxLoaderAttempts = 6
+)
+
+// loadPath resolves filename to a binary and loads it.
+//
+// It returns:
+//  * loadedELF, description of the loaded binary
+//  * arch.Context matching the binary arch
+//  * fs.Dirent of the binary file
+//  * Possibly updated argv
+func loadPath(ctx context.Context, m *mm.MemoryManager, mounts *fs.MountNamespace, root, wd *fs.Dirent, maxTraversals uint, fs *cpuid.FeatureSet, filename string, argv, envv []string) (loadedELF, arch.Context, *fs.Dirent, []string, error) {
+	for i := 0; i < maxLoaderAttempts; i++ {
+		d, f, err := openPath(ctx, mounts, root, wd, maxTraversals, filename)
+		if err != nil {
+			ctx.Infof("Error opening %s: %v", filename, err)
+			return loadedELF{}, nil, nil, nil, err
+		}
+		defer f.DecRef()
+		// We will return d in the successful case, but defer a DecRef
+		// for intermediate loops and failure cases.
+		defer d.DecRef()
+
+		// Check the header. Is this an ELF or interpreter script?
+		var hdr [4]uint8
+		// N.B. We assume that reading from a regular file cannot block.
+		_, err = readFull(ctx, f, usermem.BytesIOSequence(hdr[:]), 0)
+		// Allow unexpected EOF, as a valid executable could be only three
+		// bytes (e.g., #!a).
+		if err != nil && err != io.ErrUnexpectedEOF {
+			if err == io.EOF {
+				err = syserror.ENOEXEC
+			}
+			return loadedELF{}, nil, nil, nil, err
+		}
+
+		switch {
+		case bytes.Equal(hdr[:], []byte(elfMagic)):
+			loaded, ac, err := loadELF(ctx, m, mounts, root, wd, maxTraversals, fs, f)
+			if err != nil {
+				ctx.Infof("Error loading ELF: %v", err)
+				return loadedELF{}, nil, nil, nil, err
+			}
+			// An ELF is always terminal. Hold on to d.
+			d.IncRef()
+			return loaded, ac, d, argv, err
+		case bytes.Equal(hdr[:2], []byte(interpreterScriptMagic)):
+			newpath, newargv, err := parseInterpreterScript(ctx, filename, f, argv, envv)
+			if err != nil {
+				ctx.Infof("Error loading interpreter script: %v", err)
+				return loadedELF{}, nil, nil, nil, err
+			}
+			filename = newpath
+			argv = newargv
+		default:
+			ctx.Infof("Unknown magic: %v", hdr)
+			return loadedELF{}, nil, nil, nil, syserror.ENOEXEC
+		}
+	}
+
+	return loadedELF{}, nil, nil, nil, syserror.ELOOP
+}
+
+// Load loads filename into a MemoryManager.
+//
+// If Load returns ErrSwitchFile it should be called again with the returned
+// path and argv.
+//
+// Preconditions:
+//  * The Task MemoryManager is empty.
+//  * Load is called on the Task goroutine.
+func Load(ctx context.Context, m *mm.MemoryManager, mounts *fs.MountNamespace, root, wd *fs.Dirent, maxTraversals uint, fs *cpuid.FeatureSet, filename string, argv, envv []string, extraAuxv []arch.AuxEntry, vdso *VDSO) (abi.OS, arch.Context, string, error) {
+	// Load the binary itself.
+	loaded, ac, d, argv, err := loadPath(ctx, m, mounts, root, wd, maxTraversals, fs, filename, argv, envv)
+	if err != nil {
+		ctx.Infof("Failed to load %s: %v", filename, err)
+		return 0, nil, "", err
+	}
+	defer d.DecRef()
+
+	// Load the VDSO.
+	vdsoAddr, err := loadVDSO(ctx, m, vdso, loaded)
+	if err != nil {
+		ctx.Infof("Error loading VDSO: %v", err)
+		return 0, nil, "", err
+	}
+
+	// Setup the heap. brk starts at the next page after the end of the
+	// binary. Userspace can assume that the remainer of the page after
+	// loaded.end is available for its use.
+	e, ok := loaded.end.RoundUp()
+	if !ok {
+		ctx.Warningf("brk overflows: %#x", loaded.end)
+		return 0, nil, "", syserror.ENOEXEC
+	}
+	m.BrkSetup(ctx, e)
+
+	// Allocate our stack.
+	stack, err := allocStack(ctx, m, ac)
+	if err != nil {
+		ctx.Infof("Failed to allocate stack: %v", err)
+		return 0, nil, "", err
+	}
+
+	// Push the original filename to the stack, for AT_EXECFN.
+	execfn, err := stack.Push(filename)
+	if err != nil {
+		ctx.Infof("Failed to push exec filename: %v", err)
+		return 0, nil, "", err
+	}
+
+	// Push 16 random bytes on the stack which AT_RANDOM will point to.
+	var b [16]byte
+	if _, err := rand.Read(b[:]); err != nil {
+		ctx.Infof("Failed to read random bytes: %v", err)
+		return 0, nil, "", err
+	}
+	random, err := stack.Push(b)
+	if err != nil {
+		ctx.Infof("Failed to push random bytes: %v", err)
+		return 0, nil, "", err
+	}
+
+	// Add generic auxv entries
+	auxv := append(loaded.auxv, arch.Auxv{
+		arch.AuxEntry{linux.AT_CLKTCK, linux.CLOCKS_PER_SEC},
+		arch.AuxEntry{linux.AT_EXECFN, execfn},
+		arch.AuxEntry{linux.AT_RANDOM, random},
+		arch.AuxEntry{linux.AT_PAGESZ, usermem.PageSize},
+		arch.AuxEntry{linux.AT_SYSINFO_EHDR, vdsoAddr},
+	}...)
+	auxv = append(auxv, extraAuxv...)
+
+	sl, err := stack.Load(argv, envv, auxv)
+	if err != nil {
+		ctx.Infof("Failed to load stack: %v", err)
+		return 0, nil, "", err
+	}
+
+	m.SetArgvStart(sl.ArgvStart)
+	m.SetArgvEnd(sl.ArgvEnd)
+	m.SetEnvvStart(sl.EnvvStart)
+	m.SetEnvvEnd(sl.EnvvEnd)
+	m.SetAuxv(auxv)
+	m.SetExecutable(d)
+
+	ac.SetIP(uintptr(loaded.entry))
+	ac.SetStack(uintptr(stack.Bottom))
+
+	name := path.Base(filename)
+	if len(name) > linux.TASK_COMM_LEN-1 {
+		name = name[:linux.TASK_COMM_LEN-1]
+	}
+
+	return loaded.os, ac, name, nil
+}
diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go
new file mode 100644
index 000000000..ce4f6f5d9
--- /dev/null
+++ b/pkg/sentry/loader/vdso.go
@@ -0,0 +1,382 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package loader
+
+import (
+	"debug/elf"
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// byteReaderFileOperations implements fs.FileOperations for reading
+// from a []byte source.
+type byteReader struct {
+	fsutil.NoopRelease
+	fsutil.PipeSeek
+	fsutil.NotDirReaddir
+	fsutil.NoFsync
+	fsutil.NoopFlush
+	fsutil.NoMMap
+	fsutil.NoIoctl
+	waiter.AlwaysReady
+	data []byte
+}
+
+type fileContext struct {
+	context.Context
+}
+
+func (f *fileContext) Value(key interface{}) interface{} {
+	switch key {
+	case uniqueid.CtxGlobalUniqueID:
+		return uint64(0)
+	default:
+		return f.Context.Value(key)
+	}
+}
+
+func newByteReaderFile(data []byte) *fs.File {
+	dirent := fs.NewTransientDirent(nil)
+	flags := fs.FileFlags{Read: true, Pread: true}
+	return fs.NewFile(&fileContext{Context: context.Background()}, dirent, flags, &byteReader{
+		data: data,
+	})
+}
+
+func (b *byteReader) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	if offset < 0 {
+		return 0, syserror.EINVAL
+	}
+	if offset >= int64(len(b.data)) {
+		return 0, io.EOF
+	}
+	n, err := dst.CopyOut(ctx, b.data[offset:])
+	return int64(n), err
+}
+
+func (b *byteReader) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	panic("Write not supported")
+}
+
+// validateVDSO checks that the VDSO can be loaded by loadVDSO.
+//
+// VDSOs are special (see below). Since we are going to map the VDSO directly
+// rather than using a normal loading process, we require that the PT_LOAD
+// segments have the same layout in the ELF as they expect to have in memory.
+//
+// Namely, this means that we must verify:
+// * PT_LOAD file offsets are equivalent to the memory offset from the first
+//   segment.
+// * No extra zeroed space (memsz) is required.
+// * PT_LOAD segments are in order.
+// * No two PT_LOAD segments occupy parts of the same page.
+// * PT_LOAD segments don't extend beyond the end of the file.
+//
+// ctx may be nil if f does not need it.
+func validateVDSO(ctx context.Context, f *fs.File, size uint64) (elfInfo, error) {
+	info, err := parseHeader(ctx, f)
+	if err != nil {
+		log.Infof("Unable to parse VDSO header: %v", err)
+		return elfInfo{}, err
+	}
+
+	var first *elf.ProgHeader
+	var prev *elf.ProgHeader
+	var prevEnd usermem.Addr
+	for i, phdr := range info.phdrs {
+		if phdr.Type != elf.PT_LOAD {
+			continue
+		}
+
+		if first == nil {
+			first = &info.phdrs[i]
+			if phdr.Off != 0 {
+				log.Warningf("First PT_LOAD segment has non-zero file offset")
+				return elfInfo{}, syserror.ENOEXEC
+			}
+		}
+
+		memoryOffset := phdr.Vaddr - first.Vaddr
+		if memoryOffset != phdr.Off {
+			log.Warningf("PT_LOAD segment memory offset %#x != file offset %#x", memoryOffset, phdr.Off)
+			return elfInfo{}, syserror.ENOEXEC
+		}
+
+		// memsz larger than filesz means that extra zeroed space should be
+		// provided at the end of the segment. Since we are mapping the ELF
+		// directly, we don't want to just overwrite part of the ELF with
+		// zeroes.
+		if phdr.Memsz != phdr.Filesz {
+			log.Warningf("PT_LOAD segment memsz %#x != filesz %#x", phdr.Memsz, phdr.Filesz)
+			return elfInfo{}, syserror.ENOEXEC
+		}
+
+		start := usermem.Addr(memoryOffset)
+		end, ok := start.AddLength(phdr.Memsz)
+		if !ok {
+			log.Warningf("PT_LOAD segment size overflows: %#x + %#x", start, end)
+			return elfInfo{}, syserror.ENOEXEC
+		}
+		if uint64(end) > size {
+			log.Warningf("PT_LOAD segment end %#x extends beyond end of file %#x", end, size)
+			return elfInfo{}, syserror.ENOEXEC
+		}
+
+		if prev != nil {
+			if start < prevEnd {
+				log.Warningf("PT_LOAD segments out of order")
+				return elfInfo{}, syserror.ENOEXEC
+			}
+
+			// We mprotect entire pages, so each segment must be in
+			// its own page.
+			prevEndPage := prevEnd.RoundDown()
+			startPage := start.RoundDown()
+			if prevEndPage >= startPage {
+				log.Warningf("PT_LOAD segments share a page: %#x", prevEndPage)
+				return elfInfo{}, syserror.ENOEXEC
+			}
+		}
+		prev = &info.phdrs[i]
+		prevEnd = end
+	}
+
+	return info, nil
+}
+
+// VDSO describes a VDSO.
+//
+// NOTE: to support multiple architectures or operating systems, this
+// would need to contain a VDSO for each.
+type VDSO struct {
+	// ParamPage is the VDSO parameter page. This page should be updated to
+	// inform the VDSO for timekeeping data.
+	ParamPage *mm.SpecialMappable
+
+	// vdso is the VDSO ELF itself.
+	vdso *mm.SpecialMappable
+
+	// os is the operating system targeted by the VDSO.
+	os abi.OS
+
+	// arch is the architecture targeted by the VDSO.
+	arch arch.Arch
+
+	// phdrs are the VDSO ELF phdrs.
+	phdrs []elf.ProgHeader `state:".([]elfProgHeader)"`
+}
+
+// PrepareVDSO validates the system VDSO and returns a VDSO, containing the
+// param page for updating by the kernel.
+func PrepareVDSO(p platform.Platform) (*VDSO, error) {
+	vdsoFile := newByteReaderFile(vdsoBin)
+
+	// First make sure the VDSO is valid. vdsoFile does not use ctx, so a
+	// nil context can be passed.
+	info, err := validateVDSO(nil, vdsoFile, uint64(len(vdsoBin)))
+	if err != nil {
+		return nil, err
+	}
+
+	// Then copy it into a VDSO mapping.
+	size, ok := usermem.Addr(len(vdsoBin)).RoundUp()
+	if !ok {
+		return nil, fmt.Errorf("VDSO size overflows? %#x", len(vdsoBin))
+	}
+
+	vdso, err := p.Memory().Allocate(uint64(size), usage.System)
+	if err != nil {
+		return nil, fmt.Errorf("unable to allocate VDSO memory: %v", err)
+	}
+
+	ims, err := p.Memory().MapInternal(vdso, usermem.ReadWrite)
+	if err != nil {
+		p.Memory().DecRef(vdso)
+		return nil, fmt.Errorf("unable to map VDSO memory: %v", err)
+	}
+
+	_, err = safemem.CopySeq(ims, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(vdsoBin)))
+	if err != nil {
+		p.Memory().DecRef(vdso)
+		return nil, fmt.Errorf("unable to copy VDSO into memory: %v", err)
+	}
+
+	// Finally, allocate a param page for this VDSO.
+	paramPage, err := p.Memory().Allocate(usermem.PageSize, usage.System)
+	if err != nil {
+		p.Memory().DecRef(vdso)
+		return nil, fmt.Errorf("unable to allocate VDSO param page: %v", err)
+	}
+
+	return &VDSO{
+		ParamPage: mm.NewSpecialMappable("[vvar]", p, paramPage),
+		// TODO: Don't advertise the VDSO, as some applications may
+		// not be able to handle multiple [vdso] hints.
+		vdso:  mm.NewSpecialMappable("", p, vdso),
+		phdrs: info.phdrs,
+	}, nil
+}
+
+// loadVDSO loads the VDSO into m.
+//
+// VDSOs are special.
+//
+// VDSOs are fully position independent. However, instead of loading a VDSO
+// like a normal ELF binary, mapping only the PT_LOAD segments, the Linux
+// kernel simply directly maps the entire file into process memory, with very
+// little real ELF parsing.
+//
+// NOTE: This means that userspace can, and unfortunately does,
+// depend on parts of the ELF that would normally not be mapped.  To maintain
+// compatibility with such binaries, we load the VDSO much like Linux.
+//
+// loadVDSO takes a reference on the VDSO and parameter page FrameRegions.
+func loadVDSO(ctx context.Context, m *mm.MemoryManager, v *VDSO, bin loadedELF) (usermem.Addr, error) {
+	if v == nil {
+		// Should be used only by tests.
+		ctx.Warningf("No VDSO provided, skipping VDSO mapping")
+		return 0, nil
+	}
+
+	if v.os != bin.os {
+		ctx.Warningf("Binary ELF OS %v and VDSO ELF OS %v differ", bin.os, v.os)
+		return 0, syserror.ENOEXEC
+	}
+	if v.arch != bin.arch {
+		ctx.Warningf("Binary ELF arch %v and VDSO ELF arch %v differ", bin.arch, v.arch)
+		return 0, syserror.ENOEXEC
+	}
+
+	// Reserve address space for the VDSO and its parameter page, which is
+	// mapped just before the VDSO.
+	mapSize := v.vdso.Length() + v.ParamPage.Length()
+	addr, err := m.MMap(ctx, memmap.MMapOpts{
+		Length:  mapSize,
+		Private: true,
+	})
+	if err != nil {
+		ctx.Infof("Unable to reserve VDSO address space: %v", err)
+		return 0, err
+	}
+
+	// Now map the param page.
+	_, err = m.MMap(ctx, memmap.MMapOpts{
+		Length:          v.ParamPage.Length(),
+		MappingIdentity: v.ParamPage,
+		Mappable:        v.ParamPage,
+		Addr:            addr,
+		Fixed:           true,
+		Unmap:           true,
+		Private:         true,
+		Perms:           usermem.Read,
+		MaxPerms:        usermem.Read,
+	})
+	if err != nil {
+		ctx.Infof("Unable to map VDSO param page: %v", err)
+		return 0, err
+	}
+
+	// Now map the VDSO itself.
+	vdsoAddr, ok := addr.AddLength(v.ParamPage.Length())
+	if !ok {
+		panic(fmt.Sprintf("Part of mapped range overflows? %#x + %#x", addr, v.ParamPage.Length()))
+	}
+	_, err = m.MMap(ctx, memmap.MMapOpts{
+		Length:          v.vdso.Length(),
+		MappingIdentity: v.vdso,
+		Mappable:        v.vdso,
+		Addr:            vdsoAddr,
+		Fixed:           true,
+		Unmap:           true,
+		Private:         true,
+		Perms:           usermem.Read,
+		MaxPerms:        usermem.AnyAccess,
+	})
+	if err != nil {
+		ctx.Infof("Unable to map VDSO: %v", err)
+		return 0, err
+	}
+
+	vdsoEnd, ok := vdsoAddr.AddLength(v.vdso.Length())
+	if !ok {
+		panic(fmt.Sprintf("VDSO mapping overflows? %#x + %#x", vdsoAddr, v.vdso.Length()))
+	}
+
+	// Set additional protections for the individual segments.
+	var first *elf.ProgHeader
+	for i, phdr := range v.phdrs {
+		if phdr.Type != elf.PT_LOAD {
+			continue
+		}
+
+		if first == nil {
+			first = &v.phdrs[i]
+		}
+
+		memoryOffset := phdr.Vaddr - first.Vaddr
+		segAddr, ok := vdsoAddr.AddLength(memoryOffset)
+		if !ok {
+			ctx.Warningf("PT_LOAD segment address overflows: %#x + %#x", segAddr, memoryOffset)
+			return 0, syserror.ENOEXEC
+		}
+		segPage := segAddr.RoundDown()
+		segSize := usermem.Addr(phdr.Memsz)
+		segSize, ok = segSize.AddLength(segAddr.PageOffset())
+		if !ok {
+			ctx.Warningf("PT_LOAD segment memsize %#x + offset %#x overflows", phdr.Memsz, segAddr.PageOffset())
+			return 0, syserror.ENOEXEC
+		}
+		segSize, ok = segSize.RoundUp()
+		if !ok {
+			ctx.Warningf("PT_LOAD segment size overflows: %#x", phdr.Memsz+segAddr.PageOffset())
+			return 0, syserror.ENOEXEC
+		}
+		segEnd, ok := segPage.AddLength(uint64(segSize))
+		if !ok {
+			ctx.Warningf("PT_LOAD segment range overflows: %#x + %#x", segAddr, segSize)
+			return 0, syserror.ENOEXEC
+		}
+		if segEnd > vdsoEnd {
+			ctx.Warningf("PT_LOAD segment ends beyond VDSO: %#x > %#x", segEnd, vdsoEnd)
+			return 0, syserror.ENOEXEC
+		}
+
+		perms := progFlagsAsPerms(phdr.Flags)
+		if perms != usermem.Read {
+			if err := m.MProtect(segPage, uint64(segSize), perms, false); err != nil {
+				ctx.Warningf("Unable to set PT_LOAD segment protections %+v at [%#x, %#x): %v", perms, segAddr, segEnd, err)
+				return 0, syserror.ENOEXEC
+			}
+		}
+	}
+
+	return vdsoAddr, nil
+}
diff --git a/pkg/sentry/loader/vdso_state.go b/pkg/sentry/loader/vdso_state.go
new file mode 100644
index 000000000..92004ad9e
--- /dev/null
+++ b/pkg/sentry/loader/vdso_state.go
@@ -0,0 +1,47 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package loader
+
+import (
+	"debug/elf"
+)
+
+type elfProgHeader struct {
+	Type   elf.ProgType
+	Flags  elf.ProgFlag
+	Off    uint64
+	Vaddr  uint64
+	Paddr  uint64
+	Filesz uint64
+	Memsz  uint64
+	Align  uint64
+}
+
+// savePhdrs is invoked by stateify.
+func (v *VDSO) savePhdrs() []elfProgHeader {
+	s := make([]elfProgHeader, 0, len(v.phdrs))
+	for _, h := range v.phdrs {
+		s = append(s, elfProgHeader(h))
+	}
+	return s
+}
+
+// loadPhdrs is invoked by stateify.
+func (v *VDSO) loadPhdrs(s []elfProgHeader) {
+	v.phdrs = make([]elf.ProgHeader, 0, len(s))
+	for _, h := range s {
+		v.phdrs = append(v.phdrs, elf.ProgHeader(h))
+	}
+}
diff --git a/pkg/sentry/memmap/BUILD b/pkg/sentry/memmap/BUILD
new file mode 100644
index 000000000..7525fea45
--- /dev/null
+++ b/pkg/sentry/memmap/BUILD
@@ -0,0 +1,71 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "memmap_state",
+    srcs = [
+        "mappable_range.go",
+        "mapping_set.go",
+        "mapping_set_impl.go",
+    ],
+    out = "memmap_state.go",
+    package = "memmap",
+)
+
+go_template_instance(
+    name = "mappable_range",
+    out = "mappable_range.go",
+    package = "memmap",
+    prefix = "Mappable",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "uint64",
+    },
+)
+
+go_template_instance(
+    name = "mapping_set_impl",
+    out = "mapping_set_impl.go",
+    package = "memmap",
+    prefix = "Mapping",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "MappableRange",
+        "Value": "MappingsOfRange",
+        "Functions": "mappingSetFunctions",
+    },
+)
+
+go_library(
+    name = "memmap",
+    srcs = [
+        "mappable_range.go",
+        "mapping_set.go",
+        "mapping_set_impl.go",
+        "memmap.go",
+        "memmap_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/memmap",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/sentry/context",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "memmap_test",
+    size = "small",
+    srcs = ["mapping_set_test.go"],
+    embed = [":memmap"],
+    deps = ["//pkg/sentry/usermem"],
+)
diff --git a/pkg/sentry/memmap/mapping_set.go b/pkg/sentry/memmap/mapping_set.go
new file mode 100644
index 000000000..0cd42ffbf
--- /dev/null
+++ b/pkg/sentry/memmap/mapping_set.go
@@ -0,0 +1,245 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memmap
+
+import (
+	"fmt"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// MappingSet maps offsets into a Mappable to mappings of those offsets. It is
+// used to implement Mappable.AddMapping and RemoveMapping for Mappables that
+// may need to call MappingSpace.Invalidate.
+//
+// type MappingSet <generated by go_generics>
+
+// MappingsOfRange is the value type of MappingSet, and represents the set of
+// all mappings of the corresponding MappableRange.
+//
+// Using a map offers O(1) lookups in RemoveMapping and
+// mappingSetFunctions.Merge.
+type MappingsOfRange map[MappingOfRange]struct{}
+
+// MappingOfRange represents a mapping of a MappableRange.
+type MappingOfRange struct {
+	MappingSpace MappingSpace
+	AddrRange    usermem.AddrRange
+}
+
+func (r MappingOfRange) invalidate(opts InvalidateOpts) {
+	r.MappingSpace.Invalidate(r.AddrRange, opts)
+}
+
+// String implements fmt.Stringer.String.
+func (r MappingOfRange) String() string {
+	return fmt.Sprintf("%#v", r.AddrRange)
+}
+
+// mappingSetFunctions implements segment.Functions for MappingSet.
+type mappingSetFunctions struct{}
+
+// MinKey implements segment.Functions.MinKey.
+func (mappingSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+// MaxKey implements segment.Functions.MaxKey.
+func (mappingSetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+// ClearValue implements segment.Functions.ClearValue.
+func (mappingSetFunctions) ClearValue(v *MappingsOfRange) {
+	*v = MappingsOfRange{}
+}
+
+// Merge implements segment.Functions.Merge.
+//
+// Since each value is a map of MappingOfRanges, values can only be merged if
+// all MappingOfRanges in each map have an exact pair in the other map, forming
+// one contiguous region.
+func (mappingSetFunctions) Merge(r1 MappableRange, val1 MappingsOfRange, r2 MappableRange, val2 MappingsOfRange) (MappingsOfRange, bool) {
+	if len(val1) != len(val2) {
+		return nil, false
+	}
+
+	merged := make(MappingsOfRange, len(val1))
+
+	// Each MappingOfRange in val1 must have a matching region in val2, forming
+	// one contiguous region.
+	for k1 := range val1 {
+		// We expect val2 to to contain a key that forms a contiguous
+		// region with k1.
+		k2 := MappingOfRange{
+			MappingSpace: k1.MappingSpace,
+			AddrRange: usermem.AddrRange{
+				Start: k1.AddrRange.End,
+				End:   k1.AddrRange.End + usermem.Addr(r2.Length()),
+			},
+		}
+		if _, ok := val2[k2]; !ok {
+			return nil, false
+		}
+
+		// OK. Add it to the merged map.
+		merged[MappingOfRange{
+			MappingSpace: k1.MappingSpace,
+			AddrRange: usermem.AddrRange{
+				Start: k1.AddrRange.Start,
+				End:   k2.AddrRange.End,
+			},
+		}] = struct{}{}
+	}
+
+	return merged, true
+}
+
+// Split implements segment.Functions.Split.
+func (mappingSetFunctions) Split(r MappableRange, val MappingsOfRange, split uint64) (MappingsOfRange, MappingsOfRange) {
+	if split <= r.Start || split >= r.End {
+		panic(fmt.Sprintf("split is not within range %v", r))
+	}
+
+	m1 := make(MappingsOfRange, len(val))
+	m2 := make(MappingsOfRange, len(val))
+
+	// split is a value in MappableRange, we need the offset into the
+	// corresponding MappingsOfRange.
+	offset := usermem.Addr(split - r.Start)
+	for k := range val {
+		k1 := MappingOfRange{
+			MappingSpace: k.MappingSpace,
+			AddrRange: usermem.AddrRange{
+				Start: k.AddrRange.Start,
+				End:   k.AddrRange.Start + offset,
+			},
+		}
+		m1[k1] = struct{}{}
+
+		k2 := MappingOfRange{
+			MappingSpace: k.MappingSpace,
+			AddrRange: usermem.AddrRange{
+				Start: k.AddrRange.Start + offset,
+				End:   k.AddrRange.End,
+			},
+		}
+		m2[k2] = struct{}{}
+	}
+
+	return m1, m2
+}
+
+// subsetMapping returns the MappingOfRange that maps subsetRange, given that
+// ms maps wholeRange beginning at addr.
+//
+// For instance, suppose wholeRange = [0x0, 0x2000) and addr = 0x4000,
+// indicating that ms maps addresses [0x4000, 0x6000) to MappableRange [0x0,
+// 0x2000). Then for subsetRange = [0x1000, 0x2000), subsetMapping returns a
+// MappingOfRange for which AddrRange = [0x5000, 0x6000).
+func subsetMapping(wholeRange, subsetRange MappableRange, ms MappingSpace, addr usermem.Addr) MappingOfRange {
+	if !wholeRange.IsSupersetOf(subsetRange) {
+		panic(fmt.Sprintf("%v is not a superset of %v", wholeRange, subsetRange))
+	}
+
+	offset := subsetRange.Start - wholeRange.Start
+	start := addr + usermem.Addr(offset)
+	return MappingOfRange{
+		MappingSpace: ms,
+		AddrRange: usermem.AddrRange{
+			Start: start,
+			End:   start + usermem.Addr(subsetRange.Length()),
+		},
+	}
+}
+
+// AddMapping adds the given mapping and returns the set of MappableRanges that
+// previously had no mappings.
+//
+// Preconditions: As for Mappable.AddMapping.
+func (s *MappingSet) AddMapping(ms MappingSpace, ar usermem.AddrRange, offset uint64) []MappableRange {
+	mr := MappableRange{offset, offset + uint64(ar.Length())}
+	var mapped []MappableRange
+	seg, gap := s.Find(mr.Start)
+	for {
+		switch {
+		case seg.Ok() && seg.Start() < mr.End:
+			seg = s.Isolate(seg, mr)
+			seg.Value()[subsetMapping(mr, seg.Range(), ms, ar.Start)] = struct{}{}
+			seg, gap = seg.NextNonEmpty()
+
+		case gap.Ok() && gap.Start() < mr.End:
+			gapMR := gap.Range().Intersect(mr)
+			mapped = append(mapped, gapMR)
+			// Insert a set and continue from the above case.
+			seg, gap = s.Insert(gap, gapMR, make(MappingsOfRange)), MappingGapIterator{}
+
+		default:
+			return mapped
+		}
+	}
+}
+
+// RemoveMapping removes the given mapping and returns the set of
+// MappableRanges that now have no mappings.
+//
+// Preconditions: As for Mappable.RemoveMapping.
+func (s *MappingSet) RemoveMapping(ms MappingSpace, ar usermem.AddrRange, offset uint64) []MappableRange {
+	mr := MappableRange{offset, offset + uint64(ar.Length())}
+	var unmapped []MappableRange
+
+	seg := s.FindSegment(mr.Start)
+	if !seg.Ok() {
+		panic(fmt.Sprintf("MappingSet.RemoveMapping(%v): no segment containing %#x: %v", mr, mr.Start, s))
+	}
+	for seg.Ok() && seg.Start() < mr.End {
+		// Ensure this segment is limited to our range.
+		seg = s.Isolate(seg, mr)
+
+		// Remove this part of the mapping.
+		mappings := seg.Value()
+		delete(mappings, subsetMapping(mr, seg.Range(), ms, ar.Start))
+
+		if len(mappings) == 0 {
+			unmapped = append(unmapped, seg.Range())
+			seg = s.Remove(seg).NextSegment()
+		} else {
+			seg = seg.NextSegment()
+		}
+	}
+	s.MergeAdjacent(mr)
+	return unmapped
+}
+
+// Invalidate calls MappingSpace.Invalidate for all mappings of offsets in mr.
+func (s *MappingSet) Invalidate(mr MappableRange, opts InvalidateOpts) {
+	for seg := s.LowerBoundSegment(mr.Start); seg.Ok() && seg.Start() < mr.End; seg = seg.NextSegment() {
+		segMR := seg.Range()
+		for m := range seg.Value() {
+			region := subsetMapping(segMR, segMR.Intersect(mr), m.MappingSpace, m.AddrRange.Start)
+			region.invalidate(opts)
+		}
+	}
+}
+
+// InvalidateAll calls MappingSpace.Invalidate for all mappings of s.
+func (s *MappingSet) InvalidateAll(opts InvalidateOpts) {
+	for seg := s.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		for m := range seg.Value() {
+			m.invalidate(opts)
+		}
+	}
+}
diff --git a/pkg/sentry/memmap/mapping_set_test.go b/pkg/sentry/memmap/mapping_set_test.go
new file mode 100644
index 000000000..10668d404
--- /dev/null
+++ b/pkg/sentry/memmap/mapping_set_test.go
@@ -0,0 +1,186 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memmap
+
+import (
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type testMappingSpace struct {
+	// Ideally we'd store the full ranges that were invalidated, rather
+	// than individual calls to Invalidate, as they are an implementation
+	// detail, but this is the simplest way for now.
+	inv []usermem.AddrRange
+}
+
+func (n *testMappingSpace) reset() {
+	n.inv = []usermem.AddrRange{}
+}
+
+func (n *testMappingSpace) Invalidate(ar usermem.AddrRange, opts InvalidateOpts) {
+	n.inv = append(n.inv, ar)
+}
+
+func TestAddRemoveMapping(t *testing.T) {
+	set := MappingSet{}
+	ms := &testMappingSpace{}
+
+	mapped := set.AddMapping(ms, usermem.AddrRange{0x10000, 0x12000}, 0x1000)
+	if got, want := mapped, []MappableRange{{0x1000, 0x3000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("AddMapping: got %+v, wanted %+v", got, want)
+	}
+
+	// Mappings (usermem.AddrRanges => memmap.MappableRange):
+	// [0x10000, 0x12000) => [0x1000, 0x3000)
+	t.Log(&set)
+
+	mapped = set.AddMapping(ms, usermem.AddrRange{0x20000, 0x21000}, 0x2000)
+	if len(mapped) != 0 {
+		t.Errorf("AddMapping: got %+v, wanted []", mapped)
+	}
+
+	// Mappings:
+	// [0x10000, 0x11000) => [0x1000, 0x2000)
+	// [0x11000, 0x12000) and [0x20000, 0x21000) => [0x2000, 0x3000)
+	t.Log(&set)
+
+	mapped = set.AddMapping(ms, usermem.AddrRange{0x30000, 0x31000}, 0x4000)
+	if got, want := mapped, []MappableRange{{0x4000, 0x5000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("AddMapping: got %+v, wanted %+v", got, want)
+	}
+
+	// Mappings:
+	// [0x10000, 0x11000) => [0x1000, 0x2000)
+	// [0x11000, 0x12000) and [0x20000, 0x21000) => [0x2000, 0x3000)
+	// [0x30000, 0x31000) => [0x4000, 0x5000)
+	t.Log(&set)
+
+	mapped = set.AddMapping(ms, usermem.AddrRange{0x12000, 0x15000}, 0x3000)
+	if got, want := mapped, []MappableRange{{0x3000, 0x4000}, {0x5000, 0x6000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("AddMapping: got %+v, wanted %+v", got, want)
+	}
+
+	// Mappings:
+	// [0x10000, 0x11000) => [0x1000, 0x2000)
+	// [0x11000, 0x12000) and [0x20000, 0x21000) => [0x2000, 0x3000)
+	// [0x12000, 0x13000) => [0x3000, 0x4000)
+	// [0x13000, 0x14000) and [0x30000, 0x31000) => [0x4000, 0x5000)
+	// [0x14000, 0x15000) => [0x5000, 0x6000)
+	t.Log(&set)
+
+	unmapped := set.RemoveMapping(ms, usermem.AddrRange{0x10000, 0x11000}, 0x1000)
+	if got, want := unmapped, []MappableRange{{0x1000, 0x2000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("RemoveMapping: got %+v, wanted %+v", got, want)
+	}
+
+	// Mappings:
+	// [0x11000, 0x12000) and [0x20000, 0x21000) => [0x2000, 0x3000)
+	// [0x12000, 0x13000) => [0x3000, 0x4000)
+	// [0x13000, 0x14000) and [0x30000, 0x31000) => [0x4000, 0x5000)
+	// [0x14000, 0x15000) => [0x5000, 0x6000)
+	t.Log(&set)
+
+	unmapped = set.RemoveMapping(ms, usermem.AddrRange{0x20000, 0x21000}, 0x2000)
+	if len(unmapped) != 0 {
+		t.Errorf("RemoveMapping: got %+v, wanted []", unmapped)
+	}
+
+	// Mappings:
+	// [0x11000, 0x13000) => [0x2000, 0x4000)
+	// [0x13000, 0x14000) and [0x30000, 0x31000) => [0x4000, 0x5000)
+	// [0x14000, 0x15000) => [0x5000, 0x6000)
+	t.Log(&set)
+
+	unmapped = set.RemoveMapping(ms, usermem.AddrRange{0x11000, 0x15000}, 0x2000)
+	if got, want := unmapped, []MappableRange{{0x2000, 0x4000}, {0x5000, 0x6000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("RemoveMapping: got %+v, wanted %+v", got, want)
+	}
+
+	// Mappings:
+	// [0x30000, 0x31000) => [0x4000, 0x5000)
+	t.Log(&set)
+
+	unmapped = set.RemoveMapping(ms, usermem.AddrRange{0x30000, 0x31000}, 0x4000)
+	if got, want := unmapped, []MappableRange{{0x4000, 0x5000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("RemoveMapping: got %+v, wanted %+v", got, want)
+	}
+}
+
+func TestInvalidateWholeMapping(t *testing.T) {
+	set := MappingSet{}
+	ms := &testMappingSpace{}
+
+	set.AddMapping(ms, usermem.AddrRange{0x10000, 0x11000}, 0)
+	// Mappings:
+	// [0x10000, 0x11000) => [0, 0x1000)
+	t.Log(&set)
+	set.Invalidate(MappableRange{0, 0x1000}, InvalidateOpts{})
+	if got, want := ms.inv, []usermem.AddrRange{{0x10000, 0x11000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("Invalidate: got %+v, wanted %+v", got, want)
+	}
+}
+
+func TestInvalidatePartialMapping(t *testing.T) {
+	set := MappingSet{}
+	ms := &testMappingSpace{}
+
+	set.AddMapping(ms, usermem.AddrRange{0x10000, 0x13000}, 0)
+	// Mappings:
+	// [0x10000, 0x13000) => [0, 0x3000)
+	t.Log(&set)
+	set.Invalidate(MappableRange{0x1000, 0x2000}, InvalidateOpts{})
+	if got, want := ms.inv, []usermem.AddrRange{{0x11000, 0x12000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("Invalidate: got %+v, wanted %+v", got, want)
+	}
+}
+
+func TestInvalidateMultipleMappings(t *testing.T) {
+	set := MappingSet{}
+	ms := &testMappingSpace{}
+
+	set.AddMapping(ms, usermem.AddrRange{0x10000, 0x11000}, 0)
+	set.AddMapping(ms, usermem.AddrRange{0x20000, 0x21000}, 0x2000)
+	// Mappings:
+	// [0x10000, 0x11000) => [0, 0x1000)
+	// [0x12000, 0x13000) => [0x2000, 0x3000)
+	t.Log(&set)
+	set.Invalidate(MappableRange{0, 0x3000}, InvalidateOpts{})
+	if got, want := ms.inv, []usermem.AddrRange{{0x10000, 0x11000}, {0x20000, 0x21000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("Invalidate: got %+v, wanted %+v", got, want)
+	}
+}
+
+func TestInvalidateOverlappingMappings(t *testing.T) {
+	set := MappingSet{}
+	ms1 := &testMappingSpace{}
+	ms2 := &testMappingSpace{}
+
+	set.AddMapping(ms1, usermem.AddrRange{0x10000, 0x12000}, 0)
+	set.AddMapping(ms2, usermem.AddrRange{0x20000, 0x22000}, 0x1000)
+	// Mappings:
+	// ms1:[0x10000, 0x12000) => [0, 0x2000)
+	// ms2:[0x11000, 0x13000) => [0x1000, 0x3000)
+	t.Log(&set)
+	set.Invalidate(MappableRange{0x1000, 0x2000}, InvalidateOpts{})
+	if got, want := ms1.inv, []usermem.AddrRange{{0x11000, 0x12000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("Invalidate: ms1: got %+v, wanted %+v", got, want)
+	}
+	if got, want := ms2.inv, []usermem.AddrRange{{0x20000, 0x21000}}; !reflect.DeepEqual(got, want) {
+		t.Errorf("Invalidate: ms1: got %+v, wanted %+v", got, want)
+	}
+}
diff --git a/pkg/sentry/memmap/memmap.go b/pkg/sentry/memmap/memmap.go
new file mode 100644
index 000000000..14fed55bc
--- /dev/null
+++ b/pkg/sentry/memmap/memmap.go
@@ -0,0 +1,297 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package memmap defines semantics for memory mappings.
+package memmap
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Mappable represents a memory-mappable object, a mutable mapping from uint64
+// offsets to (platform.File, uint64 File offset) pairs.
+//
+// See mm/mm.go for Mappable's place in the lock order.
+//
+// Preconditions: For all Mappable methods, usermem.AddrRanges and
+// MappableRanges must be non-empty (Length() != 0), and usermem.Addrs and
+// Mappable offsets must be page-aligned.
+type Mappable interface {
+	// AddMapping notifies the Mappable of a mapping from addresses ar in ms to
+	// offsets [offset, offset+ar.Length()) in this Mappable.
+	//
+	// Preconditions: offset+ar.Length() does not overflow.
+	AddMapping(ctx context.Context, ms MappingSpace, ar usermem.AddrRange, offset uint64) error
+
+	// RemoveMapping notifies the Mappable of the removal of a mapping from
+	// addresses ar in ms to offsets [offset, offset+ar.Length()) in this
+	// Mappable.
+	//
+	// Preconditions: offset+ar.Length() does not overflow. The removed mapping
+	// must exist.
+	RemoveMapping(ctx context.Context, ms MappingSpace, ar usermem.AddrRange, offset uint64)
+
+	// CopyMapping notifies the Mappable of an attempt to copy a mapping in ms
+	// from srcAR to dstAR. For most Mappables, this is equivalent to
+	// AddMapping.
+	//
+	// CopyMapping is only called when a mapping is copied within a given
+	// MappingSpace; it is analogous to Linux's vm_operations_struct::mremap.
+	//
+	// Preconditions: offset+dstAR.Length() does not overflow. The mapping at
+	// srcAR must exist.
+	CopyMapping(ctx context.Context, ms MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error
+
+	// Translate returns the Mappable's current mappings for at least the range
+	// of offsets specified by required, and at most the range of offsets
+	// specified by optional. at is the set of access types that may be
+	// performed using the returned Translations. If not all required offsets
+	// are translated, it returns a non-nil error explaining why. Returned
+	// translations, and any mappings returned by platform.File.MapInternal for
+	// translated platform.Files, are valid until invalidated by a call back to
+	// MappingSpace.Invalidate or until the caller removes its mapping of the
+	// translated range.
+	//
+	// Preconditions: required.Length() > 0. optional.IsSupersetOf(required).
+	// required and optional must be page-aligned. The caller must have
+	// established a mapping for all of the queried offsets via a previous call
+	// to AddMapping. The caller is responsible for ensuring that calls to
+	// Translate synchronize with invalidation.
+	//
+	// Postconditions: See CheckTranslateResult.
+	Translate(ctx context.Context, required, optional MappableRange, at usermem.AccessType) ([]Translation, error)
+
+	// InvalidateUnsavable requests that the Mappable invalidate Translations
+	// that cannot be preserved across save/restore.
+	//
+	// Invariant: InvalidateUnsavable never races with concurrent calls to any
+	// other Mappable methods.
+	InvalidateUnsavable(ctx context.Context) error
+}
+
+// Translations are returned by Mappable.Translate.
+type Translation struct {
+	// Source is the translated range in the Mappable.
+	Source MappableRange
+
+	// File is the mapped file. When the Translation is invalidated, pages
+	// mapped by File.MapInto must be unmapped, and pages mapped by
+	// File.MapInternal become invalid.
+	File platform.File
+
+	// Offset is the offset into File at which this Translation begins.
+	Offset uint64
+}
+
+// CheckTranslateResult returns an error if (ts, terr) does not satisfy all
+// postconditions for Mappable.Translate(required, optional).
+//
+// Preconditions: As for Mappable.Translate.
+func CheckTranslateResult(required, optional MappableRange, ts []Translation, terr error) error {
+	// Verify that the inputs to Mappable.Translate were valid.
+	if !required.WellFormed() || required.Length() <= 0 {
+		panic(fmt.Sprintf("invalid required range: %v", required))
+	}
+	if !usermem.Addr(required.Start).IsPageAligned() || !usermem.Addr(required.End).IsPageAligned() {
+		panic(fmt.Sprintf("unaligned required range: %v", required))
+	}
+	if !optional.IsSupersetOf(required) {
+		panic(fmt.Sprintf("optional range %v is not a superset of required range %v", optional, required))
+	}
+	if !usermem.Addr(optional.Start).IsPageAligned() || !usermem.Addr(optional.End).IsPageAligned() {
+		panic(fmt.Sprintf("unaligned optional range: %v", optional))
+	}
+
+	// The first Translation must include required.Start.
+	if len(ts) != 0 && !ts[0].Source.Contains(required.Start) {
+		return fmt.Errorf("first Translation %+v does not cover start of required range %v", ts[0], required)
+	}
+	for i, t := range ts {
+		if !t.Source.WellFormed() || t.Source.Length() <= 0 {
+			return fmt.Errorf("Translation %+v has invalid Source", t)
+		}
+		if !usermem.Addr(t.Source.Start).IsPageAligned() || !usermem.Addr(t.Source.End).IsPageAligned() {
+			return fmt.Errorf("Translation %+v has unaligned Source", t)
+		}
+		if t.File == nil {
+			return fmt.Errorf("Translation %+v has nil File", t)
+		}
+		if !usermem.Addr(t.Offset).IsPageAligned() {
+			return fmt.Errorf("Translation %+v has unaligned Offset", t)
+		}
+		// Translations must be contiguous and in increasing order of
+		// Translation.Source.
+		if i > 0 && ts[i-1].Source.End != t.Source.Start {
+			return fmt.Errorf("Translations %+v and %+v are not contiguous", ts[i-1], t)
+		}
+		// At least part of each Translation must be required.
+		if t.Source.Intersect(required).Length() == 0 {
+			return fmt.Errorf("Translation %+v lies entirely outside required range %v", t, required)
+		}
+		// Translations must be constrained to the optional range.
+		if !optional.IsSupersetOf(t.Source) {
+			return fmt.Errorf("Translation %+v lies outside optional range %v", t, optional)
+		}
+	}
+	// If the set of Translations does not cover the entire required range,
+	// Translate must return a non-nil error explaining why.
+	if terr == nil {
+		if len(ts) == 0 {
+			return fmt.Errorf("no Translations and no error")
+		}
+		if t := ts[len(ts)-1]; !t.Source.Contains(required.End - 1) {
+			return fmt.Errorf("last Translation %+v does not reach end of required range %v, but Translate returned no error", t, required)
+		}
+	}
+	return nil
+}
+
+// BusError may be returned by implementations of Mappable.Translate for errors
+// that should result in SIGBUS delivery if they cause application page fault
+// handling to fail.
+type BusError struct {
+	// Err is the original error.
+	Err error
+}
+
+// Error implements error.Error.
+func (b *BusError) Error() string {
+	return fmt.Sprintf("BusError: %v", b.Err.Error())
+}
+
+// MappableRange represents a range of uint64 offsets into a Mappable.
+//
+// type MappableRange <generated using go_generics>
+
+// String implements fmt.Stringer.String.
+func (mr MappableRange) String() string {
+	return fmt.Sprintf("[%#x, %#x)", mr.Start, mr.End)
+}
+
+// MappingSpace represents a mutable mapping from usermem.Addrs to (Mappable,
+// uint64 offset) pairs.
+type MappingSpace interface {
+	// Invalidate is called to notify the MappingSpace that values returned by
+	// previous calls to Mappable.Translate for offsets mapped by addresses in
+	// ar are no longer valid.
+	//
+	// Invalidate must not take any locks preceding mm.MemoryManager.activeMu
+	// in the lock order.
+	//
+	// Preconditions: ar.Length() != 0. ar must be page-aligned.
+	Invalidate(ar usermem.AddrRange, opts InvalidateOpts)
+}
+
+// InvalidateOpts holds options to MappingSpace.Invalidate.
+type InvalidateOpts struct {
+	// InvalidatePrivate is true if private pages in the invalidated region
+	// should also be discarded, causing their data to be lost.
+	InvalidatePrivate bool
+}
+
+// MappingIdentity controls the lifetime of a Mappable, and provides
+// information about the Mappable for /proc/[pid]/maps. It is distinct from
+// Mappable because all Mappables that are coherent must compare equal to
+// support the implementation of shared futexes, but different
+// MappingIdentities may represent the same Mappable, in the same way that
+// multiple fs.Files may represent the same fs.Inode. (This similarity is not
+// coincidental; fs.File implements MappingIdentity, and some
+// fs.InodeOperations implement Mappable.)
+type MappingIdentity interface {
+	// MappingIdentity is reference-counted.
+	refs.RefCounter
+
+	// MappedName returns the application-visible name shown in
+	// /proc/[pid]/maps.
+	MappedName(ctx context.Context) string
+
+	// DeviceID returns the device number shown in /proc/[pid]/maps.
+	DeviceID() uint64
+
+	// InodeID returns the inode number shown in /proc/[pid]/maps.
+	InodeID() uint64
+
+	// Msync has the same semantics as fs.FileOperations.Fsync(ctx,
+	// int64(mr.Start), int64(mr.End-1), fs.SyncData).
+	// (fs.FileOperations.Fsync() takes an inclusive end, but mr.End is
+	// exclusive, hence mr.End-1.) It is defined rather than Fsync so that
+	// implementors don't need to depend on the fs package for fs.SyncType.
+	Msync(ctx context.Context, mr MappableRange) error
+}
+
+// MMapOpts specifies a request to create a memory mapping.
+type MMapOpts struct {
+	// Length is the length of the mapping.
+	Length uint64
+
+	// MappingIdentity controls the lifetime of Mappable, and provides
+	// properties of the mapping shown in /proc/[pid]/maps. If MMapOpts is used
+	// to successfully create a memory mapping, a reference is taken on
+	// MappingIdentity.
+	MappingIdentity MappingIdentity
+
+	// Mappable is the Mappable to be mapped. If Mappable is nil, the mapping
+	// is anonymous. If Mappable is not nil, it must remain valid as long as a
+	// reference is held on MappingIdentity.
+	Mappable Mappable
+
+	// Offset is the offset into Mappable to map. If Mappable is nil, Offset is
+	// ignored.
+	Offset uint64
+
+	// Addr is the suggested address for the mapping.
+	Addr usermem.Addr
+
+	// Fixed specifies whether this is a fixed mapping (it must be located at
+	// Addr).
+	Fixed bool
+
+	// Unmap specifies whether existing mappings in the range being mapped may
+	// be replaced. If Unmap is true, Fixed must be true.
+	Unmap bool
+
+	// Perms is the set of permissions to the applied to this mapping.
+	Perms usermem.AccessType
+
+	// MaxPerms limits the set of permissions that may ever apply to this
+	// mapping. If Mappable is not nil, all memmap.Translations returned by
+	// Mappable.Translate must support all accesses in MaxPerms.
+	//
+	// Preconditions: MaxAccessType should be an effective AccessType, as
+	// access cannot be limited beyond effective AccessTypes.
+	MaxPerms usermem.AccessType
+
+	// Private is true if writes to the mapping should be propagated to a copy
+	// that is exclusive to the MemoryManager.
+	Private bool
+
+	// GrowsDown is true if the mapping should be automatically expanded
+	// downward on guard page faults.
+	GrowsDown bool
+
+	// Precommit is true if the platform should eagerly commit resources to the
+	// mapping (see platform.AddressSpace.MapFile).
+	Precommit bool
+
+	// Hint is the name used for the mapping in /proc/[pid]/maps. If Hint is
+	// empty, MappingIdentity.MappedName() will be used instead.
+	//
+	// TODO: Replace entirely with MappingIdentity?
+	Hint string
+}
diff --git a/pkg/sentry/memutil/BUILD b/pkg/sentry/memutil/BUILD
new file mode 100644
index 000000000..a387a0c9f
--- /dev/null
+++ b/pkg/sentry/memutil/BUILD
@@ -0,0 +1,14 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "memutil",
+    srcs = [
+        "memutil.go",
+        "memutil_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/memutil",
+    visibility = ["//pkg/sentry:internal"],
+    deps = ["@org_golang_x_sys//unix:go_default_library"],
+)
diff --git a/pkg/sentry/memutil/memutil.go b/pkg/sentry/memutil/memutil.go
new file mode 100644
index 000000000..4f245cf3c
--- /dev/null
+++ b/pkg/sentry/memutil/memutil.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package memutil contains the utility functions for memory operations.
+package memutil
diff --git a/pkg/sentry/memutil/memutil_unsafe.go b/pkg/sentry/memutil/memutil_unsafe.go
new file mode 100644
index 000000000..32c27eb2f
--- /dev/null
+++ b/pkg/sentry/memutil/memutil_unsafe.go
@@ -0,0 +1,35 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package memutil
+
+import (
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+)
+
+// CreateMemFD creates a memfd file and returns the fd.
+func CreateMemFD(name string, flags int) (fd int, err error) {
+	p, err := syscall.BytePtrFromString(name)
+	if err != nil {
+		return -1, err
+	}
+	r0, _, e0 := syscall.Syscall(unix.SYS_MEMFD_CREATE, uintptr(unsafe.Pointer(p)), uintptr(flags), 0)
+	if e0 != 0 {
+		return -1, e0
+	}
+	return int(r0), nil
+}
diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD
new file mode 100644
index 000000000..39bde2be3
--- /dev/null
+++ b/pkg/sentry/mm/BUILD
@@ -0,0 +1,155 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "mm_state",
+    srcs = [
+        "aio_context.go",
+        "aio_context_state.go",
+        "file_refcount_set.go",
+        "io_list.go",
+        "mm.go",
+        "pma_set.go",
+        "save_restore.go",
+        "special_mappable.go",
+        "vma_set.go",
+    ],
+    out = "mm_state.go",
+    package = "mm",
+)
+
+go_template_instance(
+    name = "file_refcount_set",
+    out = "file_refcount_set.go",
+    imports = {
+        "platform": "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    },
+    package = "mm",
+    prefix = "fileRefcount",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "platform.FileRange",
+        "Value": "int32",
+        "Functions": "fileRefcountSetFunctions",
+    },
+)
+
+go_template_instance(
+    name = "vma_set",
+    out = "vma_set.go",
+    consts = {
+        "minDegree": "8",
+    },
+    imports = {
+        "usermem": "gvisor.googlesource.com/gvisor/pkg/sentry/usermem",
+    },
+    package = "mm",
+    prefix = "vma",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "usermem.Addr",
+        "Range": "usermem.AddrRange",
+        "Value": "vma",
+        "Functions": "vmaSetFunctions",
+    },
+)
+
+go_template_instance(
+    name = "pma_set",
+    out = "pma_set.go",
+    consts = {
+        "minDegree": "8",
+    },
+    imports = {
+        "usermem": "gvisor.googlesource.com/gvisor/pkg/sentry/usermem",
+    },
+    package = "mm",
+    prefix = "pma",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "usermem.Addr",
+        "Range": "usermem.AddrRange",
+        "Value": "pma",
+        "Functions": "pmaSetFunctions",
+    },
+)
+
+go_template_instance(
+    name = "io_list",
+    out = "io_list.go",
+    package = "mm",
+    prefix = "io",
+    template = "//pkg/ilist:generic_list",
+    types = {
+        "Linker": "*ioResult",
+    },
+)
+
+go_library(
+    name = "mm",
+    srcs = [
+        "address_space.go",
+        "aio_context.go",
+        "aio_context_state.go",
+        "debug.go",
+        "file_refcount_set.go",
+        "io.go",
+        "io_list.go",
+        "lifecycle.go",
+        "metadata.go",
+        "mm.go",
+        "mm_state.go",
+        "pma.go",
+        "pma_set.go",
+        "proc_pid_maps.go",
+        "save_restore.go",
+        "special_mappable.go",
+        "syscalls.go",
+        "vma.go",
+        "vma_set.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/mm",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/atomicbitops",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/proc/seqfile",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/platform/safecopy",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/sync",
+        "//pkg/syserror",
+        "//pkg/tcpip/buffer",
+    ],
+)
+
+go_test(
+    name = "mm_test",
+    size = "small",
+    srcs = ["mm_test.go"],
+    embed = [":mm"],
+    deps = [
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/context/contexttest",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/usermem",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/mm/README.md b/pkg/sentry/mm/README.md
new file mode 100644
index 000000000..067733475
--- /dev/null
+++ b/pkg/sentry/mm/README.md
@@ -0,0 +1,279 @@
+This package provides an emulation of Linux semantics for application virtual
+memory mappings.
+
+For completeness, this document also describes aspects of the memory management
+subsystem defined outside this package.
+
+# Background
+
+We begin by describing semantics for virtual memory in Linux.
+
+A virtual address space is defined as a collection of mappings from virtual
+addresses to physical memory. However, userspace applications do not configure
+mappings to physical memory directly. Instead, applications configure memory
+mappings from virtual addresses to offsets into a file using the `mmap` system
+call.[^mmap-anon] For example, a call to:
+
+    mmap(
+        /* addr = */ 0x400000,
+        /* length = */ 0x1000,
+        PROT_READ | PROT_WRITE,
+        MAP_SHARED,
+        /* fd = */ 3,
+        /* offset = */ 0);
+
+creates a mapping of length 0x1000 bytes, starting at virtual address (VA)
+0x400000, to offset 0 in the file represented by file descriptor (FD) 3. Within
+the Linux kernel, virtual memory mappings are represented by *virtual memory
+areas* (VMAs). Supposing that FD 3 represents file /tmp/foo, the state of the
+virtual memory subsystem after the `mmap` call may be depicted as:
+
+    VMA:     VA:0x400000 -> /tmp/foo:0x0
+
+Establishing a virtual memory area does not necessarily establish a mapping to a
+physical address, because Linux has not necessarily provisioned physical memory
+to store the file's contents. Thus, if the application attempts to read the
+contents of VA 0x400000, it may incur a *page fault*, a CPU exception that
+forces the kernel to create such a mapping to service the read.
+
+For a file, doing so consists of several logical phases:
+
+1. The kernel allocates physical memory to store the contents of the required
+   part of the file, and copies file contents to the allocated memory. Supposing
+   that the kernel chooses the physical memory at physical address (PA)
+   0x2fb000, the resulting state of the system is:
+
+        VMA:     VA:0x400000 -> /tmp/foo:0x0
+        Filemap:                /tmp/foo:0x0 -> PA:0x2fb000
+
+   (In Linux the state of the mapping from file offset to physical memory is
+   stored in `struct address_space`, but to avoid confusion with other notions
+   of address space we will refer to this system as filemap, named after Linux
+   kernel source file `mm/filemap.c`.)
+
+2. The kernel stores the effective mapping from virtual to physical address in a
+   *page table entry* (PTE) in the application's *page tables*, which are used
+   by the CPU's virtual memory hardware to perform address translation. The
+   resulting state of the system is:
+
+        VMA:     VA:0x400000 -> /tmp/foo:0x0
+        Filemap:                /tmp/foo:0x0 -> PA:0x2fb000
+        PTE:     VA:0x400000 -----------------> PA:0x2fb000
+
+   The PTE is required for the application to actually use the contents of the
+   mapped file as virtual memory. However, the PTE is derived from the VMA and
+   filemap state, both of which are independently mutable, such that mutations
+   to either will affect the PTE. For example:
+
+   - The application may remove the VMA using the `munmap` system call. This
+     breaks the mapping from VA:0x400000 to /tmp/foo:0x0, and consequently the
+     mapping from VA:0x400000 to PA:0x2fb000. However, it does not necessarily
+     break the mapping from /tmp/foo:0x0 to PA:0x2fb000, so a future mapping of
+     the same file offset may reuse this physical memory.
+
+   - The application may invalidate the file's contents by passing a length of 0
+     to the `ftruncate` system call. This breaks the mapping from /tmp/foo:0x0
+     to PA:0x2fb000, and consequently the mapping from VA:0x400000 to
+     PA:0x2fb000. However, it does not break the mapping from VA:0x400000 to
+     /tmp/foo:0x0, so future changes to the file's contents may again be made
+     visible at VA:0x400000 after another page fault results in the allocation
+     of a new physical address.
+
+   Note that, in order to correctly break the mapping from VA:0x400000 to
+   PA:0x2fb000 in the latter case, filemap must also store a *reverse mapping*
+   from /tmp/foo:0x0 to VA:0x400000 so that it can locate and remove the PTE.
+
+[^mmap-anon]: Memory mappings to non-files are discussed in later sections.
+
+## Private Mappings
+
+The preceding example considered VMAs created using the `MAP_SHARED` flag, which
+means that PTEs derived from the mapping should always use physical memory that
+represents the current state of the mapped file.[^mmap-dev-zero] Applications
+can alternatively pass the `MAP_PRIVATE` flag to create a *private mapping*.
+Private mappings are *copy-on-write*.
+
+Suppose that the application instead created a private mapping in the previous
+example. In Linux, the state of the system after a read page fault would be:
+
+    VMA:     VA:0x400000 -> /tmp/foo:0x0 (private)
+    Filemap:                /tmp/foo:0x0 -> PA:0x2fb000
+    PTE:     VA:0x400000 -----------------> PA:0x2fb000 (read-only)
+
+Now suppose the application attempts to write to VA:0x400000. For a shared
+mapping, the write would be propagated to PA:0x2fb000, and the kernel would be
+responsible for ensuring that the write is later propagated to the mapped file.
+For a private mapping, the write incurs another page fault since the PTE is
+marked read-only. In response, the kernel allocates physical memory to store the
+mapping's *private copy* of the file's contents, copies file contents to the
+allocated memory, and changes the PTE to map to the private copy. Supposing that
+the kernel chooses the physical memory at physical address (PA) 0x5ea000, the
+resulting state of the system is:
+
+    VMA:     VA:0x400000 -> /tmp/foo:0x0 (private)
+    Filemap:                /tmp/foo:0x0 -> PA:0x2fb000
+    PTE:     VA:0x400000 -----------------> PA:0x5ea000
+
+Note that the filemap mapping from /tmp/foo:0x0 to PA:0x2fb000 may still exist,
+but is now irrelevant to this mapping.
+
+[^mmap-dev-zero]: Modulo files with special mmap semantics such as `/dev/zero`.
+
+## Anonymous Mappings
+
+Instead of passing a file to the `mmap` system call, applications can instead
+request an *anonymous* mapping by passing the `MAP_ANONYMOUS` flag.
+Semantically, an anonymous mapping is essentially a mapping to an ephemeral file
+initially filled with zero bytes. Practically speaking, this is how shared
+anonymous mappings are implemented, but private anonymous mappings do not result
+in the creation of an ephemeral file; since there would be no way to modify the
+contents of the underlying file through a private mapping, all private anonymous
+mappings use a single shared page filled with zero bytes until copy-on-write
+occurs.
+
+# Virtual Memory in the Sentry
+
+The sentry implements application virtual memory atop a host kernel, introducing
+an additional level of indirection to the above.
+
+Consider the same scenario as in the previous section. Since the sentry handles
+application system calls, the effect of an application `mmap` system call is to
+create a VMA in the sentry (as opposed to the host kernel):
+
+    Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0
+
+When the application first incurs a page fault on this address, the host kernel
+delivers information about the page fault to the sentry in a platform-dependent
+manner, and the sentry handles the fault:
+
+1. The sentry allocates memory to store the contents of the required part of the
+   file, and copies file contents to the allocated memory. However, since the
+   sentry is implemented atop a host kernel, it does not configure mappings to
+   physical memory directly. Instead, mappable "memory" in the sentry is
+   represented by a host file descriptor and offset, since (as noted in
+   "Background") this is the memory mapping primitive provided by the host
+   kernel. In general, memory is allocated from a temporary host file using the
+   `filemem` package. Supposing that the sentry allocates offset 0x3000 from
+   host file "memory-file", the resulting state is:
+
+        Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0
+        Sentry filemap:                /tmp/foo:0x0 -> host:memory-file:0x3000
+
+2. The sentry stores the effective mapping from virtual address to host file in
+   a host VMA by invoking the `mmap` system call:
+
+        Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0
+        Sentry filemap:                /tmp/foo:0x0 -> host:memory-file:0x3000
+          Host VMA:     VA:0x400000 -----------------> host:memory-file:0x3000
+
+3. The sentry returns control to the application, which immediately incurs the
+   page fault again.[^mmap-populate] However, since a host VMA now exists for
+   the faulting virtual address, the host kernel now handles the page fault as
+   described in "Background":
+
+        Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0
+        Sentry filemap:                /tmp/foo:0x0 -> host:memory-file:0x3000
+          Host VMA:     VA:0x400000 -----------------> host:memory-file:0x3000
+          Host filemap:                                host:memory-file:0x3000 -> PA:0x2fb000
+          Host PTE:     VA:0x400000 --------------------------------------------> PA:0x2fb000
+
+Thus, from an implementation standpoint, host VMAs serve the same purpose in the
+sentry that PTEs do in Linux. As in Linux, sentry VMA and filemap state is
+independently mutable, and the desired state of host VMAs is derived from that
+state.
+
+[^mmap-populate]: The sentry could force the host kernel to establish PTEs when
+                  it creates the host VMA by passing the `MAP_POPULATE` flag to
+                  the `mmap` system call, but usually does not. This is because,
+                  to reduce the number of page faults that require handling by
+                  the sentry and (correspondingly) the number of host `mmap`
+                  system calls, the sentry usually creates host VMAs that are
+                  much larger than the single faulting page.
+
+## Private Mappings
+
+The sentry implements private mappings consistently with Linux. Before
+copy-on-write, the private mapping example given in the Background results in:
+
+    Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0 (private)
+    Sentry filemap:                /tmp/foo:0x0 -> host:memory-file:0x3000
+      Host VMA:     VA:0x400000 -----------------> host:memory-file:0x3000 (read-only)
+      Host filemap:                                host:memory-file:0x3000 -> PA:0x2fb000
+      Host PTE:     VA:0x400000 --------------------------------------------> PA:0x2fb000 (read-only)
+
+When the application attempts to write to this address, the host kernel delivers
+information about the resulting page fault to the sentry. Analogous to Linux,
+the sentry allocates memory to store the mapping's private copy of the file's
+contents, copies file contents to the allocated memory, and changes the host VMA
+to map to the private copy. Supposing that the sentry chooses the offset 0x4000
+in host file `memory-file` to store the private copy, the state of the system
+after copy-on-write is:
+
+    Sentry VMA:     VA:0x400000 -> /tmp/foo:0x0 (private)
+    Sentry filemap:                /tmp/foo:0x0 -> host:memory-file:0x3000
+      Host VMA:     VA:0x400000 -----------------> host:memory-file:0x4000
+      Host filemap:                                host:memory-file:0x4000 -> PA:0x5ea000
+      Host PTE:     VA:0x400000 --------------------------------------------> PA:0x5ea000
+
+However, this highlights an important difference between Linux and the sentry.
+In Linux, page tables are concrete (architecture-dependent) data structures
+owned by the kernel. Conversely, the sentry has the ability to create and
+destroy host VMAs using host system calls, but it does not have direct access to
+their state. Thus, as written, if the application invokes the `munmap` system
+call to remove the sentry VMA, it is non-trivial for the sentry to determine
+that it should deallocate `host:memory-file:0x4000`. This implies that the
+sentry must retain information about the host VMAs that it has created.
+
+## Anonymous Mappings
+
+The sentry implements anonymous mappings consistently with Linux, except that
+there is no shared zero page.
+
+# Implementation Constructs
+
+In Linux:
+
+- A virtual address space is represented by `struct mm_struct`.
+
+- VMAs are represented by `struct vm_area_struct`, stored in `struct
+  mm_struct::mmap`.
+
+- Mappings from file offsets to physical memory are stored in `struct
+  address_space`.
+
+- Reverse mappings from file offsets to virtual mappings are stored in `struct
+  address_space::i_mmap`.
+
+- Physical memory pages are represented by a pointer to `struct page` or an
+  index called a *page frame number* (PFN), represented by `pfn_t`.
+
+- PTEs are represented by architecture-dependent type `pte_t`, stored in a table
+  hierarchy rooted at `struct mm_struct::pgd`.
+
+In the sentry:
+
+- A virtual address space is represented by type [`mm.MemoryManager`][mm].
+
+- Sentry VMAs are represented by type [`mm.vma`][mm], stored in
+  `mm.MemoryManager.vmas`.
+
+- Mappings from sentry file offsets to host file offsets are abstracted through
+  interface method [`memmap.Mappable.Translate`][memmap].
+
+- Reverse mappings from sentry file offsets to virtual mappings are abstracted
+  through interface methods [`memmap.Mappable.AddMapping` and
+  `memmap.Mappable.RemoveMapping`][memmap].
+
+- Host files that may be mapped into host VMAs are represented by type
+  [`platform.File`][platform].
+
+- Host VMAs are represented in the sentry by type [`mm.pma`][mm] ("platform
+  mapping area"), stored in `mm.MemoryManager.pmas`.
+
+- Creation and destruction of host VMAs is abstracted through interface methods
+  [`platform.AddressSpace.MapFile` and `platform.AddressSpace.Unmap`][platform].
+
+[filemem]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/platform/filemem/filemem.go
+[memmap]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/memmap/memmap.go
+[mm]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/mm/mm.go
+[platform]: https://gvisor.googlesource.com/gvisor/+/master/pkg/sentry/platform/platform.go
diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go
new file mode 100644
index 000000000..4dd67b1ea
--- /dev/null
+++ b/pkg/sentry/mm/address_space.go
@@ -0,0 +1,223 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/atomicbitops"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// AddressSpace returns the platform.AddressSpace bound to mm.
+//
+// Preconditions: The caller must have called mm.Activate().
+func (mm *MemoryManager) AddressSpace() platform.AddressSpace {
+	if atomic.LoadInt32(&mm.active) == 0 {
+		panic("trying to use inactive address space?")
+	}
+	return mm.as
+}
+
+// Activate ensures this MemoryManager has a platform.AddressSpace.
+//
+// The caller must not hold any locks when calling Activate.
+//
+// When this MemoryManager is no longer needed by a task, it should call
+// Deactivate to release the reference.
+func (mm *MemoryManager) Activate() error {
+	// Fast path: the MemoryManager already has an active
+	// platform.AddressSpace, and we just need to indicate that we need it too.
+	if atomicbitops.IncUnlessZeroInt32(&mm.active) {
+		return nil
+	}
+
+	for {
+		// Slow path: may need to synchronize with other goroutines changing
+		// mm.active to or from zero.
+		mm.activeMu.Lock()
+		// Inline Unlock instead of using a defer for performance since this
+		// method is commonly in the hot-path.
+
+		// Check if we raced with another goroutine performing activation.
+		if atomic.LoadInt32(&mm.active) > 0 {
+			// This can't race; Deactivate can't decrease mm.active from 1 to 0
+			// without holding activeMu.
+			atomic.AddInt32(&mm.active, 1)
+			mm.activeMu.Unlock()
+			return nil
+		}
+
+		// Do we have a context? If so, then we never unmapped it. This can
+		// only be the case if !mm.p.CooperativelySchedulesAddressSpace().
+		if mm.as != nil {
+			atomic.StoreInt32(&mm.active, 1)
+			mm.activeMu.Unlock()
+			return nil
+		}
+
+		// Get a new address space. We must force unmapping by passing nil to
+		// NewAddressSpace if requested. (As in the nil interface object, not a
+		// typed nil.)
+		mappingsID := (interface{})(mm)
+		if mm.unmapAllOnActivate {
+			mappingsID = nil
+		}
+		as, c, err := mm.p.NewAddressSpace(mappingsID)
+		if err != nil {
+			mm.activeMu.Unlock()
+			return err
+		}
+		if as == nil {
+			// AddressSpace is unavailable, we must wait.
+			//
+			// activeMu must not be held while waiting, as the user
+			// of the address space we are waiting on may attempt
+			// to take activeMu.
+			//
+			// Don't call UninterruptibleSleepStart to register the
+			// wait to allow the watchdog stuck task to trigger in
+			// case a process is starved waiting for the address
+			// space.
+			mm.activeMu.Unlock()
+			<-c
+			continue
+		}
+
+		// Okay, we could restore all mappings at this point.
+		// But forget that. Let's just let them fault in.
+		mm.as = as
+
+		// Unmapping is done, if necessary.
+		mm.unmapAllOnActivate = false
+
+		// Now that m.as has been assigned, we can set m.active to a non-zero value
+		// to enable the fast path.
+		atomic.StoreInt32(&mm.active, 1)
+
+		mm.activeMu.Unlock()
+		return nil
+	}
+}
+
+// Deactivate releases a release to the MemoryManager.
+func (mm *MemoryManager) Deactivate() error {
+	// Fast path: this is not the last goroutine to deactivate the
+	// MemoryManager.
+	if atomicbitops.DecUnlessOneInt32(&mm.active) {
+		return nil
+	}
+
+	mm.activeMu.Lock()
+	// Same as Activate.
+
+	// Still active?
+	if atomic.AddInt32(&mm.active, -1) > 0 {
+		mm.activeMu.Unlock()
+		return nil
+	}
+
+	// Can we hold on to the address space?
+	if !mm.p.CooperativelySchedulesAddressSpace() {
+		mm.activeMu.Unlock()
+		return nil
+	}
+
+	// Release the address space.
+	if err := mm.as.Release(); err != nil {
+		atomic.StoreInt32(&mm.active, 1)
+		mm.activeMu.Unlock()
+		return err
+	}
+
+	// Lost it.
+	mm.as = nil
+	mm.activeMu.Unlock()
+	return nil
+}
+
+// mapASLocked maps addresses in ar into mm.as. If precommit is true, mappings
+// for all addresses in ar should be precommitted.
+//
+// Preconditions: mm.activeMu must be locked. mm.as != nil. ar.Length() != 0.
+// ar must be page-aligned. pseg.Range().Contains(ar.Start).
+func (mm *MemoryManager) mapASLocked(pseg pmaIterator, ar usermem.AddrRange, precommit bool) error {
+	// By default, map entire pmas at a time, under the assumption that there
+	// is no cost to mapping more of a pma than necessary.
+	mapAR := usermem.AddrRange{0, ^usermem.Addr(usermem.PageSize - 1)}
+	if precommit {
+		// When explicitly precommitting, only map ar, since overmapping may
+		// incur unexpected resource usage.
+		mapAR = ar
+	} else if mapUnit := mm.p.MapUnit(); mapUnit != 0 {
+		// Limit the range we map to ar, aligned to mapUnit.
+		mapMask := usermem.Addr(mapUnit - 1)
+		mapAR.Start = ar.Start &^ mapMask
+		// If rounding ar.End up overflows, just keep the existing mapAR.End.
+		if end := (ar.End + mapMask) &^ mapMask; end >= ar.End {
+			mapAR.End = end
+		}
+	}
+	if checkInvariants {
+		if !mapAR.IsSupersetOf(ar) {
+			panic(fmt.Sprintf("mapAR %#v is not a superset of ar %#v", mapAR, ar))
+		}
+	}
+
+	for {
+		pma := pseg.ValuePtr()
+		pmaAR := pseg.Range()
+		pmaMapAR := pmaAR.Intersect(mapAR)
+		perms := pma.vmaEffectivePerms
+		if pma.needCOW {
+			perms.Write = false
+		}
+		if err := pma.file.MapInto(mm.as, pmaMapAR.Start, pseg.fileRangeOf(pmaMapAR), perms, precommit); err != nil {
+			return err
+		}
+		// Since this checks ar.End and not mapAR.End, we will never map a pma
+		// that is not required.
+		if ar.End <= pmaAR.End {
+			return nil
+		}
+		pseg = pseg.NextSegment()
+	}
+}
+
+// unmapASLocked removes all AddressSpace mappings for addresses in ar.
+//
+// Preconditions: mm.activeMu must be locked.
+func (mm *MemoryManager) unmapASLocked(ar usermem.AddrRange) {
+	if mm.as == nil {
+		// No AddressSpace? Force all mappings to be unmapped on the next
+		// Activate.
+		mm.unmapAllOnActivate = true
+		return
+	}
+
+	// unmapASLocked doesn't require vmas or pmas to exist for ar, so it can be
+	// passed ranges that include addresses that can't be mapped by the
+	// application.
+	ar = ar.Intersect(mm.applicationAddrRange())
+
+	// Note that this AddressSpace may or may not be active. If the
+	// platform does not require cooperative sharing of AddressSpaces, they
+	// are retained between Deactivate/Activate calls. Despite not being
+	// active, it is still valid to perform operations on these address
+	// spaces.
+	mm.as.Unmap(ar.Start, uint64(ar.Length()))
+}
diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go
new file mode 100644
index 000000000..992bde5a5
--- /dev/null
+++ b/pkg/sentry/mm/aio_context.go
@@ -0,0 +1,377 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// aioManager creates and manages asynchronous I/O contexts.
+type aioManager struct {
+	// mu protects below.
+	mu sync.Mutex `state:"nosave"`
+
+	// aioContexts is the set of asynchronous I/O contexts.
+	contexts map[uint64]*AIOContext
+}
+
+func (a *aioManager) destroy() {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+
+	for _, ctx := range a.contexts {
+		ctx.destroy()
+	}
+}
+
+// newAIOContext creates a new context for asynchronous I/O.
+//
+// Returns false if 'id' is currently in use.
+func (a *aioManager) newAIOContext(events uint32, id uint64) bool {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+
+	if _, ok := a.contexts[id]; ok {
+		return false
+	}
+
+	a.contexts[id] = &AIOContext{
+		done:           make(chan struct{}, 1),
+		maxOutstanding: events,
+	}
+	return true
+}
+
+// destroyAIOContext destroys an asynchronous I/O context.
+//
+// False is returned if the context does not exist.
+func (a *aioManager) destroyAIOContext(id uint64) bool {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	ctx, ok := a.contexts[id]
+	if !ok {
+		return false
+	}
+	delete(a.contexts, id)
+	ctx.destroy()
+	return true
+}
+
+// lookupAIOContext looks up the given context.
+//
+// Returns false if context does not exist.
+func (a *aioManager) lookupAIOContext(id uint64) (*AIOContext, bool) {
+	a.mu.Lock()
+	defer a.mu.Unlock()
+	ctx, ok := a.contexts[id]
+	return ctx, ok
+}
+
+// ioResult is a completed I/O operation.
+type ioResult struct {
+	data interface{}
+	ioEntry
+}
+
+// AIOContext is a single asynchronous I/O context.
+type AIOContext struct {
+	// done is the notification channel used for all requests.
+	done chan struct{} `state:"nosave"`
+
+	// mu protects below.
+	mu sync.Mutex `state:"nosave"`
+
+	// results is the set of completed requests.
+	results ioList
+
+	// maxOutstanding is the maximum number of outstanding entries; this value
+	// is immutable.
+	maxOutstanding uint32
+
+	// outstanding is the number of requests outstanding; this will effectively
+	// be the number of entries in the result list or that are expected to be
+	// added to the result list.
+	outstanding uint32
+
+	// dead is set when the context is destroyed.
+	dead bool `state:"zerovalue"`
+}
+
+// destroy marks the context dead.
+func (ctx *AIOContext) destroy() {
+	ctx.mu.Lock()
+	defer ctx.mu.Unlock()
+	ctx.dead = true
+	if ctx.outstanding == 0 {
+		close(ctx.done)
+	}
+}
+
+// Prepare reserves space for a new request, returning true if available.
+// Returns false if the context is busy.
+func (ctx *AIOContext) Prepare() bool {
+	ctx.mu.Lock()
+	defer ctx.mu.Unlock()
+	if ctx.outstanding >= ctx.maxOutstanding {
+		return false
+	}
+	ctx.outstanding++
+	return true
+}
+
+// PopRequest pops a completed request if available, this function does not do
+// any blocking. Returns false if no request is available.
+func (ctx *AIOContext) PopRequest() (interface{}, bool) {
+	ctx.mu.Lock()
+	defer ctx.mu.Unlock()
+
+	// Is there anything ready?
+	if e := ctx.results.Front(); e != nil {
+		ctx.results.Remove(e)
+		ctx.outstanding--
+		if ctx.outstanding == 0 && ctx.dead {
+			close(ctx.done)
+		}
+		return e.data, true
+	}
+	return nil, false
+}
+
+// FinishRequest finishes a pending request. It queues up the data
+// and notifies listeners.
+func (ctx *AIOContext) FinishRequest(data interface{}) {
+	ctx.mu.Lock()
+	defer ctx.mu.Unlock()
+
+	// Push to the list and notify opportunistically. The channel notify
+	// here is guaranteed to be safe because outstanding must be non-zero.
+	// The done channel is only closed when outstanding reaches zero.
+	ctx.results.PushBack(&ioResult{data: data})
+
+	select {
+	case ctx.done <- struct{}{}:
+	default:
+	}
+}
+
+// WaitChannel returns a channel that is notified when an AIO request is
+// completed.
+//
+// The boolean return value indicates whether or not the context is active.
+func (ctx *AIOContext) WaitChannel() (chan struct{}, bool) {
+	ctx.mu.Lock()
+	defer ctx.mu.Unlock()
+	if ctx.outstanding == 0 && ctx.dead {
+		return nil, false
+	}
+	return ctx.done, true
+}
+
+// aioMappable implements memmap.MappingIdentity and memmap.Mappable for AIO
+// ring buffers.
+type aioMappable struct {
+	refs.AtomicRefCount
+
+	p  platform.Platform
+	fr platform.FileRange
+}
+
+var aioRingBufferSize = uint64(usermem.Addr(linux.AIORingSize).MustRoundUp())
+
+func newAIOMappable(p platform.Platform) (*aioMappable, error) {
+	fr, err := p.Memory().Allocate(aioRingBufferSize, usage.Anonymous)
+	if err != nil {
+		return nil, err
+	}
+	return &aioMappable{p: p, fr: fr}, nil
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+func (m *aioMappable) DecRef() {
+	m.AtomicRefCount.DecRefWithDestructor(func() {
+		m.p.Memory().DecRef(m.fr)
+	})
+}
+
+// MappedName implements memmap.MappingIdentity.MappedName.
+func (m *aioMappable) MappedName(ctx context.Context) string {
+	return "[aio]"
+}
+
+// DeviceID implements memmap.MappingIdentity.DeviceID.
+func (m *aioMappable) DeviceID() uint64 {
+	return 0
+}
+
+// InodeID implements memmap.MappingIdentity.InodeID.
+func (m *aioMappable) InodeID() uint64 {
+	return 0
+}
+
+// Msync implements memmap.MappingIdentity.Msync.
+func (m *aioMappable) Msync(ctx context.Context, mr memmap.MappableRange) error {
+	// Linux: aio_ring_fops.fsync == NULL
+	return syserror.EINVAL
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (m *aioMappable) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	// Don't allow mappings to be expanded (in Linux, fs/aio.c:aio_ring_mmap()
+	// sets VM_DONTEXPAND).
+	if offset != 0 || uint64(ar.Length()) != aioRingBufferSize {
+		return syserror.EFAULT
+	}
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (m *aioMappable) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (m *aioMappable) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	// Don't allow mappings to be expanded (in Linux, fs/aio.c:aio_ring_mmap()
+	// sets VM_DONTEXPAND).
+	if offset != 0 || uint64(dstAR.Length()) != aioRingBufferSize {
+		return syserror.EFAULT
+	}
+	// Require that the mapping correspond to a live AIOContext. Compare
+	// Linux's fs/aio.c:aio_ring_mremap().
+	mm, ok := ms.(*MemoryManager)
+	if !ok {
+		return syserror.EINVAL
+	}
+	am := &mm.aioManager
+	am.mu.Lock()
+	defer am.mu.Unlock()
+	oldID := uint64(srcAR.Start)
+	aioCtx, ok := am.contexts[oldID]
+	if !ok {
+		return syserror.EINVAL
+	}
+	aioCtx.mu.Lock()
+	defer aioCtx.mu.Unlock()
+	if aioCtx.dead {
+		return syserror.EINVAL
+	}
+	// Use the new ID for the AIOContext.
+	am.contexts[uint64(dstAR.Start)] = aioCtx
+	delete(am.contexts, oldID)
+	return nil
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (m *aioMappable) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	var err error
+	if required.End > m.fr.Length() {
+		err = &memmap.BusError{syserror.EFAULT}
+	}
+	if source := optional.Intersect(memmap.MappableRange{0, m.fr.Length()}); source.Length() != 0 {
+		return []memmap.Translation{
+			{
+				Source: source,
+				File:   m.p.Memory(),
+				Offset: m.fr.Start + source.Start,
+			},
+		}, err
+	}
+	return nil, err
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (m *aioMappable) InvalidateUnsavable(ctx context.Context) error {
+	return nil
+}
+
+// NewAIOContext creates a new context for asynchronous I/O.
+//
+// NewAIOContext is analogous to Linux's fs/aio.c:ioctx_alloc().
+func (mm *MemoryManager) NewAIOContext(ctx context.Context, events uint32) (uint64, error) {
+	// libaio get_ioevents() expects context "handle" to be a valid address.
+	// libaio peeks inside looking for a magic number. This function allocates
+	// a page per context and keeps it set to zeroes to ensure it will not
+	// match AIO_RING_MAGIC and make libaio happy.
+	m, err := newAIOMappable(mm.p)
+	if err != nil {
+		return 0, err
+	}
+	defer m.DecRef()
+	addr, err := mm.MMap(ctx, memmap.MMapOpts{
+		Length:          aioRingBufferSize,
+		MappingIdentity: m,
+		Mappable:        m,
+		// TODO: Linux does "do_mmap_pgoff(..., PROT_READ |
+		// PROT_WRITE, ...)" in fs/aio.c:aio_setup_ring(); why do we make this
+		// mapping read-only?
+		Perms:    usermem.Read,
+		MaxPerms: usermem.Read,
+	})
+	if err != nil {
+		return 0, err
+	}
+	id := uint64(addr)
+	if !mm.aioManager.newAIOContext(events, id) {
+		mm.MUnmap(ctx, addr, aioRingBufferSize)
+		return 0, syserror.EINVAL
+	}
+	return id, nil
+}
+
+// DestroyAIOContext destroys an asynchronous I/O context. It returns false if
+// the context does not exist.
+func (mm *MemoryManager) DestroyAIOContext(ctx context.Context, id uint64) bool {
+	if _, ok := mm.LookupAIOContext(ctx, id); !ok {
+		return false
+	}
+
+	// Only unmaps after it assured that the address is a valid aio context to
+	// prevent random memory from been unmapped.
+	//
+	// Note: It's possible to unmap this address and map something else into
+	// the same address. Then it would be unmapping memory that it doesn't own.
+	// This is, however, the way Linux implements AIO. Keeps the same [weird]
+	// semantics in case anyone relies on it.
+	mm.MUnmap(ctx, usermem.Addr(id), aioRingBufferSize)
+
+	return mm.aioManager.destroyAIOContext(id)
+}
+
+// LookupAIOContext looks up the given context. It returns false if the context
+// does not exist.
+func (mm *MemoryManager) LookupAIOContext(ctx context.Context, id uint64) (*AIOContext, bool) {
+	aioCtx, ok := mm.aioManager.lookupAIOContext(id)
+	if !ok {
+		return nil, false
+	}
+
+	// Protect against 'ids' that are inaccessible (Linux also reads 4 bytes
+	// from id).
+	var buf [4]byte
+	_, err := mm.CopyIn(ctx, usermem.Addr(id), buf[:], usermem.IOOpts{})
+	if err != nil {
+		return nil, false
+	}
+
+	return aioCtx, true
+}
diff --git a/pkg/sentry/mm/aio_context_state.go b/pkg/sentry/mm/aio_context_state.go
new file mode 100644
index 000000000..1a5e56f8e
--- /dev/null
+++ b/pkg/sentry/mm/aio_context_state.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+// afterLoad is invoked by stateify.
+func (a *AIOContext) afterLoad() {
+	a.done = make(chan struct{}, 1)
+}
diff --git a/pkg/sentry/mm/debug.go b/pkg/sentry/mm/debug.go
new file mode 100644
index 000000000..56d0490f0
--- /dev/null
+++ b/pkg/sentry/mm/debug.go
@@ -0,0 +1,98 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"bytes"
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+const (
+	// If checkInvariants is true, perform runtime checks for invariants
+	// expected by the mm package. This is normally disabled since MM is a
+	// significant hot path in general, and some such checks (notably
+	// memmap.CheckTranslateResult) are very expensive.
+	checkInvariants = false
+
+	// If logIOErrors is true, log I/O errors that originate from MM before
+	// converting them to EFAULT.
+	logIOErrors = false
+)
+
+// String implements fmt.Stringer.String.
+func (mm *MemoryManager) String() string {
+	return mm.DebugString(context.Background())
+}
+
+// DebugString returns a string containing information about mm for debugging.
+func (mm *MemoryManager) DebugString(ctx context.Context) string {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	mm.activeMu.RLock()
+	defer mm.activeMu.RUnlock()
+	return mm.debugStringLocked(ctx)
+}
+
+// Preconditions: mm.mappingMu and mm.activeMu must be locked.
+func (mm *MemoryManager) debugStringLocked(ctx context.Context) string {
+	var b bytes.Buffer
+	b.WriteString("VMAs:\n")
+	for vseg := mm.vmas.FirstSegment(); vseg.Ok(); vseg = vseg.NextSegment() {
+		b.Write(mm.vmaMapsEntryLocked(ctx, vseg))
+	}
+	b.WriteString("PMAs:\n")
+	for pseg := mm.pmas.FirstSegment(); pseg.Ok(); pseg = pseg.NextSegment() {
+		b.Write(pseg.debugStringEntryLocked())
+	}
+	return string(b.Bytes())
+}
+
+// Preconditions: mm.activeMu must be locked.
+func (pseg pmaIterator) debugStringEntryLocked() []byte {
+	var b bytes.Buffer
+
+	fmt.Fprintf(&b, "%08x-%08x ", pseg.Start(), pseg.End())
+
+	pma := pseg.ValuePtr()
+	if pma.vmaEffectivePerms.Read {
+		b.WriteByte('r')
+	} else {
+		b.WriteByte('-')
+	}
+	if pma.vmaEffectivePerms.Write {
+		if pma.needCOW {
+			b.WriteByte('c')
+		} else {
+			b.WriteByte('w')
+		}
+	} else {
+		b.WriteByte('-')
+	}
+	if pma.vmaEffectivePerms.Execute {
+		b.WriteByte('x')
+	} else {
+		b.WriteByte('-')
+	}
+	if pma.private {
+		b.WriteByte('p')
+	} else {
+		b.WriteByte('s')
+	}
+
+	fmt.Fprintf(&b, " %08x %T\n", pma.off, pma.file)
+	return b.Bytes()
+}
diff --git a/pkg/sentry/mm/io.go b/pkg/sentry/mm/io.go
new file mode 100644
index 000000000..cac81a59d
--- /dev/null
+++ b/pkg/sentry/mm/io.go
@@ -0,0 +1,604 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// There are two supported ways to copy data to/from application virtual
+// memory:
+//
+// 1. Internally-mapped copying: Determine the platform.File that backs the
+// copied-to/from virtual address, obtain a mapping of its pages, and read or
+// write to the mapping.
+//
+// 2. AddressSpace copying: If platform.Platform.SupportsAddressSpaceIO() is
+// true, AddressSpace permissions are applicable, and an AddressSpace is
+// available, copy directly through the AddressSpace, handling faults as
+// needed.
+//
+// (Given that internally-mapped copying requires that backing memory is always
+// implemented using a host file descriptor, we could also preadv/pwritev to it
+// instead. But this would incur a host syscall for each use of the mapped
+// page, whereas mmap is a one-time cost.)
+//
+// The fixed overhead of internally-mapped copying is expected to be higher
+// than that of AddressSpace copying since the former always needs to translate
+// addresses, whereas the latter only needs to do so when faults occur.
+// However, the throughput of internally-mapped copying is expected to be
+// somewhat higher than that of AddressSpace copying due to the high cost of
+// page faults and because implementations of the latter usually rely on
+// safecopy, which doesn't use AVX registers. So we prefer to use AddressSpace
+// copying (when available) for smaller copies, and switch to internally-mapped
+// copying once a size threshold is exceeded.
+const (
+	// copyMapMinBytes is the size threshold for switching to internally-mapped
+	// copying in CopyOut, CopyIn, and ZeroOut.
+	copyMapMinBytes = 32 << 10 // 32 KB
+
+	// rwMapMinBytes is the size threshold for switching to internally-mapped
+	// copying in CopyOutFrom and CopyInTo. It's lower than copyMapMinBytes
+	// since AddressSpace copying in this case requires additional buffering;
+	// see CopyOutFrom for details.
+	rwMapMinBytes = 512
+)
+
+// checkIORange is similar to usermem.Addr.ToRange, but applies bounds checks
+// consistent with Linux's arch/x86/include/asm/uaccess.h:access_ok().
+//
+// Preconditions: length >= 0.
+func (mm *MemoryManager) checkIORange(addr usermem.Addr, length int64) (usermem.AddrRange, bool) {
+	// Note that access_ok() constrains end even if length == 0.
+	ar, ok := addr.ToRange(uint64(length))
+	return ar, (ok && ar.End <= mm.layout.MaxAddr)
+}
+
+// checkIOVec applies bound checks consistent with Linux's
+// arch/x86/include/asm/uaccess.h:access_ok() to ars.
+func (mm *MemoryManager) checkIOVec(ars usermem.AddrRangeSeq) bool {
+	for !ars.IsEmpty() {
+		ar := ars.Head()
+		if _, ok := mm.checkIORange(ar.Start, int64(ar.Length())); !ok {
+			return false
+		}
+		ars = ars.Tail()
+	}
+	return true
+}
+
+func (mm *MemoryManager) asioEnabled(opts usermem.IOOpts) bool {
+	return mm.haveASIO && !opts.IgnorePermissions && opts.AddressSpaceActive
+}
+
+// translateIOError converts errors to EFAULT, as is usually reported for all
+// I/O errors originating from MM in Linux.
+func translateIOError(ctx context.Context, err error) error {
+	if err == nil {
+		return nil
+	}
+	if logIOErrors {
+		ctx.Debugf("MM I/O error: %v", err)
+	}
+	return syserror.EFAULT
+}
+
+// CopyOut implements usermem.IO.CopyOut.
+func (mm *MemoryManager) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) {
+	ar, ok := mm.checkIORange(addr, int64(len(src)))
+	if !ok {
+		return 0, syserror.EFAULT
+	}
+
+	if len(src) == 0 {
+		return 0, nil
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.asioEnabled(opts) && len(src) < copyMapMinBytes {
+		return mm.asCopyOut(ctx, addr, src)
+	}
+
+	// Go through internal mappings.
+	n64, err := mm.withInternalMappings(ctx, ar, usermem.Write, opts.IgnorePermissions, func(ims safemem.BlockSeq) (uint64, error) {
+		n, err := safemem.CopySeq(ims, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(src)))
+		return n, translateIOError(ctx, err)
+	})
+	return int(n64), err
+}
+
+func (mm *MemoryManager) asCopyOut(ctx context.Context, addr usermem.Addr, src []byte) (int, error) {
+	var done int
+	for {
+		n, err := mm.as.CopyOut(addr+usermem.Addr(done), src[done:])
+		done += n
+		if err == nil {
+			return done, nil
+		}
+		if f, ok := err.(platform.SegmentationFault); ok {
+			ar, _ := addr.ToRange(uint64(len(src)))
+			if err := mm.handleASIOFault(ctx, f.Addr, ar, usermem.Write); err != nil {
+				return done, err
+			}
+			continue
+		}
+		return done, translateIOError(ctx, err)
+	}
+}
+
+// CopyIn implements usermem.IO.CopyIn.
+func (mm *MemoryManager) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) {
+	ar, ok := mm.checkIORange(addr, int64(len(dst)))
+	if !ok {
+		return 0, syserror.EFAULT
+	}
+
+	if len(dst) == 0 {
+		return 0, nil
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.asioEnabled(opts) && len(dst) < copyMapMinBytes {
+		return mm.asCopyIn(ctx, addr, dst)
+	}
+
+	// Go through internal mappings.
+	n64, err := mm.withInternalMappings(ctx, ar, usermem.Read, opts.IgnorePermissions, func(ims safemem.BlockSeq) (uint64, error) {
+		n, err := safemem.CopySeq(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(dst)), ims)
+		return n, translateIOError(ctx, err)
+	})
+	return int(n64), err
+}
+
+func (mm *MemoryManager) asCopyIn(ctx context.Context, addr usermem.Addr, dst []byte) (int, error) {
+	var done int
+	for {
+		n, err := mm.as.CopyIn(addr+usermem.Addr(done), dst[done:])
+		done += n
+		if err == nil {
+			return done, nil
+		}
+		if f, ok := err.(platform.SegmentationFault); ok {
+			ar, _ := addr.ToRange(uint64(len(dst)))
+			if err := mm.handleASIOFault(ctx, f.Addr, ar, usermem.Read); err != nil {
+				return done, err
+			}
+			continue
+		}
+		return done, translateIOError(ctx, err)
+	}
+}
+
+// ZeroOut implements usermem.IO.ZeroOut.
+func (mm *MemoryManager) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) {
+	ar, ok := mm.checkIORange(addr, toZero)
+	if !ok {
+		return 0, syserror.EFAULT
+	}
+
+	if toZero == 0 {
+		return 0, nil
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.asioEnabled(opts) && toZero < copyMapMinBytes {
+		return mm.asZeroOut(ctx, addr, toZero)
+	}
+
+	// Go through internal mappings.
+	return mm.withInternalMappings(ctx, ar, usermem.Write, opts.IgnorePermissions, func(dsts safemem.BlockSeq) (uint64, error) {
+		n, err := safemem.ZeroSeq(dsts)
+		return n, translateIOError(ctx, err)
+	})
+}
+
+func (mm *MemoryManager) asZeroOut(ctx context.Context, addr usermem.Addr, toZero int64) (int64, error) {
+	var done int64
+	for {
+		n, err := mm.as.ZeroOut(addr+usermem.Addr(done), uintptr(toZero-done))
+		done += int64(n)
+		if err == nil {
+			return done, nil
+		}
+		if f, ok := err.(platform.SegmentationFault); ok {
+			ar, _ := addr.ToRange(uint64(toZero))
+			if err := mm.handleASIOFault(ctx, f.Addr, ar, usermem.Write); err != nil {
+				return done, err
+			}
+			continue
+		}
+		return done, translateIOError(ctx, err)
+	}
+}
+
+// CopyOutFrom implements usermem.IO.CopyOutFrom.
+func (mm *MemoryManager) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq, src safemem.Reader, opts usermem.IOOpts) (int64, error) {
+	if !mm.checkIOVec(ars) {
+		return 0, syserror.EFAULT
+	}
+
+	if ars.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.asioEnabled(opts) && ars.NumBytes() < rwMapMinBytes {
+		// We have to introduce a buffered copy, instead of just passing a
+		// safemem.BlockSeq representing addresses in the AddressSpace to src.
+		// This is because usermem.IO.CopyOutFrom() guarantees that it calls
+		// src.ReadToBlocks() at most once, which is incompatible with handling
+		// faults between calls. In the future, this is probably best resolved
+		// by introducing a CopyOutFrom variant or option that allows it to
+		// call src.ReadToBlocks() any number of times.
+		//
+		// This issue applies to CopyInTo as well.
+		buf := make([]byte, int(ars.NumBytes()))
+		bufN, bufErr := src.ReadToBlocks(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)))
+		var done int64
+		for done < int64(bufN) {
+			ar := ars.Head()
+			cplen := int64(ar.Length())
+			if cplen > int64(bufN)-done {
+				cplen = int64(bufN) - done
+			}
+			n, err := mm.asCopyOut(ctx, ar.Start, buf[int(done):int(done+cplen)])
+			done += int64(n)
+			if err != nil {
+				return done, err
+			}
+			ars = ars.Tail()
+		}
+		// Do not convert errors returned by src to EFAULT.
+		return done, bufErr
+	}
+
+	// Go through internal mappings.
+	return mm.withVecInternalMappings(ctx, ars, usermem.Write, opts.IgnorePermissions, src.ReadToBlocks)
+}
+
+// CopyInTo implements usermem.IO.CopyInTo.
+func (mm *MemoryManager) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) {
+	if !mm.checkIOVec(ars) {
+		return 0, syserror.EFAULT
+	}
+
+	if ars.NumBytes() == 0 {
+		return 0, nil
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.asioEnabled(opts) && ars.NumBytes() < rwMapMinBytes {
+		buf := make([]byte, int(ars.NumBytes()))
+		var done int
+		var bufErr error
+		for !ars.IsEmpty() {
+			ar := ars.Head()
+			var n int
+			n, bufErr = mm.asCopyIn(ctx, ar.Start, buf[done:done+int(ar.Length())])
+			done += n
+			if bufErr != nil {
+				break
+			}
+			ars = ars.Tail()
+		}
+		n, err := dst.WriteFromBlocks(safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:done])))
+		if err != nil {
+			return int64(n), err
+		}
+		// Do not convert errors returned by dst to EFAULT.
+		return int64(n), bufErr
+	}
+
+	// Go through internal mappings.
+	return mm.withVecInternalMappings(ctx, ars, usermem.Read, opts.IgnorePermissions, dst.WriteFromBlocks)
+}
+
+// SwapUint32 implements usermem.IO.SwapUint32.
+func (mm *MemoryManager) SwapUint32(ctx context.Context, addr usermem.Addr, new uint32, opts usermem.IOOpts) (uint32, error) {
+	ar, ok := mm.checkIORange(addr, 4)
+	if !ok {
+		return 0, syserror.EFAULT
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.haveASIO && opts.AddressSpaceActive && !opts.IgnorePermissions {
+		for {
+			old, err := mm.as.SwapUint32(addr, new)
+			if err == nil {
+				return old, nil
+			}
+			if f, ok := err.(platform.SegmentationFault); ok {
+				if err := mm.handleASIOFault(ctx, f.Addr, ar, usermem.ReadWrite); err != nil {
+					return 0, err
+				}
+				continue
+			}
+			return 0, translateIOError(ctx, err)
+		}
+	}
+
+	// Go through internal mappings.
+	var old uint32
+	_, err := mm.withInternalMappings(ctx, ar, usermem.ReadWrite, opts.IgnorePermissions, func(ims safemem.BlockSeq) (uint64, error) {
+		if ims.NumBlocks() != 1 || ims.NumBytes() != 4 {
+			// Atomicity is unachievable across mappings.
+			return 0, syserror.EFAULT
+		}
+		im := ims.Head()
+		var err error
+		old, err = safemem.SwapUint32(im, new)
+		if err != nil {
+			return 0, translateIOError(ctx, err)
+		}
+		return 4, nil
+	})
+	return old, err
+}
+
+// CompareAndSwapUint32 implements usermem.IO.CompareAndSwapUint32.
+func (mm *MemoryManager) CompareAndSwapUint32(ctx context.Context, addr usermem.Addr, old, new uint32, opts usermem.IOOpts) (uint32, error) {
+	ar, ok := mm.checkIORange(addr, 4)
+	if !ok {
+		return 0, syserror.EFAULT
+	}
+
+	// Do AddressSpace IO if applicable.
+	if mm.haveASIO && opts.AddressSpaceActive && !opts.IgnorePermissions {
+		for {
+			prev, err := mm.as.CompareAndSwapUint32(addr, old, new)
+			if err == nil {
+				return prev, nil
+			}
+			if f, ok := err.(platform.SegmentationFault); ok {
+				if err := mm.handleASIOFault(ctx, f.Addr, ar, usermem.ReadWrite); err != nil {
+					return 0, err
+				}
+				continue
+			}
+			return 0, translateIOError(ctx, err)
+		}
+	}
+
+	// Go through internal mappings.
+	var prev uint32
+	_, err := mm.withInternalMappings(ctx, ar, usermem.ReadWrite, opts.IgnorePermissions, func(ims safemem.BlockSeq) (uint64, error) {
+		if ims.NumBlocks() != 1 || ims.NumBytes() != 4 {
+			// Atomicity is unachievable across mappings.
+			return 0, syserror.EFAULT
+		}
+		im := ims.Head()
+		var err error
+		prev, err = safemem.CompareAndSwapUint32(im, old, new)
+		if err != nil {
+			return 0, translateIOError(ctx, err)
+		}
+		return 4, nil
+	})
+	return prev, err
+}
+
+// handleASIOFault handles a page fault at address addr for an AddressSpaceIO
+// operation spanning ioar.
+//
+// Preconditions: mm.as != nil. ioar.Length() != 0. ioar.Contains(addr).
+func (mm *MemoryManager) handleASIOFault(ctx context.Context, addr usermem.Addr, ioar usermem.AddrRange, at usermem.AccessType) error {
+	// Try to map all remaining pages in the I/O operation. This RoundUp can't
+	// overflow because otherwise it would have been caught by checkIORange.
+	end, _ := ioar.End.RoundUp()
+	ar := usermem.AddrRange{addr.RoundDown(), end}
+
+	// Don't bother trying existingPMAsLocked; in most cases, if we did have
+	// existing pmas, we wouldn't have faulted.
+
+	// Ensure that we have usable vmas. Here and below, only return early if we
+	// can't map the first (faulting) page; failure to map later pages are
+	// silently ignored. This maximizes partial success.
+	mm.mappingMu.RLock()
+	vseg, vend, err := mm.getVMAsLocked(ctx, ar, at, false)
+	if vendaddr := vend.Start(); vendaddr < ar.End {
+		if vendaddr <= ar.Start {
+			mm.mappingMu.RUnlock()
+			return translateIOError(ctx, err)
+		}
+		ar.End = vendaddr
+	}
+
+	// Ensure that we have usable pmas.
+	mm.activeMu.Lock()
+	pseg, pend, err := mm.getPMAsLocked(ctx, vseg, ar, pmaOpts{
+		breakCOW: at.Write,
+	})
+	mm.mappingMu.RUnlock()
+	if pendaddr := pend.Start(); pendaddr < ar.End {
+		if pendaddr <= ar.Start {
+			mm.activeMu.Unlock()
+			return translateIOError(ctx, err)
+		}
+		ar.End = pendaddr
+	}
+
+	// Downgrade to a read-lock on activeMu since we don't need to mutate pmas
+	// anymore.
+	mm.activeMu.DowngradeLock()
+
+	err = mm.mapASLocked(pseg, ar, false)
+	mm.activeMu.RUnlock()
+	return translateIOError(ctx, err)
+}
+
+// withInternalMappings ensures that pmas exist for all addresses in ar,
+// support access of type (at, ignorePermissions), and have internal mappings
+// cached. It then calls f with mm.activeMu locked for reading, passing
+// internal mappings for the subrange of ar for which this property holds.
+//
+// withInternalMappings takes a function returning uint64 since many safemem
+// functions have this property, but returns an int64 since this is usually
+// more useful for usermem.IO methods.
+//
+// Preconditions: 0 < ar.Length() <= math.MaxInt64.
+func (mm *MemoryManager) withInternalMappings(ctx context.Context, ar usermem.AddrRange, at usermem.AccessType, ignorePermissions bool, f func(safemem.BlockSeq) (uint64, error)) (int64, error) {
+	po := pmaOpts{
+		breakCOW: at.Write,
+	}
+
+	// If pmas are already available, we can do IO without touching mm.vmas or
+	// mm.mappingMu.
+	mm.activeMu.RLock()
+	if pseg := mm.existingPMAsLocked(ar, at, ignorePermissions, po, true /* needInternalMappings */); pseg.Ok() {
+		n, err := f(mm.internalMappingsLocked(pseg, ar))
+		mm.activeMu.RUnlock()
+		// Do not convert errors returned by f to EFAULT.
+		return int64(n), err
+	}
+	mm.activeMu.RUnlock()
+
+	// Ensure that we have usable vmas.
+	mm.mappingMu.RLock()
+	vseg, vend, verr := mm.getVMAsLocked(ctx, ar, at, ignorePermissions)
+	if vendaddr := vend.Start(); vendaddr < ar.End {
+		if vendaddr <= ar.Start {
+			mm.mappingMu.RUnlock()
+			return 0, translateIOError(ctx, verr)
+		}
+		ar.End = vendaddr
+	}
+
+	// Ensure that we have usable pmas.
+	mm.activeMu.Lock()
+	pseg, pend, perr := mm.getPMAsLocked(ctx, vseg, ar, po)
+	mm.mappingMu.RUnlock()
+	if pendaddr := pend.Start(); pendaddr < ar.End {
+		if pendaddr <= ar.Start {
+			mm.activeMu.Unlock()
+			return 0, translateIOError(ctx, perr)
+		}
+		ar.End = pendaddr
+	}
+	imend, imerr := mm.getPMAInternalMappingsLocked(pseg, ar)
+	mm.activeMu.DowngradeLock()
+	if imendaddr := imend.Start(); imendaddr < ar.End {
+		if imendaddr <= ar.Start {
+			mm.activeMu.RUnlock()
+			return 0, translateIOError(ctx, imerr)
+		}
+		ar.End = imendaddr
+	}
+
+	// Do I/O.
+	un, err := f(mm.internalMappingsLocked(pseg, ar))
+	mm.activeMu.RUnlock()
+	n := int64(un)
+
+	// Return the first error in order of progress through ar.
+	if err != nil {
+		// Do not convert errors returned by f to EFAULT.
+		return n, err
+	}
+	if imerr != nil {
+		return n, translateIOError(ctx, imerr)
+	}
+	if perr != nil {
+		return n, translateIOError(ctx, perr)
+	}
+	return n, translateIOError(ctx, verr)
+}
+
+// withVecInternalMappings ensures that pmas exist for all addresses in ars,
+// support access of type (at, ignorePermissions), and have internal mappings
+// cached. It then calls f with mm.activeMu locked for reading, passing
+// internal mappings for the subset of ars for which this property holds.
+//
+// Preconditions: !ars.IsEmpty().
+func (mm *MemoryManager) withVecInternalMappings(ctx context.Context, ars usermem.AddrRangeSeq, at usermem.AccessType, ignorePermissions bool, f func(safemem.BlockSeq) (uint64, error)) (int64, error) {
+	// withInternalMappings is faster than withVecInternalMappings because of
+	// iterator plumbing (this isn't generally practical in the vector case due
+	// to iterator invalidation between AddrRanges). Use it if possible.
+	if ars.NumRanges() == 1 {
+		return mm.withInternalMappings(ctx, ars.Head(), at, ignorePermissions, f)
+	}
+
+	po := pmaOpts{
+		breakCOW: at.Write,
+	}
+
+	// If pmas are already available, we can do IO without touching mm.vmas or
+	// mm.mappingMu.
+	mm.activeMu.RLock()
+	if mm.existingVecPMAsLocked(ars, at, ignorePermissions, po, true /* needInternalMappings */) {
+		n, err := f(mm.vecInternalMappingsLocked(ars))
+		mm.activeMu.RUnlock()
+		// Do not convert errors returned by f to EFAULT.
+		return int64(n), err
+	}
+	mm.activeMu.RUnlock()
+
+	// Ensure that we have usable vmas.
+	mm.mappingMu.RLock()
+	vars, verr := mm.getVecVMAsLocked(ctx, ars, at, ignorePermissions)
+	if vars.NumBytes() == 0 {
+		mm.mappingMu.RUnlock()
+		return 0, translateIOError(ctx, verr)
+	}
+
+	// Ensure that we have usable pmas.
+	mm.activeMu.Lock()
+	pars, perr := mm.getVecPMAsLocked(ctx, vars, po)
+	mm.mappingMu.RUnlock()
+	if pars.NumBytes() == 0 {
+		mm.activeMu.Unlock()
+		return 0, translateIOError(ctx, perr)
+	}
+	imars, imerr := mm.getVecPMAInternalMappingsLocked(pars)
+	mm.activeMu.DowngradeLock()
+	if imars.NumBytes() == 0 {
+		mm.activeMu.RUnlock()
+		return 0, translateIOError(ctx, imerr)
+	}
+
+	// Do I/O.
+	un, err := f(mm.vecInternalMappingsLocked(imars))
+	mm.activeMu.RUnlock()
+	n := int64(un)
+
+	// Return the first error in order of progress through ars.
+	if err != nil {
+		// Do not convert errors from f to EFAULT.
+		return n, err
+	}
+	if imerr != nil {
+		return n, translateIOError(ctx, imerr)
+	}
+	if perr != nil {
+		return n, translateIOError(ctx, perr)
+	}
+	return n, translateIOError(ctx, verr)
+}
+
+// truncatedAddrRangeSeq returns a copy of ars, but with the end truncated to
+// at most address end on AddrRange arsit.Head(). It is used in vector I/O paths to
+// truncate usermem.AddrRangeSeq when errors occur.
+//
+// Preconditions: !arsit.IsEmpty(). end <= arsit.Head().End.
+func truncatedAddrRangeSeq(ars, arsit usermem.AddrRangeSeq, end usermem.Addr) usermem.AddrRangeSeq {
+	ar := arsit.Head()
+	if end <= ar.Start {
+		return ars.TakeFirst64(ars.NumBytes() - arsit.NumBytes())
+	}
+	return ars.TakeFirst64(ars.NumBytes() - arsit.NumBytes() + int64(end-ar.Start))
+}
diff --git a/pkg/sentry/mm/lifecycle.go b/pkg/sentry/mm/lifecycle.go
new file mode 100644
index 000000000..de7f29b04
--- /dev/null
+++ b/pkg/sentry/mm/lifecycle.go
@@ -0,0 +1,218 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/atomicbitops"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// NewMemoryManager returns a new MemoryManager with no mappings and 1 user.
+func NewMemoryManager(p platform.Platform) *MemoryManager {
+	return &MemoryManager{
+		p:           p,
+		haveASIO:    p.SupportsAddressSpaceIO(),
+		privateRefs: &privateRefs{},
+		users:       1,
+		auxv:        arch.Auxv{},
+		aioManager:  aioManager{contexts: make(map[uint64]*AIOContext)},
+	}
+}
+
+// SetMmapLayout initializes mm's layout from the given arch.Context.
+//
+// Preconditions: mm contains no mappings and is not used concurrently.
+func (mm *MemoryManager) SetMmapLayout(ac arch.Context, r *limits.LimitSet) (arch.MmapLayout, error) {
+	layout, err := ac.NewMmapLayout(mm.p.MinUserAddress(), mm.p.MaxUserAddress(), r)
+	if err != nil {
+		return arch.MmapLayout{}, err
+	}
+	mm.layout = layout
+	return layout, nil
+}
+
+// Fork creates a copy of mm with 1 user, as for Linux syscalls fork() or
+// clone() (without CLONE_VM).
+func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	mm2 := &MemoryManager{
+		p:                    mm.p,
+		haveASIO:             mm.haveASIO,
+		layout:               mm.layout,
+		privateRefs:          mm.privateRefs,
+		users:                1,
+		usageAS:              mm.usageAS,
+		brk:                  mm.brk,
+		captureInvalidations: true,
+		argv:                 mm.argv,
+		envv:                 mm.envv,
+		auxv:                 append(arch.Auxv(nil), mm.auxv...),
+		// IncRef'd below, once we know that there isn't an error.
+		executable: mm.executable,
+		aioManager: aioManager{contexts: make(map[uint64]*AIOContext)},
+	}
+
+	// Copy vmas.
+	dstvgap := mm2.vmas.FirstGap()
+	for srcvseg := mm.vmas.FirstSegment(); srcvseg.Ok(); srcvseg = srcvseg.NextSegment() {
+		vma := srcvseg.ValuePtr()
+		vmaAR := srcvseg.Range()
+		// Inform the Mappable, if any, of the new mapping.
+		if vma.mappable != nil {
+			if err := vma.mappable.AddMapping(ctx, mm2, vmaAR, vma.off); err != nil {
+				mm2.removeVMAsLocked(ctx, mm2.applicationAddrRange())
+				return nil, err
+			}
+		}
+		if vma.id != nil {
+			vma.id.IncRef()
+		}
+		dstvgap = mm2.vmas.Insert(dstvgap, vmaAR, *vma).NextGap()
+		// We don't need to update mm2.usageAS since we copied it from mm
+		// above.
+	}
+
+	// Copy pmas. We have to lock mm.activeMu for writing to make existing
+	// private pmas copy-on-write. We also have to lock mm2.activeMu since
+	// after copying vmas above, memmap.Mappables may call mm2.Invalidate. We
+	// only copy private pmas, since in the common case where fork(2) is
+	// immediately followed by execve(2), copying non-private pmas that can be
+	// regenerated by calling memmap.Mappable.Translate is a waste of time.
+	// (Linux does the same; compare kernel/fork.c:dup_mmap() =>
+	// mm/memory.c:copy_page_range().)
+	mm2.activeMu.Lock()
+	defer mm2.activeMu.Unlock()
+	mm.activeMu.Lock()
+	defer mm.activeMu.Unlock()
+	dstpgap := mm2.pmas.FirstGap()
+	var unmapAR usermem.AddrRange
+	for srcpseg := mm.pmas.FirstSegment(); srcpseg.Ok(); srcpseg = srcpseg.NextSegment() {
+		pma := srcpseg.ValuePtr()
+		if !pma.private {
+			continue
+		}
+		if !pma.needCOW {
+			pma.needCOW = true
+			if pma.vmaEffectivePerms.Write {
+				// We don't want to unmap the whole address space, even though
+				// doing so would reduce calls to unmapASLocked(), because mm
+				// will most likely continue to be used after the fork, so
+				// unmapping pmas unnecessarily will result in extra page
+				// faults. But we do want to merge consecutive AddrRanges
+				// across pma boundaries.
+				if unmapAR.End == srcpseg.Start() {
+					unmapAR.End = srcpseg.End()
+				} else {
+					if unmapAR.Length() != 0 {
+						mm.unmapASLocked(unmapAR)
+					}
+					unmapAR = srcpseg.Range()
+				}
+			}
+		}
+		fr := srcpseg.fileRange()
+		mm2.incPrivateRef(fr)
+		srcpseg.ValuePtr().file.IncRef(fr)
+		addrRange := srcpseg.Range()
+		mm2.addRSSLocked(addrRange)
+		dstpgap = mm2.pmas.Insert(dstpgap, addrRange, *pma).NextGap()
+	}
+	if unmapAR.Length() != 0 {
+		mm.unmapASLocked(unmapAR)
+	}
+
+	// Between when we call memmap.Mappable.AddMapping while copying vmas and
+	// when we lock mm2.activeMu to copy pmas, calls to mm2.Invalidate() are
+	// ineffective because the pmas they invalidate haven't yet been copied,
+	// possibly allowing mm2 to get invalidated translations:
+	//
+	// Invalidating Mappable            mm.Fork
+	// ---------------------            -------
+	//
+	// mm2.Invalidate()
+	//                                  mm.activeMu.Lock()
+	// mm.Invalidate() /* blocks */
+	//                                  mm2.activeMu.Lock()
+	//                                  (mm copies invalidated pma to mm2)
+	//
+	// This would technically be both safe (since we only copy private pmas,
+	// which will still hold a reference on their memory) and consistent with
+	// Linux, but we avoid it anyway by setting mm2.captureInvalidations during
+	// construction, causing calls to mm2.Invalidate() to be captured in
+	// mm2.capturedInvalidations, to be replayed after pmas are copied - i.e.
+	// here.
+	mm2.captureInvalidations = false
+	for _, invArgs := range mm2.capturedInvalidations {
+		mm2.invalidateLocked(invArgs.ar, invArgs.opts.InvalidatePrivate, true)
+	}
+	mm2.capturedInvalidations = nil
+
+	if mm2.executable != nil {
+		mm2.executable.IncRef()
+	}
+	return mm2, nil
+}
+
+// IncUsers increments mm's user count and returns true. If the user count is
+// already 0, IncUsers does nothing and returns false.
+func (mm *MemoryManager) IncUsers() bool {
+	return atomicbitops.IncUnlessZeroInt32(&mm.users)
+}
+
+// DecUsers decrements mm's user count. If the user count reaches 0, all
+// mappings in mm are unmapped.
+func (mm *MemoryManager) DecUsers(ctx context.Context) {
+	if users := atomic.AddInt32(&mm.users, -1); users > 0 {
+		return
+	} else if users < 0 {
+		panic(fmt.Sprintf("Invalid MemoryManager.users: %d", users))
+	}
+
+	mm.aioManager.destroy()
+
+	mm.metadataMu.Lock()
+	exe := mm.executable
+	mm.executable = nil
+	mm.metadataMu.Unlock()
+	if exe != nil {
+		exe.DecRef()
+	}
+
+	mm.activeMu.Lock()
+	// Sanity check.
+	if atomic.LoadInt32(&mm.active) != 0 {
+		panic("active address space lost?")
+	}
+	// Make sure the AddressSpace is returned.
+	if mm.as != nil {
+		mm.as.Release()
+		mm.as = nil
+	}
+	mm.activeMu.Unlock()
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	mm.unmapLocked(ctx, mm.applicationAddrRange())
+}
diff --git a/pkg/sentry/mm/metadata.go b/pkg/sentry/mm/metadata.go
new file mode 100644
index 000000000..32d5e2ff6
--- /dev/null
+++ b/pkg/sentry/mm/metadata.go
@@ -0,0 +1,139 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// ArgvStart returns the start of the application argument vector.
+//
+// There is no guarantee that this value is sensible w.r.t. ArgvEnd.
+func (mm *MemoryManager) ArgvStart() usermem.Addr {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	return mm.argv.Start
+}
+
+// SetArgvStart sets the start of the application argument vector.
+func (mm *MemoryManager) SetArgvStart(a usermem.Addr) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.argv.Start = a
+}
+
+// ArgvEnd returns the end of the application argument vector.
+//
+// There is no guarantee that this value is sensible w.r.t. ArgvStart.
+func (mm *MemoryManager) ArgvEnd() usermem.Addr {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	return mm.argv.End
+}
+
+// SetArgvEnd sets the end of the application argument vector.
+func (mm *MemoryManager) SetArgvEnd(a usermem.Addr) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.argv.End = a
+}
+
+// EnvvStart returns the start of the application environment vector.
+//
+// There is no guarantee that this value is sensible w.r.t. EnvvEnd.
+func (mm *MemoryManager) EnvvStart() usermem.Addr {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	return mm.envv.Start
+}
+
+// SetEnvvStart sets the start of the application environment vector.
+func (mm *MemoryManager) SetEnvvStart(a usermem.Addr) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.envv.Start = a
+}
+
+// EnvvEnd returns the end of the application environment vector.
+//
+// There is no guarantee that this value is sensible w.r.t. EnvvStart.
+func (mm *MemoryManager) EnvvEnd() usermem.Addr {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	return mm.envv.End
+}
+
+// SetEnvvEnd sets the end of the application environment vector.
+func (mm *MemoryManager) SetEnvvEnd(a usermem.Addr) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.envv.End = a
+}
+
+// Auxv returns the current map of auxiliary vectors.
+func (mm *MemoryManager) Auxv() arch.Auxv {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	return append(arch.Auxv(nil), mm.auxv...)
+}
+
+// SetAuxv sets the entire map of auxiliary vectors.
+func (mm *MemoryManager) SetAuxv(auxv arch.Auxv) {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+	mm.auxv = append(arch.Auxv(nil), auxv...)
+}
+
+// Executable returns the executable, if available.
+//
+// An additional reference will be taken in the case of a non-nil executable,
+// which must be released by the caller.
+func (mm *MemoryManager) Executable() *fs.Dirent {
+	mm.metadataMu.Lock()
+	defer mm.metadataMu.Unlock()
+
+	if mm.executable == nil {
+		return nil
+	}
+
+	mm.executable.IncRef()
+	return mm.executable
+}
+
+// SetExecutable sets the executable.
+//
+// This takes a reference on d.
+func (mm *MemoryManager) SetExecutable(d *fs.Dirent) {
+	mm.metadataMu.Lock()
+
+	// Grab a new reference.
+	d.IncRef()
+
+	// Set the executable.
+	orig := mm.executable
+	mm.executable = d
+
+	mm.metadataMu.Unlock()
+
+	// Release the old reference.
+	//
+	// Do this without holding the lock, since it may wind up doing some
+	// I/O to sync the dirent, etc.
+	if orig != nil {
+		orig.DecRef()
+	}
+}
diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go
new file mode 100644
index 000000000..ce8097b7f
--- /dev/null
+++ b/pkg/sentry/mm/mm.go
@@ -0,0 +1,417 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package mm provides a memory management subsystem. See README.md for a
+// detailed overview.
+//
+// Lock order:
+//
+// fs locks, except for memmap.Mappable locks
+//   mm.MemoryManager.metadataMu
+//     mm.MemoryManager.mappingMu
+//       Locks taken by memmap.Mappable methods other than Translate
+//         mm.MemoryManager.activeMu
+//           Locks taken by memmap.Mappable.Translate
+//             mm.privateRefs.mu
+//               platform.File locks
+//         mm.aioManager.mu
+//           mm.AIOContext.mu
+//
+// Only mm.MemoryManager.Fork is permitted to lock mm.MemoryManager.activeMu in
+// multiple mm.MemoryManagers, as it does so in a well-defined order (forked
+// child first).
+package mm
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	ssync "gvisor.googlesource.com/gvisor/pkg/sync"
+)
+
+// MemoryManager implements a virtual address space.
+type MemoryManager struct {
+	// p is the platform.
+	//
+	// p is immutable.
+	p platform.Platform
+
+	// haveASIO is the cached result of p.SupportsAddressSpaceIO(). Aside from
+	// eliminating an indirect call in the hot I/O path, this makes
+	// MemoryManager.asioEnabled() a leaf function, allowing it to be inlined.
+	//
+	// haveASIO is immutable.
+	haveASIO bool `state:"nosave"`
+
+	// layout is the memory layout.
+	//
+	// layout is set by the binary loader before the MemoryManager can be used.
+	layout arch.MmapLayout
+
+	// privateRefs stores reference counts for private memory (memory whose
+	// ownership is shared by one or more pmas instead of being owned by a
+	// memmap.Mappable).
+	//
+	// NOTE: This should be replaced using refcounts on
+	// platform.File.
+	//
+	// privateRefs is immutable.
+	privateRefs *privateRefs
+
+	// users is the number of dependences on the mappings in the MemoryManager.
+	// When the number of references in users reaches zero, all mappings are
+	// unmapped.
+	//
+	// users is accessed using atomic memory operations.
+	users int32
+
+	// mappingMu is analogous to Linux's struct mm_struct::mmap_sem.
+	mappingMu ssync.DowngradableRWMutex `state:"nosave"`
+
+	// vmas stores virtual memory areas. Since vmas are stored by value,
+	// clients should usually use vmaIterator.ValuePtr() instead of
+	// vmaIterator.Value() to get a pointer to the vma rather than a copy.
+	//
+	// Invariants: vmas are always page-aligned.
+	//
+	// vmas is protected by mappingMu.
+	vmas vmaSet
+
+	// usageAS is vmas.Span(), cached to accelerate RLIMIT_AS checks.
+	//
+	// usageAS is protected by mappingMu.
+	usageAS uint64
+
+	// brk is the mm's brk, which is manipulated using the brk(2) system call.
+	// The brk is initially set up by the loader which maps an executable
+	// binary into the mm.
+	//
+	// brk is protected by mappingMu.
+	brk usermem.AddrRange
+
+	// activeMu is loosely analogous to Linux's struct
+	// mm_struct::page_table_lock.
+	activeMu ssync.DowngradableRWMutex `state:"nosave"`
+
+	// pmas stores platform mapping areas used to implement vmas. Since pmas
+	// are stored by value, clients should usually use pmaIterator.ValuePtr()
+	// instead of pmaIterator.Value() to get a pointer to the pma rather than
+	// a copy.
+	//
+	// Inserting or removing segments from pmas should happen along with a
+	// call to mm.insertRSS or mm.removeRSS.
+	//
+	// Invariants: pmas are always page-aligned. If a pma exists for a given
+	// address, a vma must also exist for that address.
+	//
+	// pmas is protected by activeMu.
+	pmas pmaSet
+
+	// curRSS is pmas.Span(), cached to accelerate updates to maxRSS. It is
+	// reported as the MemoryManager's RSS.
+	//
+	// maxRSS should be modified only via insertRSS and removeRSS, not
+	// directly.
+	//
+	// maxRSS is protected by activeMu.
+	curRSS uint64
+
+	// maxRSS is the maximum resident set size in bytes of a MemoryManager.
+	// It is tracked as the application adds and removes mappings to pmas.
+	//
+	// maxRSS should be modified only via insertRSS, not directly.
+	//
+	// maxRSS is protected by activeMu.
+	maxRSS uint64
+
+	// as is the platform.AddressSpace that pmas are mapped into. active is the
+	// number of contexts that require as to be non-nil; if active == 0, as may
+	// be nil.
+	//
+	// as is protected by activeMu. active is manipulated with atomic memory
+	// operations; transitions to and from zero are additionally protected by
+	// activeMu. (This is because such transitions may need to be atomic with
+	// changes to as.)
+	as     platform.AddressSpace `state:"nosave"`
+	active int32                 `state:"zerovalue"`
+
+	// unmapAllOnActivate indicates that the next Activate call should activate
+	// an empty AddressSpace.
+	//
+	// This is used to ensure that an AddressSpace cached in
+	// NewAddressSpace is not used after some change in the MemoryManager
+	// or VMAs has made that AddressSpace stale.
+	//
+	// unmapAllOnActivate is protected by activeMu. It must only be set when
+	// there is no active or cached AddressSpace. If as != nil, then
+	// invalidations should be propagated immediately.
+	unmapAllOnActivate bool `state:"nosave"`
+
+	// If captureInvalidations is true, calls to MM.Invalidate() are recorded
+	// in capturedInvalidations rather than being applied immediately to pmas.
+	// This is to avoid a race condition in MM.Fork(); see that function for
+	// details.
+	//
+	// Both captureInvalidations and capturedInvalidations are protected by
+	// activeMu. Neither need to be saved since captureInvalidations is only
+	// enabled during MM.Fork(), during which saving can't occur.
+	captureInvalidations  bool             `state:"zerovalue"`
+	capturedInvalidations []invalidateArgs `state:"nosave"`
+
+	metadataMu sync.Mutex `state:"nosave"`
+
+	// argv is the application argv. This is set up by the loader and may be
+	// modified by prctl(PR_SET_MM_ARG_START/PR_SET_MM_ARG_END). No
+	// requirements apply to argv; we do not require that argv.WellFormed().
+	//
+	// argv is protected by metadataMu.
+	argv usermem.AddrRange
+
+	// envv is the application envv. This is set up by the loader and may be
+	// modified by prctl(PR_SET_MM_ENV_START/PR_SET_MM_ENV_END). No
+	// requirements apply to envv; we do not require that envv.WellFormed().
+	//
+	// envv is protected by metadataMu.
+	envv usermem.AddrRange
+
+	// auxv is the ELF's auxiliary vector.
+	//
+	// auxv is protected by metadataMu.
+	auxv arch.Auxv
+
+	// executable is the executable for this MemoryManager. If executable
+	// is not nil, it holds a reference on the Dirent.
+	//
+	// executable is protected by metadataMu.
+	executable *fs.Dirent
+
+	// aioManager keeps track of AIOContexts used for async IOs. AIOManager
+	// must be cloned when CLONE_VM is used.
+	aioManager aioManager
+}
+
+// vma represents a virtual memory area.
+type vma struct {
+	// mappable is the virtual memory object mapped by this vma. If mappable is
+	// nil, the vma represents a private anonymous mapping.
+	mappable memmap.Mappable
+
+	// off is the offset into mappable at which this vma begins. If mappable is
+	// nil, off is meaningless.
+	off uint64
+
+	// To speedup VMA save/restore, we group and save the following booleans
+	// as a single integer.
+
+	// realPerms are the memory permissions on this vma, as defined by the
+	// application.
+	realPerms usermem.AccessType `state:".(int)"`
+
+	// effectivePerms are the memory permissions on this vma which are
+	// actually used to control access.
+	//
+	// Invariant: effectivePerms == realPerms.Effective().
+	effectivePerms usermem.AccessType `state:"manual"`
+
+	// maxPerms limits the set of permissions that may ever apply to this
+	// memory, as well as accesses for which usermem.IOOpts.IgnorePermissions
+	// is true (e.g. ptrace(PTRACE_POKEDATA)).
+	//
+	// Invariant: maxPerms == maxPerms.Effective().
+	maxPerms usermem.AccessType `state:"manual"`
+
+	// private is true if this is a MAP_PRIVATE mapping, such that writes to
+	// the mapping are propagated to a copy.
+	private bool `state:"manual"`
+
+	// growsDown is true if the mapping may be automatically extended downward
+	// under certain conditions. If growsDown is true, mappable must be nil.
+	//
+	// There is currently no corresponding growsUp flag; in Linux, the only
+	// architectures that can have VM_GROWSUP mappings are ia64, parisc, and
+	// metag, none of which we currently support.
+	growsDown bool `state:"manual"`
+
+	// If id is not nil, it controls the lifecycle of mappable and provides vma
+	// metadata shown in /proc/[pid]/maps, and the vma holds a reference.
+	id memmap.MappingIdentity
+
+	// If hint is non-empty, it is a description of the vma printed in
+	// /proc/[pid]/maps. hint takes priority over id.MappedName().
+	hint string
+}
+
+const (
+	vmaRealPermsRead = 1 << iota
+	vmaRealPermsWrite
+	vmaRealPermsExecute
+	vmaEffectivePermsRead
+	vmaEffectivePermsWrite
+	vmaEffectivePermsExecute
+	vmaMaxPermsRead
+	vmaMaxPermsWrite
+	vmaMaxPermsExecute
+	vmaPrivate
+	vmaGrowsDown
+)
+
+func (v *vma) saveRealPerms() int {
+	var b int
+	if v.realPerms.Read {
+		b |= vmaRealPermsRead
+	}
+	if v.realPerms.Write {
+		b |= vmaRealPermsWrite
+	}
+	if v.realPerms.Execute {
+		b |= vmaRealPermsExecute
+	}
+	if v.effectivePerms.Read {
+		b |= vmaEffectivePermsRead
+	}
+	if v.effectivePerms.Write {
+		b |= vmaEffectivePermsWrite
+	}
+	if v.effectivePerms.Execute {
+		b |= vmaEffectivePermsExecute
+	}
+	if v.maxPerms.Read {
+		b |= vmaMaxPermsRead
+	}
+	if v.maxPerms.Write {
+		b |= vmaMaxPermsWrite
+	}
+	if v.maxPerms.Execute {
+		b |= vmaMaxPermsExecute
+	}
+	if v.private {
+		b |= vmaPrivate
+	}
+	if v.growsDown {
+		b |= vmaGrowsDown
+	}
+	return b
+}
+
+func (v *vma) loadRealPerms(b int) {
+	if b&vmaRealPermsRead > 0 {
+		v.realPerms.Read = true
+	}
+	if b&vmaRealPermsWrite > 0 {
+		v.realPerms.Write = true
+	}
+	if b&vmaRealPermsExecute > 0 {
+		v.realPerms.Execute = true
+	}
+	if b&vmaEffectivePermsRead > 0 {
+		v.effectivePerms.Read = true
+	}
+	if b&vmaEffectivePermsWrite > 0 {
+		v.effectivePerms.Write = true
+	}
+	if b&vmaEffectivePermsExecute > 0 {
+		v.effectivePerms.Execute = true
+	}
+	if b&vmaMaxPermsRead > 0 {
+		v.maxPerms.Read = true
+	}
+	if b&vmaMaxPermsWrite > 0 {
+		v.maxPerms.Write = true
+	}
+	if b&vmaMaxPermsExecute > 0 {
+		v.maxPerms.Execute = true
+	}
+	if b&vmaPrivate > 0 {
+		v.private = true
+	}
+	if b&vmaGrowsDown > 0 {
+		v.growsDown = true
+	}
+}
+
+// pma represents a platform mapping area.
+type pma struct {
+	// file is the file mapped by this pma. Only pmas for which file ==
+	// platform.Platform.Memory() may be saved. pmas hold a reference to the
+	// corresponding file range while they exist.
+	file platform.File `state:"nosave"`
+
+	// off is the offset into file at which this pma begins.
+	off uint64
+
+	// vmaEffectivePerms and vmaMaxPerms are duplicated from the
+	// corresponding vma so that the IO implementation can avoid iterating
+	// mm.vmas when pmas already exist.
+	vmaEffectivePerms usermem.AccessType
+	vmaMaxPerms       usermem.AccessType
+
+	// needCOW is true if writes to the mapping must be propagated to a copy.
+	needCOW bool
+
+	// private is true if this pma represents private memory.
+	//
+	// If private is true, file must be platform.Platform.Memory(), the pma
+	// holds a reference on the mapped memory that is tracked in privateRefs,
+	// and calls to Invalidate for which
+	// memmap.InvalidateOpts.InvalidatePrivate is false should ignore the pma.
+	//
+	// If private is false, this pma caches a translation from the
+	// corresponding vma's memmap.Mappable.Translate.
+	private bool
+
+	// If internalMappings is not empty, it is the cached return value of
+	// file.MapInternal for the platform.FileRange mapped by this pma.
+	internalMappings safemem.BlockSeq `state:"nosave"`
+}
+
+type privateRefs struct {
+	mu sync.Mutex `state:"nosave"`
+
+	// refs maps offsets into Platform.Memory() to the number of pmas (or,
+	// equivalently, MemoryManagers) that share ownership of the memory at that
+	// offset.
+	refs fileRefcountSet
+}
+
+type invalidateArgs struct {
+	ar   usermem.AddrRange
+	opts memmap.InvalidateOpts
+}
+
+// fileRefcountSetFunctions implements segment.Functions for fileRefcountSet.
+type fileRefcountSetFunctions struct{}
+
+func (fileRefcountSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+func (fileRefcountSetFunctions) MaxKey() uint64 {
+	return ^uint64(0)
+}
+
+func (fileRefcountSetFunctions) ClearValue(_ *int32) {
+}
+
+func (fileRefcountSetFunctions) Merge(_ platform.FileRange, rc1 int32, _ platform.FileRange, rc2 int32) (int32, bool) {
+	return rc1, rc1 == rc2
+}
+
+func (fileRefcountSetFunctions) Split(_ platform.FileRange, rc int32, _ uint64) (int32, int32) {
+	return rc, rc
+}
diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go
new file mode 100644
index 000000000..b47aa7263
--- /dev/null
+++ b/pkg/sentry/mm/mm_test.go
@@ -0,0 +1,174 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context/contexttest"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func testMemoryManager(ctx context.Context) *MemoryManager {
+	p := platform.FromContext(ctx)
+	mm := NewMemoryManager(p)
+	mm.layout = arch.MmapLayout{
+		MinAddr:      p.MinUserAddress(),
+		MaxAddr:      p.MaxUserAddress(),
+		BottomUpBase: p.MinUserAddress(),
+		TopDownBase:  p.MaxUserAddress(),
+	}
+	return mm
+}
+
+func (mm *MemoryManager) realUsageAS() uint64 {
+	return uint64(mm.vmas.Span())
+}
+
+func TestUsageASUpdates(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm := testMemoryManager(ctx)
+	defer mm.DecUsers(ctx)
+
+	addr, err := mm.MMap(ctx, memmap.MMapOpts{
+		Length: 2 * usermem.PageSize,
+	})
+	if err != nil {
+		t.Fatalf("MMap got err %v want nil", err)
+	}
+	realUsage := mm.realUsageAS()
+	if mm.usageAS != realUsage {
+		t.Fatalf("usageAS believes %v bytes are mapped; %v bytes are actually mapped", mm.usageAS, realUsage)
+	}
+
+	mm.MUnmap(ctx, addr, usermem.PageSize)
+	realUsage = mm.realUsageAS()
+	if mm.usageAS != realUsage {
+		t.Fatalf("usageAS believes %v bytes are mapped; %v bytes are actually mapped", mm.usageAS, realUsage)
+	}
+}
+
+func TestBrkDataLimitUpdates(t *testing.T) {
+	limitSet := limits.NewLimitSet()
+	limitSet.Set(limits.Data, limits.Limit{}) // zero RLIMIT_DATA
+
+	ctx := contexttest.WithLimitSet(contexttest.Context(t), limitSet)
+	mm := testMemoryManager(ctx)
+	defer mm.DecUsers(ctx)
+
+	// Try to extend the brk by one page and expect doing so to fail.
+	oldBrk, _ := mm.Brk(ctx, 0)
+	if newBrk, _ := mm.Brk(ctx, oldBrk+usermem.PageSize); newBrk != oldBrk {
+		t.Errorf("brk() increased data segment above RLIMIT_DATA (old brk = %#x, new brk = %#x", oldBrk, newBrk)
+	}
+}
+
+// TestIOAfterUnmap ensures that IO fails after unmap.
+func TestIOAfterUnmap(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm := testMemoryManager(ctx)
+	defer mm.DecUsers(ctx)
+
+	addr, err := mm.MMap(ctx, memmap.MMapOpts{
+		Length:   usermem.PageSize,
+		Private:  true,
+		Perms:    usermem.Read,
+		MaxPerms: usermem.AnyAccess,
+	})
+	if err != nil {
+		t.Fatalf("MMap got err %v want nil", err)
+	}
+
+	// IO works before munmap.
+	b := make([]byte, 1)
+	n, err := mm.CopyIn(ctx, addr, b, usermem.IOOpts{})
+	if err != nil {
+		t.Errorf("CopyIn got err %v want nil", err)
+	}
+	if n != 1 {
+		t.Errorf("CopyIn got %d want 1", n)
+	}
+
+	err = mm.MUnmap(ctx, addr, usermem.PageSize)
+	if err != nil {
+		t.Fatalf("MUnmap got err %v want nil", err)
+	}
+
+	n, err = mm.CopyIn(ctx, addr, b, usermem.IOOpts{})
+	if err != syserror.EFAULT {
+		t.Errorf("CopyIn got err %v want EFAULT", err)
+	}
+	if n != 0 {
+		t.Errorf("CopyIn got %d want 0", n)
+	}
+}
+
+// TestIOAfterMProtect tests IO interaction with mprotect permissions.
+func TestIOAfterMProtect(t *testing.T) {
+	ctx := contexttest.Context(t)
+	mm := testMemoryManager(ctx)
+	defer mm.DecUsers(ctx)
+
+	addr, err := mm.MMap(ctx, memmap.MMapOpts{
+		Length:   usermem.PageSize,
+		Private:  true,
+		Perms:    usermem.ReadWrite,
+		MaxPerms: usermem.AnyAccess,
+	})
+	if err != nil {
+		t.Fatalf("MMap got err %v want nil", err)
+	}
+
+	// Writing works before mprotect.
+	b := make([]byte, 1)
+	n, err := mm.CopyOut(ctx, addr, b, usermem.IOOpts{})
+	if err != nil {
+		t.Errorf("CopyOut got err %v want nil", err)
+	}
+	if n != 1 {
+		t.Errorf("CopyOut got %d want 1", n)
+	}
+
+	err = mm.MProtect(addr, usermem.PageSize, usermem.Read, false)
+	if err != nil {
+		t.Errorf("MProtect got err %v want nil", err)
+	}
+
+	// Without IgnorePermissions, CopyOut should no longer succeed.
+	n, err = mm.CopyOut(ctx, addr, b, usermem.IOOpts{})
+	if err != syserror.EFAULT {
+		t.Errorf("CopyOut got err %v want EFAULT", err)
+	}
+	if n != 0 {
+		t.Errorf("CopyOut got %d want 0", n)
+	}
+
+	// With IgnorePermissions, CopyOut should succeed despite mprotect.
+	n, err = mm.CopyOut(ctx, addr, b, usermem.IOOpts{
+		IgnorePermissions: true,
+	})
+	if err != nil {
+		t.Errorf("CopyOut got err %v want nil", err)
+	}
+	if n != 1 {
+		t.Errorf("CopyOut got %d want 1", n)
+	}
+}
diff --git a/pkg/sentry/mm/pma.go b/pkg/sentry/mm/pma.go
new file mode 100644
index 000000000..35e873762
--- /dev/null
+++ b/pkg/sentry/mm/pma.go
@@ -0,0 +1,928 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/safecopy"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+type pmaOpts struct {
+	// If breakCOW is true, pmas must not be copy-on-write.
+	breakCOW bool
+}
+
+// existingPMAsLocked checks that pmas exist for all addresses in ar, and
+// support access of type (at, ignorePermissions). If so, it returns an
+// iterator to the pma containing ar.Start. Otherwise it returns a terminal
+// iterator.
+//
+// Preconditions: mm.activeMu must be locked. ar.Length() != 0.
+func (mm *MemoryManager) existingPMAsLocked(ar usermem.AddrRange, at usermem.AccessType, ignorePermissions bool, opts pmaOpts, needInternalMappings bool) pmaIterator {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	first := mm.pmas.FindSegment(ar.Start)
+	pseg := first
+	for pseg.Ok() {
+		pma := pseg.ValuePtr()
+		perms := pma.vmaEffectivePerms
+		if ignorePermissions {
+			perms = pma.vmaMaxPerms
+		}
+		if !perms.SupersetOf(at) {
+			// These are the vma's permissions, so the caller will get an error
+			// when they try to get new pmas.
+			return pmaIterator{}
+		}
+		if opts.breakCOW && pma.needCOW {
+			return pmaIterator{}
+		}
+		if needInternalMappings && pma.internalMappings.IsEmpty() {
+			return pmaIterator{}
+		}
+
+		if ar.End <= pseg.End() {
+			return first
+		}
+		pseg, _ = pseg.NextNonEmpty()
+	}
+
+	// Ran out of pmas before reaching ar.End.
+	return pmaIterator{}
+}
+
+// existingVecPMAsLocked returns true if pmas exist for all addresses in ars,
+// and support access of type (at, ignorePermissions).
+//
+// Preconditions: mm.activeMu must be locked.
+func (mm *MemoryManager) existingVecPMAsLocked(ars usermem.AddrRangeSeq, at usermem.AccessType, ignorePermissions bool, opts pmaOpts, needInternalMappings bool) bool {
+	for ; !ars.IsEmpty(); ars = ars.Tail() {
+		if ar := ars.Head(); ar.Length() != 0 && !mm.existingPMAsLocked(ar, at, ignorePermissions, opts, needInternalMappings).Ok() {
+			return false
+		}
+	}
+	return true
+}
+
+// getPMAsLocked ensures that pmas exist for all addresses in ar, subject to
+// opts. It returns:
+//
+// - An iterator to the pma containing ar.Start. If no pma contains ar.Start,
+// the iterator is unspecified.
+//
+// - An iterator to the gap after the last pma containing an address in ar. If
+// pmas exist for no addresses in ar, the iterator is to a gap that begins
+// before ar.Start.
+//
+// - An error that is non-nil if pmas exist for only a subset of ar.
+//
+// Preconditions: mm.mappingMu must be locked. mm.activeMu must be locked for
+// writing. ar.Length() != 0. vseg.Range().Contains(ar.Start). vmas must exist
+// for all addresses in ar.
+func (mm *MemoryManager) getPMAsLocked(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, opts pmaOpts) (pmaIterator, pmaGapIterator, error) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !vseg.Ok() {
+			panic("terminal vma iterator")
+		}
+		if !vseg.Range().Contains(ar.Start) {
+			panic(fmt.Sprintf("initial vma %v does not cover start of ar %v", vseg.Range(), ar))
+		}
+	}
+
+	// Page-align ar so that all AddrRanges are aligned.
+	end, ok := ar.End.RoundUp()
+	var alignerr error
+	if !ok {
+		end = ar.End.RoundDown()
+		alignerr = syserror.EFAULT
+	}
+	ar = usermem.AddrRange{ar.Start.RoundDown(), end}
+
+	pstart, pend, perr := mm.ensurePMAsLocked(ctx, vseg, ar)
+	if pend.Start() <= ar.Start {
+		return pmaIterator{}, pend, perr
+	}
+	// ensurePMAsLocked may not have pstart due to iterator invalidation. We
+	// need it, either to return it immediately or to pass to
+	// breakCopyOnWriteLocked.
+	if !pstart.Ok() {
+		pstart = mm.findOrSeekPrevUpperBoundPMA(ar.Start, pend)
+	}
+
+	var cowerr error
+	if opts.breakCOW {
+		var invalidated bool
+		pend, invalidated, cowerr = mm.breakCopyOnWriteLocked(pstart, ar)
+		if pend.Start() <= ar.Start {
+			return pmaIterator{}, pend, cowerr
+		}
+		if invalidated {
+			pstart = mm.findOrSeekPrevUpperBoundPMA(ar.Start, pend)
+		}
+	}
+
+	if cowerr != nil {
+		return pstart, pend, cowerr
+	}
+	if perr != nil {
+		return pstart, pend, perr
+	}
+	return pstart, pend, alignerr
+}
+
+// getVecPMAsLocked ensures that pmas exist for all addresses in ars. It
+// returns the subset of ars for which pmas exist. If this is not equal to ars,
+// it returns a non-nil error explaining why.
+//
+// Preconditions: mm.mappingMu must be locked. mm.activeMu must be locked for
+// writing. vmas must exist for all addresses in ars.
+func (mm *MemoryManager) getVecPMAsLocked(ctx context.Context, ars usermem.AddrRangeSeq, opts pmaOpts) (usermem.AddrRangeSeq, error) {
+	for arsit := ars; !arsit.IsEmpty(); arsit = arsit.Tail() {
+		ar := arsit.Head()
+		if ar.Length() == 0 {
+			continue
+		}
+
+		// Page-align ar so that all AddrRanges are aligned.
+		end, ok := ar.End.RoundUp()
+		var alignerr error
+		if !ok {
+			end = ar.End.RoundDown()
+			alignerr = syserror.EFAULT
+		}
+		ar = usermem.AddrRange{ar.Start.RoundDown(), end}
+
+		pstart, pend, perr := mm.ensurePMAsLocked(ctx, mm.vmas.FindSegment(ar.Start), ar)
+		if pend.Start() <= ar.Start {
+			return truncatedAddrRangeSeq(ars, arsit, pend.Start()), perr
+		}
+
+		var cowerr error
+		if opts.breakCOW {
+			if !pstart.Ok() {
+				pstart = mm.findOrSeekPrevUpperBoundPMA(ar.Start, pend)
+			}
+			pend, _, cowerr = mm.breakCopyOnWriteLocked(pstart, ar)
+		}
+
+		if cowerr != nil {
+			return truncatedAddrRangeSeq(ars, arsit, pend.Start()), cowerr
+		}
+		if perr != nil {
+			return truncatedAddrRangeSeq(ars, arsit, pend.Start()), perr
+		}
+		if alignerr != nil {
+			return truncatedAddrRangeSeq(ars, arsit, pend.Start()), alignerr
+		}
+	}
+
+	return ars, nil
+}
+
+// ensurePMAsLocked ensures that pmas exist for all addresses in ar. It returns:
+//
+// - An iterator to the pma containing ar.Start, on a best-effort basis (that
+// is, the returned iterator may be terminal, even if such a pma exists).
+// Returning this iterator on a best-effort basis allows callers that require
+// it to use it when it's cheaply available, while also avoiding the overhead
+// of retrieving it when it's not.
+//
+// - An iterator to the gap after the last pma containing an address in ar. If
+// pmas exist for no addresses in ar, the iterator is to a gap that begins
+// before ar.Start.
+//
+// - An error that is non-nil if pmas exist for only a subset of ar.
+//
+// Preconditions: mm.mappingMu must be locked. mm.activeMu must be locked for
+// writing. ar.Length() != 0. ar must be page-aligned.
+// vseg.Range().Contains(ar.Start). vmas must exist for all addresses in ar.
+func (mm *MemoryManager) ensurePMAsLocked(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange) (pmaIterator, pmaGapIterator, error) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !vseg.Range().Contains(ar.Start) {
+			panic(fmt.Sprintf("initial vma %v does not cover start of ar %v", vseg.Range(), ar))
+		}
+	}
+
+	pstart, pgap := mm.pmas.Find(ar.Start)
+	if pstart.Ok() {
+		pgap = pstart.NextGap()
+	}
+	for pgap.Start() < ar.End {
+		if pgap.Range().Length() == 0 {
+			pgap = pgap.NextGap()
+			continue
+		}
+		// A single pgap might be spanned by multiple vmas. Insert pmas to
+		// cover the first (vma, pgap) pair.
+		pgapAR := pgap.Range().Intersect(ar)
+		vseg = vseg.seekNextLowerBound(pgapAR.Start)
+		if checkInvariants {
+			if !vseg.Ok() {
+				panic(fmt.Sprintf("no vma after %#x", pgapAR.Start))
+			}
+			if pgapAR.Start < vseg.Start() {
+				panic(fmt.Sprintf("no vma in [%#x, %#x)", pgapAR.Start, vseg.Start()))
+			}
+		}
+		var err error
+		pgap, err = mm.insertPMAsLocked(ctx, vseg, pgap, ar)
+		// insertPMAsLocked most likely invalidated iterators, so pstart is now
+		// unknown.
+		pstart = pmaIterator{}
+		if err != nil {
+			return pstart, pgap, err
+		}
+	}
+	return pstart, pgap, nil
+}
+
+const (
+	// When memory is allocated for a private pma, align the allocated address
+	// range to a privateAllocUnit boundary when possible. Larger values of
+	// privateAllocUnit may reduce page faults by allowing fewer, larger pmas
+	// to be mapped, but may result in larger amounts of wasted memory in the
+	// presence of fragmentation. privateAllocUnit must be a power-of-2
+	// multiple of usermem.PageSize.
+	privateAllocUnit = usermem.HugePageSize
+
+	privateAllocMask = privateAllocUnit - 1
+)
+
+func privateAligned(ar usermem.AddrRange) usermem.AddrRange {
+	aligned := usermem.AddrRange{ar.Start &^ privateAllocMask, ar.End}
+	if end := (ar.End + privateAllocMask) &^ privateAllocMask; end >= ar.End {
+		aligned.End = end
+	}
+	if checkInvariants {
+		if !aligned.IsSupersetOf(ar) {
+			panic(fmt.Sprintf("aligned AddrRange %#v is not a superset of ar %#v", aligned, ar))
+		}
+	}
+	return aligned
+}
+
+// insertPMAsLocked inserts pmas into pgap corresponding to the vma iterated by
+// vseg, spanning at least ar. It returns:
+//
+// - An iterator to the gap after the last pma containing an address in ar. If
+// pmas exist for no addresses in ar, the iterator is to a gap that begins
+// before ar.Start.
+//
+// - An error that is non-nil if pmas exist for only a subset of ar.
+//
+// Preconditions: mm.mappingMu must be locked. mm.activeMu must be locked for
+// writing. vseg.Range().Intersect(pgap.Range()).Intersect(ar).Length() != 0.
+// ar must be page-aligned.
+func (mm *MemoryManager) insertPMAsLocked(ctx context.Context, vseg vmaIterator, pgap pmaGapIterator, ar usermem.AddrRange) (pmaGapIterator, error) {
+	optAR := vseg.Range().Intersect(pgap.Range())
+	if checkInvariants {
+		if optAR.Length() <= 0 {
+			panic(fmt.Sprintf("vseg %v and pgap %v do not overlap", vseg, pgap))
+		}
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar %v", ar))
+		}
+	}
+	vma := vseg.ValuePtr()
+
+	// Private anonymous mappings get pmas by allocating.
+	if vma.mappable == nil {
+		// Limit the range we allocate to ar, aligned to privateAllocUnit.
+		maskAR := privateAligned(ar)
+		allocAR := optAR.Intersect(maskAR)
+		mem := mm.p.Memory()
+		fr, err := mem.Allocate(uint64(allocAR.Length()), usage.Anonymous)
+		if err != nil {
+			return pgap, err
+		}
+		mm.incPrivateRef(fr)
+
+		if checkInvariants {
+			if !fr.WellFormed() || fr.Length() != uint64(allocAR.Length()) {
+				panic(fmt.Sprintf("Allocate(%v) returned invalid FileRange %v", allocAR.Length(), fr))
+			}
+		}
+
+		mm.addRSSLocked(allocAR)
+		mem.IncRef(fr)
+
+		return mm.pmas.Insert(pgap, allocAR, pma{
+			file:              mem,
+			off:               fr.Start,
+			vmaEffectivePerms: vma.effectivePerms,
+			vmaMaxPerms:       vma.maxPerms,
+			private:           true,
+			// Since we just allocated this memory and have the only reference,
+			// the new pma does not need copy-on-write.
+		}).NextGap(), nil
+	}
+
+	// Other mappings get pmas by translating. Limit the required range
+	// to ar.
+	optMR := vseg.mappableRangeOf(optAR)
+	reqAR := optAR.Intersect(ar)
+	reqMR := vseg.mappableRangeOf(reqAR)
+	perms := vma.maxPerms
+	if vma.private {
+		perms.Write = false
+	}
+	ts, err := vma.mappable.Translate(ctx, reqMR, optMR, perms)
+	if checkInvariants {
+		if err := memmap.CheckTranslateResult(reqMR, optMR, ts, err); err != nil {
+			panic(fmt.Sprintf("Mappable(%T).Translate(%v, %v): %v", vma.mappable, reqMR, optMR, err))
+		}
+	}
+
+	// Install a pma for each Translation.
+	for _, t := range ts {
+		// This is valid because memmap.Mappable.Translate is required to
+		// return Translations in increasing Translation.Source order.
+		addrRange := vseg.addrRangeOf(t.Source)
+		mm.addRSSLocked(addrRange)
+		pseg := mm.pmas.Insert(pgap, addrRange, pma{
+			file:              t.File,
+			off:               t.Offset,
+			vmaEffectivePerms: vma.effectivePerms,
+			vmaMaxPerms:       vma.maxPerms,
+			needCOW:           vma.private,
+		})
+		// The new pseg may have been merged with existing segments, only take a
+		// ref on the inserted range.
+		t.File.IncRef(pseg.fileRangeOf(addrRange))
+		pgap = pseg.NextGap()
+	}
+
+	// Even if Translate returned an error, if we got to ar.End,
+	// insertPMAsLocked succeeded.
+	if ar.End <= pgap.Start() {
+		return pgap, nil
+	}
+	return pgap, err
+}
+
+// breakCopyOnWriteLocked ensures that pmas in ar are not copy-on-write. It
+// returns:
+//
+// - An iterator to the gap after the last non-COW pma containing an address in
+// ar. If non-COW pmas exist for no addresses in ar, the iterator is to a gap
+// that begins before ar.Start.
+//
+// - A boolean that is true if iterators into mm.pmas may have been
+// invalidated.
+//
+// - An error that is non-nil if non-COW pmas exist for only a subset of ar.
+//
+// Preconditions: mm.activeMu must be locked for writing. ar.Length() != 0. ar
+// must be page-aligned. pseg.Range().Contains(ar.Start). pmas must exist for
+// all addresses in ar.
+func (mm *MemoryManager) breakCopyOnWriteLocked(pseg pmaIterator, ar usermem.AddrRange) (pmaGapIterator, bool, error) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !pseg.Range().Contains(ar.Start) {
+			panic(fmt.Sprintf("initial pma %v does not cover start of ar %v", pseg.Range(), ar))
+		}
+	}
+
+	// Limit the range we copy to ar, aligned to privateAllocUnit.
+	maskAR := privateAligned(ar)
+	var invalidatedIterators, didUnmapAS bool
+	mem := mm.p.Memory()
+	for {
+		if mm.isPMACopyOnWriteLocked(pseg) {
+			// Determine the range to copy.
+			copyAR := pseg.Range().Intersect(maskAR)
+
+			// Get internal mappings from the pma to copy from.
+			if err := pseg.getInternalMappingsLocked(); err != nil {
+				return pseg.PrevGap(), invalidatedIterators, err
+			}
+
+			// Copy contents.
+			fr, err := platform.AllocateAndFill(mem, uint64(copyAR.Length()), usage.Anonymous, &safemem.BlockSeqReader{mm.internalMappingsLocked(pseg, copyAR)})
+			if _, ok := err.(safecopy.BusError); ok {
+				// If we got SIGBUS during the copy, deliver SIGBUS to
+				// userspace (instead of SIGSEGV) if we're breaking
+				// copy-on-write due to application page fault.
+				err = &memmap.BusError{err}
+			}
+			if fr.Length() == 0 {
+				return pseg.PrevGap(), invalidatedIterators, err
+			}
+			mm.incPrivateRef(fr)
+			mem.IncRef(fr)
+
+			// Unmap all of maskAR, not just copyAR, to minimize host syscalls.
+			// AddressSpace mappings must be removed before mm.decPrivateRef().
+			if !didUnmapAS {
+				mm.unmapASLocked(maskAR)
+				didUnmapAS = true
+			}
+
+			// Replace the pma with a copy in the part of the address range
+			// where copying was successful.
+			copyAR.End = copyAR.Start + usermem.Addr(fr.Length())
+			if copyAR != pseg.Range() {
+				pseg = mm.pmas.Isolate(pseg, copyAR)
+				invalidatedIterators = true
+			}
+			pma := pseg.ValuePtr()
+			if pma.private {
+				mm.decPrivateRef(pseg.fileRange())
+			}
+			pma.file.DecRef(pseg.fileRange())
+
+			pma.file = mem
+			pma.off = fr.Start
+			pma.private = true
+			pma.needCOW = false
+			pma.internalMappings = safemem.BlockSeq{}
+
+			// Try to merge pma with its neighbors.
+			if prev := pseg.PrevSegment(); prev.Ok() {
+				if merged := mm.pmas.Merge(prev, pseg); merged.Ok() {
+					pseg = merged
+					invalidatedIterators = true
+				}
+			}
+			if next := pseg.NextSegment(); next.Ok() {
+				if merged := mm.pmas.Merge(pseg, next); merged.Ok() {
+					pseg = merged
+					invalidatedIterators = true
+				}
+			}
+
+			// If an error occurred after ar.End, breakCopyOnWriteLocked still
+			// did its job, so discard the error.
+			if err != nil && pseg.End() < ar.End {
+				return pseg.NextGap(), invalidatedIterators, err
+			}
+		}
+		// This checks against ar.End, not maskAR.End, so we will never break
+		// COW on a pma that does not intersect ar.
+		if ar.End <= pseg.End() {
+			return pseg.NextGap(), invalidatedIterators, nil
+		}
+		pseg = pseg.NextSegment()
+	}
+}
+
+// Preconditions: mm.activeMu must be locked for writing.
+func (mm *MemoryManager) isPMACopyOnWriteLocked(pseg pmaIterator) bool {
+	pma := pseg.ValuePtr()
+	if !pma.needCOW {
+		return false
+	}
+	if !pma.private {
+		return true
+	}
+	// If we have the only reference on private memory to be copied, just take
+	// ownership of it instead of copying. If we do hold the only reference,
+	// additional references can only be taken by mm.Fork(), which is excluded
+	// by mm.activeMu, so this isn't racy.
+	mm.privateRefs.mu.Lock()
+	defer mm.privateRefs.mu.Unlock()
+	fr := pseg.fileRange()
+	// This check relies on mm.privateRefs.refs being kept fully merged.
+	rseg := mm.privateRefs.refs.FindSegment(fr.Start)
+	if rseg.Ok() && rseg.Value() == 1 && fr.End <= rseg.End() {
+		pma.needCOW = false
+		return false
+	}
+	return true
+}
+
+// Invalidate implements memmap.MappingSpace.Invalidate.
+func (mm *MemoryManager) Invalidate(ar usermem.AddrRange, opts memmap.InvalidateOpts) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	mm.activeMu.Lock()
+	defer mm.activeMu.Unlock()
+	if mm.captureInvalidations {
+		mm.capturedInvalidations = append(mm.capturedInvalidations, invalidateArgs{ar, opts})
+		return
+	}
+	mm.invalidateLocked(ar, opts.InvalidatePrivate, true)
+}
+
+// invalidateLocked removes pmas and AddressSpace mappings of those pmas for
+// addresses in ar.
+//
+// Preconditions: mm.activeMu must be locked for writing. ar.Length() != 0. ar
+// must be page-aligned.
+func (mm *MemoryManager) invalidateLocked(ar usermem.AddrRange, invalidatePrivate, invalidateShared bool) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	var didUnmapAS bool
+	pseg := mm.pmas.LowerBoundSegment(ar.Start)
+	for pseg.Ok() && pseg.Start() < ar.End {
+		pma := pseg.ValuePtr()
+		if (invalidatePrivate && pma.private) || (invalidateShared && !pma.private) {
+			pseg = mm.pmas.Isolate(pseg, ar)
+			pma = pseg.ValuePtr()
+			if !didUnmapAS {
+				// Unmap all of ar, not just pseg.Range(), to minimize host
+				// syscalls. AddressSpace mappings must be removed before
+				// mm.decPrivateRef().
+				mm.unmapASLocked(ar)
+				didUnmapAS = true
+			}
+			if pma.private {
+				mm.decPrivateRef(pseg.fileRange())
+			}
+			mm.removeRSSLocked(pseg.Range())
+			pma.file.DecRef(pseg.fileRange())
+			pseg = mm.pmas.Remove(pseg).NextSegment()
+		} else {
+			pseg = pseg.NextSegment()
+		}
+	}
+}
+
+// movePMAsLocked moves all pmas in oldAR to newAR.
+//
+// Preconditions: mm.activeMu must be locked for writing. oldAR.Length() != 0.
+// oldAR.Length() == newAR.Length(). !oldAR.Overlaps(newAR).
+// mm.pmas.IsEmptyRange(newAR). oldAR and newAR must be page-aligned.
+func (mm *MemoryManager) movePMAsLocked(oldAR, newAR usermem.AddrRange) {
+	if checkInvariants {
+		if !oldAR.WellFormed() || oldAR.Length() <= 0 || !oldAR.IsPageAligned() {
+			panic(fmt.Sprintf("invalid oldAR: %v", oldAR))
+		}
+		if !newAR.WellFormed() || newAR.Length() <= 0 || !newAR.IsPageAligned() {
+			panic(fmt.Sprintf("invalid newAR: %v", newAR))
+		}
+		if oldAR.Length() != newAR.Length() {
+			panic(fmt.Sprintf("old and new address ranges have different lengths: %v, %v", oldAR, newAR))
+		}
+		if oldAR.Overlaps(newAR) {
+			panic(fmt.Sprintf("old and new address ranges overlap: %v, %v", oldAR, newAR))
+		}
+		// mm.pmas.IsEmptyRange is checked by mm.pmas.Insert.
+	}
+
+	type movedPMA struct {
+		oldAR usermem.AddrRange
+		pma   pma
+	}
+	var movedPMAs []movedPMA
+	pseg := mm.pmas.LowerBoundSegment(oldAR.Start)
+	for pseg.Ok() && pseg.Start() < oldAR.End {
+		pseg = mm.pmas.Isolate(pseg, oldAR)
+		movedPMAs = append(movedPMAs, movedPMA{
+			oldAR: pseg.Range(),
+			pma:   pseg.Value(),
+		})
+		mm.removeRSSLocked(pseg.Range())
+		pseg = mm.pmas.Remove(pseg).NextSegment()
+	}
+
+	off := newAR.Start - oldAR.Start
+	pgap := mm.pmas.FindGap(newAR.Start)
+	for i := range movedPMAs {
+		mpma := &movedPMAs[i]
+		pmaNewAR := usermem.AddrRange{mpma.oldAR.Start + off, mpma.oldAR.End + off}
+		mm.addRSSLocked(pmaNewAR)
+		pgap = mm.pmas.Insert(pgap, pmaNewAR, mpma.pma).NextGap()
+	}
+
+	mm.unmapASLocked(oldAR)
+}
+
+// getPMAInternalMappingsLocked ensures that pmas for all addresses in ar have
+// cached internal mappings. It returns:
+//
+// - An iterator to the gap after the last pma with internal mappings
+// containing an address in ar. If internal mappings exist for no addresses in
+// ar, the iterator is to a gap that begins before ar.Start.
+//
+// - An error that is non-nil if internal mappings exist for only a subset of
+// ar.
+//
+// Preconditions: mm.activeMu must be locked for writing.
+// pseg.Range().Contains(ar.Start). pmas must exist for all addresses in ar.
+// ar.Length() != 0.
+//
+// Postconditions: getPMAInternalMappingsLocked does not invalidate iterators
+// into mm.pmas.
+func (mm *MemoryManager) getPMAInternalMappingsLocked(pseg pmaIterator, ar usermem.AddrRange) (pmaGapIterator, error) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !pseg.Range().Contains(ar.Start) {
+			panic(fmt.Sprintf("initial pma %v does not cover start of ar %v", pseg.Range(), ar))
+		}
+	}
+
+	for {
+		if err := pseg.getInternalMappingsLocked(); err != nil {
+			return pseg.PrevGap(), err
+		}
+		if ar.End <= pseg.End() {
+			return pseg.NextGap(), nil
+		}
+		pseg, _ = pseg.NextNonEmpty()
+	}
+}
+
+// getVecPMAInternalMappingsLocked ensures that pmas for all addresses in ars
+// have cached internal mappings. It returns the subset of ars for which
+// internal mappings exist. If this is not equal to ars, it returns a non-nil
+// error explaining why.
+//
+// Preconditions: mm.activeMu must be locked for writing. pmas must exist for
+// all addresses in ar.
+//
+// Postconditions: getVecPMAInternalMappingsLocked does not invalidate iterators
+// into mm.pmas.
+func (mm *MemoryManager) getVecPMAInternalMappingsLocked(ars usermem.AddrRangeSeq) (usermem.AddrRangeSeq, error) {
+	for arsit := ars; !arsit.IsEmpty(); arsit = arsit.Tail() {
+		ar := arsit.Head()
+		if ar.Length() == 0 {
+			continue
+		}
+		if pend, err := mm.getPMAInternalMappingsLocked(mm.pmas.FindSegment(ar.Start), ar); err != nil {
+			return truncatedAddrRangeSeq(ars, arsit, pend.Start()), err
+		}
+	}
+	return ars, nil
+}
+
+// internalMappingsLocked returns internal mappings for addresses in ar.
+//
+// Preconditions: mm.activeMu must be locked. Internal mappings must have been
+// previously established for all addresses in ar. ar.Length() != 0.
+// pseg.Range().Contains(ar.Start).
+func (mm *MemoryManager) internalMappingsLocked(pseg pmaIterator, ar usermem.AddrRange) safemem.BlockSeq {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !pseg.Range().Contains(ar.Start) {
+			panic(fmt.Sprintf("initial pma %v does not cover start of ar %v", pseg.Range(), ar))
+		}
+	}
+
+	if ar.End <= pseg.End() {
+		// Since only one pma is involved, we can use pma.internalMappings
+		// directly, avoiding a slice allocation.
+		offset := uint64(ar.Start - pseg.Start())
+		return pseg.ValuePtr().internalMappings.DropFirst64(offset).TakeFirst64(uint64(ar.Length()))
+	}
+
+	var ims []safemem.Block
+	for {
+		pr := pseg.Range().Intersect(ar)
+		for pims := pseg.ValuePtr().internalMappings.DropFirst64(uint64(pr.Start - pseg.Start())).TakeFirst64(uint64(pr.Length())); !pims.IsEmpty(); pims = pims.Tail() {
+			ims = append(ims, pims.Head())
+		}
+		if ar.End <= pseg.End() {
+			break
+		}
+		pseg = pseg.NextSegment()
+	}
+	return safemem.BlockSeqFromSlice(ims)
+}
+
+// vecInternalMappingsLocked returns internal mappings for addresses in ars.
+//
+// Preconditions: mm.activeMu must be locked. Internal mappings must have been
+// previously established for all addresses in ars.
+func (mm *MemoryManager) vecInternalMappingsLocked(ars usermem.AddrRangeSeq) safemem.BlockSeq {
+	var ims []safemem.Block
+	for ; !ars.IsEmpty(); ars = ars.Tail() {
+		ar := ars.Head()
+		if ar.Length() == 0 {
+			continue
+		}
+		for pims := mm.internalMappingsLocked(mm.pmas.FindSegment(ar.Start), ar); !pims.IsEmpty(); pims = pims.Tail() {
+			ims = append(ims, pims.Head())
+		}
+	}
+	return safemem.BlockSeqFromSlice(ims)
+}
+
+// incPrivateRef acquires a reference on private pages in fr.
+func (mm *MemoryManager) incPrivateRef(fr platform.FileRange) {
+	mm.privateRefs.mu.Lock()
+	defer mm.privateRefs.mu.Unlock()
+	refSet := &mm.privateRefs.refs
+	seg, gap := refSet.Find(fr.Start)
+	for {
+		switch {
+		case seg.Ok() && seg.Start() < fr.End:
+			seg = refSet.Isolate(seg, fr)
+			seg.SetValue(seg.Value() + 1)
+			seg, gap = seg.NextNonEmpty()
+		case gap.Ok() && gap.Start() < fr.End:
+			seg, gap = refSet.InsertWithoutMerging(gap, gap.Range().Intersect(fr), 1).NextNonEmpty()
+		default:
+			refSet.MergeAdjacent(fr)
+			return
+		}
+	}
+}
+
+// decPrivateRef releases a reference on private pages in fr.
+func (mm *MemoryManager) decPrivateRef(fr platform.FileRange) {
+	var freed []platform.FileRange
+
+	mm.privateRefs.mu.Lock()
+	refSet := &mm.privateRefs.refs
+	seg := refSet.LowerBoundSegment(fr.Start)
+	for seg.Ok() && seg.Start() < fr.End {
+		seg = refSet.Isolate(seg, fr)
+		if old := seg.Value(); old == 1 {
+			freed = append(freed, seg.Range())
+			seg = refSet.Remove(seg).NextSegment()
+		} else {
+			seg.SetValue(old - 1)
+			seg = seg.NextSegment()
+		}
+	}
+	refSet.MergeAdjacent(fr)
+	mm.privateRefs.mu.Unlock()
+
+	mem := mm.p.Memory()
+	for _, fr := range freed {
+		mem.DecRef(fr)
+	}
+}
+
+// addRSSLocked updates the current and maximum resident set size of a
+// MemoryManager to reflect the insertion of a pma at ar.
+//
+// Preconditions: mm.activeMu must be locked for writing.
+func (mm *MemoryManager) addRSSLocked(ar usermem.AddrRange) {
+	mm.curRSS += uint64(ar.Length())
+	if mm.curRSS > mm.maxRSS {
+		mm.maxRSS = mm.curRSS
+	}
+}
+
+// removeRSSLocked updates the current resident set size of a MemoryManager to
+// reflect the removal of a pma at ar.
+//
+// Preconditions: mm.activeMu must be locked for writing.
+func (mm *MemoryManager) removeRSSLocked(ar usermem.AddrRange) {
+	mm.curRSS -= uint64(ar.Length())
+}
+
+// pmaSetFunctions implements segment.Functions for pmaSet.
+type pmaSetFunctions struct{}
+
+func (pmaSetFunctions) MinKey() usermem.Addr {
+	return 0
+}
+
+func (pmaSetFunctions) MaxKey() usermem.Addr {
+	return ^usermem.Addr(0)
+}
+
+func (pmaSetFunctions) ClearValue(pma *pma) {
+	pma.file = nil
+	pma.internalMappings = safemem.BlockSeq{}
+}
+
+func (pmaSetFunctions) Merge(ar1 usermem.AddrRange, pma1 pma, ar2 usermem.AddrRange, pma2 pma) (pma, bool) {
+	if pma1.file != pma2.file ||
+		pma1.off+uint64(ar1.Length()) != pma2.off ||
+		pma1.vmaEffectivePerms != pma2.vmaEffectivePerms ||
+		pma1.vmaMaxPerms != pma2.vmaMaxPerms ||
+		pma1.needCOW != pma2.needCOW ||
+		pma1.private != pma2.private {
+		return pma{}, false
+	}
+
+	// Discard internal mappings instead of trying to merge them, since merging
+	// them requires an allocation and getting them again from the
+	// platform.File might not.
+	pma1.internalMappings = safemem.BlockSeq{}
+	return pma1, true
+}
+
+func (pmaSetFunctions) Split(ar usermem.AddrRange, p pma, split usermem.Addr) (pma, pma) {
+	newlen1 := uint64(split - ar.Start)
+	p2 := p
+	p2.off += newlen1
+	if !p.internalMappings.IsEmpty() {
+		p.internalMappings = p.internalMappings.TakeFirst64(newlen1)
+		p2.internalMappings = p2.internalMappings.DropFirst64(newlen1)
+	}
+	return p, p2
+}
+
+// findOrSeekPrevUpperBoundPMA returns mm.pmas.UpperBoundSegment(addr), but may do
+// so by scanning linearly backward from pgap.
+//
+// Preconditions: mm.activeMu must be locked. addr <= pgap.Start().
+func (mm *MemoryManager) findOrSeekPrevUpperBoundPMA(addr usermem.Addr, pgap pmaGapIterator) pmaIterator {
+	if checkInvariants {
+		if !pgap.Ok() {
+			panic("terminal pma iterator")
+		}
+		if addr > pgap.Start() {
+			panic(fmt.Sprintf("can't seek backward to %#x from %#x", addr, pgap.Start()))
+		}
+	}
+	// Optimistically check if pgap.PrevSegment() is the PMA we're looking for,
+	// which is the case if findOrSeekPrevUpperBoundPMA is called to find the
+	// start of a range containing only a single PMA.
+	if pseg := pgap.PrevSegment(); pseg.Start() <= addr {
+		return pseg
+	}
+	return mm.pmas.UpperBoundSegment(addr)
+}
+
+// getInternalMappingsLocked ensures that pseg.ValuePtr().internalMappings is
+// non-empty.
+//
+// Preconditions: mm.activeMu must be locked for writing.
+func (pseg pmaIterator) getInternalMappingsLocked() error {
+	pma := pseg.ValuePtr()
+	if pma.internalMappings.IsEmpty() {
+		// Internal mappings are used for ignorePermissions accesses,
+		// so we need to use vma.maxPerms instead of
+		// vma.effectivePerms. However, we will never execute
+		// application code through an internal mapping, and we don't
+		// actually need a writable mapping if copy-on-write is in
+		// effect. (But get a writable mapping anyway if the pma is
+		// private, so that if breakCopyOnWriteLocked =>
+		// isPMACopyOnWriteLocked takes ownership of the pma instead of
+		// copying, it doesn't need to get a new mapping.)
+		perms := pma.vmaMaxPerms
+		perms.Execute = false
+		if pma.needCOW && !pma.private {
+			perms.Write = false
+		}
+		ims, err := pma.file.MapInternal(pseg.fileRange(), perms)
+		if err != nil {
+			return err
+		}
+		pma.internalMappings = ims
+	}
+	return nil
+}
+
+func (pseg pmaIterator) fileRange() platform.FileRange {
+	return pseg.fileRangeOf(pseg.Range())
+}
+
+// Preconditions: pseg.Range().IsSupersetOf(ar). ar.Length != 0.
+func (pseg pmaIterator) fileRangeOf(ar usermem.AddrRange) platform.FileRange {
+	if checkInvariants {
+		if !pseg.Ok() {
+			panic("terminal pma iterator")
+		}
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !pseg.Range().IsSupersetOf(ar) {
+			panic(fmt.Sprintf("ar %v out of bounds %v", ar, pseg.Range()))
+		}
+	}
+
+	pma := pseg.ValuePtr()
+	pstart := pseg.Start()
+	return platform.FileRange{pma.off + uint64(ar.Start-pstart), pma.off + uint64(ar.End-pstart)}
+}
diff --git a/pkg/sentry/mm/proc_pid_maps.go b/pkg/sentry/mm/proc_pid_maps.go
new file mode 100644
index 000000000..5840b257c
--- /dev/null
+++ b/pkg/sentry/mm/proc_pid_maps.go
@@ -0,0 +1,105 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"bytes"
+	"fmt"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/proc/seqfile"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// devMinorBits is the number of minor bits in a device number. Linux:
+	// include/linux/kdev_t.h:MINORBITS
+	devMinorBits = 20
+)
+
+// NeedsUpdate implements seqfile.SeqSource.NeedsUpdate.
+func (mm *MemoryManager) NeedsUpdate(generation int64) bool {
+	return true
+}
+
+// ReadSeqFileData is called by fs/proc.mapsData.ReadSeqFileData.
+func (mm *MemoryManager) ReadSeqFileData(ctx context.Context, handle seqfile.SeqHandle) ([]seqfile.SeqData, int64) {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	var data []seqfile.SeqData
+	var start usermem.Addr
+	if handle != nil {
+		start = *handle.(*usermem.Addr)
+	}
+	for vseg := mm.vmas.LowerBoundSegment(start); vseg.Ok(); vseg = vseg.NextSegment() {
+		// FIXME: If we use a usermem.Addr for the handle, we get
+		// "panic: autosave error: type usermem.Addr is not registered".
+		vmaAddr := vseg.End()
+		data = append(data, seqfile.SeqData{
+			Buf:    mm.vmaMapsEntryLocked(ctx, vseg),
+			Handle: &vmaAddr,
+		})
+	}
+	return data, 1
+}
+
+// vmaMapsEntryLocked returns a /proc/[pid]/maps entry for the vma iterated by
+// vseg, including the trailing newline.
+//
+// Preconditions: mm.mappingMu must be locked.
+func (mm *MemoryManager) vmaMapsEntryLocked(ctx context.Context, vseg vmaIterator) []byte {
+	vma := vseg.ValuePtr()
+	private := "p"
+	if !vma.private {
+		private = "s"
+	}
+
+	var dev, ino uint64
+	if vma.id != nil {
+		dev = vma.id.DeviceID()
+		ino = vma.id.InodeID()
+	}
+	devMajor := uint32(dev >> devMinorBits)
+	devMinor := uint32(dev & ((1 << devMinorBits) - 1))
+
+	var b bytes.Buffer
+	// Do not include the guard page: fs/proc/task_mmu.c:show_map_vma() =>
+	// stack_guard_page_start().
+	fmt.Fprintf(&b, "%08x-%08x %s%s %08x %02x:%02x %d ",
+		vseg.Start(), vseg.End(), vma.realPerms, private, vma.off, devMajor, devMinor, ino)
+
+	// Figure out our filename or hint.
+	var s string
+	if vma.hint != "" {
+		s = vma.hint
+	} else if vma.id != nil {
+		// FIXME: We are holding mm.mappingMu here, which is
+		// consistent with Linux's holding mmap_sem in
+		// fs/proc/task_mmu.c:show_map_vma() => fs/seq_file.c:seq_file_path().
+		// However, it's not clear that fs.File.MappedName() is actually
+		// consistent with this lock order.
+		s = vma.id.MappedName(ctx)
+	}
+	if s != "" {
+		// Per linux, we pad until the 74th character.
+		if pad := 73 - b.Len(); pad > 0 {
+			b.WriteString(strings.Repeat(" ", pad))
+		}
+		b.WriteString(s)
+	}
+	b.WriteString("\n")
+	return b.Bytes()
+}
diff --git a/pkg/sentry/mm/save_restore.go b/pkg/sentry/mm/save_restore.go
new file mode 100644
index 000000000..36fed8f1c
--- /dev/null
+++ b/pkg/sentry/mm/save_restore.go
@@ -0,0 +1,57 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// InvalidateUnsavable invokes memmap.Mappable.InvalidateUnsavable on all
+// Mappables mapped by mm.
+func (mm *MemoryManager) InvalidateUnsavable(ctx context.Context) error {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	for vseg := mm.vmas.FirstSegment(); vseg.Ok(); vseg = vseg.NextSegment() {
+		if vma := vseg.ValuePtr(); vma.mappable != nil {
+			if err := vma.mappable.InvalidateUnsavable(ctx); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// beforeSave is invoked by stateify.
+func (mm *MemoryManager) beforeSave() {
+	mem := mm.p.Memory()
+	for pseg := mm.pmas.FirstSegment(); pseg.Ok(); pseg = pseg.NextSegment() {
+		if pma := pseg.ValuePtr(); pma.file != mem {
+			// InvalidateUnsavable should have caused all such pmas to be
+			// invalidated.
+			panic(fmt.Sprintf("Can't save pma %#v with non-Memory file of type %T:\n%s", pseg.Range(), pma.file, mm))
+		}
+	}
+}
+
+// afterLoad is invoked by stateify.
+func (mm *MemoryManager) afterLoad() {
+	mm.haveASIO = mm.p.SupportsAddressSpaceIO()
+	mem := mm.p.Memory()
+	for pseg := mm.pmas.FirstSegment(); pseg.Ok(); pseg = pseg.NextSegment() {
+		pseg.ValuePtr().file = mem
+	}
+}
diff --git a/pkg/sentry/mm/special_mappable.go b/pkg/sentry/mm/special_mappable.go
new file mode 100644
index 000000000..9d3614034
--- /dev/null
+++ b/pkg/sentry/mm/special_mappable.go
@@ -0,0 +1,147 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SpecialMappable implements memmap.MappingIdentity and memmap.Mappable with
+// semantics similar to Linux's mm/mmap.c:_install_special_mapping(), except
+// that SpecialMappable takes ownership of the memory that it represents
+// (_install_special_mapping() does not.)
+type SpecialMappable struct {
+	refs.AtomicRefCount
+
+	p    platform.Platform
+	fr   platform.FileRange
+	name string
+}
+
+// NewSpecialMappable returns a SpecialMappable that owns fr, which represents
+// offsets in p.Memory() that contain the SpecialMappable's data. The
+// SpecialMappable will use the given name in /proc/[pid]/maps.
+//
+// Preconditions: fr.Length() != 0.
+func NewSpecialMappable(name string, p platform.Platform, fr platform.FileRange) *SpecialMappable {
+	return &SpecialMappable{p: p, fr: fr, name: name}
+}
+
+// DecRef implements refs.RefCounter.DecRef.
+func (m *SpecialMappable) DecRef() {
+	m.AtomicRefCount.DecRefWithDestructor(func() {
+		m.p.Memory().DecRef(m.fr)
+	})
+}
+
+// MappedName implements memmap.MappingIdentity.MappedName.
+func (m *SpecialMappable) MappedName(ctx context.Context) string {
+	return m.name
+}
+
+// DeviceID implements memmap.MappingIdentity.DeviceID.
+func (m *SpecialMappable) DeviceID() uint64 {
+	return 0
+}
+
+// InodeID implements memmap.MappingIdentity.InodeID.
+func (m *SpecialMappable) InodeID() uint64 {
+	return 0
+}
+
+// Msync implements memmap.MappingIdentity.Msync.
+func (m *SpecialMappable) Msync(ctx context.Context, mr memmap.MappableRange) error {
+	// Linux: vm_file is NULL, causing msync to skip it entirely.
+	return nil
+}
+
+// AddMapping implements memmap.Mappable.AddMapping.
+func (m *SpecialMappable) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) error {
+	return nil
+}
+
+// RemoveMapping implements memmap.Mappable.RemoveMapping.
+func (m *SpecialMappable) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64) {
+}
+
+// CopyMapping implements memmap.Mappable.CopyMapping.
+func (m *SpecialMappable) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64) error {
+	return nil
+}
+
+// Translate implements memmap.Mappable.Translate.
+func (m *SpecialMappable) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) {
+	var err error
+	if required.End > m.fr.Length() {
+		err = &memmap.BusError{syserror.EFAULT}
+	}
+	if source := optional.Intersect(memmap.MappableRange{0, m.fr.Length()}); source.Length() != 0 {
+		return []memmap.Translation{
+			{
+				Source: source,
+				File:   m.p.Memory(),
+				Offset: m.fr.Start + source.Start,
+			},
+		}, err
+	}
+	return nil, err
+}
+
+// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable.
+func (m *SpecialMappable) InvalidateUnsavable(ctx context.Context) error {
+	// Since data is stored in platform.Platform.Memory(), the contents of
+	// which are preserved across save/restore, we don't need to do anything.
+	return nil
+}
+
+// Platform returns the Platform whose Memory stores the SpecialMappable's
+// contents.
+func (m *SpecialMappable) Platform() platform.Platform {
+	return m.p
+}
+
+// FileRange returns the offsets into Platform().Memory() that store the
+// SpecialMappable's contents.
+func (m *SpecialMappable) FileRange() platform.FileRange {
+	return m.fr
+}
+
+// Length returns the length of the SpecialMappable.
+func (m *SpecialMappable) Length() uint64 {
+	return m.fr.Length()
+}
+
+// NewSharedAnonMappable returns a SpecialMappable that implements the
+// semantics of mmap(MAP_SHARED|MAP_ANONYMOUS) and mappings of /dev/zero.
+//
+// TODO: The use of SpecialMappable is a lazy code reuse hack. Linux
+// uses an ephemeral file created by mm/shmem.c:shmem_zero_setup(); we should
+// do the same to get non-zero device and inode IDs.
+func NewSharedAnonMappable(length uint64, p platform.Platform) (*SpecialMappable, error) {
+	if length == 0 || length != uint64(usermem.Addr(length).RoundDown()) {
+		return nil, syserror.EINVAL
+	}
+	fr, err := p.Memory().Allocate(length, usage.Anonymous)
+	if err != nil {
+		return nil, err
+	}
+	return NewSpecialMappable("/dev/zero (deleted)", p, fr), nil
+}
diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go
new file mode 100644
index 000000000..0730be65b
--- /dev/null
+++ b/pkg/sentry/mm/syscalls.go
@@ -0,0 +1,794 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+	mrand "math/rand"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// HandleUserFault handles an application page fault. sp is the faulting
+// application thread's stack pointer.
+//
+// Preconditions: mm.as != nil.
+func (mm *MemoryManager) HandleUserFault(ctx context.Context, addr usermem.Addr, at usermem.AccessType, sp usermem.Addr) error {
+	ar, ok := addr.RoundDown().ToRange(usermem.PageSize)
+	if !ok {
+		return syserror.EFAULT
+	}
+
+	// Don't bother trying existingPMAsLocked; in most cases, if we did have
+	// existing pmas, we wouldn't have faulted.
+
+	// Ensure that we have a usable vma. Here and below, since we are only
+	// asking for a single page, there is no possibility of partial success,
+	// and any error is immediately fatal.
+	mm.mappingMu.RLock()
+	vseg, _, err := mm.getVMAsLocked(ctx, ar, at, false)
+	if err != nil {
+		mm.mappingMu.RUnlock()
+		return err
+	}
+
+	// Ensure that we have a usable pma.
+	mm.activeMu.Lock()
+	pseg, _, err := mm.getPMAsLocked(ctx, vseg, ar, pmaOpts{
+		breakCOW: at.Write,
+	})
+	mm.mappingMu.RUnlock()
+	if err != nil {
+		mm.activeMu.Unlock()
+		return err
+	}
+
+	// Downgrade to a read-lock on activeMu since we don't need to mutate pmas
+	// anymore.
+	mm.activeMu.DowngradeLock()
+
+	// Map the faulted page into the active AddressSpace.
+	err = mm.mapASLocked(pseg, ar, false)
+	mm.activeMu.RUnlock()
+	return err
+}
+
+// MMap establishes a memory mapping.
+func (mm *MemoryManager) MMap(ctx context.Context, opts memmap.MMapOpts) (usermem.Addr, error) {
+	if opts.Length == 0 {
+		return 0, syserror.EINVAL
+	}
+	length, ok := usermem.Addr(opts.Length).RoundUp()
+	if !ok {
+		return 0, syserror.ENOMEM
+	}
+	opts.Length = uint64(length)
+
+	if opts.Mappable != nil {
+		// Offset must be aligned.
+		if usermem.Addr(opts.Offset).RoundDown() != usermem.Addr(opts.Offset) {
+			return 0, syserror.EINVAL
+		}
+		// Offset + length must not overflow.
+		if end := opts.Offset + opts.Length; end < opts.Offset {
+			return 0, syserror.ENOMEM
+		}
+	} else {
+		opts.Offset = 0
+		if !opts.Private {
+			if opts.MappingIdentity != nil {
+				return 0, syserror.EINVAL
+			}
+			m, err := NewSharedAnonMappable(opts.Length, platform.FromContext(ctx))
+			if err != nil {
+				return 0, err
+			}
+			opts.MappingIdentity = m
+			opts.Mappable = m
+		}
+	}
+
+	if opts.Addr.RoundDown() != opts.Addr {
+		// MAP_FIXED requires addr to be page-aligned; non-fixed mappings
+		// don't.
+		if opts.Fixed {
+			return 0, syserror.EINVAL
+		}
+		opts.Addr = opts.Addr.RoundDown()
+	}
+
+	if !opts.MaxPerms.SupersetOf(opts.Perms) {
+		return 0, syserror.EACCES
+	}
+	if opts.Unmap && !opts.Fixed {
+		return 0, syserror.EINVAL
+	}
+	if opts.GrowsDown && opts.Mappable != nil {
+		return 0, syserror.EINVAL
+	}
+
+	// Get the new vma.
+	mm.mappingMu.Lock()
+	vseg, ar, err := mm.createVMALocked(ctx, opts)
+	if err != nil {
+		mm.mappingMu.Unlock()
+		return 0, err
+	}
+
+	switch {
+	case opts.Precommit:
+		// Get pmas and map with precommit as requested.
+		mm.populateAndUnlock(ctx, vseg, ar, true)
+
+	case opts.Mappable == nil && length <= privateAllocUnit:
+		// NOTE: Get pmas and map eagerly in the hope
+		// that doing so will save on future page faults. We only do this for
+		// anonymous mappings, since otherwise the cost of
+		// memmap.Mappable.Translate is unknown; and only for small mappings,
+		// to avoid needing to allocate large amounts of memory that we may
+		// subsequently need to checkpoint.
+		mm.populateAndUnlock(ctx, vseg, ar, false)
+
+	default:
+		mm.mappingMu.Unlock()
+	}
+
+	return ar.Start, nil
+}
+
+// Preconditions: mm.mappingMu must be locked for writing.
+//
+// Postconditions: mm.mappingMu will be unlocked.
+func (mm *MemoryManager) populateAndUnlock(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, precommit bool) {
+	if !vseg.ValuePtr().effectivePerms.Any() {
+		// Linux doesn't populate inaccessible pages. See
+		// mm/gup.c:populate_vma_page_range.
+		mm.mappingMu.Unlock()
+		return
+	}
+
+	mm.activeMu.Lock()
+
+	// Even if we get a new pma, we can't actually map it if we don't have an
+	// AddressSpace.
+	if mm.as == nil {
+		mm.activeMu.Unlock()
+		mm.mappingMu.Unlock()
+		return
+	}
+
+	// Ensure that we have usable pmas.
+	mm.mappingMu.DowngradeLock()
+	pseg, _, err := mm.getPMAsLocked(ctx, vseg, ar, pmaOpts{})
+	mm.mappingMu.RUnlock()
+	if err != nil {
+		// mm/util.c:vm_mmap_pgoff() ignores the error, if any, from
+		// mm/gup.c:mm_populate(). If it matters, we'll get it again when
+		// userspace actually tries to use the failing page.
+		mm.activeMu.Unlock()
+		return
+	}
+
+	// Downgrade to a read-lock on activeMu since we don't need to mutate pmas
+	// anymore.
+	mm.activeMu.DowngradeLock()
+
+	// As above, errors are silently ignored.
+	mm.mapASLocked(pseg, ar, precommit)
+	mm.activeMu.RUnlock()
+}
+
+// MapStack allocates the initial process stack.
+func (mm *MemoryManager) MapStack(ctx context.Context) (usermem.AddrRange, error) {
+	// maxStackSize is the maximum supported process stack size in bytes.
+	//
+	// This limit exists because stack growing isn't implemented, so the entire
+	// process stack must be mapped up-front.
+	const maxStackSize = 128 << 20
+
+	stackSize := limits.FromContext(ctx).Get(limits.Stack)
+	r, ok := usermem.Addr(stackSize.Cur).RoundUp()
+	sz := uint64(r)
+	if !ok {
+		// RLIM_INFINITY rounds up to 0.
+		sz = linux.DefaultStackSoftLimit
+	} else if sz > maxStackSize {
+		ctx.Warningf("Capping stack size from RLIMIT_STACK of %v down to %v.", sz, maxStackSize)
+		sz = maxStackSize
+	} else if sz == 0 {
+		return usermem.AddrRange{}, syserror.ENOMEM
+	}
+	szaddr := usermem.Addr(sz)
+	ctx.Debugf("Allocating stack with size of %v bytes", sz)
+
+	// Determine the stack's desired location. Unlike Linux, address
+	// randomization can't be disabled.
+	stackEnd := mm.layout.MaxAddr - usermem.Addr(mrand.Int63n(int64(mm.layout.MaxStackRand))).RoundDown()
+	if stackEnd < szaddr {
+		return usermem.AddrRange{}, syserror.ENOMEM
+	}
+	stackStart := stackEnd - szaddr
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	_, ar, err := mm.createVMALocked(ctx, memmap.MMapOpts{
+		Length:    sz,
+		Addr:      stackStart,
+		Perms:     usermem.ReadWrite,
+		MaxPerms:  usermem.AnyAccess,
+		Private:   true,
+		GrowsDown: true,
+		Hint:      "[stack]",
+	})
+	return ar, err
+}
+
+// MUnmap implements the semantics of Linux's munmap(2).
+func (mm *MemoryManager) MUnmap(ctx context.Context, addr usermem.Addr, length uint64) error {
+	if addr != addr.RoundDown() {
+		return syserror.EINVAL
+	}
+	if length == 0 {
+		return syserror.EINVAL
+	}
+	la, ok := usermem.Addr(length).RoundUp()
+	if !ok {
+		return syserror.EINVAL
+	}
+	ar, ok := addr.ToRange(uint64(la))
+	if !ok {
+		return syserror.EINVAL
+	}
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	mm.unmapLocked(ctx, ar)
+	return nil
+}
+
+// MRemapOpts specifies options to MRemap.
+type MRemapOpts struct {
+	// Move controls whether MRemap moves the remapped mapping to a new address.
+	Move MRemapMoveMode
+
+	// NewAddr is the new address for the remapping. NewAddr is ignored unless
+	// Move is MMRemapMustMove.
+	NewAddr usermem.Addr
+}
+
+// MRemapMoveMode controls MRemap's moving behavior.
+type MRemapMoveMode int
+
+const (
+	// MRemapNoMove prevents MRemap from moving the remapped mapping.
+	MRemapNoMove MRemapMoveMode = iota
+
+	// MRemapMayMove allows MRemap to move the remapped mapping.
+	MRemapMayMove
+
+	// MRemapMustMove requires MRemap to move the remapped mapping to
+	// MRemapOpts.NewAddr, replacing any existing mappings in the remapped
+	// range.
+	MRemapMustMove
+)
+
+// MRemap implements the semantics of Linux's mremap(2).
+func (mm *MemoryManager) MRemap(ctx context.Context, oldAddr usermem.Addr, oldSize uint64, newSize uint64, opts MRemapOpts) (usermem.Addr, error) {
+	// "Note that old_address has to be page aligned." - mremap(2)
+	if oldAddr.RoundDown() != oldAddr {
+		return 0, syserror.EINVAL
+	}
+
+	// Linux treats an old_size that rounds up to 0 as 0, which is otherwise a
+	// valid size. However, new_size can't be 0 after rounding.
+	oldSizeAddr, _ := usermem.Addr(oldSize).RoundUp()
+	oldSize = uint64(oldSizeAddr)
+	newSizeAddr, ok := usermem.Addr(newSize).RoundUp()
+	if !ok || newSizeAddr == 0 {
+		return 0, syserror.EINVAL
+	}
+	newSize = uint64(newSizeAddr)
+
+	oldEnd, ok := oldAddr.AddLength(oldSize)
+	if !ok {
+		return 0, syserror.EINVAL
+	}
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+
+	// All cases require that a vma exists at oldAddr.
+	vseg := mm.vmas.FindSegment(oldAddr)
+	if !vseg.Ok() {
+		return 0, syserror.EFAULT
+	}
+
+	if opts.Move != MRemapMustMove {
+		// Handle noops and in-place shrinking. These cases don't care if
+		// [oldAddr, oldEnd) maps to a single vma, or is even mapped at all
+		// (aside from oldAddr).
+		if newSize <= oldSize {
+			if newSize < oldSize {
+				// If oldAddr+oldSize didn't overflow, oldAddr+newSize can't
+				// either.
+				newEnd := oldAddr + usermem.Addr(newSize)
+				mm.unmapLocked(ctx, usermem.AddrRange{newEnd, oldEnd})
+			}
+			return oldAddr, nil
+		}
+
+		// Handle in-place growing.
+
+		// Check that oldEnd maps to the same vma as oldAddr.
+		if vseg.End() < oldEnd {
+			return 0, syserror.EFAULT
+		}
+		// "Grow" the existing vma by creating a new mergeable one.
+		vma := vseg.ValuePtr()
+		var newOffset uint64
+		if vma.mappable != nil {
+			newOffset = vseg.mappableRange().End
+		}
+		_, _, err := mm.createVMALocked(ctx, memmap.MMapOpts{
+			Length:          newSize - oldSize,
+			MappingIdentity: vma.id,
+			Mappable:        vma.mappable,
+			Offset:          newOffset,
+			Addr:            oldEnd,
+			Fixed:           true,
+			Perms:           vma.realPerms,
+			MaxPerms:        vma.maxPerms,
+			Private:         vma.private,
+			GrowsDown:       vma.growsDown,
+			Hint:            vma.hint,
+		})
+		if err == nil {
+			return oldAddr, nil
+		}
+		// In-place growth failed. In the MRemapMayMove case, fall through to
+		// moving below.
+		if opts.Move == MRemapNoMove {
+			return 0, err
+		}
+	}
+
+	// Handle moving, which is the only remaining case.
+
+	// Find a destination for the move.
+	var newAR usermem.AddrRange
+	switch opts.Move {
+	case MRemapMayMove:
+		newAddr, err := mm.findAvailableLocked(newSize, findAvailableOpts{})
+		if err != nil {
+			return 0, err
+		}
+		newAR, _ = newAddr.ToRange(newSize)
+
+	case MRemapMustMove:
+		newAddr := opts.NewAddr
+		if newAddr.RoundDown() != newAddr {
+			return 0, syserror.EINVAL
+		}
+		var ok bool
+		newAR, ok = newAddr.ToRange(newSize)
+		if !ok {
+			return 0, syserror.EINVAL
+		}
+		if (usermem.AddrRange{oldAddr, oldEnd}).Overlaps(newAR) {
+			return 0, syserror.EINVAL
+		}
+
+		// Unmap any mappings at the destination.
+		mm.unmapLocked(ctx, newAR)
+
+		// If the sizes specify shrinking, unmap everything between the new and
+		// old sizes at the source.
+		if newSize < oldSize {
+			oldNewEnd := oldAddr + usermem.Addr(newSize)
+			mm.unmapLocked(ctx, usermem.AddrRange{oldNewEnd, oldEnd})
+			oldEnd = oldNewEnd
+		}
+
+		// unmapLocked may have invalidated vseg; look it up again.
+		vseg = mm.vmas.FindSegment(oldAddr)
+	}
+
+	oldAR := usermem.AddrRange{oldAddr, oldEnd}
+
+	// In the MRemapMustMove case, these checks happen after unmapping:
+	// mm/mremap.c:mremap_to() => do_munmap(), vma_to_resize().
+
+	// Check that oldEnd maps to the same vma as oldAddr.
+	if vseg.End() < oldEnd {
+		return 0, syserror.EFAULT
+	}
+
+	// Check against RLIMIT_AS.
+	newUsageAS := mm.usageAS - uint64(oldAR.Length()) + uint64(newAR.Length())
+	if limitAS := limits.FromContext(ctx).Get(limits.AS).Cur; newUsageAS > limitAS {
+		return 0, syserror.ENOMEM
+	}
+
+	if vma := vseg.ValuePtr(); vma.mappable != nil {
+		// Check that offset+length does not overflow.
+		if vma.off+uint64(newAR.Length()) < vma.off {
+			return 0, syserror.EINVAL
+		}
+		// Inform the Mappable, if any, of the copied mapping.
+		if err := vma.mappable.CopyMapping(ctx, mm, oldAR, newAR, vseg.mappableOffsetAt(oldAR.Start)); err != nil {
+			return 0, err
+		}
+	}
+
+	// Remove the existing vma before inserting the new one to minimize
+	// iterator invalidation. We do this directly (instead of calling
+	// removeVMAsLocked) because:
+	//
+	// 1. We can't drop the reference on vma.id, which will be transferred to
+	// the new vma.
+	//
+	// 2. We can't call vma.mappable.RemoveMapping, because pmas are still at
+	// oldAR, so calling RemoveMapping could cause us to miss an invalidation
+	// overlapping oldAR.
+	//
+	// Call vseg.Value() (rather than vseg.ValuePtr()) first to make a copy of
+	// the vma.
+	vseg = mm.vmas.Isolate(vseg, oldAR)
+	vma := vseg.Value()
+	mm.vmas.Remove(vseg)
+
+	// Insert the new vma, transferring the reference on vma.id.
+	mm.vmas.Add(newAR, vma)
+
+	// Move pmas. This is technically optional for non-private pmas, which
+	// could just go through memmap.Mappable.Translate again, but it's required
+	// for private pmas.
+	mm.activeMu.Lock()
+	mm.movePMAsLocked(oldAR, newAR)
+	mm.activeMu.Unlock()
+
+	// Now that pmas have been moved to newAR, we can notify vma.mappable that
+	// oldAR is no longer mapped.
+	if vma.mappable != nil {
+		vma.mappable.RemoveMapping(ctx, mm, oldAR, vma.off)
+	}
+
+	return newAR.Start, nil
+}
+
+// MProtect implements the semantics of Linux's mprotect(2).
+func (mm *MemoryManager) MProtect(addr usermem.Addr, length uint64, realPerms usermem.AccessType, growsDown bool) error {
+	if addr.RoundDown() != addr {
+		return syserror.EINVAL
+	}
+	if length == 0 {
+		return nil
+	}
+	rlength, ok := usermem.Addr(length).RoundUp()
+	if !ok {
+		return syserror.ENOMEM
+	}
+	ar, ok := addr.ToRange(uint64(rlength))
+	if !ok {
+		return syserror.ENOMEM
+	}
+	effectivePerms := realPerms.Effective()
+
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	// Non-growsDown mprotect requires that all of ar is mapped, and stops at
+	// the first non-empty gap. growsDown mprotect requires that the first vma
+	// be growsDown, but does not require it to extend all the way to ar.Start;
+	// vmas after the first must be contiguous but need not be growsDown, like
+	// the non-growsDown case.
+	vseg := mm.vmas.LowerBoundSegment(ar.Start)
+	if !vseg.Ok() {
+		return syserror.ENOMEM
+	}
+	if growsDown {
+		if !vseg.ValuePtr().growsDown {
+			return syserror.EINVAL
+		}
+		if ar.End <= vseg.Start() {
+			return syserror.ENOMEM
+		}
+		ar.Start = vseg.Start()
+	} else {
+		if ar.Start < vseg.Start() {
+			return syserror.ENOMEM
+		}
+	}
+
+	mm.activeMu.Lock()
+	defer mm.activeMu.Unlock()
+	defer func() {
+		mm.vmas.MergeRange(ar)
+		mm.vmas.MergeAdjacent(ar)
+		mm.pmas.MergeRange(ar)
+		mm.pmas.MergeAdjacent(ar)
+	}()
+	pseg := mm.pmas.LowerBoundSegment(ar.Start)
+	var didUnmapAS bool
+	for {
+		// Check for permission validity before splitting vmas, for consistency
+		// with Linux.
+		if !vseg.ValuePtr().maxPerms.SupersetOf(effectivePerms) {
+			return syserror.EACCES
+		}
+		vseg = mm.vmas.Isolate(vseg, ar)
+
+		// Update vma permissions.
+		vma := vseg.ValuePtr()
+		vma.realPerms = realPerms
+		vma.effectivePerms = effectivePerms
+
+		// Propagate vma permission changes to pmas.
+		for pseg.Ok() && pseg.Start() < vseg.End() {
+			if pseg.Range().Overlaps(vseg.Range()) {
+				pseg = mm.pmas.Isolate(pseg, vseg.Range())
+				if !effectivePerms.SupersetOf(pseg.ValuePtr().vmaEffectivePerms) && !didUnmapAS {
+					// Unmap all of ar, not just vseg.Range(), to minimize host
+					// syscalls.
+					mm.unmapASLocked(ar)
+					didUnmapAS = true
+				}
+				pseg.ValuePtr().vmaEffectivePerms = effectivePerms
+			}
+			pseg = pseg.NextSegment()
+		}
+
+		// Continue to the next vma.
+		if ar.End <= vseg.End() {
+			return nil
+		}
+		vseg, _ = vseg.NextNonEmpty()
+		if !vseg.Ok() {
+			return syserror.ENOMEM
+		}
+	}
+}
+
+// BrkSetup sets mm's brk address to addr and its brk size to 0.
+func (mm *MemoryManager) BrkSetup(ctx context.Context, addr usermem.Addr) {
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+	// Unmap the existing brk.
+	if mm.brk.Length() != 0 {
+		mm.unmapLocked(ctx, mm.brk)
+	}
+	mm.brk = usermem.AddrRange{addr, addr}
+}
+
+// Brk implements the semantics of Linux's brk(2), except that it returns an
+// error on failure.
+func (mm *MemoryManager) Brk(ctx context.Context, addr usermem.Addr) (usermem.Addr, error) {
+	mm.mappingMu.Lock()
+	defer mm.mappingMu.Unlock()
+
+	if addr < mm.brk.Start {
+		return mm.brk.End, syserror.EINVAL
+	}
+
+	// TODO: This enforces RLIMIT_DATA, but is slightly more
+	// permissive than the usual data limit. In particular, this only
+	// limits the size of the heap; a true RLIMIT_DATA limits the size of
+	// heap + data + bss. The segment sizes need to be plumbed from the
+	// loader package to fully enforce RLIMIT_DATA.
+	if uint64(addr-mm.brk.Start) > limits.FromContext(ctx).Get(limits.Data).Cur {
+		return mm.brk.End, syserror.ENOMEM
+	}
+
+	oldbrkpg, _ := mm.brk.End.RoundUp()
+	newbrkpg, ok := addr.RoundUp()
+	if !ok {
+		return mm.brk.End, syserror.EFAULT
+	}
+
+	switch {
+	case newbrkpg < oldbrkpg:
+		mm.unmapLocked(ctx, usermem.AddrRange{newbrkpg, oldbrkpg})
+
+	case oldbrkpg < newbrkpg:
+		_, _, err := mm.createVMALocked(ctx, memmap.MMapOpts{
+			Length: uint64(newbrkpg - oldbrkpg),
+			Addr:   oldbrkpg,
+			Fixed:  true,
+			// Compare Linux's
+			// arch/x86/include/asm/page_types.h:VM_DATA_DEFAULT_FLAGS.
+			Perms:    usermem.ReadWrite,
+			MaxPerms: usermem.AnyAccess,
+			Private:  true,
+			Hint:     "[heap]",
+		})
+		if err != nil {
+			return mm.brk.End, err
+		}
+	}
+
+	mm.brk.End = addr
+	return addr, nil
+}
+
+// Decommit implements the semantics of Linux's madvise(MADV_DONTNEED).
+func (mm *MemoryManager) Decommit(addr usermem.Addr, length uint64) error {
+	ar, ok := addr.ToRange(length)
+	if !ok {
+		return syserror.EINVAL
+	}
+
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	mm.activeMu.Lock()
+	defer mm.activeMu.Unlock()
+
+	// Linux's mm/madvise.c:madvise_dontneed() => mm/memory.c:zap_page_range()
+	// is analogous to our mm.invalidateLocked(ar, true, true). We inline this
+	// here, with the special case that we synchronously decommit
+	// uniquely-owned (non-copy-on-write) pages for private anonymous vma,
+	// which is the common case for MADV_DONTNEED. Invalidating these pmas, and
+	// allowing them to be reallocated when touched again, increases pma
+	// fragmentation, which may significantly reduce performance for
+	// non-vectored I/O implementations. Also, decommitting synchronously
+	// ensures that Decommit immediately reduces host memory usage.
+	var didUnmapAS bool
+	pseg := mm.pmas.LowerBoundSegment(ar.Start)
+	vseg := mm.vmas.LowerBoundSegment(ar.Start)
+	mem := mm.p.Memory()
+	for pseg.Ok() && pseg.Start() < ar.End {
+		pma := pseg.ValuePtr()
+		if pma.private && !mm.isPMACopyOnWriteLocked(pseg) {
+			psegAR := pseg.Range().Intersect(ar)
+			vseg = vseg.seekNextLowerBound(psegAR.Start)
+			if checkInvariants {
+				if !vseg.Ok() {
+					panic(fmt.Sprintf("no vma after %#x", psegAR.Start))
+				}
+				if psegAR.Start < vseg.Start() {
+					panic(fmt.Sprintf("no vma in [%#x, %#x)", psegAR.Start, vseg.Start()))
+				}
+			}
+			if vseg.Range().IsSupersetOf(psegAR) && vseg.ValuePtr().mappable == nil {
+				if err := mem.Decommit(pseg.fileRangeOf(psegAR)); err == nil {
+					pseg = pseg.NextSegment()
+					continue
+				}
+				// If an error occurs, fall through to the general
+				// invalidation case below.
+			}
+		}
+		pseg = mm.pmas.Isolate(pseg, ar)
+		pma = pseg.ValuePtr()
+		if !didUnmapAS {
+			// Unmap all of ar, not just pseg.Range(), to minimize host
+			// syscalls. AddressSpace mappings must be removed before
+			// mm.decPrivateRef().
+			mm.unmapASLocked(ar)
+			didUnmapAS = true
+		}
+		if pma.private {
+			mm.decPrivateRef(pseg.fileRange())
+		}
+		pma.file.DecRef(pseg.fileRange())
+		mm.removeRSSLocked(pseg.Range())
+
+		pseg = mm.pmas.Remove(pseg).NextSegment()
+	}
+
+	// "If there are some parts of the specified address space that are not
+	// mapped, the Linux version of madvise() ignores them and applies the call
+	// to the rest (but returns ENOMEM from the system call, as it should)." -
+	// madvise(2)
+	if mm.vmas.SpanRange(ar) != ar.Length() {
+		return syserror.ENOMEM
+	}
+	return nil
+}
+
+// Sync implements the semantics of Linux's msync(MS_SYNC).
+func (mm *MemoryManager) Sync(ctx context.Context, addr usermem.Addr, length uint64) error {
+	ar, ok := addr.ToRange(length)
+	if !ok {
+		return syserror.ENOMEM
+	}
+
+	mm.mappingMu.RLock()
+	// Can't defer mm.mappingMu.RUnlock(); see below.
+	vseg := mm.vmas.LowerBoundSegment(ar.Start)
+	if !vseg.Ok() {
+		mm.mappingMu.RUnlock()
+		return syserror.ENOMEM
+	}
+	var unmapped bool
+	lastEnd := ar.Start
+	for {
+		if !vseg.Ok() {
+			mm.mappingMu.RUnlock()
+			unmapped = true
+			break
+		}
+		if lastEnd < vseg.Start() {
+			unmapped = true
+		}
+		lastEnd = vseg.End()
+		vma := vseg.ValuePtr()
+		// It's only possible to have dirtied the Mappable through a shared
+		// mapping. Don't check if the mapping is writable, because mprotect
+		// may have changed this, and also because Linux doesn't.
+		if id := vma.id; id != nil && vma.mappable != nil && !vma.private {
+			// We can't call memmap.MappingIdentity.Msync while holding
+			// mm.mappingMu since it may take fs locks that precede it in the
+			// lock order.
+			id.IncRef()
+			mr := vseg.mappableRangeOf(vseg.Range().Intersect(ar))
+			mm.mappingMu.RUnlock()
+			err := id.Msync(ctx, mr)
+			id.DecRef()
+			if err != nil {
+				return err
+			}
+			if lastEnd >= ar.End {
+				break
+			}
+			mm.mappingMu.RLock()
+			vseg = mm.vmas.LowerBoundSegment(lastEnd)
+		} else {
+			if lastEnd >= ar.End {
+				mm.mappingMu.RUnlock()
+				break
+			}
+			vseg = vseg.NextSegment()
+		}
+	}
+
+	if unmapped {
+		return syserror.ENOMEM
+	}
+	return nil
+}
+
+// VirtualMemorySize returns the combined length in bytes of all mappings in
+// mm.
+func (mm *MemoryManager) VirtualMemorySize() uint64 {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	return uint64(mm.usageAS)
+}
+
+// VirtualMemorySizeRange returns the combined length in bytes of all mappings
+// in ar in mm.
+func (mm *MemoryManager) VirtualMemorySizeRange(ar usermem.AddrRange) uint64 {
+	mm.mappingMu.RLock()
+	defer mm.mappingMu.RUnlock()
+	return uint64(mm.vmas.SpanRange(ar))
+}
+
+// ResidentSetSize returns the value advertised as mm's RSS in bytes.
+func (mm *MemoryManager) ResidentSetSize() uint64 {
+	mm.activeMu.RLock()
+	defer mm.activeMu.RUnlock()
+	return uint64(mm.curRSS)
+}
+
+// MaxResidentSetSize returns the value advertised as mm's max RSS in bytes.
+func (mm *MemoryManager) MaxResidentSetSize() uint64 {
+	mm.activeMu.RLock()
+	defer mm.activeMu.RUnlock()
+	return uint64(mm.maxRSS)
+}
diff --git a/pkg/sentry/mm/vma.go b/pkg/sentry/mm/vma.go
new file mode 100644
index 000000000..b6af48cb7
--- /dev/null
+++ b/pkg/sentry/mm/vma.go
@@ -0,0 +1,476 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package mm
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Preconditions: mm.mappingMu must be locked for writing. opts must be valid
+// as defined by the checks in MMap.
+func (mm *MemoryManager) createVMALocked(ctx context.Context, opts memmap.MMapOpts) (vmaIterator, usermem.AddrRange, error) {
+	if opts.MaxPerms != opts.MaxPerms.Effective() {
+		panic(fmt.Sprintf("Non-effective MaxPerms %s cannot be enforced", opts.MaxPerms))
+	}
+
+	// Find a useable range.
+	addr, err := mm.findAvailableLocked(opts.Length, findAvailableOpts{
+		Addr:  opts.Addr,
+		Fixed: opts.Fixed,
+		Unmap: opts.Unmap,
+	})
+	if err != nil {
+		return vmaIterator{}, usermem.AddrRange{}, err
+	}
+	ar, _ := addr.ToRange(opts.Length)
+
+	// Check against RLIMIT_AS.
+	newUsageAS := mm.usageAS + opts.Length
+	if opts.Unmap {
+		newUsageAS -= uint64(mm.vmas.SpanRange(ar))
+	}
+	if limitAS := limits.FromContext(ctx).Get(limits.AS).Cur; newUsageAS > limitAS {
+		return vmaIterator{}, usermem.AddrRange{}, syserror.ENOMEM
+	}
+
+	// Remove overwritten mappings. This ordering is consistent with Linux:
+	// compare Linux's mm/mmap.c:mmap_region() => do_munmap(),
+	// file->f_op->mmap().
+	var vgap vmaGapIterator
+	if opts.Unmap {
+		vgap = mm.unmapLocked(ctx, ar)
+	} else {
+		vgap = mm.vmas.FindGap(ar.Start)
+	}
+
+	// Inform the Mappable, if any, of the new mapping.
+	if opts.Mappable != nil {
+		if err := opts.Mappable.AddMapping(ctx, mm, ar, opts.Offset); err != nil {
+			return vmaIterator{}, usermem.AddrRange{}, err
+		}
+	}
+
+	// Take a reference on opts.MappingIdentity before inserting the vma since
+	// vma merging can drop the reference.
+	if opts.MappingIdentity != nil {
+		opts.MappingIdentity.IncRef()
+	}
+
+	// Finally insert the vma.
+	vseg := mm.vmas.Insert(vgap, ar, vma{
+		mappable:       opts.Mappable,
+		off:            opts.Offset,
+		realPerms:      opts.Perms,
+		effectivePerms: opts.Perms.Effective(),
+		maxPerms:       opts.MaxPerms,
+		private:        opts.Private,
+		growsDown:      opts.GrowsDown,
+		id:             opts.MappingIdentity,
+		hint:           opts.Hint,
+	})
+	mm.usageAS += opts.Length
+
+	return vseg, ar, nil
+}
+
+type findAvailableOpts struct {
+	// Addr is a suggested address. Addr must be page-aligned.
+	Addr usermem.Addr
+
+	// Fixed is true if only the suggested address is acceptable.
+	Fixed bool
+
+	// Unmap is true if existing vmas and guard pages may exist in the returned
+	// range.
+	Unmap bool
+}
+
+// findAvailableLocked finds an allocatable range.
+//
+// Preconditions: mm.mappingMu must be locked.
+func (mm *MemoryManager) findAvailableLocked(length uint64, opts findAvailableOpts) (usermem.Addr, error) {
+	// Does the provided suggestion work?
+	if ar, ok := opts.Addr.ToRange(length); ok {
+		if mm.applicationAddrRange().IsSupersetOf(ar) {
+			if opts.Unmap {
+				return ar.Start, nil
+			}
+			// Check for the presence of an existing vma or guard page.
+			if vgap := mm.vmas.FindGap(ar.Start); vgap.Ok() && vgap.availableRange().IsSupersetOf(ar) {
+				return ar.Start, nil
+			}
+		}
+	}
+
+	// Fixed mappings accept only the requested address.
+	if opts.Fixed {
+		return 0, syserror.ENOMEM
+	}
+
+	// Prefer hugepage alignment if a hugepage or more is requested.
+	alignment := uint64(usermem.PageSize)
+	if length >= usermem.HugePageSize {
+		alignment = usermem.HugePageSize
+	}
+
+	if mm.layout.DefaultDirection == arch.MmapBottomUp {
+		return mm.findLowestAvailableLocked(length, alignment, usermem.AddrRange{mm.layout.BottomUpBase, mm.layout.MaxAddr})
+	}
+	return mm.findHighestAvailableLocked(length, alignment, usermem.AddrRange{mm.layout.MinAddr, mm.layout.TopDownBase})
+}
+
+func (mm *MemoryManager) applicationAddrRange() usermem.AddrRange {
+	return usermem.AddrRange{mm.layout.MinAddr, mm.layout.MaxAddr}
+}
+
+// Preconditions: mm.mappingMu must be locked.
+func (mm *MemoryManager) findLowestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) {
+	for gap := mm.vmas.LowerBoundGap(bounds.Start); gap.Ok() && gap.Start() < bounds.End; gap = gap.NextGap() {
+		if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length {
+			// Can we shift up to match the alignment?
+			if offset := uint64(gr.Start) % alignment; offset != 0 {
+				if uint64(gr.Length()) >= length+alignment-offset {
+					// Yes, we're aligned.
+					return gr.Start + usermem.Addr(alignment-offset), nil
+				}
+			}
+
+			// Either aligned perfectly, or can't align it.
+			return gr.Start, nil
+		}
+	}
+	return 0, syserror.ENOMEM
+}
+
+// Preconditions: mm.mappingMu must be locked.
+func (mm *MemoryManager) findHighestAvailableLocked(length, alignment uint64, bounds usermem.AddrRange) (usermem.Addr, error) {
+	for gap := mm.vmas.UpperBoundGap(bounds.End); gap.Ok() && gap.End() > bounds.Start; gap = gap.PrevGap() {
+		if gr := gap.availableRange().Intersect(bounds); uint64(gr.Length()) >= length {
+			// Can we shift down to match the alignment?
+			start := gr.End - usermem.Addr(length)
+			if offset := uint64(start) % alignment; offset != 0 {
+				if gr.Start <= start-usermem.Addr(offset) {
+					// Yes, we're aligned.
+					return start - usermem.Addr(offset), nil
+				}
+			}
+
+			// Either aligned perfectly, or can't align it.
+			return start, nil
+		}
+	}
+	return 0, syserror.ENOMEM
+}
+
+// getVMAsLocked ensures that vmas exist for all addresses in ar, and support
+// access of type (at, ignorePermissions). It returns:
+//
+// - An iterator to the vma containing ar.Start. If no vma contains ar.Start,
+// the iterator is unspecified.
+//
+// - An iterator to the gap after the last vma containing an address in ar. If
+// vmas exist for no addresses in ar, the iterator is to a gap that begins
+// before ar.Start.
+//
+// - An error that is non-nil if vmas exist for only a subset of ar.
+//
+// Preconditions: mm.mappingMu must be locked for reading; it may be
+// temporarily unlocked. ar.Length() != 0.
+func (mm *MemoryManager) getVMAsLocked(ctx context.Context, ar usermem.AddrRange, at usermem.AccessType, ignorePermissions bool) (vmaIterator, vmaGapIterator, error) {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	// Inline mm.vmas.LowerBoundSegment so that we have the preceding gap if
+	// !vbegin.Ok().
+	vbegin, vgap := mm.vmas.Find(ar.Start)
+	if !vbegin.Ok() {
+		vbegin = vgap.NextSegment()
+		// vseg.Ok() is checked before entering the following loop.
+	} else {
+		vgap = vbegin.PrevGap()
+	}
+
+	addr := ar.Start
+	vseg := vbegin
+	for vseg.Ok() {
+		// Loop invariants: vgap = vseg.PrevGap(); addr < vseg.End().
+		vma := vseg.ValuePtr()
+		if addr < vseg.Start() {
+			// TODO: Implement vma.growsDown here.
+			return vbegin, vgap, syserror.EFAULT
+		}
+
+		perms := vma.effectivePerms
+		if ignorePermissions {
+			perms = vma.maxPerms
+		}
+		if !perms.SupersetOf(at) {
+			return vbegin, vgap, syserror.EPERM
+		}
+
+		addr = vseg.End()
+		vgap = vseg.NextGap()
+		if addr >= ar.End {
+			return vbegin, vgap, nil
+		}
+		vseg = vgap.NextSegment()
+	}
+
+	// Ran out of vmas before ar.End.
+	return vbegin, vgap, syserror.EFAULT
+}
+
+// getVecVMAsLocked ensures that vmas exist for all addresses in ars, and
+// support access to type of (at, ignorePermissions). It retuns the subset of
+// ars for which vmas exist. If this is not equal to ars, it returns a non-nil
+// error explaining why.
+//
+// Preconditions: mm.mappingMu must be locked for reading; it may be
+// temporarily unlocked.
+//
+// Postconditions: ars is not mutated.
+func (mm *MemoryManager) getVecVMAsLocked(ctx context.Context, ars usermem.AddrRangeSeq, at usermem.AccessType, ignorePermissions bool) (usermem.AddrRangeSeq, error) {
+	for arsit := ars; !arsit.IsEmpty(); arsit = arsit.Tail() {
+		ar := arsit.Head()
+		if ar.Length() == 0 {
+			continue
+		}
+		if _, vend, err := mm.getVMAsLocked(ctx, ar, at, ignorePermissions); err != nil {
+			return truncatedAddrRangeSeq(ars, arsit, vend.Start()), err
+		}
+	}
+	return ars, nil
+}
+
+// vma extension will not shrink the number of unmapped bytes between the start
+// of a growsDown vma and the end of its predecessor non-growsDown vma below
+// guardBytes.
+//
+// guardBytes is equivalent to Linux's stack_guard_gap after upstream
+// 1be7107fbe18 "mm: larger stack guard gap, between vmas".
+const guardBytes = 256 * usermem.PageSize
+
+// unmapLocked unmaps all addresses in ar and returns the resulting gap in
+// mm.vmas.
+//
+// Preconditions: mm.mappingMu must be locked for writing. ar.Length() != 0.
+// ar must be page-aligned.
+func (mm *MemoryManager) unmapLocked(ctx context.Context, ar usermem.AddrRange) vmaGapIterator {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	// AddressSpace mappings and pmas must be invalidated before
+	// mm.removeVMAsLocked() => memmap.Mappable.RemoveMapping().
+	mm.Invalidate(ar, memmap.InvalidateOpts{InvalidatePrivate: true})
+	return mm.removeVMAsLocked(ctx, ar)
+}
+
+// removeVMAsLocked removes vmas for addresses in ar and returns the resulting
+// gap in mm.vmas. It does not remove pmas or AddressSpace mappings; clients
+// must do so before calling removeVMAsLocked.
+//
+// Preconditions: mm.mappingMu must be locked for writing. ar.Length() != 0. ar
+// must be page-aligned.
+func (mm *MemoryManager) removeVMAsLocked(ctx context.Context, ar usermem.AddrRange) vmaGapIterator {
+	if checkInvariants {
+		if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+	}
+
+	vseg, vgap := mm.vmas.Find(ar.Start)
+	if vgap.Ok() {
+		vseg = vgap.NextSegment()
+	}
+	for vseg.Ok() && vseg.Start() < ar.End {
+		vseg = mm.vmas.Isolate(vseg, ar)
+		vmaAR := vseg.Range()
+		vma := vseg.ValuePtr()
+		if vma.mappable != nil {
+			vma.mappable.RemoveMapping(ctx, mm, vmaAR, vma.off)
+		}
+		if vma.id != nil {
+			vma.id.DecRef()
+		}
+		mm.usageAS -= uint64(vmaAR.Length())
+		vgap = mm.vmas.Remove(vseg)
+		vseg = vgap.NextSegment()
+	}
+	return vgap
+}
+
+// vmaSetFunctions implements segment.Functions for vmaSet.
+type vmaSetFunctions struct{}
+
+func (vmaSetFunctions) MinKey() usermem.Addr {
+	return 0
+}
+
+func (vmaSetFunctions) MaxKey() usermem.Addr {
+	return ^usermem.Addr(0)
+}
+
+func (vmaSetFunctions) ClearValue(vma *vma) {
+	vma.mappable = nil
+	vma.id = nil
+	vma.hint = ""
+}
+
+func (vmaSetFunctions) Merge(ar1 usermem.AddrRange, vma1 vma, ar2 usermem.AddrRange, vma2 vma) (vma, bool) {
+	if vma1.mappable != vma2.mappable ||
+		(vma1.mappable != nil && vma1.off+uint64(ar1.Length()) != vma2.off) ||
+		vma1.realPerms != vma2.realPerms ||
+		vma1.maxPerms != vma2.maxPerms ||
+		vma1.private != vma2.private ||
+		vma1.growsDown != vma2.growsDown ||
+		vma1.id != vma2.id ||
+		vma1.hint != vma2.hint {
+		return vma{}, false
+	}
+
+	if vma2.id != nil {
+		vma2.id.DecRef()
+	}
+	return vma1, true
+}
+
+func (vmaSetFunctions) Split(ar usermem.AddrRange, v vma, split usermem.Addr) (vma, vma) {
+	v2 := v
+	if v2.mappable != nil {
+		v2.off += uint64(split - ar.Start)
+	}
+	if v2.id != nil {
+		v2.id.IncRef()
+	}
+	return v, v2
+}
+
+// Preconditions: vseg.ValuePtr().mappable != nil. vseg.Range().Contains(addr).
+func (vseg vmaIterator) mappableOffsetAt(addr usermem.Addr) uint64 {
+	if checkInvariants {
+		if !vseg.Ok() {
+			panic("terminal vma iterator")
+		}
+		if vseg.ValuePtr().mappable == nil {
+			panic("Mappable offset is meaningless for anonymous vma")
+		}
+		if !vseg.Range().Contains(addr) {
+			panic(fmt.Sprintf("addr %v out of bounds %v", addr, vseg.Range()))
+		}
+	}
+
+	vma := vseg.ValuePtr()
+	vstart := vseg.Start()
+	return vma.off + uint64(addr-vstart)
+}
+
+// Preconditions: vseg.ValuePtr().mappable != nil.
+func (vseg vmaIterator) mappableRange() memmap.MappableRange {
+	return vseg.mappableRangeOf(vseg.Range())
+}
+
+// Preconditions: vseg.ValuePtr().mappable != nil.
+// vseg.Range().IsSupersetOf(ar). ar.Length() != 0.
+func (vseg vmaIterator) mappableRangeOf(ar usermem.AddrRange) memmap.MappableRange {
+	if checkInvariants {
+		if !vseg.Ok() {
+			panic("terminal vma iterator")
+		}
+		if vseg.ValuePtr().mappable == nil {
+			panic("MappableRange is meaningless for anonymous vma")
+		}
+		if !ar.WellFormed() || ar.Length() <= 0 {
+			panic(fmt.Sprintf("invalid ar: %v", ar))
+		}
+		if !vseg.Range().IsSupersetOf(ar) {
+			panic(fmt.Sprintf("ar %v out of bounds %v", ar, vseg.Range()))
+		}
+	}
+
+	vma := vseg.ValuePtr()
+	vstart := vseg.Start()
+	return memmap.MappableRange{vma.off + uint64(ar.Start-vstart), vma.off + uint64(ar.End-vstart)}
+}
+
+// Preconditions: vseg.ValuePtr().mappable != nil.
+// vseg.mappableRange().IsSupersetOf(mr). mr.Length() != 0.
+func (vseg vmaIterator) addrRangeOf(mr memmap.MappableRange) usermem.AddrRange {
+	if checkInvariants {
+		if !vseg.Ok() {
+			panic("terminal vma iterator")
+		}
+		if vseg.ValuePtr().mappable == nil {
+			panic("MappableRange is meaningless for anonymous vma")
+		}
+		if !mr.WellFormed() || mr.Length() <= 0 {
+			panic(fmt.Sprintf("invalid mr: %v", mr))
+		}
+		if !vseg.mappableRange().IsSupersetOf(mr) {
+			panic(fmt.Sprintf("mr %v out of bounds %v", mr, vseg.mappableRange()))
+		}
+	}
+
+	vma := vseg.ValuePtr()
+	vstart := vseg.Start()
+	return usermem.AddrRange{vstart + usermem.Addr(mr.Start-vma.off), vstart + usermem.Addr(mr.End-vma.off)}
+}
+
+// seekNextLowerBound returns mm.vmas.LowerBoundSegment(addr), but does so by
+// scanning linearly forward from vseg.
+//
+// Preconditions: mm.mappingMu must be locked. addr >= vseg.Start().
+func (vseg vmaIterator) seekNextLowerBound(addr usermem.Addr) vmaIterator {
+	if checkInvariants {
+		if !vseg.Ok() {
+			panic("terminal vma iterator")
+		}
+		if addr < vseg.Start() {
+			panic(fmt.Sprintf("can't seek forward to %#x from %#x", addr, vseg.Start()))
+		}
+	}
+	for vseg.Ok() && addr >= vseg.End() {
+		vseg = vseg.NextSegment()
+	}
+	return vseg
+}
+
+// availableRange returns the subset of vgap.Range() in which new vmas may be
+// created without MMapOpts.Unmap == true.
+func (vgap vmaGapIterator) availableRange() usermem.AddrRange {
+	ar := vgap.Range()
+	next := vgap.NextSegment()
+	if !next.Ok() || !next.ValuePtr().growsDown {
+		return ar
+	}
+	// Exclude guard pages.
+	if ar.Length() < guardBytes {
+		return usermem.AddrRange{ar.Start, ar.Start}
+	}
+	ar.End -= guardBytes
+	return ar
+}
diff --git a/pkg/sentry/platform/BUILD b/pkg/sentry/platform/BUILD
new file mode 100644
index 000000000..d5be81f8d
--- /dev/null
+++ b/pkg/sentry/platform/BUILD
@@ -0,0 +1,51 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "platform_state",
+    srcs = [
+        "file_range.go",
+    ],
+    out = "platform_state.go",
+    package = "platform",
+)
+
+go_template_instance(
+    name = "file_range",
+    out = "file_range.go",
+    package = "platform",
+    prefix = "File",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "uint64",
+    },
+)
+
+go_library(
+    name = "platform",
+    srcs = [
+        "context.go",
+        "file_range.go",
+        "mmap_min_addr.go",
+        "platform.go",
+        "platform_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/atomicbitops",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/platform/safecopy",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/platform/context.go b/pkg/sentry/platform/context.go
new file mode 100644
index 000000000..0d200a5e2
--- /dev/null
+++ b/pkg/sentry/platform/context.go
@@ -0,0 +1,36 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package platform
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the auth package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxPlatform is a Context.Value key for a Platform.
+	CtxPlatform contextID = iota
+)
+
+// FromContext returns the Platform that is used to execute ctx's application
+// code, or nil if no such Platform exists.
+func FromContext(ctx context.Context) Platform {
+	if v := ctx.Value(CtxPlatform); v != nil {
+		return v.(Platform)
+	}
+	return nil
+}
diff --git a/pkg/sentry/platform/filemem/BUILD b/pkg/sentry/platform/filemem/BUILD
new file mode 100644
index 000000000..3c4d5b0b6
--- /dev/null
+++ b/pkg/sentry/platform/filemem/BUILD
@@ -0,0 +1,69 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "filemem_autogen_state",
+    srcs = [
+        "filemem.go",
+        "filemem_state.go",
+        "usage_set.go",
+    ],
+    out = "filemem_autogen_state.go",
+    package = "filemem",
+)
+
+go_template_instance(
+    name = "usage_set",
+    out = "usage_set.go",
+    consts = {
+        "minDegree": "10",
+    },
+    imports = {
+        "platform": "gvisor.googlesource.com/gvisor/pkg/sentry/platform",
+    },
+    package = "filemem",
+    prefix = "usage",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "uint64",
+        "Range": "platform.FileRange",
+        "Value": "usageInfo",
+        "Functions": "usageSetFunctions",
+    },
+)
+
+go_library(
+    name = "filemem",
+    srcs = [
+        "filemem.go",
+        "filemem_autogen_state.go",
+        "filemem_state.go",
+        "filemem_unsafe.go",
+        "usage_set.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/filemem",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/memutil",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "filemem_test",
+    size = "small",
+    srcs = ["filemem_test.go"],
+    embed = [":filemem"],
+    deps = ["//pkg/sentry/usermem"],
+)
diff --git a/pkg/sentry/platform/filemem/filemem.go b/pkg/sentry/platform/filemem/filemem.go
new file mode 100644
index 000000000..d79c3c7f1
--- /dev/null
+++ b/pkg/sentry/platform/filemem/filemem.go
@@ -0,0 +1,838 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package filemem provides a reusable implementation of platform.Memory.
+//
+// It enables memory to be sourced from a memfd file.
+//
+// Lock order:
+//
+// filemem.FileMem.mu
+//   filemem.FileMem.mappingsMu
+package filemem
+
+import (
+	"fmt"
+	"math"
+	"os"
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// FileMem is a platform.Memory that allocates from a host file that it owns.
+type FileMem struct {
+	// Filemem models the backing file as follows:
+	//
+	// Each page in the file can be committed or uncommitted. A page is
+	// committed if the host kernel is spending resources to store its contents
+	// and uncommitted otherwise. This definition includes pages that the host
+	// kernel has swapped; this is intentional, to ensure that accounting does
+	// not change even if host kernel swapping behavior changes, and that
+	// memory used by pseudo-swap mechanisms like zswap is still accounted.
+	//
+	// The initial contents of uncommitted pages are implicitly zero bytes. A
+	// read or write to the contents of an uncommitted page causes it to be
+	// committed. This is the only event that can cause a uncommitted page to
+	// be committed.
+	//
+	// fallocate(FALLOC_FL_PUNCH_HOLE) (FileMem.Decommit) causes committed
+	// pages to be uncommitted. This is the only event that can cause a
+	// committed page to be uncommitted.
+	//
+	// Filemem's accounting is based on identifying the set of committed pages.
+	// Since filemem does not have direct access to the MMU, tracking reads and
+	// writes to uncommitted pages to detect commitment would introduce
+	// additional page faults, which would be prohibitively expensive. Instead,
+	// filemem queries the host kernel to determine which pages are committed.
+
+	// file is the backing memory file. The file pointer is immutable.
+	file *os.File
+
+	mu sync.Mutex
+
+	// usage maps each page in the file to metadata for that page. Pages for
+	// which no segment exists in usage are both unallocated (not in use) and
+	// uncommitted.
+	//
+	// Since usage stores usageInfo objects by value, clients should usually
+	// use usageIterator.ValuePtr() instead of usageIterator.Value() to get a
+	// pointer to the usageInfo rather than a copy.
+	//
+	// usage must be kept maximally merged (that is, there should never be two
+	// adjacent segments with the same values). At least markReclaimed depends
+	// on this property.
+	//
+	// usage is protected by mu.
+	usage usageSet
+
+	// The UpdateUsage function scans all segments with knownCommitted set
+	// to false, sees which pages are committed and creates corresponding
+	// segments with knownCommitted set to true.
+	//
+	// In order to avoid unnecessary scans, usageExpected tracks the total
+	// file blocks expected. This is used to elide the scan when this
+	// matches the underlying file blocks.
+	//
+	// To track swapped pages, usageSwapped tracks the discrepency between
+	// what is observed in core and what is reported by the file. When
+	// usageSwapped is non-zero, a sweep will be performed at least every
+	// second. The start of the last sweep is recorded in usageLast.
+	//
+	// All usage attributes are all protected by mu.
+	usageExpected uint64
+	usageSwapped  uint64
+	usageLast     time.Time
+
+	// fileSize is the size of the backing memory file in bytes. fileSize is
+	// always a power-of-two multiple of chunkSize.
+	//
+	// fileSize is protected by mu.
+	fileSize int64
+
+	// destroyed is set by Destroy to instruct the reclaimer goroutine to
+	// release resources and exit. destroyed is protected by mu.
+	destroyed bool
+
+	// reclaimable is true if usage may contain reclaimable pages. reclaimable
+	// is protected by mu.
+	reclaimable bool
+
+	// reclaimCond is signaled (with mu locked) when reclaimable or destroyed
+	// transitions from false to true.
+	reclaimCond sync.Cond
+
+	// Filemem pages are mapped into the local address space on the granularity
+	// of large pieces called chunks. mappings is a []uintptr that stores, for
+	// each chunk, the start address of a mapping of that chunk in the current
+	// process' address space, or 0 if no such mapping exists. Once a chunk is
+	// mapped, it is never remapped or unmapped until the filemem is destroyed.
+	//
+	// Mutating the mappings slice or its contents requires both holding
+	// mappingsMu and using atomic memory operations. (The slice is mutated
+	// whenever the file is expanded. Per the above, the only permitted
+	// mutation of the slice's contents is the assignment of a mapping to a
+	// chunk that was previously unmapped.) Reading the slice or its contents
+	// only requires *either* holding mappingsMu or using atomic memory
+	// operations. This allows FileMem.AccessPhysical to avoid locking in the
+	// common case where chunk mappings already exist.
+
+	mappingsMu sync.Mutex
+	mappings   atomic.Value
+}
+
+// usage tracks usage information.
+type usageInfo struct {
+	// kind is the usage kind.
+	kind usage.MemoryKind
+
+	// knownCommitted indicates whether this region is known to be
+	// committed. If this is false, then the region may or may not have
+	// been touched. If it is true however, then mincore (below) has
+	// indicated that the page is present at least once.
+	knownCommitted bool
+
+	refs uint64
+}
+
+func (u *usageInfo) incRef() {
+	u.refs++
+}
+
+func (u *usageInfo) decRef() {
+	if u.refs == 0 {
+		panic("DecRef at 0 refs!")
+	}
+	u.refs--
+}
+
+const (
+	chunkShift = 24
+	chunkSize  = 1 << chunkShift // 16 MB
+	chunkMask  = chunkSize - 1
+
+	initialSize = chunkSize
+)
+
+// newFromFile creates a FileMem backed by the given file.
+func newFromFile(file *os.File) (*FileMem, error) {
+	if err := file.Truncate(initialSize); err != nil {
+		return nil, err
+	}
+	f := &FileMem{
+		fileSize: initialSize,
+		file:     file,
+	}
+	f.reclaimCond.L = &f.mu
+	f.mappings.Store(make([]uintptr, initialSize/chunkSize))
+	go f.runReclaim() // S/R-SAFE: f.mu
+
+	// The Linux kernel contains an optional feature called "Integrity
+	// Measurement Architecture" (IMA). If IMA is enabled, it will checksum
+	// binaries the first time they are mapped PROT_EXEC. This is bad news for
+	// executable pages mapped from FileMem, which can grow to terabytes in
+	// (sparse) size. If IMA attempts to checksum a file that large, it will
+	// allocate all of the sparse pages and quickly exhaust all memory.
+	//
+	// Work around IMA by immediately creating a temporary PROT_EXEC mapping,
+	// while FileMem is still small. IMA will ignore any future mappings.
+	m, _, errno := syscall.Syscall6(
+		syscall.SYS_MMAP,
+		0,
+		usermem.PageSize,
+		syscall.PROT_EXEC,
+		syscall.MAP_SHARED,
+		f.file.Fd(),
+		0)
+	if errno != 0 {
+		// This isn't fatal to filemem (IMA may not even be in use). Log the
+		// error, but don't return it.
+		log.Warningf("Failed to pre-map FileMem PROT_EXEC: %v", errno)
+	} else {
+		syscall.Syscall(
+			syscall.SYS_MUNMAP,
+			m,
+			usermem.PageSize,
+			0)
+	}
+
+	return f, nil
+}
+
+// New creates a FileMem backed by a memfd file.
+func New(name string) (*FileMem, error) {
+	fd, err := memutil.CreateMemFD(name, 0)
+	if err != nil {
+		return nil, err
+	}
+	return newFromFile(os.NewFile(uintptr(fd), name))
+}
+
+// Destroy implements platform.Memory.Destroy.
+func (f *FileMem) Destroy() {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	f.destroyed = true
+	f.reclaimCond.Signal()
+}
+
+// Allocate implements platform.Memory.Allocate.
+func (f *FileMem) Allocate(length uint64, kind usage.MemoryKind) (platform.FileRange, error) {
+	if length == 0 || length%usermem.PageSize != 0 {
+		panic(fmt.Sprintf("invalid allocation length: %#x", length))
+	}
+
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	// Align hugepage-and-larger allocations on hugepage boundaries to try
+	// to take advantage of hugetmpfs.
+	alignment := uint64(usermem.PageSize)
+	if length >= usermem.HugePageSize {
+		alignment = usermem.HugePageSize
+	}
+
+	start := findUnallocatedRange(&f.usage, length, alignment)
+	end := start + length
+	// File offsets are int64s. Since length must be strictly positive, end
+	// cannot legitimately be 0.
+	if end < start || int64(end) <= 0 {
+		return platform.FileRange{}, syserror.ENOMEM
+	}
+
+	// Expand the file if needed. Double the file size on each expansion;
+	// uncommitted pages have effectively no cost.
+	fileSize := f.fileSize
+	for int64(end) > fileSize {
+		if fileSize >= 2*fileSize {
+			// fileSize overflow.
+			return platform.FileRange{}, syserror.ENOMEM
+		}
+		fileSize *= 2
+	}
+	if fileSize > f.fileSize {
+		if err := f.file.Truncate(fileSize); err != nil {
+			return platform.FileRange{}, err
+		}
+		f.fileSize = fileSize
+		f.mappingsMu.Lock()
+		oldMappings := f.mappings.Load().([]uintptr)
+		newMappings := make([]uintptr, fileSize>>chunkShift)
+		copy(newMappings, oldMappings)
+		f.mappings.Store(newMappings)
+		f.mappingsMu.Unlock()
+	}
+
+	// Mark selected pages as in use.
+	fr := platform.FileRange{start, end}
+	if !f.usage.Add(fr, usageInfo{
+		kind: kind,
+		refs: 1,
+	}) {
+		panic(fmt.Sprintf("allocating %v: failed to insert into f.usage:\n%v", fr, &f.usage))
+	}
+	return fr, nil
+}
+
+func findUnallocatedRange(usage *usageSet, length, alignment uint64) uint64 {
+	alignMask := alignment - 1
+	var start uint64
+	for seg := usage.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		r := seg.Range()
+		if start >= r.End {
+			// start was rounded up to an alignment boundary from the end
+			// of a previous segment.
+			continue
+		}
+		// This segment represents allocated or reclaimable pages; only the
+		// range from start to the segment's beginning is allocatable, and the
+		// next allocatable range begins after the segment.
+		if r.Start > start && r.Start-start >= length {
+			break
+		}
+		start = (r.End + alignMask) &^ alignMask
+	}
+	return start
+}
+
+// fallocate(2) modes, defined in Linux's include/uapi/linux/falloc.h.
+const (
+	_FALLOC_FL_KEEP_SIZE  = 1
+	_FALLOC_FL_PUNCH_HOLE = 2
+)
+
+// Decommit implements platform.Memory.Decommit.
+func (f *FileMem) Decommit(fr platform.FileRange) error {
+	if !fr.WellFormed() || fr.Length() == 0 || fr.Start%usermem.PageSize != 0 || fr.End%usermem.PageSize != 0 {
+		panic(fmt.Sprintf("invalid range: %v", fr))
+	}
+
+	// "After a successful call, subsequent reads from this range will
+	// return zeroes. The FALLOC_FL_PUNCH_HOLE flag must be ORed with
+	// FALLOC_FL_KEEP_SIZE in mode ..." - fallocate(2)
+	err := syscall.Fallocate(
+		int(f.file.Fd()),
+		_FALLOC_FL_PUNCH_HOLE|_FALLOC_FL_KEEP_SIZE,
+		int64(fr.Start),
+		int64(fr.Length()))
+	if err != nil {
+		return err
+	}
+	f.markDecommitted(fr)
+	return nil
+}
+
+func (f *FileMem) markDecommitted(fr platform.FileRange) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	// Since we're changing the knownCommitted attribute, we need to merge
+	// across the entire range to ensure that the usage tree is minimal.
+	gap := f.usage.ApplyContiguous(fr, func(seg usageIterator) {
+		val := seg.ValuePtr()
+		if val.knownCommitted {
+			// Drop the usageExpected appropriately.
+			amount := seg.Range().Length()
+			usage.MemoryAccounting.Dec(amount, val.kind)
+			f.usageExpected -= amount
+			val.knownCommitted = false
+		}
+	})
+	if gap.Ok() {
+		panic(fmt.Sprintf("Decommit(%v): attempted to decommit unallocated pages %v:\n%v", fr, gap.Range(), &f.usage))
+	}
+	f.usage.MergeRange(fr)
+}
+
+// runReclaim implements the reclaimer goroutine, which continuously decommits
+// reclaimable frames in order to reduce memory usage.
+func (f *FileMem) runReclaim() {
+	for {
+		fr, ok := f.findReclaimable()
+		if !ok {
+			break
+		}
+
+		if err := f.Decommit(fr); err != nil {
+			log.Warningf("Reclaim failed to decommit %v: %v", fr, err)
+			// Zero the frames manually. This won't reduce memory usage, but at
+			// least ensures that the frames will be zero when reallocated.
+			f.forEachMappingSlice(fr, func(bs []byte) {
+				for i := range bs {
+					bs[i] = 0
+				}
+			})
+			// Pretend the frames were decommitted even though they weren't,
+			// since the memory accounting implementation has no idea how to
+			// deal with this.
+			f.markDecommitted(fr)
+		}
+		f.markReclaimed(fr)
+	}
+	// We only get here if findReclaimable finds f.destroyed set and returns
+	// false.
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	if !f.destroyed {
+		panic("findReclaimable broke out of reclaim loop, but f.destroyed is no longer set")
+	}
+	f.file.Close()
+	// Ensure that any attempts to use f.file.Fd() fail instead of getting a fd
+	// that has possibly been reassigned.
+	f.file = nil
+	mappings := f.mappings.Load().([]uintptr)
+	for i, m := range mappings {
+		if m != 0 {
+			_, _, errno := syscall.Syscall(syscall.SYS_MUNMAP, m, chunkSize, 0)
+			if errno != 0 {
+				log.Warningf("Failed to unmap mapping %#x for filemem chunk %d: %v", m, i, errno)
+			}
+		}
+	}
+	// Similarly, invalidate f.mappings. (atomic.Value.Store(nil) panics.)
+	f.mappings.Store([]uintptr{})
+}
+
+func (f *FileMem) findReclaimable() (platform.FileRange, bool) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	for {
+		for {
+			if f.destroyed {
+				return platform.FileRange{}, false
+			}
+			if f.reclaimable {
+				break
+			}
+			f.reclaimCond.Wait()
+		}
+		// Allocate returns the first usable range in offset order and is
+		// currently a linear scan, so reclaiming from the beginning of the
+		// file minimizes the expected latency of Allocate.
+		for seg := f.usage.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+			if seg.ValuePtr().refs == 0 {
+				return seg.Range(), true
+			}
+		}
+		f.reclaimable = false
+	}
+}
+
+func (f *FileMem) markReclaimed(fr platform.FileRange) {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	seg := f.usage.FindSegment(fr.Start)
+	// All of fr should be mapped to a single uncommitted reclaimable segment
+	// accounted to System.
+	if !seg.Ok() {
+		panic(fmt.Sprintf("Reclaimed pages %v include unreferenced pages:\n%v", fr, &f.usage))
+	}
+	if !seg.Range().IsSupersetOf(fr) {
+		panic(fmt.Sprintf("Reclaimed pages %v are not entirely contained in segment %v with state %v:\n%v", fr, seg.Range(), seg.Value(), &f.usage))
+	}
+	if got, want := seg.Value(), (usageInfo{
+		kind:           usage.System,
+		knownCommitted: false,
+		refs:           0,
+	}); got != want {
+		panic(fmt.Sprintf("Reclaimed pages %v in segment %v has incorrect state %v, wanted %v:\n%v", fr, seg.Range(), got, want, &f.usage))
+	}
+	// Deallocate reclaimed pages. Even though all of seg is reclaimable, the
+	// caller of markReclaimed may not have decommitted it, so we can only mark
+	// fr as reclaimed.
+	f.usage.Remove(f.usage.Isolate(seg, fr))
+}
+
+// MapInto implements platform.File.MapInto.
+func (f *FileMem) MapInto(as platform.AddressSpace, addr usermem.Addr, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+	if !fr.WellFormed() || fr.Length() == 0 || fr.Start%usermem.PageSize != 0 || fr.End%usermem.PageSize != 0 {
+		panic(fmt.Sprintf("invalid range: %v", fr))
+	}
+	return as.MapFile(addr, int(f.file.Fd()), fr, at, precommit)
+}
+
+// MapInternal implements platform.File.MapInternal.
+func (f *FileMem) MapInternal(fr platform.FileRange, at usermem.AccessType) (safemem.BlockSeq, error) {
+	if !fr.WellFormed() || fr.Length() == 0 {
+		panic(fmt.Sprintf("invalid range: %v", fr))
+	}
+	if at.Execute {
+		return safemem.BlockSeq{}, syserror.EACCES
+	}
+
+	chunks := ((fr.End + chunkMask) >> chunkShift) - (fr.Start >> chunkShift)
+	if chunks == 1 {
+		// Avoid an unnecessary slice allocation.
+		var seq safemem.BlockSeq
+		err := f.forEachMappingSlice(fr, func(bs []byte) {
+			seq = safemem.BlockSeqOf(safemem.BlockFromSafeSlice(bs))
+		})
+		return seq, err
+	}
+	blocks := make([]safemem.Block, 0, chunks)
+	err := f.forEachMappingSlice(fr, func(bs []byte) {
+		blocks = append(blocks, safemem.BlockFromSafeSlice(bs))
+	})
+	return safemem.BlockSeqFromSlice(blocks), err
+}
+
+// IncRef implements platform.File.IncRef.
+func (f *FileMem) IncRef(fr platform.FileRange) {
+	if !fr.WellFormed() || fr.Length() == 0 || fr.Start%usermem.PageSize != 0 || fr.End%usermem.PageSize != 0 {
+		panic(fmt.Sprintf("invalid range: %v", fr))
+	}
+
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	gap := f.usage.ApplyContiguous(fr, func(seg usageIterator) {
+		seg.ValuePtr().incRef()
+	})
+	if gap.Ok() {
+		panic(fmt.Sprintf("IncRef(%v): attempted to IncRef on unallocated pages %v:\n%v", fr, gap.Range(), &f.usage))
+	}
+}
+
+// DecRef implements platform.File.DecRef.
+func (f *FileMem) DecRef(fr platform.FileRange) {
+	if !fr.WellFormed() || fr.Length() == 0 || fr.Start%usermem.PageSize != 0 || fr.End%usermem.PageSize != 0 {
+		panic(fmt.Sprintf("invalid range: %v", fr))
+	}
+
+	var freed bool
+
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	for seg := f.usage.FindSegment(fr.Start); seg.Ok() && seg.Start() < fr.End; seg = seg.NextSegment() {
+		seg = f.usage.Isolate(seg, fr)
+		val := seg.ValuePtr()
+		val.decRef()
+		if val.refs == 0 {
+			freed = true
+			// Reclassify memory as System, until it's freed by the reclaim
+			// goroutine.
+			if val.knownCommitted {
+				usage.MemoryAccounting.Move(seg.Range().Length(), usage.System, val.kind)
+			}
+			val.kind = usage.System
+		}
+	}
+	f.usage.MergeAdjacent(fr)
+
+	if freed {
+		f.reclaimable = true
+		f.reclaimCond.Signal()
+	}
+}
+
+// Flush implements platform.Mappable.Flush.
+func (f *FileMem) Flush(ctx context.Context) error {
+	return nil
+}
+
+// forEachMappingSlice invokes fn on a sequence of byte slices that
+// collectively map all bytes in fr.
+func (f *FileMem) forEachMappingSlice(fr platform.FileRange, fn func([]byte)) error {
+	mappings := f.mappings.Load().([]uintptr)
+	for chunkStart := fr.Start &^ chunkMask; chunkStart < fr.End; chunkStart += chunkSize {
+		chunk := int(chunkStart >> chunkShift)
+		m := atomic.LoadUintptr(&mappings[chunk])
+		if m == 0 {
+			var err error
+			mappings, m, err = f.getChunkMapping(chunk)
+			if err != nil {
+				return err
+			}
+		}
+		startOff := uint64(0)
+		if chunkStart < fr.Start {
+			startOff = fr.Start - chunkStart
+		}
+		endOff := uint64(chunkSize)
+		if chunkStart+chunkSize > fr.End {
+			endOff = fr.End - chunkStart
+		}
+		fn(unsafeSlice(m, chunkSize)[startOff:endOff])
+	}
+	return nil
+}
+
+func (f *FileMem) getChunkMapping(chunk int) ([]uintptr, uintptr, error) {
+	f.mappingsMu.Lock()
+	defer f.mappingsMu.Unlock()
+	// Another thread may have replaced f.mappings altogether due to file
+	// expansion.
+	mappings := f.mappings.Load().([]uintptr)
+	// Another thread may have already mapped the chunk.
+	if m := mappings[chunk]; m != 0 {
+		return mappings, m, nil
+	}
+	m, _, errno := syscall.Syscall6(
+		syscall.SYS_MMAP,
+		0,
+		chunkSize,
+		syscall.PROT_READ|syscall.PROT_WRITE,
+		syscall.MAP_SHARED,
+		f.file.Fd(),
+		uintptr(chunk<<chunkShift))
+	if errno != 0 {
+		return nil, 0, errno
+	}
+	atomic.StoreUintptr(&mappings[chunk], m)
+	return mappings, m, nil
+}
+
+// UpdateUsage implements platform.Memory.UpdateUsage.
+func (f *FileMem) UpdateUsage() error {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+
+	// If the underlying usage matches where the usage tree already
+	// represents, then we can just avoid the entire scan (we know it's
+	// accurate).
+	currentUsage, err := f.TotalUsage()
+	if err != nil {
+		return err
+	}
+	if currentUsage == f.usageExpected && f.usageSwapped == 0 {
+		log.Debugf("UpdateUsage: skipped with usageSwapped=0.")
+		return nil
+	}
+	// If the current usage matches the expected but there's swap
+	// accounting, then ensure a scan takes place at least every second
+	// (when requested).
+	if currentUsage == f.usageExpected+f.usageSwapped && time.Now().Before(f.usageLast.Add(time.Second)) {
+		log.Debugf("UpdateUsage: skipped with usageSwapped!=0.")
+		return nil
+	}
+
+	f.usageLast = time.Now()
+	err = f.updateUsageLocked(currentUsage, mincore)
+	log.Debugf("UpdateUsage: currentUsage=%d, usageExpected=%d, usageSwapped=%d.",
+		currentUsage, f.usageExpected, f.usageSwapped)
+	log.Debugf("UpdateUsage: took %v.", time.Since(f.usageLast))
+	return err
+}
+
+// updateUsageLocked attempts to detect commitment of previous-uncommitted
+// pages by invoking checkCommitted, which is a function that, for each page i
+// in bs, sets committed[i] to 1 if the page is committed and 0 otherwise.
+//
+// Precondition: f.mu must be held.
+func (f *FileMem) updateUsageLocked(currentUsage uint64, checkCommitted func(bs []byte, committed []byte) error) error {
+	// Track if anything changed to elide the merge. In the common case, we
+	// expect all segments to be committed and no merge to occur.
+	changedAny := false
+	defer func() {
+		if changedAny {
+			f.usage.MergeAll()
+		}
+
+		// Adjust the swap usage to reflect reality.
+		if f.usageExpected < currentUsage {
+			// Since no pages may be decommitted while we hold usageMu, we
+			// know that usage may have only increased since we got the
+			// last current usage. Therefore, if usageExpected is still
+			// short of currentUsage, we must assume that the difference is
+			// in pages that have been swapped.
+			newUsageSwapped := currentUsage - f.usageExpected
+			if f.usageSwapped < newUsageSwapped {
+				usage.MemoryAccounting.Inc(newUsageSwapped-f.usageSwapped, usage.System)
+			} else {
+				usage.MemoryAccounting.Dec(f.usageSwapped-newUsageSwapped, usage.System)
+			}
+			f.usageSwapped = newUsageSwapped
+		} else if f.usageSwapped != 0 {
+			// We have more usage accounted for than the file itself.
+			// That's fine, we probably caught a race where pages were
+			// being committed while the above loop was running. Just
+			// report the higher number that we found and ignore swap.
+			usage.MemoryAccounting.Dec(f.usageSwapped, usage.System)
+			f.usageSwapped = 0
+		}
+	}()
+
+	// Reused mincore buffer, will generally be <= 4096 bytes.
+	var buf []byte
+
+	// Iterate over all usage data. There will only be usage segments
+	// present when there is an associated reference.
+	for seg := f.usage.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		val := seg.Value()
+
+		// Already known to be committed; ignore.
+		if val.knownCommitted {
+			continue
+		}
+
+		// Assume that reclaimable pages (that aren't already known to be
+		// committed) are not committed. This isn't necessarily true, even
+		// after the reclaimer does Decommit(), because the kernel may
+		// subsequently back the hugepage-sized region containing the
+		// decommitted page with a hugepage. However, it's consistent with our
+		// treatment of unallocated pages, which have the same property.
+		if val.refs == 0 {
+			continue
+		}
+
+		// Get the range for this segment. As we touch slices, the
+		// Start value will be walked along.
+		r := seg.Range()
+
+		var checkErr error
+		err := f.forEachMappingSlice(r, func(s []byte) {
+			if checkErr != nil {
+				return
+			}
+
+			// Ensure that we have sufficient buffer for the call
+			// (one byte per page). The length of each slice must
+			// be page-aligned.
+			bufLen := len(s) / usermem.PageSize
+			if len(buf) < bufLen {
+				buf = make([]byte, bufLen)
+			}
+
+			// Query for new pages in core.
+			if err := checkCommitted(s, buf); err != nil {
+				checkErr = err
+				return
+			}
+
+			// Scan each page and switch out segments.
+			populatedRun := false
+			populatedRunStart := 0
+			for i := 0; i <= bufLen; i++ {
+				// We run past the end of the slice here to
+				// simplify the logic and only set populated if
+				// we're still looking at elements.
+				populated := false
+				if i < bufLen {
+					populated = buf[i]&0x1 != 0
+				}
+
+				switch {
+				case populated == populatedRun:
+					// Keep the run going.
+					continue
+				case populated && !populatedRun:
+					// Begin the run.
+					populatedRun = true
+					populatedRunStart = i
+					// Keep going.
+					continue
+				case !populated && populatedRun:
+					// Finish the run by changing this segment.
+					runRange := platform.FileRange{
+						Start: r.Start + uint64(populatedRunStart*usermem.PageSize),
+						End:   r.Start + uint64(i*usermem.PageSize),
+					}
+					seg = f.usage.Isolate(seg, runRange)
+					seg.ValuePtr().knownCommitted = true
+					// Advance the segment only if we still
+					// have work to do in the context of
+					// the original segment from the for
+					// loop. Otherwise, the for loop itself
+					// will advance the segment
+					// appropriately.
+					if runRange.End != r.End {
+						seg = seg.NextSegment()
+					}
+					amount := runRange.Length()
+					usage.MemoryAccounting.Inc(amount, val.kind)
+					f.usageExpected += amount
+					changedAny = true
+					populatedRun = false
+				}
+			}
+
+			// Advance r.Start.
+			r.Start += uint64(len(s))
+		})
+		if checkErr != nil {
+			return checkErr
+		}
+		if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// TotalUsage implements platform.Memory.TotalUsage.
+func (f *FileMem) TotalUsage() (uint64, error) {
+	// Stat the underlying file to discover the underlying usage. stat(2)
+	// always reports the allocated block count in units of 512 bytes. This
+	// includes pages in the page cache and swapped pages.
+	var stat syscall.Stat_t
+	if err := syscall.Fstat(int(f.file.Fd()), &stat); err != nil {
+		return 0, err
+	}
+	return uint64(stat.Blocks * 512), nil
+}
+
+// TotalSize implements platform.Memory.TotalSize.
+func (f *FileMem) TotalSize() uint64 {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return uint64(f.fileSize)
+}
+
+// File returns the memory file used by f.
+func (f *FileMem) File() *os.File {
+	return f.file
+}
+
+// String implements fmt.Stringer.String.
+//
+// Note that because f.String locks f.mu, calling f.String internally
+// (including indirectly through the fmt package) risks recursive locking.
+// Within the filemem package, use f.usage directly instead.
+func (f *FileMem) String() string {
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	return f.usage.String()
+}
+
+type usageSetFunctions struct{}
+
+func (usageSetFunctions) MinKey() uint64 {
+	return 0
+}
+
+func (usageSetFunctions) MaxKey() uint64 {
+	return math.MaxUint64
+}
+
+func (usageSetFunctions) ClearValue(val *usageInfo) {
+}
+
+func (usageSetFunctions) Merge(_ platform.FileRange, val1 usageInfo, _ platform.FileRange, val2 usageInfo) (usageInfo, bool) {
+	return val1, val1 == val2
+}
+
+func (usageSetFunctions) Split(_ platform.FileRange, val usageInfo, _ uint64) (usageInfo, usageInfo) {
+	return val, val
+}
diff --git a/pkg/sentry/platform/filemem/filemem_state.go b/pkg/sentry/platform/filemem/filemem_state.go
new file mode 100644
index 000000000..5dace8fec
--- /dev/null
+++ b/pkg/sentry/platform/filemem/filemem_state.go
@@ -0,0 +1,170 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package filemem
+
+import (
+	"bytes"
+	"fmt"
+	"io"
+	"runtime"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/state"
+)
+
+// SaveTo implements platform.Memory.SaveTo.
+func (f *FileMem) SaveTo(w io.Writer) error {
+	// Wait for reclaim.
+	f.mu.Lock()
+	defer f.mu.Unlock()
+	for f.reclaimable {
+		f.reclaimCond.Signal()
+		f.mu.Unlock()
+		runtime.Gosched()
+		f.mu.Lock()
+	}
+
+	// Ensure that all pages that contain data have knownCommitted set, since
+	// we only store knownCommitted pages below.
+	zeroPage := make([]byte, usermem.PageSize)
+	err := f.updateUsageLocked(0, func(bs []byte, committed []byte) error {
+		for pgoff := 0; pgoff < len(bs); pgoff += usermem.PageSize {
+			i := pgoff / usermem.PageSize
+			pg := bs[pgoff : pgoff+usermem.PageSize]
+			if !bytes.Equal(pg, zeroPage) {
+				committed[i] = 1
+				continue
+			}
+			committed[i] = 0
+			// Reading the page caused it to be committed; decommit it to
+			// reduce memory usage.
+			//
+			// "MADV_REMOVE [...] Free up a given range of pages and its
+			// associated backing store. This is equivalent to punching a hole
+			// in the corresponding byte range of the backing store (see
+			// fallocate(2))." - madvise(2)
+			if err := syscall.Madvise(pg, syscall.MADV_REMOVE); err != nil {
+				// This doesn't impact the correctness of saved memory, it
+				// just means that we're incrementally more likely to OOM.
+				// Complain, but don't abort saving.
+				log.Warningf("Decommitting page %p while saving failed: %v", pg, err)
+			}
+		}
+		return nil
+	})
+	if err != nil {
+		return err
+	}
+
+	// Save metadata.
+	if err := state.Save(w, &f.fileSize, nil); err != nil {
+		return err
+	}
+	if err := state.Save(w, &f.usage, nil); err != nil {
+		return err
+	}
+
+	// Dump out committed pages.
+	for seg := f.usage.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		if !seg.Value().knownCommitted {
+			continue
+		}
+		// Write a header to distinguish from objects.
+		if err := state.WriteHeader(w, uint64(seg.Range().Length()), false); err != nil {
+			return err
+		}
+		// Write out data.
+		var ioErr error
+		err := f.forEachMappingSlice(seg.Range(), func(s []byte) {
+			if ioErr != nil {
+				return
+			}
+			_, ioErr = w.Write(s)
+		})
+		if ioErr != nil {
+			return ioErr
+		}
+		if err != nil {
+			return err
+		}
+
+		// Update accounting for restored pages. We need to do this here since
+		// these segments are marked as "known committed", and will be skipped
+		// over on accounting scans.
+		usage.MemoryAccounting.Inc(seg.Range().Length(), seg.Value().kind)
+	}
+
+	return nil
+}
+
+// LoadFrom implements platform.Memory.LoadFrom.
+func (f *FileMem) LoadFrom(r io.Reader) error {
+	// Load metadata.
+	if err := state.Load(r, &f.fileSize, nil); err != nil {
+		return err
+	}
+	if err := f.file.Truncate(f.fileSize); err != nil {
+		return err
+	}
+	newMappings := make([]uintptr, f.fileSize>>chunkShift)
+	f.mappings.Store(newMappings)
+	if err := state.Load(r, &f.usage, nil); err != nil {
+		return err
+	}
+
+	// Load committed pages.
+	for seg := f.usage.FirstSegment(); seg.Ok(); seg = seg.NextSegment() {
+		if !seg.Value().knownCommitted {
+			continue
+		}
+		// Verify header.
+		length, object, err := state.ReadHeader(r)
+		if err != nil {
+			return err
+		}
+		if object {
+			// Not expected.
+			return fmt.Errorf("unexpected object")
+		}
+		if expected := uint64(seg.Range().Length()); length != expected {
+			// Size mismatch.
+			return fmt.Errorf("mismatched segment: expected %d, got %d", expected, length)
+		}
+		// Read data.
+		var ioErr error
+		err = f.forEachMappingSlice(seg.Range(), func(s []byte) {
+			if ioErr != nil {
+				return
+			}
+			_, ioErr = io.ReadFull(r, s)
+		})
+		if ioErr != nil {
+			return ioErr
+		}
+		if err != nil {
+			return err
+		}
+
+		// Update accounting for restored pages. We need to do this here since
+		// these segments are marked as "known committed", and will be skipped
+		// over on accounting scans.
+		usage.MemoryAccounting.Inc(seg.End()-seg.Start(), seg.Value().kind)
+	}
+
+	return nil
+}
diff --git a/pkg/sentry/platform/filemem/filemem_test.go b/pkg/sentry/platform/filemem/filemem_test.go
new file mode 100644
index 000000000..46ffcf116
--- /dev/null
+++ b/pkg/sentry/platform/filemem/filemem_test.go
@@ -0,0 +1,122 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package filemem
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	page     = usermem.PageSize
+	hugepage = usermem.HugePageSize
+)
+
+func TestFindUnallocatedRange(t *testing.T) {
+	for _, test := range []struct {
+		desc      string
+		usage     *usageSegmentDataSlices
+		length    uint64
+		alignment uint64
+		start     uint64
+	}{
+		{
+			desc:      "Initial allocation succeeds",
+			usage:     &usageSegmentDataSlices{},
+			length:    page,
+			alignment: page,
+			start:     0,
+		},
+		{
+			desc: "Allocation begins at start of file",
+			usage: &usageSegmentDataSlices{
+				Start:  []uint64{page},
+				End:    []uint64{2 * page},
+				Values: []usageInfo{{refs: 1}},
+			},
+			length:    page,
+			alignment: page,
+			start:     0,
+		},
+		{
+			desc: "In-use frames are not allocatable",
+			usage: &usageSegmentDataSlices{
+				Start:  []uint64{0, page},
+				End:    []uint64{page, 2 * page},
+				Values: []usageInfo{{refs: 1}, {refs: 2}},
+			},
+			length:    page,
+			alignment: page,
+			start:     2 * page,
+		},
+		{
+			desc: "Reclaimable frames are not allocatable",
+			usage: &usageSegmentDataSlices{
+				Start:  []uint64{0, page, 2 * page},
+				End:    []uint64{page, 2 * page, 3 * page},
+				Values: []usageInfo{{refs: 1}, {refs: 0}, {refs: 1}},
+			},
+			length:    page,
+			alignment: page,
+			start:     3 * page,
+		},
+		{
+			desc: "Gaps between in-use frames are allocatable",
+			usage: &usageSegmentDataSlices{
+				Start:  []uint64{0, 2 * page},
+				End:    []uint64{page, 3 * page},
+				Values: []usageInfo{{refs: 1}, {refs: 1}},
+			},
+			length:    page,
+			alignment: page,
+			start:     page,
+		},
+		{
+			desc: "Inadequately-sized gaps are rejected",
+			usage: &usageSegmentDataSlices{
+				Start:  []uint64{0, 2 * page},
+				End:    []uint64{page, 3 * page},
+				Values: []usageInfo{{refs: 1}, {refs: 1}},
+			},
+			length:    2 * page,
+			alignment: page,
+			start:     3 * page,
+		},
+		{
+			desc: "Hugepage alignment is honored",
+			usage: &usageSegmentDataSlices{
+				Start: []uint64{0, hugepage + page},
+				// Hugepage-sized gap here that shouldn't be allocated from
+				// since it's incorrectly aligned.
+				End:    []uint64{page, hugepage + 2*page},
+				Values: []usageInfo{{refs: 1}, {refs: 1}},
+			},
+			length:    hugepage,
+			alignment: hugepage,
+			start:     2 * hugepage,
+		},
+	} {
+		t.Run(test.desc, func(t *testing.T) {
+			var usage usageSet
+			if err := usage.ImportSortedSlices(test.usage); err != nil {
+				t.Fatalf("Failed to initialize usage from %v: %v", test.usage, err)
+			}
+			if got, want := findUnallocatedRange(&usage, test.length, test.alignment), test.start; got != want {
+				t.Errorf("findUnallocatedRange(%v, %d, %d): got %d, wanted %d", test.usage, test.length, test.alignment, got, want)
+			}
+		})
+	}
+}
diff --git a/pkg/sentry/platform/filemem/filemem_unsafe.go b/pkg/sentry/platform/filemem/filemem_unsafe.go
new file mode 100644
index 000000000..a23b9825a
--- /dev/null
+++ b/pkg/sentry/platform/filemem/filemem_unsafe.go
@@ -0,0 +1,40 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package filemem
+
+import (
+	"reflect"
+	"syscall"
+	"unsafe"
+)
+
+func unsafeSlice(addr uintptr, length int) (slice []byte) {
+	sh := (*reflect.SliceHeader)(unsafe.Pointer(&slice))
+	sh.Data = addr
+	sh.Len = length
+	sh.Cap = length
+	return
+}
+
+func mincore(s []byte, buf []byte) error {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_MINCORE,
+		uintptr(unsafe.Pointer(&s[0])),
+		uintptr(len(s)),
+		uintptr(unsafe.Pointer(&buf[0]))); errno != 0 {
+		return errno
+	}
+	return nil
+}
diff --git a/pkg/sentry/platform/interrupt/BUILD b/pkg/sentry/platform/interrupt/BUILD
new file mode 100644
index 000000000..33dde2a31
--- /dev/null
+++ b/pkg/sentry/platform/interrupt/BUILD
@@ -0,0 +1,19 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "interrupt",
+    srcs = [
+        "interrupt.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/interrupt",
+    visibility = ["//pkg/sentry:internal"],
+)
+
+go_test(
+    name = "interrupt_test",
+    size = "small",
+    srcs = ["interrupt_test.go"],
+    embed = [":interrupt"],
+)
diff --git a/pkg/sentry/platform/interrupt/interrupt.go b/pkg/sentry/platform/interrupt/interrupt.go
new file mode 100644
index 000000000..ca4f42087
--- /dev/null
+++ b/pkg/sentry/platform/interrupt/interrupt.go
@@ -0,0 +1,96 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package interrupt provides an interrupt helper.
+package interrupt
+
+import (
+	"fmt"
+	"sync"
+)
+
+// Receiver receives interrupt notifications from a Forwarder.
+type Receiver interface {
+	// NotifyInterrupt is called when the Receiver receives an interrupt.
+	NotifyInterrupt()
+}
+
+// Forwarder is a helper for delivering delayed signal interruptions.
+//
+// This helps platform implementations with Interrupt semantics.
+type Forwarder struct {
+	// mu protects the below.
+	mu sync.Mutex
+
+	// dst is the function to be called when NotifyInterrupt() is called. If
+	// dst is nil, pending will be set instead, causing the next call to
+	// Enable() to return false.
+	dst     Receiver
+	pending bool
+}
+
+// Enable attempts to enable interrupt forwarding to r. If f has already
+// received an interrupt, Enable does nothing and returns false. Otherwise,
+// future calls to f.NotifyInterrupt() cause r.NotifyInterrupt() to be called,
+// and Enable returns true.
+//
+// Usage:
+//
+// if !f.Enable(r) {
+//     // There was an interrupt.
+//     return
+// }
+// defer f.Disable()
+//
+// Preconditions: r must not be nil. f must not already be forwarding
+// interrupts to a Receiver.
+func (f *Forwarder) Enable(r Receiver) bool {
+	if r == nil {
+		panic("nil Receiver")
+	}
+	f.mu.Lock()
+	if f.dst != nil {
+		f.mu.Unlock()
+		panic(fmt.Sprintf("already forwarding interrupts to %+v", f.dst))
+	}
+	if f.pending {
+		f.pending = false
+		f.mu.Unlock()
+		return false
+	}
+	f.dst = r
+	f.mu.Unlock()
+	return true
+}
+
+// Disable stops interrupt forwarding. If interrupt forwarding is already
+// disabled, Disable is a no-op.
+func (f *Forwarder) Disable() {
+	f.mu.Lock()
+	f.dst = nil
+	f.mu.Unlock()
+}
+
+// NotifyInterrupt implements Receiver.NotifyInterrupt. If interrupt forwarding
+// is enabled, the configured Receiver will be notified. Otherwise the
+// interrupt will be delivered to the next call to Enable.
+func (f *Forwarder) NotifyInterrupt() {
+	f.mu.Lock()
+	if f.dst != nil {
+		f.dst.NotifyInterrupt()
+	} else {
+		f.pending = true
+	}
+	f.mu.Unlock()
+}
diff --git a/pkg/sentry/platform/interrupt/interrupt_test.go b/pkg/sentry/platform/interrupt/interrupt_test.go
new file mode 100644
index 000000000..7c49eeea6
--- /dev/null
+++ b/pkg/sentry/platform/interrupt/interrupt_test.go
@@ -0,0 +1,99 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package interrupt
+
+import (
+	"testing"
+)
+
+type countingReceiver struct {
+	interrupts int
+}
+
+// NotifyInterrupt implements Receiver.NotifyInterrupt.
+func (r *countingReceiver) NotifyInterrupt() {
+	r.interrupts++
+}
+
+func TestSingleInterruptBeforeEnable(t *testing.T) {
+	var (
+		f Forwarder
+		r countingReceiver
+	)
+	f.NotifyInterrupt()
+	// The interrupt should cause the first Enable to fail.
+	if f.Enable(&r) {
+		f.Disable()
+		t.Fatalf("Enable: got true, wanted false")
+	}
+	// The failing Enable "acknowledges" the interrupt, allowing future Enables
+	// to succeed.
+	if !f.Enable(&r) {
+		t.Fatalf("Enable: got false, wanted true")
+	}
+	f.Disable()
+}
+
+func TestMultipleInterruptsBeforeEnable(t *testing.T) {
+	var (
+		f Forwarder
+		r countingReceiver
+	)
+	f.NotifyInterrupt()
+	f.NotifyInterrupt()
+	// The interrupts should cause the first Enable to fail.
+	if f.Enable(&r) {
+		f.Disable()
+		t.Fatalf("Enable: got true, wanted false")
+	}
+	// Interrupts are deduplicated while the Forwarder is disabled, so the
+	// failing Enable "acknowledges" all interrupts, allowing future Enables to
+	// succeed.
+	if !f.Enable(&r) {
+		t.Fatalf("Enable: got false, wanted true")
+	}
+	f.Disable()
+}
+
+func TestSingleInterruptAfterEnable(t *testing.T) {
+	var (
+		f Forwarder
+		r countingReceiver
+	)
+	if !f.Enable(&r) {
+		t.Fatalf("Enable: got false, wanted true")
+	}
+	defer f.Disable()
+	f.NotifyInterrupt()
+	if r.interrupts != 1 {
+		t.Errorf("interrupts: got %d, wanted 1", r.interrupts)
+	}
+}
+
+func TestMultipleInterruptsAfterEnable(t *testing.T) {
+	var (
+		f Forwarder
+		r countingReceiver
+	)
+	if !f.Enable(&r) {
+		t.Fatalf("Enable: got false, wanted true")
+	}
+	defer f.Disable()
+	f.NotifyInterrupt()
+	f.NotifyInterrupt()
+	if r.interrupts != 2 {
+		t.Errorf("interrupts: got %d, wanted 2", r.interrupts)
+	}
+}
diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD
new file mode 100644
index 000000000..d902e344a
--- /dev/null
+++ b/pkg/sentry/platform/kvm/BUILD
@@ -0,0 +1,90 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+go_template_instance(
+    name = "host_map_set",
+    out = "host_map_set.go",
+    consts = {
+        "minDegree": "15",
+    },
+    imports = {
+        "usermem": "gvisor.googlesource.com/gvisor/pkg/sentry/usermem",
+    },
+    package = "kvm",
+    prefix = "hostMap",
+    template = "//pkg/segment:generic_set",
+    types = {
+        "Key": "usermem.Addr",
+        "Range": "usermem.AddrRange",
+        "Value": "uintptr",
+        "Functions": "hostMapSetFunctions",
+    },
+)
+
+go_library(
+    name = "kvm",
+    srcs = [
+        "address_space.go",
+        "bluepill.go",
+        "bluepill_amd64.go",
+        "bluepill_amd64.s",
+        "bluepill_amd64_unsafe.go",
+        "bluepill_fault.go",
+        "bluepill_unsafe.go",
+        "context.go",
+        "host_map.go",
+        "host_map_set.go",
+        "kvm.go",
+        "kvm_amd64.go",
+        "kvm_amd64_unsafe.go",
+        "kvm_const.go",
+        "machine.go",
+        "machine_amd64.go",
+        "machine_amd64_unsafe.go",
+        "machine_unsafe.go",
+        "physical_map.go",
+        "virtual_map.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/kvm",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/cpuid",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/platform/filemem",
+        "//pkg/sentry/platform/interrupt",
+        "//pkg/sentry/platform/procid",
+        "//pkg/sentry/platform/ring0",
+        "//pkg/sentry/platform/ring0/pagetables",
+        "//pkg/sentry/platform/safecopy",
+        "//pkg/sentry/time",
+        "//pkg/sentry/usermem",
+        "//pkg/tmutex",
+    ],
+)
+
+go_test(
+    name = "kvm_test",
+    size = "small",
+    srcs = [
+        "kvm_test.go",
+        "virtual_map_test.go",
+    ],
+    embed = [":kvm"],
+    tags = [
+        "nogotsan",
+        "requires-kvm",
+    ],
+    deps = [
+        "//pkg/sentry/arch",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/platform/kvm/testutil",
+        "//pkg/sentry/platform/ring0",
+        "//pkg/sentry/platform/ring0/pagetables",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/platform/kvm/address_space.go b/pkg/sentry/platform/kvm/address_space.go
new file mode 100644
index 000000000..791f038b0
--- /dev/null
+++ b/pkg/sentry/platform/kvm/address_space.go
@@ -0,0 +1,207 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"reflect"
+	"sync"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/filemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// addressSpace is a wrapper for PageTables.
+type addressSpace struct {
+	platform.NoAddressSpaceIO
+
+	// filemem is the memory instance.
+	filemem *filemem.FileMem
+
+	// machine is the underlying machine.
+	machine *machine
+
+	// pageTables are for this particular address space.
+	pageTables *pagetables.PageTables
+
+	// dirtySet is the set of dirty vCPUs.
+	//
+	// The key is the vCPU, the value is a shared uint32 pointer that
+	// indicates whether or not the context is clean. A zero here indicates
+	// that the context should be cleaned prior to re-entry.
+	dirtySet sync.Map
+
+	// files contains files mapped in the host address space.
+	files hostMap
+}
+
+// Invalidate interrupts all dirty contexts.
+func (as *addressSpace) Invalidate() {
+	as.dirtySet.Range(func(key, value interface{}) bool {
+		c := key.(*vCPU)
+		v := value.(*uint32)
+		atomic.StoreUint32(v, 0) // Invalidation required.
+		c.Bounce()               // Force a kernel transition.
+		return true              // Keep iterating.
+	})
+}
+
+// Touch adds the given vCPU to the dirty list.
+func (as *addressSpace) Touch(c *vCPU) *uint32 {
+	value, ok := as.dirtySet.Load(c)
+	if !ok {
+		value, _ = as.dirtySet.LoadOrStore(c, new(uint32))
+	}
+	return value.(*uint32)
+}
+
+func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.AccessType) (inv bool) {
+	for m.length > 0 {
+		physical, length, ok := TranslateToPhysical(m.addr)
+		if !ok {
+			panic("unable to translate segment")
+		}
+		if length > m.length {
+			length = m.length
+		}
+
+		// Ensure that this map has physical mappings. If the page does
+		// not have physical mappings, the KVM module may inject
+		// spurious exceptions when emulation fails (i.e. it tries to
+		// emulate because the RIP is pointed at those pages).
+		as.machine.mapPhysical(physical, length)
+
+		// Install the page table mappings. Note that the ordering is
+		// important; if the pagetable mappings were installed before
+		// ensuring the physical pages were available, then some other
+		// thread could theoretically access them.
+		prev := as.pageTables.Map(addr, length, true /* user */, at, physical)
+		inv = inv || prev
+		m.addr += length
+		m.length -= length
+		addr += usermem.Addr(length)
+	}
+
+	return inv
+}
+
+func (as *addressSpace) mapHostFile(addr usermem.Addr, fd int, fr platform.FileRange, at usermem.AccessType) error {
+	// Create custom host mappings.
+	ms, err := as.files.CreateMappings(usermem.AddrRange{
+		Start: addr,
+		End:   addr + usermem.Addr(fr.End-fr.Start),
+	}, at, fd, fr.Start)
+	if err != nil {
+		return err
+	}
+
+	inv := false
+	for _, m := range ms {
+		// The host mapped slices are guaranteed to be aligned.
+		inv = inv || as.mapHost(addr, m, at)
+		addr += usermem.Addr(m.length)
+	}
+	if inv {
+		as.Invalidate()
+	}
+
+	return nil
+}
+
+func (as *addressSpace) mapFilemem(addr usermem.Addr, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+	// TODO: Lock order at the platform level is not sufficiently
+	// well-defined to guarantee that the caller (FileMem.MapInto) is not
+	// holding any locks that FileMem.MapInternal may take.
+
+	// Retrieve mappings for the underlying filemem. Note that the
+	// permissions here are largely irrelevant, since it corresponds to
+	// physical memory for the guest. We enforce the given access type
+	// below, in the guest page tables.
+	bs, err := as.filemem.MapInternal(fr, usermem.AccessType{
+		Read:  true,
+		Write: true,
+	})
+	if err != nil {
+		return err
+	}
+
+	// Save the original range for invalidation.
+	orig := usermem.AddrRange{
+		Start: addr,
+		End:   addr + usermem.Addr(fr.End-fr.Start),
+	}
+
+	inv := false
+	for !bs.IsEmpty() {
+		b := bs.Head()
+		bs = bs.Tail()
+		// Since fr was page-aligned, b should also be page-aligned. We do the
+		// lookup in our host page tables for this translation.
+		s := b.ToSlice()
+		if precommit {
+			for i := 0; i < len(s); i += usermem.PageSize {
+				_ = s[i] // Touch to commit.
+			}
+		}
+		inv = inv || as.mapHost(addr, hostMapEntry{
+			addr:   reflect.ValueOf(&s[0]).Pointer(),
+			length: uintptr(len(s)),
+		}, at)
+		addr += usermem.Addr(len(s))
+	}
+	if inv {
+		as.Invalidate()
+		as.files.DeleteMapping(orig)
+	}
+
+	return nil
+}
+
+// MapFile implements platform.AddressSpace.MapFile.
+func (as *addressSpace) MapFile(addr usermem.Addr, fd int, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+	// Create an appropriate mapping. If this is filemem, we don't create
+	// custom mappings for each in-application mapping. For files however,
+	// we create distinct mappings for each address space. Unfortunately,
+	// there's not a better way to manage this here. The file underlying
+	// this fd can change at any time, so we can't actually index the file
+	// and share between address space. Oh well. It's all refering to the
+	// same physical pages, hopefully we don't run out of address space.
+	if fd != int(as.filemem.File().Fd()) {
+		// N.B. precommit is ignored for host files.
+		return as.mapHostFile(addr, fd, fr, at)
+	}
+
+	return as.mapFilemem(addr, fr, at, precommit)
+}
+
+// Unmap unmaps the given range by calling pagetables.PageTables.Unmap.
+func (as *addressSpace) Unmap(addr usermem.Addr, length uint64) {
+	if prev := as.pageTables.Unmap(addr, uintptr(length)); prev {
+		as.Invalidate()
+		as.files.DeleteMapping(usermem.AddrRange{
+			Start: addr,
+			End:   addr + usermem.Addr(length),
+		})
+	}
+}
+
+// Release releases the page tables.
+func (as *addressSpace) Release() error {
+	as.Unmap(0, ^uint64(0))
+	as.pageTables.Release()
+	return nil
+}
diff --git a/pkg/sentry/platform/kvm/bluepill.go b/pkg/sentry/platform/kvm/bluepill.go
new file mode 100644
index 000000000..ecc33d7dd
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill.go
@@ -0,0 +1,41 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"fmt"
+	"reflect"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/safecopy"
+)
+
+// bluepill enters guest mode.
+func bluepill(*vCPU)
+
+// sighandler is the signal entry point.
+func sighandler()
+
+// savedHandler is a pointer to the previous handler.
+//
+// This is called by bluepillHandler.
+var savedHandler uintptr
+
+func init() {
+	// Install the handler.
+	if err := safecopy.ReplaceSignalHandler(syscall.SIGSEGV, reflect.ValueOf(sighandler).Pointer(), &savedHandler); err != nil {
+		panic(fmt.Sprintf("Unable to set handler for signal %d: %v", syscall.SIGSEGV, err))
+	}
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.go b/pkg/sentry/platform/kvm/bluepill_amd64.go
new file mode 100644
index 000000000..a2baefb7d
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill_amd64.go
@@ -0,0 +1,143 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+)
+
+var (
+	// bounceSignal is the signal used for bouncing KVM.
+	//
+	// We use SIGCHLD because it is not masked by the runtime, and
+	// it will be ignored properly by other parts of the kernel.
+	bounceSignal = syscall.SIGCHLD
+
+	// bounceSignalMask has only bounceSignal set.
+	bounceSignalMask = uint64(1 << (uint64(bounceSignal) - 1))
+
+	// bounce is the interrupt vector used to return to the kernel.
+	bounce = uint32(ring0.VirtualizationException)
+)
+
+// redpill on amd64 invokes a syscall with -1.
+//
+//go:nosplit
+func redpill() {
+	syscall.RawSyscall(^uintptr(0), 0, 0, 0)
+}
+
+// bluepillArchEnter is called during bluepillEnter.
+//
+//go:nosplit
+func bluepillArchEnter(context *arch.SignalContext64) (c *vCPU) {
+	c = vCPUPtr(uintptr(context.Rax))
+	regs := c.CPU.Registers()
+	regs.R8 = context.R8
+	regs.R9 = context.R9
+	regs.R10 = context.R10
+	regs.R11 = context.R11
+	regs.R12 = context.R12
+	regs.R13 = context.R13
+	regs.R14 = context.R14
+	regs.R15 = context.R15
+	regs.Rdi = context.Rdi
+	regs.Rsi = context.Rsi
+	regs.Rbp = context.Rbp
+	regs.Rbx = context.Rbx
+	regs.Rdx = context.Rdx
+	regs.Rax = context.Rax
+	regs.Rcx = context.Rcx
+	regs.Rsp = context.Rsp
+	regs.Rip = context.Rip
+	regs.Eflags = context.Eflags
+	regs.Eflags &^= uint64(ring0.KernelFlagsClear)
+	regs.Eflags |= ring0.KernelFlagsSet
+	regs.Cs = uint64(ring0.Kcode)
+	regs.Ds = uint64(ring0.Udata)
+	regs.Es = uint64(ring0.Udata)
+	regs.Fs = uint64(ring0.Udata)
+	regs.Ss = uint64(ring0.Kdata)
+
+	// ring0 uses GS exclusively, so we use GS_base to store the location
+	// of the floating point address.
+	//
+	// The address will be restored directly after running the VCPU, and
+	// will be saved again prior to halting. We rely on the fact that the
+	// SaveFloatingPointer/LoadFloatingPoint functions use the most
+	// efficient mechanism available (including compression) so the state
+	// size is guaranteed to be less than what's pointed to here.
+	regs.Gs_base = uint64(context.Fpstate)
+	return
+}
+
+// bluepillSyscall handles kernel syscalls.
+//
+//go:nosplit
+func bluepillSyscall() {
+	regs := ring0.Current().Registers()
+	if regs.Rax != ^uint64(0) {
+		regs.Rip -= 2 // Rewind.
+	}
+	ring0.SaveFloatingPoint(bytePtr(uintptr(regs.Gs_base)))
+	ring0.Halt()
+	ring0.LoadFloatingPoint(bytePtr(uintptr(regs.Gs_base)))
+}
+
+// bluepillException handles kernel exceptions.
+//
+//go:nosplit
+func bluepillException(vector ring0.Vector) {
+	regs := ring0.Current().Registers()
+	if vector == ring0.Vector(bounce) {
+		// These should not interrupt kernel execution; point the Rip
+		// to zero to ensure that we get a reasonable panic when we
+		// attempt to return.
+		regs.Rip = 0
+	}
+	ring0.SaveFloatingPoint(bytePtr(uintptr(regs.Gs_base)))
+	ring0.Halt()
+	ring0.LoadFloatingPoint(bytePtr(uintptr(regs.Gs_base)))
+}
+
+// bluepillArchExit is called during bluepillEnter.
+//
+//go:nosplit
+func bluepillArchExit(c *vCPU, context *arch.SignalContext64) {
+	regs := c.CPU.Registers()
+	context.R8 = regs.R8
+	context.R9 = regs.R9
+	context.R10 = regs.R10
+	context.R11 = regs.R11
+	context.R12 = regs.R12
+	context.R13 = regs.R13
+	context.R14 = regs.R14
+	context.R15 = regs.R15
+	context.Rdi = regs.Rdi
+	context.Rsi = regs.Rsi
+	context.Rbp = regs.Rbp
+	context.Rbx = regs.Rbx
+	context.Rdx = regs.Rdx
+	context.Rax = regs.Rax
+	context.Rcx = regs.Rcx
+	context.Rsp = regs.Rsp
+	context.Rip = regs.Rip
+	context.Eflags = regs.Eflags
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.s b/pkg/sentry/platform/kvm/bluepill_amd64.s
new file mode 100644
index 000000000..0881bd5f5
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill_amd64.s
@@ -0,0 +1,87 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// VCPU_CPU is the location of the CPU in the vCPU struct.
+//
+// This is guaranteed to be zero.
+#define VCPU_CPU 0x0
+
+// CPU_SELF is the self reference in ring0's percpu.
+//
+// This is guaranteed to be zero.
+#define CPU_SELF 0x0
+
+// Context offsets.
+//
+// Only limited use of the context is done in the assembly stub below, most is
+// done in the Go handlers. However, the RIP must be examined.
+#define CONTEXT_RAX 0x90
+#define CONTEXT_RIP 0xa8
+#define CONTEXT_FP  0xe0
+
+// CLI is the literal byte for the disable interrupts instruction.
+//
+// This is checked as the source of the fault.
+#define CLI $0xfa
+
+// See bluepill.go.
+TEXT ·bluepill(SB),NOSPLIT,$0
+begin:
+	MOVQ vcpu+0(FP), AX
+	LEAQ VCPU_CPU(AX), BX
+	BYTE CLI;
+check_vcpu:
+	MOVQ CPU_SELF(GS), CX
+	CMPQ BX, CX
+	JE right_vCPU
+wrong_vcpu:
+	CALL ·redpill(SB)
+	JMP begin
+right_vCPU:
+	RET
+
+// sighandler: see bluepill.go for documentation.
+//
+// The arguments are the following:
+//
+// 	DI - The signal number.
+// 	SI - Pointer to siginfo_t structure.
+// 	DX - Pointer to ucontext structure.
+//
+TEXT ·sighandler(SB),NOSPLIT,$0
+	// Check if the signal is from the kernel.
+	MOVQ $0x80, CX
+	CMPL CX, 0x8(SI)
+	JNE fallback
+
+	// Check if RIP is disable interrupts.
+	MOVQ CONTEXT_RIP(DX), CX
+	CMPQ CX, $0x0
+	JE fallback
+	CMPB 0(CX), CLI
+	JNE fallback
+
+	// Call the bluepillHandler.
+	PUSHQ DX                    // First argument (context).
+	CALL ·bluepillHandler(SB)   // Call the handler.
+	POPQ DX                     // Discard the argument.
+	RET
+
+fallback:
+	// Jump to the previous signal handler.
+	XORQ CX, CX
+	MOVQ ·savedHandler(SB), AX
+	JMP AX
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
new file mode 100644
index 000000000..61ca61dcb
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
@@ -0,0 +1,28 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+)
+
+// bluepillArchContext returns the arch-specific context.
+func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 {
+	return &((*arch.UContext64)(context).MContext)
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_fault.go b/pkg/sentry/platform/kvm/bluepill_fault.go
new file mode 100644
index 000000000..7c8c7bc37
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill_fault.go
@@ -0,0 +1,127 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// faultBlockSize is the size used for servicing memory faults.
+	//
+	// This should be large enough to avoid frequent faults and avoid using
+	// all available KVM slots (~512), but small enough that KVM does not
+	// complain about slot sizes (~4GB). See handleBluepillFault for how
+	// this block is used.
+	faultBlockSize = 2 << 30
+
+	// faultBlockMask is the mask for the fault blocks.
+	//
+	// This must be typed to avoid overflow complaints (ugh).
+	faultBlockMask = ^uintptr(faultBlockSize - 1)
+)
+
+// yield yields the CPU.
+//
+//go:nosplit
+func yield() {
+	syscall.RawSyscall(syscall.SYS_SCHED_YIELD, 0, 0, 0)
+}
+
+// calculateBluepillFault calculates the fault address range.
+//
+//go:nosplit
+func calculateBluepillFault(m *machine, physical uintptr) (virtualStart, physicalStart, length uintptr, ok bool) {
+	alignedPhysical := physical &^ uintptr(usermem.PageSize-1)
+	for _, pr := range physicalRegions {
+		end := pr.physical + pr.length
+		if physical < pr.physical || physical >= end {
+			continue
+		}
+
+		// Adjust the block to match our size.
+		physicalStart = alignedPhysical & faultBlockMask
+		if physicalStart < pr.physical {
+			// Bound the starting point to the start of the region.
+			physicalStart = pr.physical
+		}
+		virtualStart = pr.virtual + (physicalStart - pr.physical)
+		physicalEnd := physicalStart + faultBlockSize
+		if physicalEnd > end {
+			physicalEnd = end
+		}
+		length = physicalEnd - physicalStart
+		return virtualStart, physicalStart, length, true
+	}
+
+	return 0, 0, 0, false
+}
+
+// handleBluepillFault handles a physical fault.
+//
+// The corresponding virtual address is returned. This may throw on error.
+//
+//go:nosplit
+func handleBluepillFault(m *machine, physical uintptr) (uintptr, bool) {
+	// Paging fault: we need to map the underlying physical pages for this
+	// fault. This all has to be done in this function because we're in a
+	// signal handler context. (We can't call any functions that might
+	// split the stack.)
+	virtualStart, physicalStart, length, ok := calculateBluepillFault(m, physical)
+	if !ok {
+		return 0, false
+	}
+
+	// Set the KVM slot.
+	//
+	// First, we need to acquire the exclusive right to set a slot.  See
+	// machine.nextSlot for information about the protocol.
+	slot := atomic.SwapUint32(&m.nextSlot, ^uint32(0))
+	for slot == ^uint32(0) {
+		yield() // Race with another call.
+		slot = atomic.SwapUint32(&m.nextSlot, ^uint32(0))
+	}
+	errno := m.setMemoryRegion(int(slot), physicalStart, length, virtualStart)
+	if errno == 0 {
+		// Successfully added region; we can increment nextSlot and
+		// allow another set to proceed here.
+		atomic.StoreUint32(&m.nextSlot, slot+1)
+		return virtualStart + (physical - physicalStart), true
+	}
+
+	// Release our slot (still available).
+	atomic.StoreUint32(&m.nextSlot, slot)
+
+	switch errno {
+	case syscall.EEXIST:
+		// The region already exists. It's possible that we raced with
+		// another vCPU here. We just revert nextSlot and return true,
+		// because this must have been satisfied by some other vCPU.
+		return virtualStart + (physical - physicalStart), true
+	case syscall.EINVAL:
+		throw("set memory region failed; out of slots")
+	case syscall.ENOMEM:
+		throw("set memory region failed: out of memory")
+	case syscall.EFAULT:
+		throw("set memory region failed: invalid physical range")
+	default:
+		throw("set memory region failed: unknown reason")
+	}
+
+	panic("unreachable")
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_unsafe.go b/pkg/sentry/platform/kvm/bluepill_unsafe.go
new file mode 100644
index 000000000..85703ff18
--- /dev/null
+++ b/pkg/sentry/platform/kvm/bluepill_unsafe.go
@@ -0,0 +1,175 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+)
+
+//go:linkname throw runtime.throw
+func throw(string)
+
+// vCPUPtr returns a CPU for the given address.
+//
+//go:nosplit
+func vCPUPtr(addr uintptr) *vCPU {
+	return (*vCPU)(unsafe.Pointer(addr))
+}
+
+// bytePtr returns a bytePtr for the given address.
+//
+//go:nosplit
+func bytePtr(addr uintptr) *byte {
+	return (*byte)(unsafe.Pointer(addr))
+}
+
+// bluepillHandler is called from the signal stub.
+//
+// The world may be stopped while this is executing, and it executes on the
+// signal stack. It should only execute raw system calls and functions that are
+// explicitly marked go:nosplit.
+//
+//go:nosplit
+func bluepillHandler(context unsafe.Pointer) {
+	// Sanitize the registers; interrupts must always be disabled.
+	c := bluepillArchEnter(bluepillArchContext(context))
+
+	// Increment the number of switches.
+	atomic.AddUint32(&c.switches, 1)
+
+	// Store vCPUGuest.
+	//
+	// This is fine even if we're not in guest mode yet.  In this signal
+	// handler, we'll already have all the relevant signals blocked, so an
+	// interrupt is only deliverable when we actually execute the KVM_RUN.
+	//
+	// The state will be returned to vCPUReady by Phase2.
+	if state := atomic.SwapUintptr(&c.state, vCPUGuest); state != vCPUReady {
+		throw("vCPU not in ready state")
+	}
+
+	for {
+		_, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(c.fd), _KVM_RUN, 0)
+		if errno == syscall.EINTR {
+			// First, we process whatever pending signal
+			// interrupted KVM. Since we're in a signal handler
+			// currently, all signals are masked and the signal
+			// must have been delivered directly to this thread.
+			sig, _, errno := syscall.RawSyscall6(
+				syscall.SYS_RT_SIGTIMEDWAIT,
+				uintptr(unsafe.Pointer(&bounceSignalMask)),
+				0, // siginfo.
+				0, // timeout.
+				8, // sigset size.
+				0, 0)
+			if errno != 0 {
+				throw("error waiting for pending signal")
+			}
+			if sig != uintptr(bounceSignal) {
+				throw("unexpected signal")
+			}
+
+			// Check whether the current state of the vCPU is ready
+			// for interrupt injection. Because we don't have a
+			// PIC, we can't inject an interrupt while they are
+			// masked. We need to request a window if it's not
+			// ready.
+			if c.runData.readyForInterruptInjection == 0 {
+				c.runData.requestInterruptWindow = 1
+				continue // Rerun vCPU.
+			} else {
+				// Force injection below; the vCPU is ready.
+				c.runData.exitReason = _KVM_EXIT_IRQ_WINDOW_OPEN
+			}
+		} else if errno != 0 {
+			throw("run failed")
+		}
+
+		switch c.runData.exitReason {
+		case _KVM_EXIT_EXCEPTION:
+			throw("exception")
+		case _KVM_EXIT_IO:
+			throw("I/O")
+		case _KVM_EXIT_INTERNAL_ERROR:
+			throw("internal error")
+		case _KVM_EXIT_HYPERCALL:
+			throw("hypercall")
+		case _KVM_EXIT_DEBUG:
+			throw("debug")
+		case _KVM_EXIT_HLT:
+			// Copy out registers.
+			bluepillArchExit(c, bluepillArchContext(context))
+
+			// Notify any waiters.
+			switch state := atomic.SwapUintptr(&c.state, vCPUReady); state {
+			case vCPUGuest:
+			case vCPUWaiter:
+				c.notify() // Safe from handler.
+			default:
+				throw("invalid state")
+			}
+			return
+		case _KVM_EXIT_MMIO:
+			// Increment the fault count.
+			atomic.AddUint32(&c.faults, 1)
+
+			// For MMIO, the physical address is the first data item.
+			virtual, ok := handleBluepillFault(c.machine, uintptr(c.runData.data[0]))
+			if !ok {
+				throw("physical address not valid")
+			}
+
+			// We now need to fill in the data appropriately. KVM
+			// expects us to provide the result of the given MMIO
+			// operation in the runData struct. This is safe
+			// because, if a fault occurs here, the same fault
+			// would have occurred in guest mode. The kernel should
+			// not create invalid page table mappings.
+			data := (*[8]byte)(unsafe.Pointer(&c.runData.data[1]))
+			length := (uintptr)((uint32)(c.runData.data[2]))
+			write := (uint8)((c.runData.data[2] >> 32 & 0xff)) != 0
+			for i := uintptr(0); i < length; i++ {
+				b := bytePtr(uintptr(virtual) + i)
+				if write {
+					// Write to the given address.
+					*b = data[i]
+				} else {
+					// Read from the given address.
+					data[i] = *b
+				}
+			}
+		case _KVM_EXIT_IRQ_WINDOW_OPEN:
+			// Interrupt: we must have requested an interrupt
+			// window; set the interrupt line.
+			if _, _, errno := syscall.RawSyscall(
+				syscall.SYS_IOCTL,
+				uintptr(c.fd),
+				_KVM_INTERRUPT,
+				uintptr(unsafe.Pointer(&bounce))); errno != 0 {
+				throw("interrupt injection failed")
+			}
+			// Clear previous injection request.
+			c.runData.requestInterruptWindow = 0
+		case _KVM_EXIT_SHUTDOWN:
+			throw("shutdown")
+		case _KVM_EXIT_FAIL_ENTRY:
+			throw("entry failed")
+		default:
+			throw("unknown failure")
+		}
+	}
+}
diff --git a/pkg/sentry/platform/kvm/context.go b/pkg/sentry/platform/kvm/context.go
new file mode 100644
index 000000000..fd04a2c47
--- /dev/null
+++ b/pkg/sentry/platform/kvm/context.go
@@ -0,0 +1,81 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/interrupt"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// context is an implementation of the platform context.
+//
+// This is a thin wrapper around the machine.
+type context struct {
+	// machine is the parent machine, and is immutable.
+	machine *machine
+
+	// interrupt is the interrupt context.
+	interrupt interrupt.Forwarder
+}
+
+// Switch runs the provided context in the given address space.
+func (c *context) Switch(as platform.AddressSpace, ac arch.Context, _ int32) (*arch.SignalInfo, usermem.AccessType, error) {
+	// Extract data.
+	localAS := as.(*addressSpace)
+	regs := &ac.StateData().Regs
+	fp := (*byte)(ac.FloatingPointData())
+
+	// Grab a vCPU.
+	cpu, err := c.machine.Get()
+	if err != nil {
+		return nil, usermem.NoAccess, err
+	}
+
+	// Enable interrupts (i.e. calls to vCPU.Notify).
+	if !c.interrupt.Enable(cpu) {
+		c.machine.Put(cpu) // Already preempted.
+		return nil, usermem.NoAccess, platform.ErrContextInterrupt
+	}
+
+	// Mark the address space as dirty.
+	flags := ring0.Flags(0)
+	dirty := localAS.Touch(cpu)
+	if v := atomic.SwapUint32(dirty, 1); v == 0 {
+		flags |= ring0.FlagFlush
+	}
+	if ac.FullRestore() {
+		flags |= ring0.FlagFull
+	}
+
+	// Take the blue pill.
+	si, at, err := cpu.SwitchToUser(regs, fp, localAS.pageTables, flags)
+
+	// Release resources.
+	c.machine.Put(cpu)
+
+	// All done.
+	c.interrupt.Disable()
+	return si, at, err
+}
+
+// Interrupt interrupts the running context.
+func (c *context) Interrupt() {
+	c.interrupt.NotifyInterrupt()
+}
diff --git a/pkg/sentry/platform/kvm/host_map.go b/pkg/sentry/platform/kvm/host_map.go
new file mode 100644
index 000000000..357f8c92e
--- /dev/null
+++ b/pkg/sentry/platform/kvm/host_map.go
@@ -0,0 +1,168 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type hostMap struct {
+	// mu protects below.
+	mu sync.RWMutex
+
+	// set contains host mappings.
+	set hostMapSet
+}
+
+type hostMapEntry struct {
+	addr   uintptr
+	length uintptr
+}
+
+func (hm *hostMap) forEachEntry(r usermem.AddrRange, fn func(offset uint64, m hostMapEntry)) {
+	for seg := hm.set.FindSegment(r.Start); seg.Ok() && seg.Start() < r.End; seg = seg.NextSegment() {
+		length := uintptr(seg.Range().Length())
+		segOffset := uint64(0) // Adjusted below.
+		if seg.End() > r.End {
+			length -= uintptr(seg.End() - r.End)
+		}
+		if seg.Start() < r.Start {
+			length -= uintptr(r.Start - seg.Start())
+		} else {
+			segOffset = uint64(seg.Start() - r.Start)
+		}
+		fn(segOffset, hostMapEntry{
+			addr:   seg.Value(),
+			length: length,
+		})
+	}
+}
+
+func (hm *hostMap) createMappings(r usermem.AddrRange, at usermem.AccessType, fd int, offset uint64) (ms []hostMapEntry, err error) {
+	// Replace any existing mappings.
+	hm.forEachEntry(r, func(segOffset uint64, m hostMapEntry) {
+		_, _, errno := syscall.RawSyscall6(
+			syscall.SYS_MMAP,
+			m.addr,
+			m.length,
+			uintptr(at.Prot()),
+			syscall.MAP_FIXED|syscall.MAP_SHARED,
+			uintptr(fd),
+			uintptr(offset+segOffset))
+		if errno != 0 && err == nil {
+			err = errno
+		}
+	})
+	if err != nil {
+		return nil, err
+	}
+
+	// Add in necessary new mappings.
+	for gap := hm.set.FindGap(r.Start); gap.Ok() && gap.Start() < r.End; {
+		length := uintptr(gap.Range().Length())
+		gapOffset := uint64(0) // Adjusted below.
+		if gap.End() > r.End {
+			length -= uintptr(gap.End() - r.End)
+		}
+		if gap.Start() < r.Start {
+			length -= uintptr(r.Start - gap.Start())
+		} else {
+			gapOffset = uint64(gap.Start() - r.Start)
+		}
+
+		// Map the host file memory.
+		hostAddr, _, errno := syscall.RawSyscall6(
+			syscall.SYS_MMAP,
+			0,
+			length,
+			uintptr(at.Prot()),
+			syscall.MAP_SHARED,
+			uintptr(fd),
+			uintptr(offset+gapOffset))
+		if errno != 0 {
+			return nil, errno
+		}
+
+		// Insert into the host set and move to the next gap.
+		gap = hm.set.Insert(gap, gap.Range().Intersect(r), hostAddr).NextGap()
+	}
+
+	// Collect all slices.
+	hm.forEachEntry(r, func(_ uint64, m hostMapEntry) {
+		ms = append(ms, m)
+	})
+
+	return ms, nil
+}
+
+// CreateMappings creates a new set of host mapping entries.
+func (hm *hostMap) CreateMappings(r usermem.AddrRange, at usermem.AccessType, fd int, offset uint64) (ms []hostMapEntry, err error) {
+	hm.mu.Lock()
+	ms, err = hm.createMappings(r, at, fd, offset)
+	hm.mu.Unlock()
+	return
+}
+
+func (hm *hostMap) deleteMapping(r usermem.AddrRange) {
+	// Remove all the existing mappings.
+	hm.forEachEntry(r, func(_ uint64, m hostMapEntry) {
+		_, _, errno := syscall.RawSyscall(
+			syscall.SYS_MUNMAP,
+			m.addr,
+			m.length,
+			0)
+		if errno != 0 {
+			// Should never happen.
+			panic(fmt.Sprintf("unmap error: %v", errno))
+		}
+	})
+
+	// Knock the range out.
+	hm.set.RemoveRange(r)
+}
+
+// DeleteMapping deletes the given range.
+func (hm *hostMap) DeleteMapping(r usermem.AddrRange) {
+	hm.mu.Lock()
+	hm.deleteMapping(r)
+	hm.mu.Unlock()
+}
+
+// hostMapSetFunctions is used in the implementation of mapSet.
+type hostMapSetFunctions struct{}
+
+func (hostMapSetFunctions) MinKey() usermem.Addr    { return 0 }
+func (hostMapSetFunctions) MaxKey() usermem.Addr    { return ^usermem.Addr(0) }
+func (hostMapSetFunctions) ClearValue(val *uintptr) { *val = 0 }
+
+func (hostMapSetFunctions) Merge(r1 usermem.AddrRange, addr1 uintptr, r2 usermem.AddrRange, addr2 uintptr) (uintptr, bool) {
+	if addr1+uintptr(r1.Length()) != addr2 {
+		return 0, false
+	}
+
+	// Since the two regions are contiguous in both the key space and the
+	// value space, we can just store a single segment with the first host
+	// virtual address; the logic above operates based on the size of the
+	// segments.
+	return addr1, true
+}
+
+func (hostMapSetFunctions) Split(r usermem.AddrRange, hostAddr uintptr, split usermem.Addr) (uintptr, uintptr) {
+	return hostAddr, hostAddr + uintptr(split-r.Start)
+}
diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go
new file mode 100644
index 000000000..31928c9f0
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm.go
@@ -0,0 +1,149 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package kvm provides a kvm-based implementation of the platform interface.
+package kvm
+
+import (
+	"fmt"
+	"runtime"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/filemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// KVM represents a lightweight VM context.
+type KVM struct {
+	platform.NoCPUPreemptionDetection
+
+	// filemem is our memory source.
+	*filemem.FileMem
+
+	// machine is the backing VM.
+	machine *machine
+}
+
+var (
+	globalOnce sync.Once
+	globalErr  error
+)
+
+// New returns a new KVM-based implementation of the platform interface.
+func New() (*KVM, error) {
+	// Allocate physical memory for the vCPUs.
+	fm, err := filemem.New("kvm-memory")
+	if err != nil {
+		return nil, err
+	}
+
+	// Try opening KVM.
+	fd, err := syscall.Open("/dev/kvm", syscall.O_RDWR, 0)
+	if err != nil {
+		return nil, fmt.Errorf("opening /dev/kvm: %v", err)
+	}
+	defer syscall.Close(fd)
+
+	// Ensure global initialization is done.
+	globalOnce.Do(func() {
+		physicalInit()
+		globalErr = updateSystemValues(fd)
+		ring0.Init(cpuid.HostFeatureSet())
+	})
+	if globalErr != nil {
+		return nil, err
+	}
+
+	// Create a new VM fd.
+	vm, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(fd), _KVM_CREATE_VM, 0)
+	if errno != 0 {
+		return nil, fmt.Errorf("creating VM: %v", errno)
+	}
+
+	// Create a VM context.
+	machine, err := newMachine(int(vm), runtime.NumCPU())
+	if err != nil {
+		return nil, err
+	}
+
+	// All set.
+	return &KVM{
+		FileMem: fm,
+		machine: machine,
+	}, nil
+}
+
+// SupportsAddressSpaceIO implements platform.Platform.SupportsAddressSpaceIO.
+func (*KVM) SupportsAddressSpaceIO() bool {
+	return false
+}
+
+// CooperativelySchedulesAddressSpace implements platform.Platform.CooperativelySchedulesAddressSpace.
+func (*KVM) CooperativelySchedulesAddressSpace() bool {
+	return false
+}
+
+// MapUnit implements platform.Platform.MapUnit.
+func (*KVM) MapUnit() uint64 {
+	// We greedily creates PTEs in MapFile, so extremely large mappings can
+	// be expensive. Not _that_ expensive since we allow super pages, but
+	// even though can get out of hand if you're creating multi-terabyte
+	// mappings. For this reason, we limit mappings to an arbitrary 16MB.
+	return 16 << 20
+}
+
+// MinUserAddress returns the lowest available address.
+func (*KVM) MinUserAddress() usermem.Addr {
+	return usermem.PageSize
+}
+
+// MaxUserAddress returns the first address that may not be used.
+func (*KVM) MaxUserAddress() usermem.Addr {
+	return usermem.Addr(ring0.MaximumUserAddress)
+}
+
+// NewAddressSpace returns a new pagetable root.
+func (k *KVM) NewAddressSpace(_ interface{}) (platform.AddressSpace, <-chan struct{}, error) {
+	// Allocate page tables and install system mappings.
+	pageTables := k.machine.kernel.PageTables.New()
+	applyPhysicalRegions(func(pr physicalRegion) bool {
+		// Map the kernel in the upper half.
+		kernelVirtual := usermem.Addr(ring0.KernelStartAddress | pr.virtual)
+		pageTables.Map(kernelVirtual, pr.length, false /* kernel */, usermem.AnyAccess, pr.physical)
+		return true // Keep iterating.
+	})
+
+	// Return the new address space.
+	return &addressSpace{
+		filemem:    k.FileMem,
+		machine:    k.machine,
+		pageTables: pageTables,
+	}, nil, nil
+}
+
+// NewContext returns an interruptible context.
+func (k *KVM) NewContext() platform.Context {
+	return &context{
+		machine: k.machine,
+	}
+}
+
+// Memory returns the platform memory used to do allocations.
+func (k *KVM) Memory() platform.Memory {
+	return k.FileMem
+}
diff --git a/pkg/sentry/platform/kvm/kvm_amd64.go b/pkg/sentry/platform/kvm/kvm_amd64.go
new file mode 100644
index 000000000..3d56ed895
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm_amd64.go
@@ -0,0 +1,213 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+)
+
+// userMemoryRegion is a region of physical memory.
+//
+// This mirrors kvm_memory_region.
+type userMemoryRegion struct {
+	slot          uint32
+	flags         uint32
+	guestPhysAddr uint64
+	memorySize    uint64
+	userspaceAddr uint64
+}
+
+// userRegs represents KVM user registers.
+//
+// This mirrors kvm_regs.
+type userRegs struct {
+	RAX    uint64
+	RBX    uint64
+	RCX    uint64
+	RDX    uint64
+	RSI    uint64
+	RDI    uint64
+	RSP    uint64
+	RBP    uint64
+	R8     uint64
+	R9     uint64
+	R10    uint64
+	R11    uint64
+	R12    uint64
+	R13    uint64
+	R14    uint64
+	R15    uint64
+	RIP    uint64
+	RFLAGS uint64
+}
+
+// systemRegs represents KVM system registers.
+//
+// This mirrors kvm_sregs.
+type systemRegs struct {
+	CS              segment
+	DS              segment
+	ES              segment
+	FS              segment
+	GS              segment
+	SS              segment
+	TR              segment
+	LDT             segment
+	GDT             descriptor
+	IDT             descriptor
+	CR0             uint64
+	CR2             uint64
+	CR3             uint64
+	CR4             uint64
+	CR8             uint64
+	EFER            uint64
+	apicBase        uint64
+	interruptBitmap [(_KVM_NR_INTERRUPTS + 63) / 64]uint64
+}
+
+// segment is the expanded form of a segment register.
+//
+// This mirrors kvm_segment.
+type segment struct {
+	base     uint64
+	limit    uint32
+	selector uint16
+	typ      uint8
+	present  uint8
+	DPL      uint8
+	DB       uint8
+	S        uint8
+	L        uint8
+	G        uint8
+	AVL      uint8
+	unusable uint8
+	_        uint8
+}
+
+// Clear clears the segment and marks it unusable.
+func (s *segment) Clear() {
+	*s = segment{unusable: 1}
+}
+
+// selector is a segment selector.
+type selector uint16
+
+// tobool is a simple helper.
+func tobool(x ring0.SegmentDescriptorFlags) uint8 {
+	if x != 0 {
+		return 1
+	}
+	return 0
+}
+
+// Load loads the segment described by d into the segment s.
+//
+// The argument sel is recorded as the segment selector index.
+func (s *segment) Load(d *ring0.SegmentDescriptor, sel ring0.Selector) {
+	flag := d.Flags()
+	if flag&ring0.SegmentDescriptorPresent == 0 {
+		s.Clear()
+		return
+	}
+	s.base = uint64(d.Base())
+	s.limit = d.Limit()
+	s.typ = uint8((flag>>8)&0xF) | 1
+	s.S = tobool(flag & ring0.SegmentDescriptorSystem)
+	s.DPL = uint8(d.DPL())
+	s.present = tobool(flag & ring0.SegmentDescriptorPresent)
+	s.AVL = tobool(flag & ring0.SegmentDescriptorAVL)
+	s.L = tobool(flag & ring0.SegmentDescriptorLong)
+	s.DB = tobool(flag & ring0.SegmentDescriptorDB)
+	s.G = tobool(flag & ring0.SegmentDescriptorG)
+	if s.L != 0 {
+		s.limit = 0xffffffff
+	}
+	s.unusable = 0
+	s.selector = uint16(sel)
+}
+
+// descriptor describes a region of physical memory.
+//
+// It corresponds to the pseudo-descriptor used in the x86 LGDT and LIDT
+// instructions, and mirrors kvm_dtable.
+type descriptor struct {
+	base  uint64
+	limit uint16
+	_     [3]uint16
+}
+
+// modelControlRegister is an MSR entry.
+//
+// This mirrors kvm_msr_entry.
+type modelControlRegister struct {
+	index uint32
+	_     uint32
+	data  uint64
+}
+
+// modelControlRegisers is a collection of MSRs.
+//
+// This mirrors kvm_msrs.
+type modelControlRegisters struct {
+	nmsrs   uint32
+	_       uint32
+	entries [16]modelControlRegister
+}
+
+// runData is the run structure. This may be mapped for synchronous register
+// access (although that doesn't appear to be supported by my kernel at least).
+//
+// This mirrors kvm_run.
+type runData struct {
+	requestInterruptWindow uint8
+	_                      [7]uint8
+
+	exitReason                 uint32
+	readyForInterruptInjection uint8
+	ifFlag                     uint8
+	_                          [2]uint8
+
+	cr8      uint64
+	apicBase uint64
+
+	// This is the union data for exits. Interpretation depends entirely on
+	// the exitReason above (see vCPU code for more information).
+	data [32]uint64
+}
+
+// cpuidEntry is a single CPUID entry.
+//
+// This mirrors kvm_cpuid_entry2.
+type cpuidEntry struct {
+	function uint32
+	index    uint32
+	flags    uint32
+	eax      uint32
+	ebx      uint32
+	ecx      uint32
+	edx      uint32
+	_        [3]uint32
+}
+
+// cpuidEntries is a collection of CPUID entries.
+//
+// This mirrors kvm_cpuid2.
+type cpuidEntries struct {
+	nr      uint32
+	_       uint32
+	entries [_KVM_NR_CPUID_ENTRIES]cpuidEntry
+}
diff --git a/pkg/sentry/platform/kvm/kvm_amd64_unsafe.go b/pkg/sentry/platform/kvm/kvm_amd64_unsafe.go
new file mode 100644
index 000000000..389412d87
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm_amd64_unsafe.go
@@ -0,0 +1,93 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"fmt"
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+)
+
+var (
+	runDataSize     int
+	hasGuestPCID    bool
+	hasGuestINVPCID bool
+	pagetablesOpts  pagetables.Opts
+	cpuidSupported  = cpuidEntries{nr: _KVM_NR_CPUID_ENTRIES}
+)
+
+func updateSystemValues(fd int) error {
+	// Extract the mmap size.
+	sz, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(fd), _KVM_GET_VCPU_MMAP_SIZE, 0)
+	if errno != 0 {
+		return fmt.Errorf("getting VCPU mmap size: %v", errno)
+	}
+
+	// Save the data.
+	runDataSize = int(sz)
+
+	// Must do the dance to figure out the number of entries.
+	_, _, errno = syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(fd),
+		_KVM_GET_SUPPORTED_CPUID,
+		uintptr(unsafe.Pointer(&cpuidSupported)))
+	if errno != 0 && errno != syscall.ENOMEM {
+		// Some other error occurred.
+		return fmt.Errorf("getting supported CPUID: %v", errno)
+	}
+
+	// The number should now be correct.
+	_, _, errno = syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(fd),
+		_KVM_GET_SUPPORTED_CPUID,
+		uintptr(unsafe.Pointer(&cpuidSupported)))
+	if errno != 0 {
+		// Didn't work with the right number.
+		return fmt.Errorf("getting supported CPUID (2nd attempt): %v", errno)
+	}
+
+	// Calculate whether guestPCID is supported.
+	//
+	// FIXME: These should go through the much more pleasant
+	// cpuid package interfaces, once a way to accept raw kvm CPUID entries
+	// is plumbed (or some rough equivalent).
+	for i := 0; i < int(cpuidSupported.nr); i++ {
+		entry := cpuidSupported.entries[i]
+		if entry.function == 1 && entry.index == 0 && entry.ecx&(1<<17) != 0 {
+			hasGuestPCID = true // Found matching PCID in guest feature set.
+		}
+		if entry.function == 7 && entry.index == 0 && entry.ebx&(1<<10) != 0 {
+			hasGuestINVPCID = true // Found matching INVPCID in guest feature set.
+		}
+	}
+
+	// A basic sanity check: ensure that we don't attempt to
+	// invpcid if guest PCIDs are not supported; it's not clear
+	// what the semantics of this would be (or why some CPU or
+	// hypervisor would export this particular combination).
+	hasGuestINVPCID = hasGuestPCID && hasGuestINVPCID
+
+	// Set the pagetables to use PCID if it's available.
+	pagetablesOpts.EnablePCID = hasGuestPCID
+
+	// Success.
+	return nil
+}
diff --git a/pkg/sentry/platform/kvm/kvm_const.go b/pkg/sentry/platform/kvm/kvm_const.go
new file mode 100644
index 000000000..0ec6a4a00
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm_const.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+// KVM ioctls.
+//
+// Only the ioctls we need in Go appear here; some additional ioctls are used
+// within the assembly stubs (KVM_INTERRUPT, etc.).
+const (
+	_KVM_CREATE_VM              = 0xae01
+	_KVM_GET_VCPU_MMAP_SIZE     = 0xae04
+	_KVM_CREATE_VCPU            = 0xae41
+	_KVM_SET_TSS_ADDR           = 0xae47
+	_KVM_RUN                    = 0xae80
+	_KVM_INTERRUPT              = 0x4004ae86
+	_KVM_SET_MSRS               = 0x4008ae89
+	_KVM_SET_USER_MEMORY_REGION = 0x4020ae46
+	_KVM_SET_REGS               = 0x4090ae82
+	_KVM_SET_SREGS              = 0x4138ae84
+	_KVM_GET_SUPPORTED_CPUID    = 0xc008ae05
+	_KVM_SET_CPUID2             = 0x4008ae90
+	_KVM_SET_SIGNAL_MASK        = 0x4004ae8b
+)
+
+// KVM exit reasons.
+const (
+	_KVM_EXIT_EXCEPTION       = 0x1
+	_KVM_EXIT_IO              = 0x2
+	_KVM_EXIT_HYPERCALL       = 0x3
+	_KVM_EXIT_DEBUG           = 0x4
+	_KVM_EXIT_HLT             = 0x5
+	_KVM_EXIT_MMIO            = 0x6
+	_KVM_EXIT_IRQ_WINDOW_OPEN = 0x7
+	_KVM_EXIT_SHUTDOWN        = 0x8
+	_KVM_EXIT_FAIL_ENTRY      = 0x9
+	_KVM_EXIT_INTERNAL_ERROR  = 0x11
+)
+
+// KVM limits.
+const (
+	_KVM_NR_VCPUS         = 0x100
+	_KVM_NR_INTERRUPTS    = 0x100
+	_KVM_NR_CPUID_ENTRIES = 0x100
+)
diff --git a/pkg/sentry/platform/kvm/kvm_test.go b/pkg/sentry/platform/kvm/kvm_test.go
new file mode 100644
index 000000000..61cfdd8fd
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm_test.go
@@ -0,0 +1,415 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"math/rand"
+	"reflect"
+	"syscall"
+	"testing"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/kvm/testutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+var dummyFPState = (*byte)(arch.NewFloatingPointData())
+
+type testHarness interface {
+	Errorf(format string, args ...interface{})
+	Fatalf(format string, args ...interface{})
+}
+
+func kvmTest(t testHarness, setup func(*KVM), fn func(*vCPU) bool) {
+	// Create the machine.
+	k, err := New()
+	if err != nil {
+		t.Fatalf("error creating KVM instance: %v", err)
+	}
+	defer k.machine.Destroy()
+	defer k.FileMem.Destroy()
+
+	// Call additional setup.
+	if setup != nil {
+		setup(k)
+	}
+
+	var c *vCPU // For recovery.
+	defer func() {
+		redpill()
+		if c != nil {
+			k.machine.Put(c)
+		}
+	}()
+	for {
+		c, err = k.machine.Get()
+		if err != nil {
+			t.Fatalf("error getting vCPU: %v", err)
+		}
+		if !fn(c) {
+			break
+		}
+
+		// We put the vCPU here and clear the value so that the
+		// deferred recovery will not re-put it above.
+		k.machine.Put(c)
+		c = nil
+	}
+}
+
+func bluepillTest(t testHarness, fn func(*vCPU)) {
+	kvmTest(t, nil, func(c *vCPU) bool {
+		bluepill(c)
+		fn(c)
+		return false
+	})
+}
+
+func TestKernelSyscall(t *testing.T) {
+	bluepillTest(t, func(c *vCPU) {
+		redpill() // Leave guest mode.
+		if got := c.State(); got != vCPUReady {
+			t.Errorf("vCPU not in ready state: got %v", got)
+		}
+	})
+}
+
+func hostFault() {
+	defer func() {
+		recover()
+	}()
+	var foo *int
+	*foo = 0
+}
+
+func TestKernelFault(t *testing.T) {
+	hostFault() // Ensure recovery works.
+	bluepillTest(t, func(c *vCPU) {
+		hostFault()
+		if got := c.State(); got != vCPUReady {
+			t.Errorf("vCPU not in ready state: got %v", got)
+		}
+	})
+}
+
+func TestKernelFloatingPoint(t *testing.T) {
+	bluepillTest(t, func(c *vCPU) {
+		if !testutil.FloatingPointWorks() {
+			t.Errorf("floating point does not work, and it should!")
+		}
+	})
+}
+
+func applicationTest(t testHarness, useHostMappings bool, target func(), fn func(*vCPU, *syscall.PtraceRegs, *pagetables.PageTables) bool) {
+	// Initialize registers & page tables.
+	var (
+		regs syscall.PtraceRegs
+		pt   *pagetables.PageTables
+	)
+	testutil.SetTestTarget(&regs, target)
+	defer func() {
+		if pt != nil {
+			pt.Release()
+		}
+	}()
+
+	kvmTest(t, func(k *KVM) {
+		// Create new page tables.
+		as, _, err := k.NewAddressSpace(nil /* invalidator */)
+		if err != nil {
+			t.Fatalf("can't create new address space: %v", err)
+		}
+		pt = as.(*addressSpace).pageTables
+
+		if useHostMappings {
+			// Apply the physical mappings to these page tables.
+			// (This is normally dangerous, since they point to
+			// physical pages that may not exist. This shouldn't be
+			// done for regular user code, but is fine for test
+			// purposes.)
+			applyPhysicalRegions(func(pr physicalRegion) bool {
+				pt.Map(usermem.Addr(pr.virtual), pr.length, true /* user */, usermem.AnyAccess, pr.physical)
+				return true // Keep iterating.
+			})
+		}
+	}, func(c *vCPU) bool {
+		// Invoke the function with the extra data.
+		return fn(c, &regs, pt)
+	})
+}
+
+func TestApplicationSyscall(t *testing.T) {
+	applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); err != nil {
+			t.Errorf("application syscall with full restore failed: %v", err)
+		}
+		return false
+	})
+	applicationTest(t, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != nil {
+			t.Errorf("application syscall with partial restore failed: %v", err)
+		}
+		return false
+	})
+}
+
+func TestApplicationFault(t *testing.T) {
+	applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTouchTarget(regs, nil) // Cause fault.
+		if si, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); err != platform.ErrContextSignal || (si != nil && si.Signo != int32(syscall.SIGSEGV)) {
+			t.Errorf("application fault with full restore got (%v, %v), expected (%v, SIGSEGV)", err, si, platform.ErrContextSignal)
+		}
+		return false
+	})
+	applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTouchTarget(regs, nil) // Cause fault.
+		if si, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != platform.ErrContextSignal || (si != nil && si.Signo != int32(syscall.SIGSEGV)) {
+			t.Errorf("application fault with partial restore got (%v, %v), expected (%v, SIGSEGV)", err, si, platform.ErrContextSignal)
+		}
+		return false
+	})
+}
+
+func TestRegistersSyscall(t *testing.T) {
+	applicationTest(t, true, testutil.TwiddleRegsSyscall, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTestRegs(regs) // Fill values for all registers.
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != nil {
+			t.Errorf("application register check with partial restore got unexpected error: %v", err)
+		}
+		if err := testutil.CheckTestRegs(regs, false); err != nil {
+			t.Errorf("application register check with partial restore failed: %v", err)
+		}
+		return false
+	})
+}
+
+func TestRegistersFault(t *testing.T) {
+	applicationTest(t, true, testutil.TwiddleRegsFault, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTestRegs(regs) // Fill values for all registers.
+		if si, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); err != platform.ErrContextSignal || si.Signo != int32(syscall.SIGSEGV) {
+			t.Errorf("application register check with full restore got unexpected error: %v", err)
+		}
+		if err := testutil.CheckTestRegs(regs, true); err != nil {
+			t.Errorf("application register check with full restore failed: %v", err)
+		}
+		return false
+	})
+}
+
+func TestSegments(t *testing.T) {
+	applicationTest(t, true, testutil.TwiddleSegments, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTestSegments(regs)
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); err != nil {
+			t.Errorf("application segment check with full restore got unexpected error: %v", err)
+		}
+		if err := testutil.CheckTestSegments(regs); err != nil {
+			t.Errorf("application segment check with full restore failed: %v", err)
+		}
+		return false
+	})
+}
+
+func TestBounce(t *testing.T) {
+	applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		go func() {
+			time.Sleep(time.Millisecond)
+			c.Bounce()
+		}()
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != platform.ErrContextInterrupt {
+			t.Errorf("application partial restore: got %v, wanted %v", err, platform.ErrContextInterrupt)
+		}
+		return false
+	})
+	applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		go func() {
+			time.Sleep(time.Millisecond)
+			c.Bounce()
+		}()
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); err != platform.ErrContextInterrupt {
+			t.Errorf("application full restore: got %v, wanted %v", err, platform.ErrContextInterrupt)
+		}
+		return false
+	})
+}
+
+func TestBounceStress(t *testing.T) {
+	applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		randomSleep := func() {
+			// O(hundreds of microseconds) is appropriate to ensure
+			// different overlaps and different schedules.
+			if n := rand.Intn(1000); n > 100 {
+				time.Sleep(time.Duration(n) * time.Microsecond)
+			}
+		}
+		for i := 0; i < 1000; i++ {
+			// Start an asynchronously executing goroutine that
+			// calls Bounce at pseudo-random point in time.
+			// This should wind up calling Bounce when the
+			// kernel is in various stages of the switch.
+			go func() {
+				randomSleep()
+				c.Bounce()
+			}()
+			randomSleep()
+			// Execute the switch.
+			if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != platform.ErrContextInterrupt {
+				t.Errorf("application partial restore: got %v, wanted %v", err, platform.ErrContextInterrupt)
+			}
+			// Simulate work.
+			c.Unlock()
+			randomSleep()
+			c.Lock()
+		}
+		return false
+	})
+}
+
+func TestInvalidate(t *testing.T) {
+	var data uintptr // Used below.
+	applicationTest(t, true, testutil.Touch, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		testutil.SetTouchTarget(regs, &data) // Read legitimate value.
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != nil {
+			t.Errorf("application partial restore: got %v, wanted nil", err)
+		}
+		// Unmap the page containing data & invalidate.
+		pt.Unmap(usermem.Addr(reflect.ValueOf(&data).Pointer() & ^uintptr(usermem.PageSize-1)), usermem.PageSize)
+		c.Invalidate() // Ensure invalidation.
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != platform.ErrContextSignal {
+			t.Errorf("application partial restore: got %v, wanted %v", err, platform.ErrContextSignal)
+		}
+		return false
+	})
+}
+
+// IsFault returns true iff the given signal represents a fault.
+func IsFault(err error, si *arch.SignalInfo) bool {
+	return err == platform.ErrContextSignal && si.Signo == int32(syscall.SIGSEGV)
+}
+
+func TestEmptyAddressSpace(t *testing.T) {
+	applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if si, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); !IsFault(err, si) {
+			t.Errorf("first fault with partial restore failed got %v", err)
+			t.Logf("registers: %#v", &regs)
+		}
+		return false
+	})
+	applicationTest(t, false, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if si, _, err := c.SwitchToUser(regs, dummyFPState, pt, ring0.FlagFull); !IsFault(err, si) {
+			t.Errorf("first fault with full restore failed got %v", err)
+			t.Logf("registers: %#v", &regs)
+		}
+		return false
+	})
+}
+
+func TestWrongVCPU(t *testing.T) {
+	kvmTest(t, nil, func(c1 *vCPU) bool {
+		kvmTest(t, nil, func(c2 *vCPU) bool {
+			// Basic test, one then the other.
+			bluepill(c1)
+			bluepill(c2)
+			if c2.switches == 0 {
+				// Don't allow the test to proceed if this fails.
+				t.Fatalf("wrong vCPU#2 switches: vCPU1=%+v,vCPU2=%+v", c1, c2)
+			}
+
+			// Alternate vCPUs; we expect to need to trigger the
+			// wrong vCPU path on each switch.
+			for i := 0; i < 100; i++ {
+				bluepill(c1)
+				bluepill(c2)
+			}
+			if count := c1.switches; count < 90 {
+				t.Errorf("wrong vCPU#1 switches: vCPU1=%+v,vCPU2=%+v", c1, c2)
+			}
+			if count := c2.switches; count < 90 {
+				t.Errorf("wrong vCPU#2 switches: vCPU1=%+v,vCPU2=%+v", c1, c2)
+			}
+			return false
+		})
+		return false
+	})
+	kvmTest(t, nil, func(c1 *vCPU) bool {
+		kvmTest(t, nil, func(c2 *vCPU) bool {
+			bluepill(c1)
+			bluepill(c2)
+			return false
+		})
+		return false
+	})
+}
+
+func BenchmarkApplicationSyscall(b *testing.B) {
+	var (
+		i int // Iteration includes machine.Get() / machine.Put().
+		a int // Count for ErrContextInterrupt.
+	)
+	applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != nil {
+			if err == platform.ErrContextInterrupt {
+				a++
+				return true // Ignore.
+			}
+			b.Fatalf("benchmark failed: %v", err)
+		}
+		i++
+		return i < b.N
+	})
+	if a != 0 {
+		b.Logf("ErrContextInterrupt occurred %d times (in %d iterations).", a, a+i)
+	}
+}
+
+func BenchmarkKernelSyscall(b *testing.B) {
+	// Note that the target passed here is irrelevant, we never execute SwitchToUser.
+	applicationTest(b, true, testutil.Getpid, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		// iteration does not include machine.Get() / machine.Put().
+		for i := 0; i < b.N; i++ {
+			testutil.Getpid()
+		}
+		return false
+	})
+}
+
+func BenchmarkWorldSwitchToUserRoundtrip(b *testing.B) {
+	// see BenchmarkApplicationSyscall.
+	var (
+		i int
+		a int
+	)
+	applicationTest(b, true, testutil.SyscallLoop, func(c *vCPU, regs *syscall.PtraceRegs, pt *pagetables.PageTables) bool {
+		if _, _, err := c.SwitchToUser(regs, dummyFPState, pt, 0); err != nil {
+			if err == platform.ErrContextInterrupt {
+				a++
+				return true // Ignore.
+			}
+			b.Fatalf("benchmark failed: %v", err)
+		}
+		// This will intentionally cause the world switch. By executing
+		// a host syscall here, we force the transition between guest
+		// and host mode.
+		testutil.Getpid()
+		i++
+		return i < b.N
+	})
+	if a != 0 {
+		b.Logf("EAGAIN occurred %d times (in %d iterations).", a, a+i)
+	}
+}
diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go
new file mode 100644
index 000000000..a5be0cee3
--- /dev/null
+++ b/pkg/sentry/platform/kvm/machine.go
@@ -0,0 +1,412 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"fmt"
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/tmutex"
+)
+
+// machine contains state associated with the VM as a whole.
+type machine struct {
+	// fd is the vm fd.
+	fd int
+
+	// nextSlot is the next slot for setMemoryRegion.
+	//
+	// This must be accessed atomically. If nextSlot is ^uint32(0), then
+	// slots are currently being updated, and the caller should retry.
+	nextSlot uint32
+
+	// kernel is the set of global structures.
+	kernel *ring0.Kernel
+
+	// mappingCache is used for mapPhysical.
+	mappingCache sync.Map
+
+	// mu protects vCPUs.
+	mu sync.Mutex
+
+	// vCPUs are the machine vCPUs.
+	//
+	// This is eventually keyed by system TID, but is initially indexed by
+	// the negative vCPU id. This is merely an optimization, so while
+	// collisions here are not possible, it wouldn't matter anyways.
+	vCPUs map[uint64]*vCPU
+}
+
+const (
+	// vCPUReady is the lock value for an available vCPU.
+	//
+	// Legal transitions: vCPUGuest (bluepill).
+	vCPUReady uintptr = iota
+
+	// vCPUGuest indicates the vCPU is in guest mode.
+	//
+	// Legal transition: vCPUReady (bluepill), vCPUWaiter (wait).
+	vCPUGuest
+
+	// vCPUWaiter indicates that the vCPU should be released.
+	//
+	// Legal transition: vCPUReady (bluepill).
+	vCPUWaiter
+)
+
+// vCPU is a single KVM vCPU.
+type vCPU struct {
+	// CPU is the kernel CPU data.
+	//
+	// This must be the first element of this structure, it is referenced
+	// by the bluepill code (see bluepill_amd64.s).
+	ring0.CPU
+
+	// fd is the vCPU fd.
+	fd int
+
+	// tid is the last set tid.
+	tid uint64
+
+	// switches is a count of world switches (informational only).
+	switches uint32
+
+	// faults is a count of world faults (informational only).
+	faults uint32
+
+	// state is the vCPU state; all are described above.
+	state uintptr
+
+	// runData for this vCPU.
+	runData *runData
+
+	// machine associated with this vCPU.
+	machine *machine
+
+	// mu applies across get/put; it does not protect the above.
+	mu tmutex.Mutex
+}
+
+// newMachine returns a new VM context.
+func newMachine(vm int, vCPUs int) (*machine, error) {
+	// Create the machine.
+	m := &machine{
+		fd:    vm,
+		vCPUs: make(map[uint64]*vCPU),
+	}
+	if vCPUs > _KVM_NR_VCPUS {
+		// Hard cap at KVM's limit.
+		vCPUs = _KVM_NR_VCPUS
+	}
+	if n := 2 * runtime.NumCPU(); vCPUs > n {
+		// Cap at twice the number of physical cores. Otherwise we're
+		// just wasting memory and thrashing. (There may be scheduling
+		// issues when you've got > n active threads.)
+		vCPUs = n
+	}
+	m.kernel = ring0.New(ring0.KernelOpts{
+		PageTables: pagetables.New(m, pagetablesOpts),
+	})
+
+	// Initialize architecture state.
+	if err := m.initArchState(vCPUs); err != nil {
+		m.Destroy()
+		return nil, err
+	}
+
+	// Create all the vCPUs.
+	for id := 0; id < vCPUs; id++ {
+		// Create the vCPU.
+		fd, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(vm), _KVM_CREATE_VCPU, uintptr(id))
+		if errno != 0 {
+			m.Destroy()
+			return nil, fmt.Errorf("error creating VCPU: %v", errno)
+		}
+		c := &vCPU{
+			fd:      int(fd),
+			machine: m,
+		}
+		c.mu.Init()
+		c.CPU.Init(m.kernel)
+		c.CPU.KernelSyscall = bluepillSyscall
+		c.CPU.KernelException = bluepillException
+		m.vCPUs[uint64(-id)] = c // See above.
+
+		// Ensure the signal mask is correct.
+		if err := c.setSignalMask(); err != nil {
+			m.Destroy()
+			return nil, err
+		}
+
+		// Initialize architecture state.
+		if err := c.initArchState(); err != nil {
+			m.Destroy()
+			return nil, err
+		}
+
+		// Map the run data.
+		runData, err := mapRunData(int(fd))
+		if err != nil {
+			m.Destroy()
+			return nil, err
+		}
+		c.runData = runData
+	}
+
+	// Apply the physical mappings. Note that these mappings may point to
+	// guest physical addresses that are not actually available. These
+	// physical pages are mapped on demand, see kernel_unsafe.go.
+	applyPhysicalRegions(func(pr physicalRegion) bool {
+		// Map everything in the lower half.
+		m.kernel.PageTables.Map(usermem.Addr(pr.virtual), pr.length, false /* kernel */, usermem.AnyAccess, pr.physical)
+		// And keep everything in the upper half.
+		kernelAddr := usermem.Addr(ring0.KernelStartAddress | pr.virtual)
+		m.kernel.PageTables.Map(kernelAddr, pr.length, false /* kernel */, usermem.AnyAccess, pr.physical)
+		return true // Keep iterating.
+	})
+
+	// Ensure that the currently mapped virtual regions are actually
+	// available in the VM. Note that this doesn't guarantee no future
+	// faults, however it should guarantee that everything is available to
+	// ensure successful vCPU entry.
+	applyVirtualRegions(func(vr virtualRegion) {
+		if excludeVirtualRegion(vr) {
+			return // skip region.
+		}
+		for virtual := vr.virtual; virtual < vr.virtual+vr.length; {
+			physical, length, ok := TranslateToPhysical(virtual)
+			if !ok {
+				// This must be an invalid region that was
+				// knocked out by creation of the physical map.
+				return
+			}
+			if virtual+length > vr.virtual+vr.length {
+				// Cap the length to the end of the area.
+				length = vr.virtual + vr.length - virtual
+			}
+
+			// Ensure the physical range is mapped.
+			m.mapPhysical(physical, length)
+			virtual += length
+		}
+	})
+
+	// Ensure the machine is cleaned up properly.
+	runtime.SetFinalizer(m, (*machine).Destroy)
+	return m, nil
+}
+
+// mapPhysical checks for the mapping of a physical range, and installs one if
+// not available. This attempts to be efficient for calls in the hot path.
+//
+// This panics on error.
+func (m *machine) mapPhysical(physical, length uintptr) {
+	for end := physical + length; physical < end; {
+		_, physicalStart, length, ok := calculateBluepillFault(m, physical)
+		if !ok {
+			// Should never happen.
+			panic("mapPhysical on unknown physical address")
+		}
+
+		if _, ok := m.mappingCache.LoadOrStore(physicalStart, true); !ok {
+			// Not present in the cache; requires setting the slot.
+			if _, ok := handleBluepillFault(m, physical); !ok {
+				panic("handleBluepillFault failed")
+			}
+		}
+
+		// Move to the next chunk.
+		physical = physicalStart + length
+	}
+}
+
+// Destroy frees associated resources.
+//
+// Destroy should only be called once all active users of the machine are gone.
+// The machine object should not be used after calling Destroy.
+//
+// Precondition: all vCPUs must be returned to the machine.
+func (m *machine) Destroy() {
+	runtime.SetFinalizer(m, nil)
+
+	// Destroy vCPUs.
+	for _, c := range m.vCPUs {
+		// Ensure the vCPU is not still running in guest mode. This is
+		// possible iff teardown has been done by other threads, and
+		// somehow a single thread has not executed any system calls.
+		c.wait()
+
+		// Teardown the vCPU itself.
+		switch state := c.State(); state {
+		case vCPUReady:
+			// Note that the runData may not be mapped if an error
+			// occurs during the middle of initialization.
+			if c.runData != nil {
+				if err := unmapRunData(c.runData); err != nil {
+					panic(fmt.Sprintf("error unmapping rundata: %v", err))
+				}
+			}
+			if err := syscall.Close(int(c.fd)); err != nil {
+				panic(fmt.Sprintf("error closing vCPU fd: %v", err))
+			}
+		case vCPUGuest, vCPUWaiter:
+			// Should never happen; waited above.
+			panic("vCPU disposed in guest state")
+		default:
+			// Should never happen; not a valid state.
+			panic(fmt.Sprintf("vCPU in invalid state: %v", state))
+		}
+	}
+
+	// Release host mappings.
+	if m.kernel.PageTables != nil {
+		m.kernel.PageTables.Release()
+	}
+
+	// vCPUs are gone: teardown machine state.
+	if err := syscall.Close(m.fd); err != nil {
+		panic(fmt.Sprintf("error closing VM fd: %v", err))
+	}
+}
+
+// Get gets an available vCPU.
+func (m *machine) Get() (*vCPU, error) {
+	runtime.LockOSThread()
+	tid := procid.Current()
+	m.mu.Lock()
+
+	for {
+		// Check for an exact match.
+		if c := m.vCPUs[tid]; c != nil && c.mu.TryLock() {
+			m.mu.Unlock()
+			return c, nil
+		}
+
+		// Scan for an available vCPU.
+		for origTID, c := range m.vCPUs {
+			if c.LockInState(vCPUReady) {
+				delete(m.vCPUs, origTID)
+				m.vCPUs[tid] = c
+				m.mu.Unlock()
+
+				// We need to reload thread-local segments as
+				// we have origTID != tid and the vCPU state
+				// may be stale.
+				c.loadSegments()
+				atomic.StoreUint64(&c.tid, tid)
+				return c, nil
+			}
+		}
+
+		// Everything is busy executing user code (locked).
+		//
+		// We hold the pool lock here, so we should be able to kick something
+		// out of kernel mode and have it bounce into host mode when it tries
+		// to grab the vCPU again.
+		for _, c := range m.vCPUs {
+			if c.State() != vCPUWaiter {
+				c.Bounce()
+			}
+		}
+
+		// Give other threads an opportunity to run.
+		yield()
+	}
+}
+
+// Put puts the current vCPU.
+func (m *machine) Put(c *vCPU) {
+	c.Unlock()
+	runtime.UnlockOSThread()
+}
+
+// State returns the current state.
+func (c *vCPU) State() uintptr {
+	return atomic.LoadUintptr(&c.state)
+}
+
+// Lock locks the vCPU.
+func (c *vCPU) Lock() {
+	c.mu.Lock()
+}
+
+// Invalidate invalidates caches.
+func (c *vCPU) Invalidate() {
+}
+
+// LockInState locks the vCPU if it is in the given state and TryLock succeeds.
+func (c *vCPU) LockInState(state uintptr) bool {
+	if c.State() == state && c.mu.TryLock() {
+		if c.State() != state {
+			c.mu.Unlock()
+			return false
+		}
+		return true
+	}
+	return false
+}
+
+// Unlock unlocks the given vCPU.
+func (c *vCPU) Unlock() {
+	// Ensure we're out of guest mode, if necessary.
+	if c.State() == vCPUWaiter {
+		redpill() // Force guest mode exit.
+	}
+	c.mu.Unlock()
+}
+
+// NotifyInterrupt implements interrupt.Receiver.NotifyInterrupt.
+func (c *vCPU) NotifyInterrupt() {
+	c.Bounce()
+}
+
+// pid is used below in bounce.
+var pid = syscall.Getpid()
+
+// Bounce ensures that the vCPU bounces back to the kernel.
+//
+// In practice, this means returning EAGAIN from running user code. The vCPU
+// will be unlocked and relock, and the kernel is guaranteed to check for
+// interrupt notifications (e.g. injected via Notify) and invalidations.
+func (c *vCPU) Bounce() {
+	for {
+		if c.mu.TryLock() {
+			// We know that the vCPU must be in the kernel already,
+			// because the lock was not acquired. We specifically
+			// don't want to call bounce in this case, because it's
+			// not necessary to knock the vCPU out of guest mode.
+			c.mu.Unlock()
+			return
+		}
+
+		if state := c.State(); state == vCPUGuest || state == vCPUWaiter {
+			// We know that the vCPU was in guest mode, so a single signal
+			// interruption will guarantee that a transition takes place.
+			syscall.Tgkill(pid, int(atomic.LoadUint64(&c.tid)), bounceSignal)
+			return
+		}
+
+		// Someone holds the lock, but the vCPU is not yet transitioned
+		// into guest mode. It's in the critical section; give it time.
+		yield()
+	}
+}
diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go
new file mode 100644
index 000000000..dfa691e88
--- /dev/null
+++ b/pkg/sentry/platform/kvm/machine_amd64.go
@@ -0,0 +1,168 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"fmt"
+	"reflect"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// initArchState initializes architecture-specific state.
+func (m *machine) initArchState(vCPUs int) error {
+	// Set the legacy TSS address. This address is covered by the reserved
+	// range (up to 4GB). In fact, this is a main reason it exists.
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(m.fd),
+		_KVM_SET_TSS_ADDR,
+		uintptr(reservedMemory-(3*usermem.PageSize))); errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// initArchState initializes architecture-specific state.
+func (c *vCPU) initArchState() error {
+	var (
+		kernelSystemRegs systemRegs
+		kernelUserRegs   userRegs
+	)
+
+	// Set base control registers.
+	kernelSystemRegs.CR0 = c.CR0()
+	kernelSystemRegs.CR4 = c.CR4()
+	kernelSystemRegs.EFER = c.EFER()
+
+	// Set the IDT & GDT in the registers.
+	kernelSystemRegs.IDT.base, kernelSystemRegs.IDT.limit = c.IDT()
+	kernelSystemRegs.GDT.base, kernelSystemRegs.GDT.limit = c.GDT()
+	kernelSystemRegs.CS.Load(&ring0.KernelCodeSegment, ring0.Kcode)
+	kernelSystemRegs.DS.Load(&ring0.UserDataSegment, ring0.Udata)
+	kernelSystemRegs.ES.Load(&ring0.UserDataSegment, ring0.Udata)
+	kernelSystemRegs.SS.Load(&ring0.KernelDataSegment, ring0.Kdata)
+	kernelSystemRegs.FS.Load(&ring0.UserDataSegment, ring0.Udata)
+	kernelSystemRegs.GS.Load(&ring0.UserDataSegment, ring0.Udata)
+	tssBase, tssLimit, tss := c.TSS()
+	kernelSystemRegs.TR.Load(tss, ring0.Tss)
+	kernelSystemRegs.TR.base = tssBase
+	kernelSystemRegs.TR.limit = uint32(tssLimit)
+
+	// Point to kernel page tables.
+	kernelSystemRegs.CR3 = c.machine.kernel.PageTables.FlushCR3()
+
+	// Set the CPUID; this is required before setting system registers,
+	// since KVM will reject several CR4 bits if the CPUID does not
+	// indicate the support is available.
+	if err := c.setCPUID(); err != nil {
+		return err
+	}
+
+	// Set the entrypoint for the kernel.
+	kernelUserRegs.RIP = uint64(reflect.ValueOf(ring0.Start).Pointer())
+	kernelUserRegs.RAX = uint64(reflect.ValueOf(&c.CPU).Pointer())
+	kernelUserRegs.RFLAGS = ring0.KernelFlagsSet
+
+	// Set the system registers.
+	if err := c.setSystemRegisters(&kernelSystemRegs); err != nil {
+		return err
+	}
+
+	// Set the user registers.
+	if err := c.setUserRegisters(&kernelUserRegs); err != nil {
+		return err
+	}
+
+	// Set the time offset to the host native time.
+	return c.setSystemTime()
+}
+
+// SwitchToUser unpacks architectural-details.
+func (c *vCPU) SwitchToUser(regs *syscall.PtraceRegs, fpState *byte, pt *pagetables.PageTables, flags ring0.Flags) (*arch.SignalInfo, usermem.AccessType, error) {
+	// See below.
+	var vector ring0.Vector
+
+	// Past this point, stack growth can cause system calls (and a break
+	// from guest mode). So we need to ensure that between the bluepill
+	// call here and the switch call immediately below, no additional
+	// allocations occur.
+	entersyscall()
+	bluepill(c)
+	vector = c.CPU.SwitchToUser(regs, fpState, pt, flags)
+	exitsyscall()
+
+	// Free and clear.
+	switch vector {
+	case ring0.Debug, ring0.Breakpoint:
+		info := &arch.SignalInfo{Signo: int32(syscall.SIGTRAP)}
+		return info, usermem.AccessType{}, platform.ErrContextSignal
+
+	case ring0.PageFault:
+		bluepill(c) // Probably no-op, but may not be.
+		faultAddr := ring0.ReadCR2()
+		code, user := c.ErrorCode()
+		if !user {
+			// The last fault serviced by this CPU was not a user
+			// fault, so we can't reliably trust the faultAddr or
+			// the code provided here. We need to re-execute.
+			return nil, usermem.NoAccess, platform.ErrContextInterrupt
+		}
+		info := &arch.SignalInfo{Signo: int32(syscall.SIGSEGV)}
+		info.SetAddr(uint64(faultAddr))
+		accessType := usermem.AccessType{
+			Read:    code&(1<<1) == 0,
+			Write:   code&(1<<1) != 0,
+			Execute: code&(1<<4) != 0,
+		}
+		return info, accessType, platform.ErrContextSignal
+
+	case ring0.GeneralProtectionFault:
+		if !ring0.IsCanonical(regs.Rip) {
+			// If the RIP is non-canonical, it's a SEGV.
+			info := &arch.SignalInfo{Signo: int32(syscall.SIGSEGV)}
+			return info, usermem.AccessType{}, platform.ErrContextSignal
+		}
+		// Otherwise, we deliver a SIGBUS.
+		info := &arch.SignalInfo{Signo: int32(syscall.SIGBUS)}
+		return info, usermem.AccessType{}, platform.ErrContextSignal
+
+	case ring0.InvalidOpcode:
+		info := &arch.SignalInfo{Signo: int32(syscall.SIGILL)}
+		return info, usermem.AccessType{}, platform.ErrContextSignal
+
+	case ring0.X87FloatingPointException:
+		info := &arch.SignalInfo{Signo: int32(syscall.SIGFPE)}
+		return info, usermem.AccessType{}, platform.ErrContextSignal
+
+	case ring0.Vector(bounce):
+		redpill() // Bail and reacqire.
+		return nil, usermem.NoAccess, platform.ErrContextInterrupt
+
+	case ring0.Syscall, ring0.SyscallInt80:
+		// System call executed.
+		return nil, usermem.NoAccess, nil
+
+	default:
+		panic(fmt.Sprintf("unexpected vector: 0x%x", vector))
+	}
+}
diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go
new file mode 100644
index 000000000..c2bcb3a47
--- /dev/null
+++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go
@@ -0,0 +1,156 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package kvm
+
+import (
+	"fmt"
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/time"
+)
+
+// setMemoryRegion initializes a region.
+//
+// This may be called from bluepillHandler, and therefore returns an errno
+// directly (instead of wrapping in an error) to avoid allocations.
+//
+//go:nosplit
+func (m *machine) setMemoryRegion(slot int, physical, length, virtual uintptr) syscall.Errno {
+	userRegion := userMemoryRegion{
+		slot:          uint32(slot),
+		flags:         0,
+		guestPhysAddr: uint64(physical),
+		memorySize:    uint64(length),
+		userspaceAddr: uint64(virtual),
+	}
+
+	// Set the region.
+	_, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(m.fd),
+		_KVM_SET_USER_MEMORY_REGION,
+		uintptr(unsafe.Pointer(&userRegion)))
+	return errno
+}
+
+// loadSegments copies the current segments.
+//
+// This may be called from within the signal context and throws on error.
+//
+//go:nosplit
+func (c *vCPU) loadSegments() {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_ARCH_PRCTL,
+		linux.ARCH_GET_FS,
+		uintptr(unsafe.Pointer(&c.CPU.Registers().Fs_base)),
+		0); errno != 0 {
+		throw("getting FS segment")
+	}
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_ARCH_PRCTL,
+		linux.ARCH_GET_GS,
+		uintptr(unsafe.Pointer(&c.CPU.Registers().Gs_base)),
+		0); errno != 0 {
+		throw("getting GS segment")
+	}
+}
+
+// setUserRegisters sets user registers in the vCPU.
+func (c *vCPU) setUserRegisters(uregs *userRegs) error {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(c.fd),
+		_KVM_SET_REGS,
+		uintptr(unsafe.Pointer(uregs))); errno != 0 {
+		return fmt.Errorf("error setting user registers: %v", errno)
+	}
+	return nil
+}
+
+// setSystemRegisters sets system registers.
+func (c *vCPU) setSystemRegisters(sregs *systemRegs) error {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(c.fd),
+		_KVM_SET_SREGS,
+		uintptr(unsafe.Pointer(sregs))); errno != 0 {
+		return fmt.Errorf("error setting system registers: %v", errno)
+	}
+	return nil
+}
+
+// setCPUID sets the CPUID to be used by the guest.
+func (c *vCPU) setCPUID() error {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(c.fd),
+		_KVM_SET_CPUID2,
+		uintptr(unsafe.Pointer(&cpuidSupported))); errno != 0 {
+		return fmt.Errorf("error setting CPUID: %v", errno)
+	}
+	return nil
+}
+
+// setSystemTime sets the TSC for the vCPU.
+//
+// FIXME: This introduces a slight TSC offset between host and
+// guest, which may vary per vCPU.
+func (c *vCPU) setSystemTime() error {
+	const _MSR_IA32_TSC = 0x00000010
+	registers := modelControlRegisters{
+		nmsrs: 1,
+	}
+	registers.entries[0] = modelControlRegister{
+		index: _MSR_IA32_TSC,
+		data:  uint64(time.Rdtsc()),
+	}
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(c.fd),
+		_KVM_SET_MSRS,
+		uintptr(unsafe.Pointer(&registers))); errno != 0 {
+		return fmt.Errorf("error setting system time: %v", errno)
+	}
+	return nil
+}
+
+// setSignalMask sets the vCPU signal mask.
+//
+// This must be called prior to running the vCPU.
+func (c *vCPU) setSignalMask() error {
+	// The layout of this structure implies that it will not necessarily be
+	// the same layout chosen by the Go compiler. It gets fudged here.
+	var data struct {
+		length uint32
+		mask1  uint32
+		mask2  uint32
+		_      uint32
+	}
+	data.length = 8 // Fixed sigset size.
+	data.mask1 = ^uint32(bounceSignalMask & 0xffffffff)
+	data.mask2 = ^uint32(bounceSignalMask >> 32)
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_IOCTL,
+		uintptr(c.fd),
+		_KVM_SET_SIGNAL_MASK,
+		uintptr(unsafe.Pointer(&data))); errno != 0 {
+		return fmt.Errorf("error setting signal mask: %v", errno)
+	}
+	return nil
+}
diff --git a/pkg/sentry/platform/kvm/machine_unsafe.go b/pkg/sentry/platform/kvm/machine_unsafe.go
new file mode 100644
index 000000000..da67e23f6
--- /dev/null
+++ b/pkg/sentry/platform/kvm/machine_unsafe.go
@@ -0,0 +1,112 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"fmt"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+)
+
+//go:linkname entersyscall runtime.entersyscall
+func entersyscall()
+
+//go:linkname exitsyscall runtime.exitsyscall
+func exitsyscall()
+
+// TranslateToVirtual implements pagetables.Translater.TranslateToPhysical.
+func (m *machine) TranslateToPhysical(ptes *pagetables.PTEs) uintptr {
+	// The length doesn't matter because all these translations require
+	// only a single page, which is guaranteed to be satisfied.
+	physical, _, ok := TranslateToPhysical(uintptr(unsafe.Pointer(ptes)))
+	if !ok {
+		panic("unable to translate pagetables.Node to physical address")
+	}
+	return physical
+}
+
+// mapRunData maps the vCPU run data.
+func mapRunData(fd int) (*runData, error) {
+	r, _, errno := syscall.RawSyscall6(
+		syscall.SYS_MMAP,
+		0,
+		uintptr(runDataSize),
+		syscall.PROT_READ|syscall.PROT_WRITE,
+		syscall.MAP_SHARED,
+		uintptr(fd),
+		0)
+	if errno != 0 {
+		return nil, fmt.Errorf("error mapping runData: %v", errno)
+	}
+	return (*runData)(unsafe.Pointer(r)), nil
+}
+
+// unmapRunData unmaps the vCPU run data.
+func unmapRunData(r *runData) error {
+	if _, _, errno := syscall.RawSyscall(
+		syscall.SYS_MUNMAP,
+		uintptr(unsafe.Pointer(r)),
+		uintptr(runDataSize),
+		0); errno != 0 {
+		return fmt.Errorf("error unmapping runData: %v", errno)
+	}
+	return nil
+}
+
+// notify notifies that the vCPU has returned to host mode.
+//
+// This may be called by a signal handler and therefore throws on error.
+//
+//go:nosplit
+func (c *vCPU) notify() {
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_FUTEX,
+		uintptr(unsafe.Pointer(&c.state)),
+		linux.FUTEX_WAKE,
+		^uintptr(0), // Number of waiters.
+		0, 0, 0)
+	if errno != 0 {
+		throw("futex wake error")
+	}
+}
+
+// wait waits for the vCPU to return to host mode.
+//
+// This panics on error.
+func (c *vCPU) wait() {
+	if !atomic.CompareAndSwapUintptr(&c.state, vCPUGuest, vCPUWaiter) {
+		return // Nothing to wait for.
+	}
+	for {
+		_, _, errno := syscall.Syscall6(
+			syscall.SYS_FUTEX,
+			uintptr(unsafe.Pointer(&c.state)),
+			linux.FUTEX_WAIT,
+			uintptr(vCPUWaiter), // Expected value.
+			0, 0, 0)
+		if errno == syscall.EINTR {
+			continue
+		} else if errno == syscall.EAGAIN {
+			break
+		} else if errno != 0 {
+			panic("futex wait error")
+		}
+		break
+	}
+}
diff --git a/pkg/sentry/platform/kvm/physical_map.go b/pkg/sentry/platform/kvm/physical_map.go
new file mode 100644
index 000000000..5d55c9486
--- /dev/null
+++ b/pkg/sentry/platform/kvm/physical_map.go
@@ -0,0 +1,221 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"fmt"
+	"sort"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+const (
+	// reservedMemory is a chunk of physical memory reserved starting at
+	// physical address zero. There are some special pages in this region,
+	// so we just call the whole thing off.
+	//
+	// Other architectures may define this to be zero.
+	reservedMemory = 0x100000000
+)
+
+type region struct {
+	virtual uintptr
+	length  uintptr
+}
+
+type physicalRegion struct {
+	region
+	physical uintptr
+}
+
+// physicalRegions contains a list of available physical regions.
+//
+// The physical value used in physicalRegions is a number indicating the
+// physical offset, aligned appropriately and starting above reservedMemory.
+var physicalRegions []physicalRegion
+
+// fillAddressSpace fills the host address space with PROT_NONE mappings until
+// the number of available bits until we have a host address space size that is
+// equal to the physical address space.
+//
+// The excluded regions are returned.
+func fillAddressSpace() (excludedRegions []region) {
+	// We can cut vSize in half, because the kernel will be using the top
+	// half and we ignore it while constructing mappings. It's as if we've
+	// already excluded half the possible addresses.
+	vSize := uintptr(1) << ring0.VirtualAddressBits()
+	vSize = vSize >> 1
+
+	// We exclude reservedMemory below from our physical memory size, so it
+	// needs to be dropped here as well. Otherwise, we could end up with
+	// physical addresses that are beyond what is mapped.
+	pSize := uintptr(1) << ring0.PhysicalAddressBits()
+	pSize -= reservedMemory
+
+	// Sanity check.
+	if vSize < pSize {
+		panic(fmt.Sprintf("vSize (%x) < pSize (%x)", vSize, pSize))
+	}
+
+	// Add specifically excluded regions; see excludeVirtualRegion.
+	applyVirtualRegions(func(vr virtualRegion) {
+		if excludeVirtualRegion(vr) {
+			excludedRegions = append(excludedRegions, vr.region)
+			vSize -= vr.length
+			log.Infof("excluded: virtual [%x,%x)", vr.virtual, vr.virtual+vr.length)
+		}
+	})
+
+	// Calculate the required space and fill it.
+	//
+	// Note carefully that we add faultBlockSize to required up front, and
+	// on each iteration of the loop below (i.e. each new physical region
+	// we define), we add faultBlockSize again. This is done because the
+	// computation of physical regions will ensure proper alignments with
+	// faultBlockSize, potentially causing up to faultBlockSize bytes in
+	// internal fragmentation for each physical region. So we need to
+	// account for this properly during allocation.
+	requiredAddr, ok := usermem.Addr(vSize - pSize + faultBlockSize).RoundUp()
+	if !ok {
+		panic(fmt.Sprintf(
+			"overflow for vSize (%x) - pSize (%x) + faultBlockSize (%x)",
+			vSize, pSize, faultBlockSize))
+	}
+	required := uintptr(requiredAddr)
+	current := required // Attempted mmap size.
+	for filled := uintptr(0); filled < required && current > 0; {
+		addr, _, errno := syscall.RawSyscall6(
+			syscall.SYS_MMAP,
+			0, // Suggested address.
+			current,
+			syscall.PROT_NONE,
+			syscall.MAP_ANONYMOUS|syscall.MAP_PRIVATE|syscall.MAP_NORESERVE,
+			0, 0)
+		if errno != 0 {
+			// Attempt half the size; overflow not possible.
+			currentAddr, _ := usermem.Addr(current >> 1).RoundUp()
+			current = uintptr(currentAddr)
+			continue
+		}
+		// We filled a block.
+		filled += current
+		excludedRegions = append(excludedRegions, region{
+			virtual: addr,
+			length:  current,
+		})
+		// See comment above.
+		if filled != required {
+			required += faultBlockSize
+		}
+	}
+	if current == 0 {
+		panic("filling address space failed")
+	}
+	sort.Slice(excludedRegions, func(i, j int) bool {
+		return excludedRegions[i].virtual < excludedRegions[j].virtual
+	})
+	for _, r := range excludedRegions {
+		log.Infof("region: virtual [%x,%x)", r.virtual, r.virtual+r.length)
+	}
+	return excludedRegions
+}
+
+// computePhysicalRegions computes physical regions.
+func computePhysicalRegions(excludedRegions []region) (physicalRegions []physicalRegion) {
+	physical := uintptr(reservedMemory)
+	addValidRegion := func(virtual, length uintptr) {
+		if length == 0 {
+			return
+		}
+		if virtual == 0 {
+			virtual += usermem.PageSize
+			length -= usermem.PageSize
+		}
+		if end := virtual + length; end > ring0.MaximumUserAddress {
+			length -= (end - ring0.MaximumUserAddress)
+		}
+		if length == 0 {
+			return
+		}
+		// Round physical up to the same alignment as the virtual
+		// address (with respect to faultBlockSize).
+		if offset := virtual &^ faultBlockMask; physical&^faultBlockMask != offset {
+			if newPhysical := (physical & faultBlockMask) + offset; newPhysical > physical {
+				physical = newPhysical // Round up by only a little bit.
+			} else {
+				physical = ((physical + faultBlockSize) & faultBlockMask) + offset
+			}
+		}
+		physicalRegions = append(physicalRegions, physicalRegion{
+			region: region{
+				virtual: virtual,
+				length:  length,
+			},
+			physical: physical,
+		})
+		physical += length
+	}
+	lastExcludedEnd := uintptr(0)
+	for _, r := range excludedRegions {
+		addValidRegion(lastExcludedEnd, r.virtual-lastExcludedEnd)
+		lastExcludedEnd = r.virtual + r.length
+	}
+	addValidRegion(lastExcludedEnd, ring0.MaximumUserAddress-lastExcludedEnd)
+
+	// Dump our all physical regions.
+	for _, r := range physicalRegions {
+		log.Infof("physicalRegion: virtual [%x,%x) => physical [%x,%x)",
+			r.virtual, r.virtual+r.length, r.physical, r.physical+r.length)
+	}
+	return physicalRegions
+}
+
+// physicalInit initializes physical address mappings.
+func physicalInit() {
+	physicalRegions = computePhysicalRegions(fillAddressSpace())
+}
+
+// applyPhysicalRegions applies the given function on physical regions.
+//
+// Iteration continues as long as true is returned. The return value is the
+// return from the last call to fn, or true if there are no entries.
+//
+// Precondition: physicalInit must have been called.
+func applyPhysicalRegions(fn func(pr physicalRegion) bool) bool {
+	for _, pr := range physicalRegions {
+		if !fn(pr) {
+			return false
+		}
+	}
+	return true
+}
+
+// TranslateToPhysical translates the given virtual address.
+//
+// Precondition: physicalInit must have been called.
+func TranslateToPhysical(virtual uintptr) (physical uintptr, length uintptr, ok bool) {
+	ok = !applyPhysicalRegions(func(pr physicalRegion) bool {
+		if pr.virtual <= virtual && virtual < pr.virtual+pr.length {
+			physical = pr.physical + (virtual - pr.virtual)
+			length = pr.length - (virtual - pr.virtual)
+			return false
+		}
+		return true
+	})
+	return
+}
diff --git a/pkg/sentry/platform/kvm/testutil/BUILD b/pkg/sentry/platform/kvm/testutil/BUILD
new file mode 100644
index 000000000..8533a8d89
--- /dev/null
+++ b/pkg/sentry/platform/kvm/testutil/BUILD
@@ -0,0 +1,15 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "testutil",
+    testonly = 1,
+    srcs = [
+        "testutil.go",
+        "testutil_amd64.go",
+        "testutil_amd64.s",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/kvm/testutil",
+    visibility = ["//pkg/sentry/platform/kvm:__pkg__"],
+)
diff --git a/pkg/sentry/platform/kvm/testutil/testutil.go b/pkg/sentry/platform/kvm/testutil/testutil.go
new file mode 100644
index 000000000..8a614e25d
--- /dev/null
+++ b/pkg/sentry/platform/kvm/testutil/testutil.go
@@ -0,0 +1,75 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package testutil provides common assembly stubs for testing.
+package testutil
+
+import (
+	"fmt"
+	"strings"
+)
+
+// Getpid executes a trivial system call.
+func Getpid()
+
+// Touch touches the value in the first register.
+func Touch()
+
+// SyscallLoop executes a syscall and loops.
+func SyscallLoop()
+
+// SpinLoop spins on the CPU.
+func SpinLoop()
+
+// HaltLoop immediately halts and loops.
+func HaltLoop()
+
+// TwiddleRegsFault twiddles registers then faults.
+func TwiddleRegsFault()
+
+// TwiddleRegsSyscall twiddles registers then executes a syscall.
+func TwiddleRegsSyscall()
+
+// TwiddleSegments reads segments into known registers.
+func TwiddleSegments()
+
+// FloatingPointWorks is a floating point test.
+//
+// It returns true or false.
+func FloatingPointWorks() bool
+
+// RegisterMismatchError is used for checking registers.
+type RegisterMismatchError []string
+
+// Error returns a human-readable error.
+func (r RegisterMismatchError) Error() string {
+	return strings.Join([]string(r), ";")
+}
+
+// addRegisterMisatch allows simple chaining of register mismatches.
+func addRegisterMismatch(err error, reg string, got, expected interface{}) error {
+	errStr := fmt.Sprintf("%s got %08x, expected %08x", reg, got, expected)
+	switch r := err.(type) {
+	case nil:
+		// Return a new register mismatch.
+		return RegisterMismatchError{errStr}
+	case RegisterMismatchError:
+		// Append the error.
+		r = append(r, errStr)
+		return r
+	default:
+		// Leave as is.
+		return err
+	}
+}
diff --git a/pkg/sentry/platform/kvm/testutil/testutil_amd64.go b/pkg/sentry/platform/kvm/testutil/testutil_amd64.go
new file mode 100644
index 000000000..39286a0af
--- /dev/null
+++ b/pkg/sentry/platform/kvm/testutil/testutil_amd64.go
@@ -0,0 +1,135 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package testutil
+
+import (
+	"reflect"
+	"syscall"
+)
+
+// SetTestTarget sets the rip appropriately.
+func SetTestTarget(regs *syscall.PtraceRegs, fn func()) {
+	regs.Rip = uint64(reflect.ValueOf(fn).Pointer())
+}
+
+// SetTouchTarget sets rax appropriately.
+func SetTouchTarget(regs *syscall.PtraceRegs, target *uintptr) {
+	if target != nil {
+		regs.Rax = uint64(reflect.ValueOf(target).Pointer())
+	} else {
+		regs.Rax = 0
+	}
+}
+
+// RewindSyscall rewinds a syscall RIP.
+func RewindSyscall(regs *syscall.PtraceRegs) {
+	regs.Rip -= 2
+}
+
+// SetTestRegs initializes registers to known values.
+func SetTestRegs(regs *syscall.PtraceRegs) {
+	regs.R15 = 0x15
+	regs.R14 = 0x14
+	regs.R13 = 0x13
+	regs.R12 = 0x12
+	regs.Rbp = 0xb9
+	regs.Rbx = 0xb4
+	regs.R11 = 0x11
+	regs.R10 = 0x10
+	regs.R9 = 0x09
+	regs.R8 = 0x08
+	regs.Rax = 0x44
+	regs.Rcx = 0xc4
+	regs.Rdx = 0xd4
+	regs.Rsi = 0x51
+	regs.Rdi = 0xd1
+	regs.Rsp = 0x59
+}
+
+// CheckTestRegs checks that registers were twiddled per TwiddleRegs.
+func CheckTestRegs(regs *syscall.PtraceRegs, full bool) (err error) {
+	if need := ^uint64(0x15); regs.R15 != need {
+		err = addRegisterMismatch(err, "R15", regs.R15, need)
+	}
+	if need := ^uint64(0x14); regs.R14 != need {
+		err = addRegisterMismatch(err, "R14", regs.R14, need)
+	}
+	if need := ^uint64(0x13); regs.R13 != need {
+		err = addRegisterMismatch(err, "R13", regs.R13, need)
+	}
+	if need := ^uint64(0x12); regs.R12 != need {
+		err = addRegisterMismatch(err, "R12", regs.R12, need)
+	}
+	if need := ^uint64(0xb9); regs.Rbp != need {
+		err = addRegisterMismatch(err, "Rbp", regs.Rbp, need)
+	}
+	if need := ^uint64(0xb4); regs.Rbx != need {
+		err = addRegisterMismatch(err, "Rbx", regs.Rbx, need)
+	}
+	if need := ^uint64(0x10); regs.R10 != need {
+		err = addRegisterMismatch(err, "R10", regs.R10, need)
+	}
+	if need := ^uint64(0x09); regs.R9 != need {
+		err = addRegisterMismatch(err, "R9", regs.R9, need)
+	}
+	if need := ^uint64(0x08); regs.R8 != need {
+		err = addRegisterMismatch(err, "R8", regs.R8, need)
+	}
+	if need := ^uint64(0x44); regs.Rax != need {
+		err = addRegisterMismatch(err, "Rax", regs.Rax, need)
+	}
+	if need := ^uint64(0xd4); regs.Rdx != need {
+		err = addRegisterMismatch(err, "Rdx", regs.Rdx, need)
+	}
+	if need := ^uint64(0x51); regs.Rsi != need {
+		err = addRegisterMismatch(err, "Rsi", regs.Rsi, need)
+	}
+	if need := ^uint64(0xd1); regs.Rdi != need {
+		err = addRegisterMismatch(err, "Rdi", regs.Rdi, need)
+	}
+	if need := ^uint64(0x59); regs.Rsp != need {
+		err = addRegisterMismatch(err, "Rsp", regs.Rsp, need)
+	}
+	// Rcx & R11 are ignored if !full is set.
+	if need := ^uint64(0x11); full && regs.R11 != need {
+		err = addRegisterMismatch(err, "R11", regs.R11, need)
+	}
+	if need := ^uint64(0xc4); full && regs.Rcx != need {
+		err = addRegisterMismatch(err, "Rcx", regs.Rcx, need)
+	}
+	return
+}
+
+var fsData uint64 = 0x55
+var gsData uint64 = 0x85
+
+// SetTestSegments initializes segments to known values.
+func SetTestSegments(regs *syscall.PtraceRegs) {
+	regs.Fs_base = uint64(reflect.ValueOf(&fsData).Pointer())
+	regs.Gs_base = uint64(reflect.ValueOf(&gsData).Pointer())
+}
+
+// CheckTestSegments checks that registers were twiddled per TwiddleSegments.
+func CheckTestSegments(regs *syscall.PtraceRegs) (err error) {
+	if regs.Rax != fsData {
+		err = addRegisterMismatch(err, "Rax", regs.Rax, fsData)
+	}
+	if regs.Rbx != gsData {
+		err = addRegisterMismatch(err, "Rbx", regs.Rcx, gsData)
+	}
+	return
+}
diff --git a/pkg/sentry/platform/kvm/testutil/testutil_amd64.s b/pkg/sentry/platform/kvm/testutil/testutil_amd64.s
new file mode 100644
index 000000000..3b5ad8817
--- /dev/null
+++ b/pkg/sentry/platform/kvm/testutil/testutil_amd64.s
@@ -0,0 +1,98 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+// test_util_amd64.s provides AMD64 test functions.
+
+#include "funcdata.h"
+#include "textflag.h"
+
+TEXT ·Getpid(SB),NOSPLIT,$0
+	NO_LOCAL_POINTERS
+	MOVQ $39, AX // getpid
+	SYSCALL
+	RET
+
+TEXT ·Touch(SB),NOSPLIT,$0
+start:
+	MOVQ 0(AX), BX // deref AX
+	MOVQ $39, AX   // getpid
+	SYSCALL
+	JMP start
+
+TEXT ·HaltLoop(SB),NOSPLIT,$0
+start:
+	HLT
+	JMP start
+
+TEXT ·SyscallLoop(SB),NOSPLIT,$0
+start:
+	SYSCALL
+	JMP start
+
+TEXT ·SpinLoop(SB),NOSPLIT,$0
+start:
+	JMP start
+
+TEXT ·FloatingPointWorks(SB),NOSPLIT,$0-8
+	NO_LOCAL_POINTERS
+	MOVQ $1, AX
+	MOVQ AX, X0
+	MOVQ $39, AX // getpid
+	SYSCALL
+	MOVQ X0, AX
+	CMPQ AX, $1
+	SETEQ ret+0(FP)
+	RET
+
+#define TWIDDLE_REGS() \
+	NOTQ R15; \
+	NOTQ R14; \
+	NOTQ R13; \
+	NOTQ R12; \
+	NOTQ BP; \
+	NOTQ BX; \
+	NOTQ R11; \
+	NOTQ R10; \
+	NOTQ R9; \
+	NOTQ R8; \
+	NOTQ AX; \
+	NOTQ CX; \
+	NOTQ DX; \
+	NOTQ SI; \
+	NOTQ DI; \
+	NOTQ SP;
+
+TEXT ·TwiddleRegsSyscall(SB),NOSPLIT,$0
+	TWIDDLE_REGS()
+	SYSCALL
+	RET // never reached
+
+TEXT ·TwiddleRegsFault(SB),NOSPLIT,$0
+	TWIDDLE_REGS()
+	JMP AX // must fault
+	RET // never reached
+
+#define READ_FS() BYTE $0x64; BYTE $0x48; BYTE $0x8b; BYTE $0x00;
+#define READ_GS() BYTE $0x65; BYTE $0x48; BYTE $0x8b; BYTE $0x00;
+
+TEXT ·TwiddleSegments(SB),NOSPLIT,$0
+	MOVQ $0x0, AX
+	READ_GS()
+	MOVQ AX, BX
+	MOVQ $0x0, AX
+	READ_FS()
+	SYSCALL
+	RET // never reached
diff --git a/pkg/sentry/platform/kvm/virtual_map.go b/pkg/sentry/platform/kvm/virtual_map.go
new file mode 100644
index 000000000..0d3fbe043
--- /dev/null
+++ b/pkg/sentry/platform/kvm/virtual_map.go
@@ -0,0 +1,113 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"os"
+	"regexp"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type virtualRegion struct {
+	region
+	accessType usermem.AccessType
+	shared     bool
+	offset     uintptr
+	filename   string
+}
+
+// mapsLine matches a single line from /proc/PID/maps.
+var mapsLine = regexp.MustCompile("([0-9a-f]+)-([0-9a-f]+) ([r-][w-][x-][sp]) ([0-9a-f]+) [0-9a-f]{2}:[0-9a-f]{2,} [0-9]+\\s+(.*)")
+
+// excludeRegion returns true if these regions should be excluded from the
+// physical map. Virtual regions need to be excluded if get_user_pages will
+// fail on those addresses, preventing KVM from satisfying EPT faults.
+//
+// This includes the VVAR page because the VVAR page may be mapped as I/O
+// memory. And the VDSO page is knocked out because the VVAR page is not even
+// recorded in /proc/self/maps on older kernels; knocking out the VDSO page
+// prevents code in the VDSO from accessing the VVAR address.
+//
+// This is called by the physical map functions, not applyVirtualRegions.
+func excludeVirtualRegion(r virtualRegion) bool {
+	return r.filename == "[vvar]" || r.filename == "[vdso]"
+}
+
+// applyVirtualRegions parses the process maps file.
+//
+// Unlike mappedRegions, these are not consistent over time.
+func applyVirtualRegions(fn func(vr virtualRegion)) error {
+	// Open /proc/self/maps.
+	f, err := os.Open("/proc/self/maps")
+	if err != nil {
+		return err
+	}
+	defer f.Close()
+
+	// Parse all entries.
+	r := bufio.NewReader(f)
+	for {
+		b, err := r.ReadBytes('\n')
+		if b != nil && len(b) > 0 {
+			m := mapsLine.FindSubmatch(b)
+			if m == nil {
+				// This should not happen: kernel bug?
+				return fmt.Errorf("badly formed line: %v", string(b))
+			}
+			start, err := strconv.ParseUint(string(m[1]), 16, 64)
+			if err != nil {
+				return fmt.Errorf("bad start address: %v", string(b))
+			}
+			end, err := strconv.ParseUint(string(m[2]), 16, 64)
+			if err != nil {
+				return fmt.Errorf("bad end address: %v", string(b))
+			}
+			read := m[3][0] == 'r'
+			write := m[3][1] == 'w'
+			execute := m[3][2] == 'x'
+			shared := m[3][3] == 's'
+			offset, err := strconv.ParseUint(string(m[4]), 16, 64)
+			if err != nil {
+				return fmt.Errorf("bad offset: %v", string(b))
+			}
+			fn(virtualRegion{
+				region: region{
+					virtual: uintptr(start),
+					length:  uintptr(end - start),
+				},
+				accessType: usermem.AccessType{
+					Read:    read,
+					Write:   write,
+					Execute: execute,
+				},
+				shared:   shared,
+				offset:   uintptr(offset),
+				filename: string(m[5]),
+			})
+		}
+		if err != nil && err == io.EOF {
+			break
+		} else if err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
diff --git a/pkg/sentry/platform/kvm/virtual_map_test.go b/pkg/sentry/platform/kvm/virtual_map_test.go
new file mode 100644
index 000000000..31e5b0e61
--- /dev/null
+++ b/pkg/sentry/platform/kvm/virtual_map_test.go
@@ -0,0 +1,78 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package kvm
+
+import (
+	"syscall"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type checker struct {
+	ok bool
+}
+
+func (c *checker) Contains(addr uintptr) func(virtualRegion) {
+	c.ok = false // Reset for below calls.
+	return func(vr virtualRegion) {
+		if vr.virtual <= addr && addr < vr.virtual+vr.length {
+			c.ok = true
+		}
+	}
+}
+
+func TestParseMaps(t *testing.T) {
+	c := new(checker)
+
+	// Simple test.
+	if err := applyVirtualRegions(c.Contains(0)); err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// MMap a new page.
+	addr, _, errno := syscall.RawSyscall6(
+		syscall.SYS_MMAP, 0, usermem.PageSize,
+		syscall.PROT_READ|syscall.PROT_WRITE,
+		syscall.MAP_ANONYMOUS|syscall.MAP_PRIVATE, 0, 0)
+	if errno != 0 {
+		t.Fatalf("unexpected map error: %v", errno)
+	}
+
+	// Re-parse maps.
+	if err := applyVirtualRegions(c.Contains(addr)); err != nil {
+		syscall.RawSyscall(syscall.SYS_MUNMAP, addr, usermem.PageSize, 0)
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// Assert that it now does contain the region.
+	if !c.ok {
+		syscall.RawSyscall(syscall.SYS_MUNMAP, addr, usermem.PageSize, 0)
+		t.Fatalf("updated map does not contain 0x%08x, expected true", addr)
+	}
+
+	// Unmap the region.
+	syscall.RawSyscall(syscall.SYS_MUNMAP, addr, usermem.PageSize, 0)
+
+	// Re-parse maps.
+	if err := applyVirtualRegions(c.Contains(addr)); err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+
+	// Assert that it once again does _not_ contain the region.
+	if c.ok {
+		t.Fatalf("final map does contain 0x%08x, expected false", addr)
+	}
+}
diff --git a/pkg/sentry/platform/mmap_min_addr.go b/pkg/sentry/platform/mmap_min_addr.go
new file mode 100644
index 000000000..6398e5e01
--- /dev/null
+++ b/pkg/sentry/platform/mmap_min_addr.go
@@ -0,0 +1,60 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package platform
+
+import (
+	"fmt"
+	"io/ioutil"
+	"strconv"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// systemMMapMinAddrSource is the source file.
+const systemMMapMinAddrSource = "/proc/sys/vm/mmap_min_addr"
+
+// systemMMapMinAddr is the system's minimum map address.
+var systemMMapMinAddr uint64
+
+// SystemMMapMinAddr returns the minimum system address.
+func SystemMMapMinAddr() usermem.Addr {
+	return usermem.Addr(systemMMapMinAddr)
+}
+
+// MMapMinAddr is a size zero struct that implements MinUserAddress based on
+// the system minimum address. It is suitable for embedding in platforms that
+// rely on the system mmap, and thus require the system minimum.
+type MMapMinAddr struct {
+}
+
+// MinUserAddress implements platform.MinUserAddresss.
+func (*MMapMinAddr) MinUserAddress() usermem.Addr {
+	return SystemMMapMinAddr()
+}
+
+func init() {
+	// Open the source file.
+	b, err := ioutil.ReadFile(systemMMapMinAddrSource)
+	if err != nil {
+		panic(fmt.Sprintf("couldn't open %s: %v", systemMMapMinAddrSource, err))
+	}
+
+	// Parse the result.
+	systemMMapMinAddr, err = strconv.ParseUint(strings.TrimSpace(string(b)), 10, 64)
+	if err != nil {
+		panic(fmt.Sprintf("couldn't parse %s from %s: %v", string(b), systemMMapMinAddrSource, err))
+	}
+}
diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go
new file mode 100644
index 000000000..6219dada7
--- /dev/null
+++ b/pkg/sentry/platform/platform.go
@@ -0,0 +1,428 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package platform provides a Platform abstraction.
+//
+// See Platform for more information.
+package platform
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Platform provides abstractions for execution contexts (Context) and memory
+// management (Memory, AddressSpace).
+type Platform interface {
+	// SupportsAddressSpaceIO returns true if AddressSpaces returned by this
+	// Platform support AddressSpaceIO methods.
+	//
+	// The value returned by SupportsAddressSpaceIO is guaranteed to remain
+	// unchanged over the lifetime of the Platform.
+	SupportsAddressSpaceIO() bool
+
+	// CooperativelySchedulesAddressSpace returns true if the Platform has a
+	// limited number of AddressSpaces, such that mm.MemoryManager.Deactivate
+	// should call AddressSpace.Release when there are no goroutines that
+	// require the mm.MemoryManager to have an active AddressSpace.
+	//
+	// The value returned by CooperativelySchedulesAddressSpace is guaranteed
+	// to remain unchanged over the lifetime of the Platform.
+	CooperativelySchedulesAddressSpace() bool
+
+	// DetectsCPUPreemption returns true if Contexts returned by the Platform
+	// can reliably return ErrContextCPUPreempted.
+	DetectsCPUPreemption() bool
+
+	// MapUnit returns the alignment used for optional mappings into this
+	// platform's AddressSpaces. Higher values indicate lower per-page
+	// costs for AddressSpace.MapInto. As a special case, a MapUnit of 0
+	// indicates that the cost of AddressSpace.MapInto is effectively
+	// independent of the number of pages mapped. If MapUnit is non-zero,
+	// it must be a power-of-2 multiple of usermem.PageSize.
+	MapUnit() uint64
+
+	// MinUserAddress returns the minimum mappable address on this
+	// platform.
+	MinUserAddress() usermem.Addr
+
+	// MaxUserAddress returns the maximum mappable address on this
+	// platform.
+	MaxUserAddress() usermem.Addr
+
+	// NewAddressSpace returns a new memory context for this platform.
+	//
+	// If mappingsID is not nil, the platform may assume that (1) all calls
+	// to NewAddressSpace with the same mappingsID represent the same
+	// (mutable) set of mappings, and (2) the set of mappings has not
+	// changed since the last time AddressSpace.Release was called on an
+	// AddressSpace returned by a call to NewAddressSpace with the same
+	// mappingsID.
+	//
+	// If a new AddressSpace cannot be created immediately, a nil
+	// AddressSpace is returned, along with channel that is closed when
+	// the caller should retry a call to NewAddressSpace.
+	//
+	// In general, this blocking behavior only occurs when
+	// CooperativelySchedulesAddressSpace (above) returns false.
+	NewAddressSpace(mappingsID interface{}) (AddressSpace, <-chan struct{}, error)
+
+	// NewContext returns a new execution context.
+	NewContext() Context
+
+	// Memory returns memory for allocations.
+	Memory() Memory
+
+	// PreemptAllCPUs causes all concurrent calls to Context.Switch(), as well
+	// as the first following call to Context.Switch() for each Context, to
+	// return ErrContextCPUPreempted.
+	//
+	// PreemptAllCPUs is only supported if DetectsCPUPremption() == true.
+	// Platforms for which this does not hold may panic if PreemptAllCPUs is
+	// called.
+	PreemptAllCPUs() error
+}
+
+// NoCPUPreemptionDetection implements Platform.DetectsCPUPreemption and
+// dependent methods for Platforms that do not support this feature.
+type NoCPUPreemptionDetection struct{}
+
+// DetectsCPUPreemption implements Platform.DetectsCPUPreemption.
+func (NoCPUPreemptionDetection) DetectsCPUPreemption() bool {
+	return false
+}
+
+// PreemptAllCPUs implements Platform.PreemptAllCPUs.
+func (NoCPUPreemptionDetection) PreemptAllCPUs() error {
+	panic("This platform does not support CPU preemption detection")
+}
+
+// Context represents the execution context for a single thread.
+type Context interface {
+	// Switch resumes execution of the thread specified by the arch.Context
+	// in the provided address space. This call will block while the thread
+	// is executing.
+	//
+	// If cpu is non-negative, and it is not the number of the CPU that the
+	// thread executes on, Context should return ErrContextCPUPreempted. cpu
+	// can only be non-negative if Platform.DetectsCPUPreemption() is true;
+	// Contexts from Platforms for which this does not hold may ignore cpu, or
+	// panic if cpu is non-negative.
+	//
+	// Switch may return one of the following special errors:
+	//
+	// - nil: The Context invoked a system call.
+	//
+	// - ErrContextSignal: The Context was interrupted by a signal. The
+	// returned *arch.SignalInfo contains information about the signal. If
+	// arch.SignalInfo.Signo == SIGSEGV, the returned usermem.AccessType
+	// contains the access type of the triggering fault.
+	//
+	// - ErrContextInterrupt: The Context was interrupted by a call to
+	// Interrupt(). Switch() may return ErrContextInterrupt spuriously. In
+	// particular, most implementations of Interrupt() will cause the first
+	// following call to Switch() to return ErrContextInterrupt if there is no
+	// concurrent call to Switch().
+	//
+	// - ErrContextCPUPreempted: See the definition of that error for details.
+	Switch(as AddressSpace, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error)
+
+	// Interrupt interrupts a concurrent call to Switch(), causing it to return
+	// ErrContextInterrupt.
+	Interrupt()
+}
+
+var (
+	// ErrContextSignal is returned by Context.Switch() to indicate that the
+	// Context was interrupted by a signal.
+	ErrContextSignal = fmt.Errorf("interrupted by signal")
+
+	// ErrContextInterrupt is returned by Context.Switch() to indicate that the
+	// Context was interrupted by a call to Context.Interrupt().
+	ErrContextInterrupt = fmt.Errorf("interrupted by platform.Context.Interrupt()")
+
+	// ErrContextCPUPreempted is returned by Context.Switch() to indicate that
+	// one of the following occurred:
+	//
+	// - The CPU executing the Context is not the CPU passed to
+	// Context.Switch().
+	//
+	// - The CPU executing the Context may have executed another Context since
+	// the last time it executed this one; or the CPU has previously executed
+	// another Context, and has never executed this one.
+	//
+	// - Platform.PreemptAllCPUs() was called since the last return from
+	// Context.Switch().
+	ErrContextCPUPreempted = fmt.Errorf("interrupted by CPU preemption")
+)
+
+// SignalInterrupt is a signal reserved for use by implementations of
+// Context.Interrupt(). The sentry guarantees that it will ignore delivery of
+// this signal both to Contexts and to the sentry itself, under the assumption
+// that they originate from races with Context.Interrupt().
+//
+// NOTE: The Go runtime only guarantees that a small subset
+// of signals will be always be unblocked on all threads, one of which
+// is SIGCHLD.
+const SignalInterrupt = linux.SIGCHLD
+
+// AddressSpace represents a virtual address space in which a Context can
+// execute.
+type AddressSpace interface {
+	// MapFile creates a shared mapping of offsets in fr, from the file
+	// with file descriptor fd, at address addr. Any existing overlapping
+	// mappings are silently replaced.
+	//
+	// If precommit is true, host memory should be committed to the mapping
+	// when MapFile returns when possible. The precommit flag is advisory
+	// and implementations may choose to ignore it.
+	//
+	// Preconditions: addr and fr must be page-aligned. length > 0.
+	// at.Any() == true.
+	MapFile(addr usermem.Addr, fd int, fr FileRange, at usermem.AccessType, precommit bool) error
+
+	// Unmap unmaps the given range.
+	//
+	// Preconditions: addr is page-aligned. length > 0.
+	Unmap(addr usermem.Addr, length uint64)
+
+	// Release releases this address space. After releasing, a new AddressSpace
+	// must be acquired via platform.NewAddressSpace().
+	Release() error
+
+	// AddressSpaceIO methods are supported iff the associated platform's
+	// Platform.SupportsAddressSpaceIO() == true. AddressSpaces for which this
+	// does not hold may panic if AddressSpaceIO methods are invoked.
+	AddressSpaceIO
+}
+
+// AddressSpaceIO supports IO through the memory mappings installed in an
+// AddressSpace.
+//
+// AddressSpaceIO implementors are responsible for ensuring that address ranges
+// are application-mappable.
+type AddressSpaceIO interface {
+	// CopyOut copies len(src) bytes from src to the memory mapped at addr. It
+	// returns the number of bytes copied. If the number of bytes copied is <
+	// len(src), it returns a non-nil error explaining why.
+	CopyOut(addr usermem.Addr, src []byte) (int, error)
+
+	// CopyIn copies len(dst) bytes from the memory mapped at addr to dst.
+	// It returns the number of bytes copied. If the number of bytes copied is
+	// < len(dst), it returns a non-nil error explaining why.
+	CopyIn(addr usermem.Addr, dst []byte) (int, error)
+
+	// ZeroOut sets toZero bytes to 0, starting at addr. It returns the number
+	// of bytes zeroed. If the number of bytes zeroed is < toZero, it returns a
+	// non-nil error explaining why.
+	ZeroOut(addr usermem.Addr, toZero uintptr) (uintptr, error)
+
+	// SwapUint32 atomically sets the uint32 value at addr to new and returns
+	// the previous value.
+	//
+	// Preconditions: addr must be aligned to a 4-byte boundary.
+	SwapUint32(addr usermem.Addr, new uint32) (uint32, error)
+
+	// CompareAndSwapUint32 atomically compares the uint32 value at addr to
+	// old; if they are equal, the value in memory is replaced by new. In
+	// either case, the previous value stored in memory is returned.
+	//
+	// Preconditions: addr must be aligned to a 4-byte boundary.
+	CompareAndSwapUint32(addr usermem.Addr, old, new uint32) (uint32, error)
+}
+
+// NoAddressSpaceIO implements AddressSpaceIO methods by panicing.
+type NoAddressSpaceIO struct{}
+
+// CopyOut implements AddressSpaceIO.CopyOut.
+func (NoAddressSpaceIO) CopyOut(addr usermem.Addr, src []byte) (int, error) {
+	panic("This platform does not support AddressSpaceIO")
+}
+
+// CopyIn implements AddressSpaceIO.CopyIn.
+func (NoAddressSpaceIO) CopyIn(addr usermem.Addr, dst []byte) (int, error) {
+	panic("This platform does not support AddressSpaceIO")
+}
+
+// ZeroOut implements AddressSpaceIO.ZeroOut.
+func (NoAddressSpaceIO) ZeroOut(addr usermem.Addr, toZero uintptr) (uintptr, error) {
+	panic("This platform does not support AddressSpaceIO")
+}
+
+// SwapUint32 implements AddressSpaceIO.SwapUint32.
+func (NoAddressSpaceIO) SwapUint32(addr usermem.Addr, new uint32) (uint32, error) {
+	panic("This platform does not support AddressSpaceIO")
+}
+
+// CompareAndSwapUint32 implements AddressSpaceIO.CompareAndSwapUint32.
+func (NoAddressSpaceIO) CompareAndSwapUint32(addr usermem.Addr, old, new uint32) (uint32, error) {
+	panic("This platform does not support AddressSpaceIO")
+}
+
+// SegmentationFault is an error returned by AddressSpaceIO methods when IO
+// fails due to access of an unmapped page, or a mapped page with insufficient
+// permissions.
+type SegmentationFault struct {
+	// Addr is the address at which the fault occurred.
+	Addr usermem.Addr
+}
+
+// Error implements error.Error.
+func (f SegmentationFault) Error() string {
+	return fmt.Sprintf("segmentation fault at %#x", f.Addr)
+}
+
+// File represents a host file that may be mapped into an AddressSpace.
+type File interface {
+	// MapInto maps fr into as, starting at addr, for accesses of type at.
+	//
+	// If precommit is true, the platform should eagerly commit resources (e.g.
+	// physical memory) to the mapping. The precommit flag is advisory and
+	// implementations may choose to ignore it.
+	//
+	// Note that there is no File.Unmap; clients should use as.Unmap directly.
+	//
+	// Preconditions: fr.Start and fr.End must be page-aligned.
+	// fr.Length() > 0. at.Any() == true. Implementors may define
+	// additional requirements.
+	MapInto(as AddressSpace, addr usermem.Addr, fr FileRange, at usermem.AccessType, precommit bool) error
+
+	// MapInternal returns a mapping of the given file offsets in the invoking
+	// process' address space for reading and writing. The lifetime of the
+	// returned mapping is implementation-defined.
+	//
+	// Note that fr.Start and fr.End need not be page-aligned.
+	//
+	// Preconditions: fr.Length() > 0. Implementors may define additional
+	// requirements.
+	MapInternal(fr FileRange, at usermem.AccessType) (safemem.BlockSeq, error)
+
+	// IncRef signals that a region in the file is actively referenced through a
+	// memory map. Implementors must ensure that the contents of a referenced
+	// region remain consistent. Specifically, mappings returned by MapInternal
+	// must refer to the same underlying contents. If the implementor also
+	// implements the Memory interface, the file range must not be reused in a
+	// different allocation while it has active references.
+	//
+	// Preconditions: fr.Start and fr.End must be page-aligned. fr.Length() > 0.
+	IncRef(fr FileRange)
+
+	// DecRef reduces the frame ref count on the range specified by fr.
+	//
+	// Preconditions: fr.Start and fr.End must be page-aligned. fr.Length() >
+	// 0. DecRef()s on a region must match earlier IncRef()s.
+	DecRef(fr FileRange)
+}
+
+// FileRange represents a range of uint64 offsets into a File.
+//
+// type FileRange <generated using go_generics>
+
+// String implements fmt.Stringer.String.
+func (fr FileRange) String() string {
+	return fmt.Sprintf("[%#x, %#x)", fr.Start, fr.End)
+}
+
+// Memory represents an allocatable File that may be mapped into any
+// AddressSpace associated with the same Platform.
+type Memory interface {
+	// Memory implements File methods with the following properties:
+	//
+	// - Pages mapped by MapInto must be allocated, and must be unmapped from
+	// all AddressSpaces before they are freed.
+	//
+	// - Pages mapped by MapInternal must be allocated. Returned mappings are
+	// guaranteed to be valid until the mapped pages are freed.
+	File
+
+	// Allocate returns a range of pages of the given length, owned by the
+	// caller and with the given accounting kind. Allocated memory initially has
+	// a single reference and will automatically be freed when no references to
+	// them remain. See File.IncRef and File.DecRef.
+	//
+	// Preconditions: length must be page-aligned and non-zero.
+	Allocate(length uint64, kind usage.MemoryKind) (FileRange, error)
+
+	// Decommit releases resources associated with maintaining the contents of
+	// the given frames. If Decommit succeeds, future accesses of the
+	// decommitted frames will read zeroes.
+	//
+	// Preconditions: fr.Length() > 0.
+	Decommit(fr FileRange) error
+
+	// UpdateUsage updates the memory usage statistics. This must be called
+	// before the relevant memory statistics in usage.MemoryAccounting can
+	// be considered accurate.
+	UpdateUsage() error
+
+	// TotalUsage returns an aggregate usage for all memory statistics
+	// except Mapped (which is external to the Memory implementation). This
+	// is generally much cheaper than UpdateUsage, but will not provide a
+	// fine-grained breakdown.
+	TotalUsage() (uint64, error)
+
+	// TotalSize returns the current maximum size of the Memory in bytes. The
+	// value returned by TotalSize is permitted to change.
+	TotalSize() uint64
+
+	// Destroy releases all resources associated with the Memory.
+	//
+	// Preconditions: There are no remaining uses of any of the freed memory's
+	// frames.
+	//
+	// Postconditions: None of the Memory's methods may be called after Destroy.
+	Destroy()
+
+	// SaveTo saves the memory state to the given stream, which will
+	// generally be a statefile.
+	SaveTo(w io.Writer) error
+
+	// LoadFrom loads the memory state from the given stream, which will
+	// generally be a statefile.
+	LoadFrom(r io.Reader) error
+}
+
+// AllocateAndFill allocates memory of the given kind from mem and fills it by
+// calling r.ReadToBlocks() repeatedly until either length bytes are read or a
+// non-nil error is returned. It returns the memory filled by r, truncated down
+// to the nearest page. If this is shorter than length bytes due to an error
+// returned by r.ReadToBlocks(), it returns that error.
+//
+// Preconditions: length > 0. length must be page-aligned.
+func AllocateAndFill(mem Memory, length uint64, kind usage.MemoryKind, r safemem.Reader) (FileRange, error) {
+	fr, err := mem.Allocate(length, kind)
+	if err != nil {
+		return FileRange{}, err
+	}
+	dsts, err := mem.MapInternal(fr, usermem.Write)
+	if err != nil {
+		mem.DecRef(fr)
+		return FileRange{}, err
+	}
+	n, err := safemem.ReadFullToBlocks(r, dsts)
+	un := uint64(usermem.Addr(n).RoundDown())
+	if un < length {
+		// Free unused memory and update fr to contain only the memory that is
+		// still allocated.
+		mem.DecRef(FileRange{fr.Start + un, fr.End})
+		fr.End = fr.Start + un
+	}
+	return fr, err
+}
diff --git a/pkg/sentry/platform/procid/BUILD b/pkg/sentry/platform/procid/BUILD
new file mode 100644
index 000000000..5db4f6261
--- /dev/null
+++ b/pkg/sentry/platform/procid/BUILD
@@ -0,0 +1,32 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "procid",
+    srcs = [
+        "procid.go",
+        "procid_amd64.s",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid",
+    visibility = ["//pkg/sentry:internal"],
+)
+
+go_test(
+    name = "procid_test",
+    size = "small",
+    srcs = [
+        "procid_test.go",
+    ],
+    embed = [":procid"],
+)
+
+go_test(
+    name = "procid_net_test",
+    size = "small",
+    srcs = [
+        "procid_net_test.go",
+        "procid_test.go",
+    ],
+    embed = [":procid"],
+)
diff --git a/pkg/sentry/platform/procid/procid.go b/pkg/sentry/platform/procid/procid.go
new file mode 100644
index 000000000..5f861908f
--- /dev/null
+++ b/pkg/sentry/platform/procid/procid.go
@@ -0,0 +1,21 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package procid provides a way to get the current system thread identifier.
+package procid
+
+// Current returns the current system thread identifier.
+//
+// Precondition: This should only be called with the runtime OS thread locked.
+func Current() uint64
diff --git a/pkg/sentry/platform/procid/procid_amd64.s b/pkg/sentry/platform/procid/procid_amd64.s
new file mode 100644
index 000000000..ead4e3d91
--- /dev/null
+++ b/pkg/sentry/platform/procid/procid_amd64.s
@@ -0,0 +1,30 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+// +build go1.8
+// +build !go1.11
+
+#include "textflag.h"
+
+TEXT ·Current(SB),NOSPLIT,$0-8
+	// The offset specified here is the m_procid offset for Go1.8+.
+	// Changes to this offset should be caught by the tests, and major
+	// version changes require an explicit tag change above.
+	MOVQ TLS, AX
+	MOVQ 0(AX)(TLS*1), AX
+	MOVQ 48(AX), AX // g_m (may change in future versions)
+	MOVQ 72(AX), AX // m_procid (may change in future versions)
+	MOVQ AX, ret+0(FP)
+	RET
diff --git a/pkg/sentry/platform/procid/procid_net_test.go b/pkg/sentry/platform/procid/procid_net_test.go
new file mode 100644
index 000000000..2d1605a08
--- /dev/null
+++ b/pkg/sentry/platform/procid/procid_net_test.go
@@ -0,0 +1,21 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package procid
+
+// This file is just to force the inclusion of the "net" package, which will
+// make the test binary a cgo one.
+import (
+	_ "net"
+)
diff --git a/pkg/sentry/platform/procid/procid_test.go b/pkg/sentry/platform/procid/procid_test.go
new file mode 100644
index 000000000..5e44da36f
--- /dev/null
+++ b/pkg/sentry/platform/procid/procid_test.go
@@ -0,0 +1,85 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package procid
+
+import (
+	"os"
+	"runtime"
+	"sync"
+	"syscall"
+	"testing"
+)
+
+// runOnMain is used to send functions to run on the main (initial) thread.
+var runOnMain = make(chan func(), 10)
+
+func checkProcid(t *testing.T, start *sync.WaitGroup, done *sync.WaitGroup) {
+	defer done.Done()
+
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	start.Done()
+	start.Wait()
+
+	procID := Current()
+	tid := syscall.Gettid()
+
+	if procID != uint64(tid) {
+		t.Logf("Bad procid: expected %v, got %v", tid, procID)
+		t.Fail()
+	}
+}
+
+func TestProcidInitialized(t *testing.T) {
+	var start sync.WaitGroup
+	var done sync.WaitGroup
+
+	count := 100
+	start.Add(count + 1)
+	done.Add(count + 1)
+
+	// Run the check on the main thread.
+	//
+	// When cgo is not included, the only case when procid isn't initialized
+	// is in the main (initial) thread, so we have to test this case
+	// specifically.
+	runOnMain <- func() {
+		checkProcid(t, &start, &done)
+	}
+
+	// Run the check on a number of different threads.
+	for i := 0; i < count; i++ {
+		go checkProcid(t, &start, &done)
+	}
+
+	done.Wait()
+}
+
+func TestMain(m *testing.M) {
+	// Make sure we remain at the main (initial) thread.
+	runtime.LockOSThread()
+
+	// Start running tests in a different goroutine.
+	go func() {
+		os.Exit(m.Run())
+	}()
+
+	// Execute any functions that have been sent for execution in the main
+	// thread.
+	for f := range runOnMain {
+		f()
+	}
+}
diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD
new file mode 100644
index 000000000..16b0b3c69
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/BUILD
@@ -0,0 +1,31 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "ptrace",
+    srcs = [
+        "ptrace.go",
+        "ptrace_unsafe.go",
+        "stub_amd64.s",
+        "stub_unsafe.go",
+        "subprocess.go",
+        "subprocess_amd64.go",
+        "subprocess_linux.go",
+        "subprocess_linux_amd64_unsafe.go",
+        "subprocess_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/ptrace",
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/platform/filemem",
+        "//pkg/sentry/platform/interrupt",
+        "//pkg/sentry/platform/procid",
+        "//pkg/sentry/platform/safecopy",
+        "//pkg/sentry/usermem",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go
new file mode 100644
index 000000000..05f8b1d05
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/ptrace.go
@@ -0,0 +1,242 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package ptrace provides a ptrace-based implementation of the platform
+// interface. This is useful for development and testing purposes primarily,
+// and runs on stock kernels without special permissions.
+//
+// In a nutshell, it works as follows:
+//
+// The creation of a new address space creates a new child processes with a
+// single thread which is traced by a single goroutine.
+//
+// A context is just a collection of temporary variables. Calling Switch on a
+// context does the following:
+//
+//	Locks the runtime thread.
+//
+//	Looks up a traced subprocess thread for the current runtime thread. If
+//	none exists, the dedicated goroutine is asked to create a new stopped
+//	thread in the subprocess. This stopped subprocess thread is then traced
+//	by the current thread and this information is stored for subsequent
+//	switches.
+//
+//	The context is then bound with information about the subprocess thread
+//	so that the context may be appropriately interrupted via a signal.
+//
+//	The requested operation is performed in the traced subprocess thread
+//	(e.g. set registers, execute, return).
+//
+// FIXME: This package is currently sloppy with cleanup.
+//
+// Lock order:
+//
+// subprocess.mu
+//   context.mu
+package ptrace
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/filemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/interrupt"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+var (
+	// stubStart is the link address for our stub, and determines the
+	// maximum user address. This is valid only after a call to stubInit.
+	//
+	// We attempt to link the stub here, and adjust downward as needed.
+	stubStart uintptr = 0x7fffffff0000
+
+	// stubEnd is the first byte past the end of the stub, as with
+	// stubStart this is valid only after a call to stubInit.
+	stubEnd uintptr
+
+	// stubInitialized controls one-time stub initialization.
+	stubInitialized sync.Once
+)
+
+type context struct {
+	// signalInfo is the signal info, if and when a signal is received.
+	signalInfo arch.SignalInfo
+
+	// interrupt is the interrupt context.
+	interrupt interrupt.Forwarder
+
+	// mu protects the following fields.
+	mu sync.Mutex
+
+	// If lastFaultSP is non-nil, the last context switch was due to a fault
+	// received while executing lastFaultSP. Only context.Switch may set
+	// lastFaultSP to a non-nil value.
+	lastFaultSP *subprocess
+
+	// lastFaultAddr is the last faulting address; this is only meaningful if
+	// lastFaultSP is non-nil.
+	lastFaultAddr usermem.Addr
+
+	// lastFaultIP is the address of the last faulting instruction;
+	// this is also only meaningful if lastFaultSP is non-nil.
+	lastFaultIP usermem.Addr
+}
+
+// Switch runs the provided context in the given address space.
+func (c *context) Switch(as platform.AddressSpace, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error) {
+	s := as.(*subprocess)
+	isSyscall := s.switchToApp(c, ac)
+
+	var faultSP *subprocess
+	var faultAddr usermem.Addr
+	var faultIP usermem.Addr
+	if !isSyscall && linux.Signal(c.signalInfo.Signo) == linux.SIGSEGV {
+		faultSP = s
+		faultAddr = usermem.Addr(c.signalInfo.Addr())
+		faultIP = usermem.Addr(ac.IP())
+	}
+
+	// Update the context to reflect the outcome of this context switch.
+	c.mu.Lock()
+	lastFaultSP := c.lastFaultSP
+	lastFaultAddr := c.lastFaultAddr
+	lastFaultIP := c.lastFaultIP
+	// At this point, c may not yet be in s.contexts, so c.lastFaultSP won't be
+	// updated by s.Unmap(). This is fine; we only need to synchronize with
+	// calls to s.Unmap() that occur after the handling of this fault.
+	c.lastFaultSP = faultSP
+	c.lastFaultAddr = faultAddr
+	c.lastFaultIP = faultIP
+	c.mu.Unlock()
+
+	// Update subprocesses to reflect the outcome of this context switch.
+	if lastFaultSP != faultSP {
+		if lastFaultSP != nil {
+			lastFaultSP.mu.Lock()
+			delete(lastFaultSP.contexts, c)
+			lastFaultSP.mu.Unlock()
+		}
+		if faultSP != nil {
+			faultSP.mu.Lock()
+			faultSP.contexts[c] = struct{}{}
+			faultSP.mu.Unlock()
+		}
+	}
+
+	if isSyscall {
+		return nil, usermem.NoAccess, nil
+	}
+	if faultSP == nil {
+		// Non-fault signal.
+		return &c.signalInfo, usermem.NoAccess, platform.ErrContextSignal
+	}
+
+	// Got a page fault. Ideally, we'd get real fault type here, but ptrace
+	// doesn't expose this information. Instead, we use a simple heuristic:
+	//
+	// It was an instruction fault iff the faulting addr == instruction
+	// pointer.
+	//
+	// It was a write fault if the fault is immediately repeated.
+	at := usermem.Read
+	if faultAddr == faultIP {
+		at.Execute = true
+	}
+	if lastFaultSP == faultSP &&
+		lastFaultAddr == faultAddr &&
+		lastFaultIP == faultIP {
+		at.Write = true
+	}
+	return &c.signalInfo, at, platform.ErrContextSignal
+}
+
+// Interrupt interrupts the running guest application associated with this context.
+func (c *context) Interrupt() {
+	c.interrupt.NotifyInterrupt()
+}
+
+// PTrace represents a collection of ptrace subprocesses.
+type PTrace struct {
+	platform.MMapMinAddr
+	platform.NoCPUPreemptionDetection
+	*filemem.FileMem
+}
+
+// New returns a new ptrace-based implementation of the platform interface.
+func New() (*PTrace, error) {
+	stubInitialized.Do(func() {
+		// Initialize the stub.
+		stubInit()
+
+		// Create the master process for the global pool. This must be
+		// done before initializing any other processes.
+		master, err := newSubprocess(createStub)
+		if err != nil {
+			// Should never happen.
+			panic("unable to initialize ptrace master: " + err.Error())
+		}
+
+		// Set the master on the globalPool.
+		globalPool.master = master
+	})
+
+	fm, err := filemem.New("ptrace-memory")
+	if err != nil {
+		return nil, err
+	}
+
+	return &PTrace{FileMem: fm}, nil
+}
+
+// SupportsAddressSpaceIO implements platform.Platform.SupportsAddressSpaceIO.
+func (*PTrace) SupportsAddressSpaceIO() bool {
+	return false
+}
+
+// CooperativelySchedulesAddressSpace implements platform.Platform.CooperativelySchedulesAddressSpace.
+func (*PTrace) CooperativelySchedulesAddressSpace() bool {
+	return false
+}
+
+// MapUnit implements platform.Platform.MapUnit.
+func (*PTrace) MapUnit() uint64 {
+	// The host kernel manages page tables and arbitrary-sized mappings
+	// have effectively the same cost.
+	return 0
+}
+
+// MaxUserAddress returns the first address that may not be used by user
+// applications.
+func (*PTrace) MaxUserAddress() usermem.Addr {
+	return usermem.Addr(stubStart)
+}
+
+// NewAddressSpace returns a new subprocess.
+func (p *PTrace) NewAddressSpace(_ interface{}) (platform.AddressSpace, <-chan struct{}, error) {
+	as, err := newSubprocess(globalPool.master.createStub)
+	return as, nil, err
+}
+
+// NewContext returns an interruptible context.
+func (*PTrace) NewContext() platform.Context {
+	return &context{}
+}
+
+// Memory returns the platform memory used to do allocations.
+func (p *PTrace) Memory() platform.Memory {
+	return p.FileMem
+}
diff --git a/pkg/sentry/platform/ptrace/ptrace_unsafe.go b/pkg/sentry/platform/ptrace/ptrace_unsafe.go
new file mode 100644
index 000000000..b55b2795a
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/ptrace_unsafe.go
@@ -0,0 +1,166 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ptrace
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// GETREGSET/SETREGSET register set types.
+//
+// See include/uapi/linux/elf.h.
+const (
+	// _NT_PRFPREG is for x86 floating-point state without using xsave.
+	_NT_PRFPREG = 0x2
+
+	// _NT_X86_XSTATE is for x86 extended state using xsave.
+	_NT_X86_XSTATE = 0x202
+)
+
+// fpRegSet returns the GETREGSET/SETREGSET register set type to be used.
+func fpRegSet(useXsave bool) uintptr {
+	if useXsave {
+		return _NT_X86_XSTATE
+	}
+	return _NT_PRFPREG
+}
+
+// getRegs sets the regular register set.
+func (t *thread) getRegs(regs *syscall.PtraceRegs) error {
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_GETREGS,
+		uintptr(t.tid),
+		0,
+		uintptr(unsafe.Pointer(regs)),
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// setRegs sets the regular register set.
+func (t *thread) setRegs(regs *syscall.PtraceRegs) error {
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_SETREGS,
+		uintptr(t.tid),
+		0,
+		uintptr(unsafe.Pointer(regs)),
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// getFPRegs gets the floating-point data via the GETREGSET ptrace syscall.
+func (t *thread) getFPRegs(fpState *arch.FloatingPointData, fpLen uint64, useXsave bool) error {
+	iovec := syscall.Iovec{
+		Base: (*byte)(fpState),
+		Len:  fpLen,
+	}
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_GETREGSET,
+		uintptr(t.tid),
+		fpRegSet(useXsave),
+		uintptr(unsafe.Pointer(&iovec)),
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// setFPRegs sets the floating-point data via the SETREGSET ptrace syscall.
+func (t *thread) setFPRegs(fpState *arch.FloatingPointData, fpLen uint64, useXsave bool) error {
+	iovec := syscall.Iovec{
+		Base: (*byte)(fpState),
+		Len:  fpLen,
+	}
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_SETREGSET,
+		uintptr(t.tid),
+		fpRegSet(useXsave),
+		uintptr(unsafe.Pointer(&iovec)),
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// getSignalInfo retrieves information about the signal that caused the stop.
+func (t *thread) getSignalInfo(si *arch.SignalInfo) error {
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_GETSIGINFO,
+		uintptr(t.tid),
+		0,
+		uintptr(unsafe.Pointer(si)),
+		0, 0)
+	if errno != 0 {
+		return errno
+	}
+	return nil
+}
+
+// clone creates a new thread from this one.
+//
+// The returned thread will be stopped and available for any system thread to
+// call attach on it.
+//
+// Precondition: the OS thread must be locked and own t.
+func (t *thread) clone(initRegs *syscall.PtraceRegs) (*thread, error) {
+	r, ok := usermem.Addr(initRegs.Rsp).RoundUp()
+	if !ok {
+		return nil, syscall.EINVAL
+	}
+	rval, err := t.syscallIgnoreInterrupt(
+		initRegs,
+		syscall.SYS_CLONE,
+		arch.SyscallArgument{Value: uintptr(
+			syscall.CLONE_FILES |
+				syscall.CLONE_FS |
+				syscall.CLONE_SIGHAND |
+				syscall.CLONE_THREAD |
+				syscall.CLONE_PTRACE |
+				syscall.CLONE_VM)},
+		// The stack pointer is just made up, but we have it be
+		// something sensible so the kernel doesn't think we're
+		// up to no good. Which we are.
+		arch.SyscallArgument{Value: uintptr(r)},
+		arch.SyscallArgument{},
+		arch.SyscallArgument{},
+		// We use these registers initially, but really they
+		// could be anything. We're going to stop immediately.
+		arch.SyscallArgument{Value: uintptr(unsafe.Pointer(initRegs))})
+	if err != nil {
+		return nil, err
+	}
+
+	return &thread{
+		tgid: t.tgid,
+		tid:  int32(rval),
+		cpu:  ^uint32(0),
+	}, nil
+}
diff --git a/pkg/sentry/platform/ptrace/stub_amd64.s b/pkg/sentry/platform/ptrace/stub_amd64.s
new file mode 100644
index 000000000..9bf87b6f6
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/stub_amd64.s
@@ -0,0 +1,114 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "funcdata.h"
+#include "textflag.h"
+
+#define SYS_GETPID		39
+#define SYS_EXIT		60
+#define SYS_KILL		62
+#define SYS_GETPPID		110
+#define SYS_PRCTL		157
+
+#define SIGKILL			9
+#define SIGSTOP			19
+
+#define PR_SET_PDEATHSIG	1
+
+// stub bootstraps the child and sends itself SIGSTOP to wait for attach.
+//
+// R15 contains the expected PPID. R15 is used instead of a more typical DI
+// since syscalls will clobber DI and createStub wants to pass a new PPID to
+// grandchildren.
+//
+// This should not be used outside the context of a new ptrace child (as the
+// function is otherwise a bunch of nonsense).
+TEXT ·stub(SB),NOSPLIT,$0
+begin:
+	// N.B. This loop only executes in the context of a single-threaded
+	// fork child.
+
+	MOVQ $SYS_PRCTL, AX
+	MOVQ $PR_SET_PDEATHSIG, DI
+	MOVQ $SIGKILL, SI
+	SYSCALL
+
+	CMPQ AX, $0
+	JNE error
+
+	// If the parent already died before we called PR_SET_DEATHSIG then
+	// we'll have an unexpected PPID.
+	MOVQ $SYS_GETPPID, AX
+	SYSCALL
+
+	CMPQ AX, $0
+	JL error
+
+	CMPQ AX, R15
+	JNE parent_dead
+
+	MOVQ $SYS_GETPID, AX
+	SYSCALL
+
+	CMPQ AX, $0
+	JL error
+
+	// SIGSTOP to wait for attach.
+	//
+	// The SYSCALL instruction will be used for future syscall injection by
+	// thread.syscall.
+	MOVQ AX, DI
+	MOVQ $SYS_KILL, AX
+	MOVQ $SIGSTOP, SI
+	SYSCALL
+
+	// The tracer may "detach" and/or allow code execution here in three cases:
+	//
+	// 1. New (traced) stub threads are explicitly detached by the
+	// goroutine in newSubprocess. However, they are detached while in
+	// group-stop, so they do not execute code here.
+	//
+	// 2. If a tracer thread exits, it implicitly detaches from the stub,
+	// potentially allowing code execution here. However, the Go runtime
+	// never exits individual threads, so this case never occurs.
+	//
+	// 3. subprocess.createStub clones a new stub process that is untraced,
+	// thus executing this code. We setup the PDEATHSIG before SIGSTOPing
+	// ourselves for attach by the tracer.
+	//
+	// R15 has been updated with the expected PPID.
+	JMP begin
+
+error:
+	// Exit with -errno.
+	MOVQ AX, DI
+	NEGQ DI
+	MOVQ $SYS_EXIT, AX
+	SYSCALL
+	HLT
+
+parent_dead:
+	MOVQ $SYS_EXIT, AX
+	MOVQ $1, DI
+	SYSCALL
+	HLT
+
+// stubCall calls the stub function at the given address with the given PPID.
+//
+// This is a distinct function because stub, above, may be mapped at any
+// arbitrary location, and stub has a specific binary API (see above).
+TEXT ·stubCall(SB),NOSPLIT,$0-16
+	MOVQ addr+0(FP), AX
+	MOVQ pid+8(FP), R15
+	JMP AX
diff --git a/pkg/sentry/platform/ptrace/stub_unsafe.go b/pkg/sentry/platform/ptrace/stub_unsafe.go
new file mode 100644
index 000000000..c868a2d68
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/stub_unsafe.go
@@ -0,0 +1,98 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ptrace
+
+import (
+	"reflect"
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/safecopy"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// stub is defined in arch-specific assembly.
+func stub()
+
+// stubCall calls the stub at the given address with the given pid.
+func stubCall(addr, pid uintptr)
+
+// unsafeSlice returns a slice for the given address and length.
+func unsafeSlice(addr uintptr, length int) (slice []byte) {
+	sh := (*reflect.SliceHeader)(unsafe.Pointer(&slice))
+	sh.Data = addr
+	sh.Len = length
+	sh.Cap = length
+	return
+}
+
+// stubInit initializes the stub.
+func stubInit() {
+	// Grab the existing stub.
+	stubBegin := reflect.ValueOf(stub).Pointer()
+	stubLen := int(safecopy.FindEndAddress(stubBegin) - stubBegin)
+	stubSlice := unsafeSlice(stubBegin, stubLen)
+	mapLen := uintptr(stubLen)
+	if offset := mapLen % usermem.PageSize; offset != 0 {
+		mapLen += usermem.PageSize - offset
+	}
+
+	for stubStart > 0 {
+		// Map the target address for the stub.
+		//
+		// We don't use FIXED here because we don't want to unmap
+		// something that may have been there already. We just walk
+		// down the address space until we find a place where the stub
+		// can be placed.
+		addr, _, errno := syscall.RawSyscall6(
+			syscall.SYS_MMAP,
+			stubStart,
+			mapLen,
+			syscall.PROT_WRITE|syscall.PROT_READ,
+			syscall.MAP_PRIVATE|syscall.MAP_ANONYMOUS,
+			0 /* fd */, 0 /* offset */)
+		if addr != stubStart || errno != 0 {
+			if addr != 0 {
+				// Unmap the region we've mapped accidentally.
+				syscall.RawSyscall(syscall.SYS_MUNMAP, addr, mapLen, 0)
+			}
+
+			// Attempt to begin at a lower address.
+			stubStart -= uintptr(usermem.PageSize)
+			continue
+		}
+
+		// Copy the stub to the address.
+		targetSlice := unsafeSlice(addr, stubLen)
+		copy(targetSlice, stubSlice)
+
+		// Make the stub executable.
+		if _, _, errno := syscall.RawSyscall(
+			syscall.SYS_MPROTECT,
+			stubStart,
+			mapLen,
+			syscall.PROT_EXEC|syscall.PROT_READ); errno != 0 {
+			panic("mprotect failed: " + errno.Error())
+		}
+
+		// Set the end.
+		stubEnd = stubStart + mapLen
+		return
+	}
+
+	// This will happen only if we exhaust the entire address
+	// space, and it will take a long, long time.
+	panic("failed to map stub")
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go
new file mode 100644
index 000000000..0d6a38f15
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/subprocess.go
@@ -0,0 +1,559 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ptrace
+
+import (
+	"fmt"
+	"os"
+	"runtime"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// globalPool exists to solve two distinct problems:
+//
+// 1) Subprocesses can't always be killed properly (see Release).
+//
+// 2) Any seccomp filters that have been installed will apply to subprocesses
+// created here. Therefore we use the intermediary (master), which is created
+// on initialization of the platform.
+var globalPool struct {
+	mu        sync.Mutex
+	master    *subprocess
+	available []*subprocess
+}
+
+// thread is a traced thread; it is a thread identifier.
+//
+// This is a convenience type for defining ptrace operations.
+type thread struct {
+	tgid int32
+	tid  int32
+	cpu  uint32
+}
+
+// threadPool is a collection of threads.
+type threadPool struct {
+	// mu protects below.
+	mu sync.Mutex
+
+	// threads is the collection of threads.
+	//
+	// This map is indexed by system TID (the calling thread); which will
+	// be the tracer for the given *thread, and therefore capable of using
+	// relevant ptrace calls.
+	threads map[int32]*thread
+}
+
+// lookupOrCreate looks up a given thread or creates one.
+//
+// newThread will generally be subprocess.newThread.
+//
+// Precondition: the runtime OS thread must be locked.
+func (tp *threadPool) lookupOrCreate(currentTID int32, newThread func() *thread) *thread {
+	tp.mu.Lock()
+	t, ok := tp.threads[currentTID]
+	if !ok {
+		// Before creating a new thread, see if we can find a thread
+		// whose system tid has disappeared.
+		//
+		// TODO: Other parts of this package depend on
+		// threads never exiting.
+		for origTID, t := range tp.threads {
+			// Signal zero is an easy existence check.
+			if err := syscall.Tgkill(syscall.Getpid(), int(origTID), 0); err != nil {
+				// This thread has been abandoned; reuse it.
+				delete(tp.threads, origTID)
+				tp.threads[currentTID] = t
+				tp.mu.Unlock()
+				return t
+			}
+		}
+
+		// Create a new thread.
+		t = newThread()
+		tp.threads[currentTID] = t
+	}
+	tp.mu.Unlock()
+	return t
+}
+
+// subprocess is a collection of threads being traced.
+type subprocess struct {
+	platform.NoAddressSpaceIO
+
+	// initRegs are the initial registers for the first thread.
+	//
+	// These are used for the register set for system calls.
+	initRegs syscall.PtraceRegs
+
+	// requests is used to signal creation of new threads.
+	requests chan chan *thread
+
+	// sysemuThreads are reserved for emulation.
+	sysemuThreads threadPool
+
+	// syscallThreads are reserved for syscalls (except clone, which is
+	// handled in the dedicated goroutine corresponding to requests above).
+	syscallThreads threadPool
+
+	// mu protects the following fields.
+	mu sync.Mutex
+
+	// contexts is the set of contexts for which it's possible that
+	// context.lastFaultSP == this subprocess.
+	contexts map[*context]struct{}
+}
+
+// newSubprocess returns a useable subprocess.
+//
+// This will either be a newly created subprocess, or one from the global pool.
+// The create function will be called in the latter case, which is guaranteed
+// to happen with the runtime thread locked.
+func newSubprocess(create func() (*thread, error)) (*subprocess, error) {
+	// See Release.
+	globalPool.mu.Lock()
+	if len(globalPool.available) > 0 {
+		sp := globalPool.available[len(globalPool.available)-1]
+		globalPool.available = globalPool.available[:len(globalPool.available)-1]
+		globalPool.mu.Unlock()
+		return sp, nil
+	}
+	globalPool.mu.Unlock()
+
+	// The following goroutine is responsible for creating the first traced
+	// thread, and responding to requests to make additional threads in the
+	// traced process. The process will be killed and reaped when the
+	// request channel is closed, which happens in Release below.
+	var initRegs syscall.PtraceRegs
+	errChan := make(chan error)
+	requests := make(chan chan *thread)
+	go func() { // S/R-SAFE: Platform-related.
+		runtime.LockOSThread()
+		defer runtime.UnlockOSThread()
+
+		// Initialize the first thread.
+		firstThread, err := create()
+		if err != nil {
+			errChan <- err
+			return
+		}
+
+		// Grab registers.
+		//
+		// Note that we adjust the current register RIP value to be
+		// just before the current system call executed. This depends
+		// on the definition of the stub itself.
+		if err := firstThread.getRegs(&initRegs); err != nil {
+			panic(fmt.Sprintf("ptrace get regs failed: %v", err))
+		}
+		initRegs.Rip -= initRegsRipAdjustment
+
+		// Ready to handle requests.
+		errChan <- nil
+
+		// Wait for requests to create threads.
+		for r := range requests {
+			t, err := firstThread.clone(&initRegs)
+			if err != nil {
+				// Should not happen: not recoverable.
+				panic(fmt.Sprintf("error initializing first thread: %v", err))
+			}
+
+			// Since the new thread was created with
+			// clone(CLONE_PTRACE), it will begin execution with
+			// SIGSTOP pending and with this thread as its tracer.
+			// (Hopefully nobody tgkilled it with a signal <
+			// SIGSTOP before the SIGSTOP was delivered, in which
+			// case that signal would be delivered before SIGSTOP.)
+			if sig := t.wait(); sig != syscall.SIGSTOP {
+				panic(fmt.Sprintf("error waiting for new clone: expected SIGSTOP, got %v", sig))
+			}
+
+			// Detach the thread without suppressing the SIGSTOP,
+			// causing it to enter group-stop.
+			if _, _, errno := syscall.RawSyscall6(syscall.SYS_PTRACE, syscall.PTRACE_DETACH, uintptr(t.tid), 0, uintptr(syscall.SIGSTOP), 0, 0); errno != 0 {
+				panic(fmt.Sprintf("can't detach new clone: %v", errno))
+			}
+
+			// Return the thread.
+			r <- t
+		}
+
+		// Requests should never be closed.
+		panic("unreachable")
+	}()
+
+	// Wait until error or readiness.
+	if err := <-errChan; err != nil {
+		return nil, err
+	}
+
+	// Ready.
+	sp := &subprocess{
+		initRegs: initRegs,
+		requests: requests,
+		sysemuThreads: threadPool{
+			threads: make(map[int32]*thread),
+		},
+		syscallThreads: threadPool{
+			threads: make(map[int32]*thread),
+		},
+		contexts: make(map[*context]struct{}),
+	}
+
+	sp.unmap()
+	return sp, nil
+}
+
+// unmap unmaps non-stub regions of the process.
+//
+// This will panic on failure (which should never happen).
+func (s *subprocess) unmap() {
+	s.Unmap(0, uint64(stubStart))
+	if maximumUserAddress != stubEnd {
+		s.Unmap(usermem.Addr(stubEnd), uint64(maximumUserAddress-stubEnd))
+	}
+}
+
+// Release kills the subprocess.
+//
+// Just kidding! We can't safely co-ordinate the detaching of all the
+// tracees (since the tracers are random runtime threads, and the process
+// won't exit until tracers have been notifier).
+//
+// Therefore we simply unmap everything in the subprocess and return it to the
+// globalPool. This has the added benefit of reducing creation time for new
+// subprocesses.
+func (s *subprocess) Release() error {
+	go func() { // S/R-SAFE: Platform.
+		s.unmap()
+		globalPool.mu.Lock()
+		globalPool.available = append(globalPool.available, s)
+		globalPool.mu.Unlock()
+	}()
+	return nil
+}
+
+// newThread creates a new traced thread.
+//
+// Precondition: the OS thread must be locked.
+func (s *subprocess) newThread() *thread {
+	// Ask the first thread to create a new one.
+	r := make(chan *thread)
+	s.requests <- r
+	t := <-r
+
+	// Attach the subprocess to this one.
+	t.attach()
+
+	// Return the new thread, which is now bound.
+	return t
+}
+
+// attach attachs to the thread.
+func (t *thread) attach() {
+	if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_ATTACH, uintptr(t.tid), 0); errno != 0 {
+		panic(fmt.Sprintf("unable to attach: %v", errno))
+	}
+
+	// PTRACE_ATTACH sends SIGSTOP, and wakes the tracee if it was already
+	// stopped from the SIGSTOP queued by CLONE_PTRACE (see inner loop of
+	// newSubprocess), so we always expect to see signal-delivery-stop with
+	// SIGSTOP.
+	if sig := t.wait(); sig != syscall.SIGSTOP {
+		panic(fmt.Sprintf("wait failed: expected SIGSTOP, got %v", sig))
+	}
+
+	// Initialize options.
+	t.init()
+}
+
+// wait waits for a stop event.
+func (t *thread) wait() syscall.Signal {
+	var status syscall.WaitStatus
+
+	for {
+		r, err := syscall.Wait4(int(t.tid), &status, syscall.WALL|syscall.WUNTRACED, nil)
+		if err == syscall.EINTR || err == syscall.EAGAIN {
+			// Wait was interrupted; wait again.
+			continue
+		} else if err != nil {
+			panic(fmt.Sprintf("ptrace wait failed: %v", err))
+		}
+		if int(r) != int(t.tid) {
+			panic(fmt.Sprintf("ptrace wait returned %v, expected %v", r, t.tid))
+		}
+		if !status.Stopped() {
+			panic(fmt.Sprintf("ptrace status unexpected: got %v, wanted stopped", status))
+		}
+		if status.StopSignal() == 0 {
+			continue // Spurious stop.
+		}
+		return status.StopSignal()
+	}
+}
+
+// init initializes trace options.
+func (t *thread) init() {
+	// Set our TRACESYSGOOD option to differeniate real SIGTRAP.
+	_, _, errno := syscall.RawSyscall6(
+		syscall.SYS_PTRACE,
+		syscall.PTRACE_SETOPTIONS,
+		uintptr(t.tid),
+		0,
+		syscall.PTRACE_O_TRACESYSGOOD,
+		0, 0)
+	if errno != 0 {
+		panic(fmt.Sprintf("ptrace set options failed: %v", errno))
+	}
+}
+
+// syscall executes a system call cycle in the traced context.
+//
+// This is _not_ for use by application system calls, rather it is for use when
+// a system call must be injected into the remote context (e.g. mmap, munmap).
+// Note that clones are handled separately.
+func (t *thread) syscall(regs *syscall.PtraceRegs) (uintptr, error) {
+	// Set registers.
+	if err := t.setRegs(regs); err != nil {
+		panic(fmt.Sprintf("ptrace set regs failed: %v", err))
+	}
+
+	for {
+		// Execute the syscall instruction.
+		if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0); errno != 0 {
+			panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno))
+		}
+
+		sig := t.wait()
+		if sig == (0x80 | syscall.SIGTRAP) {
+			// Reached syscall-enter-stop.
+			break
+		} else {
+			// Some other signal caused a thread stop; ignore.
+			continue
+		}
+	}
+
+	// Complete the actual system call.
+	if _, _, errno := syscall.RawSyscall(syscall.SYS_PTRACE, syscall.PTRACE_SYSCALL, uintptr(t.tid), 0); errno != 0 {
+		panic(fmt.Sprintf("ptrace syscall-enter failed: %v", errno))
+	}
+
+	// Wait for syscall-exit-stop. "[Signal-delivery-stop] never happens
+	// between syscall-enter-stop and syscall-exit-stop; it happens *after*
+	// syscall-exit-stop.)" - ptrace(2), "Syscall-stops"
+	if sig := t.wait(); sig != (0x80 | syscall.SIGTRAP) {
+		panic(fmt.Sprintf("wait failed: expected SIGTRAP, got %v [%d]", sig, sig))
+	}
+
+	// Grab registers.
+	if err := t.getRegs(regs); err != nil {
+		panic(fmt.Sprintf("ptrace get regs failed: %v", err))
+	}
+
+	return syscallReturnValue(regs)
+}
+
+// syscallIgnoreInterrupt ignores interrupts on the system call thread and
+// restarts the syscall if the kernel indicates that should happen.
+func (t *thread) syscallIgnoreInterrupt(
+	initRegs *syscall.PtraceRegs,
+	sysno uintptr,
+	args ...arch.SyscallArgument) (uintptr, error) {
+	for {
+		regs := createSyscallRegs(initRegs, sysno, args...)
+		rval, err := t.syscall(&regs)
+		switch err {
+		case ERESTARTSYS:
+			continue
+		case ERESTARTNOINTR:
+			continue
+		case ERESTARTNOHAND:
+			continue
+		default:
+			return rval, err
+		}
+	}
+}
+
+// NotifyInterrupt implements interrupt.Receiver.NotifyInterrupt.
+func (t *thread) NotifyInterrupt() {
+	syscall.Tgkill(int(t.tgid), int(t.tid), syscall.Signal(platform.SignalInterrupt))
+}
+
+// switchToApp is called from the main SwitchToApp entrypoint.
+//
+// This function returns true on a system call, false on a signal.
+func (s *subprocess) switchToApp(c *context, ac arch.Context) bool {
+	regs := &ac.StateData().Regs
+	s.resetSysemuRegs(regs)
+
+	// Extract floating point state.
+	fpState := ac.FloatingPointData()
+	fpLen, _ := ac.FeatureSet().ExtendedStateSize()
+	useXsave := ac.FeatureSet().UseXsave()
+
+	// Lock the thread for ptrace operations.
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	// Grab our thread from the pool.
+	currentTID := int32(procid.Current())
+	t := s.sysemuThreads.lookupOrCreate(currentTID, s.newThread)
+
+	// Check for interrupts, and ensure that future interrupts will signal t.
+	if !c.interrupt.Enable(t) {
+		// Pending interrupt; simulate.
+		c.signalInfo = arch.SignalInfo{Signo: int32(platform.SignalInterrupt)}
+		return false
+	}
+	defer c.interrupt.Disable()
+
+	// Ensure that the CPU set is bound appropriately; this makes the
+	// emulation below several times faster, presumably by avoiding
+	// interprocessor wakeups and by simplifying the schedule.
+	t.bind()
+
+	// Set registers.
+	if err := t.setRegs(regs); err != nil {
+		panic(fmt.Sprintf("ptrace set regs failed: %v", err))
+	}
+	if err := t.setFPRegs(fpState, uint64(fpLen), useXsave); err != nil {
+		panic(fmt.Sprintf("ptrace set fpregs failed: %v", err))
+	}
+
+	for {
+		// Start running until the next system call.
+		if isSingleStepping(regs) {
+			if _, _, errno := syscall.RawSyscall(
+				syscall.SYS_PTRACE,
+				syscall.PTRACE_SYSEMU_SINGLESTEP,
+				uintptr(t.tid), 0); errno != 0 {
+				panic(fmt.Sprintf("ptrace sysemu failed: %v", errno))
+			}
+		} else {
+			if _, _, errno := syscall.RawSyscall(
+				syscall.SYS_PTRACE,
+				syscall.PTRACE_SYSEMU,
+				uintptr(t.tid), 0); errno != 0 {
+				panic(fmt.Sprintf("ptrace sysemu failed: %v", errno))
+			}
+		}
+
+		// Wait for the syscall-enter stop.
+		sig := t.wait()
+
+		// Refresh all registers.
+		if err := t.getRegs(regs); err != nil {
+			panic(fmt.Sprintf("ptrace get regs failed: %v", err))
+		}
+		if err := t.getFPRegs(fpState, uint64(fpLen), useXsave); err != nil {
+			panic(fmt.Sprintf("ptrace get fpregs failed: %v", err))
+		}
+
+		// Is it a system call?
+		if sig == (0x80 | syscall.SIGTRAP) {
+			// Ensure registers are sane.
+			updateSyscallRegs(regs)
+			return true
+		}
+
+		if sig == syscall.SIGSTOP {
+			// SIGSTOP was delivered to another thread in the same thread
+			// group, which initiated another group stop. Just ignore it.
+			continue
+		}
+
+		// Grab signal information.
+		if err := t.getSignalInfo(&c.signalInfo); err != nil {
+			// Should never happen.
+			panic(fmt.Sprintf("ptrace get signal info failed: %v", err))
+		}
+
+		// We have a signal. We verify however, that the signal was
+		// either delivered from the kernel or from this process. We
+		// don't respect other signals.
+		if c.signalInfo.Code > 0 {
+			return false // kernel.
+		} else if c.signalInfo.Code <= 0 && c.signalInfo.Pid() == int32(os.Getpid()) {
+			return false // this process.
+		}
+	}
+}
+
+// syscall executes the given system call without handling interruptions.
+func (s *subprocess) syscall(sysno uintptr, args ...arch.SyscallArgument) (uintptr, error) {
+	// Grab a thread.
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+	currentTID := int32(procid.Current())
+	t := s.syscallThreads.lookupOrCreate(currentTID, s.newThread)
+
+	return t.syscallIgnoreInterrupt(&s.initRegs, sysno, args...)
+}
+
+// MapFile implements platform.AddressSpace.MapFile.
+func (s *subprocess) MapFile(addr usermem.Addr, fd int, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+	var flags int
+	if precommit {
+		flags |= syscall.MAP_POPULATE
+	}
+	_, err := s.syscall(
+		syscall.SYS_MMAP,
+		arch.SyscallArgument{Value: uintptr(addr)},
+		arch.SyscallArgument{Value: uintptr(fr.Length())},
+		arch.SyscallArgument{Value: uintptr(at.Prot())},
+		arch.SyscallArgument{Value: uintptr(flags | syscall.MAP_SHARED | syscall.MAP_FIXED)},
+		arch.SyscallArgument{Value: uintptr(fd)},
+		arch.SyscallArgument{Value: uintptr(fr.Start)})
+	return err
+}
+
+// Unmap implements platform.AddressSpace.Unmap.
+func (s *subprocess) Unmap(addr usermem.Addr, length uint64) {
+	ar, ok := addr.ToRange(length)
+	if !ok {
+		panic(fmt.Sprintf("addr %#x + length %#x overflows", addr, length))
+	}
+	s.mu.Lock()
+	for c := range s.contexts {
+		c.mu.Lock()
+		if c.lastFaultSP == s && ar.Contains(c.lastFaultAddr) {
+			// Forget the last fault so that if c faults again, the fault isn't
+			// incorrectly reported as a write fault. If this is being called
+			// due to munmap() of the corresponding vma, handling of the second
+			// fault will fail anyway.
+			c.lastFaultSP = nil
+			delete(s.contexts, c)
+		}
+		c.mu.Unlock()
+	}
+	s.mu.Unlock()
+	_, err := s.syscall(
+		syscall.SYS_MUNMAP,
+		arch.SyscallArgument{Value: uintptr(addr)},
+		arch.SyscallArgument{Value: uintptr(length)})
+	if err != nil {
+		// We never expect this to happen.
+		panic(fmt.Sprintf("munmap(%x, %x)) failed: %v", addr, length, err))
+	}
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go
new file mode 100644
index 000000000..8211215df
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go
@@ -0,0 +1,104 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ptrace
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+)
+
+const (
+	// maximumUserAddress is the largest possible user address.
+	maximumUserAddress = 0x7ffffffff000
+
+	// initRegsRipAdjustment is the size of the syscall instruction.
+	initRegsRipAdjustment = 2
+)
+
+// Linux kernel errnos which "should never be seen by user programs", but will
+// be revealed to ptrace syscall exit tracing.
+//
+// These constants are used in subprocess.go.
+const (
+	ERESTARTSYS    = syscall.Errno(512)
+	ERESTARTNOINTR = syscall.Errno(513)
+	ERESTARTNOHAND = syscall.Errno(514)
+)
+
+// resetSysemuRegs sets up emulation registers.
+//
+// This should be called prior to calling sysemu.
+func (s *subprocess) resetSysemuRegs(regs *syscall.PtraceRegs) {
+	regs.Cs = s.initRegs.Cs
+	regs.Ss = s.initRegs.Ss
+	regs.Ds = s.initRegs.Ds
+	regs.Es = s.initRegs.Es
+	regs.Fs = s.initRegs.Fs
+	regs.Gs = s.initRegs.Gs
+}
+
+// createSyscallRegs sets up syscall registers.
+//
+// This should be called to generate registers for a system call.
+func createSyscallRegs(initRegs *syscall.PtraceRegs, sysno uintptr, args ...arch.SyscallArgument) syscall.PtraceRegs {
+	// Copy initial registers (RIP, segments, etc.).
+	regs := *initRegs
+
+	// Set our syscall number.
+	regs.Rax = uint64(sysno)
+	if len(args) >= 1 {
+		regs.Rdi = args[0].Uint64()
+	}
+	if len(args) >= 2 {
+		regs.Rsi = args[1].Uint64()
+	}
+	if len(args) >= 3 {
+		regs.Rdx = args[2].Uint64()
+	}
+	if len(args) >= 4 {
+		regs.R10 = args[3].Uint64()
+	}
+	if len(args) >= 5 {
+		regs.R8 = args[4].Uint64()
+	}
+	if len(args) >= 6 {
+		regs.R9 = args[5].Uint64()
+	}
+
+	return regs
+}
+
+// isSingleStepping determines if the registers indicate single-stepping.
+func isSingleStepping(regs *syscall.PtraceRegs) bool {
+	return (regs.Eflags & arch.X86TrapFlag) != 0
+}
+
+// updateSyscallRegs updates registers after finishing sysemu.
+func updateSyscallRegs(regs *syscall.PtraceRegs) {
+	// Ptrace puts -ENOSYS in rax on syscall-enter-stop.
+	regs.Rax = regs.Orig_rax
+}
+
+// syscallReturnValue extracts a sensible return from registers.
+func syscallReturnValue(regs *syscall.PtraceRegs) (uintptr, error) {
+	rval := int64(regs.Rax)
+	if rval < 0 {
+		return 0, syscall.Errno(-rval)
+	}
+	return uintptr(rval), nil
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go
new file mode 100644
index 000000000..227dd4882
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/subprocess_linux.go
@@ -0,0 +1,146 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build linux
+
+package ptrace
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/procid"
+)
+
+// createStub creates a fresh stub processes.
+//
+// Precondition: the runtime OS thread must be locked.
+func createStub() (*thread, error) {
+	// Declare all variables up front in order to ensure that there's no
+	// need for allocations between beforeFork & afterFork.
+	var (
+		pid   uintptr
+		ppid  uintptr
+		errno syscall.Errno
+	)
+
+	// Remember the current ppid for the pdeathsig race.
+	ppid, _, _ = syscall.RawSyscall(syscall.SYS_GETPID, 0, 0, 0)
+
+	// Among other things, beforeFork masks all signals.
+	beforeFork()
+	pid, _, errno = syscall.RawSyscall6(syscall.SYS_CLONE, uintptr(syscall.SIGCHLD)|syscall.CLONE_FILES, 0, 0, 0, 0, 0)
+	if errno != 0 {
+		afterFork()
+		return nil, errno
+	}
+
+	// Is this the parent?
+	if pid != 0 {
+		// Among other things, restore signal mask.
+		afterFork()
+
+		// Initialize the first thread.
+		t := &thread{
+			tgid: int32(pid),
+			tid:  int32(pid),
+			cpu:  ^uint32(0),
+		}
+		if sig := t.wait(); sig != syscall.SIGSTOP {
+			return nil, fmt.Errorf("wait failed: expected SIGSTOP, got %v", sig)
+		}
+		t.attach()
+
+		return t, nil
+	}
+
+	// afterForkInChild resets all signals to their default dispositions
+	// and restores the signal mask to its pre-fork state.
+	afterForkInChild()
+
+	// Explicitly unmask all signals to ensure that the tracer can see
+	// them.
+	errno = unmaskAllSignals()
+	if errno != 0 {
+		syscall.RawSyscall(syscall.SYS_EXIT, uintptr(errno), 0, 0)
+	}
+
+	// Call the stub; should not return.
+	stubCall(stubStart, ppid)
+	panic("unreachable")
+}
+
+// createStub creates a stub processes as a child of an existing subprocesses.
+//
+// Precondition: the runtime OS thread must be locked.
+func (s *subprocess) createStub() (*thread, error) {
+	// There's no need to lock the runtime thread here, as this can only be
+	// called from a context that is already locked.
+	currentTID := int32(procid.Current())
+	t := s.syscallThreads.lookupOrCreate(currentTID, s.newThread)
+
+	// Pass the expected PPID to the child via R15.
+	regs := s.initRegs
+	regs.R15 = uint64(t.tgid)
+
+	// Call fork in a subprocess.
+	//
+	// The new child must set up PDEATHSIG to ensure it dies if this
+	// process dies. Since this process could die at any time, this cannot
+	// be done via instrumentation from here.
+	//
+	// Instead, we create the child untraced, which will do the PDEATHSIG
+	// setup and then SIGSTOP itself for our attach below.
+	pid, err := t.syscallIgnoreInterrupt(
+		&regs,
+		syscall.SYS_CLONE,
+		arch.SyscallArgument{Value: uintptr(syscall.SIGCHLD | syscall.CLONE_FILES)},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0})
+	if err != nil {
+		return nil, err
+	}
+
+	// Wait for child to enter group-stop, so we don't stop its
+	// bootstrapping work with t.attach below.
+	//
+	// We unfortunately don't have a handy part of memory to write the wait
+	// status. If the wait succeeds, we'll assume that it was the SIGSTOP.
+	// If the child actually exited, the attach below will fail.
+	_, err = t.syscallIgnoreInterrupt(
+		&s.initRegs,
+		syscall.SYS_WAIT4,
+		arch.SyscallArgument{Value: uintptr(pid)},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: syscall.WUNTRACED},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0},
+		arch.SyscallArgument{Value: 0})
+	if err != nil {
+		return nil, err
+	}
+
+	childT := &thread{
+		tgid: int32(pid),
+		tid:  int32(pid),
+		cpu:  ^uint32(0),
+	}
+	childT.attach()
+
+	return childT, nil
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess_linux_amd64_unsafe.go b/pkg/sentry/platform/ptrace/subprocess_linux_amd64_unsafe.go
new file mode 100644
index 000000000..697431472
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/subprocess_linux_amd64_unsafe.go
@@ -0,0 +1,109 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64 linux
+
+package ptrace
+
+import (
+	"sync"
+	"sync/atomic"
+	"syscall"
+	"unsafe"
+
+	"golang.org/x/sys/unix"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// maskPool contains reusable CPU masks for setting affinity. Unfortunately,
+// runtime.NumCPU doesn't actually record the number of CPUs on the system, it
+// just records the number of CPUs available in the scheduler affinity set at
+// startup. This may a) change over time and b) gives a number far lower than
+// the maximum indexable CPU. To prevent lots of allocation in the hot path, we
+// use a pool to store large masks that we can reuse during bind.
+var maskPool = sync.Pool{
+	New: func() interface{} {
+		const maxCPUs = 1024 // Not a hard limit; see below.
+		return make([]uintptr, maxCPUs/64)
+	},
+}
+
+// unmaskAllSignals unmasks all signals on the current thread.
+//
+//go:nosplit
+func unmaskAllSignals() syscall.Errno {
+	var set linux.SignalSet
+	_, _, errno := syscall.RawSyscall6(syscall.SYS_RT_SIGPROCMASK, linux.SIG_SETMASK, uintptr(unsafe.Pointer(&set)), 0, linux.SignalSetSize, 0, 0)
+	return errno
+}
+
+// getCPU gets the current CPU.
+//
+// Precondition: the current runtime thread should be locked.
+func getCPU() (uint32, error) {
+	var cpu uintptr
+	if _, _, errno := syscall.RawSyscall(
+		unix.SYS_GETCPU,
+		uintptr(unsafe.Pointer(&cpu)),
+		0, 0); errno != 0 {
+		return 0, errno
+	}
+	return uint32(cpu), nil
+}
+
+// setCPU sets the CPU affinity.
+func (t *thread) setCPU(cpu uint32) error {
+	mask := maskPool.Get().([]uintptr)
+	n := int(cpu / 64)
+	v := uintptr(1 << uintptr(cpu%64))
+	if n >= len(mask) {
+		// See maskPool note above. We've actually exceeded the number
+		// of available cores. Grow the mask and return it.
+		mask = make([]uintptr, n+1)
+	}
+	mask[n] |= v
+	if _, _, errno := syscall.RawSyscall(
+		unix.SYS_SCHED_SETAFFINITY,
+		uintptr(t.tid),
+		uintptr(len(mask)*8),
+		uintptr(unsafe.Pointer(&mask[0]))); errno != 0 {
+		return errno
+	}
+	mask[n] &^= v
+	maskPool.Put(mask)
+	return nil
+}
+
+// bind attempts to ensure that the thread is on the same CPU as the current
+// thread. This provides no guarantees as it is fundamentally a racy operation:
+// CPU sets may change and we may be rescheduled in the middle of this
+// operation. As a result, no failures are reported.
+//
+// Precondition: the current runtime thread should be locked.
+func (t *thread) bind() {
+	currentCPU, err := getCPU()
+	if err != nil {
+		return
+	}
+	if oldCPU := atomic.SwapUint32(&t.cpu, currentCPU); oldCPU != currentCPU {
+		// Set the affinity on the thread and save the CPU for next
+		// round; we don't expect CPUs to bounce around too frequently.
+		//
+		// (It's worth noting that we could move CPUs between this point
+		// and when the tracee finishes executing. But that would be
+		// roughly the status quo anyways -- we're just maximizing our
+		// chances of colocation, not guaranteeing it.)
+		t.setCPU(currentCPU)
+	}
+}
diff --git a/pkg/sentry/platform/ptrace/subprocess_unsafe.go b/pkg/sentry/platform/ptrace/subprocess_unsafe.go
new file mode 100644
index 000000000..fe41641d3
--- /dev/null
+++ b/pkg/sentry/platform/ptrace/subprocess_unsafe.go
@@ -0,0 +1,28 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ptrace
+
+import (
+	_ "unsafe" // required for go:linkname.
+)
+
+//go:linkname beforeFork syscall.runtime_BeforeFork
+func beforeFork()
+
+//go:linkname afterFork syscall.runtime_AfterFork
+func afterFork()
+
+//go:linkname afterForkInChild syscall.runtime_AfterForkInChild
+func afterForkInChild()
diff --git a/pkg/sentry/platform/ring0/BUILD b/pkg/sentry/platform/ring0/BUILD
new file mode 100644
index 000000000..2df232a64
--- /dev/null
+++ b/pkg/sentry/platform/ring0/BUILD
@@ -0,0 +1,52 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_generics:defs.bzl", "go_template", "go_template_instance")
+
+go_template(
+    name = "defs",
+    srcs = [
+        "defs.go",
+        "defs_amd64.go",
+        "offsets_amd64.go",
+        "x86.go",
+    ],
+    visibility = [":__subpackages__"],
+)
+
+go_template_instance(
+    name = "defs_impl",
+    out = "defs_impl.go",
+    package = "ring0",
+    template = ":defs",
+)
+
+genrule(
+    name = "entry_impl_amd64",
+    srcs = ["entry_amd64.s"],
+    outs = ["entry_impl_amd64.s"],
+    cmd = "(echo -e '// build +amd64\\n' && $(location //pkg/sentry/platform/ring0/gen_offsets) && cat $(SRCS)) > $@",
+    tools = ["//pkg/sentry/platform/ring0/gen_offsets"],
+)
+
+go_library(
+    name = "ring0",
+    srcs = [
+        "defs_impl.go",
+        "entry_amd64.go",
+        "entry_impl_amd64.s",
+        "kernel.go",
+        "kernel_amd64.go",
+        "kernel_unsafe.go",
+        "lib_amd64.go",
+        "lib_amd64.s",
+        "ring0.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/cpuid",
+        "//pkg/sentry/platform/ring0/pagetables",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/platform/ring0/defs.go b/pkg/sentry/platform/ring0/defs.go
new file mode 100644
index 000000000..9d947b73d
--- /dev/null
+++ b/pkg/sentry/platform/ring0/defs.go
@@ -0,0 +1,93 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ring0
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+var (
+	// UserspaceSize is the total size of userspace.
+	UserspaceSize = uintptr(1) << (VirtualAddressBits() - 1)
+
+	// MaximumUserAddress is the largest possible user address.
+	MaximumUserAddress = (UserspaceSize - 1) & ^uintptr(usermem.PageSize-1)
+
+	// KernelStartAddress is the starting kernel address.
+	KernelStartAddress = ^uintptr(0) - (UserspaceSize - 1)
+)
+
+// Kernel is a global kernel object.
+//
+// This contains global state, shared by multiple CPUs.
+type Kernel struct {
+	KernelArchState
+}
+
+// CPU is the per-CPU struct.
+type CPU struct {
+	// self is a self reference.
+	//
+	// This is always guaranteed to be at offset zero.
+	self *CPU
+
+	// kernel is reference to the kernel that this CPU was initialized
+	// with. This reference is kept for garbage collection purposes: CPU
+	// registers may refer to objects within the Kernel object that cannot
+	// be safely freed.
+	kernel *Kernel
+
+	// CPUArchState is architecture-specific state.
+	CPUArchState
+
+	// registers is a set of registers; these may be used on kernel system
+	// calls and exceptions via the Registers function.
+	registers syscall.PtraceRegs
+
+	// KernelException handles an exception during kernel execution.
+	//
+	// Return from this call will restore registers and return to the kernel: the
+	// registers must be modified directly.
+	//
+	// If this function is not provided, a kernel exception results in halt.
+	//
+	// This must be go:nosplit, as this will be on the interrupt stack.
+	// Closures are permitted, as the pointer to the closure frame is not
+	// passed on the stack.
+	KernelException func(Vector)
+
+	// KernelSyscall is called for kernel system calls.
+	//
+	// Return from this call will restore registers and return to the kernel: the
+	// registers must be modified directly.
+	//
+	// If this function is not provided, a kernel exception results in halt.
+	//
+	// This must be go:nosplit, as this will be on the interrupt stack.
+	// Closures are permitted, as the pointer to the closure frame is not
+	// passed on the stack.
+	KernelSyscall func()
+}
+
+// Registers returns a modifiable-copy of the kernel registers.
+//
+// This is explicitly safe to call during KernelException and KernelSyscall.
+//
+//go:nosplit
+func (c *CPU) Registers() *syscall.PtraceRegs {
+	return &c.registers
+}
diff --git a/pkg/sentry/platform/ring0/defs_amd64.go b/pkg/sentry/platform/ring0/defs_amd64.go
new file mode 100644
index 000000000..bb3420125
--- /dev/null
+++ b/pkg/sentry/platform/ring0/defs_amd64.go
@@ -0,0 +1,113 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ring0
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+)
+
+// Segment indices and Selectors.
+const (
+	// Index into GDT array.
+	_          = iota // Null descriptor first.
+	_                 // Reserved (Linux is kernel 32).
+	segKcode          // Kernel code (64-bit).
+	segKdata          // Kernel data.
+	segUcode32        // User code (32-bit).
+	segUdata          // User data.
+	segUcode64        // User code (64-bit).
+	segTss            // Task segment descriptor.
+	segTssHi          // Upper bits for TSS.
+	segLast           // Last segment (terminal, not included).
+)
+
+// Selectors.
+const (
+	Kcode   Selector = segKcode << 3
+	Kdata   Selector = segKdata << 3
+	Ucode32 Selector = (segUcode32 << 3) | 3
+	Udata   Selector = (segUdata << 3) | 3
+	Ucode64 Selector = (segUcode64 << 3) | 3
+	Tss     Selector = segTss << 3
+)
+
+// Standard segments.
+var (
+	UserCodeSegment32 SegmentDescriptor
+	UserDataSegment   SegmentDescriptor
+	UserCodeSegment64 SegmentDescriptor
+	KernelCodeSegment SegmentDescriptor
+	KernelDataSegment SegmentDescriptor
+)
+
+// KernelOpts has initialization options for the kernel.
+type KernelOpts struct {
+	// PageTables are the kernel pagetables; this must be provided.
+	PageTables *pagetables.PageTables
+}
+
+// KernelArchState contains architecture-specific state.
+type KernelArchState struct {
+	KernelOpts
+
+	// globalIDT is our set of interrupt gates.
+	globalIDT idt64
+}
+
+// CPUArchState contains CPU-specific arch state.
+type CPUArchState struct {
+	// stack is the stack used for interrupts on this CPU.
+	stack [256]byte
+
+	// errorCode is the error code from the last exception.
+	errorCode uintptr
+
+	// errorType indicates the type of error code here, it is always set
+	// along with the errorCode value above.
+	//
+	// It will either by 1, which indicates a user error, or 0 indicating a
+	// kernel error. If the error code below returns false (kernel error),
+	// then it cannot provide relevant information about the last
+	// exception.
+	errorType uintptr
+
+	// gdt is the CPU's descriptor table.
+	gdt descriptorTable
+
+	// tss is the CPU's task state.
+	tss TaskState64
+}
+
+// ErrorCode returns the last error code.
+//
+// The returned boolean indicates whether the error code corresponds to the
+// last user error or not. If it does not, then fault information must be
+// ignored. This is generally the result of a kernel fault while servicing a
+// user fault.
+//
+//go:nosplit
+func (c *CPU) ErrorCode() (value uintptr, user bool) {
+	return c.errorCode, c.errorType != 0
+}
+
+func init() {
+	KernelCodeSegment.setCode64(0, 0, 0)
+	KernelDataSegment.setData(0, 0xffffffff, 0)
+	UserCodeSegment32.setCode64(0, 0, 3)
+	UserDataSegment.setData(0, 0xffffffff, 3)
+	UserCodeSegment64.setCode64(0, 0, 3)
+}
diff --git a/pkg/sentry/platform/ring0/entry_amd64.go b/pkg/sentry/platform/ring0/entry_amd64.go
new file mode 100644
index 000000000..a3e992e0d
--- /dev/null
+++ b/pkg/sentry/platform/ring0/entry_amd64.go
@@ -0,0 +1,128 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ring0
+
+import (
+	"syscall"
+)
+
+// This is an assembly function.
+//
+// The sysenter function is invoked in two situations:
+//
+//  (1) The guest kernel has executed a system call.
+//  (2) The guest application has executed a system call.
+//
+// The interrupt flag is examined to determine whether the system call was
+// executed from kernel mode or not and the appropriate stub is called.
+func sysenter()
+
+// swapgs swaps the current GS value.
+//
+// This must be called prior to sysret/iret.
+func swapgs()
+
+// sysret returns to userspace from a system call.
+//
+// The return code is the vector that interrupted execution.
+//
+// See stubs.go for a note regarding the frame size of this function.
+func sysret(*CPU, *syscall.PtraceRegs) Vector
+
+// "iret is the cadillac of CPL switching."
+//
+//				-- Neel Natu
+//
+// iret is nearly identical to sysret, except an iret is used to fully restore
+// all user state. This must be called in cases where all registers need to be
+// restored.
+func iret(*CPU, *syscall.PtraceRegs) Vector
+
+// exception is the generic exception entry.
+//
+// This is called by the individual stub definitions.
+func exception()
+
+// resume is a stub that restores the CPU kernel registers.
+//
+// This is used when processing kernel exceptions and syscalls.
+func resume()
+
+// Start is the CPU entrypoint.
+//
+// The following start conditions must be satisfied:
+//
+//  * AX should contain the CPU pointer.
+//  * c.GDT() should be loaded as the GDT.
+//  * c.IDT() should be loaded as the IDT.
+//  * c.CR0() should be the current CR0 value.
+//  * c.CR3() should be set to the kernel PageTables.
+//  * c.CR4() should be the current CR4 value.
+//  * c.EFER() should be the current EFER value.
+//
+// The CPU state will be set to c.Registers().
+func Start()
+
+// Exception stubs.
+func divideByZero()
+func debug()
+func nmi()
+func breakpoint()
+func overflow()
+func boundRangeExceeded()
+func invalidOpcode()
+func deviceNotAvailable()
+func doubleFault()
+func coprocessorSegmentOverrun()
+func invalidTSS()
+func segmentNotPresent()
+func stackSegmentFault()
+func generalProtectionFault()
+func pageFault()
+func x87FloatingPointException()
+func alignmentCheck()
+func machineCheck()
+func simdFloatingPointException()
+func virtualizationException()
+func securityException()
+func syscallInt80()
+
+// Exception handler index.
+var handlers = map[Vector]func(){
+	DivideByZero:               divideByZero,
+	Debug:                      debug,
+	NMI:                        nmi,
+	Breakpoint:                 breakpoint,
+	Overflow:                   overflow,
+	BoundRangeExceeded:         boundRangeExceeded,
+	InvalidOpcode:              invalidOpcode,
+	DeviceNotAvailable:         deviceNotAvailable,
+	DoubleFault:                doubleFault,
+	CoprocessorSegmentOverrun:  coprocessorSegmentOverrun,
+	InvalidTSS:                 invalidTSS,
+	SegmentNotPresent:          segmentNotPresent,
+	StackSegmentFault:          stackSegmentFault,
+	GeneralProtectionFault:     generalProtectionFault,
+	PageFault:                  pageFault,
+	X87FloatingPointException:  x87FloatingPointException,
+	AlignmentCheck:             alignmentCheck,
+	MachineCheck:               machineCheck,
+	SIMDFloatingPointException: simdFloatingPointException,
+	VirtualizationException:    virtualizationException,
+	SecurityException:          securityException,
+	SyscallInt80:               syscallInt80,
+}
diff --git a/pkg/sentry/platform/ring0/entry_amd64.s b/pkg/sentry/platform/ring0/entry_amd64.s
new file mode 100644
index 000000000..e8638133b
--- /dev/null
+++ b/pkg/sentry/platform/ring0/entry_amd64.s
@@ -0,0 +1,334 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "funcdata.h"
+#include "textflag.h"
+
+// NB: Offsets are programatically generated (see BUILD).
+//
+// This file is concatenated with the definitions.
+
+// Saves a register set.
+//
+// This is a macro because it may need to executed in contents where a stack is
+// not available for calls.
+//
+// The following registers are not saved: AX, SP, IP, FLAGS, all segments.
+#define REGISTERS_SAVE(reg, offset) \
+  MOVQ R15, offset+PTRACE_R15(reg); \
+  MOVQ R14, offset+PTRACE_R14(reg); \
+  MOVQ R13, offset+PTRACE_R13(reg); \
+  MOVQ R12, offset+PTRACE_R12(reg); \
+  MOVQ BP,  offset+PTRACE_RBP(reg); \
+  MOVQ BX,  offset+PTRACE_RBX(reg); \
+  MOVQ CX,  offset+PTRACE_RCX(reg); \
+  MOVQ DX,  offset+PTRACE_RDX(reg); \
+  MOVQ R11, offset+PTRACE_R11(reg); \
+  MOVQ R10, offset+PTRACE_R10(reg); \
+  MOVQ R9,  offset+PTRACE_R9(reg); \
+  MOVQ R8,  offset+PTRACE_R8(reg); \
+  MOVQ SI,  offset+PTRACE_RSI(reg); \
+  MOVQ DI,  offset+PTRACE_RDI(reg);
+
+// Loads a register set.
+//
+// This is a macro because it may need to executed in contents where a stack is
+// not available for calls.
+//
+// The following registers are not loaded: AX, SP, IP, FLAGS, all segments.
+#define REGISTERS_LOAD(reg, offset) \
+  MOVQ offset+PTRACE_R15(reg), R15; \
+  MOVQ offset+PTRACE_R14(reg), R14; \
+  MOVQ offset+PTRACE_R13(reg), R13; \
+  MOVQ offset+PTRACE_R12(reg), R12; \
+  MOVQ offset+PTRACE_RBP(reg), BP; \
+  MOVQ offset+PTRACE_RBX(reg), BX; \
+  MOVQ offset+PTRACE_RCX(reg), CX; \
+  MOVQ offset+PTRACE_RDX(reg), DX; \
+  MOVQ offset+PTRACE_R11(reg), R11; \
+  MOVQ offset+PTRACE_R10(reg), R10; \
+  MOVQ offset+PTRACE_R9(reg),  R9; \
+  MOVQ offset+PTRACE_R8(reg),  R8; \
+  MOVQ offset+PTRACE_RSI(reg), SI; \
+  MOVQ offset+PTRACE_RDI(reg), DI;
+
+// SWAP_GS swaps the kernel GS (CPU).
+#define SWAP_GS() \
+	BYTE $0x0F; BYTE $0x01; BYTE $0xf8;
+
+// IRET returns from an interrupt frame.
+#define IRET() \
+	BYTE $0x48; BYTE $0xcf;
+
+// SYSRET64 executes the sysret instruction.
+#define SYSRET64() \
+	BYTE $0x48; BYTE $0x0f; BYTE $0x07;
+
+// LOAD_KERNEL_ADDRESS loads a kernel address.
+#define LOAD_KERNEL_ADDRESS(from, to) \
+	MOVQ from, to; \
+	ORQ ·KernelStartAddress(SB), to;
+
+// LOAD_KERNEL_STACK loads the kernel stack.
+#define LOAD_KERNEL_STACK(from) \
+	LOAD_KERNEL_ADDRESS(CPU_SELF(from), SP); \
+	LEAQ CPU_STACK_TOP(SP), SP;
+
+// See kernel.go.
+TEXT ·Halt(SB),NOSPLIT,$0
+	HLT
+	RET
+
+// See kernel.go.
+TEXT ·Current(SB),NOSPLIT,$0-8
+	MOVQ CPU_SELF(GS), AX
+	MOVQ AX, ret+0(FP)
+	RET
+
+// See entry_amd64.go.
+TEXT ·swapgs(SB),NOSPLIT,$0
+	SWAP_GS()
+	RET
+
+// See entry_amd64.go.
+TEXT ·sysret(SB),NOSPLIT,$0-24
+	// Save original state.
+	LOAD_KERNEL_ADDRESS(cpu+0(FP), BX)
+	LOAD_KERNEL_ADDRESS(regs+8(FP), AX)
+	MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX)
+	MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX)
+	MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX)
+
+	// Restore user register state.
+	REGISTERS_LOAD(AX, 0)
+	MOVQ PTRACE_RIP(AX), CX    // Needed for SYSRET.
+	MOVQ PTRACE_FLAGS(AX), R11 // Needed for SYSRET.
+	MOVQ PTRACE_RSP(AX), SP    // Restore the stack directly.
+	MOVQ PTRACE_RAX(AX), AX    // Restore AX (scratch).
+	SYSRET64()
+
+// See entry_amd64.go.
+TEXT ·iret(SB),NOSPLIT,$0-24
+	// Save original state.
+	LOAD_KERNEL_ADDRESS(cpu+0(FP), BX)
+	LOAD_KERNEL_ADDRESS(regs+8(FP), AX)
+	MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX)
+	MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX)
+	MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX)
+
+	// Build an IRET frame & restore state.
+	LOAD_KERNEL_STACK(BX)
+	MOVQ PTRACE_SS(AX), BX;    PUSHQ BX
+	MOVQ PTRACE_RSP(AX), CX;   PUSHQ CX
+	MOVQ PTRACE_FLAGS(AX), DX; PUSHQ DX
+	MOVQ PTRACE_CS(AX), DI;    PUSHQ DI
+	MOVQ PTRACE_RIP(AX), SI;   PUSHQ SI
+	REGISTERS_LOAD(AX, 0)   // Restore most registers.
+	MOVQ PTRACE_RAX(AX), AX // Restore AX (scratch).
+	IRET()
+
+// See entry_amd64.go.
+TEXT ·resume(SB),NOSPLIT,$0
+	// See iret, above.
+	MOVQ CPU_REGISTERS+PTRACE_SS(GS), BX;    PUSHQ BX
+	MOVQ CPU_REGISTERS+PTRACE_RSP(GS), CX;   PUSHQ CX
+	MOVQ CPU_REGISTERS+PTRACE_FLAGS(GS), DX; PUSHQ DX
+	MOVQ CPU_REGISTERS+PTRACE_CS(GS), DI;    PUSHQ DI
+	MOVQ CPU_REGISTERS+PTRACE_RIP(GS), SI;   PUSHQ SI
+	REGISTERS_LOAD(GS, CPU_REGISTERS)
+	MOVQ CPU_REGISTERS+PTRACE_RAX(GS), AX
+	IRET()
+
+// See entry_amd64.go.
+TEXT ·Start(SB),NOSPLIT,$0
+	LOAD_KERNEL_STACK(AX) // Set the stack.
+	PUSHQ $0x0            // Previous frame pointer.
+	MOVQ SP, BP           // Set frame pointer.
+	PUSHQ AX              // First argument (CPU).
+	CALL ·start(SB)       // Call Go hook.
+	JMP ·resume(SB)       // Restore to registers.
+
+// See entry_amd64.go.
+TEXT ·sysenter(SB),NOSPLIT,$0
+	// Interrupts are always disabled while we're executing in kernel mode
+	// and always enabled while executing in user mode. Therefore, we can
+	// reliably look at the flags in R11 to determine where this syscall
+	// was from.
+	TESTL $_RFLAGS_IF, R11
+	JZ kernel
+
+user:
+	SWAP_GS()
+	XCHGQ CPU_REGISTERS+PTRACE_RSP(GS), SP // Swap stacks.
+	XCHGQ CPU_REGISTERS+PTRACE_RAX(GS), AX // Swap for AX (regs).
+	REGISTERS_SAVE(AX, 0)                  // Save all except IP, FLAGS, SP, AX.
+	MOVQ CPU_REGISTERS+PTRACE_RAX(GS), BX  // Load saved AX value.
+	MOVQ BX,  PTRACE_RAX(AX)               // Save everything else.
+	MOVQ BX,  PTRACE_ORIGRAX(AX)
+	MOVQ CX,  PTRACE_RIP(AX)
+	MOVQ R11, PTRACE_FLAGS(AX)
+	MOVQ CPU_REGISTERS+PTRACE_RSP(GS), BX; MOVQ BX, PTRACE_RSP(AX)
+	MOVQ $0, CPU_ERROR_CODE(GS) // Clear error code.
+	MOVQ $1, CPU_ERROR_TYPE(GS) // Set error type to user.
+
+	// Return to the kernel, where the frame is:
+	//
+	//	vector      (sp+24)
+	// 	regs        (sp+16)
+	// 	cpu         (sp+8)
+	// 	vcpu.Switch (sp+0)
+	//
+	MOVQ CPU_REGISTERS+PTRACE_RBP(GS), BP // Original base pointer.
+	MOVQ $Syscall, 24(SP)                 // Output vector.
+	RET
+
+kernel:
+	// We can't restore the original stack, but we can access the registers
+	// in the CPU state directly. No need for temporary juggling.
+	MOVQ AX,  CPU_REGISTERS+PTRACE_ORIGRAX(GS)
+	MOVQ AX,  CPU_REGISTERS+PTRACE_RAX(GS)
+	REGISTERS_SAVE(GS, CPU_REGISTERS)
+	MOVQ CX,  CPU_REGISTERS+PTRACE_RIP(GS)
+	MOVQ R11, CPU_REGISTERS+PTRACE_FLAGS(GS)
+	MOVQ SP,  CPU_REGISTERS+PTRACE_RSP(GS)
+	MOVQ $0, CPU_ERROR_CODE(GS) // Clear error code.
+	MOVQ $0, CPU_ERROR_TYPE(GS) // Set error type to kernel.
+
+	// Load the function stored in KernelSyscall.
+	//
+	// Note that this function needs to be executed on the stack in case
+	// the runtime decides to make use of the redzone (grumble). This also
+	// protects against any functions that might not be go:nosplit, since
+	// this will cause a failure immediately.
+	LOAD_KERNEL_STACK(GS)
+	MOVQ CPU_KERNEL_SYSCALL(GS), DX // Function data.
+	MOVQ 0(DX), AX                  // Function pointer.
+	PUSHQ BP                        // Push the frame pointer.
+	MOVQ SP, BP                     // Set frame pointer value.
+	CALL *AX                        // Call the function.
+	POPQ BP                         // Restore the frame pointer.
+	JMP ·resume(SB)
+
+// exception is a generic exception handler.
+//
+// There are two cases handled:
+//
+// 1) An exception in kernel mode: this results in saving the state at the time
+// of the exception and calling the defined hook.
+//
+// 2) An exception in guest mode: the original kernel frame is restored, and
+// the vector & error codes are pushed as return values.
+//
+// See below for the stubs that call exception.
+TEXT ·exception(SB),NOSPLIT,$0
+	// Determine whether the exception occurred in kernel mode or user
+	// mode, based on the flags. We expect the following stack:
+	//
+	//	SS          (sp+48)
+	//	SP          (sp+40)
+	//	FLAGS       (sp+32)
+	//	CS          (sp+24)
+	//	IP          (sp+16)
+	//	ERROR_CODE  (sp+8)
+	//	VECTOR      (sp+0)
+	//
+	TESTL $_RFLAGS_IF, 32(SP)
+	JZ kernel
+
+user:
+	SWAP_GS()
+	XCHGQ CPU_REGISTERS+PTRACE_RAX(GS), AX // Swap for AX (regs).
+	REGISTERS_SAVE(AX, 0)                  // Save all except IP, FLAGS, SP, AX.
+	MOVQ CPU_REGISTERS+PTRACE_RAX(GS), BX  // Load saved AX value.
+	MOVQ BX, PTRACE_RAX(AX)                // Save everything else.
+	MOVQ BX, PTRACE_ORIGRAX(AX)
+	MOVQ 16(SP), BX; MOVQ BX, PTRACE_RIP(AX)
+	MOVQ 24(SP), CX; MOVQ CX, PTRACE_CS(AX)
+	MOVQ 32(SP), DX; MOVQ DX, PTRACE_FLAGS(AX)
+	MOVQ 40(SP), DI; MOVQ DI, PTRACE_RSP(AX)
+	MOVQ 48(SP), SI; MOVQ SI, PTRACE_SS(AX)
+
+	// Copy out and return.
+	MOVQ 0(SP), BX                        // Load vector.
+	MOVQ 8(SP), CX                        // Load error code.
+	MOVQ CPU_REGISTERS+PTRACE_RSP(GS), SP // Original stack (kernel version).
+	MOVQ CPU_REGISTERS+PTRACE_RBP(GS), BP // Original base pointer.
+	MOVQ CX, CPU_ERROR_CODE(GS)           // Set error code.
+	MOVQ $1, CPU_ERROR_TYPE(GS)           // Set error type to user.
+	MOVQ BX, 24(SP)                       // Output vector.
+	RET
+
+kernel:
+	// As per above, we can save directly.
+	MOVQ AX, CPU_REGISTERS+PTRACE_RAX(GS)
+	MOVQ AX, CPU_REGISTERS+PTRACE_ORIGRAX(GS)
+	REGISTERS_SAVE(GS, CPU_REGISTERS)
+	MOVQ 16(SP), AX; MOVQ AX, CPU_REGISTERS+PTRACE_RIP(GS)
+	MOVQ 32(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_FLAGS(GS)
+	MOVQ 40(SP), CX; MOVQ CX, CPU_REGISTERS+PTRACE_RSP(GS)
+
+	// Set the error code and adjust the stack.
+	MOVQ 8(SP), AX              // Load the error code.
+	MOVQ AX, CPU_ERROR_CODE(GS) // Copy out to the CPU.
+	MOVQ $0, CPU_ERROR_TYPE(GS) // Set error type to kernel.
+	MOVQ 0(SP), BX              // BX contains the vector.
+	ADDQ $48, SP                // Drop the exception frame.
+
+	// Load the function stored in KernelException.
+	//
+	// See note above re: the kernel stack.
+	LOAD_KERNEL_STACK(GS)
+	MOVQ CPU_KERNEL_EXCEPTION(GS), DX // Function data.
+	MOVQ 0(DX), AX                    // Function pointer.
+	PUSHQ BP                          // Push the frame pointer.
+	MOVQ SP, BP                       // Set frame pointer value.
+	PUSHQ BX                          // First argument (vector).
+	CALL *AX                          // Call the function.
+	POPQ BX                           // Discard the argument.
+	POPQ BP                           // Restore the frame pointer.
+	JMP ·resume(SB)
+
+#define EXCEPTION_WITH_ERROR(value, symbol) \
+TEXT symbol,NOSPLIT,$0; \
+	PUSHQ $value; \
+	JMP ·exception(SB);
+
+#define EXCEPTION_WITHOUT_ERROR(value, symbol) \
+TEXT symbol,NOSPLIT,$0; \
+	PUSHQ $0x0; \
+	PUSHQ $value; \
+	JMP ·exception(SB);
+
+EXCEPTION_WITHOUT_ERROR(DivideByZero, ·divideByZero(SB))
+EXCEPTION_WITHOUT_ERROR(Debug, ·debug(SB))
+EXCEPTION_WITHOUT_ERROR(NMI, ·nmi(SB))
+EXCEPTION_WITHOUT_ERROR(Breakpoint, ·breakpoint(SB))
+EXCEPTION_WITHOUT_ERROR(Overflow, ·overflow(SB))
+EXCEPTION_WITHOUT_ERROR(BoundRangeExceeded, ·boundRangeExceeded(SB))
+EXCEPTION_WITHOUT_ERROR(InvalidOpcode, ·invalidOpcode(SB))
+EXCEPTION_WITHOUT_ERROR(DeviceNotAvailable, ·deviceNotAvailable(SB))
+EXCEPTION_WITH_ERROR(DoubleFault, ·doubleFault(SB))
+EXCEPTION_WITHOUT_ERROR(CoprocessorSegmentOverrun, ·coprocessorSegmentOverrun(SB))
+EXCEPTION_WITH_ERROR(InvalidTSS, ·invalidTSS(SB))
+EXCEPTION_WITH_ERROR(SegmentNotPresent, ·segmentNotPresent(SB))
+EXCEPTION_WITH_ERROR(StackSegmentFault, ·stackSegmentFault(SB))
+EXCEPTION_WITH_ERROR(GeneralProtectionFault, ·generalProtectionFault(SB))
+EXCEPTION_WITH_ERROR(PageFault, ·pageFault(SB))
+EXCEPTION_WITHOUT_ERROR(X87FloatingPointException, ·x87FloatingPointException(SB))
+EXCEPTION_WITH_ERROR(AlignmentCheck, ·alignmentCheck(SB))
+EXCEPTION_WITHOUT_ERROR(MachineCheck, ·machineCheck(SB))
+EXCEPTION_WITHOUT_ERROR(SIMDFloatingPointException, ·simdFloatingPointException(SB))
+EXCEPTION_WITHOUT_ERROR(VirtualizationException, ·virtualizationException(SB))
+EXCEPTION_WITH_ERROR(SecurityException, ·securityException(SB))
+EXCEPTION_WITHOUT_ERROR(SyscallInt80, ·syscallInt80(SB))
diff --git a/pkg/sentry/platform/ring0/gen_offsets/BUILD b/pkg/sentry/platform/ring0/gen_offsets/BUILD
new file mode 100644
index 000000000..3bce56985
--- /dev/null
+++ b/pkg/sentry/platform/ring0/gen_offsets/BUILD
@@ -0,0 +1,25 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_binary")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+go_template_instance(
+    name = "defs_impl",
+    out = "defs_impl.go",
+    package = "main",
+    template = "//pkg/sentry/platform/ring0:defs",
+)
+
+go_binary(
+    name = "gen_offsets",
+    srcs = [
+        "defs_impl.go",
+        "main.go",
+    ],
+    visibility = ["//pkg/sentry/platform/ring0:__pkg__"],
+    deps = [
+        "//pkg/cpuid",
+        "//pkg/sentry/platform/ring0/pagetables",
+        "//pkg/sentry/usermem",
+    ],
+)
diff --git a/pkg/sentry/platform/ring0/gen_offsets/main.go b/pkg/sentry/platform/ring0/gen_offsets/main.go
new file mode 100644
index 000000000..ffa7eaf77
--- /dev/null
+++ b/pkg/sentry/platform/ring0/gen_offsets/main.go
@@ -0,0 +1,24 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Binary gen_offsets is a helper for generating offset headers.
+package main
+
+import (
+	"os"
+)
+
+func main() {
+	Emit(os.Stdout)
+}
diff --git a/pkg/sentry/platform/ring0/kernel.go b/pkg/sentry/platform/ring0/kernel.go
new file mode 100644
index 000000000..b0471ab9a
--- /dev/null
+++ b/pkg/sentry/platform/ring0/kernel.go
@@ -0,0 +1,71 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ring0
+
+// New creates a new kernel.
+//
+// N.B. that constraints on KernelOpts must be satisfied.
+//
+// Init must have been called.
+func New(opts KernelOpts) *Kernel {
+	k := new(Kernel)
+	k.init(opts)
+	return k
+}
+
+// NewCPU creates a new CPU associated with this Kernel.
+//
+// Note that execution of the new CPU must begin at Start, with constraints as
+// documented. Initialization is not completed by this method alone.
+//
+// See also Init.
+func (k *Kernel) NewCPU() *CPU {
+	c := new(CPU)
+	c.Init(k)
+	return c
+}
+
+// Halt halts execution.
+func Halt()
+
+// Current returns the current CPU.
+//
+// Its use is only legal in the KernelSyscall and KernelException contexts,
+// which must all be guarded go:nosplit.
+func Current() *CPU
+
+// defaultSyscall is the default syscall hook.
+//
+//go:nosplit
+func defaultSyscall() { Halt() }
+
+// defaultException is the default exception hook.
+//
+//go:nosplit
+func defaultException(Vector) { Halt() }
+
+// Init allows the initialization of a CPU from a kernel without allocation.
+// The same constraints as NewCPU apply.
+//
+// Init allows embedding in other objects.
+func (c *CPU) Init(k *Kernel) {
+	c.self = c   // Set self reference.
+	c.kernel = k // Set kernel reference.
+	c.init()     // Perform architectural init.
+
+	// Defaults.
+	c.KernelSyscall = defaultSyscall
+	c.KernelException = defaultException
+}
diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go
new file mode 100644
index 000000000..c82613a9c
--- /dev/null
+++ b/pkg/sentry/platform/ring0/kernel_amd64.go
@@ -0,0 +1,280 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ring0
+
+import (
+	"encoding/binary"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables"
+)
+
+const (
+	// KernelFlagsSet should always be set in the kernel.
+	KernelFlagsSet = _RFLAGS_RESERVED
+
+	// UserFlagsSet are always set in userspace.
+	UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF
+
+	// KernelFlagsClear should always be clear in the kernel.
+	KernelFlagsClear = _RFLAGS_IF | _RFLAGS_NT | _RFLAGS_IOPL
+
+	// UserFlagsClear are always cleared in userspace.
+	UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL
+)
+
+// init initializes architecture-specific state.
+func (k *Kernel) init(opts KernelOpts) {
+	// Save the root page tables.
+	k.PageTables = opts.PageTables
+
+	// Setup the IDT, which is uniform.
+	for v, handler := range handlers {
+		// Note that we set all traps to use the interrupt stack, this
+		// is defined below when setting up the TSS.
+		k.globalIDT[v].setInterrupt(Kcode, uint64(kernelFunc(handler)), 0 /* dpl */, 1 /* ist */)
+	}
+}
+
+// init initializes architecture-specific state.
+func (c *CPU) init() {
+	// Null segment.
+	c.gdt[0].setNull()
+
+	// Kernel & user segments.
+	c.gdt[segKcode] = KernelCodeSegment
+	c.gdt[segKdata] = KernelDataSegment
+	c.gdt[segUcode32] = UserCodeSegment32
+	c.gdt[segUdata] = UserDataSegment
+	c.gdt[segUcode64] = UserCodeSegment64
+
+	// The task segment, this spans two entries.
+	tssBase, tssLimit, _ := c.TSS()
+	c.gdt[segTss].set(
+		uint32(tssBase),
+		uint32(tssLimit),
+		0, // Privilege level zero.
+		SegmentDescriptorPresent|
+			SegmentDescriptorAccess|
+			SegmentDescriptorWrite|
+			SegmentDescriptorExecute)
+	c.gdt[segTssHi].setHi(uint32((tssBase) >> 32))
+
+	// Set the kernel stack pointer in the TSS (virtual address).
+	stackAddr := c.StackTop()
+	c.tss.rsp0Lo = uint32(stackAddr)
+	c.tss.rsp0Hi = uint32(stackAddr >> 32)
+	c.tss.ist1Lo = uint32(stackAddr)
+	c.tss.ist1Hi = uint32(stackAddr >> 32)
+
+	// Permanently set the kernel segments.
+	c.registers.Cs = uint64(Kcode)
+	c.registers.Ds = uint64(Kdata)
+	c.registers.Es = uint64(Kdata)
+	c.registers.Ss = uint64(Kdata)
+	c.registers.Fs = uint64(Kdata)
+	c.registers.Gs = uint64(Kdata)
+}
+
+// StackTop returns the kernel's stack address.
+//
+//go:nosplit
+func (c *CPU) StackTop() uint64 {
+	return uint64(kernelAddr(&c.stack[0])) + uint64(len(c.stack))
+}
+
+// IDT returns the CPU's IDT base and limit.
+//
+//go:nosplit
+func (c *CPU) IDT() (uint64, uint16) {
+	return uint64(kernelAddr(&c.kernel.globalIDT[0])), uint16(binary.Size(&c.kernel.globalIDT) - 1)
+}
+
+// GDT returns the CPU's GDT base and limit.
+//
+//go:nosplit
+func (c *CPU) GDT() (uint64, uint16) {
+	return uint64(kernelAddr(&c.gdt[0])), uint16(8*segLast - 1)
+}
+
+// TSS returns the CPU's TSS base, limit and value.
+//
+//go:nosplit
+func (c *CPU) TSS() (uint64, uint16, *SegmentDescriptor) {
+	return uint64(kernelAddr(&c.tss)), uint16(binary.Size(&c.tss) - 1), &c.gdt[segTss]
+}
+
+// CR0 returns the CPU's CR0 value.
+//
+//go:nosplit
+func (c *CPU) CR0() uint64 {
+	return _CR0_PE | _CR0_PG | _CR0_ET
+}
+
+// CR4 returns the CPU's CR4 value.
+//
+//go:nosplit
+func (c *CPU) CR4() uint64 {
+	cr4 := uint64(_CR4_PAE | _CR4_PSE | _CR4_OSFXSR | _CR4_OSXMMEXCPT)
+	if hasPCID {
+		cr4 |= _CR4_PCIDE
+	}
+	if hasXSAVE {
+		cr4 |= _CR4_OSXSAVE
+	}
+	if hasSMEP {
+		cr4 |= _CR4_SMEP
+	}
+	if hasFSGSBASE {
+		cr4 |= _CR4_FSGSBASE
+	}
+	return cr4
+}
+
+// EFER returns the CPU's EFER value.
+//
+//go:nosplit
+func (c *CPU) EFER() uint64 {
+	return _EFER_LME | _EFER_SCE | _EFER_NX
+}
+
+// IsCanonical indicates whether addr is canonical per the amd64 spec.
+//
+//go:nosplit
+func IsCanonical(addr uint64) bool {
+	return addr <= 0x00007fffffffffff || addr > 0xffff800000000000
+}
+
+// Flags contains flags related to switch.
+type Flags uintptr
+
+const (
+	// FlagFull indicates that a full restore should be not, not a fast
+	// restore (on the syscall return path.)
+	FlagFull = 1 << iota
+
+	// FlagFlush indicates that a full TLB flush is required.
+	FlagFlush
+)
+
+// SwitchToUser performs either a sysret or an iret.
+//
+// The return value is the vector that interrupted execution.
+//
+// This function will not split the stack. Callers will probably want to call
+// runtime.entersyscall (and pair with a call to runtime.exitsyscall) prior to
+// calling this function.
+//
+// When this is done, this region is quite sensitive to things like system
+// calls. After calling entersyscall, any memory used must have been allocated
+// and no function calls without go:nosplit are permitted. Any calls made here
+// are protected appropriately (e.g. IsCanonical and CR3).
+//
+// Also note that this function transitively depends on the compiler generating
+// code that uses IP-relative addressing inside of absolute addresses. That's
+// the case for amd64, but may not be the case for other architectures.
+//
+//go:nosplit
+func (c *CPU) SwitchToUser(regs *syscall.PtraceRegs, fpState *byte, pt *pagetables.PageTables, flags Flags) (vector Vector) {
+	// Check for canonical addresses.
+	if !IsCanonical(regs.Rip) || !IsCanonical(regs.Rsp) || !IsCanonical(regs.Fs_base) || !IsCanonical(regs.Gs_base) {
+		return GeneralProtectionFault
+	}
+
+	var (
+		userCR3   uint64
+		kernelCR3 uint64
+	)
+
+	// Sanitize registers.
+	if flags&FlagFlush != 0 {
+		userCR3 = pt.FlushCR3()
+	} else {
+		userCR3 = pt.CR3()
+	}
+	regs.Eflags &= ^uint64(UserFlagsClear)
+	regs.Eflags |= UserFlagsSet
+	regs.Cs = uint64(Ucode64) // Required for iret.
+	regs.Ss = uint64(Udata)   // Ditto.
+	kernelCR3 = c.kernel.PageTables.CR3()
+
+	// Perform the switch.
+	swapgs()                    // GS will be swapped on return.
+	wrfs(uintptr(regs.Fs_base)) // Set application FS.
+	wrgs(uintptr(regs.Gs_base)) // Set application GS.
+	LoadFloatingPoint(fpState)  // Copy in floating point.
+	jumpToKernel()              // Switch to upper half.
+	writeCR3(uintptr(userCR3))  // Change to user address space.
+	if flags&FlagFull != 0 {
+		vector = iret(c, regs)
+	} else {
+		vector = sysret(c, regs)
+	}
+	writeCR3(uintptr(kernelCR3))       // Return to kernel address space.
+	jumpToUser()                       // Return to lower half.
+	SaveFloatingPoint(fpState)         // Copy out floating point.
+	wrfs(uintptr(c.registers.Fs_base)) // Restore kernel FS.
+	return
+}
+
+// start is the CPU entrypoint.
+//
+// This is called from the Start asm stub (see entry_amd64.go); on return the
+// registers in c.registers will be restored (not segments).
+//
+//go:nosplit
+func start(c *CPU) {
+	// Save per-cpu & FS segment.
+	wrgs(kernelAddr(c))
+	wrfs(uintptr(c.Registers().Fs_base))
+
+	// Initialize floating point.
+	//
+	// Note that on skylake, the valid XCR0 mask reported seems to be 0xff.
+	// This breaks down as:
+	//
+	//	bit0   - x87
+	//	bit1   - SSE
+	//	bit2   - AVX
+	//	bit3-4 - MPX
+	//	bit5-7 - AVX512
+	//
+	// For some reason, enabled MPX & AVX512 on platforms that report them
+	// seems to be cause a general protection fault. (Maybe there are some
+	// virtualization issues and these aren't exported to the guest cpuid.)
+	// This needs further investigation, but we can limit the floating
+	// point operations to x87, SSE & AVX for now.
+	fninit()
+	xsetbv(0, validXCR0Mask&0x7)
+
+	// Set the syscall target.
+	wrmsr(_MSR_LSTAR, kernelFunc(sysenter))
+	wrmsr(_MSR_SYSCALL_MASK, _RFLAGS_STEP|_RFLAGS_IF|_RFLAGS_DF|_RFLAGS_IOPL|_RFLAGS_AC|_RFLAGS_NT)
+
+	// NOTE: This depends on having the 64-bit segments immediately
+	// following the 32-bit user segments. This is simply the way the
+	// sysret instruction is designed to work (it assumes they follow).
+	wrmsr(_MSR_STAR, uintptr(uint64(Kcode)<<32|uint64(Ucode32)<<48))
+	wrmsr(_MSR_CSTAR, kernelFunc(sysenter))
+}
+
+// ReadCR2 reads the current CR2 value.
+//
+//go:nosplit
+func ReadCR2() uintptr {
+	return readCR2()
+}
diff --git a/pkg/sentry/platform/ring0/kernel_unsafe.go b/pkg/sentry/platform/ring0/kernel_unsafe.go
new file mode 100644
index 000000000..cfb3ad853
--- /dev/null
+++ b/pkg/sentry/platform/ring0/kernel_unsafe.go
@@ -0,0 +1,41 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package ring0
+
+import (
+	"unsafe"
+)
+
+// eface mirrors runtime.eface.
+type eface struct {
+	typ  uintptr
+	data unsafe.Pointer
+}
+
+// kernelAddr returns the kernel virtual address for the given object.
+//
+//go:nosplit
+func kernelAddr(obj interface{}) uintptr {
+	e := (*eface)(unsafe.Pointer(&obj))
+	return KernelStartAddress | uintptr(e.data)
+}
+
+// kernelFunc returns the address of the given function.
+//
+//go:nosplit
+func kernelFunc(fn func()) uintptr {
+	fnptr := (**uintptr)(unsafe.Pointer(&fn))
+	return KernelStartAddress | **fnptr
+}
diff --git a/pkg/sentry/platform/ring0/lib_amd64.go b/pkg/sentry/platform/ring0/lib_amd64.go
new file mode 100644
index 000000000..f1ed5bfb4
--- /dev/null
+++ b/pkg/sentry/platform/ring0/lib_amd64.go
@@ -0,0 +1,128 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ring0
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+)
+
+// LoadFloatingPoint loads floating point state by the most efficient mechanism
+// available (set by Init).
+var LoadFloatingPoint func(*byte)
+
+// SaveFloatingPoint saves floating point state by the most efficient mechanism
+// available (set by Init).
+var SaveFloatingPoint func(*byte)
+
+// fxrstor uses fxrstor64 to load floating point state.
+func fxrstor(*byte)
+
+// xrstor uses xrstor to load floating point state.
+func xrstor(*byte)
+
+// fxsave uses fxsave64 to save floating point state.
+func fxsave(*byte)
+
+// xsave uses xsave to save floating point state.
+func xsave(*byte)
+
+// xsaveopt uses xsaveopt to save floating point state.
+func xsaveopt(*byte)
+
+// wrfs sets the GS address (set by init).
+var wrfs func(addr uintptr)
+
+// wrfsbase writes to the GS base address.
+func wrfsbase(addr uintptr)
+
+// wrfsmsr writes to the GS_BASE MSR.
+func wrfsmsr(addr uintptr)
+
+// wrgs sets the GS address (set by init).
+var wrgs func(addr uintptr)
+
+// wrgsbase writes to the GS base address.
+func wrgsbase(addr uintptr)
+
+// wrgsmsr writes to the GS_BASE MSR.
+func wrgsmsr(addr uintptr)
+
+// writeCR3 writes the CR3 value.
+func writeCR3(phys uintptr)
+
+// readCR2 reads the current CR2 value.
+func readCR2() uintptr
+
+// jumpToKernel jumps to the kernel version of the current RIP.
+func jumpToKernel()
+
+// jumpToUser jumps to the user version of the current RIP.
+func jumpToUser()
+
+// fninit initializes the floating point unit.
+func fninit()
+
+// xsetbv writes to an extended control register.
+func xsetbv(reg, value uintptr)
+
+// xgetbv reads an extended control register.
+func xgetbv(reg uintptr) uintptr
+
+// wrmsr reads to the given MSR.
+func wrmsr(reg, value uintptr)
+
+// rdmsr reads the given MSR.
+func rdmsr(reg uintptr) uintptr
+
+// Mostly-constants set by Init.
+var (
+	hasSMEP       bool
+	hasPCID       bool
+	hasXSAVEOPT   bool
+	hasXSAVE      bool
+	hasFSGSBASE   bool
+	validXCR0Mask uintptr
+)
+
+// Init sets function pointers based on architectural features.
+//
+// This must be called prior to using ring0.
+func Init(featureSet *cpuid.FeatureSet) {
+	hasSMEP = featureSet.HasFeature(cpuid.X86FeatureSMEP)
+	hasPCID = featureSet.HasFeature(cpuid.X86FeaturePCID)
+	hasXSAVEOPT = featureSet.UseXsaveopt()
+	hasXSAVE = featureSet.UseXsave()
+	hasFSGSBASE = featureSet.HasFeature(cpuid.X86FeatureFSGSBase)
+	validXCR0Mask = uintptr(featureSet.ValidXCR0Mask())
+	if hasXSAVEOPT {
+		SaveFloatingPoint = xsaveopt
+		LoadFloatingPoint = xrstor
+	} else if hasXSAVE {
+		SaveFloatingPoint = xsave
+		LoadFloatingPoint = xrstor
+	} else {
+		SaveFloatingPoint = fxsave
+		LoadFloatingPoint = fxrstor
+	}
+	if hasFSGSBASE {
+		wrfs = wrfsbase
+		wrgs = wrgsbase
+	} else {
+		wrfs = wrfsmsr
+		wrgs = wrgsmsr
+	}
+}
diff --git a/pkg/sentry/platform/ring0/lib_amd64.s b/pkg/sentry/platform/ring0/lib_amd64.s
new file mode 100644
index 000000000..6f143ea5a
--- /dev/null
+++ b/pkg/sentry/platform/ring0/lib_amd64.s
@@ -0,0 +1,247 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "funcdata.h"
+#include "textflag.h"
+
+// fxrstor loads floating point state.
+//
+// The code corresponds to:
+//
+//     fxrstor64 (%rbx)
+//
+TEXT ·fxrstor(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), BX
+	MOVL $0xffffffff, AX
+	MOVL $0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x0b;
+	RET
+
+// xrstor loads floating point state.
+//
+// The code corresponds to:
+//
+//     xrstor (%rdi)
+//
+TEXT ·xrstor(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), DI
+	MOVL $0xffffffff, AX
+	MOVL $0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x2f;
+	RET
+
+// fxsave saves floating point state.
+//
+// The code corresponds to:
+//
+//     fxsave64 (%rbx)
+//
+TEXT ·fxsave(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), BX
+	MOVL $0xffffffff, AX
+	MOVL $0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x03;
+	RET
+
+// xsave saves floating point state.
+//
+// The code corresponds to:
+//
+//     xsave (%rdi)
+//
+TEXT ·xsave(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), DI
+	MOVL $0xffffffff, AX
+	MOVL $0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x27;
+	RET
+
+// xsaveopt saves floating point state.
+//
+// The code corresponds to:
+//
+//     xsaveopt (%rdi)
+//
+TEXT ·xsaveopt(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), DI
+	MOVL $0xffffffff, AX
+	MOVL $0xffffffff, DX
+	BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0x37;
+	RET
+
+// wrfsbase writes to the FS base.
+//
+// The code corresponds to:
+//
+// 	wrfsbase %rax
+//
+TEXT ·wrfsbase(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), AX
+	BYTE $0xf3; BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0xd0;
+	RET
+
+// wrfsmsr writes to the FSBASE MSR.
+//
+// The code corresponds to:
+//
+// 	wrmsr (writes EDX:EAX to the MSR in ECX)
+//
+TEXT ·wrfsmsr(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), AX
+	MOVQ AX, DX
+	SHRQ $32, DX
+	MOVQ $0xc0000100, CX // MSR_FS_BASE
+	BYTE $0x0f; BYTE $0x30;
+	RET
+
+// wrgsbase writes to the GS base.
+//
+// The code corresponds to:
+//
+// 	wrgsbase %rax
+//
+TEXT ·wrgsbase(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), AX
+	BYTE $0xf3; BYTE $0x48; BYTE $0x0f; BYTE $0xae; BYTE $0xd8;
+	RET
+
+// wrgsmsr writes to the GSBASE MSR.
+//
+// See wrfsmsr.
+TEXT ·wrgsmsr(SB),NOSPLIT,$0-8
+	MOVQ addr+0(FP), AX
+	MOVQ AX, DX
+	SHRQ $32, DX
+	MOVQ $0xc0000101, CX     // MSR_GS_BASE
+	BYTE $0x0f; BYTE $0x30;  // WRMSR
+	RET
+
+// jumpToUser changes execution to the user address.
+//
+// This works by changing the return value to the user version.
+TEXT ·jumpToUser(SB),NOSPLIT,$0
+	MOVQ 0(SP), AX
+	MOVQ ·KernelStartAddress(SB), BX
+	NOTQ BX
+	ANDQ BX, SP // Switch the stack.
+	ANDQ BX, BP // Switch the frame pointer.
+	ANDQ BX, AX // Future return value.
+	MOVQ AX, 0(SP)
+	RET
+
+// jumpToKernel changes execution to the kernel address space.
+//
+// This works by changing the return value to the kernel version.
+TEXT ·jumpToKernel(SB),NOSPLIT,$0
+	MOVQ 0(SP), AX
+	MOVQ ·KernelStartAddress(SB), BX
+	ORQ BX, SP // Switch the stack.
+	ORQ BX, BP // Switch the frame pointer.
+	ORQ BX, AX // Future return value.
+	MOVQ AX, 0(SP)
+	RET
+
+// writeCR3 writes the given CR3 value.
+//
+// The code corresponds to:
+//
+// 	mov %rax, %cr3
+//
+TEXT ·writeCR3(SB),NOSPLIT,$0-8
+	MOVQ cr3+0(FP), AX
+	BYTE $0x0f; BYTE $0x22; BYTE $0xd8;
+	RET
+
+// readCR3 reads the current CR3 value.
+//
+// The code corresponds to:
+//
+// 	mov %cr3, %rax
+//
+TEXT ·readCR3(SB),NOSPLIT,$0-8
+	BYTE $0x0f; BYTE $0x20; BYTE $0xd8;
+	MOVQ AX, ret+0(FP)
+	RET
+
+// readCR2 reads the current CR2 value.
+//
+// The code corresponds to:
+//
+// 	mov %cr2, %rax
+//
+TEXT ·readCR2(SB),NOSPLIT,$0-8
+	BYTE $0x0f; BYTE $0x20; BYTE $0xd0;
+	MOVQ AX, ret+0(FP)
+	RET
+
+// fninit initializes the floating point unit.
+//
+// The code corresponds to:
+//
+// 	fninit
+TEXT ·fninit(SB),NOSPLIT,$0
+	BYTE $0xdb; BYTE $0xe3;
+	RET
+
+// xsetbv writes to an extended control register.
+//
+// The code corresponds to:
+//
+// 	xsetbv
+//
+TEXT ·xsetbv(SB),NOSPLIT,$0-16
+	MOVL reg+0(FP), CX
+	MOVL value+8(FP), AX
+	MOVL value+12(FP), DX
+	BYTE $0x0f; BYTE $0x01; BYTE $0xd1;
+	RET
+
+// xgetbv reads an extended control register.
+//
+// The code corresponds to:
+//
+// 	xgetbv
+//
+TEXT ·xgetbv(SB),NOSPLIT,$0-16
+	MOVL reg+0(FP), CX
+	BYTE $0x0f; BYTE $0x01; BYTE $0xd0;
+	MOVL AX, ret+8(FP)
+	MOVL DX, ret+12(FP)
+	RET
+
+// wrmsr writes to a control register.
+//
+// The code corresponds to:
+//
+// 	wrmsr
+//
+TEXT ·wrmsr(SB),NOSPLIT,$0-16
+	MOVL reg+0(FP), CX
+	MOVL value+8(FP), AX
+	MOVL value+12(FP), DX
+	BYTE $0x0f; BYTE $0x30;
+	RET
+
+// rdmsr reads a control register.
+//
+// The code corresponds to:
+//
+// 	rdmsr
+//
+TEXT ·rdmsr(SB),NOSPLIT,$0-16
+	MOVL reg+0(FP), CX
+	BYTE $0x0f; BYTE $0x32;
+	MOVL AX, ret+8(FP)
+	MOVL DX, ret+12(FP)
+	RET
diff --git a/pkg/sentry/platform/ring0/offsets_amd64.go b/pkg/sentry/platform/ring0/offsets_amd64.go
new file mode 100644
index 000000000..9acd442ba
--- /dev/null
+++ b/pkg/sentry/platform/ring0/offsets_amd64.go
@@ -0,0 +1,93 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package ring0
+
+import (
+	"fmt"
+	"io"
+	"reflect"
+	"syscall"
+)
+
+// Emit prints architecture-specific offsets.
+func Emit(w io.Writer) {
+	fmt.Fprintf(w, "// Automatically generated, do not edit.\n")
+
+	c := &CPU{}
+	fmt.Fprintf(w, "\n// CPU offsets.\n")
+	fmt.Fprintf(w, "#define CPU_SELF             0x%02x\n", reflect.ValueOf(&c.self).Pointer()-reflect.ValueOf(c).Pointer())
+	fmt.Fprintf(w, "#define CPU_REGISTERS        0x%02x\n", reflect.ValueOf(&c.registers).Pointer()-reflect.ValueOf(c).Pointer())
+	fmt.Fprintf(w, "#define CPU_STACK_TOP        0x%02x\n", reflect.ValueOf(&c.stack[0]).Pointer()-reflect.ValueOf(c).Pointer()+uintptr(len(c.stack)))
+	fmt.Fprintf(w, "#define CPU_ERROR_CODE       0x%02x\n", reflect.ValueOf(&c.errorCode).Pointer()-reflect.ValueOf(c).Pointer())
+	fmt.Fprintf(w, "#define CPU_ERROR_TYPE       0x%02x\n", reflect.ValueOf(&c.errorType).Pointer()-reflect.ValueOf(c).Pointer())
+	fmt.Fprintf(w, "#define CPU_KERNEL_EXCEPTION 0x%02x\n", reflect.ValueOf(&c.KernelException).Pointer()-reflect.ValueOf(c).Pointer())
+	fmt.Fprintf(w, "#define CPU_KERNEL_SYSCALL   0x%02x\n", reflect.ValueOf(&c.KernelSyscall).Pointer()-reflect.ValueOf(c).Pointer())
+
+	fmt.Fprintf(w, "\n// Bits.\n")
+	fmt.Fprintf(w, "#define _RFLAGS_IF 0x%02x\n", _RFLAGS_IF)
+
+	fmt.Fprintf(w, "\n// Vectors.\n")
+	fmt.Fprintf(w, "#define DivideByZero               0x%02x\n", DivideByZero)
+	fmt.Fprintf(w, "#define Debug                      0x%02x\n", Debug)
+	fmt.Fprintf(w, "#define NMI                        0x%02x\n", NMI)
+	fmt.Fprintf(w, "#define Breakpoint                 0x%02x\n", Breakpoint)
+	fmt.Fprintf(w, "#define Overflow                   0x%02x\n", Overflow)
+	fmt.Fprintf(w, "#define BoundRangeExceeded         0x%02x\n", BoundRangeExceeded)
+	fmt.Fprintf(w, "#define InvalidOpcode              0x%02x\n", InvalidOpcode)
+	fmt.Fprintf(w, "#define DeviceNotAvailable         0x%02x\n", DeviceNotAvailable)
+	fmt.Fprintf(w, "#define DoubleFault                0x%02x\n", DoubleFault)
+	fmt.Fprintf(w, "#define CoprocessorSegmentOverrun  0x%02x\n", CoprocessorSegmentOverrun)
+	fmt.Fprintf(w, "#define InvalidTSS                 0x%02x\n", InvalidTSS)
+	fmt.Fprintf(w, "#define SegmentNotPresent          0x%02x\n", SegmentNotPresent)
+	fmt.Fprintf(w, "#define StackSegmentFault          0x%02x\n", StackSegmentFault)
+	fmt.Fprintf(w, "#define GeneralProtectionFault     0x%02x\n", GeneralProtectionFault)
+	fmt.Fprintf(w, "#define PageFault                  0x%02x\n", PageFault)
+	fmt.Fprintf(w, "#define X87FloatingPointException  0x%02x\n", X87FloatingPointException)
+	fmt.Fprintf(w, "#define AlignmentCheck             0x%02x\n", AlignmentCheck)
+	fmt.Fprintf(w, "#define MachineCheck               0x%02x\n", MachineCheck)
+	fmt.Fprintf(w, "#define SIMDFloatingPointException 0x%02x\n", SIMDFloatingPointException)
+	fmt.Fprintf(w, "#define VirtualizationException    0x%02x\n", VirtualizationException)
+	fmt.Fprintf(w, "#define SecurityException          0x%02x\n", SecurityException)
+	fmt.Fprintf(w, "#define SyscallInt80               0x%02x\n", SyscallInt80)
+	fmt.Fprintf(w, "#define Syscall                    0x%02x\n", Syscall)
+
+	p := &syscall.PtraceRegs{}
+	fmt.Fprintf(w, "\n// Ptrace registers.\n")
+	fmt.Fprintf(w, "#define PTRACE_R15      0x%02x\n", reflect.ValueOf(&p.R15).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R14      0x%02x\n", reflect.ValueOf(&p.R14).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R13      0x%02x\n", reflect.ValueOf(&p.R13).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R12      0x%02x\n", reflect.ValueOf(&p.R12).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RBP      0x%02x\n", reflect.ValueOf(&p.Rbp).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RBX      0x%02x\n", reflect.ValueOf(&p.Rbx).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R11      0x%02x\n", reflect.ValueOf(&p.R11).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R10      0x%02x\n", reflect.ValueOf(&p.R10).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R9       0x%02x\n", reflect.ValueOf(&p.R9).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_R8       0x%02x\n", reflect.ValueOf(&p.R8).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RAX      0x%02x\n", reflect.ValueOf(&p.Rax).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RCX      0x%02x\n", reflect.ValueOf(&p.Rcx).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RDX      0x%02x\n", reflect.ValueOf(&p.Rdx).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RSI      0x%02x\n", reflect.ValueOf(&p.Rsi).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RDI      0x%02x\n", reflect.ValueOf(&p.Rdi).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_ORIGRAX  0x%02x\n", reflect.ValueOf(&p.Orig_rax).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RIP      0x%02x\n", reflect.ValueOf(&p.Rip).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_CS       0x%02x\n", reflect.ValueOf(&p.Cs).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_FLAGS    0x%02x\n", reflect.ValueOf(&p.Eflags).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_RSP      0x%02x\n", reflect.ValueOf(&p.Rsp).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_SS       0x%02x\n", reflect.ValueOf(&p.Ss).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_FS       0x%02x\n", reflect.ValueOf(&p.Fs_base).Pointer()-reflect.ValueOf(p).Pointer())
+	fmt.Fprintf(w, "#define PTRACE_GS       0x%02x\n", reflect.ValueOf(&p.Gs_base).Pointer()-reflect.ValueOf(p).Pointer())
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/BUILD b/pkg/sentry/platform/ring0/pagetables/BUILD
new file mode 100644
index 000000000..c0c481ab3
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/BUILD
@@ -0,0 +1,32 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "pagetables",
+    srcs = [
+        "pagetables.go",
+        "pagetables_amd64.go",
+        "pagetables_unsafe.go",
+        "pagetables_x86.go",
+        "pcids_x86.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/ring0/pagetables",
+    visibility = [
+        "//pkg/sentry/platform/kvm:__subpackages__",
+        "//pkg/sentry/platform/ring0:__subpackages__",
+    ],
+    deps = ["//pkg/sentry/usermem"],
+)
+
+go_test(
+    name = "pagetables_test",
+    size = "small",
+    srcs = [
+        "pagetables_test.go",
+        "pagetables_x86_test.go",
+        "pcids_x86_test.go",
+    ],
+    embed = [":pagetables"],
+    deps = ["//pkg/sentry/usermem"],
+)
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables.go b/pkg/sentry/platform/ring0/pagetables/pagetables.go
new file mode 100644
index 000000000..3cbf0bfa5
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables.go
@@ -0,0 +1,193 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package pagetables provides a generic implementation of pagetables.
+package pagetables
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// Node is a single node within a set of page tables.
+type Node struct {
+	// unalignedData has unaligned data. Unfortunately, we can't really
+	// rely on the allocator to give us what we want here. So we just throw
+	// it at the wall and use the portion that matches. Gross. This may be
+	// changed in the future to use a different allocation mechanism.
+	//
+	// Access must happen via functions found in pagetables_unsafe.go.
+	unalignedData [(2 * usermem.PageSize) - 1]byte
+
+	// physical is the translated address of these entries.
+	//
+	// This is filled in at creation time.
+	physical uintptr
+}
+
+// PageTables is a set of page tables.
+type PageTables struct {
+	mu sync.Mutex
+
+	// root is the pagetable root.
+	root *Node
+
+	// translater is the translater passed at creation.
+	translater Translater
+
+	// archPageTables includes architecture-specific features.
+	archPageTables
+
+	// allNodes is a set of nodes indexed by translater address.
+	allNodes map[uintptr]*Node
+}
+
+// Translater translates to guest physical addresses.
+type Translater interface {
+	// TranslateToPhysical translates the given pointer object into a
+	// "physical" address. We do not require that it translates back, the
+	// reverse mapping is maintained internally.
+	TranslateToPhysical(*PTEs) uintptr
+}
+
+// New returns new PageTables.
+func New(t Translater, opts Opts) *PageTables {
+	p := &PageTables{
+		translater: t,
+		allNodes:   make(map[uintptr]*Node),
+	}
+	p.root = p.allocNode()
+	p.init(opts)
+	return p
+}
+
+// New returns a new set of PageTables derived from the given one.
+//
+// This function should always be preferred to New if there are existing
+// pagetables, as this function preserves architectural constraints relevant to
+// managing multiple sets of pagetables.
+func (p *PageTables) New() *PageTables {
+	np := &PageTables{
+		translater: p.translater,
+		allNodes:   make(map[uintptr]*Node),
+	}
+	np.root = np.allocNode()
+	np.initFrom(&p.archPageTables)
+	return np
+}
+
+// setPageTable sets the given index as a page table.
+func (p *PageTables) setPageTable(n *Node, index int, child *Node) {
+	phys := p.translater.TranslateToPhysical(child.PTEs())
+	p.allNodes[phys] = child
+	pte := &n.PTEs()[index]
+	pte.setPageTable(phys)
+}
+
+// clearPageTable clears the given entry.
+func (p *PageTables) clearPageTable(n *Node, index int) {
+	pte := &n.PTEs()[index]
+	physical := pte.Address()
+	pte.Clear()
+	delete(p.allNodes, physical)
+}
+
+// getPageTable returns the page table entry.
+func (p *PageTables) getPageTable(n *Node, index int) *Node {
+	pte := &n.PTEs()[index]
+	physical := pte.Address()
+	child := p.allNodes[physical]
+	return child
+}
+
+// Map installs a mapping with the given physical address.
+//
+// True is returned iff there was a previous mapping in the range.
+//
+// Precondition: addr & length must be aligned, their sum must not overflow.
+func (p *PageTables) Map(addr usermem.Addr, length uintptr, user bool, at usermem.AccessType, physical uintptr) bool {
+	if at == usermem.NoAccess {
+		return p.Unmap(addr, length)
+	}
+	prev := false
+	p.mu.Lock()
+	end, ok := addr.AddLength(uint64(length))
+	if !ok {
+		panic("pagetables.Map: overflow")
+	}
+	p.iterateRange(uintptr(addr), uintptr(end), true, func(s, e uintptr, pte *PTE, align uintptr) {
+		p := physical + (s - uintptr(addr))
+		prev = prev || (pte.Valid() && (p != pte.Address() || at.Write != pte.Writeable() || at.Execute != pte.Executable()))
+		if p&align != 0 {
+			// We will install entries at a smaller granulaity if
+			// we don't install a valid entry here, however we must
+			// zap any existing entry to ensure this happens.
+			pte.Clear()
+			return
+		}
+		pte.Set(p, at.Write, at.Execute, user)
+	})
+	p.mu.Unlock()
+	return prev
+}
+
+// Unmap unmaps the given range.
+//
+// True is returned iff there was a previous mapping in the range.
+func (p *PageTables) Unmap(addr usermem.Addr, length uintptr) bool {
+	p.mu.Lock()
+	count := 0
+	p.iterateRange(uintptr(addr), uintptr(addr)+length, false, func(s, e uintptr, pte *PTE, align uintptr) {
+		pte.Clear()
+		count++
+	})
+	p.mu.Unlock()
+	return count > 0
+}
+
+// Release releases this address space.
+//
+// This must be called to release the PCID.
+func (p *PageTables) Release() {
+	// Clear all pages.
+	p.Unmap(0, ^uintptr(0))
+	p.release()
+}
+
+// Lookup returns the physical address for the given virtual address.
+func (p *PageTables) Lookup(addr usermem.Addr) (physical uintptr, accessType usermem.AccessType) {
+	mask := uintptr(usermem.PageSize - 1)
+	off := uintptr(addr) & mask
+	addr = addr &^ usermem.Addr(mask)
+	p.iterateRange(uintptr(addr), uintptr(addr+usermem.PageSize), false, func(s, e uintptr, pte *PTE, align uintptr) {
+		if !pte.Valid() {
+			return
+		}
+		physical = pte.Address() + (s - uintptr(addr)) + off
+		accessType = usermem.AccessType{
+			Read:    true,
+			Write:   pte.Writeable(),
+			Execute: pte.Executable(),
+		}
+	})
+	return physical, accessType
+}
+
+// allocNode allocates a new page.
+func (p *PageTables) allocNode() *Node {
+	n := new(Node)
+	n.physical = p.translater.TranslateToPhysical(n.PTEs())
+	return n
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go b/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go
new file mode 100644
index 000000000..b89665c96
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_amd64.go
@@ -0,0 +1,397 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package pagetables
+
+import (
+	"fmt"
+	"sync/atomic"
+)
+
+// Address constraints.
+//
+// The lowerTop and upperBottom currently apply to four-level pagetables;
+// additional refactoring would be necessary to support five-level pagetables.
+const (
+	lowerTop    = 0x00007fffffffffff
+	upperBottom = 0xffff800000000000
+
+	pteShift = 12
+	pmdShift = 21
+	pudShift = 30
+	pgdShift = 39
+
+	pteMask = 0x1ff << pteShift
+	pmdMask = 0x1ff << pmdShift
+	pudMask = 0x1ff << pudShift
+	pgdMask = 0x1ff << pgdShift
+
+	pteSize = 1 << pteShift
+	pmdSize = 1 << pmdShift
+	pudSize = 1 << pudShift
+	pgdSize = 1 << pgdShift
+)
+
+// Bits in page table entries.
+const (
+	present        = 0x001
+	writable       = 0x002
+	user           = 0x004
+	writeThrough   = 0x008
+	cacheDisable   = 0x010
+	accessed       = 0x020
+	dirty          = 0x040
+	super          = 0x080
+	executeDisable = 1 << 63
+)
+
+// PTE is a page table entry.
+type PTE uint64
+
+// Clear clears this PTE, including super page information.
+func (p *PTE) Clear() {
+	atomic.StoreUint64((*uint64)(p), 0)
+}
+
+// Valid returns true iff this entry is valid.
+func (p *PTE) Valid() bool {
+	return atomic.LoadUint64((*uint64)(p))&present != 0
+}
+
+// Writeable returns true iff the page is writable.
+func (p *PTE) Writeable() bool {
+	return atomic.LoadUint64((*uint64)(p))&writable != 0
+}
+
+// User returns true iff the page is user-accessible.
+func (p *PTE) User() bool {
+	return atomic.LoadUint64((*uint64)(p))&user != 0
+}
+
+// Executable returns true iff the page is executable.
+func (p *PTE) Executable() bool {
+	return atomic.LoadUint64((*uint64)(p))&executeDisable == 0
+}
+
+// SetSuper sets this page as a super page.
+//
+// The page must not be valid or a panic will result.
+func (p *PTE) SetSuper() {
+	if p.Valid() {
+		// This is not allowed.
+		panic("SetSuper called on valid page!")
+	}
+	atomic.StoreUint64((*uint64)(p), super)
+}
+
+// IsSuper returns true iff this page is a super page.
+func (p *PTE) IsSuper() bool {
+	return atomic.LoadUint64((*uint64)(p))&super != 0
+}
+
+// Set sets this PTE value.
+func (p *PTE) Set(addr uintptr, write, execute bool, userAccessible bool) {
+	v := uint64(addr)&^uint64(0xfff) | present | accessed
+	if userAccessible {
+		v |= user
+	}
+	if !execute {
+		v |= executeDisable
+	}
+	if write {
+		v |= writable | dirty
+	}
+	if p.IsSuper() {
+		v |= super
+	}
+	atomic.StoreUint64((*uint64)(p), v)
+}
+
+// setPageTable sets this PTE value and forces the write bit and super bit to
+// be cleared. This is used explicitly for breaking super pages.
+func (p *PTE) setPageTable(addr uintptr) {
+	v := uint64(addr)&^uint64(0xfff) | present | user | writable | accessed | dirty
+	atomic.StoreUint64((*uint64)(p), v)
+}
+
+// Address extracts the address. This should only be used if Valid returns true.
+func (p *PTE) Address() uintptr {
+	return uintptr(atomic.LoadUint64((*uint64)(p)) & ^uint64(executeDisable|0xfff))
+}
+
+// entriesPerPage is the number of PTEs per page.
+const entriesPerPage = 512
+
+// PTEs is a collection of entries.
+type PTEs [entriesPerPage]PTE
+
+// next returns the next address quantized by the given size.
+func next(start uint64, size uint64) uint64 {
+	start &= ^(size - 1)
+	start += size
+	return start
+}
+
+// iterateRange iterates over all appropriate levels of page tables for the given range.
+//
+// If alloc is set, then Set _must_ be called on all given PTEs. The exception
+// is super pages. If a valid super page cannot be installed, then the walk
+// will continue to individual entries.
+//
+// This algorithm will attempt to maximize the use of super pages whenever
+// possible. Whether a super page is provided will be clear through the range
+// provided in the callback.
+//
+// Note that if alloc set, then no gaps will be present. However, if alloc is
+// not set, then the iteration will likely be full of gaps.
+//
+// Note that this function should generally be avoided in favor of Map, Unmap,
+// etc. when not necessary.
+//
+// Precondition: startAddr and endAddr must be page-aligned.
+//
+// Precondition: startStart must be less than endAddr.
+//
+// Precondition: If alloc is set, then startAddr and endAddr should not span
+// non-canonical ranges. If they do, a panic will result.
+func (p *PageTables) iterateRange(startAddr, endAddr uintptr, alloc bool, fn func(s, e uintptr, pte *PTE, align uintptr)) {
+	start := uint64(startAddr)
+	end := uint64(endAddr)
+	if start%pteSize != 0 {
+		panic(fmt.Sprintf("unaligned start: %v", start))
+	}
+	if start > end {
+		panic(fmt.Sprintf("start > end (%v > %v))", start, end))
+	}
+
+	// Deal with cases where we traverse the "gap".
+	//
+	// These are all explicitly disallowed if alloc is set, and we must
+	// traverse an entry for each address explicitly.
+	switch {
+	case start < lowerTop && end > lowerTop && end < upperBottom:
+		if alloc {
+			panic(fmt.Sprintf("alloc [%x, %x) spans non-canonical range", start, end))
+		}
+		p.iterateRange(startAddr, lowerTop, false, fn)
+		return
+	case start < lowerTop && end > lowerTop:
+		if alloc {
+			panic(fmt.Sprintf("alloc [%x, %x) spans non-canonical range", start, end))
+		}
+		p.iterateRange(startAddr, lowerTop, false, fn)
+		p.iterateRange(upperBottom, endAddr, false, fn)
+		return
+	case start > lowerTop && end < upperBottom:
+		if alloc {
+			panic(fmt.Sprintf("alloc [%x, %x) spans non-canonical range", start, end))
+		}
+		return
+	case start > lowerTop && start < upperBottom && end > upperBottom:
+		if alloc {
+			panic(fmt.Sprintf("alloc [%x, %x) spans non-canonical range", start, end))
+		}
+		p.iterateRange(upperBottom, endAddr, false, fn)
+		return
+	}
+
+	for pgdIndex := int((start & pgdMask) >> pgdShift); start < end && pgdIndex < entriesPerPage; pgdIndex++ {
+		pgdEntry := &p.root.PTEs()[pgdIndex]
+		if !pgdEntry.Valid() {
+			if !alloc {
+				// Skip over this entry.
+				start = next(start, pgdSize)
+				continue
+			}
+
+			// Allocate a new pgd.
+			p.setPageTable(p.root, pgdIndex, p.allocNode())
+		}
+
+		// Map the next level.
+		pudNode := p.getPageTable(p.root, pgdIndex)
+		clearPUDEntries := 0
+
+		for pudIndex := int((start & pudMask) >> pudShift); start < end && pudIndex < entriesPerPage; pudIndex++ {
+			pudEntry := &(pudNode.PTEs()[pudIndex])
+			if !pudEntry.Valid() {
+				if !alloc {
+					// Skip over this entry.
+					clearPUDEntries++
+					start = next(start, pudSize)
+					continue
+				}
+
+				// This level has 1-GB super pages. Is this
+				// entire region contained in a single PUD
+				// entry? If so, we can skip allocating a new
+				// page for the pmd.
+				if start&(pudSize-1) == 0 && end-start >= pudSize {
+					pudEntry.SetSuper()
+					fn(uintptr(start), uintptr(start+pudSize), pudEntry, pudSize-1)
+					if pudEntry.Valid() {
+						start = next(start, pudSize)
+						continue
+					}
+				}
+
+				// Allocate a new pud.
+				p.setPageTable(pudNode, pudIndex, p.allocNode())
+
+			} else if pudEntry.IsSuper() {
+				// Does this page need to be split?
+				if start&(pudSize-1) != 0 || end < next(start, pudSize) {
+					currentAddr := uint64(pudEntry.Address())
+					writeable := pudEntry.Writeable()
+					executable := pudEntry.Executable()
+					user := pudEntry.User()
+
+					// Install the relevant entries.
+					pmdNode := p.allocNode()
+					pmdEntries := pmdNode.PTEs()
+					for index := 0; index < entriesPerPage; index++ {
+						pmdEntry := &pmdEntries[index]
+						pmdEntry.SetSuper()
+						pmdEntry.Set(uintptr(currentAddr), writeable, executable, user)
+						currentAddr += pmdSize
+					}
+
+					// Reset to point to the new page.
+					p.setPageTable(pudNode, pudIndex, pmdNode)
+				} else {
+					// A super page to be checked directly.
+					fn(uintptr(start), uintptr(start+pudSize), pudEntry, pudSize-1)
+
+					// Might have been cleared.
+					if !pudEntry.Valid() {
+						clearPUDEntries++
+					}
+
+					// Note that the super page was changed.
+					start = next(start, pudSize)
+					continue
+				}
+			}
+
+			// Map the next level, since this is valid.
+			pmdNode := p.getPageTable(pudNode, pudIndex)
+			clearPMDEntries := 0
+
+			for pmdIndex := int((start & pmdMask) >> pmdShift); start < end && pmdIndex < entriesPerPage; pmdIndex++ {
+				pmdEntry := &pmdNode.PTEs()[pmdIndex]
+				if !pmdEntry.Valid() {
+					if !alloc {
+						// Skip over this entry.
+						clearPMDEntries++
+						start = next(start, pmdSize)
+						continue
+					}
+
+					// This level has 2-MB huge pages. If this
+					// region is contined in a single PMD entry?
+					// As above, we can skip allocating a new page.
+					if start&(pmdSize-1) == 0 && end-start >= pmdSize {
+						pmdEntry.SetSuper()
+						fn(uintptr(start), uintptr(start+pmdSize), pmdEntry, pmdSize-1)
+						if pmdEntry.Valid() {
+							start = next(start, pmdSize)
+							continue
+						}
+					}
+
+					// Allocate a new pmd.
+					p.setPageTable(pmdNode, pmdIndex, p.allocNode())
+
+				} else if pmdEntry.IsSuper() {
+					// Does this page need to be split?
+					if start&(pmdSize-1) != 0 || end < next(start, pmdSize) {
+						currentAddr := uint64(pmdEntry.Address())
+						writeable := pmdEntry.Writeable()
+						executable := pmdEntry.Executable()
+						user := pmdEntry.User()
+
+						// Install the relevant entries.
+						pteNode := p.allocNode()
+						pteEntries := pteNode.PTEs()
+						for index := 0; index < entriesPerPage; index++ {
+							pteEntry := &pteEntries[index]
+							pteEntry.Set(uintptr(currentAddr), writeable, executable, user)
+							currentAddr += pteSize
+						}
+
+						// Reset to point to the new page.
+						p.setPageTable(pmdNode, pmdIndex, pteNode)
+					} else {
+						// A huge page to be checked directly.
+						fn(uintptr(start), uintptr(start+pmdSize), pmdEntry, pmdSize-1)
+
+						// Might have been cleared.
+						if !pmdEntry.Valid() {
+							clearPMDEntries++
+						}
+
+						// Note that the huge page was changed.
+						start = next(start, pmdSize)
+						continue
+					}
+				}
+
+				// Map the next level, since this is valid.
+				pteNode := p.getPageTable(pmdNode, pmdIndex)
+				clearPTEEntries := 0
+
+				for pteIndex := int((start & pteMask) >> pteShift); start < end && pteIndex < entriesPerPage; pteIndex++ {
+					pteEntry := &pteNode.PTEs()[pteIndex]
+					if !pteEntry.Valid() && !alloc {
+						clearPTEEntries++
+						start += pteSize
+						continue
+					}
+
+					// At this point, we are guaranteed that start%pteSize == 0.
+					fn(uintptr(start), uintptr(start+pteSize), pteEntry, pteSize-1)
+					if !pteEntry.Valid() {
+						if alloc {
+							panic("PTE not set after iteration with alloc=true!")
+						}
+						clearPTEEntries++
+					}
+
+					// Note that the pte was changed.
+					start += pteSize
+					continue
+				}
+
+				// Check if we no longer need this page.
+				if clearPTEEntries == entriesPerPage {
+					p.clearPageTable(pmdNode, pmdIndex)
+					clearPMDEntries++
+				}
+			}
+
+			// Check if we no longer need this page.
+			if clearPMDEntries == entriesPerPage {
+				p.clearPageTable(pudNode, pudIndex)
+				clearPUDEntries++
+			}
+		}
+
+		// Check if we no longer need this page.
+		if clearPUDEntries == entriesPerPage {
+			p.clearPageTable(p.root, pgdIndex)
+		}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_test.go b/pkg/sentry/platform/ring0/pagetables/pagetables_test.go
new file mode 100644
index 000000000..9cbc0e3b0
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_test.go
@@ -0,0 +1,161 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pagetables
+
+import (
+	"reflect"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+type reflectTranslater struct{}
+
+func (r reflectTranslater) TranslateToPhysical(ptes *PTEs) uintptr {
+	return reflect.ValueOf(ptes).Pointer()
+}
+
+type mapping struct {
+	start     uintptr
+	length    uintptr
+	addr      uintptr
+	writeable bool
+}
+
+func checkMappings(t *testing.T, pt *PageTables, m []mapping) {
+	var (
+		current int
+		found   []mapping
+		failed  string
+	)
+
+	// Iterate over all the mappings.
+	pt.iterateRange(0, ^uintptr(0), false, func(s, e uintptr, pte *PTE, align uintptr) {
+		found = append(found, mapping{
+			start:     s,
+			length:    e - s,
+			addr:      pte.Address(),
+			writeable: pte.Writeable(),
+		})
+		if failed != "" {
+			// Don't keep looking for errors.
+			return
+		}
+
+		if current >= len(m) {
+			failed = "more mappings than expected"
+		} else if m[current].start != s {
+			failed = "start didn't match expected"
+		} else if m[current].length != (e - s) {
+			failed = "end didn't match expected"
+		} else if m[current].addr != pte.Address() {
+			failed = "address didn't match expected"
+		} else if m[current].writeable != pte.Writeable() {
+			failed = "writeable didn't match"
+		}
+		current++
+	})
+
+	// Were we expected additional mappings?
+	if failed == "" && current != len(m) {
+		failed = "insufficient mappings found"
+	}
+
+	// Emit a meaningful error message on failure.
+	if failed != "" {
+		t.Errorf("%s; got %#v, wanted %#v", failed, found, m)
+	}
+}
+
+func TestAllocFree(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+	pt.Release()
+}
+
+func TestUnmap(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map and unmap one entry.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+	pt.Unmap(0x400000, pteSize)
+
+	checkMappings(t, pt, nil)
+	pt.Release()
+}
+
+func TestReadOnly(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map one entry.
+	pt.Map(0x400000, pteSize, true, usermem.Read, pteSize*42)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, false},
+	})
+	pt.Release()
+}
+
+func TestReadWrite(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map one entry.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, true},
+	})
+	pt.Release()
+}
+
+func TestSerialEntries(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map two sequential entries.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+	pt.Map(0x401000, pteSize, true, usermem.ReadWrite, pteSize*47)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, true},
+		{0x401000, pteSize, pteSize * 47, true},
+	})
+	pt.Release()
+}
+
+func TestSpanningEntries(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Span a pgd with two pages.
+	pt.Map(0x00007efffffff000, 2*pteSize, true, usermem.Read, pteSize*42)
+
+	checkMappings(t, pt, []mapping{
+		{0x00007efffffff000, pteSize, pteSize * 42, false},
+		{0x00007f0000000000, pteSize, pteSize * 43, false},
+	})
+	pt.Release()
+}
+
+func TestSparseEntries(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map two entries in different pgds.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+	pt.Map(0x00007f0000000000, pteSize, true, usermem.Read, pteSize*47)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, true},
+		{0x00007f0000000000, pteSize, pteSize * 47, false},
+	})
+	pt.Release()
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_unsafe.go b/pkg/sentry/platform/ring0/pagetables/pagetables_unsafe.go
new file mode 100644
index 000000000..a2b44fb79
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_unsafe.go
@@ -0,0 +1,31 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package pagetables
+
+import (
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// PTEs returns aligned PTE entries.
+func (n *Node) PTEs() *PTEs {
+	addr := uintptr(unsafe.Pointer(&n.unalignedData[0]))
+	offset := addr & (usermem.PageSize - 1)
+	if offset != 0 {
+		offset = usermem.PageSize - offset
+	}
+	return (*PTEs)(unsafe.Pointer(addr + offset))
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go
new file mode 100644
index 000000000..dac66373f
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_x86.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+// Opts are pagetable options.
+type Opts struct {
+	EnablePCID bool
+}
+
+// archPageTables has x86-specific features.
+type archPageTables struct {
+	// pcids is the PCID database.
+	pcids *PCIDs
+
+	// pcid is the globally unique identifier, or zero if none were
+	// available or pcids is nil.
+	pcid uint16
+}
+
+// init initializes arch-specific features.
+func (a *archPageTables) init(opts Opts) {
+	if opts.EnablePCID {
+		a.pcids = NewPCIDs()
+		a.pcid = a.pcids.allocate()
+	}
+}
+
+// initFrom initializes arch-specific features from an existing entry.'
+func (a *archPageTables) initFrom(other *archPageTables) {
+	a.pcids = other.pcids // Refer to the same PCID database.
+	if a.pcids != nil {
+		a.pcid = a.pcids.allocate()
+	}
+}
+
+// release is called from Release.
+func (a *archPageTables) release() {
+	// Return the PCID.
+	if a.pcids != nil {
+		a.pcids.free(a.pcid)
+	}
+}
+
+// CR3 returns the CR3 value for these tables.
+//
+// This may be called in interrupt contexts.
+//
+//go:nosplit
+func (p *PageTables) CR3() uint64 {
+	// Bit 63 is set to avoid flushing the PCID (per SDM 4.10.4.1).
+	const noFlushBit uint64 = 0x8000000000000000
+	if p.pcid != 0 {
+		return noFlushBit | uint64(p.root.physical) | uint64(p.pcid)
+	}
+	return uint64(p.root.physical)
+}
+
+// FlushCR3 returns the CR3 value that flushes the TLB.
+//
+// This may be called in interrupt contexts.
+//
+//go:nosplit
+func (p *PageTables) FlushCR3() uint64 {
+	return uint64(p.root.physical) | uint64(p.pcid)
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_x86_test.go b/pkg/sentry/platform/ring0/pagetables/pagetables_x86_test.go
new file mode 100644
index 000000000..1fc403c48
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_x86_test.go
@@ -0,0 +1,79 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+import (
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+func Test2MAnd4K(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map a small page and a huge page.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+	pt.Map(0x00007f0000000000, 1<<21, true, usermem.Read, pmdSize*47)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, true},
+		{0x00007f0000000000, pmdSize, pmdSize * 47, false},
+	})
+	pt.Release()
+}
+
+func Test1GAnd4K(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map a small page and a super page.
+	pt.Map(0x400000, pteSize, true, usermem.ReadWrite, pteSize*42)
+	pt.Map(0x00007f0000000000, pudSize, true, usermem.Read, pudSize*47)
+
+	checkMappings(t, pt, []mapping{
+		{0x400000, pteSize, pteSize * 42, true},
+		{0x00007f0000000000, pudSize, pudSize * 47, false},
+	})
+	pt.Release()
+}
+
+func TestSplit1GPage(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map a super page and knock out the middle.
+	pt.Map(0x00007f0000000000, pudSize, true, usermem.Read, pudSize*42)
+	pt.Unmap(usermem.Addr(0x00007f0000000000+pteSize), pudSize-(2*pteSize))
+
+	checkMappings(t, pt, []mapping{
+		{0x00007f0000000000, pteSize, pudSize * 42, false},
+		{0x00007f0000000000 + pudSize - pteSize, pteSize, pudSize*42 + pudSize - pteSize, false},
+	})
+	pt.Release()
+}
+
+func TestSplit2MPage(t *testing.T) {
+	pt := New(reflectTranslater{}, Opts{})
+
+	// Map a huge page and knock out the middle.
+	pt.Map(0x00007f0000000000, pmdSize, true, usermem.Read, pmdSize*42)
+	pt.Unmap(usermem.Addr(0x00007f0000000000+pteSize), pmdSize-(2*pteSize))
+
+	checkMappings(t, pt, []mapping{
+		{0x00007f0000000000, pteSize, pmdSize * 42, false},
+		{0x00007f0000000000 + pmdSize - pteSize, pteSize, pmdSize*42 + pmdSize - pteSize, false},
+	})
+	pt.Release()
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86.go b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go
new file mode 100644
index 000000000..509e8c0d9
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86.go
@@ -0,0 +1,74 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+import (
+	"sync"
+)
+
+// maxPCID is the maximum allowed PCID.
+const maxPCID = 4095
+
+// PCIDs is a simple PCID database.
+type PCIDs struct {
+	mu sync.Mutex
+
+	// last is the last fresh PCID given out (not including the available
+	// pool). If last >= maxPCID, then the only PCIDs available in the
+	// available pool below.
+	last uint16
+
+	// available are PCIDs that have been freed.
+	available map[uint16]struct{}
+}
+
+// NewPCIDs returns a new PCID set.
+func NewPCIDs() *PCIDs {
+	return &PCIDs{
+		available: make(map[uint16]struct{}),
+	}
+}
+
+// allocate returns an unused PCID, or zero if all are taken.
+func (p *PCIDs) allocate() uint16 {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if len(p.available) > 0 {
+		for id := range p.available {
+			delete(p.available, id)
+			return id
+		}
+	}
+	if id := p.last + 1; id <= maxPCID {
+		p.last = id
+		return id
+	}
+	// Nothing available.
+	return 0
+}
+
+// free returns a PCID to the pool.
+//
+// It is safe to call free with a zero pcid. That is, you may always call free
+// with anything returned by allocate.
+func (p *PCIDs) free(id uint16) {
+	p.mu.Lock()
+	defer p.mu.Unlock()
+	if id != 0 {
+		p.available[id] = struct{}{}
+	}
+}
diff --git a/pkg/sentry/platform/ring0/pagetables/pcids_x86_test.go b/pkg/sentry/platform/ring0/pagetables/pcids_x86_test.go
new file mode 100644
index 000000000..0b555cd76
--- /dev/null
+++ b/pkg/sentry/platform/ring0/pagetables/pcids_x86_test.go
@@ -0,0 +1,65 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package pagetables
+
+import (
+	"testing"
+)
+
+func TestMaxPCID(t *testing.T) {
+	p := NewPCIDs()
+	for i := 0; i < maxPCID; i++ {
+		if id := p.allocate(); id != uint16(i+1) {
+			t.Errorf("got %d, expected %d", id, i+1)
+		}
+	}
+	if id := p.allocate(); id != 0 {
+		if id != 0 {
+			t.Errorf("got %d, expected 0", id)
+		}
+	}
+}
+
+func TestFirstPCID(t *testing.T) {
+	p := NewPCIDs()
+	if id := p.allocate(); id != 1 {
+		t.Errorf("got %d, expected 1", id)
+	}
+}
+
+func TestFreePCID(t *testing.T) {
+	p := NewPCIDs()
+	p.free(0)
+	if id := p.allocate(); id != 1 {
+		t.Errorf("got %d, expected 1 (not zero)", id)
+	}
+}
+
+func TestReusePCID(t *testing.T) {
+	p := NewPCIDs()
+	id := p.allocate()
+	if id != 1 {
+		t.Errorf("got %d, expected 1", id)
+	}
+	p.free(id)
+	if id := p.allocate(); id != 1 {
+		t.Errorf("got %d, expected 1", id)
+	}
+	if id := p.allocate(); id != 2 {
+		t.Errorf("got %d, expected 2", id)
+	}
+}
diff --git a/pkg/sentry/platform/ring0/ring0.go b/pkg/sentry/platform/ring0/ring0.go
new file mode 100644
index 000000000..4991031c5
--- /dev/null
+++ b/pkg/sentry/platform/ring0/ring0.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package ring0 provides basic operating system-level stubs.
+package ring0
diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go
new file mode 100644
index 000000000..e16f6c599
--- /dev/null
+++ b/pkg/sentry/platform/ring0/x86.go
@@ -0,0 +1,242 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build i386 amd64
+
+package ring0
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/cpuid"
+)
+
+// Useful bits.
+const (
+	_CR0_PE = 1 << 0
+	_CR0_ET = 1 << 4
+	_CR0_PG = 1 << 31
+
+	_CR4_PSE        = 1 << 4
+	_CR4_PAE        = 1 << 5
+	_CR4_PGE        = 1 << 7
+	_CR4_OSFXSR     = 1 << 9
+	_CR4_OSXMMEXCPT = 1 << 10
+	_CR4_FSGSBASE   = 1 << 16
+	_CR4_PCIDE      = 1 << 17
+	_CR4_OSXSAVE    = 1 << 18
+	_CR4_SMEP       = 1 << 20
+
+	_RFLAGS_AC       = 1 << 18
+	_RFLAGS_NT       = 1 << 14
+	_RFLAGS_IOPL     = 3 << 12
+	_RFLAGS_DF       = 1 << 10
+	_RFLAGS_IF       = 1 << 9
+	_RFLAGS_STEP     = 1 << 8
+	_RFLAGS_RESERVED = 1 << 1
+
+	_EFER_SCE = 0x001
+	_EFER_LME = 0x100
+	_EFER_NX  = 0x800
+
+	_MSR_STAR         = 0xc0000081
+	_MSR_LSTAR        = 0xc0000082
+	_MSR_CSTAR        = 0xc0000083
+	_MSR_SYSCALL_MASK = 0xc0000084
+)
+
+// Vector is an exception vector.
+type Vector uintptr
+
+// Exception vectors.
+const (
+	DivideByZero Vector = iota
+	Debug
+	NMI
+	Breakpoint
+	Overflow
+	BoundRangeExceeded
+	InvalidOpcode
+	DeviceNotAvailable
+	DoubleFault
+	CoprocessorSegmentOverrun
+	InvalidTSS
+	SegmentNotPresent
+	StackSegmentFault
+	GeneralProtectionFault
+	PageFault
+	_
+	X87FloatingPointException
+	AlignmentCheck
+	MachineCheck
+	SIMDFloatingPointException
+	VirtualizationException
+	SecurityException = 0x1e
+	SyscallInt80      = 0x80
+	_NR_INTERRUPTS    = SyscallInt80 + 1
+)
+
+// System call vectors.
+const (
+	Syscall Vector = _NR_INTERRUPTS
+)
+
+// VirtualAddressBits returns the number bits available for virtual addresses.
+//
+// Note that sign-extension semantics apply to the highest order bit.
+//
+// FIXME: This should use the cpuid passed to Init.
+func VirtualAddressBits() uint32 {
+	ax, _, _, _ := cpuid.HostID(0x80000008, 0)
+	return (ax >> 8) & 0xff
+}
+
+// PhysicalAddressBits returns the number of bits available for physical addresses.
+//
+// FIXME: This should use the cpuid passed to Init.
+func PhysicalAddressBits() uint32 {
+	ax, _, _, _ := cpuid.HostID(0x80000008, 0)
+	return ax & 0xff
+}
+
+// Selector is a segment Selector.
+type Selector uint16
+
+// SegmentDescriptor is a segment descriptor.
+type SegmentDescriptor struct {
+	bits [2]uint32
+}
+
+// descriptorTable is a collection of descriptors.
+type descriptorTable [32]SegmentDescriptor
+
+// SegmentDescriptorFlags are typed flags within a descriptor.
+type SegmentDescriptorFlags uint32
+
+// SegmentDescriptorFlag declarations.
+const (
+	SegmentDescriptorAccess     SegmentDescriptorFlags = 1 << 8  // Access bit (always set).
+	SegmentDescriptorWrite                             = 1 << 9  // Write permission.
+	SegmentDescriptorExpandDown                        = 1 << 10 // Grows down, not used.
+	SegmentDescriptorExecute                           = 1 << 11 // Execute permission.
+	SegmentDescriptorSystem                            = 1 << 12 // Zero => system, 1 => user code/data.
+	SegmentDescriptorPresent                           = 1 << 15 // Present.
+	SegmentDescriptorAVL                               = 1 << 20 // Available.
+	SegmentDescriptorLong                              = 1 << 21 // Long mode.
+	SegmentDescriptorDB                                = 1 << 22 // 16 or 32-bit.
+	SegmentDescriptorG                                 = 1 << 23 // Granularity: page or byte.
+)
+
+// Base returns the descriptor's base linear address.
+func (d *SegmentDescriptor) Base() uint32 {
+	return d.bits[1]&0xFF000000 | (d.bits[1]&0x000000FF)<<16 | d.bits[0]>>16
+}
+
+// Limit returns the descriptor size.
+func (d *SegmentDescriptor) Limit() uint32 {
+	l := d.bits[0]&0xFFFF | d.bits[1]&0xF0000
+	if d.bits[1]&uint32(SegmentDescriptorG) != 0 {
+		l <<= 12
+		l |= 0xFFF
+	}
+	return l
+}
+
+// Flags returns descriptor flags.
+func (d *SegmentDescriptor) Flags() SegmentDescriptorFlags {
+	return SegmentDescriptorFlags(d.bits[1] & 0x00F09F00)
+}
+
+// DPL returns the descriptor privilege level.
+func (d *SegmentDescriptor) DPL() int {
+	return int((d.bits[1] >> 13) & 3)
+}
+
+func (d *SegmentDescriptor) setNull() {
+	d.bits[0] = 0
+	d.bits[1] = 0
+}
+
+func (d *SegmentDescriptor) set(base, limit uint32, dpl int, flags SegmentDescriptorFlags) {
+	flags |= SegmentDescriptorPresent
+	if limit>>12 != 0 {
+		limit >>= 12
+		flags |= SegmentDescriptorG
+	}
+	d.bits[0] = base<<16 | limit&0xFFFF
+	d.bits[1] = base&0xFF000000 | (base>>16)&0xFF | limit&0x000F0000 | uint32(flags) | uint32(dpl)<<13
+}
+
+func (d *SegmentDescriptor) setCode32(base, limit uint32, dpl int) {
+	d.set(base, limit, dpl,
+		SegmentDescriptorDB|
+			SegmentDescriptorExecute|
+			SegmentDescriptorSystem)
+}
+
+func (d *SegmentDescriptor) setCode64(base, limit uint32, dpl int) {
+	d.set(base, limit, dpl,
+		SegmentDescriptorG|
+			SegmentDescriptorLong|
+			SegmentDescriptorExecute|
+			SegmentDescriptorSystem)
+}
+
+func (d *SegmentDescriptor) setData(base, limit uint32, dpl int) {
+	d.set(base, limit, dpl,
+		SegmentDescriptorWrite|
+			SegmentDescriptorSystem)
+}
+
+// setHi is only used for the TSS segment, which is magically 64-bits.
+func (d *SegmentDescriptor) setHi(base uint32) {
+	d.bits[0] = base
+	d.bits[1] = 0
+}
+
+// Gate64 is a 64-bit task, trap, or interrupt gate.
+type Gate64 struct {
+	bits [4]uint32
+}
+
+// idt64 is a 64-bit interrupt descriptor table.
+type idt64 [_NR_INTERRUPTS]Gate64
+
+func (g *Gate64) setInterrupt(cs Selector, rip uint64, dpl int, ist int) {
+	g.bits[0] = uint32(cs)<<16 | uint32(rip)&0xFFFF
+	g.bits[1] = uint32(rip)&0xFFFF0000 | SegmentDescriptorPresent | uint32(dpl)<<13 | 14<<8 | uint32(ist)&0x7
+	g.bits[2] = uint32(rip >> 32)
+}
+
+func (g *Gate64) setTrap(cs Selector, rip uint64, dpl int, ist int) {
+	g.setInterrupt(cs, rip, dpl, ist)
+	g.bits[1] |= 1 << 8
+}
+
+// TaskState64 is a 64-bit task state structure.
+type TaskState64 struct {
+	_              uint32
+	rsp0Lo, rsp0Hi uint32
+	rsp1Lo, rsp1Hi uint32
+	rsp2Lo, rsp2Hi uint32
+	_              [2]uint32
+	ist1Lo, ist1Hi uint32
+	ist2Lo, ist2Hi uint32
+	ist3Lo, ist3Hi uint32
+	ist4Lo, ist4Hi uint32
+	ist5Lo, ist5Hi uint32
+	ist6Lo, ist6Hi uint32
+	ist7Lo, ist7Hi uint32
+	_              [2]uint32
+	_              uint16
+	ioPerm         uint16
+}
diff --git a/pkg/sentry/platform/safecopy/BUILD b/pkg/sentry/platform/safecopy/BUILD
new file mode 100644
index 000000000..8b9f29403
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/BUILD
@@ -0,0 +1,28 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "safecopy",
+    srcs = [
+        "atomic_amd64.s",
+        "memclr_amd64.s",
+        "memcpy_amd64.s",
+        "safecopy.go",
+        "safecopy_unsafe.go",
+        "sighandler_amd64.s",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/platform/safecopy",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "safecopy_test",
+    srcs = [
+        "safecopy_test.go",
+    ],
+    embed = [":safecopy"],
+)
diff --git a/pkg/sentry/platform/safecopy/atomic_amd64.s b/pkg/sentry/platform/safecopy/atomic_amd64.s
new file mode 100644
index 000000000..69947dec3
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/atomic_amd64.s
@@ -0,0 +1,108 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// handleSwapUint32Fault returns the value stored in DI. Control is transferred
+// to it when swapUint32 below receives SIGSEGV or SIGBUS, with the signal
+// number stored in DI.
+//
+// It must have the same frame configuration as swapUint32 so that it can undo
+// any potential call frame set up by the assembler.
+TEXT handleSwapUint32Fault(SB), NOSPLIT, $0-24
+  MOVL DI, sig+20(FP)
+  RET
+
+// swapUint32 atomically stores new into *addr and returns (the previous *addr
+// value, 0). If a SIGSEGV or SIGBUS signal is received during the swap, the
+// value of old is unspecified, and sig is the number of the signal that was
+// received.
+//
+// Preconditions: addr must be aligned to a 4-byte boundary.
+//
+//func swapUint32(ptr unsafe.Pointer, new uint32) (old uint32, sig int32)
+TEXT ·swapUint32(SB), NOSPLIT, $0-24
+  // Store 0 as the returned signal number. If we run to completion,
+  // this is the value the caller will see; if a signal is received,
+  // handleSwapUint32Fault will store a different value in this address.
+  MOVL $0, sig+20(FP)
+
+  MOVQ addr+0(FP), DI
+  MOVL new+8(FP), AX
+  XCHGL AX, 0(DI)
+  MOVL AX, old+16(FP)
+  RET
+
+// handleSwapUint64Fault returns the value stored in DI. Control is transferred
+// to it when swapUint64 below receives SIGSEGV or SIGBUS, with the signal
+// number stored in DI.
+//
+// It must have the same frame configuration as swapUint64 so that it can undo
+// any potential call frame set up by the assembler.
+TEXT handleSwapUint64Fault(SB), NOSPLIT, $0-28
+  MOVL DI, sig+24(FP)
+  RET
+
+// swapUint64 atomically stores new into *addr and returns (the previous *addr
+// value, 0). If a SIGSEGV or SIGBUS signal is received during the swap, the
+// value of old is unspecified, and sig is the number of the signal that was
+// received.
+//
+// Preconditions: addr must be aligned to a 8-byte boundary.
+//
+//func swapUint64(ptr unsafe.Pointer, new uint64) (old uint64, sig int32)
+TEXT ·swapUint64(SB), NOSPLIT, $0-28
+  // Store 0 as the returned signal number. If we run to completion,
+  // this is the value the caller will see; if a signal is received,
+  // handleSwapUint64Fault will store a different value in this address.
+  MOVL $0, sig+24(FP)
+
+  MOVQ addr+0(FP), DI
+  MOVQ new+8(FP), AX
+  XCHGQ AX, 0(DI)
+  MOVQ AX, old+16(FP)
+  RET
+
+// handleCompareAndSwapUint32Fault returns the value stored in DI. Control is
+// transferred to it when swapUint64 below receives SIGSEGV or SIGBUS, with the
+// signal number stored in DI.
+//
+// It must have the same frame configuration as compareAndSwapUint32 so that it
+// can undo any potential call frame set up by the assembler.
+TEXT handleCompareAndSwapUint32Fault(SB), NOSPLIT, $0-24
+  MOVL DI, sig+20(FP)
+  RET
+
+// compareAndSwapUint32 is like sync/atomic.CompareAndSwapUint32, but returns
+// (the value previously stored at addr, 0). If a SIGSEGV or SIGBUS signal is
+// received during the operation, the value of prev is unspecified, and sig is
+// the number of the signal that was received.
+//
+// Preconditions: addr must be aligned to a 4-byte boundary.
+//
+//func compareAndSwapUint32(ptr unsafe.Pointer, old, new uint32) (prev uint32, sig int32)
+TEXT ·compareAndSwapUint32(SB), NOSPLIT, $0-24
+  // Store 0 as the returned signal number. If we run to completion, this is
+  // the value the caller will see; if a signal is received,
+  // handleCompareAndSwapUint32Fault will store a different value in this
+  // address.
+  MOVL $0, sig+20(FP)
+
+  MOVQ addr+0(FP), DI
+  MOVL old+8(FP), AX
+  MOVL new+12(FP), DX
+  LOCK
+  CMPXCHGL DX, 0(DI)
+  MOVL AX, prev+16(FP)
+  RET
diff --git a/pkg/sentry/platform/safecopy/memclr_amd64.s b/pkg/sentry/platform/safecopy/memclr_amd64.s
new file mode 100644
index 000000000..7d1019f60
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/memclr_amd64.s
@@ -0,0 +1,157 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// handleMemclrFault returns (the value stored in AX, the value stored in DI).
+// Control is transferred to it when memclr below receives SIGSEGV or SIGBUS,
+// with the faulting address stored in AX and the signal number stored in DI.
+//
+// It must have the same frame configuration as memclr so that it can undo any
+// potential call frame set up by the assembler.
+TEXT handleMemclrFault(SB), NOSPLIT, $0-28
+	MOVQ	AX, addr+16(FP)
+	MOVL	DI, sig+24(FP)
+	RET
+
+// memclr sets the n bytes following ptr to zeroes. If a SIGSEGV or SIGBUS
+// signal is received during the write, it returns the address that caused the
+// fault and the number of the signal that was received. Otherwise, it returns
+// an unspecified address and a signal number of 0.
+//
+// Data is written in order, such that if a fault happens at address p, it is
+// safe to assume that all data before p-maxRegisterSize has already been
+// successfully written.
+//
+// The code is derived from runtime.memclrNoHeapPointers.
+//
+// func memclr(ptr unsafe.Pointer, n uintptr) (fault unsafe.Pointer, sig int32)
+TEXT ·memclr(SB), NOSPLIT, $0-28
+	// Store 0 as the returned signal number. If we run to completion,
+	// this is the value the caller will see; if a signal is received,
+	// handleMemclrFault will store a different value in this address.
+	MOVL	$0, sig+24(FP)
+
+	MOVQ	ptr+0(FP), DI
+	MOVQ	n+8(FP), BX
+	XORQ	AX, AX
+
+	// MOVOU seems always faster than REP STOSQ.
+tail:
+	TESTQ	BX, BX
+	JEQ	_0
+	CMPQ	BX, $2
+	JBE	_1or2
+	CMPQ	BX, $4
+	JBE	_3or4
+	CMPQ	BX, $8
+	JB	_5through7
+	JE	_8
+	CMPQ	BX, $16
+	JBE	_9through16
+	PXOR	X0, X0
+	CMPQ	BX, $32
+	JBE	_17through32
+	CMPQ	BX, $64
+	JBE	_33through64
+	CMPQ	BX, $128
+	JBE	_65through128
+	CMPQ	BX, $256
+	JBE	_129through256
+	// TODO: use branch table and BSR to make this just a single dispatch
+	// TODO: for really big clears, use MOVNTDQ, even without AVX2.
+
+loop:
+	MOVOU	X0, 0(DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, 128(DI)
+	MOVOU	X0, 144(DI)
+	MOVOU	X0, 160(DI)
+	MOVOU	X0, 176(DI)
+	MOVOU	X0, 192(DI)
+	MOVOU	X0, 208(DI)
+	MOVOU	X0, 224(DI)
+	MOVOU	X0, 240(DI)
+	SUBQ	$256, BX
+	ADDQ	$256, DI
+	CMPQ	BX, $256
+	JAE	loop
+	JMP	tail
+
+_1or2:
+	MOVB	AX, (DI)
+	MOVB	AX, -1(DI)(BX*1)
+	RET
+_0:
+	RET
+_3or4:
+	MOVW	AX, (DI)
+	MOVW	AX, -2(DI)(BX*1)
+	RET
+_5through7:
+	MOVL	AX, (DI)
+	MOVL	AX, -4(DI)(BX*1)
+	RET
+_8:
+	// We need a separate case for 8 to make sure we clear pointers atomically.
+	MOVQ	AX, (DI)
+	RET
+_9through16:
+	MOVQ	AX, (DI)
+	MOVQ	AX, -8(DI)(BX*1)
+	RET
+_17through32:
+	MOVOU	X0, (DI)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+_33through64:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+_65through128:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
+_129through256:
+	MOVOU	X0, (DI)
+	MOVOU	X0, 16(DI)
+	MOVOU	X0, 32(DI)
+	MOVOU	X0, 48(DI)
+	MOVOU	X0, 64(DI)
+	MOVOU	X0, 80(DI)
+	MOVOU	X0, 96(DI)
+	MOVOU	X0, 112(DI)
+	MOVOU	X0, -128(DI)(BX*1)
+	MOVOU	X0, -112(DI)(BX*1)
+	MOVOU	X0, -96(DI)(BX*1)
+	MOVOU	X0, -80(DI)(BX*1)
+	MOVOU	X0, -64(DI)(BX*1)
+	MOVOU	X0, -48(DI)(BX*1)
+	MOVOU	X0, -32(DI)(BX*1)
+	MOVOU	X0, -16(DI)(BX*1)
+	RET
diff --git a/pkg/sentry/platform/safecopy/memcpy_amd64.s b/pkg/sentry/platform/safecopy/memcpy_amd64.s
new file mode 100644
index 000000000..96ef2eefc
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/memcpy_amd64.s
@@ -0,0 +1,242 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// handleMemcpyFault returns (the value stored in AX, the value stored in DI).
+// Control is transferred to it when memcpy below receives SIGSEGV or SIGBUS,
+// with the faulting address stored in AX and the signal number stored in DI.
+//
+// It must have the same frame configuration as memcpy so that it can undo any
+// potential call frame set up by the assembler.
+TEXT handleMemcpyFault(SB), NOSPLIT, $0-36
+	MOVQ	AX, addr+24(FP)
+	MOVL	DI, sig+32(FP)
+	RET
+
+// memcpy copies data from src to dst. If a SIGSEGV or SIGBUS signal is received
+// during the copy, it returns the address that caused the fault and the number
+// of the signal that was received. Otherwise, it returns an unspecified address
+// and a signal number of 0.
+//
+// Data is copied in order, such that if a fault happens at address p, it is
+// safe to assume that all data before p-maxRegisterSize has already been
+// successfully copied.
+//
+// The code is derived from the forward copying part of runtime.memmove.
+//
+// func memcpy(dst, src unsafe.Pointer, n uintptr) (fault unsafe.Pointer, sig int32)
+TEXT ·memcpy(SB), NOSPLIT, $0-36
+	// Store 0 as the returned signal number. If we run to completion,
+	// this is the value the caller will see; if a signal is received,
+	// handleMemcpyFault will store a different value in this address.
+	MOVL	$0, sig+32(FP)
+
+	MOVQ	to+0(FP), DI
+	MOVQ	from+8(FP), SI
+	MOVQ	n+16(FP), BX
+
+	// REP instructions have a high startup cost, so we handle small sizes
+	// with some straightline code. The REP MOVSQ instruction is really fast
+	// for large sizes. The cutover is approximately 2K.
+tail:
+	// move_129through256 or smaller work whether or not the source and the
+	// destination memory regions overlap because they load all data into
+	// registers before writing it back.  move_256through2048 on the other
+	// hand can be used only when the memory regions don't overlap or the copy
+	// direction is forward.
+	TESTQ	BX, BX
+	JEQ	move_0
+	CMPQ	BX, $2
+	JBE	move_1or2
+	CMPQ	BX, $4
+	JBE	move_3or4
+	CMPQ	BX, $8
+	JB	move_5through7
+	JE	move_8
+	CMPQ	BX, $16
+	JBE	move_9through16
+	CMPQ	BX, $32
+	JBE	move_17through32
+	CMPQ	BX, $64
+	JBE	move_33through64
+	CMPQ	BX, $128
+	JBE	move_65through128
+	CMPQ	BX, $256
+	JBE	move_129through256
+	// TODO: use branch table and BSR to make this just a single dispatch
+
+/*
+ * forward copy loop
+ */
+	CMPQ	BX, $2048
+	JLS	move_256through2048
+
+	// Check alignment
+	MOVL	SI, AX
+	ORL	DI, AX
+	TESTL	$7, AX
+	JEQ	fwdBy8
+
+	// Do 1 byte at a time
+	MOVQ	BX, CX
+	REP;	MOVSB
+	RET
+
+fwdBy8:
+	// Do 8 bytes at a time
+	MOVQ	BX, CX
+	SHRQ	$3, CX
+	ANDQ	$7, BX
+	REP;	MOVSQ
+	JMP	tail
+
+move_1or2:
+	MOVB	(SI), AX
+	MOVB	AX, (DI)
+	MOVB	-1(SI)(BX*1), CX
+	MOVB	CX, -1(DI)(BX*1)
+	RET
+move_0:
+	RET
+move_3or4:
+	MOVW	(SI), AX
+	MOVW	AX, (DI)
+	MOVW	-2(SI)(BX*1), CX
+	MOVW	CX, -2(DI)(BX*1)
+	RET
+move_5through7:
+	MOVL	(SI), AX
+	MOVL	AX, (DI)
+	MOVL	-4(SI)(BX*1), CX
+	MOVL	CX, -4(DI)(BX*1)
+	RET
+move_8:
+	// We need a separate case for 8 to make sure we write pointers atomically.
+	MOVQ	(SI), AX
+	MOVQ	AX, (DI)
+	RET
+move_9through16:
+	MOVQ	(SI), AX
+	MOVQ	AX, (DI)
+	MOVQ	-8(SI)(BX*1), CX
+	MOVQ	CX, -8(DI)(BX*1)
+	RET
+move_17through32:
+	MOVOU	(SI), X0
+	MOVOU	X0, (DI)
+	MOVOU	-16(SI)(BX*1), X1
+	MOVOU	X1, -16(DI)(BX*1)
+	RET
+move_33through64:
+	MOVOU	(SI), X0
+	MOVOU	X0, (DI)
+	MOVOU	16(SI), X1
+	MOVOU	X1, 16(DI)
+	MOVOU	-32(SI)(BX*1), X2
+	MOVOU	X2, -32(DI)(BX*1)
+	MOVOU	-16(SI)(BX*1), X3
+	MOVOU	X3, -16(DI)(BX*1)
+	RET
+move_65through128:
+	MOVOU	(SI), X0
+	MOVOU	X0, (DI)
+	MOVOU	16(SI), X1
+	MOVOU	X1, 16(DI)
+	MOVOU	32(SI), X2
+	MOVOU	X2, 32(DI)
+	MOVOU	48(SI), X3
+	MOVOU	X3, 48(DI)
+	MOVOU	-64(SI)(BX*1), X4
+	MOVOU	X4, -64(DI)(BX*1)
+	MOVOU	-48(SI)(BX*1), X5
+	MOVOU	X5, -48(DI)(BX*1)
+	MOVOU	-32(SI)(BX*1), X6
+	MOVOU	X6, -32(DI)(BX*1)
+	MOVOU	-16(SI)(BX*1), X7
+	MOVOU	X7, -16(DI)(BX*1)
+	RET
+move_129through256:
+	MOVOU	(SI), X0
+	MOVOU	X0, (DI)
+	MOVOU	16(SI), X1
+	MOVOU	X1, 16(DI)
+	MOVOU	32(SI), X2
+	MOVOU	X2, 32(DI)
+	MOVOU	48(SI), X3
+	MOVOU	X3, 48(DI)
+	MOVOU	64(SI), X4
+	MOVOU	X4, 64(DI)
+	MOVOU	80(SI), X5
+	MOVOU	X5, 80(DI)
+	MOVOU	96(SI), X6
+	MOVOU	X6, 96(DI)
+	MOVOU	112(SI), X7
+	MOVOU	X7, 112(DI)
+	MOVOU	-128(SI)(BX*1), X8
+	MOVOU	X8, -128(DI)(BX*1)
+	MOVOU	-112(SI)(BX*1), X9
+	MOVOU	X9, -112(DI)(BX*1)
+	MOVOU	-96(SI)(BX*1), X10
+	MOVOU	X10, -96(DI)(BX*1)
+	MOVOU	-80(SI)(BX*1), X11
+	MOVOU	X11, -80(DI)(BX*1)
+	MOVOU	-64(SI)(BX*1), X12
+	MOVOU	X12, -64(DI)(BX*1)
+	MOVOU	-48(SI)(BX*1), X13
+	MOVOU	X13, -48(DI)(BX*1)
+	MOVOU	-32(SI)(BX*1), X14
+	MOVOU	X14, -32(DI)(BX*1)
+	MOVOU	-16(SI)(BX*1), X15
+	MOVOU	X15, -16(DI)(BX*1)
+	RET
+move_256through2048:
+	SUBQ	$256, BX
+	MOVOU	(SI), X0
+	MOVOU	X0, (DI)
+	MOVOU	16(SI), X1
+	MOVOU	X1, 16(DI)
+	MOVOU	32(SI), X2
+	MOVOU	X2, 32(DI)
+	MOVOU	48(SI), X3
+	MOVOU	X3, 48(DI)
+	MOVOU	64(SI), X4
+	MOVOU	X4, 64(DI)
+	MOVOU	80(SI), X5
+	MOVOU	X5, 80(DI)
+	MOVOU	96(SI), X6
+	MOVOU	X6, 96(DI)
+	MOVOU	112(SI), X7
+	MOVOU	X7, 112(DI)
+	MOVOU	128(SI), X8
+	MOVOU	X8, 128(DI)
+	MOVOU	144(SI), X9
+	MOVOU	X9, 144(DI)
+	MOVOU	160(SI), X10
+	MOVOU	X10, 160(DI)
+	MOVOU	176(SI), X11
+	MOVOU	X11, 176(DI)
+	MOVOU	192(SI), X12
+	MOVOU	X12, 192(DI)
+	MOVOU	208(SI), X13
+	MOVOU	X13, 208(DI)
+	MOVOU	224(SI), X14
+	MOVOU	X14, 224(DI)
+	MOVOU	240(SI), X15
+	MOVOU	X15, 240(DI)
+	CMPQ	BX, $256
+	LEAQ	256(SI), SI
+	LEAQ	256(DI), DI
+	JGE	move_256through2048
+	JMP	tail
diff --git a/pkg/sentry/platform/safecopy/safecopy.go b/pkg/sentry/platform/safecopy/safecopy.go
new file mode 100644
index 000000000..90a2aad7b
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/safecopy.go
@@ -0,0 +1,140 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package safecopy provides an efficient implementation of functions to access
+// memory that may result in SIGSEGV or SIGBUS being sent to the accessor.
+package safecopy
+
+import (
+	"fmt"
+	"reflect"
+	"runtime"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// SegvError is returned when a safecopy function receives SIGSEGV.
+type SegvError struct {
+	// Addr is the address at which the SIGSEGV occurred.
+	Addr uintptr
+}
+
+// Error implements error.Error.
+func (e SegvError) Error() string {
+	return fmt.Sprintf("SIGSEGV at %#x", e.Addr)
+}
+
+// BusError is returned when a safecopy function receives SIGBUS.
+type BusError struct {
+	// Addr is the address at which the SIGBUS occurred.
+	Addr uintptr
+}
+
+// Error implements error.Error.
+func (e BusError) Error() string {
+	return fmt.Sprintf("SIGBUS at %#x", e.Addr)
+}
+
+// AlignmentError is returned when a safecopy function is passed an address
+// that does not meet alignment requirements.
+type AlignmentError struct {
+	// Addr is the invalid address.
+	Addr uintptr
+
+	// Alignment is the required alignment.
+	Alignment uintptr
+}
+
+// Error implements error.Error.
+func (e AlignmentError) Error() string {
+	return fmt.Sprintf("address %#x is not aligned to a %d-byte boundary", e.Addr, e.Alignment)
+}
+
+var (
+	// The begin and end addresses below are for the functions that are
+	// checked by the signal handler.
+	memcpyBegin               uintptr
+	memcpyEnd                 uintptr
+	memclrBegin               uintptr
+	memclrEnd                 uintptr
+	swapUint32Begin           uintptr
+	swapUint32End             uintptr
+	swapUint64Begin           uintptr
+	swapUint64End             uintptr
+	compareAndSwapUint32Begin uintptr
+	compareAndSwapUint32End   uintptr
+
+	// savedSigSegVHandler is a pointer to the SIGSEGV handler that was
+	// configured before we replaced it with our own. We still call into it
+	// when we get a SIGSEGV that is not interesting to us.
+	savedSigSegVHandler uintptr
+
+	// same a above, but for SIGBUS signals.
+	savedSigBusHandler uintptr
+)
+
+// signalHandler is our replacement signal handler for SIGSEGV and SIGBUS
+// signals.
+func signalHandler()
+
+// FindEndAddress returns the end address (one byte beyond the last) of the
+// function that contains the specified address (begin).
+func FindEndAddress(begin uintptr) uintptr {
+	f := runtime.FuncForPC(begin)
+	if f != nil {
+		for p := begin; ; p++ {
+			g := runtime.FuncForPC(p)
+			if f != g {
+				return p
+			}
+		}
+	}
+	return begin
+}
+
+// initializeAddresses initializes the addresses used by the signal handler.
+func initializeAddresses() {
+	// The following functions are written in assembly language, so they won't
+	// be inlined by the existing compiler/linker. Tests will fail if this
+	// assumption is violated.
+	memcpyBegin = reflect.ValueOf(memcpy).Pointer()
+	memcpyEnd = FindEndAddress(memcpyBegin)
+	memclrBegin = reflect.ValueOf(memclr).Pointer()
+	memclrEnd = FindEndAddress(memclrBegin)
+	swapUint32Begin = reflect.ValueOf(swapUint32).Pointer()
+	swapUint32End = FindEndAddress(swapUint32Begin)
+	swapUint64Begin = reflect.ValueOf(swapUint64).Pointer()
+	swapUint64End = FindEndAddress(swapUint64Begin)
+	compareAndSwapUint32Begin = reflect.ValueOf(compareAndSwapUint32).Pointer()
+	compareAndSwapUint32End = FindEndAddress(compareAndSwapUint32Begin)
+}
+
+func init() {
+	initializeAddresses()
+	if err := ReplaceSignalHandler(syscall.SIGSEGV, reflect.ValueOf(signalHandler).Pointer(), &savedSigSegVHandler); err != nil {
+		panic(fmt.Sprintf("Unable to set handler for SIGSEGV: %v", err))
+	}
+	if err := ReplaceSignalHandler(syscall.SIGBUS, reflect.ValueOf(signalHandler).Pointer(), &savedSigBusHandler); err != nil {
+		panic(fmt.Sprintf("Unable to set handler for SIGBUS: %v", err))
+	}
+	syserror.AddErrorUnwrapper(func(e error) (syscall.Errno, bool) {
+		switch e.(type) {
+		case SegvError, BusError, AlignmentError:
+			return syscall.EFAULT, true
+		default:
+			return 0, false
+		}
+	})
+}
diff --git a/pkg/sentry/platform/safecopy/safecopy_test.go b/pkg/sentry/platform/safecopy/safecopy_test.go
new file mode 100644
index 000000000..67df36121
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/safecopy_test.go
@@ -0,0 +1,617 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safecopy
+
+import (
+	"bytes"
+	"fmt"
+	"io/ioutil"
+	"math/rand"
+	"os"
+	"runtime/debug"
+	"syscall"
+	"testing"
+	"unsafe"
+)
+
+// Size of a page in bytes. Cloned from usermem.PageSize to avoid a circular
+// dependency.
+const pageSize = 4096
+
+func initRandom(b []byte) {
+	for i := range b {
+		b[i] = byte(rand.Intn(256))
+	}
+}
+
+func randBuf(size int) []byte {
+	b := make([]byte, size)
+	initRandom(b)
+	return b
+}
+
+func TestCopyInSuccess(t *testing.T) {
+	// Test that CopyIn does not return an error when all pages are accessible.
+	const bufLen = 8192
+	a := randBuf(bufLen)
+	b := make([]byte, bufLen)
+
+	n, err := CopyIn(b, unsafe.Pointer(&a[0]))
+	if n != bufLen {
+		t.Errorf("Unexpected copy length, got %v, want %v", n, bufLen)
+	}
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if !bytes.Equal(a, b) {
+		t.Errorf("Buffers are not equal when they should be: %v %v", a, b)
+	}
+}
+
+func TestCopyOutSuccess(t *testing.T) {
+	// Test that CopyOut does not return an error when all pages are
+	// accessible.
+	const bufLen = 8192
+	a := randBuf(bufLen)
+	b := make([]byte, bufLen)
+
+	n, err := CopyOut(unsafe.Pointer(&b[0]), a)
+	if n != bufLen {
+		t.Errorf("Unexpected copy length, got %v, want %v", n, bufLen)
+	}
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if !bytes.Equal(a, b) {
+		t.Errorf("Buffers are not equal when they should be: %v %v", a, b)
+	}
+}
+
+func TestCopySuccess(t *testing.T) {
+	// Test that Copy does not return an error when all pages are accessible.
+	const bufLen = 8192
+	a := randBuf(bufLen)
+	b := make([]byte, bufLen)
+
+	n, err := Copy(unsafe.Pointer(&b[0]), unsafe.Pointer(&a[0]), bufLen)
+	if n != bufLen {
+		t.Errorf("Unexpected copy length, got %v, want %v", n, bufLen)
+	}
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if !bytes.Equal(a, b) {
+		t.Errorf("Buffers are not equal when they should be: %v %v", a, b)
+	}
+}
+
+func TestZeroOutSuccess(t *testing.T) {
+	// Test that ZeroOut does not return an error when all pages are
+	// accessible.
+	const bufLen = 8192
+	a := make([]byte, bufLen)
+	b := randBuf(bufLen)
+
+	n, err := ZeroOut(unsafe.Pointer(&b[0]), bufLen)
+	if n != bufLen {
+		t.Errorf("Unexpected copy length, got %v, want %v", n, bufLen)
+	}
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if !bytes.Equal(a, b) {
+		t.Errorf("Buffers are not equal when they should be: %v %v", a, b)
+	}
+}
+
+func TestSwapUint32Success(t *testing.T) {
+	// Test that SwapUint32 does not return an error when the page is
+	// accessible.
+	before := uint32(rand.Int31())
+	after := uint32(rand.Int31())
+	val := before
+
+	old, err := SwapUint32(unsafe.Pointer(&val), after)
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if old != before {
+		t.Errorf("Unexpected old value: got %v, want %v", old, before)
+	}
+	if val != after {
+		t.Errorf("Unexpected new value: got %v, want %v", val, after)
+	}
+}
+
+func TestSwapUint32AlignmentError(t *testing.T) {
+	// Test that SwapUint32 returns an AlignmentError when passed an unaligned
+	// address.
+	data := new(struct{ val uint64 })
+	addr := uintptr(unsafe.Pointer(&data.val)) + 1
+	want := AlignmentError{Addr: addr, Alignment: 4}
+	if _, err := SwapUint32(unsafe.Pointer(addr), 1); err != want {
+		t.Errorf("Unexpected error: got %v, want %v", err, want)
+	}
+}
+
+func TestSwapUint64Success(t *testing.T) {
+	// Test that SwapUint64 does not return an error when the page is
+	// accessible.
+	before := uint64(rand.Int63())
+	after := uint64(rand.Int63())
+	// "The first word in ... an allocated struct or slice can be relied upon
+	// to be 64-bit aligned." - sync/atomic docs
+	data := new(struct{ val uint64 })
+	data.val = before
+
+	old, err := SwapUint64(unsafe.Pointer(&data.val), after)
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if old != before {
+		t.Errorf("Unexpected old value: got %v, want %v", old, before)
+	}
+	if data.val != after {
+		t.Errorf("Unexpected new value: got %v, want %v", data.val, after)
+	}
+}
+
+func TestSwapUint64AlignmentError(t *testing.T) {
+	// Test that SwapUint64 returns an AlignmentError when passed an unaligned
+	// address.
+	data := new(struct{ val1, val2 uint64 })
+	addr := uintptr(unsafe.Pointer(&data.val1)) + 1
+	want := AlignmentError{Addr: addr, Alignment: 8}
+	if _, err := SwapUint64(unsafe.Pointer(addr), 1); err != want {
+		t.Errorf("Unexpected error: got %v, want %v", err, want)
+	}
+}
+
+func TestCompareAndSwapUint32Success(t *testing.T) {
+	// Test that CompareAndSwapUint32 does not return an error when the page is
+	// accessible.
+	before := uint32(rand.Int31())
+	after := uint32(rand.Int31())
+	val := before
+
+	old, err := CompareAndSwapUint32(unsafe.Pointer(&val), before, after)
+	if err != nil {
+		t.Errorf("Unexpected error: %v", err)
+	}
+	if old != before {
+		t.Errorf("Unexpected old value: got %v, want %v", old, before)
+	}
+	if val != after {
+		t.Errorf("Unexpected new value: got %v, want %v", val, after)
+	}
+}
+
+func TestCompareAndSwapUint32AlignmentError(t *testing.T) {
+	// Test that CompareAndSwapUint32 returns an AlignmentError when passed an
+	// unaligned address.
+	data := new(struct{ val uint64 })
+	addr := uintptr(unsafe.Pointer(&data.val)) + 1
+	want := AlignmentError{Addr: addr, Alignment: 4}
+	if _, err := CompareAndSwapUint32(unsafe.Pointer(addr), 0, 1); err != want {
+		t.Errorf("Unexpected error: got %v, want %v", err, want)
+	}
+}
+
+// withSegvErrorTestMapping calls fn with a two-page mapping. The first page
+// contains random data, and the second page generates SIGSEGV when accessed.
+func withSegvErrorTestMapping(t *testing.T, fn func(m []byte)) {
+	mapping, err := syscall.Mmap(-1, 0, 2*pageSize, syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_ANONYMOUS|syscall.MAP_PRIVATE)
+	if err != nil {
+		t.Fatalf("Mmap failed: %v", err)
+	}
+	defer syscall.Munmap(mapping)
+	if err := syscall.Mprotect(mapping[pageSize:], syscall.PROT_NONE); err != nil {
+		t.Fatalf("Mprotect failed: %v", err)
+	}
+	initRandom(mapping[:pageSize])
+
+	fn(mapping)
+}
+
+// withBusErrorTestMapping calls fn with a two-page mapping. The first page
+// contains random data, and the second page generates SIGBUS when accessed.
+func withBusErrorTestMapping(t *testing.T, fn func(m []byte)) {
+	f, err := ioutil.TempFile("", "sigbus_test")
+	if err != nil {
+		t.Fatalf("TempFile failed: %v", err)
+	}
+	defer f.Close()
+	if err := f.Truncate(pageSize); err != nil {
+		t.Fatalf("Truncate failed: %v", err)
+	}
+	mapping, err := syscall.Mmap(int(f.Fd()), 0, 2*pageSize, syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_SHARED)
+	if err != nil {
+		t.Fatalf("Mmap failed: %v", err)
+	}
+	defer syscall.Munmap(mapping)
+	initRandom(mapping[:pageSize])
+
+	fn(mapping)
+}
+
+func TestCopyInSegvError(t *testing.T) {
+	// Test that CopyIn returns a SegvError when reaching a page that signals
+	// SIGSEGV.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withSegvErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				src := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				dst := randBuf(pageSize)
+				n, err := CopyIn(dst, src)
+				if n != bytesBeforeFault {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (SegvError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := dst[:bytesBeforeFault], mapping[pageSize-bytesBeforeFault:pageSize]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopyInBusError(t *testing.T) {
+	// Test that CopyIn returns a BusError when reaching a page that signals
+	// SIGBUS.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGBUS", bytesBeforeFault), func(t *testing.T) {
+			withBusErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				src := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				dst := randBuf(pageSize)
+				n, err := CopyIn(dst, src)
+				if n != bytesBeforeFault {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (BusError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := dst[:bytesBeforeFault], mapping[pageSize-bytesBeforeFault:pageSize]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopyOutSegvError(t *testing.T) {
+	// Test that CopyOut returns a SegvError when reaching a page that signals
+	// SIGSEGV.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withSegvErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				src := randBuf(pageSize)
+				n, err := CopyOut(dst, src)
+				if n != bytesBeforeFault {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (SegvError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], src[:bytesBeforeFault]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopyOutBusError(t *testing.T) {
+	// Test that CopyOut returns a BusError when reaching a page that signals
+	// SIGBUS.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withBusErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				src := randBuf(pageSize)
+				n, err := CopyOut(dst, src)
+				if n != bytesBeforeFault {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (BusError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], src[:bytesBeforeFault]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopySourceSegvError(t *testing.T) {
+	// Test that Copy returns a SegvError when copying from a page that signals
+	// SIGSEGV.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withSegvErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				src := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				dst := randBuf(pageSize)
+				n, err := Copy(unsafe.Pointer(&dst[0]), src, pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (SegvError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := dst[:bytesBeforeFault], mapping[pageSize-bytesBeforeFault:pageSize]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopySourceBusError(t *testing.T) {
+	// Test that Copy returns a BusError when copying from a page that signals
+	// SIGBUS.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGBUS", bytesBeforeFault), func(t *testing.T) {
+			withBusErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				src := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				dst := randBuf(pageSize)
+				n, err := Copy(unsafe.Pointer(&dst[0]), src, pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (BusError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := dst[:bytesBeforeFault], mapping[pageSize-bytesBeforeFault:pageSize]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopyDestinationSegvError(t *testing.T) {
+	// Test that Copy returns a SegvError when copying to a page that signals
+	// SIGSEGV.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withSegvErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				src := randBuf(pageSize)
+				n, err := Copy(dst, unsafe.Pointer(&src[0]), pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (SegvError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], src[:bytesBeforeFault]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestCopyDestinationBusError(t *testing.T) {
+	// Test that Copy returns a BusError when copying to a page that signals
+	// SIGBUS.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting copy %d bytes before SIGBUS", bytesBeforeFault), func(t *testing.T) {
+			withBusErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				src := randBuf(pageSize)
+				n, err := Copy(dst, unsafe.Pointer(&src[0]), pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected copy length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (BusError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], src[:bytesBeforeFault]; !bytes.Equal(got, want) {
+					t.Errorf("Buffers are not equal when they should be: %v %v", got, want)
+				}
+			})
+		})
+	}
+}
+
+func TestZeroOutSegvError(t *testing.T) {
+	// Test that ZeroOut returns a SegvError when reaching a page that signals
+	// SIGSEGV.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting write %d bytes before SIGSEGV", bytesBeforeFault), func(t *testing.T) {
+			withSegvErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				n, err := ZeroOut(dst, pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected write length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (SegvError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], make([]byte, bytesBeforeFault); !bytes.Equal(got, want) {
+					t.Errorf("Non-zero bytes in written part of mapping: %v", got)
+				}
+			})
+		})
+	}
+}
+
+func TestZeroOutBusError(t *testing.T) {
+	// Test that ZeroOut returns a BusError when reaching a page that signals
+	// SIGBUS.
+	for bytesBeforeFault := 0; bytesBeforeFault <= 2*maxRegisterSize; bytesBeforeFault++ {
+		t.Run(fmt.Sprintf("starting write %d bytes before SIGBUS", bytesBeforeFault), func(t *testing.T) {
+			withBusErrorTestMapping(t, func(mapping []byte) {
+				secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+				dst := unsafe.Pointer(secondPage - uintptr(bytesBeforeFault))
+				n, err := ZeroOut(dst, pageSize)
+				if n != uintptr(bytesBeforeFault) {
+					t.Errorf("Unexpected write length: got %v, want %v", n, bytesBeforeFault)
+				}
+				if want := (BusError{secondPage}); err != want {
+					t.Errorf("Unexpected error: got %v, want %v", err, want)
+				}
+				if got, want := mapping[pageSize-bytesBeforeFault:pageSize], make([]byte, bytesBeforeFault); !bytes.Equal(got, want) {
+					t.Errorf("Non-zero bytes in written part of mapping: %v", got)
+				}
+			})
+		})
+	}
+}
+
+func TestSwapUint32SegvError(t *testing.T) {
+	// Test that SwapUint32 returns a SegvError when reaching a page that
+	// signals SIGSEGV.
+	withSegvErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := SwapUint32(unsafe.Pointer(secondPage), 1)
+		if want := (SegvError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func TestSwapUint32BusError(t *testing.T) {
+	// Test that SwapUint32 returns a BusError when reaching a page that
+	// signals SIGBUS.
+	withBusErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := SwapUint32(unsafe.Pointer(secondPage), 1)
+		if want := (BusError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func TestSwapUint64SegvError(t *testing.T) {
+	// Test that SwapUint64 returns a SegvError when reaching a page that
+	// signals SIGSEGV.
+	withSegvErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := SwapUint64(unsafe.Pointer(secondPage), 1)
+		if want := (SegvError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func TestSwapUint64BusError(t *testing.T) {
+	// Test that SwapUint64 returns a BusError when reaching a page that
+	// signals SIGBUS.
+	withBusErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := SwapUint64(unsafe.Pointer(secondPage), 1)
+		if want := (BusError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func TestCompareAndSwapUint32SegvError(t *testing.T) {
+	// Test that CompareAndSwapUint32 returns a SegvError when reaching a page
+	// that signals SIGSEGV.
+	withSegvErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := CompareAndSwapUint32(unsafe.Pointer(secondPage), 0, 1)
+		if want := (SegvError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func TestCompareAndSwapUint32BusError(t *testing.T) {
+	// Test that CompareAndSwapUint32 returns a BusError when reaching a page
+	// that signals SIGBUS.
+	withBusErrorTestMapping(t, func(mapping []byte) {
+		secondPage := uintptr(unsafe.Pointer(&mapping[0])) + pageSize
+		_, err := CompareAndSwapUint32(unsafe.Pointer(secondPage), 0, 1)
+		if want := (BusError{secondPage}); err != want {
+			t.Errorf("Unexpected error: got %v, want %v", err, want)
+		}
+	})
+}
+
+func testCopy(dst, src []byte) (panicked bool) {
+	defer func() {
+		if r := recover(); r != nil {
+			panicked = true
+		}
+	}()
+	debug.SetPanicOnFault(true)
+	copy(dst, src)
+	return
+}
+
+func TestSegVOnMemmove(t *testing.T) {
+	// Test that SIGSEGVs received by runtime.memmove when *not* doing
+	// CopyIn or CopyOut work gets propagated to the runtime.
+	const bufLen = pageSize
+	a, err := syscall.Mmap(-1, 0, bufLen, syscall.PROT_NONE, syscall.MAP_ANON|syscall.MAP_PRIVATE)
+	if err != nil {
+		t.Fatalf("Mmap failed: %v", err)
+
+	}
+	defer syscall.Munmap(a)
+	b := randBuf(bufLen)
+
+	if !testCopy(b, a) {
+		t.Fatalf("testCopy didn't panic when it should have")
+	}
+
+	if !testCopy(a, b) {
+		t.Fatalf("testCopy didn't panic when it should have")
+	}
+}
+
+func TestSigbusOnMemmove(t *testing.T) {
+	// Test that SIGBUS received by runtime.memmove when *not* doing
+	// CopyIn or CopyOut work gets propagated to the runtime.
+	const bufLen = pageSize
+	f, err := ioutil.TempFile("", "sigbus_test")
+	if err != nil {
+		t.Fatalf("TempFile failed: %v", err)
+	}
+	os.Remove(f.Name())
+	defer f.Close()
+
+	a, err := syscall.Mmap(int(f.Fd()), 0, bufLen, syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_SHARED)
+	if err != nil {
+		t.Fatalf("Mmap failed: %v", err)
+
+	}
+	defer syscall.Munmap(a)
+	b := randBuf(bufLen)
+
+	if !testCopy(b, a) {
+		t.Fatalf("testCopy didn't panic when it should have")
+	}
+
+	if !testCopy(a, b) {
+		t.Fatalf("testCopy didn't panic when it should have")
+	}
+}
diff --git a/pkg/sentry/platform/safecopy/safecopy_unsafe.go b/pkg/sentry/platform/safecopy/safecopy_unsafe.go
new file mode 100644
index 000000000..72f243f8d
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/safecopy_unsafe.go
@@ -0,0 +1,315 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safecopy
+
+import (
+	"fmt"
+	"syscall"
+	"unsafe"
+)
+
+// maxRegisterSize is the maximum register size used in memcpy and memclr. It
+// is used to decide by how much to rewind the copy (for memcpy) or zeroing
+// (for memclr) before proceeding.
+const maxRegisterSize = 16
+
+// memcpy copies data from src to dst. If a SIGSEGV or SIGBUS signal is received
+// during the copy, it returns the address that caused the fault and the number
+// of the signal that was received. Otherwise, it returns an unspecified address
+// and a signal number of 0.
+//
+// Data is copied in order, such that if a fault happens at address p, it is
+// safe to assume that all data before p-maxRegisterSize has already been
+// successfully copied.
+//
+//go:noescape
+func memcpy(dst, src unsafe.Pointer, n uintptr) (fault unsafe.Pointer, sig int32)
+
+// memclr sets the n bytes following ptr to zeroes. If a SIGSEGV or SIGBUS
+// signal is received during the write, it returns the address that caused the
+// fault and the number of the signal that was received. Otherwise, it returns
+// an unspecified address and a signal number of 0.
+//
+// Data is written in order, such that if a fault happens at address p, it is
+// safe to assume that all data before p-maxRegisterSize has already been
+// successfully written.
+//
+//go:noescape
+func memclr(ptr unsafe.Pointer, n uintptr) (fault unsafe.Pointer, sig int32)
+
+// swapUint32 atomically stores new into *ptr and returns (the previous *ptr
+// value, 0). If a SIGSEGV or SIGBUS signal is received during the swap, the
+// value of old is unspecified, and sig is the number of the signal that was
+// received.
+//
+// Preconditions: ptr must be aligned to a 4-byte boundary.
+//
+//go:noescape
+func swapUint32(ptr unsafe.Pointer, new uint32) (old uint32, sig int32)
+
+// swapUint64 atomically stores new into *ptr and returns (the previous *ptr
+// value, 0). If a SIGSEGV or SIGBUS signal is received during the swap, the
+// value of old is unspecified, and sig is the number of the signal that was
+// received.
+//
+// Preconditions: ptr must be aligned to a 8-byte boundary.
+//
+//go:noescape
+func swapUint64(ptr unsafe.Pointer, new uint64) (old uint64, sig int32)
+
+// compareAndSwapUint32 is like sync/atomic.CompareAndSwapUint32, but returns
+// (the value previously stored at ptr, 0). If a SIGSEGV or SIGBUS signal is
+// received during the operation, the value of prev is unspecified, and sig is
+// the number of the signal that was received.
+//
+// Preconditions: ptr must be aligned to a 4-byte boundary.
+//
+//go:noescape
+func compareAndSwapUint32(ptr unsafe.Pointer, old, new uint32) (prev uint32, sig int32)
+
+// CopyIn copies len(dst) bytes from src to dst. It returns the number of bytes
+// copied and an error if SIGSEGV or SIGBUS is received while reading from src.
+func CopyIn(dst []byte, src unsafe.Pointer) (int, error) {
+	toCopy := uintptr(len(dst))
+	if len(dst) == 0 {
+		return 0, nil
+	}
+
+	fault, sig := memcpy(unsafe.Pointer(&dst[0]), src, toCopy)
+	if sig == 0 {
+		return len(dst), nil
+	}
+
+	if faultN, srcN := uintptr(fault), uintptr(src); faultN < srcN && faultN >= srcN+toCopy {
+		panic(fmt.Sprintf("CopyIn faulted at %#x, which is outside source [%#x, %#x)", faultN, srcN, srcN+toCopy))
+	}
+
+	// memcpy might have ended the copy up to maxRegisterSize bytes before
+	// fault, if an instruction caused a memory access that straddled two
+	// pages, and the second one faulted. Try to copy up to the fault.
+	faultN, srcN := uintptr(fault), uintptr(src)
+	var done int
+	if faultN-srcN > maxRegisterSize {
+		done = int(faultN - srcN - maxRegisterSize)
+	}
+	n, err := CopyIn(dst[done:int(faultN-srcN)], unsafe.Pointer(srcN+uintptr(done)))
+	done += n
+	if err != nil {
+		return done, err
+	}
+	return done, errorFromFaultSignal(fault, sig)
+}
+
+// CopyOut copies len(src) bytes from src to dst. If returns the number of
+// bytes done and an error if SIGSEGV or SIGBUS is received while writing to
+// dst.
+func CopyOut(dst unsafe.Pointer, src []byte) (int, error) {
+	toCopy := uintptr(len(src))
+	if toCopy == 0 {
+		return 0, nil
+	}
+
+	fault, sig := memcpy(dst, unsafe.Pointer(&src[0]), toCopy)
+	if sig == 0 {
+		return len(src), nil
+	}
+
+	if faultN, dstN := uintptr(fault), uintptr(dst); faultN < dstN && faultN >= dstN+toCopy {
+		panic(fmt.Sprintf("CopyOut faulted at %#x, which is outside destination [%#x, %#x)", faultN, dstN, dstN+toCopy))
+	}
+
+	// memcpy might have ended the copy up to maxRegisterSize bytes before
+	// fault, if an instruction caused a memory access that straddled two
+	// pages, and the second one faulted. Try to copy up to the fault.
+	faultN, dstN := uintptr(fault), uintptr(dst)
+	var done int
+	if faultN-dstN > maxRegisterSize {
+		done = int(faultN - dstN - maxRegisterSize)
+	}
+	n, err := CopyOut(unsafe.Pointer(dstN+uintptr(done)), src[done:int(faultN-dstN)])
+	done += n
+	if err != nil {
+		return done, err
+	}
+	return done, errorFromFaultSignal(fault, sig)
+}
+
+// Copy copies toCopy bytes from src to dst. It returns the number of bytes
+// copied and an error if SIGSEGV or SIGBUS is received while reading from src
+// or writing to dst.
+//
+// Data is copied in order; if [src, src+toCopy) and [dst, dst+toCopy) overlap,
+// the resulting contents of dst are unspecified.
+func Copy(dst, src unsafe.Pointer, toCopy uintptr) (uintptr, error) {
+	if toCopy == 0 {
+		return 0, nil
+	}
+
+	fault, sig := memcpy(dst, src, toCopy)
+	if sig == 0 {
+		return toCopy, nil
+	}
+
+	// Did the fault occur while reading from src or writing to dst?
+	faultN, srcN, dstN := uintptr(fault), uintptr(src), uintptr(dst)
+	faultAfterSrc := ^uintptr(0)
+	if faultN >= srcN {
+		faultAfterSrc = faultN - srcN
+	}
+	faultAfterDst := ^uintptr(0)
+	if faultN >= dstN {
+		faultAfterDst = faultN - dstN
+	}
+	if faultAfterSrc >= toCopy && faultAfterDst >= toCopy {
+		panic(fmt.Sprintf("Copy faulted at %#x, which is outside source [%#x, %#x) and destination [%#x, %#x)", faultN, srcN, srcN+toCopy, dstN, dstN+toCopy))
+	}
+	faultedAfter := faultAfterSrc
+	if faultedAfter > faultAfterDst {
+		faultedAfter = faultAfterDst
+	}
+
+	// memcpy might have ended the copy up to maxRegisterSize bytes before
+	// fault, if an instruction caused a memory access that straddled two
+	// pages, and the second one faulted. Try to copy up to the fault.
+	var done uintptr
+	if faultedAfter > maxRegisterSize {
+		done = faultedAfter - maxRegisterSize
+	}
+	n, err := Copy(unsafe.Pointer(dstN+done), unsafe.Pointer(srcN+done), faultedAfter-done)
+	done += n
+	if err != nil {
+		return done, err
+	}
+	return done, errorFromFaultSignal(fault, sig)
+}
+
+// ZeroOut writes toZero zero bytes to dst. It returns the number of bytes
+// written and an error if SIGSEGV or SIGBUS is received while writing to dst.
+func ZeroOut(dst unsafe.Pointer, toZero uintptr) (uintptr, error) {
+	if toZero == 0 {
+		return 0, nil
+	}
+
+	fault, sig := memclr(dst, toZero)
+	if sig == 0 {
+		return toZero, nil
+	}
+
+	if faultN, dstN := uintptr(fault), uintptr(dst); faultN < dstN && faultN >= dstN+toZero {
+		panic(fmt.Sprintf("ZeroOut faulted at %#x, which is outside destination [%#x, %#x)", faultN, dstN, dstN+toZero))
+	}
+
+	// memclr might have ended the write up to maxRegisterSize bytes before
+	// fault, if an instruction caused a memory access that straddled two
+	// pages, and the second one faulted. Try to write up to the fault.
+	faultN, dstN := uintptr(fault), uintptr(dst)
+	var done uintptr
+	if faultN-dstN > maxRegisterSize {
+		done = faultN - dstN - maxRegisterSize
+	}
+	n, err := ZeroOut(unsafe.Pointer(dstN+done), faultN-dstN-done)
+	done += n
+	if err != nil {
+		return done, err
+	}
+	return done, errorFromFaultSignal(fault, sig)
+}
+
+// SwapUint32 is equivalent to sync/atomic.SwapUint32, except that it returns
+// an error if SIGSEGV or SIGBUS is received while accessing ptr, or if ptr is
+// not aligned to a 4-byte boundary.
+func SwapUint32(ptr unsafe.Pointer, new uint32) (uint32, error) {
+	if addr := uintptr(ptr); addr&3 != 0 {
+		return 0, AlignmentError{addr, 4}
+	}
+	old, sig := swapUint32(ptr, new)
+	return old, errorFromFaultSignal(ptr, sig)
+}
+
+// SwapUint64 is equivalent to sync/atomic.SwapUint64, except that it returns
+// an error if SIGSEGV or SIGBUS is received while accessing ptr, or if ptr is
+// not aligned to an 8-byte boundary.
+func SwapUint64(ptr unsafe.Pointer, new uint64) (uint64, error) {
+	if addr := uintptr(ptr); addr&7 != 0 {
+		return 0, AlignmentError{addr, 8}
+	}
+	old, sig := swapUint64(ptr, new)
+	return old, errorFromFaultSignal(ptr, sig)
+}
+
+// CompareAndSwapUint32 is equivalent to atomicbitops.CompareAndSwapUint32,
+// except that it returns an error if SIGSEGV or SIGBUS is received while
+// accessing ptr, or if ptr is not aligned to a 4-byte boundary.
+func CompareAndSwapUint32(ptr unsafe.Pointer, old, new uint32) (uint32, error) {
+	if addr := uintptr(ptr); addr&3 != 0 {
+		return 0, AlignmentError{addr, 4}
+	}
+	prev, sig := compareAndSwapUint32(ptr, old, new)
+	return prev, errorFromFaultSignal(ptr, sig)
+}
+
+func errorFromFaultSignal(addr unsafe.Pointer, sig int32) error {
+	switch sig {
+	case 0:
+		return nil
+	case int32(syscall.SIGSEGV):
+		return SegvError{uintptr(addr)}
+	case int32(syscall.SIGBUS):
+		return BusError{uintptr(addr)}
+	default:
+		panic(fmt.Sprintf("safecopy got unexpected signal %d at address %#x", sig, addr))
+	}
+}
+
+// ReplaceSignalHandler replaces the existing signal handler for the provided
+// signal with the one that handles faults in safecopy-protected functions.
+//
+// It stores the value of the previously set handler in previous.
+//
+// This function will be called on initialization in order to install safecopy
+// handlers for appropriate signals. These handlers will call the previous
+// handler however, and if this is function is being used externally then the
+// same courtesy is expected.
+func ReplaceSignalHandler(sig syscall.Signal, handler uintptr, previous *uintptr) error {
+	var sa struct {
+		handler  uintptr
+		flags    uint64
+		restorer uintptr
+		mask     uint64
+	}
+	const maskLen = 8
+
+	// Get the existing signal handler information, and save the current
+	// handler. Once we replace it, we will use this pointer to fall back to
+	// it when we receive other signals.
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(sig), 0, uintptr(unsafe.Pointer(&sa)), maskLen, 0, 0); e != 0 {
+		return e
+	}
+
+	// Fail if there isn't a previous handler.
+	if sa.handler == 0 {
+		return fmt.Errorf("previous handler for signal %x isn't set", sig)
+	}
+
+	*previous = sa.handler
+
+	// Install our own handler.
+	sa.handler = handler
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(sig), uintptr(unsafe.Pointer(&sa)), 0, maskLen, 0, 0); e != 0 {
+		return e
+	}
+
+	return nil
+}
diff --git a/pkg/sentry/platform/safecopy/sighandler_amd64.s b/pkg/sentry/platform/safecopy/sighandler_amd64.s
new file mode 100644
index 000000000..a65cb0c26
--- /dev/null
+++ b/pkg/sentry/platform/safecopy/sighandler_amd64.s
@@ -0,0 +1,124 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// The signals handled by sigHandler.
+#define SIGBUS  7
+#define SIGSEGV 11
+
+// Offsets to the registers in context->uc_mcontext.gregs[].
+#define REG_RDI 0x68
+#define REG_RAX 0x90
+#define REG_IP  0xa8
+
+// Offset to the si_addr field of siginfo.
+#define SI_CODE 0x08
+#define SI_ADDR 0x10
+
+// signalHandler is the signal handler for SIGSEGV and SIGBUS signals. It must
+// not be set up as a handler to any other signals.
+//
+// If the instruction causing the signal is within a safecopy-protected
+// function, the signal is handled such that execution resumes in the
+// appropriate fault handling stub with AX containing the faulting address and
+// DI containing the signal number. Otherwise control is transferred to the
+// previously configured signal handler (savedSigSegvHandler or
+// savedSigBusHandler).
+//
+// This function cannot be written in go because it runs whenever a signal is
+// received by the thread (preempting whatever was running), which includes when
+// garbage collector has stopped or isn't expecting any interactions (like
+// barriers).
+//
+// The arguments are the following:
+// DI - The signal number.
+// SI - Pointer to siginfo_t structure.
+// DX - Pointer to ucontext structure.
+TEXT ·signalHandler(SB),NOSPLIT,$0
+	// Check if the signal is from the kernel.
+	MOVQ $0x0, CX
+	CMPL CX, SI_CODE(SI)
+	JGE original_handler
+
+	// Check if RIP is within the area we care about.
+	MOVQ REG_IP(DX), CX
+	CMPQ CX, ·memcpyBegin(SB)
+	JB not_memcpy
+	CMPQ CX, ·memcpyEnd(SB)
+	JAE not_memcpy
+
+	// Modify the context such that execution will resume in the fault
+	// handler.
+	LEAQ handleMemcpyFault(SB), CX
+	JMP handle_fault
+
+not_memcpy:
+	CMPQ CX, ·memclrBegin(SB)
+	JB not_memclr
+	CMPQ CX, ·memclrEnd(SB)
+	JAE not_memclr
+
+	LEAQ handleMemclrFault(SB), CX
+	JMP handle_fault
+
+not_memclr:
+	CMPQ CX, ·swapUint32Begin(SB)
+	JB not_swapuint32
+	CMPQ CX, ·swapUint32End(SB)
+	JAE not_swapuint32
+
+	LEAQ handleSwapUint32Fault(SB), CX
+	JMP handle_fault
+
+not_swapuint32:
+	CMPQ CX, ·swapUint64Begin(SB)
+	JB not_swapuint64
+	CMPQ CX, ·swapUint64End(SB)
+	JAE not_swapuint64
+
+	LEAQ handleSwapUint64Fault(SB), CX
+	JMP handle_fault
+
+not_swapuint64:
+	CMPQ CX, ·compareAndSwapUint32Begin(SB)
+	JB not_casuint32
+	CMPQ CX, ·compareAndSwapUint32End(SB)
+	JAE not_casuint32
+
+	LEAQ handleCompareAndSwapUint32Fault(SB), CX
+	JMP handle_fault
+
+not_casuint32:
+original_handler:
+	// Jump to the previous signal handler, which is likely the golang one.
+	XORQ CX, CX
+	MOVQ ·savedSigBusHandler(SB), AX
+	CMPL DI, $SIGSEGV
+	CMOVQEQ ·savedSigSegVHandler(SB), AX
+	JMP AX
+
+handle_fault:
+	// Entered with the address of the fault handler in RCX; store it in
+	// RIP.
+	MOVQ CX, REG_IP(DX)
+
+	// Store the faulting address in RAX.
+	MOVQ SI_ADDR(SI), CX
+	MOVQ CX, REG_RAX(DX)
+
+	// Store the signal number in EDI.
+	MOVL DI, REG_RDI(DX)
+
+	RET
diff --git a/pkg/sentry/safemem/BUILD b/pkg/sentry/safemem/BUILD
new file mode 100644
index 000000000..dc4cfce41
--- /dev/null
+++ b/pkg/sentry/safemem/BUILD
@@ -0,0 +1,28 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+
+go_library(
+    name = "safemem",
+    srcs = [
+        "block_unsafe.go",
+        "io.go",
+        "safemem.go",
+        "seq_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/safemem",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/platform/safecopy",
+    ],
+)
+
+go_test(
+    name = "safemem_test",
+    size = "small",
+    srcs = [
+        "io_test.go",
+        "seq_test.go",
+    ],
+    embed = [":safemem"],
+)
diff --git a/pkg/sentry/safemem/block_unsafe.go b/pkg/sentry/safemem/block_unsafe.go
new file mode 100644
index 000000000..0b58f6497
--- /dev/null
+++ b/pkg/sentry/safemem/block_unsafe.go
@@ -0,0 +1,269 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safemem
+
+import (
+	"fmt"
+	"reflect"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform/safecopy"
+)
+
+// A Block is a range of contiguous bytes, similar to []byte but with the
+// following differences:
+//
+// - The memory represented by a Block may require the use of safecopy to
+// access.
+//
+// - Block does not carry a capacity and cannot be expanded.
+//
+// Blocks are immutable and may be copied by value. The zero value of Block
+// represents an empty range, analogous to a nil []byte.
+type Block struct {
+	// [start, start+length) is the represented memory.
+	//
+	// start is an unsafe.Pointer to ensure that Block prevents the represented
+	// memory from being garbage-collected.
+	start  unsafe.Pointer
+	length int
+
+	// needSafecopy is true if accessing the represented memory requires the
+	// use of safecopy.
+	needSafecopy bool
+}
+
+// BlockFromSafeSlice returns a Block equivalent to slice, which is safe to
+// access without safecopy.
+func BlockFromSafeSlice(slice []byte) Block {
+	return blockFromSlice(slice, false)
+}
+
+// BlockFromUnsafeSlice returns a Block equivalent to bs, which is not safe to
+// access without safecopy.
+func BlockFromUnsafeSlice(slice []byte) Block {
+	return blockFromSlice(slice, true)
+}
+
+func blockFromSlice(slice []byte, needSafecopy bool) Block {
+	if len(slice) == 0 {
+		return Block{}
+	}
+	return Block{
+		start:        unsafe.Pointer(&slice[0]),
+		length:       len(slice),
+		needSafecopy: needSafecopy,
+	}
+}
+
+// BlockFromSafePointer returns a Block equivalent to [ptr, ptr+len), which is
+// safe to access without safecopy.
+//
+// Preconditions: ptr+len does not overflow.
+func BlockFromSafePointer(ptr unsafe.Pointer, len int) Block {
+	return blockFromPointer(ptr, len, false)
+}
+
+// BlockFromUnsafePointer returns a Block equivalent to [ptr, ptr+len), which
+// is not safe to access without safecopy.
+//
+// Preconditions: ptr+len does not overflow.
+func BlockFromUnsafePointer(ptr unsafe.Pointer, len int) Block {
+	return blockFromPointer(ptr, len, true)
+}
+
+func blockFromPointer(ptr unsafe.Pointer, len int, needSafecopy bool) Block {
+	if uptr := uintptr(ptr); uptr+uintptr(len) < uptr {
+		panic(fmt.Sprintf("ptr %#x + len %#x overflows", ptr, len))
+	}
+	return Block{
+		start:        ptr,
+		length:       len,
+		needSafecopy: needSafecopy,
+	}
+}
+
+// DropFirst returns a Block equivalent to b, but with the first n bytes
+// omitted. It is analogous to the [n:] operation on a slice, except that if n
+// > b.Len(), DropFirst returns an empty Block instead of panicking.
+//
+// Preconditions: n >= 0.
+func (b Block) DropFirst(n int) Block {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return b.DropFirst64(uint64(n))
+}
+
+// DropFirst64 is equivalent to DropFirst but takes a uint64.
+func (b Block) DropFirst64(n uint64) Block {
+	if n >= uint64(b.length) {
+		return Block{}
+	}
+	return Block{
+		start:        unsafe.Pointer(uintptr(b.start) + uintptr(n)),
+		length:       b.length - int(n),
+		needSafecopy: b.needSafecopy,
+	}
+}
+
+// TakeFirst returns a Block equivalent to the first n bytes of b. It is
+// analogous to the [:n] operation on a slice, except that if n > b.Len(),
+// TakeFirst returns a copy of b instead of panicking.
+//
+// Preconditions: n >= 0.
+func (b Block) TakeFirst(n int) Block {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return b.TakeFirst64(uint64(n))
+}
+
+// TakeFirst64 is equivalent to TakeFirst but takes a uint64.
+func (b Block) TakeFirst64(n uint64) Block {
+	if n == 0 {
+		return Block{}
+	}
+	if n >= uint64(b.length) {
+		return b
+	}
+	return Block{
+		start:        b.start,
+		length:       int(n),
+		needSafecopy: b.needSafecopy,
+	}
+}
+
+// ToSlice returns a []byte equivalent to b.
+func (b Block) ToSlice() []byte {
+	var bs []byte
+	hdr := (*reflect.SliceHeader)(unsafe.Pointer(&bs))
+	hdr.Data = uintptr(b.start)
+	hdr.Len = b.length
+	hdr.Cap = b.length
+	return bs
+}
+
+// Addr returns b's start address as a uintptr. It returns uintptr instead of
+// unsafe.Pointer so that code using safemem cannot obtain unsafe.Pointers
+// without importing the unsafe package explicitly.
+//
+// Note that a uintptr is not recognized as a pointer by the garbage collector,
+// such that if there are no uses of b after a call to b.Addr() and the address
+// is to Go-managed memory, the returned uintptr does not prevent garbage
+// collection of the pointee.
+func (b Block) Addr() uintptr {
+	return uintptr(b.start)
+}
+
+// Len returns b's length in bytes.
+func (b Block) Len() int {
+	return b.length
+}
+
+// NeedSafecopy returns true if accessing b.ToSlice() requires the use of safecopy.
+func (b Block) NeedSafecopy() bool {
+	return b.needSafecopy
+}
+
+// String implements fmt.Stringer.String.
+func (b Block) String() string {
+	if uintptr(b.start) == 0 && b.length == 0 {
+		return "<nil>"
+	}
+	var suffix string
+	if b.needSafecopy {
+		suffix = "*"
+	}
+	return fmt.Sprintf("[%#x-%#x)%s", uintptr(b.start), uintptr(b.start)+uintptr(b.length), suffix)
+}
+
+// Copy copies src.Len() or dst.Len() bytes, whichever is less, from src
+// to dst and returns the number of bytes copied.
+//
+// If src and dst overlap, the data stored in dst is unspecified.
+func Copy(dst, src Block) (int, error) {
+	if !dst.needSafecopy && !src.needSafecopy {
+		return copy(dst.ToSlice(), src.ToSlice()), nil
+	}
+
+	n := dst.length
+	if n > src.length {
+		n = src.length
+	}
+	if n == 0 {
+		return 0, nil
+	}
+
+	switch {
+	case dst.needSafecopy && !src.needSafecopy:
+		return safecopy.CopyOut(dst.start, src.TakeFirst(n).ToSlice())
+	case !dst.needSafecopy && src.needSafecopy:
+		return safecopy.CopyIn(dst.TakeFirst(n).ToSlice(), src.start)
+	case dst.needSafecopy && src.needSafecopy:
+		n64, err := safecopy.Copy(dst.start, src.start, uintptr(n))
+		return int(n64), err
+	default:
+		panic("unreachable")
+	}
+}
+
+// Zero sets all bytes in dst to 0 and returns the number of bytes zeroed.
+func Zero(dst Block) (int, error) {
+	if !dst.needSafecopy {
+		bs := dst.ToSlice()
+		for i := range bs {
+			bs[i] = 0
+		}
+		return len(bs), nil
+	}
+
+	n64, err := safecopy.ZeroOut(dst.start, uintptr(dst.length))
+	return int(n64), err
+}
+
+// Safecopy atomics are no slower than non-safecopy atomics, so use the former
+// even when !b.needSafecopy to get consistent alignment checking.
+
+// SwapUint32 invokes safecopy.SwapUint32 on the first 4 bytes of b.
+//
+// Preconditions: b.Len() >= 4.
+func SwapUint32(b Block, new uint32) (uint32, error) {
+	if b.length < 4 {
+		panic(fmt.Sprintf("insufficient length: %d", b.length))
+	}
+	return safecopy.SwapUint32(b.start, new)
+}
+
+// SwapUint64 invokes safecopy.SwapUint64 on the first 8 bytes of b.
+//
+// Preconditions: b.Len() >= 8.
+func SwapUint64(b Block, new uint64) (uint64, error) {
+	if b.length < 8 {
+		panic(fmt.Sprintf("insufficient length: %d", b.length))
+	}
+	return safecopy.SwapUint64(b.start, new)
+}
+
+// CompareAndSwapUint32 invokes safecopy.CompareAndSwapUint32 on the first 4
+// bytes of b.
+//
+// Preconditions: b.Len() >= 4.
+func CompareAndSwapUint32(b Block, old, new uint32) (uint32, error) {
+	if b.length < 4 {
+		panic(fmt.Sprintf("insufficient length: %d", b.length))
+	}
+	return safecopy.CompareAndSwapUint32(b.start, old, new)
+}
diff --git a/pkg/sentry/safemem/io.go b/pkg/sentry/safemem/io.go
new file mode 100644
index 000000000..fd917648b
--- /dev/null
+++ b/pkg/sentry/safemem/io.go
@@ -0,0 +1,339 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safemem
+
+import (
+	"errors"
+	"io"
+	"math"
+)
+
+// ErrEndOfBlockSeq is returned by BlockSeqWriter when attempting to write
+// beyond the end of the BlockSeq.
+var ErrEndOfBlockSeq = errors.New("write beyond end of BlockSeq")
+
+// Reader represents a streaming byte source like io.Reader.
+type Reader interface {
+	// ReadToBlocks reads up to dsts.NumBytes() bytes into dsts and returns the
+	// number of bytes read. It may return a partial read without an error
+	// (i.e. (n, nil) where 0 < n < dsts.NumBytes()). It should not return a
+	// full read with an error (i.e. (dsts.NumBytes(), err) where err != nil);
+	// note that this differs from io.Reader.Read (in particular, io.EOF should
+	// not be returned if ReadToBlocks successfully reads dsts.NumBytes()
+	// bytes.)
+	ReadToBlocks(dsts BlockSeq) (uint64, error)
+}
+
+// Writer represents a streaming byte sink like io.Writer.
+type Writer interface {
+	// WriteFromBlocks writes up to srcs.NumBytes() bytes from srcs and returns
+	// the number of bytes written. It may return a partial write without an
+	// error (i.e. (n, nil) where 0 < n < srcs.NumBytes()). It should not
+	// return a full write with an error (i.e. srcs.NumBytes(), err) where err
+	// != nil).
+	WriteFromBlocks(srcs BlockSeq) (uint64, error)
+}
+
+// ReadFullToBlocks repeatedly invokes r.ReadToBlocks until dsts.NumBytes()
+// bytes have been read or ReadToBlocks returns an error.
+func ReadFullToBlocks(r Reader, dsts BlockSeq) (uint64, error) {
+	var done uint64
+	for !dsts.IsEmpty() {
+		n, err := r.ReadToBlocks(dsts)
+		done += n
+		if err != nil {
+			return done, err
+		}
+		dsts = dsts.DropFirst64(n)
+	}
+	return done, nil
+}
+
+// WriteFullFromBlocks repeatedly invokes w.WriteFromBlocks until
+// srcs.NumBytes() bytes have been written or WriteFromBlocks returns an error.
+func WriteFullFromBlocks(w Writer, srcs BlockSeq) (uint64, error) {
+	var done uint64
+	for !srcs.IsEmpty() {
+		n, err := w.WriteFromBlocks(srcs)
+		done += n
+		if err != nil {
+			return done, err
+		}
+		srcs = srcs.DropFirst64(n)
+	}
+	return done, nil
+}
+
+// BlockSeqReader implements Reader by reading from a BlockSeq.
+type BlockSeqReader struct {
+	Blocks BlockSeq
+}
+
+// ReadToBlocks implements Reader.ReadToBlocks.
+func (r *BlockSeqReader) ReadToBlocks(dsts BlockSeq) (uint64, error) {
+	n, err := CopySeq(dsts, r.Blocks)
+	r.Blocks = r.Blocks.DropFirst64(n)
+	if err != nil {
+		return n, err
+	}
+	if n < dsts.NumBytes() {
+		return n, io.EOF
+	}
+	return n, nil
+}
+
+// BlockSeqWriter implements Writer by writing to a BlockSeq.
+type BlockSeqWriter struct {
+	Blocks BlockSeq
+}
+
+// WriteFromBlocks implements Writer.WriteFromBlocks.
+func (w *BlockSeqWriter) WriteFromBlocks(srcs BlockSeq) (uint64, error) {
+	n, err := CopySeq(w.Blocks, srcs)
+	w.Blocks = w.Blocks.DropFirst64(n)
+	if err != nil {
+		return n, err
+	}
+	if n < srcs.NumBytes() {
+		return n, ErrEndOfBlockSeq
+	}
+	return n, nil
+}
+
+// ReaderFunc implements Reader for a function with the semantics of
+// Reader.ReadToBlocks.
+type ReaderFunc func(dsts BlockSeq) (uint64, error)
+
+// ReadToBlocks implements Reader.ReadToBlocks.
+func (f ReaderFunc) ReadToBlocks(dsts BlockSeq) (uint64, error) {
+	return f(dsts)
+}
+
+// WriterFunc implements Writer for a function with the semantics of
+// Writer.WriteFromBlocks.
+type WriterFunc func(srcs BlockSeq) (uint64, error)
+
+// WriteFromBlocks implements Writer.WriteFromBlocks.
+func (f WriterFunc) WriteFromBlocks(srcs BlockSeq) (uint64, error) {
+	return f(srcs)
+}
+
+// ToIOReader implements io.Reader for a (safemem.)Reader.
+//
+// ToIOReader will return a successful partial read iff Reader.ReadToBlocks does
+// so.
+type ToIOReader struct {
+	Reader Reader
+}
+
+// Read implements io.Reader.Read.
+func (r ToIOReader) Read(dst []byte) (int, error) {
+	n, err := r.Reader.ReadToBlocks(BlockSeqOf(BlockFromSafeSlice(dst)))
+	return int(n), err
+}
+
+// ToIOWriter implements io.Writer for a (safemem.)Writer.
+type ToIOWriter struct {
+	Writer Writer
+}
+
+// Write implements io.Writer.Write.
+func (w ToIOWriter) Write(src []byte) (int, error) {
+	// io.Writer does not permit partial writes.
+	n, err := WriteFullFromBlocks(w.Writer, BlockSeqOf(BlockFromSafeSlice(src)))
+	return int(n), err
+}
+
+// FromIOReader implements Reader for an io.Reader by repeatedly invoking
+// io.Reader.Read until it returns an error or partial read.
+//
+// FromIOReader will return a successful partial read iff Reader.Read does so.
+type FromIOReader struct {
+	Reader io.Reader
+}
+
+// ReadToBlocks implements Reader.ReadToBlocks.
+func (r FromIOReader) ReadToBlocks(dsts BlockSeq) (uint64, error) {
+	var buf []byte
+	var done uint64
+	for !dsts.IsEmpty() {
+		dst := dsts.Head()
+		var n int
+		var err error
+		n, buf, err = r.readToBlock(dst, buf)
+		done += uint64(n)
+		if n != dst.Len() {
+			return done, err
+		}
+		dsts = dsts.Tail()
+		if err != nil {
+			if dsts.IsEmpty() && err == io.EOF {
+				return done, nil
+			}
+			return done, err
+		}
+	}
+	return done, nil
+}
+
+func (r FromIOReader) readToBlock(dst Block, buf []byte) (int, []byte, error) {
+	// io.Reader isn't safecopy-aware, so we have to buffer Blocks that require
+	// safecopy.
+	if !dst.NeedSafecopy() {
+		n, err := r.Reader.Read(dst.ToSlice())
+		return n, buf, err
+	}
+	if len(buf) < dst.Len() {
+		buf = make([]byte, dst.Len())
+	}
+	rn, rerr := r.Reader.Read(buf[:dst.Len()])
+	wbn, wberr := Copy(dst, BlockFromSafeSlice(buf[:rn]))
+	if wberr != nil {
+		return wbn, buf, wberr
+	}
+	return wbn, buf, rerr
+}
+
+// FromIOWriter implements Writer for an io.Writer by repeatedly invoking
+// io.Writer.Write until it returns an error or partial write.
+//
+// FromIOWriter will tolerate implementations of io.Writer.Write that return
+// partial writes with a nil error in contravention of io.Writer's
+// requirements, since Writer is permitted to do so. FromIOWriter will return a
+// successful partial write iff Writer.Write does so.
+type FromIOWriter struct {
+	Writer io.Writer
+}
+
+// WriteFromBlocks implements Writer.WriteFromBlocks.
+func (w FromIOWriter) WriteFromBlocks(srcs BlockSeq) (uint64, error) {
+	var buf []byte
+	var done uint64
+	for !srcs.IsEmpty() {
+		src := srcs.Head()
+		var n int
+		var err error
+		n, buf, err = w.writeFromBlock(src, buf)
+		done += uint64(n)
+		if n != src.Len() || err != nil {
+			return done, err
+		}
+		srcs = srcs.Tail()
+	}
+	return done, nil
+}
+
+func (w FromIOWriter) writeFromBlock(src Block, buf []byte) (int, []byte, error) {
+	// io.Writer isn't safecopy-aware, so we have to buffer Blocks that require
+	// safecopy.
+	if !src.NeedSafecopy() {
+		n, err := w.Writer.Write(src.ToSlice())
+		return n, buf, err
+	}
+	if len(buf) < src.Len() {
+		buf = make([]byte, src.Len())
+	}
+	bufn, buferr := Copy(BlockFromSafeSlice(buf[:src.Len()]), src)
+	wn, werr := w.Writer.Write(buf[:bufn])
+	if werr != nil {
+		return wn, buf, werr
+	}
+	return wn, buf, buferr
+}
+
+// FromVecReaderFunc implements Reader for a function that reads data into a
+// [][]byte and returns the number of bytes read as an int64.
+type FromVecReaderFunc struct {
+	ReadVec func(dsts [][]byte) (int64, error)
+}
+
+// ReadToBlocks implements Reader.ReadToBlocks.
+//
+// ReadToBlocks calls r.ReadVec at most once.
+func (r FromVecReaderFunc) ReadToBlocks(dsts BlockSeq) (uint64, error) {
+	if dsts.IsEmpty() {
+		return 0, nil
+	}
+	// Ensure that we don't pass a [][]byte with a total length > MaxInt64.
+	dsts = dsts.TakeFirst64(uint64(math.MaxInt64))
+	dstSlices := make([][]byte, 0, dsts.NumBlocks())
+	// Buffer Blocks that require safecopy.
+	for tmp := dsts; !tmp.IsEmpty(); tmp = tmp.Tail() {
+		dst := tmp.Head()
+		if dst.NeedSafecopy() {
+			dstSlices = append(dstSlices, make([]byte, dst.Len()))
+		} else {
+			dstSlices = append(dstSlices, dst.ToSlice())
+		}
+	}
+	rn, rerr := r.ReadVec(dstSlices)
+	dsts = dsts.TakeFirst64(uint64(rn))
+	var done uint64
+	var i int
+	for !dsts.IsEmpty() {
+		dst := dsts.Head()
+		if dst.NeedSafecopy() {
+			n, err := Copy(dst, BlockFromSafeSlice(dstSlices[i]))
+			done += uint64(n)
+			if err != nil {
+				return done, err
+			}
+		} else {
+			done += uint64(dst.Len())
+		}
+		dsts = dsts.Tail()
+		i++
+	}
+	return done, rerr
+}
+
+// FromVecWriterFunc implements Writer for a function that writes data from a
+// [][]byte and returns the number of bytes written.
+type FromVecWriterFunc struct {
+	WriteVec func(srcs [][]byte) (int64, error)
+}
+
+// WriteFromBlocks implements Writer.WriteFromBlocks.
+//
+// WriteFromBlocks calls w.WriteVec at most once.
+func (w FromVecWriterFunc) WriteFromBlocks(srcs BlockSeq) (uint64, error) {
+	if srcs.IsEmpty() {
+		return 0, nil
+	}
+	// Ensure that we don't pass a [][]byte with a total length > MaxInt64.
+	srcs = srcs.TakeFirst64(uint64(math.MaxInt64))
+	srcSlices := make([][]byte, 0, srcs.NumBlocks())
+	// Buffer Blocks that require safecopy.
+	var buferr error
+	for tmp := srcs; !tmp.IsEmpty(); tmp = tmp.Tail() {
+		src := tmp.Head()
+		if src.NeedSafecopy() {
+			slice := make([]byte, src.Len())
+			n, err := Copy(BlockFromSafeSlice(slice), src)
+			srcSlices = append(srcSlices, slice[:n])
+			if err != nil {
+				buferr = err
+				break
+			}
+		} else {
+			srcSlices = append(srcSlices, src.ToSlice())
+		}
+	}
+	n, err := w.WriteVec(srcSlices)
+	if err != nil {
+		return uint64(n), err
+	}
+	return uint64(n), buferr
+}
diff --git a/pkg/sentry/safemem/io_test.go b/pkg/sentry/safemem/io_test.go
new file mode 100644
index 000000000..edac4c1d7
--- /dev/null
+++ b/pkg/sentry/safemem/io_test.go
@@ -0,0 +1,199 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safemem
+
+import (
+	"bytes"
+	"io"
+	"testing"
+)
+
+func makeBlocks(slices ...[]byte) []Block {
+	blocks := make([]Block, 0, len(slices))
+	for _, s := range slices {
+		blocks = append(blocks, BlockFromSafeSlice(s))
+	}
+	return blocks
+}
+
+func TestFromIOReaderFullRead(t *testing.T) {
+	r := FromIOReader{bytes.NewBufferString("foobar")}
+	dsts := makeBlocks(make([]byte, 3), make([]byte, 3))
+	n, err := r.ReadToBlocks(BlockSeqFromSlice(dsts))
+	if wantN := uint64(6); n != wantN || err != nil {
+		t.Errorf("ReadToBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	for i, want := range [][]byte{[]byte("foo"), []byte("bar")} {
+		if got := dsts[i].ToSlice(); !bytes.Equal(got, want) {
+			t.Errorf("dsts[%d]: got %q, wanted %q", i, got, want)
+		}
+	}
+}
+
+type eofHidingReader struct {
+	Reader io.Reader
+}
+
+func (r eofHidingReader) Read(dst []byte) (int, error) {
+	n, err := r.Reader.Read(dst)
+	if err == io.EOF {
+		return n, nil
+	}
+	return n, err
+}
+
+func TestFromIOReaderPartialRead(t *testing.T) {
+	r := FromIOReader{eofHidingReader{bytes.NewBufferString("foob")}}
+	dsts := makeBlocks(make([]byte, 3), make([]byte, 3))
+	n, err := r.ReadToBlocks(BlockSeqFromSlice(dsts))
+	// FromIOReader should stop after the eofHidingReader returns (1, nil)
+	// for a 3-byte read.
+	if wantN := uint64(4); n != wantN || err != nil {
+		t.Errorf("ReadToBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	for i, want := range [][]byte{[]byte("foo"), []byte("b\x00\x00")} {
+		if got := dsts[i].ToSlice(); !bytes.Equal(got, want) {
+			t.Errorf("dsts[%d]: got %q, wanted %q", i, got, want)
+		}
+	}
+}
+
+type singleByteReader struct {
+	Reader io.Reader
+}
+
+func (r singleByteReader) Read(dst []byte) (int, error) {
+	if len(dst) == 0 {
+		return r.Reader.Read(dst)
+	}
+	return r.Reader.Read(dst[:1])
+}
+
+func TestSingleByteReader(t *testing.T) {
+	r := FromIOReader{singleByteReader{bytes.NewBufferString("foobar")}}
+	dsts := makeBlocks(make([]byte, 3), make([]byte, 3))
+	n, err := r.ReadToBlocks(BlockSeqFromSlice(dsts))
+	// FromIOReader should stop after the singleByteReader returns (1, nil)
+	// for a 3-byte read.
+	if wantN := uint64(1); n != wantN || err != nil {
+		t.Errorf("ReadToBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	for i, want := range [][]byte{[]byte("f\x00\x00"), []byte("\x00\x00\x00")} {
+		if got := dsts[i].ToSlice(); !bytes.Equal(got, want) {
+			t.Errorf("dsts[%d]: got %q, wanted %q", i, got, want)
+		}
+	}
+}
+
+func TestReadFullToBlocks(t *testing.T) {
+	r := FromIOReader{singleByteReader{bytes.NewBufferString("foobar")}}
+	dsts := makeBlocks(make([]byte, 3), make([]byte, 3))
+	n, err := ReadFullToBlocks(r, BlockSeqFromSlice(dsts))
+	// ReadFullToBlocks should call into FromIOReader => singleByteReader
+	// repeatedly until dsts is exhausted.
+	if wantN := uint64(6); n != wantN || err != nil {
+		t.Errorf("ReadFullToBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	for i, want := range [][]byte{[]byte("foo"), []byte("bar")} {
+		if got := dsts[i].ToSlice(); !bytes.Equal(got, want) {
+			t.Errorf("dsts[%d]: got %q, wanted %q", i, got, want)
+		}
+	}
+}
+
+func TestFromIOWriterFullWrite(t *testing.T) {
+	srcs := makeBlocks([]byte("foo"), []byte("bar"))
+	var dst bytes.Buffer
+	w := FromIOWriter{&dst}
+	n, err := w.WriteFromBlocks(BlockSeqFromSlice(srcs))
+	if wantN := uint64(6); n != wantN || err != nil {
+		t.Errorf("WriteFromBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst.Bytes(), []byte("foobar"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
+
+type limitedWriter struct {
+	Writer io.Writer
+	Done   int
+	Limit  int
+}
+
+func (w *limitedWriter) Write(src []byte) (int, error) {
+	count := len(src)
+	if count > (w.Limit - w.Done) {
+		count = w.Limit - w.Done
+	}
+	n, err := w.Writer.Write(src[:count])
+	w.Done += n
+	return n, err
+}
+
+func TestFromIOWriterPartialWrite(t *testing.T) {
+	srcs := makeBlocks([]byte("foo"), []byte("bar"))
+	var dst bytes.Buffer
+	w := FromIOWriter{&limitedWriter{&dst, 0, 4}}
+	n, err := w.WriteFromBlocks(BlockSeqFromSlice(srcs))
+	// FromIOWriter should stop after the limitedWriter returns (1, nil) for a
+	// 3-byte write.
+	if wantN := uint64(4); n != wantN || err != nil {
+		t.Errorf("WriteFromBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst.Bytes(), []byte("foob"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
+
+type singleByteWriter struct {
+	Writer io.Writer
+}
+
+func (w singleByteWriter) Write(src []byte) (int, error) {
+	if len(src) == 0 {
+		return w.Writer.Write(src)
+	}
+	return w.Writer.Write(src[:1])
+}
+
+func TestSingleByteWriter(t *testing.T) {
+	srcs := makeBlocks([]byte("foo"), []byte("bar"))
+	var dst bytes.Buffer
+	w := FromIOWriter{singleByteWriter{&dst}}
+	n, err := w.WriteFromBlocks(BlockSeqFromSlice(srcs))
+	// FromIOWriter should stop after the singleByteWriter returns (1, nil)
+	// for a 3-byte write.
+	if wantN := uint64(1); n != wantN || err != nil {
+		t.Errorf("WriteFromBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst.Bytes(), []byte("f"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
+
+func TestWriteFullToBlocks(t *testing.T) {
+	srcs := makeBlocks([]byte("foo"), []byte("bar"))
+	var dst bytes.Buffer
+	w := FromIOWriter{singleByteWriter{&dst}}
+	n, err := WriteFullFromBlocks(w, BlockSeqFromSlice(srcs))
+	// WriteFullToBlocks should call into FromIOWriter => singleByteWriter
+	// repeatedly until srcs is exhausted.
+	if wantN := uint64(6); n != wantN || err != nil {
+		t.Errorf("WriteFullFromBlocks: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst.Bytes(), []byte("foobar"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
diff --git a/pkg/sentry/safemem/safemem.go b/pkg/sentry/safemem/safemem.go
new file mode 100644
index 000000000..2f8002004
--- /dev/null
+++ b/pkg/sentry/safemem/safemem.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package safemem provides the Block and BlockSeq types.
+package safemem
diff --git a/pkg/sentry/safemem/seq_test.go b/pkg/sentry/safemem/seq_test.go
new file mode 100644
index 000000000..3e83b3851
--- /dev/null
+++ b/pkg/sentry/safemem/seq_test.go
@@ -0,0 +1,196 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safemem
+
+import (
+	"bytes"
+	"reflect"
+	"testing"
+)
+
+type blockSeqTest struct {
+	desc string
+
+	pieces     []string
+	haveOffset bool
+	offset     uint64
+	haveLimit  bool
+	limit      uint64
+
+	want string
+}
+
+func (t blockSeqTest) NonEmptyByteSlices() [][]byte {
+	// t is a value, so we can mutate it freely.
+	slices := make([][]byte, 0, len(t.pieces))
+	for _, str := range t.pieces {
+		if t.haveOffset {
+			strOff := t.offset
+			if strOff > uint64(len(str)) {
+				strOff = uint64(len(str))
+			}
+			str = str[strOff:]
+			t.offset -= strOff
+		}
+		if t.haveLimit {
+			strLim := t.limit
+			if strLim > uint64(len(str)) {
+				strLim = uint64(len(str))
+			}
+			str = str[:strLim]
+			t.limit -= strLim
+		}
+		if len(str) != 0 {
+			slices = append(slices, []byte(str))
+		}
+	}
+	return slices
+}
+
+func (t blockSeqTest) BlockSeq() BlockSeq {
+	blocks := make([]Block, 0, len(t.pieces))
+	for _, str := range t.pieces {
+		blocks = append(blocks, BlockFromSafeSlice([]byte(str)))
+	}
+	bs := BlockSeqFromSlice(blocks)
+	if t.haveOffset {
+		bs = bs.DropFirst64(t.offset)
+	}
+	if t.haveLimit {
+		bs = bs.TakeFirst64(t.limit)
+	}
+	return bs
+}
+
+var blockSeqTests = []blockSeqTest{
+	{
+		desc: "Empty sequence",
+	},
+	{
+		desc:   "Sequence of length 1",
+		pieces: []string{"foobar"},
+		want:   "foobar",
+	},
+	{
+		desc:   "Sequence of length 2",
+		pieces: []string{"foo", "bar"},
+		want:   "foobar",
+	},
+	{
+		desc:   "Empty Blocks",
+		pieces: []string{"", "foo", "", "", "bar", ""},
+		want:   "foobar",
+	},
+	{
+		desc:       "Sequence with non-zero offset",
+		pieces:     []string{"foo", "bar"},
+		haveOffset: true,
+		offset:     2,
+		want:       "obar",
+	},
+	{
+		desc:      "Sequence with non-maximal limit",
+		pieces:    []string{"foo", "bar"},
+		haveLimit: true,
+		limit:     5,
+		want:      "fooba",
+	},
+	{
+		desc:       "Sequence with offset and limit",
+		pieces:     []string{"foo", "bar"},
+		haveOffset: true,
+		offset:     2,
+		haveLimit:  true,
+		limit:      3,
+		want:       "oba",
+	},
+}
+
+func TestBlockSeqNumBytes(t *testing.T) {
+	for _, test := range blockSeqTests {
+		t.Run(test.desc, func(t *testing.T) {
+			if got, want := test.BlockSeq().NumBytes(), uint64(len(test.want)); got != want {
+				t.Errorf("NumBytes: got %d, wanted %d", got, want)
+			}
+		})
+	}
+}
+
+func TestBlockSeqIterBlocks(t *testing.T) {
+	// Tests BlockSeq iteration using Head/Tail.
+	for _, test := range blockSeqTests {
+		t.Run(test.desc, func(t *testing.T) {
+			srcs := test.BlockSeq()
+			// "Note that a non-nil empty slice and a nil slice ... are not
+			// deeply equal." - reflect
+			slices := make([][]byte, 0, 0)
+			for !srcs.IsEmpty() {
+				src := srcs.Head()
+				slices = append(slices, src.ToSlice())
+				nextSrcs := srcs.Tail()
+				if got, want := nextSrcs.NumBytes(), srcs.NumBytes()-uint64(src.Len()); got != want {
+					t.Fatalf("%v.Tail(): got %v (%d bytes), wanted %d bytes", srcs, nextSrcs, got, want)
+				}
+				srcs = nextSrcs
+			}
+			if wantSlices := test.NonEmptyByteSlices(); !reflect.DeepEqual(slices, wantSlices) {
+				t.Errorf("Accumulated slices: got %v, wanted %v", slices, wantSlices)
+			}
+		})
+	}
+}
+
+func TestBlockSeqIterBytes(t *testing.T) {
+	// Tests BlockSeq iteration using Head/DropFirst.
+	for _, test := range blockSeqTests {
+		t.Run(test.desc, func(t *testing.T) {
+			srcs := test.BlockSeq()
+			var dst bytes.Buffer
+			for !srcs.IsEmpty() {
+				src := srcs.Head()
+				var b [1]byte
+				n, err := Copy(BlockFromSafeSlice(b[:]), src)
+				if n != 1 || err != nil {
+					t.Fatalf("Copy: got (%v, %v), wanted (1, nil)", n, err)
+				}
+				dst.WriteByte(b[0])
+				nextSrcs := srcs.DropFirst(1)
+				if got, want := nextSrcs.NumBytes(), srcs.NumBytes()-1; got != want {
+					t.Fatalf("%v.DropFirst(1): got %v (%d bytes), wanted %d bytes", srcs, nextSrcs, got, want)
+				}
+				srcs = nextSrcs
+			}
+			if got := string(dst.Bytes()); got != test.want {
+				t.Errorf("Copied string: got %q, wanted %q", got, test.want)
+			}
+		})
+	}
+}
+
+func TestBlockSeqDropBeyondLimit(t *testing.T) {
+	blocks := []Block{BlockFromSafeSlice([]byte("123")), BlockFromSafeSlice([]byte("4"))}
+	bs := BlockSeqFromSlice(blocks)
+	if got, want := bs.NumBytes(), uint64(4); got != want {
+		t.Errorf("%v.NumBytes(): got %d, wanted %d", bs, got, want)
+	}
+	bs = bs.TakeFirst(1)
+	if got, want := bs.NumBytes(), uint64(1); got != want {
+		t.Errorf("%v.NumBytes(): got %d, wanted %d", bs, got, want)
+	}
+	bs = bs.DropFirst(2)
+	if got, want := bs.NumBytes(), uint64(0); got != want {
+		t.Errorf("%v.NumBytes(): got %d, wanted %d", bs, got, want)
+	}
+}
diff --git a/pkg/sentry/safemem/seq_unsafe.go b/pkg/sentry/safemem/seq_unsafe.go
new file mode 100644
index 000000000..e0d29a0b3
--- /dev/null
+++ b/pkg/sentry/safemem/seq_unsafe.go
@@ -0,0 +1,299 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package safemem
+
+import (
+	"bytes"
+	"fmt"
+	"reflect"
+	"unsafe"
+)
+
+// A BlockSeq represents a sequence of Blocks, each of which has non-zero
+// length.
+//
+// BlockSeqs are immutable and may be copied by value. The zero value of
+// BlockSeq represents an empty sequence.
+type BlockSeq struct {
+	// If length is 0, then the BlockSeq is empty. Invariants: data == 0;
+	// offset == 0; limit == 0.
+	//
+	// If length is -1, then the BlockSeq represents the single Block{data,
+	// limit, false}. Invariants: offset == 0; limit > 0; limit does not
+	// overflow the range of an int.
+	//
+	// If length is -2, then the BlockSeq represents the single Block{data,
+	// limit, true}. Invariants: offset == 0; limit > 0; limit does not
+	// overflow the range of an int.
+	//
+	// Otherwise, length >= 2, and the BlockSeq represents the `length` Blocks
+	// in the array of Blocks starting at address `data`, starting at `offset`
+	// bytes into the first Block and limited to the following `limit` bytes.
+	// Invariants: data != 0; offset < len(data[0]); limit > 0; offset+limit <=
+	// the combined length of all Blocks in the array; the first Block in the
+	// array has non-zero length.
+	//
+	// length is never 1; sequences consisting of a single Block are always
+	// stored inline (with length < 0).
+	data   unsafe.Pointer
+	length int
+	offset int
+	limit  uint64
+}
+
+// BlockSeqOf returns a BlockSeq representing the single Block b.
+func BlockSeqOf(b Block) BlockSeq {
+	bs := BlockSeq{
+		data:   b.start,
+		length: -1,
+		limit:  uint64(b.length),
+	}
+	if b.needSafecopy {
+		bs.length = -2
+	}
+	return bs
+}
+
+// BlockSeqFromSlice returns a BlockSeq representing all Blocks in slice.
+// If slice contains Blocks with zero length, BlockSeq will skip them during
+// iteration.
+//
+// Whether the returned BlockSeq shares memory with slice is unspecified;
+// clients should avoid mutating slices passed to BlockSeqFromSlice.
+//
+// Preconditions: The combined length of all Blocks in slice <= math.MaxUint64.
+func BlockSeqFromSlice(slice []Block) BlockSeq {
+	slice = skipEmpty(slice)
+	var limit uint64
+	for _, b := range slice {
+		sum := limit + uint64(b.Len())
+		if sum < limit {
+			panic("BlockSeq length overflows uint64")
+		}
+		limit = sum
+	}
+	return blockSeqFromSliceLimited(slice, limit)
+}
+
+// Preconditions: The combined length of all Blocks in slice <= limit. If
+// len(slice) != 0, the first Block in slice has non-zero length, and limit >
+// 0.
+func blockSeqFromSliceLimited(slice []Block, limit uint64) BlockSeq {
+	switch len(slice) {
+	case 0:
+		return BlockSeq{}
+	case 1:
+		return BlockSeqOf(slice[0].TakeFirst64(limit))
+	default:
+		return BlockSeq{
+			data:   unsafe.Pointer(&slice[0]),
+			length: len(slice),
+			limit:  limit,
+		}
+	}
+}
+
+func skipEmpty(slice []Block) []Block {
+	for i, b := range slice {
+		if b.Len() != 0 {
+			return slice[i:]
+		}
+	}
+	return nil
+}
+
+// IsEmpty returns true if bs contains no Blocks.
+//
+// Invariants: bs.IsEmpty() == (bs.NumBlocks() == 0) == (bs.NumBytes() == 0).
+// (Of these, prefer to use bs.IsEmpty().)
+func (bs BlockSeq) IsEmpty() bool {
+	return bs.length == 0
+}
+
+// NumBlocks returns the number of Blocks in bs.
+func (bs BlockSeq) NumBlocks() int {
+	// In general, we have to count: if bs represents a windowed slice then the
+	// slice may contain Blocks with zero length, and bs.length may be larger
+	// than the actual number of Blocks due to bs.limit.
+	var n int
+	for !bs.IsEmpty() {
+		n++
+		bs = bs.Tail()
+	}
+	return n
+}
+
+// NumBytes returns the sum of Block.Len() for all Blocks in bs.
+func (bs BlockSeq) NumBytes() uint64 {
+	return bs.limit
+}
+
+// Head returns the first Block in bs.
+//
+// Preconditions: !bs.IsEmpty().
+func (bs BlockSeq) Head() Block {
+	if bs.length == 0 {
+		panic("empty BlockSeq")
+	}
+	if bs.length < 0 {
+		return bs.internalBlock()
+	}
+	return (*Block)(bs.data).DropFirst(bs.offset).TakeFirst64(bs.limit)
+}
+
+// Preconditions: bs.length < 0.
+func (bs BlockSeq) internalBlock() Block {
+	return Block{
+		start:        bs.data,
+		length:       int(bs.limit),
+		needSafecopy: bs.length == -2,
+	}
+}
+
+// Tail returns a BlockSeq consisting of all Blocks in bs after the first.
+//
+// Preconditions: !bs.IsEmpty().
+func (bs BlockSeq) Tail() BlockSeq {
+	if bs.length == 0 {
+		panic("empty BlockSeq")
+	}
+	if bs.length < 0 {
+		return BlockSeq{}
+	}
+	head := (*Block)(bs.data).DropFirst(bs.offset)
+	headLen := uint64(head.Len())
+	if headLen >= bs.limit {
+		// The head Block exhausts the limit, so the tail is empty.
+		return BlockSeq{}
+	}
+	var extSlice []Block
+	extSliceHdr := (*reflect.SliceHeader)(unsafe.Pointer(&extSlice))
+	extSliceHdr.Data = uintptr(bs.data)
+	extSliceHdr.Len = bs.length
+	extSliceHdr.Cap = bs.length
+	tailSlice := skipEmpty(extSlice[1:])
+	tailLimit := bs.limit - headLen
+	return blockSeqFromSliceLimited(tailSlice, tailLimit)
+}
+
+// DropFirst returns a BlockSeq equivalent to bs, but with the first n bytes
+// omitted. If n > bs.NumBytes(), DropFirst returns an empty BlockSeq.
+//
+// Preconditions: n >= 0.
+func (bs BlockSeq) DropFirst(n int) BlockSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return bs.DropFirst64(uint64(n))
+}
+
+// DropFirst64 is equivalent to DropFirst but takes an uint64.
+func (bs BlockSeq) DropFirst64(n uint64) BlockSeq {
+	if n >= bs.limit {
+		return BlockSeq{}
+	}
+	for {
+		// Calling bs.Head() here is surprisingly expensive, so inline getting
+		// the head's length.
+		var headLen uint64
+		if bs.length < 0 {
+			headLen = bs.limit
+		} else {
+			headLen = uint64((*Block)(bs.data).Len() - bs.offset)
+		}
+		if n < headLen {
+			// Dropping ends partway through the head Block.
+			if bs.length < 0 {
+				return BlockSeqOf(bs.internalBlock().DropFirst64(n))
+			}
+			bs.offset += int(n)
+			bs.limit -= n
+			return bs
+		}
+		n -= headLen
+		bs = bs.Tail()
+	}
+}
+
+// TakeFirst returns a BlockSeq equivalent to the first n bytes of bs. If n >
+// bs.NumBytes(), TakeFirst returns a BlockSeq equivalent to bs.
+//
+// Preconditions: n >= 0.
+func (bs BlockSeq) TakeFirst(n int) BlockSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return bs.TakeFirst64(uint64(n))
+}
+
+// TakeFirst64 is equivalent to TakeFirst but takes a uint64.
+func (bs BlockSeq) TakeFirst64(n uint64) BlockSeq {
+	if n == 0 {
+		return BlockSeq{}
+	}
+	if bs.limit > n {
+		bs.limit = n
+	}
+	return bs
+}
+
+// String implements fmt.Stringer.String.
+func (bs BlockSeq) String() string {
+	var buf bytes.Buffer
+	buf.WriteByte('[')
+	var sep string
+	for !bs.IsEmpty() {
+		buf.WriteString(sep)
+		sep = " "
+		buf.WriteString(bs.Head().String())
+		bs = bs.Tail()
+	}
+	buf.WriteByte(']')
+	return buf.String()
+}
+
+// CopySeq copies srcs.NumBytes() or dsts.NumBytes() bytes, whichever is less,
+// from srcs to dsts and returns the number of bytes copied.
+//
+// If srcs and dsts overlap, the data stored in dsts is unspecified.
+func CopySeq(dsts, srcs BlockSeq) (uint64, error) {
+	var done uint64
+	for !dsts.IsEmpty() && !srcs.IsEmpty() {
+		dst := dsts.Head()
+		src := srcs.Head()
+		n, err := Copy(dst, src)
+		done += uint64(n)
+		if err != nil {
+			return done, err
+		}
+		dsts = dsts.DropFirst(n)
+		srcs = srcs.DropFirst(n)
+	}
+	return done, nil
+}
+
+// ZeroSeq sets all bytes in dsts to 0 and returns the number of bytes zeroed.
+func ZeroSeq(dsts BlockSeq) (uint64, error) {
+	var done uint64
+	for !dsts.IsEmpty() {
+		n, err := Zero(dsts.Head())
+		done += uint64(n)
+		if err != nil {
+			return done, err
+		}
+		dsts = dsts.DropFirst(n)
+	}
+	return done, nil
+}
diff --git a/pkg/sentry/sighandling/BUILD b/pkg/sentry/sighandling/BUILD
new file mode 100644
index 000000000..daaad7c90
--- /dev/null
+++ b/pkg/sentry/sighandling/BUILD
@@ -0,0 +1,18 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "sighandling",
+    srcs = [
+        "sighandling.go",
+        "sighandling_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/sighandling",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/kernel",
+    ],
+)
diff --git a/pkg/sentry/sighandling/sighandling.go b/pkg/sentry/sighandling/sighandling.go
new file mode 100644
index 000000000..1a94b535b
--- /dev/null
+++ b/pkg/sentry/sighandling/sighandling.go
@@ -0,0 +1,116 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package sighandling contains helpers for handling signals to applications.
+package sighandling
+
+import (
+	"os"
+	"os/signal"
+	"reflect"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// numSignals is the number of normal (non-realtime) signals on Linux.
+const numSignals = 32
+
+// forwardSignals listens for incoming signals and delivers them to k. It stops
+// when the stop channel is closed.
+func forwardSignals(k *kernel.Kernel, sigchans []chan os.Signal, stop chan struct{}) {
+	// Build a select case.
+	sc := []reflect.SelectCase{{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(stop)}}
+	for _, sigchan := range sigchans {
+		sc = append(sc, reflect.SelectCase{Dir: reflect.SelectRecv, Chan: reflect.ValueOf(sigchan)})
+	}
+
+	for {
+		// Wait for a notification.
+		index, _, ok := reflect.Select(sc)
+
+		// Was it the stop channel?
+		if index == 0 {
+			if !ok {
+				break
+			}
+			continue
+		}
+
+		// How about a different close?
+		if !ok {
+			panic("signal channel closed unexpectedly")
+		}
+
+		// Otherwise, it was a signal on channel N. Index 0 represents the stop
+		// channel, so index N represents the channel for signal N.
+		if !k.SendExternalSignal(&arch.SignalInfo{Signo: int32(index)}, "sentry") {
+			// Kernel is not ready to receive signals.
+			//
+			// Kill ourselves if this signal would have killed the
+			// process before StartForwarding was called. i.e., all
+			// _SigKill signals; see Go
+			// src/runtime/sigtab_linux_generic.go.
+			//
+			// Otherwise ignore the signal.
+			//
+			// TODO: Convert Go's runtime.raise from
+			// tkill to tgkill so StartForwarding doesn't need to
+			// be called until after filter installation.
+			switch linux.Signal(index) {
+			case linux.SIGHUP, linux.SIGINT, linux.SIGTERM:
+				dieFromSignal(linux.Signal(index))
+			}
+		}
+	}
+
+	// Close all individual channels.
+	for _, sigchan := range sigchans {
+		signal.Stop(sigchan)
+		close(sigchan)
+	}
+}
+
+// StartForwarding ensures that synchronous signals are forwarded to k and
+// returns a callback that stops signal forwarding.
+func StartForwarding(k *kernel.Kernel) func() {
+	stop := make(chan struct{})
+
+	// Register individual channels. One channel per standard signal is
+	// required as os.Notify() is non-blocking and may drop signals. To avoid
+	// this, standard signals have to be queued separately. Channel size 1 is
+	// enough for standard signals as their semantics allow de-duplication.
+	//
+	// External real-time signals are not supported. We rely on the go-runtime
+	// for their handling.
+	var sigchans []chan os.Signal
+	for sig := 1; sig <= numSignals+1; sig++ {
+		sigchan := make(chan os.Signal, 1)
+		sigchans = append(sigchans, sigchan)
+
+		// SignalPanic is handled by Run.
+		if linux.Signal(sig) == kernel.SignalPanic {
+			continue
+		}
+
+		signal.Notify(sigchan, syscall.Signal(sig))
+	}
+	// Start up our listener.
+	go forwardSignals(k, sigchans, stop) // S/R-SAFE: synchronized by Kernel.extMu
+
+	// ... shouldn't this wait until the forwardSignals goroutine returns?
+	return func() { close(stop) }
+}
diff --git a/pkg/sentry/sighandling/sighandling_unsafe.go b/pkg/sentry/sighandling/sighandling_unsafe.go
new file mode 100644
index 000000000..a455b919f
--- /dev/null
+++ b/pkg/sentry/sighandling/sighandling_unsafe.go
@@ -0,0 +1,74 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package sighandling
+
+import (
+	"fmt"
+	"runtime"
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// TODO: Move to pkg/abi/linux along with definitions in
+// pkg/sentry/arch.
+type sigaction struct {
+	handler  uintptr
+	flags    uint64
+	restorer uintptr
+	mask     uint64
+}
+
+// IgnoreChildStop sets the SA_NOCLDSTOP flag, causing child processes to not
+// generate SIGCHLD when they stop.
+func IgnoreChildStop() error {
+	var sa sigaction
+
+	// Get the existing signal handler information, and set the flag.
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(syscall.SIGCHLD), 0, uintptr(unsafe.Pointer(&sa)), linux.SignalSetSize, 0, 0); e != 0 {
+		return e
+	}
+	sa.flags |= linux.SA_NOCLDSTOP
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(syscall.SIGCHLD), uintptr(unsafe.Pointer(&sa)), 0, linux.SignalSetSize, 0, 0); e != 0 {
+		return e
+	}
+
+	return nil
+}
+
+// dieFromSignal kills the current process with sig.
+//
+// Preconditions: The default action of sig is termination.
+func dieFromSignal(sig linux.Signal) {
+	runtime.LockOSThread()
+	defer runtime.UnlockOSThread()
+
+	sa := sigaction{handler: linux.SIG_DFL}
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGACTION, uintptr(sig), uintptr(unsafe.Pointer(&sa)), 0, linux.SignalSetSize, 0, 0); e != 0 {
+		panic(fmt.Sprintf("rt_sigaction failed: %v", e))
+	}
+
+	set := linux.MakeSignalSet(sig)
+	if _, _, e := syscall.RawSyscall6(syscall.SYS_RT_SIGPROCMASK, linux.SIG_UNBLOCK, uintptr(unsafe.Pointer(&set)), 0, linux.SignalSetSize, 0, 0); e != 0 {
+		panic(fmt.Sprintf("rt_sigprocmask failed: %v", e))
+	}
+
+	if err := syscall.Tgkill(syscall.Getpid(), syscall.Gettid(), syscall.Signal(sig)); err != nil {
+		panic(fmt.Sprintf("tgkill failed: %v", err))
+	}
+
+	panic("failed to die")
+}
diff --git a/pkg/sentry/socket/BUILD b/pkg/sentry/socket/BUILD
new file mode 100644
index 000000000..87e32df37
--- /dev/null
+++ b/pkg/sentry/socket/BUILD
@@ -0,0 +1,37 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "socket_state",
+    srcs = [
+        "socket.go",
+    ],
+    out = "socket_state_autogen.go",
+    package = "socket",
+)
+
+go_library(
+    name = "socket",
+    srcs = [
+        "socket.go",
+        "socket_state_autogen.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/tcpip/transport/unix",
+    ],
+)
diff --git a/pkg/sentry/socket/control/BUILD b/pkg/sentry/socket/control/BUILD
new file mode 100644
index 000000000..25de2f655
--- /dev/null
+++ b/pkg/sentry/socket/control/BUILD
@@ -0,0 +1,39 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "control_state",
+    srcs = [
+        "control.go",
+    ],
+    out = "control_state.go",
+    imports = [
+        "gvisor.googlesource.com/gvisor/pkg/sentry/fs",
+    ],
+    package = "control",
+)
+
+go_library(
+    name = "control",
+    srcs = [
+        "control.go",
+        "control_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/control",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+    ],
+)
diff --git a/pkg/sentry/socket/control/control.go b/pkg/sentry/socket/control/control.go
new file mode 100644
index 000000000..cb34cbc85
--- /dev/null
+++ b/pkg/sentry/socket/control/control.go
@@ -0,0 +1,370 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package control provides internal representations of socket control
+// messages.
+package control
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+const maxInt = int(^uint(0) >> 1)
+
+// SCMCredentials represents a SCM_CREDENTIALS socket control message.
+type SCMCredentials interface {
+	unix.CredentialsControlMessage
+
+	// Credentials returns properly namespaced values for the sender's pid, uid
+	// and gid.
+	Credentials(t *kernel.Task) (kernel.ThreadID, auth.UID, auth.GID)
+}
+
+// SCMRights represents a SCM_RIGHTS socket control message.
+type SCMRights interface {
+	unix.RightsControlMessage
+
+	// Files returns up to max RightsFiles.
+	Files(ctx context.Context, max int) RightsFiles
+}
+
+// RightsFiles represents a SCM_RIGHTS socket control message. A reference is
+// maintained for each fs.File and is release either when an FD is created or
+// when the Release method is called.
+type RightsFiles []*fs.File
+
+// NewSCMRights creates a new SCM_RIGHTS socket control message representation
+// using local sentry FDs.
+func NewSCMRights(t *kernel.Task, fds []int32) (SCMRights, error) {
+	files := make(RightsFiles, 0, len(fds))
+	for _, fd := range fds {
+		file, _ := t.FDMap().GetDescriptor(kdefs.FD(fd))
+		if file == nil {
+			files.Release()
+			return nil, syserror.EBADF
+		}
+		files = append(files, file)
+	}
+	return &files, nil
+}
+
+// Files implements SCMRights.Files.
+func (fs *RightsFiles) Files(ctx context.Context, max int) RightsFiles {
+	n := max
+	if l := len(*fs); n > l {
+		n = l
+	}
+	rf := (*fs)[:n]
+	*fs = (*fs)[n:]
+	return rf
+}
+
+// Clone implements unix.RightsControlMessage.Clone.
+func (fs *RightsFiles) Clone() unix.RightsControlMessage {
+	nfs := append(RightsFiles(nil), *fs...)
+	for _, nf := range nfs {
+		nf.IncRef()
+	}
+	return &nfs
+}
+
+// Release implements unix.RightsControlMessage.Release.
+func (fs *RightsFiles) Release() {
+	for _, f := range *fs {
+		f.DecRef()
+	}
+	*fs = nil
+}
+
+// rightsFDs gets up to the specified maximum number of FDs.
+func rightsFDs(t *kernel.Task, rights SCMRights, cloexec bool, max int) []int32 {
+	files := rights.Files(t, max)
+	fds := make([]int32, 0, len(files))
+	for i := 0; i < max && len(files) > 0; i++ {
+		fd, err := t.FDMap().NewFDFrom(0, files[0], kernel.FDFlags{cloexec}, t.ThreadGroup().Limits())
+		files[0].DecRef()
+		files = files[1:]
+		if err != nil {
+			t.Warningf("Error inserting FD: %v", err)
+			// This is what Linux does.
+			break
+		}
+
+		fds = append(fds, int32(fd))
+	}
+	return fds
+}
+
+// PackRights packs as many FDs as will fit into the unused capacity of buf.
+func PackRights(t *kernel.Task, rights SCMRights, cloexec bool, buf []byte) []byte {
+	maxFDs := (cap(buf) - len(buf) - linux.SizeOfControlMessageHeader) / 4
+	// Linux does not return any FDs if none fit.
+	if maxFDs <= 0 {
+		return buf
+	}
+	fds := rightsFDs(t, rights, cloexec, maxFDs)
+	align := t.Arch().Width()
+	return putCmsg(buf, linux.SCM_RIGHTS, align, fds)
+}
+
+// scmCredentials represents an SCM_CREDENTIALS socket control message.
+type scmCredentials struct {
+	t    *kernel.Task
+	kuid auth.KUID
+	kgid auth.KGID
+}
+
+// NewSCMCredentials creates a new SCM_CREDENTIALS socket control message
+// representation.
+func NewSCMCredentials(t *kernel.Task, cred linux.ControlMessageCredentials) (SCMCredentials, error) {
+	tcred := t.Credentials()
+	kuid, err := tcred.UseUID(auth.UID(cred.UID))
+	if err != nil {
+		return nil, err
+	}
+	kgid, err := tcred.UseGID(auth.GID(cred.GID))
+	if err != nil {
+		return nil, err
+	}
+	if kernel.ThreadID(cred.PID) != t.ThreadGroup().ID() && !t.HasCapabilityIn(linux.CAP_SYS_ADMIN, t.PIDNamespace().UserNamespace()) {
+		return nil, syserror.EPERM
+	}
+	return &scmCredentials{t, kuid, kgid}, nil
+}
+
+// Equals implements unix.CredentialsControlMessage.Equals.
+func (c *scmCredentials) Equals(oc unix.CredentialsControlMessage) bool {
+	if oc, _ := oc.(*scmCredentials); oc != nil && *c == *oc {
+		return true
+	}
+	return false
+}
+
+func putUint64(buf []byte, n uint64) []byte {
+	usermem.ByteOrder.PutUint64(buf[len(buf):len(buf)+8], n)
+	return buf[:len(buf)+8]
+}
+
+func putUint32(buf []byte, n uint32) []byte {
+	usermem.ByteOrder.PutUint32(buf[len(buf):len(buf)+4], n)
+	return buf[:len(buf)+4]
+}
+
+// putCmsg writes a control message header and as much data as will fit into
+// the unused capacity of a buffer.
+func putCmsg(buf []byte, msgType uint32, align uint, data []int32) []byte {
+	space := AlignDown(cap(buf)-len(buf), 4)
+
+	// We can't write to space that doesn't exist, so if we are going to align
+	// the available space, we must align down.
+	//
+	// align must be >= 4 and each data int32 is 4 bytes. The length of the
+	// header is already aligned, so if we align to the with of the data there
+	// are two cases:
+	// 1. The aligned length is less than the length of the header. The
+	// unaligned length was also less than the length of the header, so we
+	// can't write anything.
+	// 2. The aligned length is greater than or equal to the length of the
+	// header. We can write the header plus zero or more datas. We can't write
+	// a partial int32, so the length of the message will be
+	// min(aligned length, header + datas).
+	if space < linux.SizeOfControlMessageHeader {
+		return buf
+	}
+
+	length := 4*len(data) + linux.SizeOfControlMessageHeader
+	if length > space {
+		length = space
+	}
+	buf = putUint64(buf, uint64(length))
+	buf = putUint32(buf, linux.SOL_SOCKET)
+	buf = putUint32(buf, msgType)
+	for _, d := range data {
+		if len(buf)+4 > cap(buf) {
+			break
+		}
+		buf = putUint32(buf, uint32(d))
+	}
+	return alignSlice(buf, align)
+}
+
+// Credentials implements SCMCredentials.Credentials.
+func (c *scmCredentials) Credentials(t *kernel.Task) (kernel.ThreadID, auth.UID, auth.GID) {
+	// "When a process's user and group IDs are passed over a UNIX domain
+	// socket to a process in a different user namespace (see the description
+	// of SCM_CREDENTIALS in unix(7)), they are translated into the
+	// corresponding values as per the receiving process's user and group ID
+	// mappings." - user_namespaces(7)
+	pid := t.PIDNamespace().IDOfTask(c.t)
+	uid := c.kuid.In(t.UserNamespace()).OrOverflow()
+	gid := c.kgid.In(t.UserNamespace()).OrOverflow()
+
+	return pid, uid, gid
+}
+
+// PackCredentials packs the credentials in the control message (or default
+// credentials if none) into a buffer.
+func PackCredentials(t *kernel.Task, creds SCMCredentials, buf []byte) []byte {
+	align := t.Arch().Width()
+
+	// Default credentials if none are available.
+	pid := kernel.ThreadID(0)
+	uid := auth.UID(auth.NobodyKUID)
+	gid := auth.GID(auth.NobodyKGID)
+
+	if creds != nil {
+		pid, uid, gid = creds.Credentials(t)
+	}
+	c := []int32{int32(pid), int32(uid), int32(gid)}
+	return putCmsg(buf, linux.SCM_CREDENTIALS, align, c)
+}
+
+// AlignUp rounds a length up to an alignment. align must be a power of 2.
+func AlignUp(length int, align uint) int {
+	return (length + int(align) - 1) & ^(int(align) - 1)
+}
+
+// AlignDown rounds a down to an alignment. align must be a power of 2.
+func AlignDown(length int, align uint) int {
+	return length & ^(int(align) - 1)
+}
+
+// alignSlice extends a slice's length (up to the capacity) to align it.
+func alignSlice(buf []byte, align uint) []byte {
+	aligned := AlignUp(len(buf), align)
+	if aligned > cap(buf) {
+		// Linux allows unaligned data if there isn't room for alignment.
+		// Since there isn't room for alignment, there isn't room for any
+		// additional messages either.
+		return buf
+	}
+	return buf[:aligned]
+}
+
+// Parse parses a raw socket control message into portable objects.
+func Parse(t *kernel.Task, socketOrEndpoint interface{}, buf []byte) (unix.ControlMessages, error) {
+	var (
+		fds linux.ControlMessageRights
+
+		haveCreds bool
+		creds     linux.ControlMessageCredentials
+	)
+
+	for i := 0; i < len(buf); {
+		if i+linux.SizeOfControlMessageHeader > len(buf) {
+			return unix.ControlMessages{}, syserror.EINVAL
+		}
+
+		var h linux.ControlMessageHeader
+		binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageHeader], usermem.ByteOrder, &h)
+
+		if h.Length < uint64(linux.SizeOfControlMessageHeader) {
+			return unix.ControlMessages{}, syserror.EINVAL
+		}
+		if h.Length > uint64(len(buf)-i) {
+			return unix.ControlMessages{}, syserror.EINVAL
+		}
+		if h.Level != linux.SOL_SOCKET {
+			return unix.ControlMessages{}, syserror.EINVAL
+		}
+
+		i += linux.SizeOfControlMessageHeader
+		length := int(h.Length) - linux.SizeOfControlMessageHeader
+
+		// The use of t.Arch().Width() is analogous to Linux's use of
+		// sizeof(long) in CMSG_ALIGN.
+		width := t.Arch().Width()
+
+		switch h.Type {
+		case linux.SCM_RIGHTS:
+			rightsSize := AlignDown(length, linux.SizeOfControlMessageRight)
+			numRights := rightsSize / linux.SizeOfControlMessageRight
+
+			if len(fds)+numRights > linux.SCM_MAX_FD {
+				return unix.ControlMessages{}, syserror.EINVAL
+			}
+
+			for j := i; j < i+rightsSize; j += linux.SizeOfControlMessageRight {
+				fds = append(fds, int32(usermem.ByteOrder.Uint32(buf[j:j+linux.SizeOfControlMessageRight])))
+			}
+
+			i += AlignUp(length, width)
+
+		case linux.SCM_CREDENTIALS:
+			if length < linux.SizeOfControlMessageCredentials {
+				return unix.ControlMessages{}, syserror.EINVAL
+			}
+
+			binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageCredentials], usermem.ByteOrder, &creds)
+			haveCreds = true
+			i += AlignUp(length, width)
+
+		default:
+			// Unknown message type.
+			return unix.ControlMessages{}, syserror.EINVAL
+		}
+	}
+
+	var credentials SCMCredentials
+	if haveCreds {
+		var err error
+		if credentials, err = NewSCMCredentials(t, creds); err != nil {
+			return unix.ControlMessages{}, err
+		}
+	} else {
+		credentials = makeCreds(t, socketOrEndpoint)
+	}
+
+	var rights SCMRights
+	if len(fds) > 0 {
+		var err error
+		if rights, err = NewSCMRights(t, fds); err != nil {
+			return unix.ControlMessages{}, err
+		}
+	}
+
+	if credentials == nil && rights == nil {
+		return unix.ControlMessages{}, nil
+	}
+
+	return unix.ControlMessages{Credentials: credentials, Rights: rights}, nil
+}
+
+func makeCreds(t *kernel.Task, socketOrEndpoint interface{}) SCMCredentials {
+	if t == nil || socketOrEndpoint == nil {
+		return nil
+	}
+	if cr, ok := socketOrEndpoint.(unix.Credentialer); ok && (cr.Passcred() || cr.ConnectedPasscred()) {
+		tcred := t.Credentials()
+		return &scmCredentials{t, tcred.EffectiveKUID, tcred.EffectiveKGID}
+	}
+	return nil
+}
+
+// New creates default control messages if needed.
+func New(t *kernel.Task, socketOrEndpoint interface{}, rights SCMRights) unix.ControlMessages {
+	return unix.ControlMessages{
+		Credentials: makeCreds(t, socketOrEndpoint),
+		Rights:      rights,
+	}
+}
diff --git a/pkg/sentry/socket/epsocket/BUILD b/pkg/sentry/socket/epsocket/BUILD
new file mode 100644
index 000000000..0e463a92a
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/BUILD
@@ -0,0 +1,61 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "epsocket_state",
+    srcs = [
+        "epsocket.go",
+        "save_restore.go",
+        "stack.go",
+    ],
+    out = "epsocket_state.go",
+    package = "epsocket",
+)
+
+go_library(
+    name = "epsocket",
+    srcs = [
+        "device.go",
+        "epsocket.go",
+        "epsocket_state.go",
+        "provider.go",
+        "save_restore.go",
+        "stack.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/epsocket",
+    visibility = [
+        "//pkg/sentry:internal",
+    ],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/log",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/buffer",
+        "//pkg/tcpip/network/ipv4",
+        "//pkg/tcpip/network/ipv6",
+        "//pkg/tcpip/stack",
+        "//pkg/tcpip/transport/tcp",
+        "//pkg/tcpip/transport/udp",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/socket/epsocket/device.go b/pkg/sentry/socket/epsocket/device.go
new file mode 100644
index 000000000..17f2c9559
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epsocket
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// epsocketDevice is the endpoint socket virtual device.
+var epsocketDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/socket/epsocket/epsocket.go b/pkg/sentry/socket/epsocket/epsocket.go
new file mode 100644
index 000000000..3fc3ea58f
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/epsocket.go
@@ -0,0 +1,1230 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package epsocket provides an implementation of the socket.Socket interface
+// that is backed by a tcpip.Endpoint.
+//
+// It does not depend on any particular endpoint implementation, and thus can
+// be used to expose certain endpoints to the sentry while leaving others out,
+// for example, TCP endpoints and Unix-domain endpoints.
+//
+// Lock ordering: netstack => mm: ioSequencePayload copies user memory inside
+// tcpip.Endpoint.Write(). Netstack is allowed to (and does) hold locks during
+// this operation.
+package epsocket
+
+import (
+	"bytes"
+	"math"
+	"strings"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/buffer"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+const sizeOfInt32 int = 4
+
+// ntohs converts a 16-bit number from network byte order to host byte order. It
+// assumes that the host is little endian.
+func ntohs(v uint16) uint16 {
+	return v<<8 | v>>8
+}
+
+// htons converts a 16-bit number from host byte order to network byte order. It
+// assumes that the host is little endian.
+func htons(v uint16) uint16 {
+	return ntohs(v)
+}
+
+// commonEndpoint represents the intersection of a tcpip.Endpoint and a
+// unix.Endpoint.
+type commonEndpoint interface {
+	// GetLocalAddress implements tcpip.Endpoint.GetLocalAddress and
+	// unix.Endpoint.GetLocalAddress.
+	GetLocalAddress() (tcpip.FullAddress, *tcpip.Error)
+
+	// GetRemoteAddress implements tcpip.Endpoint.GetRemoteAddress and
+	// unix.Endpoint.GetRemoteAddress.
+	GetRemoteAddress() (tcpip.FullAddress, *tcpip.Error)
+
+	// Readiness implements tcpip.Endpoint.Readiness and
+	// unix.Endpoint.Readiness.
+	Readiness(mask waiter.EventMask) waiter.EventMask
+
+	// SetSockOpt implements tcpip.Endpoint.SetSockOpt and
+	// unix.Endpoint.SetSockOpt.
+	SetSockOpt(interface{}) *tcpip.Error
+
+	// GetSockOpt implements tcpip.Endpoint.GetSockOpt and
+	// unix.Endpoint.GetSockOpt.
+	GetSockOpt(interface{}) *tcpip.Error
+}
+
+// SocketOperations encapsulates all the state needed to represent a network stack
+// endpoint in the kernel context.
+type SocketOperations struct {
+	socket.ReceiveTimeout
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	*waiter.Queue
+
+	family   int
+	stack    inet.Stack
+	Endpoint tcpip.Endpoint
+	skType   unix.SockType
+
+	// readMu protects access to readView, control, and sender.
+	readMu   sync.Mutex `state:"nosave"`
+	readView buffer.View
+	sender   tcpip.FullAddress
+}
+
+// New creates a new endpoint socket.
+func New(t *kernel.Task, family int, skType unix.SockType, queue *waiter.Queue, endpoint tcpip.Endpoint) *fs.File {
+	dirent := socket.NewDirent(t, epsocketDevice)
+	return fs.NewFile(t, dirent, fs.FileFlags{Read: true, Write: true}, &SocketOperations{
+		Queue:    queue,
+		family:   family,
+		stack:    t.NetworkContext(),
+		Endpoint: endpoint,
+		skType:   skType,
+	})
+}
+
+var sockAddrInetSize = int(binary.Size(linux.SockAddrInet{}))
+var sockAddrInet6Size = int(binary.Size(linux.SockAddrInet6{}))
+
+// GetAddress reads an sockaddr struct from the given address and converts it
+// to the FullAddress format. It supports AF_UNIX, AF_INET and AF_INET6
+// addresses.
+func GetAddress(sfamily int, addr []byte) (tcpip.FullAddress, *syserr.Error) {
+	// Make sure we have at least 2 bytes for the address family.
+	if len(addr) < 2 {
+		return tcpip.FullAddress{}, syserr.ErrInvalidArgument
+	}
+
+	family := usermem.ByteOrder.Uint16(addr)
+	if family != uint16(sfamily) {
+		return tcpip.FullAddress{}, syserr.ErrAddressFamilyNotSupported
+	}
+
+	// Get the rest of the fields based on the address family.
+	switch family {
+	case linux.AF_UNIX:
+		path := addr[2:]
+		// Drop the terminating NUL (if one exists) and everything after it.
+		// Skip the first byte, which is NUL for abstract paths.
+		if len(path) > 1 {
+			if n := bytes.IndexByte(path[1:], 0); n >= 0 {
+				path = path[:n+1]
+			}
+		}
+		return tcpip.FullAddress{
+			Addr: tcpip.Address(path),
+		}, nil
+
+	case linux.AF_INET:
+		var a linux.SockAddrInet
+		if len(addr) < sockAddrInetSize {
+			return tcpip.FullAddress{}, syserr.ErrBadAddress
+		}
+		binary.Unmarshal(addr[:sockAddrInetSize], usermem.ByteOrder, &a)
+
+		out := tcpip.FullAddress{
+			Addr: tcpip.Address(a.Addr[:]),
+			Port: ntohs(a.Port),
+		}
+		if out.Addr == "\x00\x00\x00\x00" {
+			out.Addr = ""
+		}
+		return out, nil
+
+	case linux.AF_INET6:
+		var a linux.SockAddrInet6
+		if len(addr) < sockAddrInet6Size {
+			return tcpip.FullAddress{}, syserr.ErrBadAddress
+		}
+		binary.Unmarshal(addr[:sockAddrInet6Size], usermem.ByteOrder, &a)
+
+		out := tcpip.FullAddress{
+			Addr: tcpip.Address(a.Addr[:]),
+			Port: ntohs(a.Port),
+		}
+		if isLinkLocal(out.Addr) {
+			out.NIC = tcpip.NICID(a.Scope_id)
+		}
+		if out.Addr == tcpip.Address(strings.Repeat("\x00", 16)) {
+			out.Addr = ""
+		}
+		return out, nil
+
+	default:
+		return tcpip.FullAddress{}, syserr.ErrAddressFamilyNotSupported
+	}
+}
+
+func (s *SocketOperations) isPacketBased() bool {
+	return s.skType == linux.SOCK_DGRAM || s.skType == linux.SOCK_SEQPACKET || s.skType == linux.SOCK_RDM
+}
+
+// fetchReadView updates the readView field of the socket if it's currently
+// empty. It assumes that the socket is locked.
+func (s *SocketOperations) fetchReadView() *syserr.Error {
+	if len(s.readView) > 0 {
+		return nil
+	}
+
+	s.readView = nil
+	s.sender = tcpip.FullAddress{}
+
+	v, err := s.Endpoint.Read(&s.sender)
+	if err != nil {
+		return syserr.TranslateNetstackError(err)
+	}
+
+	s.readView = v
+
+	return nil
+}
+
+// Release implements fs.FileOperations.Release.
+func (s *SocketOperations) Release() {
+	s.Endpoint.Close()
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *SocketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	n, _, _, err := s.nonBlockingRead(ctx, dst, false, false, false)
+	if err == syserr.ErrWouldBlock {
+		return int64(n), syserror.ErrWouldBlock
+	}
+	if err != nil {
+		return 0, err.ToError()
+	}
+	return int64(n), nil
+}
+
+// ioSequencePayload implements tcpip.Payload. It copies user memory bytes on demand
+// based on the requested size.
+type ioSequencePayload struct {
+	ctx context.Context
+	src usermem.IOSequence
+}
+
+// Get implements tcpip.Payload.
+func (i *ioSequencePayload) Get(size int) ([]byte, *tcpip.Error) {
+	if size > i.Size() {
+		size = i.Size()
+	}
+	v := buffer.NewView(size)
+	if _, err := i.src.CopyIn(i.ctx, v); err != nil {
+		return nil, tcpip.ErrBadAddress
+	}
+	return v, nil
+}
+
+// Size implements tcpip.Payload.
+func (i *ioSequencePayload) Size() int {
+	return int(i.src.NumBytes())
+}
+
+// Write implements fs.FileOperations.Write.
+func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	f := &ioSequencePayload{ctx: ctx, src: src}
+	n, err := s.Endpoint.Write(f, tcpip.WriteOptions{})
+	if err == tcpip.ErrWouldBlock {
+		return int64(n), syserror.ErrWouldBlock
+	}
+	return int64(n), syserr.TranslateNetstackError(err).ToError()
+}
+
+// Readiness returns a mask of ready events for socket s.
+func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	r := s.Endpoint.Readiness(mask)
+
+	// Check our cached value iff the caller asked for readability and the
+	// endpoint itself is currently not readable.
+	if (mask & ^r & waiter.EventIn) != 0 {
+		s.readMu.Lock()
+		if len(s.readView) > 0 {
+			r |= waiter.EventIn
+		}
+		s.readMu.Unlock()
+	}
+
+	return r
+}
+
+// Connect implements the linux syscall connect(2) for sockets backed by
+// tpcip.Endpoint.
+func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	addr, err := GetAddress(s.family, sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// Always return right away in the non-blocking case.
+	if !blocking {
+		return syserr.TranslateNetstackError(s.Endpoint.Connect(addr))
+	}
+
+	// Register for notification when the endpoint becomes writable, then
+	// initiate the connection.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+
+	if err := s.Endpoint.Connect(addr); err != tcpip.ErrConnectStarted && err != tcpip.ErrAlreadyConnecting {
+		return syserr.TranslateNetstackError(err)
+	}
+
+	// It's pending, so we have to wait for a notification, and fetch the
+	// result once the wait completes.
+	if err := t.Block(ch); err != nil {
+		return syserr.FromError(err)
+	}
+
+	// Call Connect() again after blocking to find connect's result.
+	return syserr.TranslateNetstackError(s.Endpoint.Connect(addr))
+}
+
+// Bind implements the linux syscall bind(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	addr, err := GetAddress(s.family, sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// Issue the bind request to the endpoint.
+	return syserr.TranslateNetstackError(s.Endpoint.Bind(addr, nil))
+}
+
+// Listen implements the linux syscall listen(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	return syserr.TranslateNetstackError(s.Endpoint.Listen(backlog))
+}
+
+// blockingAccept implements a blocking version of accept(2), that is, if no
+// connections are ready to be accept, it will block until one becomes ready.
+func (s *SocketOperations) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) {
+	// Register for notifications.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	// Try to accept the connection again; if it fails, then wait until we
+	// get a notification.
+	for {
+		if ep, wq, err := s.Endpoint.Accept(); err != tcpip.ErrWouldBlock {
+			return ep, wq, syserr.TranslateNetstackError(err)
+		}
+
+		if err := t.Block(ch); err != nil {
+			return nil, nil, syserr.FromError(err)
+		}
+	}
+}
+
+// Accept implements the linux syscall accept(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	// Issue the accept request to get the new endpoint.
+	ep, wq, err := s.Endpoint.Accept()
+	if err != nil {
+		if err != tcpip.ErrWouldBlock || !blocking {
+			return 0, nil, 0, syserr.TranslateNetstackError(err)
+		}
+
+		var err *syserr.Error
+		ep, wq, err = s.blockingAccept(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	ns := New(t, s.family, s.skType, wq, ep)
+	defer ns.DecRef()
+
+	if flags&linux.SOCK_NONBLOCK != 0 {
+		flags := ns.Flags()
+		flags.NonBlocking = true
+		ns.SetFlags(flags.Settable())
+	}
+
+	var addr interface{}
+	var addrLen uint32
+	if peerRequested {
+		// Get address of the peer and write it to peer slice.
+		var err *syserr.Error
+		addr, addrLen, err = ns.FileOperations.(*SocketOperations).GetPeerName(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	fdFlags := kernel.FDFlags{
+		CloseOnExec: flags&linux.SOCK_CLOEXEC != 0,
+	}
+	fd, e := t.FDMap().NewFDFrom(0, ns, fdFlags, t.ThreadGroup().Limits())
+
+	return fd, addr, addrLen, syserr.FromError(e)
+}
+
+// ConvertShutdown converts Linux shutdown flags into tcpip shutdown flags.
+func ConvertShutdown(how int) (tcpip.ShutdownFlags, *syserr.Error) {
+	var f tcpip.ShutdownFlags
+	switch how {
+	case linux.SHUT_RD:
+		f = tcpip.ShutdownRead
+	case linux.SHUT_WR:
+		f = tcpip.ShutdownWrite
+	case linux.SHUT_RDWR:
+		f = tcpip.ShutdownRead | tcpip.ShutdownWrite
+	default:
+		return 0, syserr.ErrInvalidArgument
+	}
+	return f, nil
+}
+
+// Shutdown implements the linux syscall shutdown(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	f, err := ConvertShutdown(how)
+	if err != nil {
+		return err
+	}
+
+	// Issue shutdown request.
+	return syserr.TranslateNetstackError(s.Endpoint.Shutdown(f))
+}
+
+// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name, outLen int) (interface{}, *syserr.Error) {
+	return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outLen)
+}
+
+// GetSockOpt can be used to implement the linux syscall getsockopt(2) for
+// sockets backed by a commonEndpoint.
+func GetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, family int, skType unix.SockType, level, name, outLen int) (interface{}, *syserr.Error) {
+	switch level {
+	case syscall.SOL_SOCKET:
+		switch name {
+		case linux.SO_TYPE:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+			return int32(skType), nil
+
+		case linux.SO_ERROR:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			// Get the last error and convert it.
+			err := ep.GetSockOpt(tcpip.ErrorOption{})
+			if err == nil {
+				return int32(0), nil
+			}
+			return int32(syserr.ToLinux(syserr.TranslateNetstackError(err)).Number()), nil
+
+		case linux.SO_PEERCRED:
+			if family != linux.AF_UNIX || outLen < syscall.SizeofUcred {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			tcred := t.Credentials()
+			return syscall.Ucred{
+				Pid: int32(t.ThreadGroup().ID()),
+				Uid: uint32(tcred.EffectiveKUID.In(tcred.UserNamespace).OrOverflow()),
+				Gid: uint32(tcred.EffectiveKGID.In(tcred.UserNamespace).OrOverflow()),
+			}, nil
+
+		case linux.SO_PASSCRED:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var v tcpip.PasscredOption
+			if err := ep.GetSockOpt(&v); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			return int32(v), nil
+
+		case linux.SO_SNDBUF:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var size tcpip.SendBufferSizeOption
+			if err := ep.GetSockOpt(&size); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			if size > math.MaxInt32 {
+				size = math.MaxInt32
+			}
+
+			return int32(size), nil
+
+		case linux.SO_RCVBUF:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var size tcpip.ReceiveBufferSizeOption
+			if err := ep.GetSockOpt(&size); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			if size > math.MaxInt32 {
+				size = math.MaxInt32
+			}
+
+			return int32(size), nil
+
+		case linux.SO_REUSEADDR:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var v tcpip.ReuseAddressOption
+			if err := ep.GetSockOpt(&v); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			return int32(v), nil
+
+		case linux.SO_KEEPALIVE:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+			return int32(0), nil
+
+		case linux.SO_LINGER:
+			if outLen < syscall.SizeofLinger {
+				return nil, syserr.ErrInvalidArgument
+			}
+			return syscall.Linger{}, nil
+
+		case linux.SO_RCVTIMEO:
+			if outLen < linux.SizeOfTimeval {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			return linux.NsecToTimeval(s.RecvTimeout()), nil
+		}
+
+	case syscall.SOL_TCP:
+		switch name {
+		case syscall.TCP_NODELAY:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var v tcpip.NoDelayOption
+			if err := ep.GetSockOpt(&v); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			return int32(v), nil
+
+		case syscall.TCP_INFO:
+			var v tcpip.TCPInfoOption
+			if err := ep.GetSockOpt(&v); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			// TODO: Translate fields once they are added to
+			// tcpip.TCPInfoOption.
+			info := linux.TCPInfo{}
+
+			// Linux truncates the output binary to outLen.
+			ib := binary.Marshal(nil, usermem.ByteOrder, &info)
+			if len(ib) > outLen {
+				ib = ib[:outLen]
+			}
+
+			return ib, nil
+		}
+
+	case syscall.SOL_IPV6:
+		switch name {
+		case syscall.IPV6_V6ONLY:
+			if outLen < sizeOfInt32 {
+				return nil, syserr.ErrInvalidArgument
+			}
+
+			var v tcpip.V6OnlyOption
+			if err := ep.GetSockOpt(&v); err != nil {
+				return nil, syserr.TranslateNetstackError(err)
+			}
+
+			return int32(v), nil
+		}
+	}
+
+	return nil, syserr.ErrProtocolNotAvailable
+}
+
+// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error {
+	return SetSockOpt(t, s, s.Endpoint, level, name, optVal)
+}
+
+// SetSockOpt can be used to implement the linux syscall setsockopt(2) for
+// sockets backed by a commonEndpoint.
+func SetSockOpt(t *kernel.Task, s socket.Socket, ep commonEndpoint, level int, name int, optVal []byte) *syserr.Error {
+	switch level {
+	case syscall.SOL_SOCKET:
+		switch name {
+		case linux.SO_SNDBUF:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.SendBufferSizeOption(v)))
+
+		case linux.SO_RCVBUF:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReceiveBufferSizeOption(v)))
+
+		case linux.SO_REUSEADDR:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.ReuseAddressOption(v)))
+
+		case linux.SO_PASSCRED:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.PasscredOption(v)))
+
+		case linux.SO_RCVTIMEO:
+			if len(optVal) < linux.SizeOfTimeval {
+				return syserr.ErrInvalidArgument
+			}
+
+			var v linux.Timeval
+			binary.Unmarshal(optVal[:linux.SizeOfTimeval], usermem.ByteOrder, &v)
+			s.SetRecvTimeout(v.ToNsecCapped())
+			return nil
+		}
+
+	case syscall.SOL_TCP:
+		switch name {
+		case syscall.TCP_NODELAY:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.NoDelayOption(v)))
+		}
+	case syscall.SOL_IPV6:
+		switch name {
+		case syscall.IPV6_V6ONLY:
+			if len(optVal) < sizeOfInt32 {
+				return syserr.ErrInvalidArgument
+			}
+
+			v := usermem.ByteOrder.Uint32(optVal)
+			return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.V6OnlyOption(v)))
+		}
+	}
+
+	// FIXME: Disallow IP-level multicast group options by
+	// default. These will need to be supported by appropriately plumbing
+	// the level through to the network stack (if at all). However, we
+	// still allow setting TTL, and multicast-enable/disable type options.
+	if level == 0 {
+		const (
+			_IP_ADD_MEMBERSHIP = 35
+			_MCAST_JOIN_GROUP  = 42
+		)
+		if name == _IP_ADD_MEMBERSHIP || name == _MCAST_JOIN_GROUP {
+			return syserr.ErrInvalidArgument
+		}
+	}
+
+	// Default to the old behavior; hand off to network stack.
+	return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{}))
+}
+
+// isLinkLocal determines if the given IPv6 address is link-local. This is the
+// case when it has the fe80::/10 prefix. This check is used to determine when
+// the NICID is relevant for a given IPv6 address.
+func isLinkLocal(addr tcpip.Address) bool {
+	return len(addr) >= 2 && addr[0] == 0xfe && addr[1]&0xc0 == 0x80
+}
+
+// ConvertAddress converts the given address to a native format.
+func ConvertAddress(family int, addr tcpip.FullAddress) (interface{}, uint32) {
+	switch family {
+	case linux.AF_UNIX:
+		var out linux.SockAddrUnix
+		out.Family = linux.AF_UNIX
+		for i := 0; i < len([]byte(addr.Addr)); i++ {
+			out.Path[i] = int8(addr.Addr[i])
+		}
+		// Linux just returns the header for empty addresses.
+		if len(addr.Addr) == 0 {
+			return out, 2
+		}
+		// Linux returns the used length of the address struct (including the
+		// null terminator) for filesystem paths. The Family field is 2 bytes.
+		// It is sometimes allowed to exclude the null terminator if the
+		// address length is the max. Abstract paths always return the full
+		// length.
+		if out.Path[0] == 0 || len([]byte(addr.Addr)) == len(out.Path) {
+			return out, uint32(binary.Size(out))
+		}
+		return out, uint32(3 + len(addr.Addr))
+	case linux.AF_INET:
+		var out linux.SockAddrInet
+		copy(out.Addr[:], addr.Addr)
+		out.Family = linux.AF_INET
+		out.Port = htons(addr.Port)
+		return out, uint32(binary.Size(out))
+	case linux.AF_INET6:
+		var out linux.SockAddrInet6
+		if len(addr.Addr) == 4 {
+			// Copy address is v4-mapped format.
+			copy(out.Addr[12:], addr.Addr)
+			out.Addr[10] = 0xff
+			out.Addr[11] = 0xff
+		} else {
+			copy(out.Addr[:], addr.Addr)
+		}
+		out.Family = linux.AF_INET6
+		out.Port = htons(addr.Port)
+		if isLinkLocal(addr.Addr) {
+			out.Scope_id = uint32(addr.NIC)
+		}
+		return out, uint32(binary.Size(out))
+	default:
+		return nil, 0
+	}
+}
+
+// GetSockName implements the linux syscall getsockname(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr, err := s.Endpoint.GetLocalAddress()
+	if err != nil {
+		return nil, 0, syserr.TranslateNetstackError(err)
+	}
+
+	a, l := ConvertAddress(s.family, addr)
+	return a, l, nil
+}
+
+// GetPeerName implements the linux syscall getpeername(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr, err := s.Endpoint.GetRemoteAddress()
+	if err != nil {
+		return nil, 0, syserr.TranslateNetstackError(err)
+	}
+
+	a, l := ConvertAddress(s.family, addr)
+	return a, l, nil
+}
+
+// coalescingRead is the fast path for non-blocking, non-peek, stream-based
+// case. It coalesces as many packets as possible before returning to the
+// caller.
+func (s *SocketOperations) coalescingRead(ctx context.Context, dst usermem.IOSequence, discard bool) (int, *syserr.Error) {
+	var err *syserr.Error
+	var copied int
+
+	// Copy as many views as possible into the user-provided buffer.
+	for dst.NumBytes() != 0 {
+		err = s.fetchReadView()
+		if err != nil {
+			break
+		}
+
+		var n int
+		var e error
+		if discard {
+			n = len(s.readView)
+			if int64(n) > dst.NumBytes() {
+				n = int(dst.NumBytes())
+			}
+		} else {
+			n, e = dst.CopyOut(ctx, s.readView)
+		}
+		copied += n
+		s.readView.TrimFront(n)
+		dst = dst.DropFirst(n)
+		if e != nil {
+			err = syserr.FromError(e)
+			break
+		}
+	}
+
+	// If we managed to copy something, we must deliver it.
+	if copied > 0 {
+		return copied, nil
+	}
+
+	return 0, err
+}
+
+// nonBlockingRead issues a non-blocking read.
+func (s *SocketOperations) nonBlockingRead(ctx context.Context, dst usermem.IOSequence, peek, trunc, senderRequested bool) (int, interface{}, uint32, *syserr.Error) {
+	isPacket := s.isPacketBased()
+
+	// Fast path for regular reads from stream (e.g., TCP) endpoints. Note
+	// that senderRequested is ignored for stream sockets.
+	if !peek && !isPacket {
+		// TCP sockets discard the data if MSG_TRUNC is set.
+		//
+		// This behavior is documented in man 7 tcp:
+		// Since version 2.4, Linux supports the use of MSG_TRUNC in the flags
+		// argument of recv(2) (and recvmsg(2)). This flag causes the received
+		// bytes of data to be discarded, rather than passed back in a
+		// caller-supplied  buffer.
+		s.readMu.Lock()
+		n, err := s.coalescingRead(ctx, dst, trunc)
+		s.readMu.Unlock()
+		return n, nil, 0, err
+	}
+
+	s.readMu.Lock()
+	defer s.readMu.Unlock()
+
+	if err := s.fetchReadView(); err != nil {
+		return 0, nil, 0, err
+	}
+
+	if !isPacket && peek && trunc {
+		// MSG_TRUNC with MSG_PEEK on a TCP socket returns the
+		// amount that could be read.
+		var rql tcpip.ReceiveQueueSizeOption
+		if err := s.Endpoint.GetSockOpt(&rql); err != nil {
+			return 0, nil, 0, syserr.TranslateNetstackError(err)
+		}
+		available := len(s.readView) + int(rql)
+		bufLen := int(dst.NumBytes())
+		if available < bufLen {
+			return available, nil, 0, nil
+		}
+		return bufLen, nil, 0, nil
+	}
+
+	n, err := dst.CopyOut(ctx, s.readView)
+	var addr interface{}
+	var addrLen uint32
+	if isPacket && senderRequested {
+		addr, addrLen = ConvertAddress(s.family, s.sender)
+	}
+
+	if peek {
+		if l := len(s.readView); trunc && l > n {
+			// isPacket must be true.
+			return l, addr, addrLen, syserr.FromError(err)
+		}
+
+		if isPacket || err != nil {
+			return int(n), addr, addrLen, syserr.FromError(err)
+		}
+
+		// We need to peek beyond the first message.
+		dst = dst.DropFirst(n)
+		num, err := dst.CopyOutFrom(ctx, safemem.FromVecReaderFunc{func(dsts [][]byte) (int64, error) {
+			n, err := s.Endpoint.Peek(dsts)
+			if err != nil {
+				return int64(n), syserr.TranslateNetstackError(err).ToError()
+			}
+			return int64(n), nil
+		}})
+		n += int(num)
+		if err == syserror.ErrWouldBlock && n > 0 {
+			// We got some data, so no need to return an error.
+			err = nil
+		}
+		return int(n), nil, 0, syserr.FromError(err)
+	}
+
+	var msgLen int
+	if isPacket {
+		msgLen = len(s.readView)
+		s.readView = nil
+	} else {
+		msgLen = int(n)
+		s.readView.TrimFront(int(n))
+	}
+
+	if trunc {
+		return msgLen, addr, addrLen, syserr.FromError(err)
+	}
+
+	return int(n), addr, addrLen, syserr.FromError(err)
+}
+
+// RecvMsg implements the linux syscall recvmsg(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, senderAddr interface{}, senderAddrLen uint32, controlMessages unix.ControlMessages, err *syserr.Error) {
+	trunc := flags&linux.MSG_TRUNC != 0
+
+	peek := flags&linux.MSG_PEEK != 0
+	if senderRequested && !s.isPacketBased() {
+		// Stream sockets ignore the sender address.
+		senderRequested = false
+	}
+	n, senderAddr, senderAddrLen, err = s.nonBlockingRead(t, dst, peek, trunc, senderRequested)
+	if err != syserr.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		return
+	}
+
+	// We'll have to block. Register for notifications and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		n, senderAddr, senderAddrLen, err = s.nonBlockingRead(t, dst, peek, trunc, senderRequested)
+		if err != syserr.ErrWouldBlock {
+			return
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syserror.ETIMEDOUT {
+				return 0, nil, 0, unix.ControlMessages{}, syserr.ErrTryAgain
+			}
+			return 0, nil, 0, unix.ControlMessages{}, syserr.FromError(err)
+		}
+	}
+}
+
+// SendMsg implements the linux syscall sendmsg(2) for sockets backed by
+// tcpip.Endpoint.
+func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	// Reject control messages.
+	if !controlMessages.Empty() {
+		return 0, syserr.ErrInvalidArgument
+	}
+
+	var addr *tcpip.FullAddress
+	if len(to) > 0 {
+		addrBuf, err := GetAddress(s.family, to)
+		if err != nil {
+			return 0, err
+		}
+
+		addr = &addrBuf
+	}
+
+	v := buffer.NewView(int(src.NumBytes()))
+
+	// Copy all the data into the buffer.
+	if _, err := src.CopyIn(t, v); err != nil {
+		return 0, syserr.FromError(err)
+	}
+
+	opts := tcpip.WriteOptions{
+		To:          addr,
+		More:        flags&linux.MSG_MORE != 0,
+		EndOfRecord: flags&linux.MSG_EOR != 0,
+	}
+
+	n, err := s.Endpoint.Write(tcpip.SlicePayload(v), opts)
+	if err != tcpip.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		return int(n), syserr.TranslateNetstackError(err)
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+
+	v.TrimFront(int(n))
+	total := n
+	for {
+		n, err = s.Endpoint.Write(tcpip.SlicePayload(v), opts)
+		v.TrimFront(int(n))
+		total += n
+		if err != tcpip.ErrWouldBlock {
+			return int(total), syserr.TranslateNetstackError(err)
+		}
+
+		if err := t.Block(ch); err != nil {
+			return int(total), syserr.FromError(err)
+		}
+	}
+}
+
+// interfaceIoctl implements interface requests.
+func (s *SocketOperations) interfaceIoctl(ctx context.Context, io usermem.IO, arg int, ifr *linux.IFReq) *syserr.Error {
+	var (
+		iface inet.Interface
+		index int32
+		found bool
+	)
+
+	// Find the relevant device.
+	for index, iface = range s.stack.Interfaces() {
+		if iface.Name == ifr.Name() {
+			found = true
+			break
+		}
+	}
+	if !found {
+		return syserr.ErrNoDevice
+	}
+
+	switch arg {
+	case syscall.SIOCGIFINDEX:
+		// Copy out the index to the data.
+		usermem.ByteOrder.PutUint32(ifr.Data[:], uint32(index))
+
+	case syscall.SIOCGIFHWADDR:
+		// Copy the hardware address out.
+		ifr.Data[0] = 6 // IEEE802.2 arp type.
+		ifr.Data[1] = 0
+		n := copy(ifr.Data[2:], iface.Addr)
+		for i := 2 + n; i < len(ifr.Data); i++ {
+			ifr.Data[i] = 0 // Clear padding.
+		}
+		usermem.ByteOrder.PutUint16(ifr.Data[:2], uint16(n))
+
+	case syscall.SIOCGIFFLAGS:
+		// TODO: Implement. For now, return only that the
+		// device is up so that ifconfig prints it.
+		usermem.ByteOrder.PutUint16(ifr.Data[:2], linux.IFF_UP)
+
+	case syscall.SIOCGIFADDR:
+		// Copy the IPv4 address out.
+		for _, addr := range s.stack.InterfaceAddrs()[index] {
+			// This ioctl is only compatible with AF_INET addresses.
+			if addr.Family != linux.AF_INET {
+				continue
+			}
+			copy(ifr.Data[4:8], addr.Addr)
+			break
+		}
+
+	case syscall.SIOCGIFMETRIC:
+		// Gets the metric of the device. As per netdevice(7), this
+		// always just sets ifr_metric to 0.
+		usermem.ByteOrder.PutUint32(ifr.Data[:4], 0)
+	case syscall.SIOCGIFMTU:
+		// Gets the MTU of the device.
+		// TODO: Implement.
+
+	case syscall.SIOCGIFMAP:
+		// Gets the hardware parameters of the device.
+		// TODO: Implement.
+
+	case syscall.SIOCGIFTXQLEN:
+		// Gets the transmit queue length of the device.
+		// TODO: Implement.
+
+	case syscall.SIOCGIFDSTADDR:
+		// Gets the destination address of a point-to-point device.
+		// TODO: Implement.
+
+	case syscall.SIOCGIFBRDADDR:
+		// Gets the broadcast address of a device.
+		// TODO: Implement.
+
+	case syscall.SIOCGIFNETMASK:
+		// Gets the network mask of a device.
+		// TODO: Implement.
+
+	default:
+		// Not a valid call.
+		return syserr.ErrInvalidArgument
+	}
+
+	return nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *SocketOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch arg := int(args[1].Int()); arg {
+	case syscall.SIOCGIFFLAGS,
+		syscall.SIOCGIFADDR,
+		syscall.SIOCGIFBRDADDR,
+		syscall.SIOCGIFDSTADDR,
+		syscall.SIOCGIFHWADDR,
+		syscall.SIOCGIFINDEX,
+		syscall.SIOCGIFMAP,
+		syscall.SIOCGIFMETRIC,
+		syscall.SIOCGIFMTU,
+		syscall.SIOCGIFNETMASK,
+		syscall.SIOCGIFTXQLEN:
+
+		var ifr linux.IFReq
+		if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &ifr, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, err
+		}
+		if err := s.interfaceIoctl(ctx, io, arg, &ifr); err != nil {
+			return 0, err.ToError()
+		}
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &ifr, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	case syscall.SIOCGIFCONF:
+		// Return a list of interface addresses or the buffer size
+		// necessary to hold the list.
+		var ifc linux.IFConf
+		if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &ifc, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, err
+		}
+
+		if err := s.ifconfIoctl(ctx, io, &ifc); err != nil {
+			return 0, err
+		}
+
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), ifc, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+
+		return 0, err
+	}
+
+	return Ioctl(ctx, s.Endpoint, io, args)
+}
+
+// ifconfIoctl populates a struct ifconf for the SIOCGIFCONF ioctl.
+func (s *SocketOperations) ifconfIoctl(ctx context.Context, io usermem.IO, ifc *linux.IFConf) error {
+	// If Ptr is NULL, return the necessary buffer size via Len.
+	// Otherwise, write up to Len bytes starting at Ptr containing ifreq
+	// structs.
+	if ifc.Ptr == 0 {
+		ifc.Len = int32(len(s.stack.Interfaces())) * int32(linux.SizeOfIFReq)
+		return nil
+	}
+
+	max := ifc.Len
+	ifc.Len = 0
+	for key, ifaceAddrs := range s.stack.InterfaceAddrs() {
+		iface := s.stack.Interfaces()[key]
+		for _, ifaceAddr := range ifaceAddrs {
+			// Don't write past the end of the buffer.
+			if ifc.Len+int32(linux.SizeOfIFReq) > max {
+				break
+			}
+			if ifaceAddr.Family != linux.AF_INET {
+				continue
+			}
+
+			// Populate ifr.ifr_addr.
+			ifr := linux.IFReq{}
+			ifr.SetName(iface.Name)
+			usermem.ByteOrder.PutUint16(ifr.Data[0:2], uint16(ifaceAddr.Family))
+			usermem.ByteOrder.PutUint16(ifr.Data[2:4], 0)
+			copy(ifr.Data[4:8], ifaceAddr.Addr[:4])
+
+			// Copy the ifr to userspace.
+			dst := uintptr(ifc.Ptr) + uintptr(ifc.Len)
+			ifc.Len += int32(linux.SizeOfIFReq)
+			if _, err := usermem.CopyObjectOut(ctx, io, usermem.Addr(dst), ifr, usermem.IOOpts{
+				AddressSpaceActive: true,
+			}); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl for sockets backed by a
+// commonEndpoint.
+func Ioctl(ctx context.Context, ep commonEndpoint, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// Switch on ioctl request.
+	switch int(args[1].Int()) {
+	case linux.TIOCINQ:
+		var v tcpip.ReceiveQueueSizeOption
+		if err := ep.GetSockOpt(&v); err != nil {
+			return 0, syserr.TranslateNetstackError(err).ToError()
+		}
+
+		if v > math.MaxInt32 {
+			v = math.MaxInt32
+		}
+		// Copy result to user-space.
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+
+	case linux.TIOCOUTQ:
+		var v tcpip.SendQueueSizeOption
+		if err := ep.GetSockOpt(&v); err != nil {
+			return 0, syserr.TranslateNetstackError(err).ToError()
+		}
+
+		if v > math.MaxInt32 {
+			v = math.MaxInt32
+		}
+
+		// Copy result to user-space.
+		_, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+	}
+
+	return 0, syserror.ENOTTY
+}
diff --git a/pkg/sentry/socket/epsocket/provider.go b/pkg/sentry/socket/epsocket/provider.go
new file mode 100644
index 000000000..5616435b3
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/provider.go
@@ -0,0 +1,113 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epsocket
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/network/ipv4"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/network/ipv6"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/tcp"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/udp"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// provider is an inet socket provider.
+type provider struct {
+	family   int
+	netProto tcpip.NetworkProtocolNumber
+}
+
+// GetTransportProtocol figures out transport protocol. Currently only TCP and
+// UDP are supported.
+func GetTransportProtocol(stype unix.SockType, protocol int) (tcpip.TransportProtocolNumber, *syserr.Error) {
+	switch stype {
+	case linux.SOCK_STREAM:
+		if protocol != 0 && protocol != syscall.IPPROTO_TCP {
+			return 0, syserr.ErrInvalidArgument
+		}
+		return tcp.ProtocolNumber, nil
+
+	case linux.SOCK_DGRAM:
+		if protocol != 0 && protocol != syscall.IPPROTO_UDP {
+			return 0, syserr.ErrInvalidArgument
+		}
+		return udp.ProtocolNumber, nil
+
+	default:
+		return 0, syserr.ErrInvalidArgument
+	}
+}
+
+// Socket creates a new socket object for the AF_INET or AF_INET6 family.
+func (p *provider) Socket(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Fail right away if we don't have a stack.
+	stack := t.NetworkContext()
+	if stack == nil {
+		// Don't propagate an error here. Instead, allow the socket
+		// code to continue searching for another provider.
+		return nil, nil
+	}
+	eps, ok := stack.(*Stack)
+	if !ok {
+		return nil, nil
+	}
+
+	// Figure out the transport protocol.
+	transProto, err := GetTransportProtocol(stype, protocol)
+	if err != nil {
+		return nil, err
+	}
+
+	// Create the endpoint.
+	wq := &waiter.Queue{}
+	ep, e := eps.Stack.NewEndpoint(transProto, p.netProto, wq)
+	if e != nil {
+		return nil, syserr.TranslateNetstackError(e)
+	}
+
+	return New(t, p.family, stype, wq, ep), nil
+}
+
+// Pair just returns nil sockets (not supported).
+func (*provider) Pair(*kernel.Task, unix.SockType, int) (*fs.File, *fs.File, *syserr.Error) {
+	return nil, nil, nil
+}
+
+// init registers socket providers for AF_INET and AF_INET6.
+func init() {
+	// Providers backed by netstack.
+	p := []provider{
+		{
+			family:   linux.AF_INET,
+			netProto: ipv4.ProtocolNumber,
+		},
+		{
+			family:   linux.AF_INET6,
+			netProto: ipv6.ProtocolNumber,
+		},
+	}
+
+	for i := range p {
+		socket.RegisterProvider(p[i].family, &p[i])
+	}
+}
diff --git a/pkg/sentry/socket/epsocket/save_restore.go b/pkg/sentry/socket/epsocket/save_restore.go
new file mode 100644
index 000000000..2613f90de
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/save_restore.go
@@ -0,0 +1,27 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epsocket
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/stack"
+)
+
+// afterLoad is invoked by stateify.
+func (s *Stack) afterLoad() {
+	s.Stack = stack.StackFromEnv // FIXME
+	if s.Stack == nil {
+		panic("can't restore without netstack/tcpip/stack.Stack")
+	}
+}
diff --git a/pkg/sentry/socket/epsocket/stack.go b/pkg/sentry/socket/epsocket/stack.go
new file mode 100644
index 000000000..ec1d96ccb
--- /dev/null
+++ b/pkg/sentry/socket/epsocket/stack.go
@@ -0,0 +1,132 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package epsocket
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/network/ipv4"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/network/ipv6"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/stack"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/tcp"
+)
+
+// Stack implements inet.Stack for netstack/tcpip/stack.Stack.
+type Stack struct {
+	Stack *stack.Stack `state:"manual"`
+}
+
+// SupportsIPv6 implements Stack.SupportsIPv6.
+func (s *Stack) SupportsIPv6() bool {
+	return s.Stack.CheckNetworkProtocol(ipv6.ProtocolNumber)
+}
+
+// Interfaces implements inet.Stack.Interfaces.
+func (s *Stack) Interfaces() map[int32]inet.Interface {
+	is := make(map[int32]inet.Interface)
+	for id, ni := range s.Stack.NICInfo() {
+		is[int32(id)] = inet.Interface{
+			Name: ni.Name,
+			Addr: []byte(ni.LinkAddress),
+			// TODO: Other fields.
+		}
+	}
+	return is
+}
+
+// InterfaceAddrs implements inet.Stack.InterfaceAddrs.
+func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr {
+	nicAddrs := make(map[int32][]inet.InterfaceAddr)
+	for id, ni := range s.Stack.NICInfo() {
+		var addrs []inet.InterfaceAddr
+		for _, a := range ni.ProtocolAddresses {
+			var family uint8
+			switch a.Protocol {
+			case ipv4.ProtocolNumber:
+				family = linux.AF_INET
+			case ipv6.ProtocolNumber:
+				family = linux.AF_INET6
+			default:
+				log.Warningf("Unknown network protocol in %+v", a)
+				continue
+			}
+
+			addrs = append(addrs, inet.InterfaceAddr{
+				Family:    family,
+				PrefixLen: uint8(len(a.Address) * 8),
+				Addr:      []byte(a.Address),
+				// TODO: Other fields.
+			})
+		}
+		nicAddrs[int32(id)] = addrs
+	}
+	return nicAddrs
+}
+
+// TCPReceiveBufferSize implements inet.Stack.TCPReceiveBufferSize.
+func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) {
+	var rs tcp.ReceiveBufferSizeOption
+	err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &rs)
+	return inet.TCPBufferSize{
+		Min:     rs.Min,
+		Default: rs.Default,
+		Max:     rs.Max,
+	}, syserr.TranslateNetstackError(err).ToError()
+}
+
+// SetTCPReceiveBufferSize implements inet.Stack.SetTCPReceiveBufferSize.
+func (s *Stack) SetTCPReceiveBufferSize(size inet.TCPBufferSize) error {
+	rs := tcp.ReceiveBufferSizeOption{
+		Min:     size.Min,
+		Default: size.Default,
+		Max:     size.Max,
+	}
+	return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, rs)).ToError()
+}
+
+// TCPSendBufferSize implements inet.Stack.TCPSendBufferSize.
+func (s *Stack) TCPSendBufferSize() (inet.TCPBufferSize, error) {
+	var ss tcp.SendBufferSizeOption
+	err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &ss)
+	return inet.TCPBufferSize{
+		Min:     ss.Min,
+		Default: ss.Default,
+		Max:     ss.Max,
+	}, syserr.TranslateNetstackError(err).ToError()
+}
+
+// SetTCPSendBufferSize implements inet.Stack.SetTCPSendBufferSize.
+func (s *Stack) SetTCPSendBufferSize(size inet.TCPBufferSize) error {
+	ss := tcp.SendBufferSizeOption{
+		Min:     size.Min,
+		Default: size.Default,
+		Max:     size.Max,
+	}
+	return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, ss)).ToError()
+}
+
+// TCPSACKEnabled implements inet.Stack.TCPSACKEnabled.
+func (s *Stack) TCPSACKEnabled() (bool, error) {
+	var sack tcp.SACKEnabled
+	err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &sack)
+	return bool(sack), syserr.TranslateNetstackError(err).ToError()
+}
+
+// SetTCPSACKEnabled implements inet.Stack.SetTCPSACKEnabled.
+func (s *Stack) SetTCPSACKEnabled(enabled bool) error {
+	return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.SACKEnabled(enabled))).ToError()
+}
diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD
new file mode 100644
index 000000000..60ec265ba
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/BUILD
@@ -0,0 +1,53 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "hostinet_state",
+    srcs = [
+        "save_restore.go",
+        "socket.go",
+        "stack.go",
+    ],
+    out = "hostinet_autogen_state.go",
+    package = "hostinet",
+)
+
+go_library(
+    name = "hostinet",
+    srcs = [
+        "device.go",
+        "hostinet.go",
+        "hostinet_autogen_state.go",
+        "save_restore.go",
+        "socket.go",
+        "socket_unsafe.go",
+        "stack.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/hostinet",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/log",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+        "//pkg/waiter/fdnotifier",
+    ],
+)
diff --git a/pkg/sentry/socket/hostinet/device.go b/pkg/sentry/socket/hostinet/device.go
new file mode 100644
index 000000000..a9a673316
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/device.go
@@ -0,0 +1,19 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostinet
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+var socketDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/socket/hostinet/hostinet.go b/pkg/sentry/socket/hostinet/hostinet.go
new file mode 100644
index 000000000..67c6c8066
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/hostinet.go
@@ -0,0 +1,17 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package hostinet implements AF_INET and AF_INET6 sockets using the host's
+// network stack.
+package hostinet
diff --git a/pkg/sentry/socket/hostinet/save_restore.go b/pkg/sentry/socket/hostinet/save_restore.go
new file mode 100644
index 000000000..0821a794a
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/save_restore.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostinet
+
+// beforeSave is invoked by stateify.
+func (*socketOperations) beforeSave() {
+	panic("host.socketOperations is not savable")
+}
diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go
new file mode 100644
index 000000000..defa3db2c
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/socket.go
@@ -0,0 +1,562 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostinet
+
+import (
+	"fmt"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+	"gvisor.googlesource.com/gvisor/pkg/waiter/fdnotifier"
+)
+
+const (
+	sizeofInt32 = 4
+
+	// sizeofSockaddr is the size in bytes of the largest sockaddr type
+	// supported by this package.
+	sizeofSockaddr = syscall.SizeofSockaddrInet6 // sizeof(sockaddr_in6) > sizeof(sockaddr_in)
+)
+
+// socketOperations implements fs.FileOperations and socket.Socket for a socket
+// implemented using a host socket.
+type socketOperations struct {
+	socket.ReceiveTimeout
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+
+	fd    int // must be O_NONBLOCK
+	queue waiter.Queue
+}
+
+func newSocketFile(ctx context.Context, fd int, nonblock bool) (*fs.File, *syserr.Error) {
+	s := &socketOperations{fd: fd}
+	if err := fdnotifier.AddFD(int32(fd), &s.queue); err != nil {
+		return nil, syserr.FromError(err)
+	}
+	dirent := socket.NewDirent(ctx, socketDevice)
+	return fs.NewFile(ctx, dirent, fs.FileFlags{NonBlocking: nonblock, Read: true, Write: true}, s), nil
+}
+
+// Release implements fs.FileOperations.Release.
+func (s *socketOperations) Release() {
+	fdnotifier.RemoveFD(int32(s.fd))
+	syscall.Close(s.fd)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *socketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return fdnotifier.NonBlockingPoll(int32(s.fd), mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *socketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.queue.EventRegister(e, mask)
+	fdnotifier.UpdateFD(int32(s.fd))
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *socketOperations) EventUnregister(e *waiter.Entry) {
+	s.queue.EventUnregister(e)
+	fdnotifier.UpdateFD(int32(s.fd))
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *socketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	n, err := dst.CopyOutFrom(ctx, safemem.ReaderFunc(func(dsts safemem.BlockSeq) (uint64, error) {
+		// Refuse to do anything if any part of dst.Addrs was unusable.
+		if uint64(dst.NumBytes()) != dsts.NumBytes() {
+			return 0, nil
+		}
+		if dsts.IsEmpty() {
+			return 0, nil
+		}
+		if dsts.NumBlocks() == 1 {
+			// Skip allocating []syscall.Iovec.
+			n, err := syscall.Read(s.fd, dsts.Head().ToSlice())
+			if err != nil {
+				return 0, translateIOSyscallError(err)
+			}
+			return uint64(n), nil
+		}
+		return readv(s.fd, iovecsFromBlockSeq(dsts))
+	}))
+	return int64(n), err
+}
+
+// Write implements fs.FileOperations.Write.
+func (s *socketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	n, err := src.CopyInTo(ctx, safemem.WriterFunc(func(srcs safemem.BlockSeq) (uint64, error) {
+		// Refuse to do anything if any part of src.Addrs was unusable.
+		if uint64(src.NumBytes()) != srcs.NumBytes() {
+			return 0, nil
+		}
+		if srcs.IsEmpty() {
+			return 0, nil
+		}
+		if srcs.NumBlocks() == 1 {
+			// Skip allocating []syscall.Iovec.
+			n, err := syscall.Write(s.fd, srcs.Head().ToSlice())
+			if err != nil {
+				return 0, translateIOSyscallError(err)
+			}
+			return uint64(n), nil
+		}
+		return writev(s.fd, iovecsFromBlockSeq(srcs))
+	}))
+	return int64(n), err
+}
+
+// Connect implements socket.Socket.Connect.
+func (s *socketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	if len(sockaddr) > sizeofSockaddr {
+		sockaddr = sockaddr[:sizeofSockaddr]
+	}
+
+	_, _, errno := syscall.Syscall(syscall.SYS_CONNECT, uintptr(s.fd), uintptr(firstBytePtr(sockaddr)), uintptr(len(sockaddr)))
+
+	if errno == 0 {
+		return nil
+	}
+	if errno != syscall.EINPROGRESS || !blocking {
+		return syserr.FromError(translateIOSyscallError(errno))
+	}
+
+	// "EINPROGRESS: The socket is nonblocking and the connection cannot be
+	// completed immediately. It is possible to select(2) or poll(2) for
+	// completion by selecting the socket for writing. After select(2)
+	// indicates writability, use getsockopt(2) to read the SO_ERROR option at
+	// level SOL-SOCKET to determine whether connect() completed successfully
+	// (SO_ERROR is zero) or unsuccessfully (SO_ERROR is one of the usual error
+	// codes listed here, explaining the reason for the failure)." - connect(2)
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+	if s.Readiness(waiter.EventOut)&waiter.EventOut == 0 {
+		if err := t.Block(ch); err != nil {
+			return syserr.FromError(err)
+		}
+	}
+	val, err := syscall.GetsockoptInt(s.fd, syscall.SOL_SOCKET, syscall.SO_ERROR)
+	if err != nil {
+		return syserr.FromError(err)
+	}
+	if val != 0 {
+		return syserr.FromError(syscall.Errno(uintptr(val)))
+	}
+	return nil
+}
+
+// Accept implements socket.Socket.Accept.
+func (s *socketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	var peerAddr []byte
+	var peerAddrlen uint32
+	var peerAddrPtr *byte
+	var peerAddrlenPtr *uint32
+	if peerRequested {
+		peerAddr = make([]byte, sizeofSockaddr)
+		peerAddrlen = uint32(len(peerAddr))
+		peerAddrPtr = &peerAddr[0]
+		peerAddrlenPtr = &peerAddrlen
+	}
+
+	// Conservatively ignore all flags specified by the application and add
+	// SOCK_NONBLOCK since socketOperations requires it.
+	fd, syscallErr := accept4(s.fd, peerAddrPtr, peerAddrlenPtr, syscall.SOCK_NONBLOCK)
+	if blocking {
+		var ch chan struct{}
+		for syscallErr == syserror.ErrWouldBlock {
+			if ch != nil {
+				if syscallErr = t.Block(ch); syscallErr != nil {
+					break
+				}
+			} else {
+				var e waiter.Entry
+				e, ch = waiter.NewChannelEntry(nil)
+				s.EventRegister(&e, waiter.EventIn)
+				defer s.EventUnregister(&e)
+			}
+			fd, syscallErr = accept4(s.fd, peerAddrPtr, peerAddrlenPtr, syscall.SOCK_NONBLOCK)
+		}
+	}
+
+	if peerRequested {
+		peerAddr = peerAddr[:peerAddrlen]
+	}
+	if syscallErr != nil {
+		return 0, peerAddr, peerAddrlen, syserr.FromError(syscallErr)
+	}
+
+	f, err := newSocketFile(t, fd, flags&syscall.SOCK_NONBLOCK != 0)
+	if err != nil {
+		syscall.Close(fd)
+		return 0, nil, 0, err
+	}
+	defer f.DecRef()
+
+	fdFlags := kernel.FDFlags{
+		CloseOnExec: flags&syscall.SOCK_CLOEXEC != 0,
+	}
+	kfd, kerr := t.FDMap().NewFDFrom(0, f, fdFlags, t.ThreadGroup().Limits())
+	return kfd, peerAddr, peerAddrlen, syserr.FromError(kerr)
+}
+
+// Bind implements socket.Socket.Bind.
+func (s *socketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	if len(sockaddr) > sizeofSockaddr {
+		sockaddr = sockaddr[:sizeofSockaddr]
+	}
+
+	_, _, errno := syscall.Syscall(syscall.SYS_BIND, uintptr(s.fd), uintptr(firstBytePtr(sockaddr)), uintptr(len(sockaddr)))
+	if errno != 0 {
+		return syserr.FromError(errno)
+	}
+	return nil
+}
+
+// Listen implements socket.Socket.Listen.
+func (s *socketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	return syserr.FromError(syscall.Listen(s.fd, backlog))
+}
+
+// Shutdown implements socket.Socket.Shutdown.
+func (s *socketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	switch how {
+	case syscall.SHUT_RD, syscall.SHUT_WR, syscall.SHUT_RDWR:
+		return syserr.FromError(syscall.Shutdown(s.fd, how))
+	default:
+		return syserr.ErrInvalidArgument
+	}
+}
+
+// GetSockOpt implements socket.Socket.GetSockOpt.
+func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outLen int) (interface{}, *syserr.Error) {
+	if outLen < 0 {
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	// Whitelist options and constrain option length.
+	var optlen int
+	switch level {
+	case syscall.SOL_IPV6:
+		switch name {
+		case syscall.IPV6_V6ONLY:
+			optlen = sizeofInt32
+		}
+	case syscall.SOL_SOCKET:
+		switch name {
+		case syscall.SO_ERROR, syscall.SO_KEEPALIVE, syscall.SO_SNDBUF, syscall.SO_RCVBUF, syscall.SO_REUSEADDR, syscall.SO_TYPE:
+			optlen = sizeofInt32
+		case syscall.SO_LINGER:
+			optlen = syscall.SizeofLinger
+		}
+	case syscall.SOL_TCP:
+		switch name {
+		case syscall.TCP_NODELAY:
+			optlen = sizeofInt32
+		case syscall.TCP_INFO:
+			optlen = int(linux.SizeOfTCPInfo)
+		}
+	}
+	if optlen == 0 {
+		return nil, syserr.ErrProtocolNotAvailable // ENOPROTOOPT
+	}
+	if outLen < optlen {
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	opt, err := getsockopt(s.fd, level, name, optlen)
+	if err != nil {
+		return nil, syserr.FromError(err)
+	}
+	return opt, nil
+}
+
+// SetSockOpt implements socket.Socket.SetSockOpt.
+func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error {
+	// Whitelist options and constrain option length.
+	var optlen int
+	switch level {
+	case syscall.SOL_IPV6:
+		switch name {
+		case syscall.IPV6_V6ONLY:
+			optlen = sizeofInt32
+		}
+	case syscall.SOL_SOCKET:
+		switch name {
+		case syscall.SO_SNDBUF, syscall.SO_RCVBUF, syscall.SO_REUSEADDR:
+			optlen = sizeofInt32
+		}
+	case syscall.SOL_TCP:
+		switch name {
+		case syscall.TCP_NODELAY:
+			optlen = sizeofInt32
+		}
+	}
+	if optlen == 0 {
+		// Pretend to accept socket options we don't understand. This seems
+		// dangerous, but it's what netstack does...
+		return nil
+	}
+	if len(opt) < optlen {
+		return syserr.ErrInvalidArgument
+	}
+	opt = opt[:optlen]
+
+	_, _, errno := syscall.Syscall6(syscall.SYS_SETSOCKOPT, uintptr(s.fd), uintptr(level), uintptr(name), uintptr(firstBytePtr(opt)), uintptr(len(opt)), 0)
+	if errno != 0 {
+		return syserr.FromError(errno)
+	}
+	return nil
+}
+
+// RecvMsg implements socket.Socket.RecvMsg.
+func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, interface{}, uint32, unix.ControlMessages, *syserr.Error) {
+	// Whitelist flags.
+	//
+	// FIXME: We can't support MSG_ERRQUEUE because it uses ancillary
+	// messages that netstack/tcpip/transport/unix doesn't understand. Kill the
+	// Socket interface's dependence on netstack.
+	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_PEEK|syscall.MSG_TRUNC) != 0 {
+		return 0, nil, 0, unix.ControlMessages{}, syserr.ErrInvalidArgument
+	}
+
+	var senderAddr []byte
+	if senderRequested {
+		senderAddr = make([]byte, sizeofSockaddr)
+	}
+
+	recvmsgToBlocks := safemem.ReaderFunc(func(dsts safemem.BlockSeq) (uint64, error) {
+		// Refuse to do anything if any part of dst.Addrs was unusable.
+		if uint64(dst.NumBytes()) != dsts.NumBytes() {
+			return 0, nil
+		}
+		if dsts.IsEmpty() {
+			return 0, nil
+		}
+
+		// We always do a non-blocking recv*().
+		sysflags := flags | syscall.MSG_DONTWAIT
+
+		if dsts.NumBlocks() == 1 {
+			// Skip allocating []syscall.Iovec.
+			return recvfrom(s.fd, dsts.Head().ToSlice(), sysflags, &senderAddr)
+		}
+
+		iovs := iovecsFromBlockSeq(dsts)
+		msg := syscall.Msghdr{
+			Iov:    &iovs[0],
+			Iovlen: uint64(len(iovs)),
+		}
+		if len(senderAddr) != 0 {
+			msg.Name = &senderAddr[0]
+			msg.Namelen = uint32(len(senderAddr))
+		}
+		n, err := recvmsg(s.fd, &msg, sysflags)
+		if err != nil {
+			return 0, err
+		}
+		senderAddr = senderAddr[:msg.Namelen]
+		return n, nil
+	})
+
+	var ch chan struct{}
+	n, err := dst.CopyOutFrom(t, recvmsgToBlocks)
+	if flags&syscall.MSG_DONTWAIT == 0 {
+		for err == syserror.ErrWouldBlock {
+			// We only expect blocking to come from the actual syscall, in which
+			// case it can't have returned any data.
+			if n != 0 {
+				panic(fmt.Sprintf("CopyOutFrom: got (%d, %v), wanted (0, %v)", n, err, err))
+			}
+			if ch != nil {
+				if err = t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+					break
+				}
+			} else {
+				var e waiter.Entry
+				e, ch = waiter.NewChannelEntry(nil)
+				s.EventRegister(&e, waiter.EventIn)
+				defer s.EventUnregister(&e)
+			}
+			n, err = dst.CopyOutFrom(t, recvmsgToBlocks)
+		}
+	}
+
+	return int(n), senderAddr, uint32(len(senderAddr)), unix.ControlMessages{}, syserr.FromError(err)
+}
+
+// SendMsg implements socket.Socket.SendMsg.
+func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	// Whitelist flags.
+	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_EOR|syscall.MSG_FASTOPEN|syscall.MSG_MORE|syscall.MSG_NOSIGNAL) != 0 {
+		return 0, syserr.ErrInvalidArgument
+	}
+
+	sendmsgFromBlocks := safemem.WriterFunc(func(srcs safemem.BlockSeq) (uint64, error) {
+		// Refuse to do anything if any part of src.Addrs was unusable.
+		if uint64(src.NumBytes()) != srcs.NumBytes() {
+			return 0, nil
+		}
+		if srcs.IsEmpty() {
+			return 0, nil
+		}
+
+		// We always do a non-blocking send*().
+		sysflags := flags | syscall.MSG_DONTWAIT
+
+		if srcs.NumBlocks() == 1 {
+			// Skip allocating []syscall.Iovec.
+			src := srcs.Head()
+			n, _, errno := syscall.Syscall6(syscall.SYS_SENDTO, uintptr(s.fd), src.Addr(), uintptr(src.Len()), uintptr(sysflags), uintptr(firstBytePtr(to)), uintptr(len(to)))
+			if errno != 0 {
+				return 0, translateIOSyscallError(errno)
+			}
+			return uint64(n), nil
+		}
+
+		iovs := iovecsFromBlockSeq(srcs)
+		msg := syscall.Msghdr{
+			Iov:    &iovs[0],
+			Iovlen: uint64(len(iovs)),
+		}
+		if len(to) != 0 {
+			msg.Name = &to[0]
+			msg.Namelen = uint32(len(to))
+		}
+		return sendmsg(s.fd, &msg, sysflags)
+	})
+
+	var ch chan struct{}
+	n, err := src.CopyInTo(t, sendmsgFromBlocks)
+	if flags&syscall.MSG_DONTWAIT == 0 {
+		for err == syserror.ErrWouldBlock {
+			// We only expect blocking to come from the actual syscall, in which
+			// case it can't have returned any data.
+			if n != 0 {
+				panic(fmt.Sprintf("CopyInTo: got (%d, %v), wanted (0, %v)", n, err, err))
+			}
+			if ch != nil {
+				if err = t.Block(ch); err != nil {
+					break
+				}
+			} else {
+				var e waiter.Entry
+				e, ch = waiter.NewChannelEntry(nil)
+				s.EventRegister(&e, waiter.EventOut)
+				defer s.EventUnregister(&e)
+			}
+			n, err = src.CopyInTo(t, sendmsgFromBlocks)
+		}
+	}
+
+	return int(n), syserr.FromError(err)
+}
+
+func iovecsFromBlockSeq(bs safemem.BlockSeq) []syscall.Iovec {
+	iovs := make([]syscall.Iovec, 0, bs.NumBlocks())
+	for ; !bs.IsEmpty(); bs = bs.Tail() {
+		b := bs.Head()
+		iovs = append(iovs, syscall.Iovec{
+			Base: &b.ToSlice()[0],
+			Len:  uint64(b.Len()),
+		})
+		// We don't need to care about b.NeedSafecopy(), because the host
+		// kernel will handle such address ranges just fine (by returning
+		// EFAULT).
+	}
+	return iovs
+}
+
+func translateIOSyscallError(err error) error {
+	if err == syscall.EAGAIN || err == syscall.EWOULDBLOCK {
+		return syserror.ErrWouldBlock
+	}
+	return err
+}
+
+type socketProvider struct {
+	family int
+}
+
+// Socket implements socket.Provider.Socket.
+func (p *socketProvider) Socket(t *kernel.Task, stypeflags unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Check that we are using the host network stack.
+	stack := t.NetworkContext()
+	if stack == nil {
+		return nil, nil
+	}
+	if _, ok := stack.(*Stack); !ok {
+		return nil, nil
+	}
+
+	// Only accept TCP and UDP.
+	stype := int(stypeflags) & linux.SOCK_TYPE_MASK
+	switch stype {
+	case syscall.SOCK_STREAM:
+		switch protocol {
+		case 0, syscall.IPPROTO_TCP:
+			// ok
+		default:
+			return nil, nil
+		}
+	case syscall.SOCK_DGRAM:
+		switch protocol {
+		case 0, syscall.IPPROTO_UDP:
+			// ok
+		default:
+			return nil, nil
+		}
+	default:
+		return nil, nil
+	}
+
+	// Conservatively ignore all flags specified by the application and add
+	// SOCK_NONBLOCK since socketOperations requires it. Pass a protocol of 0
+	// to simplify the syscall filters, since 0 and IPPROTO_* are equivalent.
+	fd, err := syscall.Socket(p.family, stype|syscall.SOCK_NONBLOCK, 0)
+	if err != nil {
+		return nil, syserr.FromError(err)
+	}
+	return newSocketFile(t, fd, stypeflags&syscall.SOCK_NONBLOCK != 0)
+}
+
+// Pair implements socket.Provider.Pair.
+func (p *socketProvider) Pair(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *fs.File, *syserr.Error) {
+	// Not supported by AF_INET/AF_INET6.
+	return nil, nil, nil
+}
+
+func init() {
+	for _, family := range []int{syscall.AF_INET, syscall.AF_INET6} {
+		socket.RegisterProvider(family, &socketProvider{family})
+	}
+}
diff --git a/pkg/sentry/socket/hostinet/socket_unsafe.go b/pkg/sentry/socket/hostinet/socket_unsafe.go
new file mode 100644
index 000000000..f8bb75636
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/socket_unsafe.go
@@ -0,0 +1,138 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostinet
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func firstBytePtr(bs []byte) unsafe.Pointer {
+	if bs == nil {
+		return nil
+	}
+	return unsafe.Pointer(&bs[0])
+}
+
+// Preconditions: len(dsts) != 0.
+func readv(fd int, dsts []syscall.Iovec) (uint64, error) {
+	n, _, errno := syscall.Syscall(syscall.SYS_READV, uintptr(fd), uintptr(unsafe.Pointer(&dsts[0])), uintptr(len(dsts)))
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	return uint64(n), nil
+}
+
+// Preconditions: len(srcs) != 0.
+func writev(fd int, srcs []syscall.Iovec) (uint64, error) {
+	n, _, errno := syscall.Syscall(syscall.SYS_WRITEV, uintptr(fd), uintptr(unsafe.Pointer(&srcs[0])), uintptr(len(srcs)))
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	return uint64(n), nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *socketOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	switch cmd := uintptr(args[1].Int()); cmd {
+	case syscall.TIOCINQ, syscall.TIOCOUTQ:
+		var val int32
+		if _, _, errno := syscall.Syscall(syscall.SYS_IOCTL, uintptr(s.fd), cmd, uintptr(unsafe.Pointer(&val))); errno != 0 {
+			return 0, translateIOSyscallError(errno)
+		}
+		var buf [4]byte
+		usermem.ByteOrder.PutUint32(buf[:], uint32(val))
+		_, err := io.CopyOut(ctx, args[2].Pointer(), buf[:], usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		return 0, err
+
+	default:
+		return 0, syserror.ENOTTY
+	}
+}
+
+func accept4(fd int, addr *byte, addrlen *uint32, flags int) (int, error) {
+	afd, _, errno := syscall.Syscall6(syscall.SYS_ACCEPT4, uintptr(fd), uintptr(unsafe.Pointer(addr)), uintptr(unsafe.Pointer(addrlen)), uintptr(flags), 0, 0)
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	return int(afd), nil
+}
+
+func getsockopt(fd int, level, name int, optlen int) ([]byte, error) {
+	opt := make([]byte, optlen)
+	optlen32 := int32(len(opt))
+	_, _, errno := syscall.Syscall6(syscall.SYS_GETSOCKOPT, uintptr(fd), uintptr(level), uintptr(name), uintptr(firstBytePtr(opt)), uintptr(unsafe.Pointer(&optlen32)), 0)
+	if errno != 0 {
+		return nil, errno
+	}
+	return opt[:optlen32], nil
+}
+
+// GetSockName implements socket.Socket.GetSockName.
+func (s *socketOperations) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr := make([]byte, sizeofSockaddr)
+	addrlen := uint32(len(addr))
+	_, _, errno := syscall.Syscall(syscall.SYS_GETSOCKNAME, uintptr(s.fd), uintptr(unsafe.Pointer(&addr[0])), uintptr(unsafe.Pointer(&addrlen)))
+	if errno != 0 {
+		return nil, 0, syserr.FromError(errno)
+	}
+	return addr[:addrlen], addrlen, nil
+}
+
+// GetPeerName implements socket.Socket.GetPeerName.
+func (s *socketOperations) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr := make([]byte, sizeofSockaddr)
+	addrlen := uint32(len(addr))
+	_, _, errno := syscall.Syscall(syscall.SYS_GETPEERNAME, uintptr(s.fd), uintptr(unsafe.Pointer(&addr[0])), uintptr(unsafe.Pointer(&addrlen)))
+	if errno != 0 {
+		return nil, 0, syserr.FromError(errno)
+	}
+	return addr[:addrlen], addrlen, nil
+}
+
+func recvfrom(fd int, dst []byte, flags int, from *[]byte) (uint64, error) {
+	fromLen := uint32(len(*from))
+	n, _, errno := syscall.Syscall6(syscall.SYS_RECVFROM, uintptr(fd), uintptr(firstBytePtr(dst)), uintptr(len(dst)), uintptr(flags), uintptr(firstBytePtr(*from)), uintptr(unsafe.Pointer(&fromLen)))
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	*from = (*from)[:fromLen]
+	return uint64(n), nil
+}
+
+func recvmsg(fd int, msg *syscall.Msghdr, flags int) (uint64, error) {
+	n, _, errno := syscall.Syscall(syscall.SYS_RECVMSG, uintptr(fd), uintptr(unsafe.Pointer(msg)), uintptr(flags))
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	return uint64(n), nil
+}
+
+func sendmsg(fd int, msg *syscall.Msghdr, flags int) (uint64, error) {
+	n, _, errno := syscall.Syscall(syscall.SYS_SENDMSG, uintptr(fd), uintptr(unsafe.Pointer(msg)), uintptr(flags))
+	if errno != 0 {
+		return 0, translateIOSyscallError(errno)
+	}
+	return uint64(n), nil
+}
diff --git a/pkg/sentry/socket/hostinet/stack.go b/pkg/sentry/socket/hostinet/stack.go
new file mode 100644
index 000000000..44c3b9a3f
--- /dev/null
+++ b/pkg/sentry/socket/hostinet/stack.go
@@ -0,0 +1,244 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package hostinet
+
+import (
+	"fmt"
+	"io/ioutil"
+	"os"
+	"strings"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+var defaultRecvBufSize = inet.TCPBufferSize{
+	Min:     4096,
+	Default: 87380,
+	Max:     6291456,
+}
+
+var defaultSendBufSize = inet.TCPBufferSize{
+	Min:     4096,
+	Default: 16384,
+	Max:     4194304,
+}
+
+// Stack implements inet.Stack for host sockets.
+type Stack struct {
+	// Stack is immutable.
+	interfaces     map[int32]inet.Interface
+	interfaceAddrs map[int32][]inet.InterfaceAddr
+	supportsIPv6   bool
+	tcpRecvBufSize inet.TCPBufferSize
+	tcpSendBufSize inet.TCPBufferSize
+	tcpSACKEnabled bool
+}
+
+// NewStack returns an empty Stack containing no configuration.
+func NewStack() *Stack {
+	return &Stack{
+		interfaces:     make(map[int32]inet.Interface),
+		interfaceAddrs: make(map[int32][]inet.InterfaceAddr),
+	}
+}
+
+// Configure sets up the stack using the current state of the host network.
+func (s *Stack) Configure() error {
+	if err := addHostInterfaces(s); err != nil {
+		return err
+	}
+
+	if _, err := os.Stat("/proc/net/if_inet6"); err == nil {
+		s.supportsIPv6 = true
+	}
+
+	s.tcpRecvBufSize = defaultRecvBufSize
+	if tcpRMem, err := readTCPBufferSizeFile("/proc/sys/net/ipv4/tcp_rmem"); err == nil {
+		s.tcpRecvBufSize = tcpRMem
+	} else {
+		log.Warningf("Failed to read TCP receive buffer size, using default values")
+	}
+
+	s.tcpSendBufSize = defaultSendBufSize
+	if tcpWMem, err := readTCPBufferSizeFile("/proc/sys/net/ipv4/tcp_wmem"); err == nil {
+		s.tcpSendBufSize = tcpWMem
+	} else {
+		log.Warningf("Failed to read TCP send buffer size, using default values")
+	}
+
+	s.tcpSACKEnabled = false
+	if sack, err := ioutil.ReadFile("/proc/sys/net/ipv4/tcp_sack"); err == nil {
+		s.tcpSACKEnabled = strings.TrimSpace(string(sack)) != "0"
+	} else {
+		log.Warningf("Failed to read if TCP SACK if enabled, setting to false")
+	}
+
+	return nil
+}
+
+// ExtractHostInterfaces will populate an interface map and
+// interfaceAddrs map with the results of the equivalent
+// netlink messages.
+func ExtractHostInterfaces(links []syscall.NetlinkMessage, addrs []syscall.NetlinkMessage, interfaces map[int32]inet.Interface, interfaceAddrs map[int32][]inet.InterfaceAddr) error {
+	for _, link := range links {
+		if link.Header.Type != syscall.RTM_NEWLINK {
+			continue
+		}
+		if len(link.Data) < syscall.SizeofIfInfomsg {
+			return fmt.Errorf("RTM_GETLINK returned RTM_NEWLINK message with invalid data length (%d bytes, expected at least %d bytes)", len(link.Data), syscall.SizeofIfInfomsg)
+		}
+		var ifinfo syscall.IfInfomsg
+		binary.Unmarshal(link.Data[:syscall.SizeofIfInfomsg], usermem.ByteOrder, &ifinfo)
+		inetIF := inet.Interface{
+			DeviceType: ifinfo.Type,
+			Flags:      ifinfo.Flags,
+		}
+		// Not clearly documented: syscall.ParseNetlinkRouteAttr will check the
+		// syscall.NetlinkMessage.Header.Type and skip the struct ifinfomsg
+		// accordingly.
+		attrs, err := syscall.ParseNetlinkRouteAttr(&link)
+		if err != nil {
+			return fmt.Errorf("RTM_GETLINK returned RTM_NEWLINK message with invalid rtattrs: %v", err)
+		}
+		for _, attr := range attrs {
+			switch attr.Attr.Type {
+			case syscall.IFLA_ADDRESS:
+				inetIF.Addr = attr.Value
+			case syscall.IFLA_IFNAME:
+				inetIF.Name = string(attr.Value[:len(attr.Value)-1])
+			}
+		}
+		interfaces[ifinfo.Index] = inetIF
+	}
+
+	for _, addr := range addrs {
+		if addr.Header.Type != syscall.RTM_NEWADDR {
+			continue
+		}
+		if len(addr.Data) < syscall.SizeofIfAddrmsg {
+			return fmt.Errorf("RTM_GETADDR returned RTM_NEWADDR message with invalid data length (%d bytes, expected at least %d bytes)", len(addr.Data), syscall.SizeofIfAddrmsg)
+		}
+		var ifaddr syscall.IfAddrmsg
+		binary.Unmarshal(addr.Data[:syscall.SizeofIfAddrmsg], usermem.ByteOrder, &ifaddr)
+		inetAddr := inet.InterfaceAddr{
+			Family:    ifaddr.Family,
+			PrefixLen: ifaddr.Prefixlen,
+			Flags:     ifaddr.Flags,
+		}
+		attrs, err := syscall.ParseNetlinkRouteAttr(&addr)
+		if err != nil {
+			return fmt.Errorf("RTM_GETADDR returned RTM_NEWADDR message with invalid rtattrs: %v", err)
+		}
+		for _, attr := range attrs {
+			switch attr.Attr.Type {
+			case syscall.IFA_ADDRESS:
+				inetAddr.Addr = attr.Value
+			}
+		}
+		interfaceAddrs[int32(ifaddr.Index)] = append(interfaceAddrs[int32(ifaddr.Index)], inetAddr)
+	}
+
+	return nil
+}
+
+func addHostInterfaces(s *Stack) error {
+	links, err := doNetlinkRouteRequest(syscall.RTM_GETLINK)
+	if err != nil {
+		return fmt.Errorf("RTM_GETLINK failed: %v", err)
+	}
+
+	addrs, err := doNetlinkRouteRequest(syscall.RTM_GETADDR)
+	if err != nil {
+		return fmt.Errorf("RTM_GETADDR failed: %v", err)
+	}
+
+	return ExtractHostInterfaces(links, addrs, s.interfaces, s.interfaceAddrs)
+}
+
+func doNetlinkRouteRequest(req int) ([]syscall.NetlinkMessage, error) {
+	data, err := syscall.NetlinkRIB(req, syscall.AF_UNSPEC)
+	if err != nil {
+		return nil, err
+	}
+	return syscall.ParseNetlinkMessage(data)
+}
+
+func readTCPBufferSizeFile(filename string) (inet.TCPBufferSize, error) {
+	contents, err := ioutil.ReadFile(filename)
+	if err != nil {
+		return inet.TCPBufferSize{}, fmt.Errorf("failed to read %s: %v", filename, err)
+	}
+	ioseq := usermem.BytesIOSequence(contents)
+	fields := make([]int32, 3)
+	if n, err := usermem.CopyInt32StringsInVec(context.Background(), ioseq.IO, ioseq.Addrs, fields, ioseq.Opts); n != ioseq.NumBytes() || err != nil {
+		return inet.TCPBufferSize{}, fmt.Errorf("failed to parse %s (%q): got %v after %d/%d bytes", filename, contents, err, n, ioseq.NumBytes())
+	}
+	return inet.TCPBufferSize{
+		Min:     int(fields[0]),
+		Default: int(fields[1]),
+		Max:     int(fields[2]),
+	}, nil
+}
+
+// Interfaces implements inet.Stack.Interfaces.
+func (s *Stack) Interfaces() map[int32]inet.Interface {
+	return s.interfaces
+}
+
+// InterfaceAddrs implements inet.Stack.InterfaceAddrs.
+func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr {
+	return s.interfaceAddrs
+}
+
+// SupportsIPv6 implements inet.Stack.SupportsIPv6.
+func (s *Stack) SupportsIPv6() bool {
+	return s.supportsIPv6
+}
+
+// TCPReceiveBufferSize implements inet.Stack.TCPReceiveBufferSize.
+func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) {
+	return s.tcpRecvBufSize, nil
+}
+
+// SetTCPReceiveBufferSize implements inet.Stack.SetTCPReceiveBufferSize.
+func (s *Stack) SetTCPReceiveBufferSize(size inet.TCPBufferSize) error {
+	return syserror.EACCES
+}
+
+// TCPSendBufferSize implements inet.Stack.TCPSendBufferSize.
+func (s *Stack) TCPSendBufferSize() (inet.TCPBufferSize, error) {
+	return s.tcpSendBufSize, nil
+}
+
+// SetTCPSendBufferSize implements inet.Stack.SetTCPSendBufferSize.
+func (s *Stack) SetTCPSendBufferSize(size inet.TCPBufferSize) error {
+	return syserror.EACCES
+}
+
+// TCPSACKEnabled implements inet.Stack.TCPSACKEnabled.
+func (s *Stack) TCPSACKEnabled() (bool, error) {
+	return s.tcpSACKEnabled, nil
+}
+
+// SetTCPSACKEnabled implements inet.Stack.SetTCPSACKEnabled.
+func (s *Stack) SetTCPSACKEnabled(enabled bool) error {
+	return syserror.EACCES
+}
diff --git a/pkg/sentry/socket/netlink/BUILD b/pkg/sentry/socket/netlink/BUILD
new file mode 100644
index 000000000..9df3ab17c
--- /dev/null
+++ b/pkg/sentry/socket/netlink/BUILD
@@ -0,0 +1,47 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "netlink_state",
+    srcs = [
+        "socket.go",
+    ],
+    out = "netlink_state.go",
+    package = "netlink",
+)
+
+go_library(
+    name = "netlink",
+    srcs = [
+        "message.go",
+        "netlink_state.go",
+        "provider.go",
+        "socket.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/socket/netlink/port",
+        "//pkg/sentry/socket/unix",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/socket/netlink/message.go b/pkg/sentry/socket/netlink/message.go
new file mode 100644
index 000000000..b902d7ec9
--- /dev/null
+++ b/pkg/sentry/socket/netlink/message.go
@@ -0,0 +1,159 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netlink
+
+import (
+	"fmt"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// alignUp rounds a length up to an alignment.
+//
+// Preconditions: align is a power of two.
+func alignUp(length int, align uint) int {
+	return (length + int(align) - 1) &^ (int(align) - 1)
+}
+
+// Message contains a complete serialized netlink message.
+type Message struct {
+	buf []byte
+}
+
+// NewMessage creates a new Message containing the passed header.
+//
+// The header length will be updated by Finalize.
+func NewMessage(hdr linux.NetlinkMessageHeader) *Message {
+	return &Message{
+		buf: binary.Marshal(nil, usermem.ByteOrder, hdr),
+	}
+}
+
+// Finalize returns the []byte containing the entire message, with the total
+// length set in the message header. The Message must not be modified after
+// calling Finalize.
+func (m *Message) Finalize() []byte {
+	// Update length, which is the first 4 bytes of the header.
+	usermem.ByteOrder.PutUint32(m.buf, uint32(len(m.buf)))
+
+	// Align the message. Note that the message length in the header (set
+	// above) is the useful length of the message, not the total aligned
+	// length. See net/netlink/af_netlink.c:__nlmsg_put.
+	aligned := alignUp(len(m.buf), linux.NLMSG_ALIGNTO)
+	m.putZeros(aligned - len(m.buf))
+	return m.buf
+}
+
+// putZeros adds n zeros to the message.
+func (m *Message) putZeros(n int) {
+	for n > 0 {
+		m.buf = append(m.buf, 0)
+		n--
+	}
+}
+
+// Put serializes v into the message.
+func (m *Message) Put(v interface{}) {
+	m.buf = binary.Marshal(m.buf, usermem.ByteOrder, v)
+}
+
+// PutAttr adds v to the message as a netlink attribute.
+//
+// Preconditions: The serialized attribute (linux.NetlinkAttrHeaderSize +
+// binary.Size(v) fits in math.MaxUint16 bytes.
+func (m *Message) PutAttr(atype uint16, v interface{}) {
+	l := linux.NetlinkAttrHeaderSize + int(binary.Size(v))
+	if l > math.MaxUint16 {
+		panic(fmt.Sprintf("attribute too large: %d", l))
+	}
+
+	m.Put(linux.NetlinkAttrHeader{
+		Type:   atype,
+		Length: uint16(l),
+	})
+	m.Put(v)
+
+	// Align the attribute.
+	aligned := alignUp(l, linux.NLA_ALIGNTO)
+	m.putZeros(aligned - l)
+}
+
+// PutAttrString adds s to the message as a netlink attribute.
+func (m *Message) PutAttrString(atype uint16, s string) {
+	l := linux.NetlinkAttrHeaderSize + len(s) + 1
+	m.Put(linux.NetlinkAttrHeader{
+		Type:   atype,
+		Length: uint16(l),
+	})
+
+	// String + NUL-termination.
+	m.Put([]byte(s))
+	m.putZeros(1)
+
+	// Align the attribute.
+	aligned := alignUp(l, linux.NLA_ALIGNTO)
+	m.putZeros(aligned - l)
+}
+
+// MessageSet contains a series of netlink messages.
+type MessageSet struct {
+	// Multi indicates that this a multi-part message, to be terminated by
+	// NLMSG_DONE. NLMSG_DONE is sent even if the set contains only one
+	// Message.
+	//
+	// If Multi is set, all added messages will have NLM_F_MULTI set.
+	Multi bool
+
+	// PortID is the destination port for all messages.
+	PortID int32
+
+	// Seq is the sequence counter for all messages in the set.
+	Seq uint32
+
+	// Messages contains the messages in the set.
+	Messages []*Message
+}
+
+// NewMessageSet creates a new MessageSet.
+//
+// portID is the destination port to set as PortID in all messages.
+//
+// seq is the sequence counter to set as seq in all messages in the set.
+func NewMessageSet(portID int32, seq uint32) *MessageSet {
+	return &MessageSet{
+		PortID: portID,
+		Seq:    seq,
+	}
+}
+
+// AddMessage adds a new message to the set and returns it for further
+// additions.
+//
+// The passed header will have Seq, PortID and the multi flag set
+// automatically.
+func (ms *MessageSet) AddMessage(hdr linux.NetlinkMessageHeader) *Message {
+	hdr.Seq = ms.Seq
+	hdr.PortID = uint32(ms.PortID)
+	if ms.Multi {
+		hdr.Flags |= linux.NLM_F_MULTI
+	}
+
+	m := NewMessage(hdr)
+	ms.Messages = append(ms.Messages, m)
+	return m
+}
diff --git a/pkg/sentry/socket/netlink/port/BUILD b/pkg/sentry/socket/netlink/port/BUILD
new file mode 100644
index 000000000..7340b95c9
--- /dev/null
+++ b/pkg/sentry/socket/netlink/port/BUILD
@@ -0,0 +1,28 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "port_state",
+    srcs = ["port.go"],
+    out = "port_state.go",
+    package = "port",
+)
+
+go_library(
+    name = "port",
+    srcs = [
+        "port.go",
+        "port_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink/port",
+    visibility = ["//pkg/sentry:internal"],
+    deps = ["//pkg/state"],
+)
+
+go_test(
+    name = "port_test",
+    srcs = ["port_test.go"],
+    embed = [":port"],
+)
diff --git a/pkg/sentry/socket/netlink/port/port.go b/pkg/sentry/socket/netlink/port/port.go
new file mode 100644
index 000000000..4ccf0b84c
--- /dev/null
+++ b/pkg/sentry/socket/netlink/port/port.go
@@ -0,0 +1,114 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package port provides port ID allocation for netlink sockets.
+//
+// A netlink port is any int32 value. Positive ports are typically equivalent
+// to the PID of the binding process. If that port is unavailable, negative
+// ports are searched to find a free port that will not conflict with other
+// PIDS.
+package port
+
+import (
+	"fmt"
+	"math"
+	"math/rand"
+	"sync"
+)
+
+// maxPorts is a sanity limit on the maximum number of ports to allocate per
+// protocol.
+const maxPorts = 10000
+
+// Manager allocates netlink port IDs.
+type Manager struct {
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// ports contains a map of allocated ports for each protocol.
+	ports map[int]map[int32]struct{}
+}
+
+// New creates a new Manager.
+func New() *Manager {
+	return &Manager{
+		ports: make(map[int]map[int32]struct{}),
+	}
+}
+
+// Allocate reserves a new port ID for protocol. hint will be taken if
+// available.
+func (m *Manager) Allocate(protocol int, hint int32) (int32, bool) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	proto, ok := m.ports[protocol]
+	if !ok {
+		proto = make(map[int32]struct{})
+		// Port 0 is reserved for the kernel.
+		proto[0] = struct{}{}
+		m.ports[protocol] = proto
+	}
+
+	if len(proto) >= maxPorts {
+		return 0, false
+	}
+
+	if _, ok := proto[hint]; !ok {
+		// Hint is available, reserve it.
+		proto[hint] = struct{}{}
+		return hint, true
+	}
+
+	// Search for any free port in [math.MinInt32, -4096). The positive
+	// port space is left open for pid-based allocations. This behavior is
+	// consistent with Linux.
+	start := int32(math.MinInt32 + rand.Int63n(math.MaxInt32-4096+1))
+	curr := start
+	for {
+		if _, ok := proto[curr]; !ok {
+			proto[curr] = struct{}{}
+			return curr, true
+		}
+
+		curr--
+		if curr >= -4096 {
+			curr = -4097
+		}
+		if curr == start {
+			// Nothing found. We should always find a free port
+			// because maxPorts < -4096 - MinInt32.
+			panic(fmt.Sprintf("No free port found in %+v", proto))
+		}
+	}
+}
+
+// Release frees the specified port for protocol.
+//
+// Preconditions: port is already allocated.
+func (m *Manager) Release(protocol int, port int32) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	proto, ok := m.ports[protocol]
+	if !ok {
+		panic(fmt.Sprintf("Released port %d for protocol %d which has no allocations", port, protocol))
+	}
+
+	if _, ok := proto[port]; !ok {
+		panic(fmt.Sprintf("Released port %d for protocol %d is not allocated", port, protocol))
+	}
+
+	delete(proto, port)
+}
diff --git a/pkg/sentry/socket/netlink/port/port_test.go b/pkg/sentry/socket/netlink/port/port_test.go
new file mode 100644
index 000000000..34565e2f9
--- /dev/null
+++ b/pkg/sentry/socket/netlink/port/port_test.go
@@ -0,0 +1,82 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package port
+
+import (
+	"testing"
+)
+
+func TestAllocateHint(t *testing.T) {
+	m := New()
+
+	// We can get the hint port.
+	p, ok := m.Allocate(0, 1)
+	if !ok {
+		t.Errorf("m.Allocate got !ok want ok")
+	}
+	if p != 1 {
+		t.Errorf("m.Allocate(0, 1) got %d want 1", p)
+	}
+
+	// Hint is taken.
+	p, ok = m.Allocate(0, 1)
+	if !ok {
+		t.Errorf("m.Allocate got !ok want ok")
+	}
+	if p == 1 {
+		t.Errorf("m.Allocate(0, 1) got 1 want anything else")
+	}
+
+	// Hint is available for a different protocol.
+	p, ok = m.Allocate(1, 1)
+	if !ok {
+		t.Errorf("m.Allocate got !ok want ok")
+	}
+	if p != 1 {
+		t.Errorf("m.Allocate(1, 1) got %d want 1", p)
+	}
+
+	m.Release(0, 1)
+
+	// Hint is available again after release.
+	p, ok = m.Allocate(0, 1)
+	if !ok {
+		t.Errorf("m.Allocate got !ok want ok")
+	}
+	if p != 1 {
+		t.Errorf("m.Allocate(0, 1) got %d want 1", p)
+	}
+}
+
+func TestAllocateExhausted(t *testing.T) {
+	m := New()
+
+	// Fill all ports (0 is already reserved).
+	for i := int32(1); i < maxPorts; i++ {
+		p, ok := m.Allocate(0, i)
+		if !ok {
+			t.Fatalf("m.Allocate got !ok want ok")
+		}
+		if p != i {
+			t.Fatalf("m.Allocate(0, %d) got %d want %d", i, p, i)
+		}
+	}
+
+	// Now no more can be allocated.
+	p, ok := m.Allocate(0, 1)
+	if ok {
+		t.Errorf("m.Allocate got %d, ok want !ok", p)
+	}
+}
diff --git a/pkg/sentry/socket/netlink/provider.go b/pkg/sentry/socket/netlink/provider.go
new file mode 100644
index 000000000..36800da4d
--- /dev/null
+++ b/pkg/sentry/socket/netlink/provider.go
@@ -0,0 +1,104 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package netlink
+
+import (
+	"fmt"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// Protocol is the implementation of a netlink socket protocol.
+type Protocol interface {
+	// Protocol returns the Linux netlink protocol value.
+	Protocol() int
+
+	// ProcessMessage processes a single message from userspace.
+	//
+	// If err == nil, any messages added to ms will be sent back to the
+	// other end of the socket. Setting ms.Multi will cause an NLMSG_DONE
+	// message to be sent even if ms contains no messages.
+	ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *MessageSet) *syserr.Error
+}
+
+// Provider is a function that creates a new Protocol for a specific netlink
+// protocol.
+//
+// Note that this is distinct from socket.Provider, which is used for all
+// socket families.
+type Provider func(t *kernel.Task) (Protocol, *syserr.Error)
+
+// protocols holds a map of all known address protocols and their provider.
+var protocols = make(map[int]Provider)
+
+// RegisterProvider registers the provider of a given address protocol so that
+// netlink sockets of that type can be created via socket(2).
+//
+// Preconditions: May only be called before any netlink sockets are created.
+func RegisterProvider(protocol int, provider Provider) {
+	if p, ok := protocols[protocol]; ok {
+		panic(fmt.Sprintf("Netlink protocol %d already provided by %+v", protocol, p))
+	}
+
+	protocols[protocol] = provider
+}
+
+// socketProvider implements socket.Provider.
+type socketProvider struct {
+}
+
+// Socket implements socket.Provider.Socket.
+func (*socketProvider) Socket(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Netlink sockets must be specified as datagram or raw, but they
+	// behave the same regardless of type.
+	if stype != unix.SockDgram && stype != unix.SockRaw {
+		return nil, syserr.ErrSocketNotSupported
+	}
+
+	provider, ok := protocols[protocol]
+	if !ok {
+		return nil, syserr.ErrProtocolNotSupported
+	}
+
+	p, err := provider(t)
+	if err != nil {
+		return nil, err
+	}
+
+	s, err := NewSocket(t, p)
+	if err != nil {
+		return nil, err
+	}
+
+	d := socket.NewDirent(t, netlinkSocketDevice)
+	return fs.NewFile(t, d, fs.FileFlags{Read: true, Write: true}, s), nil
+}
+
+// Pair implements socket.Provider.Pair by returning an error.
+func (*socketProvider) Pair(*kernel.Task, unix.SockType, int) (*fs.File, *fs.File, *syserr.Error) {
+	// Netlink sockets never supports creating socket pairs.
+	return nil, nil, syserr.ErrNotSupported
+}
+
+// init registers the socket provider.
+func init() {
+	socket.RegisterProvider(linux.AF_NETLINK, &socketProvider{})
+}
diff --git a/pkg/sentry/socket/netlink/route/BUILD b/pkg/sentry/socket/netlink/route/BUILD
new file mode 100644
index 000000000..ff3f7b7a4
--- /dev/null
+++ b/pkg/sentry/socket/netlink/route/BUILD
@@ -0,0 +1,33 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "route_state",
+    srcs = ["protocol.go"],
+    out = "route_state.go",
+    package = "route",
+)
+
+go_library(
+    name = "route",
+    srcs = [
+        "protocol.go",
+        "route_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink/route",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/socket/netlink",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+    ],
+)
diff --git a/pkg/sentry/socket/netlink/route/protocol.go b/pkg/sentry/socket/netlink/route/protocol.go
new file mode 100644
index 000000000..d611519d4
--- /dev/null
+++ b/pkg/sentry/socket/netlink/route/protocol.go
@@ -0,0 +1,189 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package route provides a NETLINK_ROUTE socket protocol.
+package route
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+)
+
+// commandKind describes the operational class of a message type.
+//
+// The route message types use the lower 2 bits of the type to describe class
+// of command.
+type commandKind int
+
+const (
+	kindNew commandKind = 0x0
+	kindDel             = 0x1
+	kindGet             = 0x2
+	kindSet             = 0x3
+)
+
+func typeKind(typ uint16) commandKind {
+	return commandKind(typ & 0x3)
+}
+
+// Protocol implements netlink.Protocol.
+type Protocol struct {
+	// stack is the network stack that this provider describes.
+	//
+	// May be nil.
+	stack inet.Stack
+}
+
+var _ netlink.Protocol = (*Protocol)(nil)
+
+// NewProtocol creates a NETLINK_ROUTE netlink.Protocol.
+func NewProtocol(t *kernel.Task) (netlink.Protocol, *syserr.Error) {
+	return &Protocol{
+		stack: t.NetworkContext(),
+	}, nil
+}
+
+// Protocol implements netlink.Protocol.Protocol.
+func (p *Protocol) Protocol() int {
+	return linux.NETLINK_ROUTE
+}
+
+// dumpLinks handles RTM_GETLINK + NLM_F_DUMP requests.
+func (p *Protocol) dumpLinks(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// NLM_F_DUMP + RTM_GETLINK messages are supposed to include an
+	// ifinfomsg. However, Linux <3.9 only checked for rtgenmsg, and some
+	// userspace applications (including glibc) still include rtgenmsg.
+	// Linux has a workaround based on the total message length.
+	//
+	// We don't bother to check for either, since we don't support any
+	// extra attributes that may be included anyways.
+	//
+	// The message may also contain netlink attribute IFLA_EXT_MASK, which
+	// we don't support.
+
+	// The RTM_GETLINK dump response is a set of messages each containing
+	// an InterfaceInfoMessage followed by a set of netlink attributes.
+
+	// We always send back an NLMSG_DONE.
+	ms.Multi = true
+
+	if p.stack == nil {
+		// No network devices.
+		return nil
+	}
+
+	for id, i := range p.stack.Interfaces() {
+		m := ms.AddMessage(linux.NetlinkMessageHeader{
+			Type: linux.RTM_NEWLINK,
+		})
+
+		m.Put(linux.InterfaceInfoMessage{
+			Family: linux.AF_UNSPEC,
+			Type:   i.DeviceType,
+			Index:  id,
+			Flags:  i.Flags,
+		})
+
+		m.PutAttrString(linux.IFLA_IFNAME, i.Name)
+
+		// TODO: There are many more attributes, such as
+		// MAC address.
+	}
+
+	return nil
+}
+
+// dumpAddrs handles RTM_GETADDR + NLM_F_DUMP requests.
+func (p *Protocol) dumpAddrs(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// RTM_GETADDR dump requests need not contain anything more than the
+	// netlink header and 1 byte protocol family common to all
+	// NETLINK_ROUTE requests.
+	//
+	// TODO: Filter output by passed protocol family.
+
+	// The RTM_GETADDR dump response is a set of RTM_NEWADDR messages each
+	// containing an InterfaceAddrMessage followed by a set of netlink
+	// attributes.
+
+	// We always send back an NLMSG_DONE.
+	ms.Multi = true
+
+	if p.stack == nil {
+		// No network devices.
+		return nil
+	}
+
+	for id, as := range p.stack.InterfaceAddrs() {
+		for _, a := range as {
+			m := ms.AddMessage(linux.NetlinkMessageHeader{
+				Type: linux.RTM_NEWADDR,
+			})
+
+			m.Put(linux.InterfaceAddrMessage{
+				Family:    a.Family,
+				PrefixLen: a.PrefixLen,
+				Index:     uint32(id),
+			})
+
+			m.PutAttr(linux.IFA_ADDRESS, []byte(a.Addr))
+
+			// TODO: There are many more attributes.
+		}
+	}
+
+	return nil
+}
+
+// ProcessMessage implements netlink.Protocol.ProcessMessage.
+func (p *Protocol) ProcessMessage(ctx context.Context, hdr linux.NetlinkMessageHeader, data []byte, ms *netlink.MessageSet) *syserr.Error {
+	// All messages start with a 1 byte protocol family.
+	if len(data) < 1 {
+		// Linux ignores messages missing the protocol family. See
+		// net/core/rtnetlink.c:rtnetlink_rcv_msg.
+		return nil
+	}
+
+	// Non-GET message types require CAP_NET_ADMIN.
+	if typeKind(hdr.Type) != kindGet {
+		creds := auth.CredentialsFromContext(ctx)
+		if !creds.HasCapability(linux.CAP_NET_ADMIN) {
+			return syserr.ErrPermissionDenied
+		}
+	}
+
+	// TODO: Only the dump variant of the types below are
+	// supported.
+	if hdr.Flags&linux.NLM_F_DUMP != linux.NLM_F_DUMP {
+		return syserr.ErrNotSupported
+	}
+
+	switch hdr.Type {
+	case linux.RTM_GETLINK:
+		return p.dumpLinks(ctx, hdr, data, ms)
+	case linux.RTM_GETADDR:
+		return p.dumpAddrs(ctx, hdr, data, ms)
+	default:
+		return syserr.ErrNotSupported
+	}
+}
+
+// init registers the NETLINK_ROUTE provider.
+func init() {
+	netlink.RegisterProvider(linux.NETLINK_ROUTE, NewProtocol)
+}
diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go
new file mode 100644
index 000000000..2d0e59ceb
--- /dev/null
+++ b/pkg/sentry/socket/netlink/socket.go
@@ -0,0 +1,517 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package netlink provides core functionality for netlink sockets.
+package netlink
+
+import (
+	"sync"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink/port"
+	sunix "gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// defaultSendBufferSize is the default size for the send buffer.
+const defaultSendBufferSize = 16 * 1024
+
+// netlinkSocketDevice is the netlink socket virtual device.
+var netlinkSocketDevice = device.NewAnonDevice()
+
+// Socket is the base socket type for netlink sockets.
+//
+// This implementation only supports userspace sending and receiving messages
+// to/from the kernel.
+//
+// Socket implements socket.Socket.
+type Socket struct {
+	socket.ReceiveTimeout
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+
+	// ports provides netlink port allocation.
+	ports *port.Manager
+
+	// protocol is the netlink protocol implementation.
+	protocol Protocol
+
+	// ep is a datagram unix endpoint used to buffer messages sent from the
+	// kernel to userspace. RecvMsg reads messages from this endpoint.
+	ep unix.Endpoint
+
+	// connection is the kernel's connection to ep, used to write messages
+	// sent to userspace.
+	connection unix.ConnectedEndpoint
+
+	// mu protects the fields below.
+	mu sync.Mutex `state:"nosave"`
+
+	// bound indicates that portid is valid.
+	bound bool
+
+	// portID is the port ID allocated for this socket.
+	portID int32
+
+	// sendBufferSize is the send buffer "size". We don't actually have a
+	// fixed buffer but only consume this many bytes.
+	sendBufferSize uint64
+}
+
+var _ socket.Socket = (*Socket)(nil)
+
+// NewSocket creates a new Socket.
+func NewSocket(t *kernel.Task, protocol Protocol) (*Socket, *syserr.Error) {
+	// Datagram endpoint used to buffer kernel -> user messages.
+	ep := unix.NewConnectionless()
+
+	// Bind the endpoint for good measure so we can connect to it. The
+	// bound address will never be exposed.
+	if terr := ep.Bind(tcpip.FullAddress{Addr: "dummy"}, nil); terr != nil {
+		ep.Close()
+		return nil, syserr.TranslateNetstackError(terr)
+	}
+
+	// Create a connection from which the kernel can write messages.
+	connection, terr := ep.(unix.BoundEndpoint).UnidirectionalConnect()
+	if terr != nil {
+		ep.Close()
+		return nil, syserr.TranslateNetstackError(terr)
+	}
+
+	return &Socket{
+		ports:          t.Kernel().NetlinkPorts(),
+		protocol:       protocol,
+		ep:             ep,
+		connection:     connection,
+		sendBufferSize: defaultSendBufferSize,
+	}, nil
+}
+
+// Release implements fs.FileOperations.Release.
+func (s *Socket) Release() {
+	s.connection.Release()
+	s.ep.Close()
+
+	if s.bound {
+		s.ports.Release(s.protocol.Protocol(), s.portID)
+	}
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *Socket) Readiness(mask waiter.EventMask) waiter.EventMask {
+	// ep holds messages to be read and thus handles EventIn readiness.
+	ready := s.ep.Readiness(mask)
+
+	if mask&waiter.EventOut == waiter.EventOut {
+		// sendMsg handles messages synchronously and is thus always
+		// ready for writing.
+		ready |= waiter.EventOut
+	}
+
+	return ready
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *Socket) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.ep.EventRegister(e, mask)
+	// Writable readiness never changes, so no registration is needed.
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *Socket) EventUnregister(e *waiter.Entry) {
+	s.ep.EventUnregister(e)
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *Socket) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	// TODO: no ioctls supported.
+	return 0, syserror.ENOTTY
+}
+
+// ExtractSockAddr extracts the SockAddrNetlink from b.
+func ExtractSockAddr(b []byte) (*linux.SockAddrNetlink, *syserr.Error) {
+	if len(b) < linux.SockAddrNetlinkSize {
+		return nil, syserr.ErrBadAddress
+	}
+
+	var sa linux.SockAddrNetlink
+	binary.Unmarshal(b[:linux.SockAddrNetlinkSize], usermem.ByteOrder, &sa)
+
+	if sa.Family != linux.AF_NETLINK {
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	return &sa, nil
+}
+
+// bindPort binds this socket to a port, preferring 'port' if it is available.
+//
+// port of 0 defaults to the ThreadGroup ID.
+//
+// Preconditions: mu is held.
+func (s *Socket) bindPort(t *kernel.Task, port int32) *syserr.Error {
+	if s.bound {
+		// Re-binding is only allowed if the port doesn't change.
+		if port != s.portID {
+			return syserr.ErrInvalidArgument
+		}
+
+		return nil
+	}
+
+	if port == 0 {
+		port = int32(t.ThreadGroup().ID())
+	}
+	port, ok := s.ports.Allocate(s.protocol.Protocol(), port)
+	if !ok {
+		return syserr.ErrBusy
+	}
+
+	s.portID = port
+	s.bound = true
+	return nil
+}
+
+// Bind implements socket.Socket.Bind.
+func (s *Socket) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	a, err := ExtractSockAddr(sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// No support for multicast groups yet.
+	if a.Groups != 0 {
+		return syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	return s.bindPort(t, int32(a.PortID))
+}
+
+// Connect implements socket.Socket.Connect.
+func (s *Socket) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	a, err := ExtractSockAddr(sockaddr)
+	if err != nil {
+		return err
+	}
+
+	// No support for multicast groups yet.
+	if a.Groups != 0 {
+		return syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	if a.PortID == 0 {
+		// Netlink sockets default to connected to the kernel, but
+		// connecting anyways automatically binds if not already bound.
+		if !s.bound {
+			// Pass port 0 to get an auto-selected port ID.
+			return s.bindPort(t, 0)
+		}
+		return nil
+	}
+
+	// We don't support non-kernel destination ports. Linux returns EPERM
+	// if applications attempt to do this without NL_CFG_F_NONROOT_SEND, so
+	// we emulate that.
+	return syserr.ErrPermissionDenied
+}
+
+// Accept implements socket.Socket.Accept.
+func (s *Socket) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	// Netlink sockets never support accept.
+	return 0, nil, 0, syserr.ErrNotSupported
+}
+
+// Listen implements socket.Socket.Listen.
+func (s *Socket) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	// Netlink sockets never support listen.
+	return syserr.ErrNotSupported
+}
+
+// Shutdown implements socket.Socket.Shutdown.
+func (s *Socket) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	// Netlink sockets never support shutdown.
+	return syserr.ErrNotSupported
+}
+
+// GetSockOpt implements socket.Socket.GetSockOpt.
+func (s *Socket) GetSockOpt(t *kernel.Task, level int, name int, outLen int) (interface{}, *syserr.Error) {
+	// TODO: no sockopts supported.
+	return nil, syserr.ErrProtocolNotAvailable
+}
+
+// SetSockOpt implements socket.Socket.SetSockOpt.
+func (s *Socket) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error {
+	// TODO: no sockopts supported.
+	return syserr.ErrProtocolNotAvailable
+}
+
+// GetSockName implements socket.Socket.GetSockName.
+func (s *Socket) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	sa := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		PortID: uint32(s.portID),
+	}
+	return sa, uint32(binary.Size(sa)), nil
+}
+
+// GetPeerName implements socket.Socket.GetPeerName.
+func (s *Socket) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	sa := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		// TODO: Support non-kernel peers. For now the peer
+		// must be the kernel.
+		PortID: 0,
+	}
+	return sa, uint32(binary.Size(sa)), nil
+}
+
+// RecvMsg implements socket.Socket.RecvMsg.
+func (s *Socket) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, interface{}, uint32, unix.ControlMessages, *syserr.Error) {
+	from := linux.SockAddrNetlink{
+		Family: linux.AF_NETLINK,
+		PortID: 0,
+	}
+	fromLen := uint32(binary.Size(from))
+
+	trunc := flags&linux.MSG_TRUNC != 0
+
+	r := sunix.EndpointReader{
+		Endpoint: s.ep,
+		Peek:     flags&linux.MSG_PEEK != 0,
+	}
+
+	if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		if trunc {
+			n = int64(r.MsgSize)
+		}
+		return int(n), from, fromLen, unix.ControlMessages{}, syserr.FromError(err)
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// receive all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock {
+			if trunc {
+				n = int64(r.MsgSize)
+			}
+			return int(n), from, fromLen, unix.ControlMessages{}, syserr.FromError(err)
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syserror.ETIMEDOUT {
+				return 0, nil, 0, unix.ControlMessages{}, syserr.ErrTryAgain
+			}
+			return 0, nil, 0, unix.ControlMessages{}, syserr.FromError(err)
+		}
+	}
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *Socket) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	return dst.CopyOutFrom(ctx, &sunix.EndpointReader{
+		Endpoint: s.ep,
+	})
+}
+
+// sendResponse sends the response messages in ms back to userspace.
+func (s *Socket) sendResponse(ctx context.Context, ms *MessageSet) *syserr.Error {
+	// Linux combines multiple netlink messages into a single datagram.
+	bufs := make([][]byte, 0, len(ms.Messages))
+	for _, m := range ms.Messages {
+		bufs = append(bufs, m.Finalize())
+	}
+
+	if len(bufs) > 0 {
+		// RecvMsg never receives the address, so we don't need to send
+		// one.
+		_, notify, terr := s.connection.Send(bufs, unix.ControlMessages{}, tcpip.FullAddress{})
+		// If the buffer is full, we simply drop messages, just like
+		// Linux.
+		if terr != nil && terr != tcpip.ErrWouldBlock {
+			return syserr.TranslateNetstackError(terr)
+		}
+		if notify {
+			s.connection.SendNotify()
+		}
+	}
+
+	// N.B. multi-part messages should still send NLMSG_DONE even if
+	// MessageSet contains no messages.
+	//
+	// N.B. NLMSG_DONE is always sent in a different datagram. See
+	// net/netlink/af_netlink.c:netlink_dump.
+	if ms.Multi {
+		m := NewMessage(linux.NetlinkMessageHeader{
+			Type:   linux.NLMSG_DONE,
+			Flags:  linux.NLM_F_MULTI,
+			Seq:    ms.Seq,
+			PortID: uint32(ms.PortID),
+		})
+
+		_, notify, terr := s.connection.Send([][]byte{m.Finalize()}, unix.ControlMessages{}, tcpip.FullAddress{})
+		if terr != nil && terr != tcpip.ErrWouldBlock {
+			return syserr.TranslateNetstackError(terr)
+		}
+		if notify {
+			s.connection.SendNotify()
+		}
+	}
+
+	return nil
+}
+
+// processMessages handles each message in buf, passing it to the protocol
+// handler for final handling.
+func (s *Socket) processMessages(ctx context.Context, buf []byte) *syserr.Error {
+	for len(buf) > 0 {
+		if len(buf) < linux.NetlinkMessageHeaderSize {
+			// Linux ignores messages that are too short. See
+			// net/netlink/af_netlink.c:netlink_rcv_skb.
+			break
+		}
+
+		var hdr linux.NetlinkMessageHeader
+		binary.Unmarshal(buf[:linux.NetlinkMessageHeaderSize], usermem.ByteOrder, &hdr)
+
+		if hdr.Length < linux.NetlinkMessageHeaderSize || uint64(hdr.Length) > uint64(len(buf)) {
+			// Linux ignores malformed messages. See
+			// net/netlink/af_netlink.c:netlink_rcv_skb.
+			break
+		}
+
+		// Data from this message.
+		data := buf[linux.NetlinkMessageHeaderSize:hdr.Length]
+
+		// Advance to the next message.
+		next := alignUp(int(hdr.Length), linux.NLMSG_ALIGNTO)
+		if next >= len(buf)-1 {
+			next = len(buf) - 1
+		}
+		buf = buf[next:]
+
+		// Ignore control messages.
+		if hdr.Type < linux.NLMSG_MIN_TYPE {
+			continue
+		}
+
+		// TODO: ACKs not supported yet.
+		if hdr.Flags&linux.NLM_F_ACK == linux.NLM_F_ACK {
+			return syserr.ErrNotSupported
+		}
+
+		ms := NewMessageSet(s.portID, hdr.Seq)
+		if err := s.protocol.ProcessMessage(ctx, hdr, data, ms); err != nil {
+			return err
+		}
+
+		if err := s.sendResponse(ctx, ms); err != nil {
+			return err
+		}
+	}
+
+	return nil
+}
+
+// sendMsg is the core of message send, used for SendMsg and Write.
+func (s *Socket) sendMsg(ctx context.Context, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	dstPort := int32(0)
+
+	if len(to) != 0 {
+		a, err := ExtractSockAddr(to)
+		if err != nil {
+			return 0, err
+		}
+
+		// No support for multicast groups yet.
+		if a.Groups != 0 {
+			return 0, syserr.ErrPermissionDenied
+		}
+
+		dstPort = int32(a.PortID)
+	}
+
+	if dstPort != 0 {
+		// Non-kernel destinations not supported yet. Treat as if
+		// NL_CFG_F_NONROOT_SEND is not set.
+		return 0, syserr.ErrPermissionDenied
+	}
+
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	// For simplicity, and consistency with Linux, we copy in the entire
+	// message up front.
+	if uint64(src.NumBytes()) > s.sendBufferSize {
+		return 0, syserr.ErrMessageTooLong
+	}
+
+	buf := make([]byte, src.NumBytes())
+	n, err := src.CopyIn(ctx, buf)
+	if err != nil {
+		// Don't partially consume messages.
+		return 0, syserr.FromError(err)
+	}
+
+	if err := s.processMessages(ctx, buf); err != nil {
+		return 0, err
+	}
+
+	return n, nil
+}
+
+// SendMsg implements socket.Socket.SendMsg.
+func (s *Socket) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	return s.sendMsg(t, src, to, flags, controlMessages)
+}
+
+// Write implements fs.FileOperations.Write.
+func (s *Socket) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	n, err := s.sendMsg(ctx, src, nil, 0, unix.ControlMessages{})
+	return int64(n), err.ToError()
+}
diff --git a/pkg/sentry/socket/rpcinet/BUILD b/pkg/sentry/socket/rpcinet/BUILD
new file mode 100644
index 000000000..b0351b363
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/BUILD
@@ -0,0 +1,59 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "rpcinet",
+    srcs = [
+        "device.go",
+        "rpcinet.go",
+        "socket.go",
+        "stack.go",
+        "stack_unsafe.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        ":syscall_rpc_go_proto",
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/socket/hostinet",
+        "//pkg/sentry/socket/rpcinet/conn",
+        "//pkg/sentry/socket/rpcinet/notifier",
+        "//pkg/sentry/usermem",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip/buffer",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/unet",
+        "//pkg/waiter",
+    ],
+)
+
+proto_library(
+    name = "syscall_rpc_proto",
+    srcs = ["syscall_rpc.proto"],
+    visibility = [
+        "//visibility:public",
+    ],
+)
+
+go_proto_library(
+    name = "syscall_rpc_go_proto",
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/syscall_rpc_go_proto",
+    proto = ":syscall_rpc_proto",
+    visibility = [
+        "//visibility:public",
+    ],
+)
diff --git a/pkg/sentry/socket/rpcinet/conn/BUILD b/pkg/sentry/socket/rpcinet/conn/BUILD
new file mode 100644
index 000000000..4923dee4b
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/conn/BUILD
@@ -0,0 +1,17 @@
+package(licenses = ["notice"])  # BSD
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "conn",
+    srcs = ["conn.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/conn",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/binary",
+        "//pkg/sentry/socket/rpcinet:syscall_rpc_go_proto",
+        "//pkg/syserr",
+        "//pkg/unet",
+        "@com_github_golang_protobuf//proto:go_default_library",
+    ],
+)
diff --git a/pkg/sentry/socket/rpcinet/conn/conn.go b/pkg/sentry/socket/rpcinet/conn/conn.go
new file mode 100644
index 000000000..ea6ec87ed
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/conn/conn.go
@@ -0,0 +1,167 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package conn is an RPC connection to a syscall RPC server.
+package conn
+
+import (
+	"fmt"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"github.com/golang/protobuf/proto"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/syscall_rpc_go_proto"
+)
+
+type request struct {
+	response     []byte
+	ready        chan struct{}
+	ignoreResult bool
+}
+
+// RPCConnection represents a single RPC connection to a syscall gofer.
+type RPCConnection struct {
+	// reqID is the ID of the last request and must be accessed atomically.
+	reqID uint64
+
+	sendMu sync.Mutex
+	socket *unet.Socket
+
+	reqMu    sync.Mutex
+	requests map[uint64]request
+}
+
+// NewRPCConnection initializes a RPC connection to a socket gofer.
+func NewRPCConnection(s *unet.Socket) *RPCConnection {
+	conn := &RPCConnection{socket: s, requests: map[uint64]request{}}
+	go func() { // S/R-FIXME
+		var nums [16]byte
+		for {
+			for n := 0; n < len(nums); {
+				nn, err := conn.socket.Read(nums[n:])
+				if err != nil {
+					panic(fmt.Sprint("error reading length from socket rpc gofer: ", err))
+				}
+				n += nn
+			}
+
+			b := make([]byte, binary.LittleEndian.Uint64(nums[:8]))
+			id := binary.LittleEndian.Uint64(nums[8:])
+
+			for n := 0; n < len(b); {
+				nn, err := conn.socket.Read(b[n:])
+				if err != nil {
+					panic(fmt.Sprint("error reading request from socket rpc gofer: ", err))
+				}
+				n += nn
+			}
+
+			conn.reqMu.Lock()
+			r := conn.requests[id]
+			if r.ignoreResult {
+				delete(conn.requests, id)
+			} else {
+				r.response = b
+				conn.requests[id] = r
+			}
+			conn.reqMu.Unlock()
+			close(r.ready)
+		}
+	}()
+	return conn
+}
+
+// NewRequest makes a request to the RPC gofer and returns the request ID and a
+// channel which will be closed once the request completes.
+func (c *RPCConnection) NewRequest(req pb.SyscallRequest, ignoreResult bool) (uint64, chan struct{}) {
+	b, err := proto.Marshal(&req)
+	if err != nil {
+		panic(fmt.Sprint("invalid proto: ", err))
+	}
+
+	id := atomic.AddUint64(&c.reqID, 1)
+	ch := make(chan struct{})
+
+	c.reqMu.Lock()
+	c.requests[id] = request{ready: ch, ignoreResult: ignoreResult}
+	c.reqMu.Unlock()
+
+	c.sendMu.Lock()
+	defer c.sendMu.Unlock()
+
+	var nums [16]byte
+	binary.LittleEndian.PutUint64(nums[:8], uint64(len(b)))
+	binary.LittleEndian.PutUint64(nums[8:], id)
+	for n := 0; n < len(nums); {
+		nn, err := c.socket.Write(nums[n:])
+		if err != nil {
+			panic(fmt.Sprint("error writing length and ID to socket gofer: ", err))
+		}
+		n += nn
+	}
+
+	for n := 0; n < len(b); {
+		nn, err := c.socket.Write(b[n:])
+		if err != nil {
+			panic(fmt.Sprint("error writing request to socket gofer: ", err))
+		}
+		n += nn
+	}
+
+	return id, ch
+}
+
+// RPCReadFile will execute the ReadFile helper RPC method which avoids the
+// common pattern of open(2), read(2), close(2) by doing all three operations
+// as a single RPC. It will read the entire file or return EFBIG if the file
+// was too large.
+func (c *RPCConnection) RPCReadFile(path string) ([]byte, *syserr.Error) {
+	req := &pb.SyscallRequest_ReadFile{&pb.ReadFileRequest{
+		Path: path,
+	}}
+
+	id, ch := c.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-ch
+
+	res := c.Request(id).Result.(*pb.SyscallResponse_ReadFile).ReadFile.Result
+	if e, ok := res.(*pb.ReadFileResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.ReadFileResponse_Data).Data, nil
+}
+
+// Request retrieves the request corresponding to the given request ID.
+//
+// The channel returned by NewRequest must have been closed before Request can
+// be called. This will happen automatically, do not manually close the
+// channel.
+func (c *RPCConnection) Request(id uint64) pb.SyscallResponse {
+	c.reqMu.Lock()
+	r := c.requests[id]
+	delete(c.requests, id)
+	c.reqMu.Unlock()
+
+	var resp pb.SyscallResponse
+	if err := proto.Unmarshal(r.response, &resp); err != nil {
+		panic(fmt.Sprint("invalid proto: ", err))
+	}
+
+	return resp
+}
diff --git a/pkg/sentry/socket/rpcinet/device.go b/pkg/sentry/socket/rpcinet/device.go
new file mode 100644
index 000000000..f7b63436e
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/device.go
@@ -0,0 +1,19 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package rpcinet
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+var socketDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/socket/rpcinet/notifier/BUILD b/pkg/sentry/socket/rpcinet/notifier/BUILD
new file mode 100644
index 000000000..6f3b06a05
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/notifier/BUILD
@@ -0,0 +1,15 @@
+package(licenses = ["notice"])  # BSD
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "notifier",
+    srcs = ["notifier.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/notifier",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/sentry/socket/rpcinet:syscall_rpc_go_proto",
+        "//pkg/sentry/socket/rpcinet/conn",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/socket/rpcinet/notifier/notifier.go b/pkg/sentry/socket/rpcinet/notifier/notifier.go
new file mode 100644
index 000000000..f88a908ed
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/notifier/notifier.go
@@ -0,0 +1,230 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package notifier implements an FD notifier implementation over RPC.
+package notifier
+
+import (
+	"fmt"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/conn"
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/syscall_rpc_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+type fdInfo struct {
+	queue   *waiter.Queue
+	waiting bool
+}
+
+// Notifier holds all the state necessary to issue notifications when IO events
+// occur in the observed FDs.
+type Notifier struct {
+	// rpcConn is the connection that is used for sending RPCs.
+	rpcConn *conn.RPCConnection
+
+	// epFD is the epoll file descriptor used to register for io
+	// notifications.
+	epFD uint32
+
+	// mu protects fdMap.
+	mu sync.Mutex
+
+	// fdMap maps file descriptors to their notification queues and waiting
+	// status.
+	fdMap map[uint32]*fdInfo
+}
+
+// NewRPCNotifier creates a new notifier object.
+func NewRPCNotifier(cn *conn.RPCConnection) (*Notifier, error) {
+	id, c := cn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_EpollCreate1{&pb.EpollCreate1Request{}}}, false /* ignoreResult */)
+	<-c
+
+	res := cn.Request(id).Result.(*pb.SyscallResponse_EpollCreate1).EpollCreate1.Result
+	if e, ok := res.(*pb.EpollCreate1Response_ErrorNumber); ok {
+		return nil, syscall.Errno(e.ErrorNumber)
+	}
+
+	w := &Notifier{
+		rpcConn: cn,
+		epFD:    res.(*pb.EpollCreate1Response_Fd).Fd,
+		fdMap:   make(map[uint32]*fdInfo),
+	}
+
+	go w.waitAndNotify() // S/R-FIXME
+
+	return w, nil
+}
+
+// waitFD waits on mask for fd. The fdMap mutex must be hold.
+func (n *Notifier) waitFD(fd uint32, fi *fdInfo, mask waiter.EventMask) error {
+	if !fi.waiting && mask == 0 {
+		return nil
+	}
+
+	e := pb.EpollEvent{
+		Events: uint32(mask) | -syscall.EPOLLET,
+		Fd:     fd,
+	}
+
+	switch {
+	case !fi.waiting && mask != 0:
+		id, c := n.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_EpollCtl{&pb.EpollCtlRequest{Epfd: n.epFD, Op: syscall.EPOLL_CTL_ADD, Fd: fd, Event: &e}}}, false /* ignoreResult */)
+		<-c
+
+		e := n.rpcConn.Request(id).Result.(*pb.SyscallResponse_EpollCtl).EpollCtl.ErrorNumber
+		if e != 0 {
+			return syscall.Errno(e)
+		}
+
+		fi.waiting = true
+	case fi.waiting && mask == 0:
+		id, c := n.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_EpollCtl{&pb.EpollCtlRequest{Epfd: n.epFD, Op: syscall.EPOLL_CTL_DEL, Fd: fd}}}, false /* ignoreResult */)
+		<-c
+		n.rpcConn.Request(id)
+
+		fi.waiting = false
+	case fi.waiting && mask != 0:
+		id, c := n.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_EpollCtl{&pb.EpollCtlRequest{Epfd: n.epFD, Op: syscall.EPOLL_CTL_MOD, Fd: fd, Event: &e}}}, false /* ignoreResult */)
+		<-c
+
+		e := n.rpcConn.Request(id).Result.(*pb.SyscallResponse_EpollCtl).EpollCtl.ErrorNumber
+		if e != 0 {
+			return syscall.Errno(e)
+		}
+	}
+
+	return nil
+}
+
+// addFD adds an FD to the list of FDs observed by n.
+func (n *Notifier) addFD(fd uint32, queue *waiter.Queue) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	// Panic if we're already notifying on this FD.
+	if _, ok := n.fdMap[fd]; ok {
+		panic(fmt.Sprintf("File descriptor %d added twice", fd))
+	}
+
+	// We have nothing to wait for at the moment. Just add it to the map.
+	n.fdMap[fd] = &fdInfo{queue: queue}
+}
+
+// updateFD updates the set of events the FD needs to be notified on.
+func (n *Notifier) updateFD(fd uint32) error {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	if fi, ok := n.fdMap[fd]; ok {
+		return n.waitFD(fd, fi, fi.queue.Events())
+	}
+
+	return nil
+}
+
+// RemoveFD removes an FD from the list of FDs observed by n.
+func (n *Notifier) removeFD(fd uint32) {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	// Remove from map, then from epoll object.
+	n.waitFD(fd, n.fdMap[fd], 0)
+	delete(n.fdMap, fd)
+}
+
+// hasFD returns true if the FD is in the list of observed FDs.
+func (n *Notifier) hasFD(fd uint32) bool {
+	n.mu.Lock()
+	defer n.mu.Unlock()
+
+	_, ok := n.fdMap[fd]
+	return ok
+}
+
+// waitAndNotify loops waiting for io event notifications from the epoll
+// object. Once notifications arrive, they are dispatched to the
+// registered queue.
+func (n *Notifier) waitAndNotify() error {
+	for {
+		id, c := n.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_EpollWait{&pb.EpollWaitRequest{Fd: n.epFD, NumEvents: 100, Msec: -1}}}, false /* ignoreResult */)
+		<-c
+
+		res := n.rpcConn.Request(id).Result.(*pb.SyscallResponse_EpollWait).EpollWait.Result
+		if e, ok := res.(*pb.EpollWaitResponse_ErrorNumber); ok {
+			err := syscall.Errno(e.ErrorNumber)
+			// NOTE: I don't think epoll_wait can return EAGAIN but I'm being
+			// conseratively careful here since exiting the notification thread
+			// would be really bad.
+			if err == syscall.EINTR || err == syscall.EAGAIN {
+				continue
+			}
+			return err
+		}
+
+		n.mu.Lock()
+		for _, e := range res.(*pb.EpollWaitResponse_Events).Events.Events {
+			if fi, ok := n.fdMap[e.Fd]; ok {
+				fi.queue.Notify(waiter.EventMask(e.Events))
+			}
+		}
+		n.mu.Unlock()
+	}
+}
+
+// AddFD adds an FD to the list of observed FDs.
+func (n *Notifier) AddFD(fd uint32, queue *waiter.Queue) error {
+	n.addFD(fd, queue)
+	return nil
+}
+
+// UpdateFD updates the set of events the FD needs to be notified on.
+func (n *Notifier) UpdateFD(fd uint32) error {
+	return n.updateFD(fd)
+}
+
+// RemoveFD removes an FD from the list of observed FDs.
+func (n *Notifier) RemoveFD(fd uint32) {
+	n.removeFD(fd)
+}
+
+// HasFD returns true if the FD is in the list of observed FDs.
+//
+// This should only be used by tests to assert that FDs are correctly
+// registered.
+func (n *Notifier) HasFD(fd uint32) bool {
+	return n.hasFD(fd)
+}
+
+// NonBlockingPoll polls the given fd in non-blocking fashion. It is used just
+// to query the FD's current state; this method will block on the RPC response
+// although the syscall is non-blocking.
+func (n *Notifier) NonBlockingPoll(fd uint32, mask waiter.EventMask) waiter.EventMask {
+	for {
+		id, c := n.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Poll{&pb.PollRequest{Fd: fd, Events: uint32(mask)}}}, false /* ignoreResult */)
+		<-c
+
+		res := n.rpcConn.Request(id).Result.(*pb.SyscallResponse_Poll).Poll.Result
+		if e, ok := res.(*pb.PollResponse_ErrorNumber); ok {
+			if syscall.Errno(e.ErrorNumber) == syscall.EINTR {
+				continue
+			}
+			return mask
+		}
+
+		return waiter.EventMask(res.(*pb.PollResponse_Events).Events)
+	}
+}
diff --git a/pkg/sentry/socket/rpcinet/rpcinet.go b/pkg/sentry/socket/rpcinet/rpcinet.go
new file mode 100644
index 000000000..10b0dedc2
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/rpcinet.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package rpcinet implements sockets using an RPC for each syscall.
+package rpcinet
diff --git a/pkg/sentry/socket/rpcinet/socket.go b/pkg/sentry/socket/rpcinet/socket.go
new file mode 100644
index 000000000..574d99ba5
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/socket.go
@@ -0,0 +1,567 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package rpcinet
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/conn"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/notifier"
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/syscall_rpc_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/buffer"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// socketOperations implements fs.FileOperations and socket.Socket for a socket
+// implemented using a host socket.
+type socketOperations struct {
+	socket.ReceiveTimeout
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+
+	fd       uint32 // must be O_NONBLOCK
+	wq       *waiter.Queue
+	rpcConn  *conn.RPCConnection
+	notifier *notifier.Notifier
+}
+
+// Verify that we actually implement socket.Socket.
+var _ = socket.Socket(&socketOperations{})
+
+// New creates a new RPC socket.
+func newSocketFile(ctx context.Context, stack *Stack, family int, skType int, protocol int) (*fs.File, *syserr.Error) {
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Socket{&pb.SocketRequest{Family: int64(family), Type: int64(skType | syscall.SOCK_NONBLOCK), Protocol: int64(protocol)}}}, false /* ignoreResult */)
+	<-c
+
+	res := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_Socket).Socket.Result
+	if e, ok := res.(*pb.SocketResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+	fd := res.(*pb.SocketResponse_Fd).Fd
+
+	var wq waiter.Queue
+	stack.notifier.AddFD(fd, &wq)
+
+	dirent := socket.NewDirent(ctx, socketDevice)
+	return fs.NewFile(ctx, dirent, fs.FileFlags{Read: true, Write: true}, &socketOperations{
+		wq:       &wq,
+		fd:       fd,
+		rpcConn:  stack.rpcConn,
+		notifier: stack.notifier,
+	}), nil
+}
+
+func isBlockingErrno(err error) bool {
+	return err == syscall.EAGAIN || err == syscall.EWOULDBLOCK
+}
+
+func translateIOSyscallError(err error) error {
+	if isBlockingErrno(err) {
+		return syserror.ErrWouldBlock
+	}
+	return err
+}
+
+// Release implements fs.FileOperations.Release.
+func (s *socketOperations) Release() {
+	s.notifier.RemoveFD(s.fd)
+
+	// We always need to close the FD.
+	_, _ = s.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Close{&pb.CloseRequest{Fd: s.fd}}}, true /* ignoreResult */)
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *socketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return s.notifier.NonBlockingPoll(s.fd, mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *socketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.wq.EventRegister(e, mask)
+	s.notifier.UpdateFD(s.fd)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *socketOperations) EventUnregister(e *waiter.Entry) {
+	s.wq.EventUnregister(e)
+	s.notifier.UpdateFD(s.fd)
+}
+
+func rpcRead(t *kernel.Task, req *pb.SyscallRequest_Read) (*pb.ReadResponse_Data, *syserr.Error) {
+	s := t.NetworkContext().(*Stack)
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Read).Read.Result
+	if e, ok := res.(*pb.ReadResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.ReadResponse_Data), nil
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *socketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	req := &pb.SyscallRequest_Read{&pb.ReadRequest{
+		Fd:     s.fd,
+		Length: uint32(dst.NumBytes()),
+	}}
+
+	res, se := rpcRead(ctx.(*kernel.Task), req)
+	if se == nil {
+		n, e := dst.CopyOut(ctx, res.Data)
+		return int64(n), e
+	}
+	if se != syserr.ErrWouldBlock {
+		return 0, se.ToError()
+	}
+
+	// We'll have to block. Register for notifications and read again when ready.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		res, se := rpcRead(ctx.(*kernel.Task), req)
+		if se == nil {
+			n, e := dst.CopyOut(ctx, res.Data)
+			return int64(n), e
+		}
+		if se != syserr.ErrWouldBlock {
+			return 0, se.ToError()
+		}
+
+		if err := ctx.(*kernel.Task).Block(ch); err != nil {
+			return 0, err
+		}
+	}
+}
+
+func rpcWrite(t *kernel.Task, req *pb.SyscallRequest_Write) (uint32, *syserr.Error) {
+	s := t.NetworkContext().(*Stack)
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Write).Write.Result
+	if e, ok := res.(*pb.WriteResponse_ErrorNumber); ok {
+		return 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.WriteResponse_Length).Length, nil
+}
+
+// Write implements fs.FileOperations.Write.
+func (s *socketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	t := ctx.(*kernel.Task)
+	v := buffer.NewView(int(src.NumBytes()))
+
+	// Copy all the data into the buffer.
+	if _, err := src.CopyIn(t, v); err != nil {
+		return 0, err
+	}
+
+	n, err := rpcWrite(t, &pb.SyscallRequest_Write{&pb.WriteRequest{Fd: s.fd, Data: v}})
+	return int64(n), err.ToError()
+}
+
+func rpcConnect(t *kernel.Task, fd uint32, sockaddr []byte) *syserr.Error {
+	s := t.NetworkContext().(*Stack)
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Connect{&pb.ConnectRequest{Fd: uint32(fd), Address: sockaddr}}}, false /* ignoreResult */)
+	<-c
+
+	if e := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Connect).Connect.ErrorNumber; e != 0 {
+		return syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+// Connect implements socket.Socket.Connect.
+func (s *socketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	if !blocking {
+		return rpcConnect(t, s.fd, sockaddr)
+	}
+
+	// Register for notification when the endpoint becomes writable, then
+	// initiate the connection.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+
+	if err := rpcConnect(t, s.fd, sockaddr); err != syserr.ErrConnectStarted && err != syserr.ErrAlreadyConnecting {
+		return err
+	}
+
+	// It's pending, so we have to wait for a notification, and fetch the
+	// result once the wait completes.
+	if err := t.Block(ch); err != nil {
+		return syserr.FromError(err)
+	}
+
+	// Call Connect() again after blocking to find connect's result.
+	return rpcConnect(t, s.fd, sockaddr)
+}
+
+func rpcAccept(t *kernel.Task, fd uint32, peer bool) (*pb.AcceptResponse_ResultPayload, *syserr.Error) {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Accept{&pb.AcceptRequest{Fd: fd, Peer: peer, Flags: syscall.SOCK_NONBLOCK}}}, false /* ignoreResult */)
+	<-c
+
+	res := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_Accept).Accept.Result
+	if e, ok := res.(*pb.AcceptResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+	return res.(*pb.AcceptResponse_Payload).Payload, nil
+}
+
+// Accept implements socket.Socket.Accept.
+func (s *socketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	payload, se := rpcAccept(t, s.fd, peerRequested)
+
+	// Check if we need to block.
+	if blocking && se == syserr.ErrWouldBlock {
+		// Register for notifications.
+		e, ch := waiter.NewChannelEntry(nil)
+		s.EventRegister(&e, waiter.EventIn)
+		defer s.EventUnregister(&e)
+
+		// Try to accept the connection again; if it fails, then wait until we
+		// get a notification.
+		for {
+			if payload, se = rpcAccept(t, s.fd, peerRequested); se != syserr.ErrWouldBlock {
+				break
+			}
+
+			if err := t.Block(ch); err != nil {
+				return 0, nil, 0, syserr.FromError(err)
+			}
+		}
+	}
+
+	// Handle any error from accept.
+	if se != nil {
+		return 0, nil, 0, se
+	}
+
+	var wq waiter.Queue
+	s.notifier.AddFD(payload.Fd, &wq)
+
+	dirent := socket.NewDirent(t, socketDevice)
+	file := fs.NewFile(t, dirent, fs.FileFlags{Read: true, Write: true, NonBlocking: flags&linux.SOCK_NONBLOCK != 0}, &socketOperations{
+		wq:       &wq,
+		fd:       payload.Fd,
+		notifier: s.notifier,
+	})
+
+	fdFlags := kernel.FDFlags{
+		CloseOnExec: flags&linux.SOCK_CLOEXEC != 0,
+	}
+	fd, err := t.FDMap().NewFDFrom(0, file, fdFlags, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, 0, syserr.FromError(err)
+	}
+
+	return fd, payload.Address.Address, payload.Address.Length, nil
+}
+
+// Bind implements socket.Socket.Bind.
+func (s *socketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Bind{&pb.BindRequest{Fd: s.fd, Address: sockaddr}}}, false /* ignoreResult */)
+	<-c
+
+	if e := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_Bind).Bind.ErrorNumber; e != 0 {
+		syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+// Listen implements socket.Socket.Listen.
+func (s *socketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Listen{&pb.ListenRequest{Fd: s.fd, Backlog: int64(backlog)}}}, false /* ignoreResult */)
+	<-c
+
+	if e := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_Listen).Listen.ErrorNumber; e != 0 {
+		syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+// Shutdown implements socket.Socket.Shutdown.
+func (s *socketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Shutdown{&pb.ShutdownRequest{Fd: s.fd, How: int64(how)}}}, false /* ignoreResult */)
+	<-c
+
+	if e := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_Shutdown).Shutdown.ErrorNumber; e != 0 {
+		return syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+// GetSockOpt implements socket.Socket.GetSockOpt.
+func (s *socketOperations) GetSockOpt(t *kernel.Task, level int, name int, outLen int) (interface{}, *syserr.Error) {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_GetSockOpt{&pb.GetSockOptRequest{Fd: s.fd, Level: int64(level), Name: int64(name), Length: uint32(outLen)}}}, false /* ignoreResult */)
+	<-c
+
+	res := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_GetSockOpt).GetSockOpt.Result
+	if e, ok := res.(*pb.GetSockOptResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.GetSockOptResponse_Opt).Opt, nil
+}
+
+// SetSockOpt implements socket.Socket.SetSockOpt.
+func (s *socketOperations) SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_SetSockOpt{&pb.SetSockOptRequest{Fd: s.fd, Level: int64(level), Name: int64(name), Opt: opt}}}, false /* ignoreResult */)
+	<-c
+
+	if e := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_SetSockOpt).SetSockOpt.ErrorNumber; e != 0 {
+		syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+// GetPeerName implements socket.Socket.GetPeerName.
+func (s *socketOperations) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_GetPeerName{&pb.GetPeerNameRequest{Fd: s.fd}}}, false /* ignoreResult */)
+	<-c
+
+	res := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_GetPeerName).GetPeerName.Result
+	if e, ok := res.(*pb.GetPeerNameResponse_ErrorNumber); ok {
+		return nil, 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	addr := res.(*pb.GetPeerNameResponse_Address).Address
+	return addr.Address, addr.Length, nil
+}
+
+// GetSockName implements socket.Socket.GetSockName.
+func (s *socketOperations) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	stack := t.NetworkContext().(*Stack)
+	id, c := stack.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_GetSockName{&pb.GetSockNameRequest{Fd: s.fd}}}, false /* ignoreResult */)
+	<-c
+
+	res := stack.rpcConn.Request(id).Result.(*pb.SyscallResponse_GetSockName).GetSockName.Result
+	if e, ok := res.(*pb.GetSockNameResponse_ErrorNumber); ok {
+		return nil, 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	addr := res.(*pb.GetSockNameResponse_Address).Address
+	return addr.Address, addr.Length, nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *socketOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return 0, syserror.ENOTTY
+}
+
+func rpcRecvMsg(t *kernel.Task, req *pb.SyscallRequest_Recvmsg) (*pb.RecvmsgResponse_ResultPayload, *syserr.Error) {
+	s := t.NetworkContext().(*Stack)
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Recvmsg).Recvmsg.Result
+	if e, ok := res.(*pb.RecvmsgResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.RecvmsgResponse_Payload).Payload, nil
+}
+
+// RecvMsg implements socket.Socket.RecvMsg.
+func (s *socketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (int, interface{}, uint32, unix.ControlMessages, *syserr.Error) {
+	req := &pb.SyscallRequest_Recvmsg{&pb.RecvmsgRequest{
+		Fd:     s.fd,
+		Length: uint32(dst.NumBytes()),
+		Sender: senderRequested,
+		Trunc:  flags&linux.MSG_TRUNC != 0,
+		Peek:   flags&linux.MSG_PEEK != 0,
+	}}
+
+	res, err := rpcRecvMsg(t, req)
+	if err == nil {
+		n, e := dst.CopyOut(t, res.Data)
+		return int(n), res.Address.GetAddress(), res.Address.GetLength(), unix.ControlMessages{}, syserr.FromError(e)
+	}
+	if err != syserr.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		return 0, nil, 0, unix.ControlMessages{}, err
+	}
+
+	// We'll have to block. Register for notifications and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		res, err := rpcRecvMsg(t, req)
+		if err == nil {
+			n, e := dst.CopyOut(t, res.Data)
+			return int(n), res.Address.GetAddress(), res.Address.GetLength(), unix.ControlMessages{}, syserr.FromError(e)
+		}
+		if err != syserr.ErrWouldBlock {
+			return 0, nil, 0, unix.ControlMessages{}, err
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syserror.ETIMEDOUT {
+				return 0, nil, 0, unix.ControlMessages{}, syserr.ErrTryAgain
+			}
+			return 0, nil, 0, unix.ControlMessages{}, syserr.FromError(err)
+		}
+	}
+}
+
+func rpcSendMsg(t *kernel.Task, req *pb.SyscallRequest_Sendmsg) (uint32, *syserr.Error) {
+	s := t.NetworkContext().(*Stack)
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Sendmsg).Sendmsg.Result
+	if e, ok := res.(*pb.SendmsgResponse_ErrorNumber); ok {
+		return 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.SendmsgResponse_Length).Length, nil
+}
+
+// SendMsg implements socket.Socket.SendMsg.
+func (s *socketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	// Whitelist flags.
+	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_EOR|syscall.MSG_FASTOPEN|syscall.MSG_MORE|syscall.MSG_NOSIGNAL) != 0 {
+		return 0, syserr.ErrInvalidArgument
+	}
+
+	// Reject control messages.
+	if !controlMessages.Empty() {
+		return 0, syserr.ErrInvalidArgument
+	}
+
+	v := buffer.NewView(int(src.NumBytes()))
+
+	// Copy all the data into the buffer.
+	if _, err := src.CopyIn(t, v); err != nil {
+		return 0, syserr.FromError(err)
+	}
+
+	// TODO: this needs to change to map directly to a SendMsg syscall
+	// in the RPC.
+	req := &pb.SyscallRequest_Sendmsg{&pb.SendmsgRequest{
+		Fd:          uint32(s.fd),
+		Data:        v,
+		Address:     to,
+		More:        flags&linux.MSG_MORE != 0,
+		EndOfRecord: flags&linux.MSG_EOR != 0,
+	}}
+
+	n, err := rpcSendMsg(t, req)
+	if err != syserr.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		return int(n), err
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+
+	for {
+		n, err := rpcSendMsg(t, req)
+		if err != syserr.ErrWouldBlock {
+			return int(n), err
+		}
+
+		if err := t.Block(ch); err != nil {
+			return 0, syserr.FromError(err)
+		}
+	}
+}
+
+type socketProvider struct {
+	family int
+}
+
+// Socket implements socket.Provider.Socket.
+func (p *socketProvider) Socket(t *kernel.Task, stypeflags unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Check that we are using the RPC network stack.
+	stack := t.NetworkContext()
+	if stack == nil {
+		return nil, nil
+	}
+
+	s, ok := stack.(*Stack)
+	if !ok {
+		return nil, nil
+	}
+
+	// Only accept TCP and UDP.
+	//
+	// Try to restrict the flags we will accept to minimize backwards
+	// incompatability with netstack.
+	stype := int(stypeflags) & linux.SOCK_TYPE_MASK
+	switch stype {
+	case syscall.SOCK_STREAM:
+		switch protocol {
+		case 0, syscall.IPPROTO_TCP:
+			// ok
+		default:
+			return nil, nil
+		}
+	case syscall.SOCK_DGRAM:
+		switch protocol {
+		case 0, syscall.IPPROTO_UDP:
+			// ok
+		default:
+			return nil, nil
+		}
+	default:
+		return nil, nil
+	}
+
+	return newSocketFile(t, s, p.family, stype, 0)
+}
+
+// Pair implements socket.Provider.Pair.
+func (p *socketProvider) Pair(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *fs.File, *syserr.Error) {
+	// Not supported by AF_INET/AF_INET6.
+	return nil, nil, nil
+}
+
+func init() {
+	for _, family := range []int{syscall.AF_INET, syscall.AF_INET6} {
+		socket.RegisterProvider(family, &socketProvider{family})
+	}
+}
diff --git a/pkg/sentry/socket/rpcinet/stack.go b/pkg/sentry/socket/rpcinet/stack.go
new file mode 100644
index 000000000..503e0e932
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/stack.go
@@ -0,0 +1,175 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package rpcinet
+
+import (
+	"fmt"
+	"strings"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/hostinet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/conn"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/notifier"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/unet"
+)
+
+// Stack implements inet.Stack for RPC backed sockets.
+type Stack struct {
+	// We intentionally do not allow these values to be changed to remain
+	// consistent with the other networking stacks.
+	interfaces     map[int32]inet.Interface
+	interfaceAddrs map[int32][]inet.InterfaceAddr
+	supportsIPv6   bool
+	tcpRecvBufSize inet.TCPBufferSize
+	tcpSendBufSize inet.TCPBufferSize
+	tcpSACKEnabled bool
+	rpcConn        *conn.RPCConnection
+	notifier       *notifier.Notifier
+}
+
+func readTCPBufferSizeFile(conn *conn.RPCConnection, filename string) (inet.TCPBufferSize, error) {
+	contents, se := conn.RPCReadFile(filename)
+	if se != nil {
+		return inet.TCPBufferSize{}, fmt.Errorf("failed to read %s: %v", filename, se)
+	}
+	ioseq := usermem.BytesIOSequence(contents)
+	fields := make([]int32, 3)
+	if n, err := usermem.CopyInt32StringsInVec(context.Background(), ioseq.IO, ioseq.Addrs, fields, ioseq.Opts); n != ioseq.NumBytes() || err != nil {
+		return inet.TCPBufferSize{}, fmt.Errorf("failed to parse %s (%q): got %v after %d/%d bytes", filename, contents, err, n, ioseq.NumBytes())
+	}
+	return inet.TCPBufferSize{
+		Min:     int(fields[0]),
+		Default: int(fields[1]),
+		Max:     int(fields[2]),
+	}, nil
+}
+
+// NewStack returns a Stack containing the current state of the host network
+// stack.
+func NewStack(fd int32) (*Stack, error) {
+	sock, err := unet.NewSocket(int(fd))
+	if err != nil {
+		return nil, err
+	}
+
+	stack := &Stack{
+		interfaces:     make(map[int32]inet.Interface),
+		interfaceAddrs: make(map[int32][]inet.InterfaceAddr),
+		rpcConn:        conn.NewRPCConnection(sock),
+	}
+
+	var e error
+	stack.notifier, e = notifier.NewRPCNotifier(stack.rpcConn)
+	if e != nil {
+		return nil, e
+	}
+
+	// Load the configuration values from procfs.
+	tcpRMem, e := readTCPBufferSizeFile(stack.rpcConn, "/proc/sys/net/ipv4/tcp_rmem")
+	if e != nil {
+		return nil, e
+	}
+	stack.tcpRecvBufSize = tcpRMem
+
+	tcpWMem, e := readTCPBufferSizeFile(stack.rpcConn, "/proc/sys/net/ipv4/tcp_wmem")
+	if e != nil {
+		return nil, e
+	}
+	stack.tcpSendBufSize = tcpWMem
+
+	ipv6, se := stack.rpcConn.RPCReadFile("/proc/net/if_inet6")
+	if len(string(ipv6)) > 0 {
+		stack.supportsIPv6 = true
+	}
+
+	sackFile := "/proc/sys/net/ipv4/tcp_sack"
+	sack, se := stack.rpcConn.RPCReadFile(sackFile)
+	if se != nil {
+		return nil, fmt.Errorf("failed to read %s: %v", sackFile, se)
+	}
+	stack.tcpSACKEnabled = strings.TrimSpace(string(sack)) != "0"
+
+	links, err := stack.DoNetlinkRouteRequest(syscall.RTM_GETLINK)
+	if err != nil {
+		return nil, fmt.Errorf("RTM_GETLINK failed: %v", err)
+	}
+
+	addrs, err := stack.DoNetlinkRouteRequest(syscall.RTM_GETADDR)
+	if err != nil {
+		return nil, fmt.Errorf("RTM_GETADDR failed: %v", err)
+	}
+
+	e = hostinet.ExtractHostInterfaces(links, addrs, stack.interfaces, stack.interfaceAddrs)
+	if e != nil {
+		return nil, e
+	}
+
+	return stack, nil
+}
+
+// Interfaces implements inet.Stack.Interfaces.
+func (s *Stack) Interfaces() map[int32]inet.Interface {
+	return s.interfaces
+}
+
+// InterfaceAddrs implements inet.Stack.InterfaceAddrs.
+func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr {
+	return s.interfaceAddrs
+}
+
+// SupportsIPv6 implements inet.Stack.SupportsIPv6.
+func (s *Stack) SupportsIPv6() bool {
+	return s.supportsIPv6
+}
+
+// TCPReceiveBufferSize implements inet.Stack.TCPReceiveBufferSize.
+func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) {
+	return s.tcpRecvBufSize, nil
+}
+
+// SetTCPReceiveBufferSize implements inet.Stack.SetTCPReceiveBufferSize.
+func (s *Stack) SetTCPReceiveBufferSize(size inet.TCPBufferSize) error {
+	// To keep all the supported stacks consistent we don't allow changing this
+	// value even though it would be possible via an RPC.
+	return syserror.EACCES
+}
+
+// TCPSendBufferSize implements inet.Stack.TCPSendBufferSize.
+func (s *Stack) TCPSendBufferSize() (inet.TCPBufferSize, error) {
+	return s.tcpSendBufSize, nil
+}
+
+// SetTCPSendBufferSize implements inet.Stack.SetTCPSendBufferSize.
+func (s *Stack) SetTCPSendBufferSize(size inet.TCPBufferSize) error {
+	// To keep all the supported stacks consistent we don't allow changing this
+	// value even though it would be possible via an RPC.
+	return syserror.EACCES
+}
+
+// TCPSACKEnabled implements inet.Stack.TCPSACKEnabled.
+func (s *Stack) TCPSACKEnabled() (bool, error) {
+	return s.tcpSACKEnabled, nil
+}
+
+// SetTCPSACKEnabled implements inet.Stack.SetTCPSACKEnabled.
+func (s *Stack) SetTCPSACKEnabled(enabled bool) error {
+	// To keep all the supported stacks consistent we don't allow changing this
+	// value even though it would be possible via an RPC.
+	return syserror.EACCES
+}
diff --git a/pkg/sentry/socket/rpcinet/stack_unsafe.go b/pkg/sentry/socket/rpcinet/stack_unsafe.go
new file mode 100644
index 000000000..9a896c623
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/stack_unsafe.go
@@ -0,0 +1,193 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package rpcinet
+
+import (
+	"syscall"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/socket/rpcinet/syscall_rpc_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+)
+
+// NewNetlinkRouteRequest builds a netlink message for getting the RIB,
+// the routing information base.
+func newNetlinkRouteRequest(proto, seq, family int) []byte {
+	rr := &syscall.NetlinkRouteRequest{}
+	rr.Header.Len = uint32(syscall.NLMSG_HDRLEN + syscall.SizeofRtGenmsg)
+	rr.Header.Type = uint16(proto)
+	rr.Header.Flags = syscall.NLM_F_DUMP | syscall.NLM_F_REQUEST
+	rr.Header.Seq = uint32(seq)
+	rr.Data.Family = uint8(family)
+	return netlinkRRtoWireFormat(rr)
+}
+
+func netlinkRRtoWireFormat(rr *syscall.NetlinkRouteRequest) []byte {
+	b := make([]byte, rr.Header.Len)
+	*(*uint32)(unsafe.Pointer(&b[0:4][0])) = rr.Header.Len
+	*(*uint16)(unsafe.Pointer(&b[4:6][0])) = rr.Header.Type
+	*(*uint16)(unsafe.Pointer(&b[6:8][0])) = rr.Header.Flags
+	*(*uint32)(unsafe.Pointer(&b[8:12][0])) = rr.Header.Seq
+	*(*uint32)(unsafe.Pointer(&b[12:16][0])) = rr.Header.Pid
+	b[16] = byte(rr.Data.Family)
+	return b
+}
+
+func (s *Stack) getNetlinkFd() (uint32, *syserr.Error) {
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Socket{&pb.SocketRequest{Family: int64(syscall.AF_NETLINK), Type: int64(syscall.SOCK_RAW | syscall.SOCK_NONBLOCK), Protocol: int64(syscall.NETLINK_ROUTE)}}}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Socket).Socket.Result
+	if e, ok := res.(*pb.SocketResponse_ErrorNumber); ok {
+		return 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+	return res.(*pb.SocketResponse_Fd).Fd, nil
+}
+
+func (s *Stack) bindNetlinkFd(fd uint32, sockaddr []byte) *syserr.Error {
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Bind{&pb.BindRequest{Fd: fd, Address: sockaddr}}}, false /* ignoreResult */)
+	<-c
+
+	if e := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Bind).Bind.ErrorNumber; e != 0 {
+		return syserr.FromHost(syscall.Errno(e))
+	}
+	return nil
+}
+
+func (s *Stack) closeNetlinkFd(fd uint32) {
+	_, _ = s.rpcConn.NewRequest(pb.SyscallRequest{Args: &pb.SyscallRequest_Close{&pb.CloseRequest{Fd: fd}}}, true /* ignoreResult */)
+}
+
+func (s *Stack) rpcSendMsg(req *pb.SyscallRequest_Sendmsg) (uint32, *syserr.Error) {
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Sendmsg).Sendmsg.Result
+	if e, ok := res.(*pb.SendmsgResponse_ErrorNumber); ok {
+		return 0, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.SendmsgResponse_Length).Length, nil
+}
+
+func (s *Stack) sendMsg(fd uint32, buf []byte, to []byte, flags int) (int, *syserr.Error) {
+	// Whitelist flags.
+	if flags&^(syscall.MSG_DONTWAIT|syscall.MSG_EOR|syscall.MSG_FASTOPEN|syscall.MSG_MORE|syscall.MSG_NOSIGNAL) != 0 {
+		return 0, syserr.ErrInvalidArgument
+	}
+
+	req := &pb.SyscallRequest_Sendmsg{&pb.SendmsgRequest{
+		Fd:          fd,
+		Data:        buf,
+		Address:     to,
+		More:        flags&linux.MSG_MORE != 0,
+		EndOfRecord: flags&linux.MSG_EOR != 0,
+	}}
+
+	n, err := s.rpcSendMsg(req)
+	return int(n), err
+}
+
+func (s *Stack) rpcRecvMsg(req *pb.SyscallRequest_Recvmsg) (*pb.RecvmsgResponse_ResultPayload, *syserr.Error) {
+	id, c := s.rpcConn.NewRequest(pb.SyscallRequest{Args: req}, false /* ignoreResult */)
+	<-c
+
+	res := s.rpcConn.Request(id).Result.(*pb.SyscallResponse_Recvmsg).Recvmsg.Result
+	if e, ok := res.(*pb.RecvmsgResponse_ErrorNumber); ok {
+		return nil, syserr.FromHost(syscall.Errno(e.ErrorNumber))
+	}
+
+	return res.(*pb.RecvmsgResponse_Payload).Payload, nil
+}
+
+func (s *Stack) recvMsg(fd, l, flags uint32) ([]byte, *syserr.Error) {
+	req := &pb.SyscallRequest_Recvmsg{&pb.RecvmsgRequest{
+		Fd:     fd,
+		Length: l,
+		Sender: false,
+		Trunc:  flags&linux.MSG_TRUNC != 0,
+		Peek:   flags&linux.MSG_PEEK != 0,
+	}}
+
+	res, err := s.rpcRecvMsg(req)
+	if err != nil {
+		return nil, err
+	}
+	return res.Data, nil
+}
+
+func (s *Stack) netlinkRequest(proto, family int) ([]byte, error) {
+	fd, err := s.getNetlinkFd()
+	if err != nil {
+		return nil, err.ToError()
+	}
+	defer s.closeNetlinkFd(fd)
+
+	lsa := syscall.SockaddrNetlink{Family: syscall.AF_NETLINK}
+	b := binary.Marshal(nil, usermem.ByteOrder, &lsa)
+	if err := s.bindNetlinkFd(fd, b); err != nil {
+		return nil, err.ToError()
+	}
+
+	wb := newNetlinkRouteRequest(proto, 1, family)
+	_, err = s.sendMsg(fd, wb, b, 0)
+	if err != nil {
+		return nil, err.ToError()
+	}
+
+	var tab []byte
+done:
+	for {
+		rb, err := s.recvMsg(fd, uint32(syscall.Getpagesize()), 0)
+		nr := len(rb)
+		if err != nil {
+			return nil, err.ToError()
+		}
+
+		if nr < syscall.NLMSG_HDRLEN {
+			return nil, syserr.ErrInvalidArgument.ToError()
+		}
+
+		tab = append(tab, rb...)
+		msgs, e := syscall.ParseNetlinkMessage(rb)
+		if e != nil {
+			return nil, e
+		}
+
+		for _, m := range msgs {
+			if m.Header.Type == syscall.NLMSG_DONE {
+				break done
+			}
+			if m.Header.Type == syscall.NLMSG_ERROR {
+				return nil, syserr.ErrInvalidArgument.ToError()
+			}
+		}
+	}
+
+	return tab, nil
+}
+
+// DoNetlinkRouteRequest returns routing information base, also known as RIB,
+// which consists of network facility information, states and parameters.
+func (s *Stack) DoNetlinkRouteRequest(req int) ([]syscall.NetlinkMessage, error) {
+	data, err := s.netlinkRequest(req, syscall.AF_UNSPEC)
+	if err != nil {
+		return nil, err
+	}
+	return syscall.ParseNetlinkMessage(data)
+}
diff --git a/pkg/sentry/socket/rpcinet/syscall_rpc.proto b/pkg/sentry/socket/rpcinet/syscall_rpc.proto
new file mode 100644
index 000000000..b845b1bce
--- /dev/null
+++ b/pkg/sentry/socket/rpcinet/syscall_rpc.proto
@@ -0,0 +1,351 @@
+syntax = "proto3";
+
+// package syscall_rpc is a set of networking related system calls that can be
+// forwarded to a socket gofer.
+//
+// TODO: Document individual RPCs.
+package syscall_rpc;
+
+message SendmsgRequest {
+  uint32 fd = 1;
+  bytes data = 2;
+  bytes address = 3;
+  bool more = 4;
+  bool end_of_record = 5;
+}
+
+message SendmsgResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 length = 2;
+  }
+}
+
+message IOCtlRequest {
+  uint32 fd = 1;
+  uint32 cmd = 2;
+  uint64 arg = 3;
+}
+
+message IOCtlResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint64 value = 2;
+  }
+}
+
+message RecvmsgRequest {
+  uint32 fd = 1;
+  uint32 length = 2;
+  bool sender = 3;
+  bool peek = 4;
+  bool trunc = 5;
+}
+
+message OpenRequest {
+  bytes path = 1;
+  uint32 flags = 2;
+  uint32 mode = 3;
+}
+
+message OpenResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 fd = 2;
+  }
+}
+
+message ReadRequest {
+  uint32 fd = 1;
+  uint32 length = 2;
+}
+
+message ReadResponse {
+  oneof result {
+    uint32 error_number = 1;
+    bytes data = 2;
+  }
+}
+
+message ReadFileRequest {
+  string path = 1;
+}
+
+message ReadFileResponse {
+  oneof result {
+    uint32 error_number = 1;
+    bytes data = 2;
+  }
+}
+
+message WriteRequest {
+  uint32 fd = 1;
+  bytes data = 2;
+}
+
+message WriteResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 length = 2;
+  }
+}
+
+message WriteFileRequest {
+  string path = 1;
+  bytes content = 2;
+}
+
+message WriteFileResponse {
+  uint32 error_number = 1;
+  uint32 written = 2;
+}
+
+message AddressResponse {
+  bytes address = 1;
+  uint32 length = 2;
+}
+
+message RecvmsgResponse {
+  message ResultPayload {
+    bytes data = 1;
+    AddressResponse address = 2;
+    uint32 length = 3;
+  }
+  oneof result {
+    uint32 error_number = 1;
+    ResultPayload payload = 2;
+  }
+}
+
+message BindRequest {
+  uint32 fd = 1;
+  bytes address = 2;
+}
+
+message BindResponse {
+  uint32 error_number = 1;
+}
+
+message AcceptRequest {
+  uint32 fd = 1;
+  bool peer = 2;
+  int64 flags = 3;
+}
+
+message AcceptResponse {
+  message ResultPayload {
+    uint32 fd = 1;
+    AddressResponse address = 2;
+  }
+  oneof result {
+    uint32 error_number = 1;
+    ResultPayload payload = 2;
+  }
+}
+
+message ConnectRequest {
+  uint32 fd = 1;
+  bytes address = 2;
+}
+
+message ConnectResponse {
+  uint32 error_number = 1;
+}
+
+message ListenRequest {
+  uint32 fd = 1;
+  int64 backlog = 2;
+}
+
+message ListenResponse {
+  uint32 error_number = 1;
+}
+
+message ShutdownRequest {
+  uint32 fd = 1;
+  int64 how = 2;
+}
+
+message ShutdownResponse {
+  uint32 error_number = 1;
+}
+
+message CloseRequest {
+  uint32 fd = 1;
+}
+
+message CloseResponse {
+  uint32 error_number = 1;
+}
+
+message GetSockOptRequest {
+  uint32 fd = 1;
+  int64 level = 2;
+  int64 name = 3;
+  uint32 length = 4;
+}
+
+message GetSockOptResponse {
+  oneof result {
+    uint32 error_number = 1;
+    bytes opt = 2;
+  }
+}
+
+message SetSockOptRequest {
+  uint32 fd = 1;
+  int64 level = 2;
+  int64 name = 3;
+  bytes opt = 4;
+}
+
+message SetSockOptResponse {
+  uint32 error_number = 1;
+}
+
+message GetSockNameRequest {
+  uint32 fd = 1;
+}
+
+message GetSockNameResponse {
+  oneof result {
+    uint32 error_number = 1;
+    AddressResponse address = 2;
+  }
+}
+
+message GetPeerNameRequest {
+  uint32 fd = 1;
+}
+
+message GetPeerNameResponse {
+  oneof result {
+    uint32 error_number = 1;
+    AddressResponse address = 2;
+  }
+}
+
+message SocketRequest {
+  int64 family = 1;
+  int64 type = 2;
+  int64 protocol = 3;
+}
+
+message SocketResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 fd = 2;
+  }
+}
+
+message EpollWaitRequest {
+  uint32 fd = 1;
+  uint32 num_events = 2;
+  sint64 msec = 3;
+}
+
+message EpollEvent {
+  uint32 fd = 1;
+  uint32 events = 2;
+}
+
+message EpollEvents {
+  repeated EpollEvent events = 1;
+}
+
+message EpollWaitResponse {
+  oneof result {
+    uint32 error_number = 1;
+    EpollEvents events = 2;
+  }
+}
+
+message EpollCtlRequest {
+  uint32 epfd = 1;
+  int64 op = 2;
+  uint32 fd = 3;
+  EpollEvent event = 4;
+}
+
+message EpollCtlResponse {
+  uint32 error_number = 1;
+}
+
+message EpollCreate1Request {
+  int64 flag = 1;
+}
+
+message EpollCreate1Response {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 fd = 2;
+  }
+}
+
+message PollRequest {
+  uint32 fd = 1;
+  uint32 events = 2;
+}
+
+message PollResponse {
+  oneof result {
+    uint32 error_number = 1;
+    uint32 events = 2;
+  }
+}
+
+message SyscallRequest {
+  oneof args {
+    SocketRequest socket = 1;
+    SendmsgRequest sendmsg = 2;
+    RecvmsgRequest recvmsg = 3;
+    BindRequest bind = 4;
+    AcceptRequest accept = 5;
+    ConnectRequest connect = 6;
+    ListenRequest listen = 7;
+    ShutdownRequest shutdown = 8;
+    CloseRequest close = 9;
+    GetSockOptRequest get_sock_opt = 10;
+    SetSockOptRequest set_sock_opt = 11;
+    GetSockNameRequest get_sock_name = 12;
+    GetPeerNameRequest get_peer_name = 13;
+    EpollWaitRequest epoll_wait = 14;
+    EpollCtlRequest epoll_ctl = 15;
+    EpollCreate1Request epoll_create1 = 16;
+    PollRequest poll = 17;
+    ReadRequest read = 18;
+    WriteRequest write = 19;
+    OpenRequest open = 20;
+    IOCtlRequest ioctl = 21;
+    WriteFileRequest write_file = 22;
+    ReadFileRequest read_file = 23;
+  }
+}
+
+message SyscallResponse {
+  oneof result {
+    SocketResponse socket = 1;
+    SendmsgResponse sendmsg = 2;
+    RecvmsgResponse recvmsg = 3;
+    BindResponse bind = 4;
+    AcceptResponse accept = 5;
+    ConnectResponse connect = 6;
+    ListenResponse listen = 7;
+    ShutdownResponse shutdown = 8;
+    CloseResponse close = 9;
+    GetSockOptResponse get_sock_opt = 10;
+    SetSockOptResponse set_sock_opt = 11;
+    GetSockNameResponse get_sock_name = 12;
+    GetPeerNameResponse get_peer_name = 13;
+    EpollWaitResponse epoll_wait = 14;
+    EpollCtlResponse epoll_ctl = 15;
+    EpollCreate1Response epoll_create1 = 16;
+    PollResponse poll = 17;
+    ReadResponse read = 18;
+    WriteResponse write = 19;
+    OpenResponse open = 20;
+    IOCtlResponse ioctl = 21;
+    WriteFileResponse write_file = 22;
+    ReadFileResponse read_file = 23;
+  }
+}
diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go
new file mode 100644
index 000000000..be3026bfa
--- /dev/null
+++ b/pkg/sentry/socket/socket.go
@@ -0,0 +1,205 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package socket provides the interfaces that need to be provided by socket
+// implementations and providers, as well as per family demultiplexing of socket
+// creation.
+package socket
+
+import (
+	"fmt"
+	"sync/atomic"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/device"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// Socket is the interface containing socket syscalls used by the syscall layer
+// to redirect them to the appropriate implementation.
+type Socket interface {
+	fs.FileOperations
+
+	// Connect implements the connect(2) linux syscall.
+	Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error
+
+	// Accept implements the accept4(2) linux syscall.
+	// Returns fd, real peer address length and error. Real peer address
+	// length is only set if len(peer) > 0.
+	Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error)
+
+	// Bind implements the bind(2) linux syscall.
+	Bind(t *kernel.Task, sockaddr []byte) *syserr.Error
+
+	// Listen implements the listen(2) linux syscall.
+	Listen(t *kernel.Task, backlog int) *syserr.Error
+
+	// Shutdown implements the shutdown(2) linux syscall.
+	Shutdown(t *kernel.Task, how int) *syserr.Error
+
+	// GetSockOpt implements the getsockopt(2) linux syscall.
+	GetSockOpt(t *kernel.Task, level int, name int, outLen int) (interface{}, *syserr.Error)
+
+	// SetSockOpt implements the setsockopt(2) linux syscall.
+	SetSockOpt(t *kernel.Task, level int, name int, opt []byte) *syserr.Error
+
+	// GetSockName implements the getsockname(2) linux syscall.
+	//
+	// addrLen is the address length to be returned to the application, not
+	// necessarily the actual length of the address.
+	GetSockName(t *kernel.Task) (addr interface{}, addrLen uint32, err *syserr.Error)
+
+	// GetPeerName implements the getpeername(2) linux syscall.
+	//
+	// addrLen is the address length to be returned to the application, not
+	// necessarily the actual length of the address.
+	GetPeerName(t *kernel.Task) (addr interface{}, addrLen uint32, err *syserr.Error)
+
+	// RecvMsg implements the recvmsg(2) linux syscall.
+	//
+	// senderAddrLen is the address length to be returned to the application,
+	// not necessarily the actual length of the address.
+	RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, senderAddr interface{}, senderAddrLen uint32, controlMessages unix.ControlMessages, err *syserr.Error)
+
+	// SendMsg implements the sendmsg(2) linux syscall. SendMsg does not take
+	// ownership of the ControlMessage on error.
+	SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (n int, err *syserr.Error)
+
+	// SetRecvTimeout sets the timeout (in ns) for recv operations. Zero means
+	// no timeout.
+	SetRecvTimeout(nanoseconds int64)
+
+	// RecvTimeout gets the current timeout (in ns) for recv operations. Zero
+	// means no timeout.
+	RecvTimeout() int64
+}
+
+// Provider is the interface implemented by providers of sockets for specific
+// address families (e.g., AF_INET).
+type Provider interface {
+	// Socket creates a new socket.
+	//
+	// If a nil Socket _and_ a nil error is returned, it means that the
+	// protocol is not supported. A non-nil error should only be returned
+	// if the protocol is supported, but an error occurs during creation.
+	Socket(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *syserr.Error)
+
+	// Pair creates a pair of connected sockets.
+	//
+	// See Socket for error information.
+	Pair(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *fs.File, *syserr.Error)
+}
+
+// families holds a map of all known address families and their providers.
+var families = make(map[int][]Provider)
+
+// RegisterProvider registers the provider of a given address family so that
+// sockets of that type can be created via socket() and/or socketpair()
+// syscalls.
+func RegisterProvider(family int, provider Provider) {
+	families[family] = append(families[family], provider)
+}
+
+// New creates a new socket with the given family, type and protocol.
+func New(t *kernel.Task, family int, stype unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	for _, p := range families[family] {
+		s, err := p.Socket(t, stype, protocol)
+		if err != nil {
+			return nil, err
+		}
+		if s != nil {
+			return s, nil
+		}
+	}
+
+	return nil, syserr.ErrAddressFamilyNotSupported
+}
+
+// Pair creates a new connected socket pair with the given family, type and
+// protocol.
+func Pair(t *kernel.Task, family int, stype unix.SockType, protocol int) (*fs.File, *fs.File, *syserr.Error) {
+	providers, ok := families[family]
+	if !ok {
+		return nil, nil, syserr.ErrAddressFamilyNotSupported
+	}
+
+	for _, p := range providers {
+		s, t, err := p.Pair(t, stype, protocol)
+		if err != nil {
+			return nil, nil, err
+		}
+		if s != nil && t != nil {
+			return s, t, nil
+		}
+	}
+
+	return nil, nil, syserr.ErrSocketNotSupported
+}
+
+// NewDirent returns a sockfs fs.Dirent that resides on device d.
+func NewDirent(ctx context.Context, d *device.Device) *fs.Dirent {
+	ino := d.NextIno()
+	// There is no real filesystem backing this pipe, so we pass in a nil
+	// Filesystem.
+	inode := fs.NewInode(fsutil.NewSimpleInodeOperations(fsutil.InodeSimpleAttributes{
+		FSType: linux.SOCKFS_MAGIC,
+		UAttr: fs.WithCurrentTime(ctx, fs.UnstableAttr{
+			Owner: fs.FileOwnerFromContext(ctx),
+			Perms: fs.FilePermissions{
+				User: fs.PermMask{Read: true, Write: true},
+			},
+			Links: 1,
+		}),
+	}), fs.NewNonCachingMountSource(nil, fs.MountSourceFlags{}), fs.StableAttr{
+		Type:      fs.Socket,
+		DeviceID:  d.DeviceID(),
+		InodeID:   ino,
+		BlockSize: usermem.PageSize,
+	})
+
+	// Dirent name matches net/socket.c:sockfs_dname.
+	return fs.NewDirent(inode, fmt.Sprintf("socket:[%d]", ino))
+}
+
+// ReceiveTimeout stores a timeout for receive calls.
+//
+// It is meant to be embedded into Socket implementations to help satisfy the
+// interface.
+//
+// Care must be taken when copying ReceiveTimeout as it contains atomic
+// variables.
+type ReceiveTimeout struct {
+	// ns is length of the timeout in nanoseconds.
+	//
+	// ns must be accessed atomically.
+	ns int64
+}
+
+// SetRecvTimeout implements Socket.SetRecvTimeout.
+func (rt *ReceiveTimeout) SetRecvTimeout(nanoseconds int64) {
+	atomic.StoreInt64(&rt.ns, nanoseconds)
+}
+
+// RecvTimeout implements Socket.RecvTimeout.
+func (rt *ReceiveTimeout) RecvTimeout() int64 {
+	return atomic.LoadInt64(&rt.ns)
+}
diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD
new file mode 100644
index 000000000..1ec6eb7ed
--- /dev/null
+++ b/pkg/sentry/socket/unix/BUILD
@@ -0,0 +1,48 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "unix_state",
+    srcs = [
+        "unix.go",
+    ],
+    out = "unix_state.go",
+    package = "unix",
+)
+
+go_library(
+    name = "unix",
+    srcs = [
+        "device.go",
+        "io.go",
+        "unix.go",
+        "unix_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/socket/unix",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/refs",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/device",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/socket/control",
+        "//pkg/sentry/socket/epsocket",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/socket/unix/device.go b/pkg/sentry/socket/unix/device.go
new file mode 100644
index 000000000..e8bcc7a9f
--- /dev/null
+++ b/pkg/sentry/socket/unix/device.go
@@ -0,0 +1,20 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package unix
+
+import "gvisor.googlesource.com/gvisor/pkg/sentry/device"
+
+// unixSocketDevice is the unix socket virtual device.
+var unixSocketDevice = device.NewAnonDevice()
diff --git a/pkg/sentry/socket/unix/io.go b/pkg/sentry/socket/unix/io.go
new file mode 100644
index 000000000..0ca2e35d0
--- /dev/null
+++ b/pkg/sentry/socket/unix/io.go
@@ -0,0 +1,88 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package unix
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// EndpointWriter implements safemem.Writer that writes to a unix.Endpoint.
+//
+// EndpointWriter is not thread-safe.
+type EndpointWriter struct {
+	// Endpoint is the unix.Endpoint to write to.
+	Endpoint unix.Endpoint
+
+	// Control is the control messages to send.
+	Control unix.ControlMessages
+
+	// To is the endpoint to send to. May be nil.
+	To unix.BoundEndpoint
+}
+
+// WriteFromBlocks implements safemem.Writer.WriteFromBlocks.
+func (w *EndpointWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) {
+	return safemem.FromVecWriterFunc{func(bufs [][]byte) (int64, error) {
+		n, err := w.Endpoint.SendMsg(bufs, w.Control, w.To)
+		if err != nil {
+			return int64(n), syserr.TranslateNetstackError(err).ToError()
+		}
+		return int64(n), nil
+	}}.WriteFromBlocks(srcs)
+}
+
+// EndpointReader implements safemem.Reader that reads from a unix.Endpoint.
+//
+// EndpointReader is not thread-safe.
+type EndpointReader struct {
+	// Endpoint is the unix.Endpoint to read from.
+	Endpoint unix.Endpoint
+
+	// Creds indicates if credential control messages are requested.
+	Creds bool
+
+	// NumRights is the number of SCM_RIGHTS FDs requested.
+	NumRights uintptr
+
+	// Peek indicates that the data should not be consumed from the
+	// endpoint.
+	Peek bool
+
+	// MsgSize is the size of the message that was read from. For stream
+	// sockets, it is the amount read.
+	MsgSize uintptr
+
+	// From, if not nil, will be set with the address read from.
+	From *tcpip.FullAddress
+
+	// Control contains the received control messages.
+	Control unix.ControlMessages
+}
+
+// ReadToBlocks implements safemem.Reader.ReadToBlocks.
+func (r *EndpointReader) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) {
+	return safemem.FromVecReaderFunc{func(bufs [][]byte) (int64, error) {
+		n, ms, c, err := r.Endpoint.RecvMsg(bufs, r.Creds, r.NumRights, r.Peek, r.From)
+		r.Control = c
+		r.MsgSize = ms
+		if err != nil {
+			return int64(n), syserr.TranslateNetstackError(err).ToError()
+		}
+		return int64(n), nil
+	}}.ReadToBlocks(dsts)
+}
diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go
new file mode 100644
index 000000000..a4b414851
--- /dev/null
+++ b/pkg/sentry/socket/unix/unix.go
@@ -0,0 +1,571 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package unix provides an implementation of the socket.Socket interface for
+// the AF_UNIX protocol family.
+package unix
+
+import (
+	"strings"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/refs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/fsutil"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/control"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/epsocket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserr"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// SocketOperations is a Unix socket. It is similar to an epsocket, except it is backed
+// by a unix.Endpoint instead of a tcpip.Endpoint.
+type SocketOperations struct {
+	refs.AtomicRefCount
+	socket.ReceiveTimeout
+	fsutil.PipeSeek      `state:"nosave"`
+	fsutil.NotDirReaddir `state:"nosave"`
+	fsutil.NoFsync       `state:"nosave"`
+	fsutil.NoopFlush     `state:"nosave"`
+	fsutil.NoMMap        `state:"nosave"`
+	ep                   unix.Endpoint
+}
+
+// New creates a new unix socket.
+func New(ctx context.Context, endpoint unix.Endpoint) *fs.File {
+	dirent := socket.NewDirent(ctx, unixSocketDevice)
+	return NewWithDirent(ctx, dirent, endpoint, fs.FileFlags{Read: true, Write: true})
+}
+
+// NewWithDirent creates a new unix socket using an existing dirent.
+func NewWithDirent(ctx context.Context, d *fs.Dirent, ep unix.Endpoint, flags fs.FileFlags) *fs.File {
+	return fs.NewFile(ctx, d, flags, &SocketOperations{
+		ep: ep,
+	})
+}
+
+// DecRef implements RefCounter.DecRef.
+func (s *SocketOperations) DecRef() {
+	s.DecRefWithDestructor(func() {
+		s.ep.Close()
+	})
+}
+
+// Release implemements fs.FileOperations.Release.
+func (s *SocketOperations) Release() {
+	// Release only decrements a reference on s because s may be referenced in
+	// the abstract socket namespace.
+	s.DecRef()
+}
+
+// Endpoint extracts the unix.Endpoint.
+func (s *SocketOperations) Endpoint() unix.Endpoint {
+	return s.ep
+}
+
+// extractPath extracts and validates the address.
+func extractPath(sockaddr []byte) (string, *syserr.Error) {
+	addr, err := epsocket.GetAddress(linux.AF_UNIX, sockaddr)
+	if err != nil {
+		return "", err
+	}
+
+	// The address is trimmed by GetAddress.
+	p := string(addr.Addr)
+	if p == "" {
+		// Not allowed.
+		return "", syserr.ErrInvalidArgument
+	}
+	if p[len(p)-1] == '/' {
+		// Weird, they tried to bind '/a/b/c/'?
+		return "", syserr.ErrIsDir
+	}
+
+	return p, nil
+}
+
+// GetPeerName implements the linux syscall getpeername(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) GetPeerName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr, err := s.ep.GetRemoteAddress()
+	if err != nil {
+		return nil, 0, syserr.TranslateNetstackError(err)
+	}
+
+	a, l := epsocket.ConvertAddress(linux.AF_UNIX, addr)
+	return a, l, nil
+}
+
+// GetSockName implements the linux syscall getsockname(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) GetSockName(t *kernel.Task) (interface{}, uint32, *syserr.Error) {
+	addr, err := s.ep.GetLocalAddress()
+	if err != nil {
+		return nil, 0, syserr.TranslateNetstackError(err)
+	}
+
+	a, l := epsocket.ConvertAddress(linux.AF_UNIX, addr)
+	return a, l, nil
+}
+
+// Ioctl implements fs.FileOperations.Ioctl.
+func (s *SocketOperations) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) {
+	return epsocket.Ioctl(ctx, s.ep, io, args)
+}
+
+// GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name, outLen int) (interface{}, *syserr.Error) {
+	return epsocket.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outLen)
+}
+
+// Listen implements the linux syscall listen(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) Listen(t *kernel.Task, backlog int) *syserr.Error {
+	return syserr.TranslateNetstackError(s.ep.Listen(backlog))
+}
+
+// blockingAccept implements a blocking version of accept(2), that is, if no
+// connections are ready to be accept, it will block until one becomes ready.
+func (s *SocketOperations) blockingAccept(t *kernel.Task) (unix.Endpoint, *syserr.Error) {
+	// Register for notifications.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	// Try to accept the connection; if it fails, then wait until we get a
+	// notification.
+	for {
+		if ep, err := s.ep.Accept(); err != tcpip.ErrWouldBlock {
+			return ep, syserr.TranslateNetstackError(err)
+		}
+
+		if err := t.Block(ch); err != nil {
+			return nil, syserr.FromError(err)
+		}
+	}
+}
+
+// Accept implements the linux syscall accept(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (kdefs.FD, interface{}, uint32, *syserr.Error) {
+	// Issue the accept request to get the new endpoint.
+	ep, err := s.ep.Accept()
+	if err != nil {
+		if err != tcpip.ErrWouldBlock || !blocking {
+			return 0, nil, 0, syserr.TranslateNetstackError(err)
+		}
+
+		var err *syserr.Error
+		ep, err = s.blockingAccept(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	ns := New(t, ep)
+	defer ns.DecRef()
+
+	if flags&linux.SOCK_NONBLOCK != 0 {
+		flags := ns.Flags()
+		flags.NonBlocking = true
+		ns.SetFlags(flags.Settable())
+	}
+
+	var addr interface{}
+	var addrLen uint32
+	if peerRequested {
+		// Get address of the peer.
+		var err *syserr.Error
+		addr, addrLen, err = ns.FileOperations.(*SocketOperations).GetPeerName(t)
+		if err != nil {
+			return 0, nil, 0, err
+		}
+	}
+
+	fdFlags := kernel.FDFlags{
+		CloseOnExec: flags&linux.SOCK_CLOEXEC != 0,
+	}
+	fd, e := t.FDMap().NewFDFrom(0, ns, fdFlags, t.ThreadGroup().Limits())
+	if e != nil {
+		return 0, nil, 0, syserr.FromError(e)
+	}
+
+	return fd, addr, addrLen, nil
+}
+
+// Bind implements the linux syscall bind(2) for unix sockets.
+func (s *SocketOperations) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error {
+	p, e := extractPath(sockaddr)
+	if e != nil {
+		return e
+	}
+
+	bep, ok := s.ep.(unix.BoundEndpoint)
+	if !ok {
+		// This socket can't be bound.
+		return syserr.ErrInvalidArgument
+	}
+
+	return syserr.TranslateNetstackError(s.ep.Bind(tcpip.FullAddress{Addr: tcpip.Address(p)}, func() *tcpip.Error {
+		// Is it abstract?
+		if p[0] == 0 {
+			if t.IsNetworkNamespaced() {
+				return tcpip.ErrInvalidEndpointState
+			}
+			if err := t.AbstractSockets().Bind(p[1:], bep, s); err != nil {
+				// tcpip.ErrPortInUse corresponds to EADDRINUSE.
+				return tcpip.ErrPortInUse
+			}
+		} else {
+			// The parent and name.
+			var d *fs.Dirent
+			var name string
+
+			cwd := t.FSContext().WorkingDirectory()
+			defer cwd.DecRef()
+
+			// Is there no slash at all?
+			if !strings.Contains(p, "/") {
+				d = cwd
+				name = p
+			} else {
+				root := t.FSContext().RootDirectory()
+				defer root.DecRef()
+				// Find the last path component, we know that something follows
+				// that final slash, otherwise extractPath() would have failed.
+				lastSlash := strings.LastIndex(p, "/")
+				subPath := p[:lastSlash]
+				if subPath == "" {
+					// Fix up subpath in case file is in root.
+					subPath = "/"
+				}
+				var err error
+				d, err = t.MountNamespace().FindInode(t, root, cwd, subPath, fs.DefaultTraversalLimit)
+				if err != nil {
+					// No path available.
+					return tcpip.ErrNoSuchFile
+				}
+				defer d.DecRef()
+				name = p[lastSlash+1:]
+			}
+
+			// Create the socket.
+			if err := d.Bind(t, t.FSContext().RootDirectory(), name, bep, fs.FilePermissions{User: fs.PermMask{Read: true}}); err != nil {
+				return tcpip.ErrPortInUse
+			}
+		}
+
+		return nil
+	}))
+}
+
+// extractEndpoint retrieves the unix.BoundEndpoint associated with a Unix
+// socket path. The Release must be called on the unix.BoundEndpoint when the
+// caller is done with it.
+func extractEndpoint(t *kernel.Task, sockaddr []byte) (unix.BoundEndpoint, *syserr.Error) {
+	path, err := extractPath(sockaddr)
+	if err != nil {
+		return nil, err
+	}
+
+	// Is it abstract?
+	if path[0] == 0 {
+		if t.IsNetworkNamespaced() {
+			return nil, syserr.ErrInvalidArgument
+		}
+
+		ep := t.AbstractSockets().BoundEndpoint(path[1:])
+		if ep == nil {
+			// No socket found.
+			return nil, syserr.ErrConnectionRefused
+		}
+
+		return ep, nil
+	}
+
+	// Find the node in the filesystem.
+	root := t.FSContext().RootDirectory()
+	cwd := t.FSContext().WorkingDirectory()
+	d, e := t.MountNamespace().FindInode(t, root, cwd, path, fs.DefaultTraversalLimit)
+	cwd.DecRef()
+	root.DecRef()
+	if e != nil {
+		return nil, syserr.FromError(e)
+	}
+
+	// Extract the endpoint if one is there.
+	ep := d.Inode.BoundEndpoint(path)
+	d.DecRef()
+	if ep == nil {
+		// No socket!
+		return nil, syserr.ErrConnectionRefused
+	}
+
+	return ep, nil
+}
+
+// Connect implements the linux syscall connect(2) for unix sockets.
+func (s *SocketOperations) Connect(t *kernel.Task, sockaddr []byte, blocking bool) *syserr.Error {
+	ep, err := extractEndpoint(t, sockaddr)
+	if err != nil {
+		return err
+	}
+	defer ep.Release()
+
+	// Connect the server endpoint.
+	return syserr.TranslateNetstackError(s.ep.Connect(ep))
+}
+
+// Writev implements fs.FileOperations.Write.
+func (s *SocketOperations) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, _ int64) (int64, error) {
+	t := kernel.TaskFromContext(ctx)
+	ctrl := control.New(t, s.ep, nil)
+
+	if src.NumBytes() == 0 {
+		nInt, tcpipError := s.ep.SendMsg([][]byte{}, ctrl, nil)
+		return int64(nInt), syserr.TranslateNetstackError(tcpipError).ToError()
+	}
+
+	return src.CopyInTo(ctx, &EndpointWriter{
+		Endpoint: s.ep,
+		Control:  ctrl,
+		To:       nil,
+	})
+}
+
+// SendMsg implements the linux syscall sendmsg(2) for unix sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) SendMsg(t *kernel.Task, src usermem.IOSequence, to []byte, flags int, controlMessages unix.ControlMessages) (int, *syserr.Error) {
+	w := EndpointWriter{
+		Endpoint: s.ep,
+		Control:  controlMessages,
+		To:       nil,
+	}
+	if len(to) > 0 {
+		ep, err := extractEndpoint(t, to)
+		if err != nil {
+			return 0, err
+		}
+		defer ep.Release()
+		w.To = ep
+	}
+
+	if n, err := src.CopyInTo(t, &w); err != syserror.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		return int(n), syserr.FromError(err)
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventOut)
+	defer s.EventUnregister(&e)
+
+	for {
+		if n, err := src.CopyInTo(t, &w); err != syserror.ErrWouldBlock {
+			return int(n), syserr.FromError(err)
+		}
+
+		if err := t.Block(ch); err != nil {
+			return 0, syserr.FromError(err)
+		}
+	}
+}
+
+// Passcred implements unix.Credentialer.Passcred.
+func (s *SocketOperations) Passcred() bool {
+	return s.ep.Passcred()
+}
+
+// ConnectedPasscred implements unix.Credentialer.ConnectedPasscred.
+func (s *SocketOperations) ConnectedPasscred() bool {
+	return s.ep.ConnectedPasscred()
+}
+
+// Readiness implements waiter.Waitable.Readiness.
+func (s *SocketOperations) Readiness(mask waiter.EventMask) waiter.EventMask {
+	return s.ep.Readiness(mask)
+}
+
+// EventRegister implements waiter.Waitable.EventRegister.
+func (s *SocketOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) {
+	s.ep.EventRegister(e, mask)
+}
+
+// EventUnregister implements waiter.Waitable.EventUnregister.
+func (s *SocketOperations) EventUnregister(e *waiter.Entry) {
+	s.ep.EventUnregister(e)
+}
+
+// SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVal []byte) *syserr.Error {
+	return epsocket.SetSockOpt(t, s, s.ep, level, name, optVal)
+}
+
+// Shutdown implements the linux syscall shutdown(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) Shutdown(t *kernel.Task, how int) *syserr.Error {
+	f, err := epsocket.ConvertShutdown(how)
+	if err != nil {
+		return err
+	}
+
+	// Issue shutdown request.
+	return syserr.TranslateNetstackError(s.ep.Shutdown(f))
+}
+
+// Read implements fs.FileOperations.Read.
+func (s *SocketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) {
+	if dst.NumBytes() == 0 {
+		return 0, nil
+	}
+	return dst.CopyOutFrom(ctx, &EndpointReader{
+		Endpoint:  s.ep,
+		NumRights: 0,
+		Peek:      false,
+		From:      nil,
+	})
+}
+
+// RecvMsg implements the linux syscall recvmsg(2) for sockets backed by
+// a unix.Endpoint.
+func (s *SocketOperations) RecvMsg(t *kernel.Task, dst usermem.IOSequence, flags int, haveDeadline bool, deadline ktime.Time, senderRequested bool, controlDataLen uint64) (n int, senderAddr interface{}, senderAddrLen uint32, controlMessages unix.ControlMessages, err *syserr.Error) {
+	trunc := flags&linux.MSG_TRUNC != 0
+	peek := flags&linux.MSG_PEEK != 0
+
+	// Calculate the number of FDs for which we have space and if we are
+	// requesting credentials.
+	var wantCreds bool
+	rightsLen := int(controlDataLen) - syscall.SizeofCmsghdr
+	if s.Passcred() {
+		// Credentials take priority if they are enabled and there is space.
+		wantCreds = rightsLen > 0
+		credLen := syscall.CmsgSpace(syscall.SizeofUcred)
+		rightsLen -= credLen
+	}
+	// FDs are 32 bit (4 byte) ints.
+	numRights := rightsLen / 4
+	if numRights < 0 {
+		numRights = 0
+	}
+
+	r := EndpointReader{
+		Endpoint:  s.ep,
+		Creds:     wantCreds,
+		NumRights: uintptr(numRights),
+		Peek:      peek,
+	}
+	if senderRequested {
+		r.From = &tcpip.FullAddress{}
+	}
+	if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock || flags&linux.MSG_DONTWAIT != 0 {
+		var from interface{}
+		var fromLen uint32
+		if r.From != nil {
+			from, fromLen = epsocket.ConvertAddress(linux.AF_UNIX, *r.From)
+		}
+		if trunc {
+			n = int64(r.MsgSize)
+		}
+		return int(n), from, fromLen, r.Control, syserr.FromError(err)
+	}
+
+	// We'll have to block. Register for notification and keep trying to
+	// send all the data.
+	e, ch := waiter.NewChannelEntry(nil)
+	s.EventRegister(&e, waiter.EventIn)
+	defer s.EventUnregister(&e)
+
+	for {
+		if n, err := dst.CopyOutFrom(t, &r); err != syserror.ErrWouldBlock {
+			var from interface{}
+			var fromLen uint32
+			if r.From != nil {
+				from, fromLen = epsocket.ConvertAddress(linux.AF_UNIX, *r.From)
+			}
+			if trunc {
+				n = int64(r.MsgSize)
+			}
+			return int(n), from, fromLen, r.Control, syserr.FromError(err)
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syserror.ETIMEDOUT {
+				return 0, nil, 0, unix.ControlMessages{}, syserr.ErrTryAgain
+			}
+			return 0, nil, 0, unix.ControlMessages{}, syserr.FromError(err)
+		}
+	}
+}
+
+// provider is a unix domain socket provider.
+type provider struct{}
+
+// Socket returns a new unix domain socket.
+func (*provider) Socket(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *syserr.Error) {
+	// Check arguments.
+	if protocol != 0 {
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	// Create the endpoint and socket.
+	var ep unix.Endpoint
+	switch stype {
+	case linux.SOCK_DGRAM:
+		ep = unix.NewConnectionless()
+	case linux.SOCK_STREAM, linux.SOCK_SEQPACKET:
+		ep = unix.NewConnectioned(stype, t.Kernel())
+	default:
+		return nil, syserr.ErrInvalidArgument
+	}
+
+	return New(t, ep), nil
+}
+
+// Pair creates a new pair of AF_UNIX connected sockets.
+func (*provider) Pair(t *kernel.Task, stype unix.SockType, protocol int) (*fs.File, *fs.File, *syserr.Error) {
+	// Check arguments.
+	if protocol != 0 {
+		return nil, nil, syserr.ErrInvalidArgument
+	}
+
+	switch stype {
+	case linux.SOCK_STREAM, linux.SOCK_DGRAM, linux.SOCK_SEQPACKET:
+	default:
+		return nil, nil, syserr.ErrInvalidArgument
+	}
+
+	// Create the endpoints and sockets.
+	ep1, ep2 := unix.NewPair(stype, t.Kernel())
+	s1 := New(t, ep1)
+	s2 := New(t, ep2)
+
+	return s1, s2, nil
+}
+
+func init() {
+	socket.RegisterProvider(linux.AF_UNIX, &provider{})
+}
diff --git a/pkg/sentry/state/BUILD b/pkg/sentry/state/BUILD
new file mode 100644
index 000000000..7148df395
--- /dev/null
+++ b/pkg/sentry/state/BUILD
@@ -0,0 +1,21 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "state",
+    srcs = [
+        "state.go",
+        "state_metadata.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/state",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/log",
+        "//pkg/sentry/inet",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/platform",
+        "//pkg/sentry/watchdog",
+        "//pkg/state/statefile",
+    ],
+)
diff --git a/pkg/sentry/state/state.go b/pkg/sentry/state/state.go
new file mode 100644
index 000000000..5bec4e018
--- /dev/null
+++ b/pkg/sentry/state/state.go
@@ -0,0 +1,113 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package state provides high-level state wrappers.
+package state
+
+import (
+	"fmt"
+	"io"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/inet"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/platform"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/watchdog"
+	"gvisor.googlesource.com/gvisor/pkg/state/statefile"
+)
+
+// ErrStateFile is returned when the state file cannot be opened.
+type ErrStateFile struct {
+	err error
+}
+
+// Error implements error.Error().
+func (e ErrStateFile) Error() string {
+	return fmt.Sprintf("failed to open statefile: %v", e.err)
+}
+
+// SaveOpts contains save-related options.
+type SaveOpts struct {
+	// Destination is the save target.
+	Destination io.Writer
+
+	// Key is used for state integrity check.
+	Key []byte
+
+	// Metadata is save metadata.
+	Metadata map[string]string
+
+	// CompressionLevel is the compression level to use.
+	//
+	// See statefile.NewWriter for details.
+	CompressionLevel int
+
+	// Callback is called prior to unpause, with any save error.
+	Callback func(err error)
+}
+
+// Save saves the system state.
+func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error {
+	log.Infof("Sandbox save started, pausing all tasks.")
+	k.Pause()
+	defer k.Unpause()
+	defer log.Infof("Tasks resumed after save.")
+
+	w.Stop()
+	defer w.Start()
+
+	// Supplement the metadata.
+	if opts.Metadata == nil {
+		opts.Metadata = make(map[string]string)
+	}
+	addSaveMetadata(opts.Metadata)
+
+	// Open the statefile.
+	wc, err := statefile.NewWriter(opts.Destination, opts.Key, opts.Metadata, opts.CompressionLevel)
+	if err != nil {
+		err = ErrStateFile{err}
+	} else {
+		// Save the kernel.
+		err = k.SaveTo(wc)
+		if closeErr := wc.Close(); err == nil && closeErr != nil {
+			err = closeErr
+		}
+		if err != nil {
+			err = ErrStateFile{err}
+		}
+	}
+	opts.Callback(err)
+	return err
+}
+
+// LoadOpts contains load-related options.
+type LoadOpts struct {
+	// Destination is the load source.
+	Source io.Reader
+
+	// Key is used for state integrity check.
+	Key []byte
+}
+
+// Load loads the given kernel, setting the provided platform and stack.
+func (opts LoadOpts) Load(k *kernel.Kernel, p platform.Platform, n inet.Stack) error {
+	// Open the file.
+	r, _, err := statefile.NewReader(opts.Source, opts.Key)
+	if err != nil {
+		return ErrStateFile{err}
+	}
+
+	// Restore the Kernel object graph.
+	return k.LoadFrom(r, p, n)
+}
diff --git a/pkg/sentry/state/state_metadata.go b/pkg/sentry/state/state_metadata.go
new file mode 100644
index 000000000..ac374f428
--- /dev/null
+++ b/pkg/sentry/state/state_metadata.go
@@ -0,0 +1,29 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package state
+
+import (
+	"fmt"
+	"time"
+)
+
+// The save metadata keys for timestamp.
+const (
+	metadataTimestamp = "timestamp"
+)
+
+func addSaveMetadata(m map[string]string) {
+	m[metadataTimestamp] = fmt.Sprintf("%v", time.Now())
+}
diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD
new file mode 100644
index 000000000..c5946a564
--- /dev/null
+++ b/pkg/sentry/strace/BUILD
@@ -0,0 +1,48 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "strace",
+    srcs = [
+        "clone.go",
+        "futex.go",
+        "linux64.go",
+        "open.go",
+        "ptrace.go",
+        "socket.go",
+        "strace.go",
+        "syscalls.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/strace",
+    visibility = ["//:sandbox"],
+    deps = [
+        ":strace_go_proto",
+        "//pkg/abi",
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/bits",
+        "//pkg/eventchannel",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/socket/control",
+        "//pkg/sentry/socket/epsocket",
+        "//pkg/sentry/socket/netlink",
+        "//pkg/sentry/syscalls/linux",
+        "//pkg/sentry/usermem",
+    ],
+)
+
+proto_library(
+    name = "strace_proto",
+    srcs = ["strace.proto"],
+    visibility = ["//visibility:public"],
+)
+
+go_proto_library(
+    name = "strace_go_proto",
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/strace/strace_go_proto",
+    proto = ":strace_proto",
+    visibility = ["//visibility:public"],
+)
diff --git a/pkg/sentry/strace/clone.go b/pkg/sentry/strace/clone.go
new file mode 100644
index 000000000..b82ca1ad1
--- /dev/null
+++ b/pkg/sentry/strace/clone.go
@@ -0,0 +1,113 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+)
+
+// CloneFlagSet is the set of clone(2) flags.
+var CloneFlagSet = abi.FlagSet{
+	{
+		Flag: syscall.CLONE_VM,
+		Name: "CLONE_VM",
+	},
+	{
+		Flag: syscall.CLONE_FS,
+		Name: "CLONE_FS",
+	},
+	{
+		Flag: syscall.CLONE_FILES,
+		Name: "CLONE_FILES",
+	},
+	{
+		Flag: syscall.CLONE_SIGHAND,
+		Name: "CLONE_SIGHAND",
+	},
+	{
+		Flag: syscall.CLONE_PTRACE,
+		Name: "CLONE_PTRACE",
+	},
+	{
+		Flag: syscall.CLONE_VFORK,
+		Name: "CLONE_VFORK",
+	},
+	{
+		Flag: syscall.CLONE_PARENT,
+		Name: "CLONE_PARENT",
+	},
+	{
+		Flag: syscall.CLONE_THREAD,
+		Name: "CLONE_THREAD",
+	},
+	{
+		Flag: syscall.CLONE_NEWNS,
+		Name: "CLONE_NEWNS",
+	},
+	{
+		Flag: syscall.CLONE_SYSVSEM,
+		Name: "CLONE_SYSVSEM",
+	},
+	{
+		Flag: syscall.CLONE_SETTLS,
+		Name: "CLONE_SETTLS",
+	},
+	{
+		Flag: syscall.CLONE_PARENT_SETTID,
+		Name: "CLONE_PARENT_SETTID",
+	},
+	{
+		Flag: syscall.CLONE_CHILD_CLEARTID,
+		Name: "CLONE_CHILD_CLEARTID",
+	},
+	{
+		Flag: syscall.CLONE_DETACHED,
+		Name: "CLONE_DETACHED",
+	},
+	{
+		Flag: syscall.CLONE_UNTRACED,
+		Name: "CLONE_UNTRACED",
+	},
+	{
+		Flag: syscall.CLONE_CHILD_SETTID,
+		Name: "CLONE_CHILD_SETTID",
+	},
+	{
+		Flag: syscall.CLONE_NEWUTS,
+		Name: "CLONE_NEWUTS",
+	},
+	{
+		Flag: syscall.CLONE_NEWIPC,
+		Name: "CLONE_NEWIPC",
+	},
+	{
+		Flag: syscall.CLONE_NEWUSER,
+		Name: "CLONE_NEWUSER",
+	},
+	{
+		Flag: syscall.CLONE_NEWPID,
+		Name: "CLONE_NEWPID",
+	},
+	{
+		Flag: syscall.CLONE_NEWNET,
+		Name: "CLONE_NEWNET",
+	},
+	{
+		Flag: syscall.CLONE_IO,
+		Name: "CLONE_IO",
+	},
+}
diff --git a/pkg/sentry/strace/futex.go b/pkg/sentry/strace/futex.go
new file mode 100644
index 000000000..3da108cb7
--- /dev/null
+++ b/pkg/sentry/strace/futex.go
@@ -0,0 +1,91 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+)
+
+// FutexCmd are the possible futex(2) commands.
+var FutexCmd = abi.ValueSet{
+	{
+		Value: linux.FUTEX_WAIT,
+		Name:  "FUTEX_WAIT",
+	},
+	{
+		Value: linux.FUTEX_WAKE,
+		Name:  "FUTEX_WAKE",
+	},
+	{
+		Value: linux.FUTEX_FD,
+		Name:  "FUTEX_FD",
+	},
+	{
+		Value: linux.FUTEX_REQUEUE,
+		Name:  "FUTEX_REQUEUE",
+	},
+	{
+		Value: linux.FUTEX_CMP_REQUEUE,
+		Name:  "FUTEX_CMP_REQUEUE",
+	},
+	{
+		Value: linux.FUTEX_WAKE_OP,
+		Name:  "FUTEX_WAKE_OP",
+	},
+	{
+		Value: linux.FUTEX_LOCK_PI,
+		Name:  "FUTEX_LOCK_PI",
+	},
+	{
+		Value: linux.FUTEX_UNLOCK_PI,
+		Name:  "FUTEX_UNLOCK_PI",
+	},
+	{
+		Value: linux.FUTEX_TRYLOCK_PI,
+		Name:  "FUTEX_TRYLOCK_PI",
+	},
+	{
+		Value: linux.FUTEX_WAIT_BITSET,
+		Name:  "FUTEX_WAIT_BITSET",
+	},
+	{
+		Value: linux.FUTEX_WAKE_BITSET,
+		Name:  "FUTEX_WAKE_BITSET",
+	},
+	{
+		Value: linux.FUTEX_WAIT_REQUEUE_PI,
+		Name:  "FUTEX_WAIT_REQUEUE_PI",
+	},
+	{
+		Value: linux.FUTEX_CMP_REQUEUE_PI,
+		Name:  "FUTEX_CMP_REQUEUE_PI",
+	},
+}
+
+func futex(op uint64) string {
+	cmd := op &^ (linux.FUTEX_PRIVATE_FLAG | linux.FUTEX_CLOCK_REALTIME)
+	clockRealtime := (op & linux.FUTEX_CLOCK_REALTIME) == linux.FUTEX_CLOCK_REALTIME
+	private := (op & linux.FUTEX_PRIVATE_FLAG) == linux.FUTEX_PRIVATE_FLAG
+
+	s := FutexCmd.Parse(cmd)
+	if clockRealtime {
+		s += "|FUTEX_CLOCK_REALTIME"
+	}
+	if private {
+		s += "|FUTEX_PRIVATE_FLAG"
+	}
+	return s
+}
diff --git a/pkg/sentry/strace/linux64.go b/pkg/sentry/strace/linux64.go
new file mode 100644
index 000000000..90ea8c36f
--- /dev/null
+++ b/pkg/sentry/strace/linux64.go
@@ -0,0 +1,338 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+// linuxAMD64 provides a mapping of the Linux amd64 syscalls and their argument
+// types for display / formatting.
+var linuxAMD64 = SyscallMap{
+	0:   makeSyscallInfo("read", Hex, ReadBuffer, Hex),
+	1:   makeSyscallInfo("write", Hex, WriteBuffer, Hex),
+	2:   makeSyscallInfo("open", Path, OpenFlags, Mode),
+	3:   makeSyscallInfo("close", Hex),
+	4:   makeSyscallInfo("stat", Path, Stat),
+	5:   makeSyscallInfo("fstat", Hex, Stat),
+	6:   makeSyscallInfo("lstat", Path, Stat),
+	7:   makeSyscallInfo("poll", Hex, Hex, Hex),
+	8:   makeSyscallInfo("lseek", Hex, Hex, Hex),
+	9:   makeSyscallInfo("mmap", Hex, Hex, Hex, Hex, Hex, Hex),
+	10:  makeSyscallInfo("mprotect", Hex, Hex, Hex),
+	11:  makeSyscallInfo("munmap", Hex, Hex),
+	12:  makeSyscallInfo("brk", Hex),
+	13:  makeSyscallInfo("rt_sigaction", Hex, Hex, Hex),
+	14:  makeSyscallInfo("rt_sigprocmask", Hex, Hex, Hex, Hex),
+	15:  makeSyscallInfo("rt_sigreturn"),
+	16:  makeSyscallInfo("ioctl", Hex, Hex, Hex),
+	17:  makeSyscallInfo("pread64", Hex, ReadBuffer, Hex, Hex),
+	18:  makeSyscallInfo("pwrite64", Hex, WriteBuffer, Hex, Hex),
+	19:  makeSyscallInfo("readv", Hex, ReadIOVec, Hex),
+	20:  makeSyscallInfo("writev", Hex, WriteIOVec, Hex),
+	21:  makeSyscallInfo("access", Path, Oct),
+	22:  makeSyscallInfo("pipe", PipeFDs),
+	23:  makeSyscallInfo("select", Hex, Hex, Hex, Hex, Timeval),
+	24:  makeSyscallInfo("sched_yield"),
+	25:  makeSyscallInfo("mremap", Hex, Hex, Hex, Hex, Hex),
+	26:  makeSyscallInfo("msync", Hex, Hex, Hex),
+	27:  makeSyscallInfo("mincore", Hex, Hex, Hex),
+	28:  makeSyscallInfo("madvise", Hex, Hex, Hex),
+	29:  makeSyscallInfo("shmget", Hex, Hex, Hex),
+	30:  makeSyscallInfo("shmat", Hex, Hex, Hex),
+	31:  makeSyscallInfo("shmctl", Hex, Hex, Hex),
+	32:  makeSyscallInfo("dup", Hex),
+	33:  makeSyscallInfo("dup2", Hex, Hex),
+	34:  makeSyscallInfo("pause"),
+	35:  makeSyscallInfo("nanosleep", Timespec, PostTimespec),
+	36:  makeSyscallInfo("getitimer", Hex, PostItimerVal),
+	37:  makeSyscallInfo("alarm", Hex),
+	38:  makeSyscallInfo("setitimer", Hex, ItimerVal, PostItimerVal),
+	39:  makeSyscallInfo("getpid"),
+	40:  makeSyscallInfo("sendfile", Hex, Hex, Hex, Hex),
+	41:  makeSyscallInfo("socket", SockFamily, SockType, SockProtocol),
+	42:  makeSyscallInfo("connect", Hex, SockAddr, Hex),
+	43:  makeSyscallInfo("accept", Hex, PostSockAddr, SockLen),
+	44:  makeSyscallInfo("sendto", Hex, Hex, Hex, Hex, SockAddr, Hex),
+	45:  makeSyscallInfo("recvfrom", Hex, Hex, Hex, Hex, PostSockAddr, SockLen),
+	46:  makeSyscallInfo("sendmsg", Hex, SendMsgHdr, Hex),
+	47:  makeSyscallInfo("recvmsg", Hex, RecvMsgHdr, Hex),
+	48:  makeSyscallInfo("shutdown", Hex, Hex),
+	49:  makeSyscallInfo("bind", Hex, SockAddr, Hex),
+	50:  makeSyscallInfo("listen", Hex, Hex),
+	51:  makeSyscallInfo("getsockname", Hex, PostSockAddr, SockLen),
+	52:  makeSyscallInfo("getpeername", Hex, PostSockAddr, SockLen),
+	53:  makeSyscallInfo("socketpair", SockFamily, SockType, SockProtocol, Hex),
+	54:  makeSyscallInfo("setsockopt", Hex, Hex, Hex, Hex, Hex),
+	55:  makeSyscallInfo("getsockopt", Hex, Hex, Hex, Hex, Hex),
+	56:  makeSyscallInfo("clone", CloneFlags, Hex, Hex, Hex, Hex),
+	57:  makeSyscallInfo("fork"),
+	58:  makeSyscallInfo("vfork"),
+	59:  makeSyscallInfo("execve", Path, ExecveStringVector, ExecveStringVector),
+	60:  makeSyscallInfo("exit", Hex),
+	61:  makeSyscallInfo("wait4", Hex, Hex, Hex, Rusage),
+	62:  makeSyscallInfo("kill", Hex, Hex),
+	63:  makeSyscallInfo("uname", Uname),
+	64:  makeSyscallInfo("semget", Hex, Hex, Hex),
+	65:  makeSyscallInfo("semop", Hex, Hex, Hex),
+	66:  makeSyscallInfo("semctl", Hex, Hex, Hex, Hex),
+	67:  makeSyscallInfo("shmdt", Hex),
+	68:  makeSyscallInfo("msgget", Hex, Hex),
+	69:  makeSyscallInfo("msgsnd", Hex, Hex, Hex, Hex),
+	70:  makeSyscallInfo("msgrcv", Hex, Hex, Hex, Hex, Hex),
+	71:  makeSyscallInfo("msgctl", Hex, Hex, Hex),
+	72:  makeSyscallInfo("fcntl", Hex, Hex, Hex),
+	73:  makeSyscallInfo("flock", Hex, Hex),
+	74:  makeSyscallInfo("fsync", Hex),
+	75:  makeSyscallInfo("fdatasync", Hex),
+	76:  makeSyscallInfo("truncate", Path, Hex),
+	77:  makeSyscallInfo("ftruncate", Hex, Hex),
+	78:  makeSyscallInfo("getdents", Hex, Hex, Hex),
+	79:  makeSyscallInfo("getcwd", Hex, Hex),
+	80:  makeSyscallInfo("chdir", Path),
+	81:  makeSyscallInfo("fchdir", Hex),
+	82:  makeSyscallInfo("rename", Path, Path),
+	83:  makeSyscallInfo("mkdir", Path, Oct),
+	84:  makeSyscallInfo("rmdir", Path),
+	85:  makeSyscallInfo("creat", Path, Oct),
+	86:  makeSyscallInfo("link", Path, Path),
+	87:  makeSyscallInfo("unlink", Path),
+	88:  makeSyscallInfo("symlink", Path, Path),
+	89:  makeSyscallInfo("readlink", Path, ReadBuffer, Hex),
+	90:  makeSyscallInfo("chmod", Path, Mode),
+	91:  makeSyscallInfo("fchmod", Hex, Mode),
+	92:  makeSyscallInfo("chown", Path, Hex, Hex),
+	93:  makeSyscallInfo("fchown", Hex, Hex, Hex),
+	94:  makeSyscallInfo("lchown", Hex, Hex, Hex),
+	95:  makeSyscallInfo("umask", Hex),
+	96:  makeSyscallInfo("gettimeofday", Timeval, Hex),
+	97:  makeSyscallInfo("getrlimit", Hex, Hex),
+	98:  makeSyscallInfo("getrusage", Hex, Rusage),
+	99:  makeSyscallInfo("sysinfo", Hex),
+	100: makeSyscallInfo("times", Hex),
+	101: makeSyscallInfo("ptrace", PtraceRequest, Hex, Hex, Hex),
+	102: makeSyscallInfo("getuid"),
+	103: makeSyscallInfo("syslog", Hex, Hex, Hex),
+	104: makeSyscallInfo("getgid"),
+	105: makeSyscallInfo("setuid", Hex),
+	106: makeSyscallInfo("setgid", Hex),
+	107: makeSyscallInfo("geteuid"),
+	108: makeSyscallInfo("getegid"),
+	109: makeSyscallInfo("setpgid", Hex, Hex),
+	110: makeSyscallInfo("getppid"),
+	111: makeSyscallInfo("getpgrp"),
+	112: makeSyscallInfo("setsid"),
+	113: makeSyscallInfo("setreuid", Hex, Hex),
+	114: makeSyscallInfo("setregid", Hex, Hex),
+	115: makeSyscallInfo("getgroups", Hex, Hex),
+	116: makeSyscallInfo("setgroups", Hex, Hex),
+	117: makeSyscallInfo("setresuid", Hex, Hex, Hex),
+	118: makeSyscallInfo("getresuid", Hex, Hex, Hex),
+	119: makeSyscallInfo("setresgid", Hex, Hex, Hex),
+	120: makeSyscallInfo("getresgid", Hex, Hex, Hex),
+	121: makeSyscallInfo("getpgid", Hex),
+	122: makeSyscallInfo("setfsuid", Hex),
+	123: makeSyscallInfo("setfsgid", Hex),
+	124: makeSyscallInfo("getsid", Hex),
+	125: makeSyscallInfo("capget", Hex, Hex),
+	126: makeSyscallInfo("capset", Hex, Hex),
+	127: makeSyscallInfo("rt_sigpending", Hex),
+	128: makeSyscallInfo("rt_sigtimedwait", Hex, Hex, Timespec, Hex),
+	129: makeSyscallInfo("rt_sigqueueinfo", Hex, Hex, Hex),
+	130: makeSyscallInfo("rt_sigsuspend", Hex),
+	131: makeSyscallInfo("sigaltstack", Hex, Hex),
+	132: makeSyscallInfo("utime", Path, Utimbuf),
+	133: makeSyscallInfo("mknod", Path, Mode, Hex),
+	134: makeSyscallInfo("uselib", Hex),
+	135: makeSyscallInfo("personality", Hex),
+	136: makeSyscallInfo("ustat", Hex, Hex),
+	137: makeSyscallInfo("statfs", Path, Hex),
+	138: makeSyscallInfo("fstatfs", Hex, Hex),
+	139: makeSyscallInfo("sysfs", Hex, Hex, Hex),
+	140: makeSyscallInfo("getpriority", Hex, Hex),
+	141: makeSyscallInfo("setpriority", Hex, Hex, Hex),
+	142: makeSyscallInfo("sched_setparam", Hex, Hex),
+	143: makeSyscallInfo("sched_getparam", Hex, Hex),
+	144: makeSyscallInfo("sched_setscheduler", Hex, Hex, Hex),
+	145: makeSyscallInfo("sched_getscheduler", Hex),
+	146: makeSyscallInfo("sched_get_priority_max", Hex),
+	147: makeSyscallInfo("sched_get_priority_min", Hex),
+	148: makeSyscallInfo("sched_rr_get_interval", Hex, Hex),
+	149: makeSyscallInfo("mlock", Hex, Hex),
+	150: makeSyscallInfo("munlock", Hex, Hex),
+	151: makeSyscallInfo("mlockall", Hex),
+	152: makeSyscallInfo("munlockall"),
+	153: makeSyscallInfo("vhangup"),
+	154: makeSyscallInfo("modify_ldt", Hex, Hex, Hex),
+	155: makeSyscallInfo("pivot_root", Hex, Hex),
+	156: makeSyscallInfo("_sysctl", Hex),
+	157: makeSyscallInfo("prctl", Hex, Hex, Hex, Hex, Hex),
+	158: makeSyscallInfo("arch_prctl", Hex, Hex),
+	159: makeSyscallInfo("adjtimex", Hex),
+	160: makeSyscallInfo("setrlimit", Hex, Hex),
+	161: makeSyscallInfo("chroot", Path),
+	162: makeSyscallInfo("sync"),
+	163: makeSyscallInfo("acct", Hex),
+	164: makeSyscallInfo("settimeofday", Timeval, Hex),
+	165: makeSyscallInfo("mount", Path, Path, Path, Hex, Path),
+	166: makeSyscallInfo("umount2", Path, Hex),
+	167: makeSyscallInfo("swapon", Hex, Hex),
+	168: makeSyscallInfo("swapoff", Hex),
+	169: makeSyscallInfo("reboot", Hex, Hex, Hex, Hex),
+	170: makeSyscallInfo("sethostname", Hex, Hex),
+	171: makeSyscallInfo("setdomainname", Hex, Hex),
+	172: makeSyscallInfo("iopl", Hex),
+	173: makeSyscallInfo("ioperm", Hex, Hex, Hex),
+	174: makeSyscallInfo("create_module", Path, Hex),
+	175: makeSyscallInfo("init_module", Hex, Hex, Hex),
+	176: makeSyscallInfo("delete_module", Hex, Hex),
+	177: makeSyscallInfo("get_kernel_syms", Hex),
+	// 178: query_module (only present in Linux < 2.6)
+	179: makeSyscallInfo("quotactl", Hex, Hex, Hex, Hex),
+	180: makeSyscallInfo("nfsservctl", Hex, Hex, Hex),
+	// 181: getpmsg (not implemented in the Linux kernel)
+	// 182: putpmsg (not implemented in the Linux kernel)
+	// 183: afs_syscall (not implemented in the Linux kernel)
+	// 184: tuxcall (not implemented in the Linux kernel)
+	// 185: security (not implemented in the Linux kernel)
+	186: makeSyscallInfo("gettid"),
+	187: makeSyscallInfo("readahead", Hex, Hex, Hex),
+	188: makeSyscallInfo("setxattr", Path, Path, Hex, Hex, Hex),
+	189: makeSyscallInfo("lsetxattr", Path, Path, Hex, Hex, Hex),
+	190: makeSyscallInfo("fsetxattr", Hex, Path, Hex, Hex, Hex),
+	191: makeSyscallInfo("getxattr", Path, Path, Hex, Hex),
+	192: makeSyscallInfo("lgetxattr", Path, Path, Hex, Hex),
+	193: makeSyscallInfo("fgetxattr", Hex, Path, Hex, Hex),
+	194: makeSyscallInfo("listxattr", Path, Path, Hex),
+	195: makeSyscallInfo("llistxattr", Path, Path, Hex),
+	196: makeSyscallInfo("flistxattr", Hex, Path, Hex),
+	197: makeSyscallInfo("removexattr", Path, Path),
+	198: makeSyscallInfo("lremovexattr", Path, Path),
+	199: makeSyscallInfo("fremovexattr", Hex, Path),
+	200: makeSyscallInfo("tkill", Hex, Hex),
+	201: makeSyscallInfo("time", Hex),
+	202: makeSyscallInfo("futex", Hex, FutexOp, Hex, Timespec, Hex, Hex),
+	203: makeSyscallInfo("sched_setaffinity", Hex, Hex, Hex),
+	204: makeSyscallInfo("sched_getaffinity", Hex, Hex, Hex),
+	205: makeSyscallInfo("set_thread_area", Hex),
+	206: makeSyscallInfo("io_setup", Hex, Hex),
+	207: makeSyscallInfo("io_destroy", Hex),
+	208: makeSyscallInfo("io_getevents", Hex, Hex, Hex, Hex, Timespec),
+	209: makeSyscallInfo("io_submit", Hex, Hex, Hex),
+	210: makeSyscallInfo("io_cancel", Hex, Hex, Hex),
+	211: makeSyscallInfo("get_thread_area", Hex),
+	212: makeSyscallInfo("lookup_dcookie", Hex, Hex, Hex),
+	213: makeSyscallInfo("epoll_create", Hex),
+	// 214: epoll_ctl_old (not implemented in the Linux kernel)
+	// 215: epoll_wait_old (not implemented in the Linux kernel)
+	216: makeSyscallInfo("remap_file_pages", Hex, Hex, Hex, Hex, Hex),
+	217: makeSyscallInfo("getdents64", Hex, Hex, Hex),
+	218: makeSyscallInfo("set_tid_address", Hex),
+	219: makeSyscallInfo("restart_syscall"),
+	220: makeSyscallInfo("semtimedop", Hex, Hex, Hex, Hex),
+	221: makeSyscallInfo("fadvise64", Hex, Hex, Hex, Hex),
+	222: makeSyscallInfo("timer_create", Hex, Hex, Hex),
+	223: makeSyscallInfo("timer_settime", Hex, Hex, Hex, Hex),
+	224: makeSyscallInfo("timer_gettime", Hex, Hex),
+	225: makeSyscallInfo("timer_getoverrun", Hex),
+	226: makeSyscallInfo("timer_delete", Hex),
+	227: makeSyscallInfo("clock_settime", Hex, Timespec),
+	228: makeSyscallInfo("clock_gettime", Hex, PostTimespec),
+	229: makeSyscallInfo("clock_getres", Hex, PostTimespec),
+	230: makeSyscallInfo("clock_nanosleep", Hex, Hex, Timespec, PostTimespec),
+	231: makeSyscallInfo("exit_group", Hex),
+	232: makeSyscallInfo("epoll_wait", Hex, Hex, Hex, Hex),
+	233: makeSyscallInfo("epoll_ctl", Hex, Hex, Hex, Hex),
+	234: makeSyscallInfo("tgkill", Hex, Hex, Hex),
+	235: makeSyscallInfo("utimes", Path, Timeval),
+	// 236: vserver (not implemented in the Linux kernel)
+	237: makeSyscallInfo("mbind", Hex, Hex, Hex, Hex, Hex, Hex),
+	238: makeSyscallInfo("set_mempolicy", Hex, Hex, Hex),
+	239: makeSyscallInfo("get_mempolicy", Hex, Hex, Hex, Hex, Hex),
+	240: makeSyscallInfo("mq_open", Hex, Hex, Hex, Hex),
+	241: makeSyscallInfo("mq_unlink", Hex),
+	242: makeSyscallInfo("mq_timedsend", Hex, Hex, Hex, Hex, Hex),
+	243: makeSyscallInfo("mq_timedreceive", Hex, Hex, Hex, Hex, Hex),
+	244: makeSyscallInfo("mq_notify", Hex, Hex),
+	245: makeSyscallInfo("mq_getsetattr", Hex, Hex, Hex),
+	246: makeSyscallInfo("kexec_load", Hex, Hex, Hex, Hex),
+	247: makeSyscallInfo("waitid", Hex, Hex, Hex, Hex, Rusage),
+	248: makeSyscallInfo("add_key", Hex, Hex, Hex, Hex, Hex),
+	249: makeSyscallInfo("request_key", Hex, Hex, Hex, Hex),
+	250: makeSyscallInfo("keyctl", Hex, Hex, Hex, Hex, Hex),
+	251: makeSyscallInfo("ioprio_set", Hex, Hex, Hex),
+	252: makeSyscallInfo("ioprio_get", Hex, Hex),
+	253: makeSyscallInfo("inotify_init"),
+	254: makeSyscallInfo("inotify_add_watch", Hex, Hex, Hex),
+	255: makeSyscallInfo("inotify_rm_watch", Hex, Hex),
+	256: makeSyscallInfo("migrate_pages", Hex, Hex, Hex, Hex),
+	257: makeSyscallInfo("openat", Hex, Path, Hex, Mode),
+	258: makeSyscallInfo("mkdirat", Hex, Path, Hex),
+	259: makeSyscallInfo("mknodat", Hex, Path, Mode, Hex),
+	260: makeSyscallInfo("fchownat", Hex, Path, Hex, Hex, Hex),
+	261: makeSyscallInfo("futimesat", Hex, Path, Hex),
+	262: makeSyscallInfo("newfstatat", Hex, Path, Stat, Hex),
+	263: makeSyscallInfo("unlinkat", Hex, Path, Hex),
+	264: makeSyscallInfo("renameat", Hex, Path, Hex, Path),
+	265: makeSyscallInfo("linkat", Hex, Path, Hex, Path, Hex),
+	266: makeSyscallInfo("symlinkat", Path, Hex, Path),
+	267: makeSyscallInfo("readlinkat", Hex, Path, ReadBuffer, Hex),
+	268: makeSyscallInfo("fchmodat", Hex, Path, Mode),
+	269: makeSyscallInfo("faccessat", Hex, Path, Oct, Hex),
+	270: makeSyscallInfo("pselect6", Hex, Hex, Hex, Hex, Hex, Hex),
+	271: makeSyscallInfo("ppoll", Hex, Hex, Timespec, Hex, Hex),
+	272: makeSyscallInfo("unshare", Hex),
+	273: makeSyscallInfo("set_robust_list", Hex, Hex),
+	274: makeSyscallInfo("get_robust_list", Hex, Hex, Hex),
+	275: makeSyscallInfo("splice", Hex, Hex, Hex, Hex, Hex, Hex),
+	276: makeSyscallInfo("tee", Hex, Hex, Hex, Hex),
+	277: makeSyscallInfo("sync_file_range", Hex, Hex, Hex, Hex),
+	278: makeSyscallInfo("vmsplice", Hex, Hex, Hex, Hex),
+	279: makeSyscallInfo("move_pages", Hex, Hex, Hex, Hex, Hex, Hex),
+	280: makeSyscallInfo("utimensat", Hex, Path, UTimeTimespec, Hex),
+	281: makeSyscallInfo("epoll_pwait", Hex, Hex, Hex, Hex, Hex, Hex),
+	282: makeSyscallInfo("signalfd", Hex, Hex, Hex),
+	283: makeSyscallInfo("timerfd_create", Hex, Hex),
+	284: makeSyscallInfo("eventfd", Hex),
+	285: makeSyscallInfo("fallocate", Hex, Hex, Hex, Hex),
+	286: makeSyscallInfo("timerfd_settime", Hex, Hex, Hex, Hex),
+	287: makeSyscallInfo("timerfd_gettime", Hex, Hex),
+	288: makeSyscallInfo("accept4", Hex, PostSockAddr, SockLen, SockFlags),
+	289: makeSyscallInfo("signalfd4", Hex, Hex, Hex, Hex),
+	290: makeSyscallInfo("eventfd2", Hex, Hex),
+	291: makeSyscallInfo("epoll_create1", Hex),
+	292: makeSyscallInfo("dup3", Hex, Hex, Hex),
+	293: makeSyscallInfo("pipe2", PipeFDs, Hex),
+	294: makeSyscallInfo("inotify_init1", Hex),
+	295: makeSyscallInfo("preadv", Hex, ReadIOVec, Hex, Hex),
+	296: makeSyscallInfo("pwritev", Hex, WriteIOVec, Hex, Hex),
+	297: makeSyscallInfo("rt_tgsigqueueinfo", Hex, Hex, Hex, Hex),
+	298: makeSyscallInfo("perf_event_open", Hex, Hex, Hex, Hex, Hex),
+	299: makeSyscallInfo("recvmmsg", Hex, Hex, Hex, Hex, Hex),
+	300: makeSyscallInfo("fanotify_init", Hex, Hex),
+	301: makeSyscallInfo("fanotify_mark", Hex, Hex, Hex, Hex, Hex),
+	302: makeSyscallInfo("prlimit64", Hex, Hex, Hex, Hex),
+	303: makeSyscallInfo("name_to_handle_at", Hex, Hex, Hex, Hex, Hex),
+	304: makeSyscallInfo("open_by_handle_at", Hex, Hex, Hex),
+	305: makeSyscallInfo("clock_adjtime", Hex, Hex),
+	306: makeSyscallInfo("syncfs", Hex),
+	307: makeSyscallInfo("sendmmsg", Hex, Hex, Hex, Hex),
+	308: makeSyscallInfo("setns", Hex, Hex),
+	309: makeSyscallInfo("getcpu", Hex, Hex, Hex),
+	310: makeSyscallInfo("process_vm_readv", Hex, ReadIOVec, Hex, IOVec, Hex, Hex),
+	311: makeSyscallInfo("process_vm_writev", Hex, IOVec, Hex, WriteIOVec, Hex, Hex),
+	312: makeSyscallInfo("kcmp", Hex, Hex, Hex, Hex, Hex),
+	313: makeSyscallInfo("finit_module", Hex, Hex, Hex),
+	314: makeSyscallInfo("sched_setattr", Hex, Hex, Hex),
+	315: makeSyscallInfo("sched_getattr", Hex, Hex, Hex),
+	316: makeSyscallInfo("renameat2", Hex, Path, Hex, Path, Hex),
+	317: makeSyscallInfo("seccomp", Hex, Hex, Hex),
+}
diff --git a/pkg/sentry/strace/open.go b/pkg/sentry/strace/open.go
new file mode 100644
index 000000000..839d5eda7
--- /dev/null
+++ b/pkg/sentry/strace/open.go
@@ -0,0 +1,105 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+)
+
+// OpenMode represents the mode to open(2) a file.
+var OpenMode = abi.ValueSet{
+	{
+		Value: syscall.O_RDWR,
+		Name:  "O_RDWR",
+	},
+	{
+		Value: syscall.O_WRONLY,
+		Name:  "O_WRONLY",
+	},
+	{
+		Value: syscall.O_RDONLY,
+		Name:  "O_RDONLY",
+	},
+}
+
+// OpenFlagSet is the set of open(2) flags.
+var OpenFlagSet = abi.FlagSet{
+	{
+		Flag: syscall.O_APPEND,
+		Name: "O_APPEND",
+	},
+	{
+		Flag: syscall.O_ASYNC,
+		Name: "O_ASYNC",
+	},
+	{
+		Flag: syscall.O_CLOEXEC,
+		Name: "O_CLOEXEC",
+	},
+	{
+		Flag: syscall.O_CREAT,
+		Name: "O_CREAT",
+	},
+	{
+		Flag: syscall.O_DIRECT,
+		Name: "O_DIRECT",
+	},
+	{
+		Flag: syscall.O_DIRECTORY,
+		Name: "O_DIRECTORY",
+	},
+	{
+		Flag: syscall.O_EXCL,
+		Name: "O_EXCL",
+	},
+	{
+		Flag: syscall.O_NOATIME,
+		Name: "O_NOATIME",
+	},
+	{
+		Flag: syscall.O_NOCTTY,
+		Name: "O_NOCTTY",
+	},
+	{
+		Flag: syscall.O_NOFOLLOW,
+		Name: "O_NOFOLLOW",
+	},
+	{
+		Flag: syscall.O_NONBLOCK,
+		Name: "O_NONBLOCK",
+	},
+	{
+		Flag: 0x200000, // O_PATH
+		Name: "O_PATH",
+	},
+	{
+		Flag: syscall.O_SYNC,
+		Name: "O_SYNC",
+	},
+	{
+		Flag: syscall.O_TRUNC,
+		Name: "O_TRUNC",
+	},
+}
+
+func open(val uint64) string {
+	s := OpenMode.Parse(val & syscall.O_ACCMODE)
+	if flags := OpenFlagSet.Parse(val &^ syscall.O_ACCMODE); flags != "" {
+		s += "|" + flags
+	}
+	return s
+}
diff --git a/pkg/sentry/strace/ptrace.go b/pkg/sentry/strace/ptrace.go
new file mode 100644
index 000000000..a0dabb27a
--- /dev/null
+++ b/pkg/sentry/strace/ptrace.go
@@ -0,0 +1,178 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// PtraceRequestSet are the possible ptrace(2) requests.
+var PtraceRequestSet = abi.ValueSet{
+	{
+		Value: syscall.PTRACE_TRACEME,
+		Name:  "PTRACE_TRACEME",
+	},
+	{
+		Value: syscall.PTRACE_PEEKTEXT,
+		Name:  "PTRACE_PEEKTEXT",
+	},
+	{
+		Value: syscall.PTRACE_PEEKDATA,
+		Name:  "PTRACE_PEEKDATA",
+	},
+	{
+		Value: syscall.PTRACE_PEEKUSR,
+		Name:  "PTRACE_PEEKUSR",
+	},
+	{
+		Value: syscall.PTRACE_POKETEXT,
+		Name:  "PTRACE_POKETEXT",
+	},
+	{
+		Value: syscall.PTRACE_POKEDATA,
+		Name:  "PTRACE_POKEDATA",
+	},
+	{
+		Value: syscall.PTRACE_POKEUSR,
+		Name:  "PTRACE_POKEUSR",
+	},
+	{
+		Value: syscall.PTRACE_CONT,
+		Name:  "PTRACE_CONT",
+	},
+	{
+		Value: syscall.PTRACE_KILL,
+		Name:  "PTRACE_KILL",
+	},
+	{
+		Value: syscall.PTRACE_SINGLESTEP,
+		Name:  "PTRACE_SINGLESTEP",
+	},
+	{
+		Value: syscall.PTRACE_ATTACH,
+		Name:  "PTRACE_ATTACH",
+	},
+	{
+		Value: syscall.PTRACE_DETACH,
+		Name:  "PTRACE_DETACH",
+	},
+	{
+		Value: syscall.PTRACE_SYSCALL,
+		Name:  "PTRACE_SYSCALL",
+	},
+	{
+		Value: syscall.PTRACE_SETOPTIONS,
+		Name:  "PTRACE_SETOPTIONS",
+	},
+	{
+		Value: syscall.PTRACE_GETEVENTMSG,
+		Name:  "PTRACE_GETEVENTMSG",
+	},
+	{
+		Value: syscall.PTRACE_GETSIGINFO,
+		Name:  "PTRACE_GETSIGINFO",
+	},
+	{
+		Value: syscall.PTRACE_SETSIGINFO,
+		Name:  "PTRACE_SETSIGINFO",
+	},
+	{
+		Value: syscall.PTRACE_GETREGSET,
+		Name:  "PTRACE_GETREGSET",
+	},
+	{
+		Value: syscall.PTRACE_SETREGSET,
+		Name:  "PTRACE_SETREGSET",
+	},
+	{
+		Value: kernel.PTRACE_SEIZE,
+		Name:  "PTRACE_SEIZE",
+	},
+	{
+		Value: kernel.PTRACE_INTERRUPT,
+		Name:  "PTRACE_INTERRUPT",
+	},
+	{
+		Value: kernel.PTRACE_LISTEN,
+		Name:  "PTRACE_LISTEN",
+	},
+	{
+		Value: kernel.PTRACE_PEEKSIGINFO,
+		Name:  "PTRACE_PEEKSIGINFO",
+	},
+	{
+		Value: kernel.PTRACE_GETSIGMASK,
+		Name:  "PTRACE_GETSIGMASK",
+	},
+	{
+		Value: kernel.PTRACE_SETSIGMASK,
+		Name:  "PTRACE_SETSIGMASK",
+	},
+	{
+		Value: syscall.PTRACE_GETREGS,
+		Name:  "PTRACE_GETREGS",
+	},
+	{
+		Value: syscall.PTRACE_SETREGS,
+		Name:  "PTRACE_SETREGS",
+	},
+	{
+		Value: syscall.PTRACE_GETFPREGS,
+		Name:  "PTRACE_GETFPREGS",
+	},
+	{
+		Value: syscall.PTRACE_SETFPREGS,
+		Name:  "PTRACE_SETFPREGS",
+	},
+	{
+		Value: syscall.PTRACE_GETFPXREGS,
+		Name:  "PTRACE_GETFPXREGS",
+	},
+	{
+		Value: syscall.PTRACE_SETFPXREGS,
+		Name:  "PTRACE_SETFPXREGS",
+	},
+	{
+		Value: syscall.PTRACE_OLDSETOPTIONS,
+		Name:  "PTRACE_OLDSETOPTIONS",
+	},
+	{
+		Value: syscall.PTRACE_GET_THREAD_AREA,
+		Name:  "PTRACE_GET_THREAD_AREA",
+	},
+	{
+		Value: syscall.PTRACE_SET_THREAD_AREA,
+		Name:  "PTRACE_SET_THREAD_AREA",
+	},
+	{
+		Value: syscall.PTRACE_ARCH_PRCTL,
+		Name:  "PTRACE_ARCH_PRCTL",
+	},
+	{
+		Value: syscall.PTRACE_SYSEMU,
+		Name:  "PTRACE_SYSEMU",
+	},
+	{
+		Value: syscall.PTRACE_SYSEMU_SINGLESTEP,
+		Name:  "PTRACE_SYSEMU_SINGLESTEP",
+	},
+	{
+		Value: syscall.PTRACE_SINGLEBLOCK,
+		Name:  "PTRACE_SINGLEBLOCK",
+	},
+}
diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go
new file mode 100644
index 000000000..48c072e96
--- /dev/null
+++ b/pkg/sentry/strace/socket.go
@@ -0,0 +1,674 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"fmt"
+	"strings"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/control"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/epsocket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/netlink"
+	slinux "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// SocketFamily are the possible socket(2) families.
+var SocketFamily = abi.ValueSet{
+	{
+		Value: linux.AF_UNSPEC,
+		Name:  "AF_UNSPEC",
+	},
+	{
+		Value: linux.AF_UNIX,
+		Name:  "AF_UNIX",
+	},
+	{
+		Value: linux.AF_INET,
+		Name:  "AF_INET",
+	},
+	{
+		Value: linux.AF_AX25,
+		Name:  "AF_AX25",
+	},
+	{
+		Value: linux.AF_IPX,
+		Name:  "AF_IPX",
+	},
+	{
+		Value: linux.AF_APPLETALK,
+		Name:  "AF_APPLETALK",
+	},
+	{
+		Value: linux.AF_NETROM,
+		Name:  "AF_NETROM",
+	},
+	{
+		Value: linux.AF_BRIDGE,
+		Name:  "AF_BRIDGE",
+	},
+	{
+		Value: linux.AF_ATMPVC,
+		Name:  "AF_ATMPVC",
+	},
+	{
+		Value: linux.AF_X25,
+		Name:  "AF_X25",
+	},
+	{
+		Value: linux.AF_INET6,
+		Name:  "AF_INET6",
+	},
+	{
+		Value: linux.AF_ROSE,
+		Name:  "AF_ROSE",
+	},
+	{
+		Value: linux.AF_DECnet,
+		Name:  "AF_DECnet",
+	},
+	{
+		Value: linux.AF_NETBEUI,
+		Name:  "AF_NETBEUI",
+	},
+	{
+		Value: linux.AF_SECURITY,
+		Name:  "AF_SECURITY",
+	},
+	{
+		Value: linux.AF_KEY,
+		Name:  "AF_KEY",
+	},
+	{
+		Value: linux.AF_NETLINK,
+		Name:  "AF_NETLINK",
+	},
+	{
+		Value: linux.AF_PACKET,
+		Name:  "AF_PACKET",
+	},
+	{
+		Value: linux.AF_ASH,
+		Name:  "AF_ASH",
+	},
+	{
+		Value: linux.AF_ECONET,
+		Name:  "AF_ECONET",
+	},
+	{
+		Value: linux.AF_ATMSVC,
+		Name:  "AF_ATMSVC",
+	},
+	{
+		Value: linux.AF_RDS,
+		Name:  "AF_RDS",
+	},
+	{
+		Value: linux.AF_SNA,
+		Name:  "AF_SNA",
+	},
+	{
+		Value: linux.AF_IRDA,
+		Name:  "AF_IRDA",
+	},
+	{
+		Value: linux.AF_PPPOX,
+		Name:  "AF_PPPOX",
+	},
+	{
+		Value: linux.AF_WANPIPE,
+		Name:  "AF_WANPIPE",
+	},
+	{
+		Value: linux.AF_LLC,
+		Name:  "AF_LLC",
+	},
+	{
+		Value: linux.AF_IB,
+		Name:  "AF_IB",
+	},
+	{
+		Value: linux.AF_MPLS,
+		Name:  "AF_MPLS",
+	},
+	{
+		Value: linux.AF_CAN,
+		Name:  "AF_CAN",
+	},
+	{
+		Value: linux.AF_TIPC,
+		Name:  "AF_TIPC",
+	},
+	{
+		Value: linux.AF_BLUETOOTH,
+		Name:  "AF_BLUETOOTH",
+	},
+	{
+		Value: linux.AF_IUCV,
+		Name:  "AF_IUCV",
+	},
+	{
+		Value: linux.AF_RXRPC,
+		Name:  "AF_RXRPC",
+	},
+	{
+		Value: linux.AF_ISDN,
+		Name:  "AF_ISDN",
+	},
+	{
+		Value: linux.AF_PHONET,
+		Name:  "AF_PHONET",
+	},
+	{
+		Value: linux.AF_IEEE802154,
+		Name:  "AF_IEEE802154",
+	},
+	{
+		Value: linux.AF_CAIF,
+		Name:  "AF_CAIF",
+	},
+	{
+		Value: linux.AF_ALG,
+		Name:  "AF_ALG",
+	},
+	{
+		Value: linux.AF_NFC,
+		Name:  "AF_NFC",
+	},
+	{
+		Value: linux.AF_VSOCK,
+		Name:  "AF_VSOCK",
+	},
+}
+
+// SocketType are the possible socket(2) types.
+var SocketType = abi.ValueSet{
+	{
+		Value: linux.SOCK_STREAM,
+		Name:  "SOCK_STREAM",
+	},
+	{
+		Value: linux.SOCK_DGRAM,
+		Name:  "SOCK_DGRAM",
+	},
+	{
+		Value: linux.SOCK_RAW,
+		Name:  "SOCK_RAW",
+	},
+	{
+		Value: linux.SOCK_RDM,
+		Name:  "SOCK_RDM",
+	},
+	{
+		Value: linux.SOCK_SEQPACKET,
+		Name:  "SOCK_SEQPACKET",
+	},
+	{
+		Value: linux.SOCK_DCCP,
+		Name:  "SOCK_DCCP",
+	},
+	{
+		Value: linux.SOCK_PACKET,
+		Name:  "SOCK_PACKET",
+	},
+}
+
+// SocketFlagSet are the possible socket(2) flags.
+var SocketFlagSet = abi.FlagSet{
+	{
+		Flag: linux.SOCK_CLOEXEC,
+		Name: "SOCK_CLOEXEC",
+	},
+	{
+		Flag: linux.SOCK_NONBLOCK,
+		Name: "SOCK_NONBLOCK",
+	},
+}
+
+// ipProtocol are the possible socket(2) types for INET and INET6 sockets.
+var ipProtocol = abi.ValueSet{
+	{
+		Value: linux.IPPROTO_IP,
+		Name:  "IPPROTO_IP",
+	},
+	{
+		Value: linux.IPPROTO_ICMP,
+		Name:  "IPPROTO_ICMP",
+	},
+	{
+		Value: linux.IPPROTO_IGMP,
+		Name:  "IPPROTO_IGMP",
+	},
+	{
+		Value: linux.IPPROTO_IPIP,
+		Name:  "IPPROTO_IPIP",
+	},
+	{
+		Value: linux.IPPROTO_TCP,
+		Name:  "IPPROTO_TCP",
+	},
+	{
+		Value: linux.IPPROTO_EGP,
+		Name:  "IPPROTO_EGP",
+	},
+	{
+		Value: linux.IPPROTO_PUP,
+		Name:  "IPPROTO_PUP",
+	},
+	{
+		Value: linux.IPPROTO_UDP,
+		Name:  "IPPROTO_UDP",
+	},
+	{
+		Value: linux.IPPROTO_IDP,
+		Name:  "IPPROTO_IDP",
+	},
+	{
+		Value: linux.IPPROTO_TP,
+		Name:  "IPPROTO_TP",
+	},
+	{
+		Value: linux.IPPROTO_DCCP,
+		Name:  "IPPROTO_DCCP",
+	},
+	{
+		Value: linux.IPPROTO_IPV6,
+		Name:  "IPPROTO_IPV6",
+	},
+	{
+		Value: linux.IPPROTO_RSVP,
+		Name:  "IPPROTO_RSVP",
+	},
+	{
+		Value: linux.IPPROTO_GRE,
+		Name:  "IPPROTO_GRE",
+	},
+	{
+		Value: linux.IPPROTO_ESP,
+		Name:  "IPPROTO_ESP",
+	},
+	{
+		Value: linux.IPPROTO_AH,
+		Name:  "IPPROTO_AH",
+	},
+	{
+		Value: linux.IPPROTO_MTP,
+		Name:  "IPPROTO_MTP",
+	},
+	{
+		Value: linux.IPPROTO_BEETPH,
+		Name:  "IPPROTO_BEETPH",
+	},
+	{
+		Value: linux.IPPROTO_ENCAP,
+		Name:  "IPPROTO_ENCAP",
+	},
+	{
+		Value: linux.IPPROTO_PIM,
+		Name:  "IPPROTO_PIM",
+	},
+	{
+		Value: linux.IPPROTO_COMP,
+		Name:  "IPPROTO_COMP",
+	},
+	{
+		Value: linux.IPPROTO_SCTP,
+		Name:  "IPPROTO_SCTP",
+	},
+	{
+		Value: linux.IPPROTO_UDPLITE,
+		Name:  "IPPROTO_UDPLITE",
+	},
+	{
+		Value: linux.IPPROTO_MPLS,
+		Name:  "IPPROTO_MPLS",
+	},
+	{
+		Value: linux.IPPROTO_RAW,
+		Name:  "IPPROTO_RAW",
+	},
+}
+
+// SocketProtocol are the possible socket(2) protocols for each protocol family.
+var SocketProtocol = map[int32]abi.ValueSet{
+	linux.AF_INET:  ipProtocol,
+	linux.AF_INET6: ipProtocol,
+	linux.AF_NETLINK: {
+		{
+			Value: linux.NETLINK_ROUTE,
+			Name:  "NETLINK_ROUTE",
+		},
+		{
+			Value: linux.NETLINK_UNUSED,
+			Name:  "NETLINK_UNUSED",
+		},
+		{
+			Value: linux.NETLINK_USERSOCK,
+			Name:  "NETLINK_USERSOCK",
+		},
+		{
+			Value: linux.NETLINK_FIREWALL,
+			Name:  "NETLINK_FIREWALL",
+		},
+		{
+			Value: linux.NETLINK_SOCK_DIAG,
+			Name:  "NETLINK_SOCK_DIAG",
+		},
+		{
+			Value: linux.NETLINK_NFLOG,
+			Name:  "NETLINK_NFLOG",
+		},
+		{
+			Value: linux.NETLINK_XFRM,
+			Name:  "NETLINK_XFRM",
+		},
+		{
+			Value: linux.NETLINK_SELINUX,
+			Name:  "NETLINK_SELINUX",
+		},
+		{
+			Value: linux.NETLINK_ISCSI,
+			Name:  "NETLINK_ISCSI",
+		},
+		{
+			Value: linux.NETLINK_AUDIT,
+			Name:  "NETLINK_AUDIT",
+		},
+		{
+			Value: linux.NETLINK_FIB_LOOKUP,
+			Name:  "NETLINK_FIB_LOOKUP",
+		},
+		{
+			Value: linux.NETLINK_CONNECTOR,
+			Name:  "NETLINK_CONNECTOR",
+		},
+		{
+			Value: linux.NETLINK_NETFILTER,
+			Name:  "NETLINK_NETFILTER",
+		},
+		{
+			Value: linux.NETLINK_IP6_FW,
+			Name:  "NETLINK_IP6_FW",
+		},
+		{
+			Value: linux.NETLINK_DNRTMSG,
+			Name:  "NETLINK_DNRTMSG",
+		},
+		{
+			Value: linux.NETLINK_KOBJECT_UEVENT,
+			Name:  "NETLINK_KOBJECT_UEVENT",
+		},
+		{
+			Value: linux.NETLINK_GENERIC,
+			Name:  "NETLINK_GENERIC",
+		},
+		{
+			Value: linux.NETLINK_SCSITRANSPORT,
+			Name:  "NETLINK_SCSITRANSPORT",
+		},
+		{
+			Value: linux.NETLINK_ECRYPTFS,
+			Name:  "NETLINK_ECRYPTFS",
+		},
+		{
+			Value: linux.NETLINK_RDMA,
+			Name:  "NETLINK_RDMA",
+		},
+		{
+			Value: linux.NETLINK_CRYPTO,
+			Name:  "NETLINK_CRYPTO",
+		},
+	},
+}
+
+var controlMessageType = map[int32]string{
+	linux.SCM_RIGHTS:      "SCM_RIGHTS",
+	linux.SCM_CREDENTIALS: "SCM_CREDENTIALS",
+}
+
+func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) string {
+	if length > maxBytes {
+		return fmt.Sprintf("%#x (error decoding control: invalid length (%d))", addr, length)
+	}
+
+	buf := make([]byte, length)
+	if _, err := t.CopyIn(addr, &buf); err != nil {
+		return fmt.Sprintf("%#x (error decoding control: %v)", addr, err)
+	}
+
+	var strs []string
+
+	for i := 0; i < len(buf); {
+		if i+linux.SizeOfControlMessageHeader > len(buf) {
+			strs = append(strs, "{invalid control message (too short)}")
+			break
+		}
+
+		var h linux.ControlMessageHeader
+		binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageHeader], usermem.ByteOrder, &h)
+		i += linux.SizeOfControlMessageHeader
+
+		var skipData bool
+		level := "SOL_SOCKET"
+		if h.Level != linux.SOL_SOCKET {
+			skipData = true
+			level = fmt.Sprint(h.Level)
+		}
+
+		typ, ok := controlMessageType[h.Type]
+		if !ok {
+			skipData = true
+			typ = fmt.Sprint(h.Type)
+		}
+
+		if h.Length > uint64(len(buf)-i) {
+			strs = append(strs, fmt.Sprintf(
+				"{level=%s, type=%s, length=%d, content extends beyond buffer}",
+				level,
+				typ,
+				h.Length,
+			))
+			break
+		}
+
+		width := t.Arch().Width()
+		length := int(h.Length) - linux.SizeOfControlMessageHeader
+
+		if skipData {
+			strs = append(strs, fmt.Sprintf("{level=%s, type=%s, length=%d}", level, typ, h.Length))
+			i += control.AlignUp(i+length, width)
+			continue
+		}
+
+		switch h.Type {
+		case linux.SCM_RIGHTS:
+			rightsSize := control.AlignDown(length, linux.SizeOfControlMessageRight)
+
+			numRights := rightsSize / linux.SizeOfControlMessageRight
+			fds := make(linux.ControlMessageRights, numRights)
+			binary.Unmarshal(buf[i:i+rightsSize], usermem.ByteOrder, &fds)
+
+			rights := make([]string, 0, len(fds))
+			for _, fd := range fds {
+				rights = append(rights, fmt.Sprint(fd))
+			}
+
+			strs = append(strs, fmt.Sprintf(
+				"{level=%s, type=%s, length=%d, content: %s}",
+				level,
+				typ,
+				h.Length,
+				strings.Join(rights, ","),
+			))
+
+			i += control.AlignUp(length, width)
+
+		case linux.SCM_CREDENTIALS:
+			if length < linux.SizeOfControlMessageCredentials {
+				strs = append(strs, fmt.Sprintf(
+					"{level=%s, type=%s, length=%d, content too short}",
+					level,
+					typ,
+					h.Length,
+				))
+				i += control.AlignUp(length, width)
+				break
+			}
+
+			var creds linux.ControlMessageCredentials
+			binary.Unmarshal(buf[i:i+linux.SizeOfControlMessageCredentials], usermem.ByteOrder, &creds)
+
+			strs = append(strs, fmt.Sprintf(
+				"{level=%s, type=%s, length=%d, pid: %d, uid: %d, gid: %d}",
+				level,
+				typ,
+				h.Length,
+				creds.PID,
+				creds.UID,
+				creds.GID,
+			))
+
+			i += control.AlignUp(length, width)
+
+		default:
+			panic("unreachable")
+		}
+	}
+
+	return fmt.Sprintf("%#x %s", addr, strings.Join(strs, ", "))
+}
+
+func msghdr(t *kernel.Task, addr usermem.Addr, printContent bool, maxBytes uint64) string {
+	var msg slinux.MessageHeader64
+	if err := slinux.CopyInMessageHeader64(t, addr, &msg); err != nil {
+		return fmt.Sprintf("%#x (error decoding msghdr: %v)", addr, err)
+	}
+	s := fmt.Sprintf(
+		"%#x {name=%#x, namelen=%d, iovecs=%s",
+		addr,
+		msg.Name,
+		msg.NameLen,
+		iovecs(t, usermem.Addr(msg.Iov), int(msg.IovLen), printContent, maxBytes),
+	)
+	if printContent {
+		s = fmt.Sprintf("%s, control={%s}", s, cmsghdr(t, usermem.Addr(msg.Control), msg.ControlLen, maxBytes))
+	} else {
+		s = fmt.Sprintf("%s, control=%#x, control_len=%d", s, msg.Control, msg.ControlLen)
+	}
+	return fmt.Sprintf("%s, flags=%d}", s, msg.Flags)
+}
+
+func sockAddr(t *kernel.Task, addr usermem.Addr, length uint32) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	b, err := slinux.CaptureAddress(t, addr, length)
+	if err != nil {
+		return fmt.Sprintf("%#x {error reading address: %v}", addr, err)
+	}
+
+	// Extract address family.
+	if len(b) < 2 {
+		return fmt.Sprintf("%#x {address too short: %d bytes}", addr, len(b))
+	}
+	family := usermem.ByteOrder.Uint16(b)
+
+	familyStr := SocketFamily.Parse(uint64(family))
+
+	switch family {
+	case linux.AF_INET, linux.AF_INET6, linux.AF_UNIX:
+		fa, err := epsocket.GetAddress(int(family), b)
+		if err != nil {
+			return fmt.Sprintf("%#x {Family: %s, error extracting address: %v}", addr, familyStr, err)
+		}
+
+		if family == linux.AF_UNIX {
+			return fmt.Sprintf("%#x {Family: %s, Addr: %q}", addr, familyStr, string(fa.Addr))
+		}
+
+		return fmt.Sprintf("%#x {Family: %s, Addr: %v, Port: %d}", addr, familyStr, fa.Addr, fa.Port)
+	case linux.AF_NETLINK:
+		sa, err := netlink.ExtractSockAddr(b)
+		if err != nil {
+			return fmt.Sprintf("%#x {Family: %s, error extracting address: %v}", addr, familyStr, err)
+		}
+		return fmt.Sprintf("%#x {Family: %s, PortID: %d, Groups: %d}", addr, familyStr, sa.PortID, sa.Groups)
+	default:
+		return fmt.Sprintf("%#x {Family: %s, family addr format unknown}", addr, familyStr)
+	}
+}
+
+func postSockAddr(t *kernel.Task, addr usermem.Addr, lengthPtr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	if lengthPtr == 0 {
+		return fmt.Sprintf("%#x {length null}", addr)
+	}
+
+	l, err := copySockLen(t, lengthPtr)
+	if err != nil {
+		return fmt.Sprintf("%#x {error reading length: %v}", addr, err)
+	}
+
+	return sockAddr(t, addr, l)
+}
+
+func copySockLen(t *kernel.Task, addr usermem.Addr) (uint32, error) {
+	// socklen_t is 32-bits.
+	var l uint32
+	_, err := t.CopyIn(addr, &l)
+	return l, err
+}
+
+func sockLenPointer(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+	l, err := copySockLen(t, addr)
+	if err != nil {
+		return fmt.Sprintf("%#x {error reading length: %v}", addr, err)
+	}
+	return fmt.Sprintf("%#x {length=%v}", addr, l)
+}
+
+func sockType(stype int32) string {
+	s := SocketType.Parse(uint64(stype & linux.SOCK_TYPE_MASK))
+	if flags := SocketFlagSet.Parse(uint64(stype &^ linux.SOCK_TYPE_MASK)); flags != "" {
+		s += "|" + flags
+	}
+	return s
+}
+
+func sockProtocol(family, protocol int32) string {
+	protocols, ok := SocketProtocol[family]
+	if !ok {
+		return fmt.Sprintf("%#x", protocol)
+	}
+	return protocols.Parse(uint64(protocol))
+}
+
+func sockFlags(flags int32) string {
+	if flags == 0 {
+		return "0"
+	}
+	return SocketFlagSet.Parse(uint64(flags))
+}
diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go
new file mode 100644
index 000000000..4cd16d2f8
--- /dev/null
+++ b/pkg/sentry/strace/strace.go
@@ -0,0 +1,666 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package strace implements the logic to print out the input and the return value
+// of each traced syscall.
+package strace
+
+import (
+	"encoding/binary"
+	"fmt"
+	"strconv"
+	"strings"
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+	"gvisor.googlesource.com/gvisor/pkg/eventchannel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	pb "gvisor.googlesource.com/gvisor/pkg/sentry/strace/strace_go_proto"
+	slinux "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// DefaultLogMaximumSize is the default LogMaximumSize.
+const DefaultLogMaximumSize = 1024
+
+// LogMaximumSize determines the maximum display size for data blobs (read,
+// write, etc.).
+var LogMaximumSize uint = DefaultLogMaximumSize
+
+// EventMaximumSize determines the maximum size for data blobs (read, write,
+// etc.) sent over the event channel. Default is 0 because most clients cannot
+// do anything useful with binary text dump of byte array arguments.
+var EventMaximumSize uint
+
+func iovecs(t *kernel.Task, addr usermem.Addr, iovcnt int, printContent bool, maxBytes uint64) string {
+	if iovcnt < 0 || iovcnt > linux.UIO_MAXIOV {
+		return fmt.Sprintf("%#x (error decoding iovecs: invalid iovcnt)", addr)
+	}
+	ars, err := t.CopyInIovecs(addr, iovcnt)
+	if err != nil {
+		return fmt.Sprintf("%#x (error decoding iovecs: %v)", addr, err)
+	}
+
+	var totalBytes uint64
+	var truncated bool
+	iovs := make([]string, iovcnt)
+	for i := 0; !ars.IsEmpty(); i, ars = i+1, ars.Tail() {
+		ar := ars.Head()
+		if ar.Length() == 0 || !printContent {
+			iovs[i] = fmt.Sprintf("{base=%#x, len=%d}", ar.Start, ar.Length())
+			continue
+		}
+
+		size := uint64(ar.Length())
+		if truncated || totalBytes+size > maxBytes {
+			truncated = true
+			size = maxBytes - totalBytes
+		} else {
+			totalBytes += uint64(ar.Length())
+		}
+
+		b := make([]byte, size)
+		amt, err := t.CopyIn(ar.Start, b)
+		if err != nil {
+			iovs[i] = fmt.Sprintf("{base=%#x, len=%d, %q..., error decoding string: %v}", ar.Start, ar.Length(), b[:amt], err)
+			continue
+		}
+
+		dot := ""
+		if truncated {
+			// Indicate truncation.
+			dot = "..."
+		}
+		iovs[i] = fmt.Sprintf("{base=%#x, len=%d, %q%s}", ar.Start, ar.Length(), b[:amt], dot)
+	}
+
+	return fmt.Sprintf("%#x %s", addr, strings.Join(iovs, ", "))
+}
+
+func dump(t *kernel.Task, addr usermem.Addr, size uint, maximumBlobSize uint) string {
+	origSize := size
+	if size > maximumBlobSize {
+		size = maximumBlobSize
+	}
+	if size == 0 {
+		return ""
+	}
+
+	b := make([]byte, size)
+	amt, err := t.CopyIn(addr, b)
+	if err != nil {
+		return fmt.Sprintf("%#x (error decoding string: %s)", addr, err)
+	}
+
+	dot := ""
+	if uint(amt) < origSize {
+		// ... if we truncated the dump.
+		dot = "..."
+	}
+
+	return fmt.Sprintf("%#x %q%s", addr, b[:amt], dot)
+}
+
+func path(t *kernel.Task, addr usermem.Addr) string {
+	path, err := t.CopyInString(addr, syscall.PathMax)
+	if err != nil {
+		return fmt.Sprintf("%#x (error decoding path: %s)", addr, err)
+	}
+	return fmt.Sprintf("%#x %s", addr, path)
+}
+
+func fdpair(t *kernel.Task, addr usermem.Addr) string {
+	var fds [2]int32
+	_, err := t.CopyIn(addr, &fds)
+	if err != nil {
+		return fmt.Sprintf("%#x (error decoding fds: %s)", addr, err)
+	}
+
+	return fmt.Sprintf("%#x [%d %d]", addr, fds[0], fds[1])
+}
+
+func uname(t *kernel.Task, addr usermem.Addr) string {
+	var u linux.UtsName
+	if _, err := t.CopyIn(addr, &u); err != nil {
+		return fmt.Sprintf("%#x (error decoding utsname: %s)", addr, err)
+	}
+
+	return fmt.Sprintf("%#x %s", addr, u)
+}
+
+func utimensTimespec(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var tim linux.Timespec
+	if _, err := t.CopyIn(addr, &tim); err != nil {
+		return fmt.Sprintf("%#x (error decoding timespec: %s)", addr, err)
+	}
+
+	var ns string
+	switch tim.Nsec {
+	case linux.UTIME_NOW:
+		ns = "UTIME_NOW"
+	case linux.UTIME_OMIT:
+		ns = "UTIME_OMIT"
+	default:
+		ns = fmt.Sprintf("%v", tim.Nsec)
+	}
+	return fmt.Sprintf("%#x {sec=%v nsec=%s}", addr, tim.Sec, ns)
+}
+
+func timespec(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var tim linux.Timespec
+	if _, err := t.CopyIn(addr, &tim); err != nil {
+		return fmt.Sprintf("%#x (error decoding timespec: %s)", addr, err)
+	}
+	return fmt.Sprintf("%#x {sec=%v nsec=%v}", addr, tim.Sec, tim.Nsec)
+}
+
+func timeval(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var tim linux.Timeval
+	if _, err := t.CopyIn(addr, &tim); err != nil {
+		return fmt.Sprintf("%#x (error decoding timeval: %s)", addr, err)
+	}
+
+	return fmt.Sprintf("%#x {sec=%v usec=%v}", addr, tim.Sec, tim.Usec)
+}
+
+func utimbuf(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var utim syscall.Utimbuf
+	if _, err := t.CopyIn(addr, &utim); err != nil {
+		return fmt.Sprintf("%#x (error decoding utimbuf: %s)", addr, err)
+	}
+
+	return fmt.Sprintf("%#x {actime=%v, modtime=%v}", addr, utim.Actime, utim.Modtime)
+}
+
+func stat(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var stat linux.Stat
+	if _, err := t.CopyIn(addr, &stat); err != nil {
+		return fmt.Sprintf("%#x (error decoding stat: %s)", addr, err)
+	}
+	return fmt.Sprintf("%#x {dev=%d, ino=%d, mode=%s, nlink=%d, uid=%d, gid=%d, rdev=%d, size=%d, blksize=%d, blocks=%d, atime=%s, mtime=%s, ctime=%s}", addr, stat.Dev, stat.Ino, linux.FileMode(stat.Mode), stat.Nlink, stat.UID, stat.GID, stat.Rdev, stat.Size, stat.Blksize, stat.Blocks, time.Unix(stat.ATime.Sec, stat.ATime.Nsec), time.Unix(stat.MTime.Sec, stat.MTime.Nsec), time.Unix(stat.CTime.Sec, stat.CTime.Nsec))
+}
+
+func itimerval(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	interval := timeval(t, addr)
+	value := timeval(t, addr+usermem.Addr(binary.Size(linux.Timeval{})))
+	return fmt.Sprintf("%#x {interval=%s, value=%s}", addr, interval, value)
+}
+
+func stringVector(t *kernel.Task, addr usermem.Addr) string {
+	vec, err := t.CopyInVector(addr, slinux.ExecMaxElemSize, slinux.ExecMaxTotalSize)
+	if err != nil {
+		return fmt.Sprintf("%#x {error copying vector: %v}", addr, err)
+	}
+	s := fmt.Sprintf("%#x [", addr)
+	for i, v := range vec {
+		if i != 0 {
+			s += ", "
+		}
+		s += fmt.Sprintf("%q", v)
+	}
+	s += "]"
+	return s
+}
+
+func rusage(t *kernel.Task, addr usermem.Addr) string {
+	if addr == 0 {
+		return "null"
+	}
+
+	var ru linux.Rusage
+	if _, err := t.CopyIn(addr, &ru); err != nil {
+		return fmt.Sprintf("%#x (error decoding rusage: %s)", addr, err)
+	}
+	return fmt.Sprintf("%#x %+v", addr, ru)
+}
+
+// pre fills in the pre-execution arguments for a system call. If an argument
+// cannot be interpreted before the system call is executed, then a hex value
+// will be used. Note that a full output slice will always be provided, that is
+// len(return) == len(args).
+func (i *SyscallInfo) pre(t *kernel.Task, args arch.SyscallArguments, maximumBlobSize uint) []string {
+	var output []string
+
+	for arg := range args {
+		if arg >= len(i.format) {
+			break
+		}
+		switch i.format[arg] {
+		case WriteBuffer:
+			output = append(output, dump(t, args[arg].Pointer(), args[arg+1].SizeT(), maximumBlobSize))
+		case WriteIOVec:
+			output = append(output, iovecs(t, args[arg].Pointer(), int(args[arg+1].Int()), true /* content */, uint64(maximumBlobSize)))
+		case IOVec:
+			output = append(output, iovecs(t, args[arg].Pointer(), int(args[arg+1].Int()), false /* content */, uint64(maximumBlobSize)))
+		case SendMsgHdr:
+			output = append(output, msghdr(t, args[arg].Pointer(), true /* content */, uint64(maximumBlobSize)))
+		case RecvMsgHdr:
+			output = append(output, msghdr(t, args[arg].Pointer(), false /* content */, uint64(maximumBlobSize)))
+		case Path:
+			output = append(output, path(t, args[arg].Pointer()))
+		case ExecveStringVector:
+			output = append(output, stringVector(t, args[arg].Pointer()))
+		case SockAddr:
+			output = append(output, sockAddr(t, args[arg].Pointer(), uint32(args[arg+1].Uint64())))
+		case SockLen:
+			output = append(output, sockLenPointer(t, args[arg].Pointer()))
+		case SockFamily:
+			output = append(output, SocketFamily.Parse(uint64(args[arg].Int())))
+		case SockType:
+			output = append(output, sockType(args[arg].Int()))
+		case SockProtocol:
+			output = append(output, sockProtocol(args[arg-2].Int(), args[arg].Int()))
+		case SockFlags:
+			output = append(output, sockFlags(args[arg].Int()))
+		case Timespec:
+			output = append(output, timespec(t, args[arg].Pointer()))
+		case UTimeTimespec:
+			output = append(output, utimensTimespec(t, args[arg].Pointer()))
+		case ItimerVal:
+			output = append(output, itimerval(t, args[arg].Pointer()))
+		case Timeval:
+			output = append(output, timeval(t, args[arg].Pointer()))
+		case Utimbuf:
+			output = append(output, utimbuf(t, args[arg].Pointer()))
+		case CloneFlags:
+			output = append(output, CloneFlagSet.Parse(uint64(args[arg].Uint())))
+		case OpenFlags:
+			output = append(output, open(uint64(args[arg].Uint())))
+		case Mode:
+			output = append(output, linux.FileMode(args[arg].ModeT()).String())
+		case FutexOp:
+			output = append(output, futex(uint64(args[arg].Uint())))
+		case PtraceRequest:
+			output = append(output, PtraceRequestSet.Parse(args[arg].Uint64()))
+		case Oct:
+			output = append(output, "0o"+strconv.FormatUint(args[arg].Uint64(), 8))
+		case Hex:
+			fallthrough
+		default:
+			output = append(output, "0x"+strconv.FormatUint(args[arg].Uint64(), 16))
+		}
+	}
+
+	return output
+}
+
+// post fills in the post-execution arguments for a system call. This modifies
+// the given output slice in place with arguments that may only be interpreted
+// after the system call has been executed.
+func (i *SyscallInfo) post(t *kernel.Task, args arch.SyscallArguments, rval uintptr, output []string, maximumBlobSize uint) {
+	for arg := range output {
+		if arg >= len(i.format) {
+			break
+		}
+		switch i.format[arg] {
+		case ReadBuffer:
+			output[arg] = dump(t, args[arg].Pointer(), uint(rval), maximumBlobSize)
+		case ReadIOVec:
+			printLength := uint64(rval)
+			if printLength > uint64(maximumBlobSize) {
+				printLength = uint64(maximumBlobSize)
+			}
+			output[arg] = iovecs(t, args[arg].Pointer(), int(args[arg+1].Int()), true /* content */, printLength)
+		case WriteIOVec, IOVec, WriteBuffer:
+			// We already have a big blast from write.
+			output[arg] = "..."
+		case SendMsgHdr:
+			output[arg] = msghdr(t, args[arg].Pointer(), false /* content */, uint64(maximumBlobSize))
+		case RecvMsgHdr:
+			output[arg] = msghdr(t, args[arg].Pointer(), true /* content */, uint64(maximumBlobSize))
+		case PipeFDs:
+			output[arg] = fdpair(t, args[arg].Pointer())
+		case Uname:
+			output[arg] = uname(t, args[arg].Pointer())
+		case Stat:
+			output[arg] = stat(t, args[arg].Pointer())
+		case PostSockAddr:
+			output[arg] = postSockAddr(t, args[arg].Pointer(), args[arg+1].Pointer())
+		case SockLen:
+			output[arg] = sockLenPointer(t, args[arg].Pointer())
+		case PostTimespec:
+			output[arg] = timespec(t, args[arg].Pointer())
+		case PostItimerVal:
+			output[arg] = itimerval(t, args[arg].Pointer())
+		case Timeval:
+			output[arg] = timeval(t, args[arg].Pointer())
+		case Rusage:
+			output[arg] = rusage(t, args[arg].Pointer())
+		}
+	}
+}
+
+// printEntry prints the given system call entry.
+func (i *SyscallInfo) printEnter(t *kernel.Task, args arch.SyscallArguments) []string {
+	output := i.pre(t, args, LogMaximumSize)
+
+	switch len(output) {
+	case 0:
+		t.Infof("%s E %s()", t.Name(), i.name)
+	case 1:
+		t.Infof("%s E %s(%s)", t.Name(), i.name,
+			output[0])
+	case 2:
+		t.Infof("%s E %s(%s, %s)", t.Name(), i.name,
+			output[0], output[1])
+	case 3:
+		t.Infof("%s E %s(%s, %s, %s)", t.Name(), i.name,
+			output[0], output[1], output[2])
+	case 4:
+		t.Infof("%s E %s(%s, %s, %s, %s)", t.Name(), i.name,
+			output[0], output[1], output[2], output[3])
+	case 5:
+		t.Infof("%s E %s(%s, %s, %s, %s, %s)", t.Name(), i.name,
+			output[0], output[1], output[2], output[3], output[4])
+	case 6:
+		t.Infof("%s E %s(%s, %s, %s, %s, %s, %s)", t.Name(), i.name,
+			output[0], output[1], output[2], output[3], output[4], output[5])
+	}
+
+	return output
+}
+
+// printExit prints the given system call exit.
+func (i *SyscallInfo) printExit(t *kernel.Task, elapsed time.Duration, output []string, args arch.SyscallArguments, retval uintptr, err error, errno int) {
+	var rval string
+	if err == nil {
+		// Fill in the output after successful execution.
+		i.post(t, args, retval, output, LogMaximumSize)
+		rval = fmt.Sprintf("%#x (%v)", retval, elapsed)
+	} else {
+		rval = fmt.Sprintf("%#x errno=%d (%s) (%v)", retval, errno, err, elapsed)
+	}
+
+	switch len(output) {
+	case 0:
+		t.Infof("%s X %s() = %s", t.Name(), i.name,
+			rval)
+	case 1:
+		t.Infof("%s X %s(%s) = %s", t.Name(), i.name,
+			output[0], rval)
+	case 2:
+		t.Infof("%s X %s(%s, %s) = %s", t.Name(), i.name,
+			output[0], output[1], rval)
+	case 3:
+		t.Infof("%s X %s(%s, %s, %s) = %s", t.Name(), i.name,
+			output[0], output[1], output[2], rval)
+	case 4:
+		t.Infof("%s X %s(%s, %s, %s, %s) = %s", t.Name(), i.name,
+			output[0], output[1], output[2], output[3], rval)
+	case 5:
+		t.Infof("%s X %s(%s, %s, %s, %s, %s) = %s", t.Name(), i.name,
+			output[0], output[1], output[2], output[3], output[4], rval)
+	case 6:
+		t.Infof("%s X %s(%s, %s, %s, %s, %s, %s) = %s", t.Name(), i.name,
+			output[0], output[1], output[2], output[3], output[4], output[5], rval)
+	}
+}
+
+// sendEnter sends the syscall enter to event log.
+func (i *SyscallInfo) sendEnter(t *kernel.Task, args arch.SyscallArguments) []string {
+	output := i.pre(t, args, EventMaximumSize)
+
+	event := pb.Strace{
+		Process:  t.Name(),
+		Function: i.name,
+		Info: &pb.Strace_Enter{
+			Enter: &pb.StraceEnter{},
+		},
+	}
+	for _, arg := range output {
+		event.Args = append(event.Args, arg)
+	}
+	eventchannel.Emit(&event)
+
+	return output
+}
+
+// sendExit sends the syscall exit to event log.
+func (i *SyscallInfo) sendExit(t *kernel.Task, elapsed time.Duration, output []string, args arch.SyscallArguments, rval uintptr, err error, errno int) {
+	if err == nil {
+		// Fill in the output after successful execution.
+		i.post(t, args, rval, output, EventMaximumSize)
+	}
+
+	exit := &pb.StraceExit{
+		Return:    fmt.Sprintf("%#x", rval),
+		ElapsedNs: elapsed.Nanoseconds(),
+	}
+	if err != nil {
+		exit.Error = err.Error()
+		exit.ErrNo = int64(errno)
+	}
+	event := pb.Strace{
+		Process:  t.Name(),
+		Function: i.name,
+		Info:     &pb.Strace_Exit{Exit: exit},
+	}
+	for _, arg := range output {
+		event.Args = append(event.Args, arg)
+	}
+	eventchannel.Emit(&event)
+}
+
+type syscallContext struct {
+	info        SyscallInfo
+	args        arch.SyscallArguments
+	start       time.Time
+	logOutput   []string
+	eventOutput []string
+	flags       uint32
+}
+
+// SyscallEnter implements kernel.Stracer.SyscallEnter. It logs the syscall
+// entry trace.
+func (s SyscallMap) SyscallEnter(t *kernel.Task, sysno uintptr, args arch.SyscallArguments, flags uint32) interface{} {
+	info, ok := s[sysno]
+	if !ok {
+		info = SyscallInfo{
+			name:   fmt.Sprintf("sys_%d", sysno),
+			format: defaultFormat,
+		}
+	}
+
+	var output, eventOutput []string
+	if bits.IsOn32(flags, kernel.StraceEnableLog) {
+		output = info.printEnter(t, args)
+	}
+	if bits.IsOn32(flags, kernel.StraceEnableEvent) {
+		eventOutput = info.sendEnter(t, args)
+	}
+
+	return &syscallContext{
+		info:        info,
+		args:        args,
+		start:       time.Now(),
+		logOutput:   output,
+		eventOutput: eventOutput,
+		flags:       flags,
+	}
+}
+
+// SyscallExit implements kernel.Stracer.SyscallExit. It logs the syscall
+// exit trace.
+func (s SyscallMap) SyscallExit(context interface{}, t *kernel.Task, sysno, rval uintptr, err error) {
+	errno := t.ExtractErrno(err, int(sysno))
+	c := context.(*syscallContext)
+
+	elapsed := time.Since(c.start)
+	if bits.IsOn32(c.flags, kernel.StraceEnableLog) {
+		c.info.printExit(t, elapsed, c.logOutput, c.args, rval, err, errno)
+	}
+	if bits.IsOn32(c.flags, kernel.StraceEnableEvent) {
+		c.info.sendExit(t, elapsed, c.eventOutput, c.args, rval, err, errno)
+	}
+}
+
+// ConvertToSysnoMap converts the names to a map keyed on the syscall number and value set to true.
+// The map is in a convenient format to call SyscallFlagsTable.Enable().
+func (s SyscallMap) ConvertToSysnoMap(syscalls []string) (map[uintptr]bool, error) {
+	if syscalls == nil {
+		// Sentinel: no list.
+		return nil, nil
+	}
+
+	l := make(map[uintptr]bool)
+	for _, sc := range syscalls {
+		// Try to match this system call.
+		sysno, ok := s.ConvertToSysno(sc)
+		if !ok {
+			return nil, fmt.Errorf("syscall %q not found", sc)
+		}
+		l[sysno] = true
+	}
+
+	// Success.
+	return l, nil
+}
+
+// ConvertToSysno converts the name to system call number. Returns false
+// if syscall with same name is not found.
+func (s SyscallMap) ConvertToSysno(syscall string) (uintptr, bool) {
+	for sysno, info := range s {
+		if info.name != "" && info.name == syscall {
+			return sysno, true
+		}
+	}
+	return 0, false
+}
+
+// Name returns the syscall name.
+func (s SyscallMap) Name(sysno uintptr) string {
+	if info, ok := s[sysno]; ok {
+		return info.name
+	}
+	return fmt.Sprintf("sys_%d", sysno)
+}
+
+// Initialize prepares all syscall tables for use by this package.
+//
+// N.B. This is not in an init function because we can't be sure all syscall
+// tables are registered with the kernel when init runs.
+//
+// TODO: remove kernel package dependencies from this package and
+// have the kernel package self-initialize all syscall tables.
+func Initialize() {
+	for _, table := range kernel.SyscallTables() {
+		// Is this known?
+		sys, ok := Lookup(table.OS, table.Arch)
+		if !ok {
+			continue
+		}
+
+		table.Stracer = sys
+	}
+}
+
+// SinkType defines where to send straces to.
+type SinkType uint32
+
+const (
+	// SinkTypeLog sends straces to text log
+	SinkTypeLog SinkType = 1 << iota
+
+	// SinkTypeEvent sends strace to event log
+	SinkTypeEvent
+)
+
+func convertToSyscallFlag(sinks SinkType) uint32 {
+	ret := uint32(0)
+	if bits.IsOn32(uint32(sinks), uint32(SinkTypeLog)) {
+		ret |= kernel.StraceEnableLog
+	}
+	if bits.IsOn32(uint32(sinks), uint32(SinkTypeEvent)) {
+		ret |= kernel.StraceEnableEvent
+	}
+	return ret
+}
+
+// Enable enables the syscalls in whitelist in all syscall tables.
+//
+// Preconditions: Initialize has been called.
+func Enable(whitelist []string, sinks SinkType) error {
+	flags := convertToSyscallFlag(sinks)
+	for _, table := range kernel.SyscallTables() {
+		// Is this known?
+		sys, ok := Lookup(table.OS, table.Arch)
+		if !ok {
+			continue
+		}
+
+		// Convert to a set of system calls numbers.
+		wl, err := sys.ConvertToSysnoMap(whitelist)
+		if err != nil {
+			return err
+		}
+
+		table.FeatureEnable.Enable(flags, wl, true)
+	}
+
+	// Done.
+	return nil
+}
+
+// Disable will disable Strace for all system calls and missing syscalls.
+//
+// Preconditions: Initialize has been called.
+func Disable(sinks SinkType) {
+	flags := convertToSyscallFlag(sinks)
+	for _, table := range kernel.SyscallTables() {
+		// Strace will be disabled for all syscalls including missing.
+		table.FeatureEnable.Enable(flags, nil, false)
+	}
+}
+
+// EnableAll enables all syscalls in all syscall tables.
+//
+// Preconditions: Initialize has been called.
+func EnableAll(sinks SinkType) {
+	flags := convertToSyscallFlag(sinks)
+	for _, table := range kernel.SyscallTables() {
+		// Is this known?
+		if _, ok := Lookup(table.OS, table.Arch); !ok {
+			continue
+		}
+
+		table.FeatureEnable.EnableAll(flags)
+	}
+}
diff --git a/pkg/sentry/strace/strace.proto b/pkg/sentry/strace/strace.proto
new file mode 100644
index 000000000..914e8c7b0
--- /dev/null
+++ b/pkg/sentry/strace/strace.proto
@@ -0,0 +1,50 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package gvisor;
+
+message Strace {
+  // Process name that made the syscall.
+  string process = 1;
+
+  // Syscall function name.
+  string function = 2;
+
+  // List of syscall arguments formatted as strings.
+  repeated string args = 3;
+
+  oneof info {
+    StraceEnter enter = 4;
+    StraceExit exit = 5;
+  }
+}
+
+message StraceEnter {
+}
+
+message StraceExit {
+  // Return value formatted as string.
+  string return = 1;
+
+  // Formatted error string in case syscall failed.
+  string error = 2;
+
+  // Value of errno upon syscall exit.
+  int64 err_no = 3;  // errno is a macro and gets expanded :-(
+
+  // Time elapsed between syscall enter and exit.
+  int64 elapsed_ns = 4;
+}
diff --git a/pkg/sentry/strace/syscalls.go b/pkg/sentry/strace/syscalls.go
new file mode 100644
index 000000000..d0e661706
--- /dev/null
+++ b/pkg/sentry/strace/syscalls.go
@@ -0,0 +1,217 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package strace
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// FormatSpecifier values describe how an individual syscall argument should be
+// formatted.
+type FormatSpecifier int
+
+// Valid FormatSpecifiers.
+//
+// Unless otherwise specified, values are formatted before syscall execution
+// and not updated after syscall execution (the same value is output).
+const (
+	// Hex is just a hexadecimal number.
+	Hex FormatSpecifier = iota
+
+	// Oct is just an octal number.
+	Oct
+
+	// ReadBuffer is a buffer for a read-style call. The syscall return
+	// value is used for the length.
+	//
+	// Formatted after syscall execution.
+	ReadBuffer
+
+	// WriteBuffer is a buffer for a write-style call. The following arg is
+	// used for the length.
+	//
+	// Contents omitted after syscall execution.
+	WriteBuffer
+
+	// ReadIOVec is a pointer to a struct iovec for a writev-style call.
+	// The following arg is used for the length. The return value is used
+	// for the total length.
+	//
+	// Complete contents only formatted after syscall execution.
+	ReadIOVec
+
+	// WriteIOVec is a pointer to a struct iovec for a writev-style call.
+	// The following arg is used for the length.
+	//
+	// Complete contents only formatted before syscall execution, omitted
+	// after.
+	WriteIOVec
+
+	// IOVec is a generic pointer to a struct iovec. Contents are not dumped.
+	IOVec
+
+	// SendMsgHdr is a pointer to a struct msghdr for a sendmsg-style call.
+	// Contents formatted only before syscall execution, omitted after.
+	SendMsgHdr
+
+	// RecvMsgHdr is a pointer to a struct msghdr for a recvmsg-style call.
+	// Contents formatted only after syscall execution.
+	RecvMsgHdr
+
+	// Path is a pointer to a char* path.
+	Path
+
+	// ExecveStringVector is a NULL-terminated array of strings. Enforces
+	// the maximum execve array length.
+	ExecveStringVector
+
+	// PipeFDs is an array of two FDs, formatted after syscall execution.
+	PipeFDs
+
+	// Uname is a pointer to a struct uname, formatted after syscall exection.
+	Uname
+
+	// Stat is a pointer to a struct stat, formatted after syscall execution.
+	Stat
+
+	// SockAddr is a pointer to a struct sockaddr. The following arg is
+	// used for length.
+	SockAddr
+
+	// PostSockAddr is a pointer to a struct sockaddr, formatted after
+	// syscall execution. The following arg is a pointer to the socklen_t
+	// length.
+	PostSockAddr
+
+	// SockLen is a pointer to a socklen_t, formatted before and after
+	// syscall execution.
+	SockLen
+
+	// SockFamily is a socket protocol family value.
+	SockFamily
+
+	// SockType is a socket type and flags value.
+	SockType
+
+	// SockProtocol is a socket protocol value. Argument n-2 is the socket
+	// protocol family.
+	SockProtocol
+
+	// SockFlags are socket flags.
+	SockFlags
+
+	// Timespec is a pointer to a struct timespec.
+	Timespec
+
+	// PostTimespec is a pointer to a struct timespec, formatted after
+	// syscall execution.
+	PostTimespec
+
+	// UTimeTimespec is a pointer to a struct timespec. Formatting includes
+	// UTIME_NOW and UTIME_OMIT.
+	UTimeTimespec
+
+	// ItimerVal is a pointer to a struct itimerval.
+	ItimerVal
+
+	// ItimerVal is a pointer to a struct itimerval, formatted after
+	// syscall execution.
+	PostItimerVal
+
+	// Timeval is a pointer to a struct timeval, formatted before and after
+	// syscall execution.
+	Timeval
+
+	// Utimbuf is a pointer to a struct utimbuf.
+	Utimbuf
+
+	// CloneFlags are clone(2) flags.
+	CloneFlags
+
+	// OpenFlags are open(2) flags.
+	OpenFlags
+
+	// Mode is a mode_t.
+	Mode
+
+	// FutexOp is the futex(2) operation.
+	FutexOp
+
+	// PtraceRequest is the ptrace(2) request.
+	PtraceRequest
+
+	// Rusage is a struct rusage, formatted after syscall execution.
+	Rusage
+)
+
+// defaultFormat is the syscall argument format to use if the actual format is
+// not known. It formats all six arguments as hex.
+var defaultFormat = []FormatSpecifier{Hex, Hex, Hex, Hex, Hex, Hex}
+
+// SyscallInfo captures the name and printing format of a syscall.
+type SyscallInfo struct {
+	// name is the name of the syscall.
+	name string
+
+	// format contains the format specifiers for each argument.
+	//
+	// Syscall calls can have up to six arguments. Arguments without a
+	// corresponding entry in format will not be printed.
+	format []FormatSpecifier
+}
+
+// makeSyscallInfo returns a SyscallInfo for a syscall.
+func makeSyscallInfo(name string, f ...FormatSpecifier) SyscallInfo {
+	return SyscallInfo{name: name, format: f}
+}
+
+// SyscallMap maps syscalls into names and printing formats.
+type SyscallMap map[uintptr]SyscallInfo
+
+var _ kernel.Stracer = (SyscallMap)(nil)
+
+// syscallTable contains the syscalls for a specific OS/Arch.
+type syscallTable struct {
+	// os is the operating system this table targets.
+	os abi.OS
+
+	// arch is the architecture this table targets.
+	arch arch.Arch
+
+	// syscalls contains the syscall mappings.
+	syscalls SyscallMap
+}
+
+// syscallTables contains all syscall tables.
+var syscallTables = []syscallTable{
+	{
+		os:       abi.Linux,
+		arch:     arch.AMD64,
+		syscalls: linuxAMD64,
+	},
+}
+
+// Lookup returns the SyscallMap for the OS/Arch combination. The returned map
+// must not be changed.
+func Lookup(os abi.OS, a arch.Arch) (SyscallMap, bool) {
+	for _, s := range syscallTables {
+		if s.os == os && s.arch == a {
+			return s.syscalls, true
+		}
+	}
+	return nil, false
+}
diff --git a/pkg/sentry/syscalls/BUILD b/pkg/sentry/syscalls/BUILD
new file mode 100644
index 000000000..d667b42c8
--- /dev/null
+++ b/pkg/sentry/syscalls/BUILD
@@ -0,0 +1,43 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//proto:def.bzl", "go_proto_library")
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "syscalls",
+    srcs = [
+        "epoll.go",
+        "polling.go",
+        "syscalls.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls",
+    visibility = ["//:sandbox"],
+    deps = [
+        ":unimplemented_syscall_go_proto",
+        "//pkg/abi/linux",
+        "//pkg/eventchannel",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/epoll",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/time",
+        "//pkg/syserror",
+        "//pkg/waiter",
+    ],
+)
+
+proto_library(
+    name = "unimplemented_syscall_proto",
+    srcs = ["unimplemented_syscall.proto"],
+    visibility = ["//visibility:public"],
+    deps = ["//pkg/sentry/arch:registers_proto"],
+)
+
+go_proto_library(
+    name = "unimplemented_syscall_go_proto",
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls/unimplemented_syscall_go_proto",
+    proto = ":unimplemented_syscall_proto",
+    visibility = ["//visibility:public"],
+    deps = ["//pkg/sentry/arch:registers_go_proto"],
+)
diff --git a/pkg/sentry/syscalls/epoll.go b/pkg/sentry/syscalls/epoll.go
new file mode 100644
index 000000000..01dd6fa71
--- /dev/null
+++ b/pkg/sentry/syscalls/epoll.go
@@ -0,0 +1,174 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syscalls
+
+import (
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/epoll"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// CreateEpoll implements the epoll_create(2) linux syscall.
+func CreateEpoll(t *kernel.Task, closeOnExec bool) (kdefs.FD, error) {
+	file := epoll.NewEventPoll(t)
+	defer file.DecRef()
+
+	flags := kernel.FDFlags{
+		CloseOnExec: closeOnExec,
+	}
+	fd, err := t.FDMap().NewFDFrom(0, file, flags, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, err
+	}
+
+	return fd, nil
+}
+
+// AddEpoll implements the epoll_ctl(2) linux syscall when op is EPOLL_CTL_ADD.
+func AddEpoll(t *kernel.Task, epfd kdefs.FD, fd kdefs.FD, flags epoll.EntryFlags, mask waiter.EventMask, userData [2]int32) error {
+	// Get epoll from the file descriptor.
+	epollfile := t.FDMap().GetFile(epfd)
+	if epollfile == nil {
+		return syscall.EBADF
+	}
+	defer epollfile.DecRef()
+
+	// Get the target file id.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the epollPoll operations.
+	e, ok := epollfile.FileOperations.(*epoll.EventPoll)
+	if !ok {
+		return syscall.EBADF
+	}
+
+	// Try to add the entry.
+	return e.AddEntry(epoll.FileIdentifier{file, fd}, flags, mask, userData)
+}
+
+// UpdateEpoll implements the epoll_ctl(2) linux syscall when op is EPOLL_CTL_MOD.
+func UpdateEpoll(t *kernel.Task, epfd kdefs.FD, fd kdefs.FD, flags epoll.EntryFlags, mask waiter.EventMask, userData [2]int32) error {
+	// Get epoll from the file descriptor.
+	epollfile := t.FDMap().GetFile(epfd)
+	if epollfile == nil {
+		return syscall.EBADF
+	}
+	defer epollfile.DecRef()
+
+	// Get the target file id.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the epollPoll operations.
+	e, ok := epollfile.FileOperations.(*epoll.EventPoll)
+	if !ok {
+		return syscall.EBADF
+	}
+
+	// Try to update the entry.
+	return e.UpdateEntry(epoll.FileIdentifier{file, fd}, flags, mask, userData)
+}
+
+// RemoveEpoll implements the epoll_ctl(2) linux syscall when op is EPOLL_CTL_DEL.
+func RemoveEpoll(t *kernel.Task, epfd kdefs.FD, fd kdefs.FD) error {
+	// Get epoll from the file descriptor.
+	epollfile := t.FDMap().GetFile(epfd)
+	if epollfile == nil {
+		return syscall.EBADF
+	}
+	defer epollfile.DecRef()
+
+	// Get the target file id.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the epollPoll operations.
+	e, ok := epollfile.FileOperations.(*epoll.EventPoll)
+	if !ok {
+		return syscall.EBADF
+	}
+
+	// Try to remove the entry.
+	return e.RemoveEntry(epoll.FileIdentifier{file, fd})
+}
+
+// WaitEpoll implements the epoll_wait(2) linux syscall.
+func WaitEpoll(t *kernel.Task, fd kdefs.FD, max int, timeout int) ([]epoll.Event, error) {
+	// Get epoll from the file descriptor.
+	epollfile := t.FDMap().GetFile(fd)
+	if epollfile == nil {
+		return nil, syscall.EBADF
+	}
+	defer epollfile.DecRef()
+
+	// Extract the epollPoll operations.
+	e, ok := epollfile.FileOperations.(*epoll.EventPoll)
+	if !ok {
+		return nil, syscall.EBADF
+	}
+
+	// Try to read events and return right away if we got them or if the
+	// caller requested a non-blocking "wait".
+	r := e.ReadEvents(max)
+	if len(r) != 0 || timeout == 0 {
+		return r, nil
+	}
+
+	// We'll have to wait. Set up the timer if a timeout was specified and
+	// and register with the epoll object for readability events.
+	var haveDeadline bool
+	var deadline ktime.Time
+	if timeout > 0 {
+		timeoutDur := time.Duration(timeout) * time.Millisecond
+		deadline = t.Kernel().MonotonicClock().Now().Add(timeoutDur)
+		haveDeadline = true
+	}
+
+	w, ch := waiter.NewChannelEntry(nil)
+	e.EventRegister(&w, waiter.EventIn)
+	defer e.EventUnregister(&w)
+
+	// Try to read the events again until we succeed, timeout or get
+	// interrupted.
+	for {
+		r = e.ReadEvents(max)
+		if len(r) != 0 {
+			return r, nil
+		}
+
+		if err := t.BlockWithDeadline(ch, haveDeadline, deadline); err != nil {
+			if err == syscall.ETIMEDOUT {
+				return nil, nil
+			}
+
+			return nil, err
+		}
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD
new file mode 100644
index 000000000..bc67ebf30
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/BUILD
@@ -0,0 +1,103 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "linux_state",
+    srcs = [
+        "sys_aio.go",
+        "sys_futex.go",
+        "sys_poll.go",
+        "sys_time.go",
+    ],
+    out = "linux_state.go",
+    package = "linux",
+)
+
+go_library(
+    name = "linux",
+    srcs = [
+        "error.go",
+        "flags.go",
+        "linux64.go",
+        "linux_state.go",
+        "sigset.go",
+        "sys_aio.go",
+        "sys_capability.go",
+        "sys_epoll.go",
+        "sys_eventfd.go",
+        "sys_file.go",
+        "sys_futex.go",
+        "sys_getdents.go",
+        "sys_identity.go",
+        "sys_inotify.go",
+        "sys_lseek.go",
+        "sys_mmap.go",
+        "sys_mount.go",
+        "sys_pipe.go",
+        "sys_poll.go",
+        "sys_prctl.go",
+        "sys_random.go",
+        "sys_read.go",
+        "sys_rlimit.go",
+        "sys_rusage.go",
+        "sys_sched.go",
+        "sys_sem.go",
+        "sys_signal.go",
+        "sys_socket.go",
+        "sys_stat.go",
+        "sys_sync.go",
+        "sys_sysinfo.go",
+        "sys_syslog.go",
+        "sys_thread.go",
+        "sys_time.go",
+        "sys_timer.go",
+        "sys_timerfd.go",
+        "sys_tls.go",
+        "sys_utsname.go",
+        "sys_write.go",
+        "timespec.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls/linux",
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/abi",
+        "//pkg/abi/linux",
+        "//pkg/binary",
+        "//pkg/bpf",
+        "//pkg/eventchannel",
+        "//pkg/log",
+        "//pkg/metric",
+        "//pkg/sentry/arch",
+        "//pkg/sentry/context",
+        "//pkg/sentry/fs",
+        "//pkg/sentry/fs/anon",
+        "//pkg/sentry/fs/fsutil",
+        "//pkg/sentry/fs/lock",
+        "//pkg/sentry/fs/timerfd",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/auth",
+        "//pkg/sentry/kernel/epoll",
+        "//pkg/sentry/kernel/eventfd",
+        "//pkg/sentry/kernel/kdefs",
+        "//pkg/sentry/kernel/pipe",
+        "//pkg/sentry/kernel/sched",
+        "//pkg/sentry/kernel/semaphore",
+        "//pkg/sentry/kernel/time",
+        "//pkg/sentry/limits",
+        "//pkg/sentry/memmap",
+        "//pkg/sentry/mm",
+        "//pkg/sentry/safemem",
+        "//pkg/sentry/socket",
+        "//pkg/sentry/socket/control",
+        "//pkg/sentry/syscalls",
+        "//pkg/sentry/usage",
+        "//pkg/sentry/usermem",
+        "//pkg/state",
+        "//pkg/syserr",
+        "//pkg/syserror",
+        "//pkg/tcpip/transport/unix",
+        "//pkg/waiter",
+    ],
+)
diff --git a/pkg/sentry/syscalls/linux/error.go b/pkg/sentry/syscalls/linux/error.go
new file mode 100644
index 000000000..013b385bc
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/error.go
@@ -0,0 +1,117 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"io"
+	"sync"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/metric"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+var (
+	partialResultMetric = metric.MustCreateNewUint64Metric("/syscalls/partial_result", true /* sync */, "Whether or not a partial result has occurred for this sandbox.")
+	partialResultOnce   sync.Once
+)
+
+// handleIOError handles special error cases for partial results. For some
+// errors, we may consume the error and return only the partial read/write.
+//
+// op and f are used only for panics.
+func handleIOError(t *kernel.Task, partialResult bool, err, intr error, op string, f *fs.File) error {
+	switch err {
+	case nil:
+		// Typical successful syscall.
+		return nil
+	case io.EOF:
+		// EOF is always consumed. If this is a partial read/write
+		// (result != 0), the application will see that, otherwise
+		// they will see 0.
+		return nil
+	case syserror.ErrExceedsFileSizeLimit:
+		// Ignore partialResult because this error only applies to
+		// normal files, and for those files we cannot accumulate
+		// write results.
+		//
+		// Do not consume the error and return it as EFBIG.
+		// Simultaneously send a SIGXFSZ per setrlimit(2).
+		t.SendSignal(&arch.SignalInfo{
+			Signo: int32(syscall.SIGXFSZ),
+			Code:  arch.SignalInfoKernel,
+		})
+		return syscall.EFBIG
+	case syserror.ErrInterrupted:
+		// The syscall was interrupted. Return nil if it completed
+		// partially, otherwise return the error code that the syscall
+		// needs (to indicate to the kernel what it should do).
+		if partialResult {
+			return nil
+		}
+		return intr
+	}
+
+	if !partialResult {
+		// Typical syscall error.
+		return err
+	}
+
+	switch err {
+	case syserror.EINTR:
+		// Syscall interrupted, but completed a partial
+		// read/write.  Like ErrWouldBlock, since we have a
+		// partial read/write, we consume the error and return
+		// the partial result.
+		return nil
+	case syserror.EFAULT:
+		// EFAULT is only shown the user if nothing was
+		// read/written. If we read something (this case), they see
+		// a partial read/write. They will then presumably try again
+		// with an incremented buffer, which will EFAULT with
+		// result == 0.
+		return nil
+	case syserror.EPIPE:
+		// Writes to a pipe or socket will return EPIPE if the other
+		// side is gone. The partial write is returned. EPIPE will be
+		// returned on the next call.
+		//
+		// TODO: In some cases SIGPIPE should also be sent
+		// to the application.
+		return nil
+	case syserror.ErrWouldBlock:
+		// Syscall would block, but completed a partial read/write.
+		// This case should only be returned by IssueIO for nonblocking
+		// files. Since we have a partial read/write, we consume
+		// ErrWouldBlock, returning the partial result.
+		return nil
+	}
+
+	switch err.(type) {
+	case kernel.SyscallRestartErrno:
+		// Identical to the EINTR case.
+		return nil
+	}
+
+	// An unknown error is encountered with a partial read/write.
+	name, _ := f.Dirent.FullName(nil /* ignore chroot */)
+	log.Traceback("Invalid request partialResult %v and err (type %T) %v for %s operation on %q, %T", partialResult, err, err, op, name, f.FileOperations)
+	partialResultOnce.Do(partialResultMetric.Increment)
+	return nil
+}
diff --git a/pkg/sentry/syscalls/linux/flags.go b/pkg/sentry/syscalls/linux/flags.go
new file mode 100644
index 000000000..82bfd7c2a
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/flags.go
@@ -0,0 +1,95 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// flagsToPermissions returns a Permissions object from Linux flags.
+// This includes truncate permission if O_TRUNC is set in the mask.
+func flagsToPermissions(mask uint) (p fs.PermMask) {
+	switch mask & syscall.O_ACCMODE {
+	case syscall.O_WRONLY:
+		p.Write = true
+	case syscall.O_RDWR:
+		p.Write = true
+		p.Read = true
+	case syscall.O_RDONLY:
+		p.Read = true
+	}
+	return
+}
+
+// fdFlagsToLinux converts a kernel.FDFlags object to a Linux representation.
+func fdFlagsToLinux(flags kernel.FDFlags) (mask uint) {
+	if flags.CloseOnExec {
+		mask |= syscall.FD_CLOEXEC
+	}
+	return
+}
+
+// flagsToLinux converts a FileFlags object to a Linux representation.
+func flagsToLinux(flags fs.FileFlags) (mask uint) {
+	if flags.Direct {
+		mask |= syscall.O_DIRECT
+	}
+	if flags.NonBlocking {
+		mask |= syscall.O_NONBLOCK
+	}
+	if flags.Sync {
+		mask |= syscall.O_SYNC
+	}
+	if flags.Append {
+		mask |= syscall.O_APPEND
+	}
+	if flags.Directory {
+		mask |= syscall.O_DIRECTORY
+	}
+	switch {
+	case flags.Read && flags.Write:
+		mask |= syscall.O_RDWR
+	case flags.Write:
+		mask |= syscall.O_WRONLY
+	case flags.Read:
+		mask |= syscall.O_RDONLY
+	}
+	return
+}
+
+// linuxToFlags converts linux file flags to a FileFlags object.
+func linuxToFlags(mask uint) (flags fs.FileFlags) {
+	return fs.FileFlags{
+		Direct:      mask&syscall.O_DIRECT != 0,
+		Sync:        mask&syscall.O_SYNC != 0,
+		NonBlocking: mask&syscall.O_NONBLOCK != 0,
+		Read:        (mask & syscall.O_ACCMODE) != syscall.O_WRONLY,
+		Write:       (mask & syscall.O_ACCMODE) != syscall.O_RDONLY,
+		Append:      mask&syscall.O_APPEND != 0,
+		Directory:   mask&syscall.O_DIRECTORY != 0,
+	}
+}
+
+// linuxToSettableFlags converts linux file flags to a SettableFileFlags object.
+func linuxToSettableFlags(mask uint) fs.SettableFileFlags {
+	return fs.SettableFileFlags{
+		Direct:      mask&syscall.O_DIRECT != 0,
+		NonBlocking: mask&syscall.O_NONBLOCK != 0,
+		Append:      mask&syscall.O_APPEND != 0,
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go
new file mode 100644
index 000000000..44db2d582
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/linux64.go
@@ -0,0 +1,376 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package linux provides syscall tables for amd64 Linux.
+//
+// NOTE: Linux i386 support has been removed.
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi"
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/syscalls"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// AUDIT_ARCH_X86_64 identifies the Linux syscall API on AMD64, and is taken
+// from <linux/audit.h>.
+const _AUDIT_ARCH_X86_64 = 0xc000003e
+
+// AMD64 is a table of Linux amd64 syscall API with the corresponding syscall
+// numbers from Linux 3.11. The entries commented out are those syscalls we
+// don't currently support.
+var AMD64 = &kernel.SyscallTable{
+	OS:   abi.Linux,
+	Arch: arch.AMD64,
+	Version: kernel.Version{
+		Sysname: "Linux",
+		Release: "3.11.10",
+		Version: "#1 SMP Fri Nov 29 10:47:50 PST 2013",
+	},
+	AuditNumber: _AUDIT_ARCH_X86_64,
+	Table: map[uintptr]kernel.SyscallFn{
+		0:  Read,
+		1:  Write,
+		2:  Open,
+		3:  Close,
+		4:  Stat,
+		5:  Fstat,
+		6:  Lstat,
+		7:  Poll,
+		8:  Lseek,
+		9:  Mmap,
+		10: Mprotect,
+		11: Munmap,
+		12: Brk,
+		13: RtSigaction,
+		14: RtSigprocmask,
+		15: RtSigreturn,
+		16: Ioctl,
+		17: Pread64,
+		18: Pwrite64,
+		19: Readv,
+		20: Writev,
+		21: Access,
+		22: Pipe,
+		23: Select,
+		24: SchedYield,
+		25: Mremap,
+		26: Msync,
+		27: Mincore,
+		28: Madvise,
+		//     29: Shmget, TODO
+		//     30: Shmat, TODO
+		//     31: Shmctl, TODO
+		32: Dup,
+		33: Dup2,
+		34: Pause,
+		35: Nanosleep,
+		36: Getitimer,
+		37: Alarm,
+		38: Setitimer,
+		39: Getpid,
+		40: Sendfile,
+		41: Socket,
+		42: Connect,
+		43: Accept,
+		44: SendTo,
+		45: RecvFrom,
+		46: SendMsg,
+		47: RecvMsg,
+		48: Shutdown,
+		49: Bind,
+		50: Listen,
+		51: GetSockName,
+		52: GetPeerName,
+		53: SocketPair,
+		54: SetSockOpt,
+		55: GetSockOpt,
+		56: Clone,
+		57: Fork,
+		58: Vfork,
+		59: Execve,
+		60: Exit,
+		61: Wait4,
+		62: Kill,
+		63: Uname,
+		64: Semget,
+		65: Semop,
+		66: Semctl,
+		//     67: Shmdt, TODO
+		//     68: Msgget, TODO
+		//     69: Msgsnd, TODO
+		//     70: Msgrcv, TODO
+		//     71: Msgctl, TODO
+		72:  Fcntl,
+		73:  Flock,
+		74:  Fsync,
+		75:  Fdatasync,
+		76:  Truncate,
+		77:  Ftruncate,
+		78:  Getdents,
+		79:  Getcwd,
+		80:  Chdir,
+		81:  Fchdir,
+		82:  Rename,
+		83:  Mkdir,
+		84:  Rmdir,
+		85:  Creat,
+		86:  Link,
+		87:  Unlink,
+		88:  Symlink,
+		89:  Readlink,
+		90:  Chmod,
+		91:  Fchmod,
+		92:  Chown,
+		93:  Fchown,
+		94:  Lchown,
+		95:  Umask,
+		96:  Gettimeofday,
+		97:  Getrlimit,
+		98:  Getrusage,
+		99:  Sysinfo,
+		100: Times,
+		101: Ptrace,
+		102: Getuid,
+		103: Syslog,
+		104: Getgid,
+		105: Setuid,
+		106: Setgid,
+		107: Geteuid,
+		108: Getegid,
+		109: Setpgid,
+		110: Getppid,
+		111: Getpgrp,
+		112: Setsid,
+		113: Setreuid,
+		114: Setregid,
+		115: Getgroups,
+		116: Setgroups,
+		117: Setresuid,
+		118: Getresuid,
+		119: Setresgid,
+		120: Getresgid,
+		121: Getpgid,
+		//     122: Setfsuid, TODO
+		//     123: Setfsgid, TODO
+		124: Getsid,
+		125: Capget,
+		126: Capset,
+		127: RtSigpending,
+		128: RtSigtimedwait,
+		129: RtSigqueueinfo,
+		130: RtSigsuspend,
+		131: Sigaltstack,
+		132: Utime,
+		133: Mknod,
+		134: syscalls.Error(syscall.ENOSYS),          // Uselib, obsolete
+		135: syscalls.ErrorWithEvent(syscall.EINVAL), // SetPersonality, unable to change personality
+		136: syscalls.ErrorWithEvent(syscall.ENOSYS), // Ustat, needs filesystem support
+		137: Statfs,
+		138: Fstatfs,
+		//     139: Sysfs, TODO
+		140: Getpriority,
+		141: Setpriority,
+		142: syscalls.CapError(linux.CAP_SYS_NICE), // SchedSetparam, requires cap_sys_nice
+		143: SchedGetparam,
+		144: SchedSetscheduler,
+		145: SchedGetscheduler,
+		146: SchedGetPriorityMax,
+		147: SchedGetPriorityMin,
+		148: syscalls.ErrorWithEvent(syscall.EPERM),      // SchedRrGetInterval,
+		149: syscalls.Error(nil),                         // Mlock, TODO
+		150: syscalls.Error(nil),                         // Munlock, TODO
+		151: syscalls.Error(nil),                         // Mlockall, TODO
+		152: syscalls.Error(nil),                         // Munlockall, TODO
+		153: syscalls.CapError(linux.CAP_SYS_TTY_CONFIG), // Vhangup,
+		154: syscalls.Error(syscall.EPERM),               // ModifyLdt,
+		155: syscalls.Error(syscall.EPERM),               // PivotRoot,
+		156: syscalls.Error(syscall.EPERM),               // Sysctl, syscall is "worthless"
+		157: Prctl,
+		158: ArchPrctl,
+		159: syscalls.CapError(linux.CAP_SYS_TIME), // Adjtimex, requires cap_sys_time
+		160: Setrlimit,
+		161: Chroot,
+		162: Sync,
+		163: syscalls.CapError(linux.CAP_SYS_PACCT), // Acct, requires cap_sys_pacct
+		164: syscalls.CapError(linux.CAP_SYS_TIME),  // Settimeofday, requires cap_sys_time
+		165: Mount,
+		166: Umount2,
+		167: syscalls.CapError(linux.CAP_SYS_ADMIN), // Swapon, requires cap_sys_admin
+		168: syscalls.CapError(linux.CAP_SYS_ADMIN), // Swapoff, requires cap_sys_admin
+		169: syscalls.CapError(linux.CAP_SYS_BOOT),  // Reboot, requires cap_sys_boot
+		170: Sethostname,
+		171: Setdomainname,
+		172: syscalls.CapError(linux.CAP_SYS_RAWIO),  // Iopl, requires cap_sys_rawio
+		173: syscalls.CapError(linux.CAP_SYS_RAWIO),  // Ioperm, requires cap_sys_rawio
+		174: syscalls.CapError(linux.CAP_SYS_MODULE), // CreateModule, requires cap_sys_module
+		175: syscalls.CapError(linux.CAP_SYS_MODULE), // InitModule, requires cap_sys_module
+		176: syscalls.CapError(linux.CAP_SYS_MODULE), // DeleteModule, requires cap_sys_module
+		177: syscalls.Error(syscall.ENOSYS),          // GetKernelSyms, not supported in > 2.6
+		178: syscalls.Error(syscall.ENOSYS),          // QueryModule, not supported in > 2.6
+		179: syscalls.CapError(linux.CAP_SYS_ADMIN),  // Quotactl, requires cap_sys_admin (most operations)
+		180: syscalls.Error(syscall.ENOSYS),          // Nfsservctl, does not exist > 3.1
+		181: syscalls.Error(syscall.ENOSYS),          // Getpmsg, not implemented in Linux
+		182: syscalls.Error(syscall.ENOSYS),          // Putpmsg, not implemented in Linux
+		183: syscalls.Error(syscall.ENOSYS),          // AfsSyscall, not implemented in Linux
+		184: syscalls.Error(syscall.ENOSYS),          // Tuxcall, not implemented in Linux
+		185: syscalls.Error(syscall.ENOSYS),          // Security, not implemented in Linux
+		186: Gettid,
+		187: nil,                                      // Readahead, TODO
+		188: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Setxattr, requires filesystem support
+		189: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Lsetxattr, requires filesystem support
+		190: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Fsetxattr, requires filesystem support
+		191: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Getxattr, requires filesystem support
+		192: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Lgetxattr, requires filesystem support
+		193: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Fgetxattr, requires filesystem support
+		194: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Listxattr, requires filesystem support
+		195: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Llistxattr, requires filesystem support
+		196: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Flistxattr, requires filesystem support
+		197: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Removexattr, requires filesystem support
+		198: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Lremovexattr, requires filesystem support
+		199: syscalls.ErrorWithEvent(syscall.ENOTSUP), // Fremovexattr, requires filesystem support
+		200: Tkill,
+		201: Time,
+		202: Futex,
+		203: SchedSetaffinity,
+		204: SchedGetaffinity,
+		205: syscalls.Error(syscall.ENOSYS), // SetThreadArea, expected to return ENOSYS on 64-bit
+		206: IoSetup,
+		207: IoDestroy,
+		208: IoGetevents,
+		209: IoSubmit,
+		210: IoCancel,
+		211: syscalls.Error(syscall.ENOSYS),         // GetThreadArea, expected to return ENOSYS on 64-bit
+		212: syscalls.CapError(linux.CAP_SYS_ADMIN), // LookupDcookie, requires cap_sys_admin
+		213: EpollCreate,
+		214: syscalls.ErrorWithEvent(syscall.ENOSYS), // EpollCtlOld, deprecated (afaik, unused)
+		215: syscalls.ErrorWithEvent(syscall.ENOSYS), // EpollWaitOld, deprecated (afaik, unused)
+		216: syscalls.ErrorWithEvent(syscall.ENOSYS), // RemapFilePages, deprecated since 3.16
+		217: Getdents64,
+		218: SetTidAddress,
+		219: RestartSyscall,
+		//     220: Semtimedop, TODO
+		221: Fadvise64,
+		//     222: TimerCreate, TODO
+		//     223: TimerSettime, TODO
+		//     224: TimerGettime, TODO
+		//     225: TimerGetoverrun, TODO
+		//     226: TimerDelete, TODO
+		227: ClockSettime,
+		228: ClockGettime,
+		229: ClockGetres,
+		230: ClockNanosleep,
+		231: ExitGroup,
+		232: EpollWait,
+		233: EpollCtl,
+		234: Tgkill,
+		235: Utimes,
+		236: syscalls.Error(syscall.ENOSYS),        // Vserver, not implemented by Linux
+		237: syscalls.CapError(linux.CAP_SYS_NICE), // Mbind, may require cap_sys_nice TODO
+		238: SetMempolicy,
+		239: GetMempolicy,
+		//     240: MqOpen, TODO
+		//     241: MqUnlink, TODO
+		//     242: MqTimedsend, TODO
+		//     243: MqTimedreceive, TODO
+		//     244: MqNotify, TODO
+		//     245: MqGetsetattr, TODO
+		246: syscalls.CapError(linux.CAP_SYS_BOOT), // kexec_load, requires cap_sys_boot
+		247: Waitid,
+		248: syscalls.Error(syscall.EACCES),         // AddKey, not available to user
+		249: syscalls.Error(syscall.EACCES),         // RequestKey, not available to user
+		250: syscalls.Error(syscall.EACCES),         // Keyctl, not available to user
+		251: syscalls.CapError(linux.CAP_SYS_ADMIN), // IoprioSet, requires cap_sys_nice or cap_sys_admin (depending)
+		252: syscalls.CapError(linux.CAP_SYS_ADMIN), // IoprioGet, requires cap_sys_nice or cap_sys_admin (depending)
+		253: InotifyInit,
+		254: InotifyAddWatch,
+		255: InotifyRmWatch,
+		256: syscalls.CapError(linux.CAP_SYS_NICE), // MigratePages, requires cap_sys_nice
+		257: Openat,
+		258: Mkdirat,
+		259: Mknodat,
+		260: Fchownat,
+		261: Futimesat,
+		262: Fstatat,
+		263: Unlinkat,
+		264: Renameat,
+		265: Linkat,
+		266: Symlinkat,
+		267: Readlinkat,
+		268: Fchmodat,
+		269: Faccessat,
+		270: Pselect,
+		271: Ppoll,
+		272: Unshare,
+		273: syscalls.Error(syscall.ENOSYS), // SetRobustList, obsolete
+		274: syscalls.Error(syscall.ENOSYS), // GetRobustList, obsolete
+		//     275: Splice, TODO
+		//     276: Tee, TODO
+		//     277: SyncFileRange, TODO
+		//     278: Vmsplice, TODO
+		279: syscalls.CapError(linux.CAP_SYS_NICE), // MovePages, requires cap_sys_nice (mostly)
+		280: Utimensat,
+		281: EpollPwait,
+		//     282: Signalfd, TODO
+		283: TimerfdCreate,
+		284: Eventfd,
+		285: Fallocate,
+		286: TimerfdSettime,
+		287: TimerfdGettime,
+		288: Accept4,
+		//     289: Signalfd4, TODO
+		290: Eventfd2,
+		291: EpollCreate1,
+		292: Dup3,
+		293: Pipe2,
+		294: InotifyInit1,
+		295: Preadv,
+		296: Pwritev,
+		297: RtTgsigqueueinfo,
+		298: syscalls.ErrorWithEvent(syscall.ENODEV), // PerfEventOpen, no support for perf counters
+		299: RecvMMsg,
+		300: syscalls.ErrorWithEvent(syscall.ENOSYS), // FanotifyInit, needs CONFIG_FANOTIFY
+		301: syscalls.ErrorWithEvent(syscall.ENOSYS), // FanotifyMark, needs CONFIG_FANOTIFY
+		302: Prlimit64,
+		303: syscalls.ErrorWithEvent(syscall.EOPNOTSUPP), // NameToHandleAt, needs filesystem support
+		304: syscalls.ErrorWithEvent(syscall.EOPNOTSUPP), // OpenByHandleAt, needs filesystem support
+		305: syscalls.CapError(linux.CAP_SYS_TIME),       // ClockAdjtime, requires cap_sys_time
+		306: Syncfs,
+		307: SendMMsg,
+		//     308: Setns, TODO
+		309: Getcpu,
+		//     310: ProcessVmReadv, TODO may require cap_sys_ptrace
+		//     311: ProcessVmWritev, TODO may require cap_sys_ptrace
+		312: syscalls.CapError(linux.CAP_SYS_PTRACE), // Kcmp, requires cap_sys_ptrace
+		313: syscalls.CapError(linux.CAP_SYS_MODULE), // FinitModule, requires cap_sys_module
+		// "Backports."
+		318: GetRandom,
+	},
+
+	Emulate: map[usermem.Addr]uintptr{
+		0xffffffffff600000: 96,  // vsyscall gettimeofday(2)
+		0xffffffffff600400: 201, // vsyscall time(2)
+		0xffffffffff600800: 309, // vsyscall getcpu(2)
+	},
+	Missing: func(t *kernel.Task, sysno uintptr, args arch.SyscallArguments) (uintptr, error) {
+		syscalls.UnimplementedEvent(t)
+		return 0, syserror.ENOSYS
+	},
+}
diff --git a/pkg/sentry/syscalls/linux/sigset.go b/pkg/sentry/syscalls/linux/sigset.go
new file mode 100644
index 000000000..bfb541634
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sigset.go
@@ -0,0 +1,69 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// copyInSigSet copies in a sigset_t, checks its size, and ensures that KILL and
+// STOP are clear.
+func copyInSigSet(t *kernel.Task, sigSetAddr usermem.Addr, size uint) (linux.SignalSet, error) {
+	if size != linux.SignalSetSize {
+		return 0, syscall.EINVAL
+	}
+	b := t.CopyScratchBuffer(8)
+	if _, err := t.CopyInBytes(sigSetAddr, b); err != nil {
+		return 0, err
+	}
+	mask := usermem.ByteOrder.Uint64(b[:])
+	return linux.SignalSet(mask) &^ kernel.UnblockableSignals, nil
+}
+
+// copyOutSigSet copies out a sigset_t.
+func copyOutSigSet(t *kernel.Task, sigSetAddr usermem.Addr, mask linux.SignalSet) error {
+	b := t.CopyScratchBuffer(8)
+	usermem.ByteOrder.PutUint64(b, uint64(mask))
+	_, err := t.CopyOutBytes(sigSetAddr, b)
+	return err
+}
+
+// copyInSigSetWithSize copies in a structure as below
+//
+//   struct {
+//       sigset_t* sigset_addr;
+//       size_t sizeof_sigset;
+//   };
+//
+// and returns sigset_addr and size.
+func copyInSigSetWithSize(t *kernel.Task, addr usermem.Addr) (usermem.Addr, uint, error) {
+	switch t.Arch().Width() {
+	case 8:
+		in := t.CopyScratchBuffer(16)
+		if _, err := t.CopyInBytes(addr, in); err != nil {
+			return 0, 0, err
+		}
+		maskAddr := usermem.Addr(usermem.ByteOrder.Uint64(in[0:]))
+		maskSize := uint(usermem.ByteOrder.Uint64(in[8:]))
+		return maskAddr, maskSize, nil
+	default:
+		return 0, 0, syserror.ENOSYS
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/sys_aio.go b/pkg/sentry/syscalls/linux/sys_aio.go
new file mode 100644
index 000000000..80407a082
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_aio.go
@@ -0,0 +1,402 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"encoding/binary"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/eventfd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// I/O commands.
+const (
+	_IOCB_CMD_PREAD   = 0
+	_IOCB_CMD_PWRITE  = 1
+	_IOCB_CMD_FSYNC   = 2
+	_IOCB_CMD_FDSYNC  = 3
+	_IOCB_CMD_NOOP    = 6
+	_IOCB_CMD_PREADV  = 7
+	_IOCB_CMD_PWRITEV = 8
+)
+
+// I/O flags.
+const (
+	_IOCB_FLAG_RESFD = 1
+)
+
+// ioCallback describes an I/O request.
+//
+// The priority field is currently ignored in the implementation below. Also
+// note that the IOCB_FLAG_RESFD feature is not supported.
+type ioCallback struct {
+	Data      uint64
+	Key       uint32
+	Reserved1 uint32
+
+	OpCode  uint16
+	ReqPrio int16
+	FD      uint32
+
+	Buf    uint64
+	Bytes  uint64
+	Offset int64
+
+	Reserved2 uint64
+	Flags     uint32
+
+	// eventfd to signal if IOCB_FLAG_RESFD is set in flags.
+	ResFD uint32
+}
+
+// ioEvent describes an I/O result.
+type ioEvent struct {
+	Data    uint64
+	Obj     uint64
+	Result  int64
+	Result2 int64
+}
+
+// ioEventSize is the size of an ioEvent encoded.
+var ioEventSize = binary.Size(ioEvent{})
+
+// IoSetup implements linux syscall io_setup(2).
+func IoSetup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	nrEvents := args[0].Int()
+	idAddr := args[1].Pointer()
+
+	// Linux uses the native long as the aio ID.
+	//
+	// The context pointer _must_ be zero initially.
+	var idIn uint64
+	if _, err := t.CopyIn(idAddr, &idIn); err != nil {
+		return 0, nil, err
+	}
+	if idIn != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	id, err := t.MemoryManager().NewAIOContext(t, uint32(nrEvents))
+	if err != nil {
+		return 0, nil, err
+	}
+
+	// Copy out the new ID.
+	if _, err := t.CopyOut(idAddr, &id); err != nil {
+		t.MemoryManager().DestroyAIOContext(t, id)
+		return 0, nil, err
+	}
+
+	return 0, nil, nil
+}
+
+// IoDestroy implements linux syscall io_destroy(2).
+func IoDestroy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	id := args[0].Uint64()
+
+	// Destroy the given context.
+	if !t.MemoryManager().DestroyAIOContext(t, id) {
+		// Does not exist.
+		return 0, nil, syserror.EINVAL
+	}
+	// FIXME: Linux blocks until all AIO to the destroyed context is
+	// done.
+	return 0, nil, nil
+}
+
+// IoGetevents implements linux syscall io_getevents(2).
+func IoGetevents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	id := args[0].Uint64()
+	minEvents := args[1].Int()
+	events := args[2].Int()
+	eventsAddr := args[3].Pointer()
+	timespecAddr := args[4].Pointer()
+
+	// Sanity check arguments.
+	if minEvents > events {
+		return 0, nil, syserror.EINVAL
+	}
+
+	ctx, ok := t.MemoryManager().LookupAIOContext(t, id)
+	if !ok {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Setup the timeout.
+	var haveDeadline bool
+	var deadline ktime.Time
+	if timespecAddr != 0 {
+		d, err := copyTimespecIn(t, timespecAddr)
+		if err != nil {
+			return 0, nil, err
+		}
+		if !d.Valid() {
+			return 0, nil, syserror.EINVAL
+		}
+		deadline = t.Kernel().MonotonicClock().Now().Add(d.ToDuration())
+		haveDeadline = true
+	}
+
+	// Loop over all requests.
+	for count := int32(0); count < events; count++ {
+		// Get a request, per semantics.
+		var v interface{}
+		if count >= minEvents {
+			var ok bool
+			v, ok = ctx.PopRequest()
+			if !ok {
+				return uintptr(count), nil, nil
+			}
+		} else {
+			var err error
+			v, err = waitForRequest(ctx, t, haveDeadline, deadline)
+			if err != nil {
+				if count > 0 || err == syserror.ETIMEDOUT {
+					return uintptr(count), nil, nil
+				}
+				return 0, nil, syserror.ConvertIntr(err, syserror.EINTR)
+			}
+		}
+
+		ev := v.(*ioEvent)
+
+		// Copy out the result.
+		if _, err := t.CopyOut(eventsAddr, ev); err != nil {
+			if count > 0 {
+				return uintptr(count), nil, nil
+			}
+			// Nothing done.
+			return 0, nil, err
+		}
+
+		// Keep rolling.
+		eventsAddr += usermem.Addr(ioEventSize)
+	}
+
+	// Everything finished.
+	return uintptr(events), nil, nil
+}
+
+func waitForRequest(ctx *mm.AIOContext, t *kernel.Task, haveDeadline bool, deadline ktime.Time) (interface{}, error) {
+	for {
+		if v, ok := ctx.PopRequest(); ok {
+			// Request was readly available. Just return it.
+			return v, nil
+		}
+
+		// Need to wait for request completion.
+		done, active := ctx.WaitChannel()
+		if !active {
+			// Context has been destroyed.
+			return nil, syserror.EINVAL
+		}
+		if err := t.BlockWithDeadline(done, haveDeadline, deadline); err != nil {
+			return nil, err
+		}
+	}
+}
+
+// memoryFor returns appropriate memory for the given callback.
+func memoryFor(t *kernel.Task, cb *ioCallback) (usermem.IOSequence, error) {
+	bytes := int(cb.Bytes)
+	if bytes < 0 {
+		// Linux also requires that this field fit in ssize_t.
+		return usermem.IOSequence{}, syserror.EINVAL
+	}
+
+	// Since this I/O will be asynchronous with respect to t's task goroutine,
+	// we have no guarantee that t's AddressSpace will be active during the
+	// I/O.
+	switch cb.OpCode {
+	case _IOCB_CMD_PREAD, _IOCB_CMD_PWRITE:
+		return t.SingleIOSequence(usermem.Addr(cb.Buf), bytes, usermem.IOOpts{
+			AddressSpaceActive: false,
+		})
+
+	case _IOCB_CMD_PREADV, _IOCB_CMD_PWRITEV:
+		return t.IovecsIOSequence(usermem.Addr(cb.Buf), bytes, usermem.IOOpts{
+			AddressSpaceActive: false,
+		})
+
+	case _IOCB_CMD_FSYNC, _IOCB_CMD_FDSYNC, _IOCB_CMD_NOOP:
+		return usermem.IOSequence{}, nil
+
+	default:
+		// Not a supported command.
+		return usermem.IOSequence{}, syserror.EINVAL
+	}
+}
+
+func performCallback(t *kernel.Task, file *fs.File, cbAddr usermem.Addr, cb *ioCallback, ioseq usermem.IOSequence, ctx *mm.AIOContext, eventFile *fs.File) {
+	ev := &ioEvent{
+		Data: cb.Data,
+		Obj:  uint64(cbAddr),
+	}
+
+	// Construct a context.Context that will not be interrupted if t is
+	// interrupted.
+	c := t.AsyncContext()
+
+	var err error
+	switch cb.OpCode {
+	case _IOCB_CMD_PREAD, _IOCB_CMD_PREADV:
+		ev.Result, err = file.Preadv(c, ioseq, cb.Offset)
+	case _IOCB_CMD_PWRITE, _IOCB_CMD_PWRITEV:
+		ev.Result, err = file.Pwritev(c, ioseq, cb.Offset)
+	case _IOCB_CMD_FSYNC:
+		err = file.Fsync(c, 0, fs.FileMaxOffset, fs.SyncAll)
+	case _IOCB_CMD_FDSYNC:
+		err = file.Fsync(c, 0, fs.FileMaxOffset, fs.SyncData)
+	}
+
+	// Update the result.
+	if err != nil {
+		err = handleIOError(t, ev.Result != 0 /* partial */, err, nil /* never interrupted */, "aio", file)
+		ev.Result = -int64(t.ExtractErrno(err, 0))
+	}
+
+	file.DecRef()
+
+	// Queue the result for delivery.
+	ctx.FinishRequest(ev)
+
+	// Notify the event file if one was specified. This needs to happen
+	// *after* queueing the result to avoid racing with the thread we may
+	// wake up.
+	if eventFile != nil {
+		eventFile.FileOperations.(*eventfd.EventOperations).Signal(1)
+		eventFile.DecRef()
+	}
+}
+
+// submitCallback processes a single callback.
+func submitCallback(t *kernel.Task, id uint64, cb *ioCallback, cbAddr usermem.Addr) error {
+	file := t.FDMap().GetFile(kdefs.FD(cb.FD))
+	if file == nil {
+		// File not found.
+		return syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Was there an eventFD? Extract it.
+	var eventFile *fs.File
+	if cb.Flags&_IOCB_FLAG_RESFD != 0 {
+		eventFile := t.FDMap().GetFile(kdefs.FD(cb.ResFD))
+		if eventFile == nil {
+			// Bad FD.
+			return syserror.EBADF
+		}
+		defer eventFile.DecRef()
+
+		// Check that it is an eventfd.
+		if _, ok := eventFile.FileOperations.(*eventfd.EventOperations); !ok {
+			// Not an event FD.
+			return syserror.EINVAL
+		}
+	}
+
+	ioseq, err := memoryFor(t, cb)
+	if err != nil {
+		return err
+	}
+
+	// Prepare the request.
+	ctx, ok := t.MemoryManager().LookupAIOContext(t, id)
+	if !ok {
+		return syserror.EINVAL
+	}
+	if ready := ctx.Prepare(); !ready {
+		// Context is busy.
+		return syserror.EAGAIN
+	}
+
+	if eventFile != nil {
+		// The request is set. Make sure there's a ref on the file.
+		//
+		// This is necessary when the callback executes on completion,
+		// which is also what will release this reference.
+		eventFile.IncRef()
+	}
+
+	// Perform the request asynchronously.
+	file.IncRef()
+	fs.Async(func() { performCallback(t, file, cbAddr, cb, ioseq, ctx, eventFile) })
+
+	// All set.
+	return nil
+}
+
+// IoSubmit implements linux syscall io_submit(2).
+func IoSubmit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	id := args[0].Uint64()
+	nrEvents := args[1].Int()
+	addr := args[2].Pointer()
+
+	for i := int32(0); i < nrEvents; i++ {
+		// Copy in the address.
+		cbAddrNative := t.Arch().Native(0)
+		if _, err := t.CopyIn(addr, cbAddrNative); err != nil {
+			if i > 0 {
+				// Some successful.
+				return uintptr(i), nil, nil
+			}
+			// Nothing done.
+			return 0, nil, err
+		}
+
+		// Copy in this callback.
+		var cb ioCallback
+		cbAddr := usermem.Addr(t.Arch().Value(cbAddrNative))
+		if _, err := t.CopyIn(cbAddr, &cb); err != nil {
+
+			if i > 0 {
+				// Some have been successful.
+				return uintptr(i), nil, nil
+			}
+			// Nothing done.
+			return 0, nil, err
+		}
+
+		// Process this callback.
+		if err := submitCallback(t, id, &cb, cbAddr); err != nil {
+			if i > 0 {
+				// Partial success.
+				return uintptr(i), nil, nil
+			}
+			// Nothing done.
+			return 0, nil, err
+		}
+
+		// Advance to the next one.
+		addr += usermem.Addr(t.Arch().Width())
+	}
+
+	return uintptr(nrEvents), nil, nil
+}
+
+// IoCancel implements linux syscall io_cancel(2).
+//
+// It is not presently supported (ENOSYS indicates no support on this
+// architecture).
+func IoCancel(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, syserror.ENOSYS
+}
diff --git a/pkg/sentry/syscalls/linux/sys_capability.go b/pkg/sentry/syscalls/linux/sys_capability.go
new file mode 100644
index 000000000..89c81ac90
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_capability.go
@@ -0,0 +1,149 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func lookupCaps(t *kernel.Task, tid kernel.ThreadID) (permitted, inheritable, effective auth.CapabilitySet, err error) {
+	if tid < 0 {
+		err = syserror.EINVAL
+		return
+	}
+	if tid > 0 {
+		t = t.PIDNamespace().TaskWithID(tid)
+	}
+	if t == nil {
+		err = syserror.ESRCH
+		return
+	}
+	creds := t.Credentials()
+	permitted, inheritable, effective = creds.PermittedCaps, creds.InheritableCaps, creds.EffectiveCaps
+	return
+}
+
+// Capget implements Linux syscall capget.
+func Capget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	hdrAddr := args[0].Pointer()
+	dataAddr := args[1].Pointer()
+
+	var hdr linux.CapUserHeader
+	if _, err := t.CopyIn(hdrAddr, &hdr); err != nil {
+		return 0, nil, err
+	}
+	// hdr.Pid doesn't need to be valid if this capget() is a "version probe"
+	// (hdr.Version is unrecognized and dataAddr is null), so we can't do the
+	// lookup yet.
+	switch hdr.Version {
+	case linux.LINUX_CAPABILITY_VERSION_1:
+		if dataAddr == 0 {
+			return 0, nil, nil
+		}
+		p, i, e, err := lookupCaps(t, kernel.ThreadID(hdr.Pid))
+		if err != nil {
+			return 0, nil, err
+		}
+		data := linux.CapUserData{
+			Effective:   uint32(e),
+			Permitted:   uint32(p),
+			Inheritable: uint32(i),
+		}
+		_, err = t.CopyOut(dataAddr, &data)
+		return 0, nil, err
+
+	case linux.LINUX_CAPABILITY_VERSION_2, linux.LINUX_CAPABILITY_VERSION_3:
+		if dataAddr == 0 {
+			return 0, nil, nil
+		}
+		p, i, e, err := lookupCaps(t, kernel.ThreadID(hdr.Pid))
+		if err != nil {
+			return 0, nil, err
+		}
+		data := [2]linux.CapUserData{
+			{
+				Effective:   uint32(e),
+				Permitted:   uint32(p),
+				Inheritable: uint32(i),
+			},
+			{
+				Effective:   uint32(e >> 32),
+				Permitted:   uint32(p >> 32),
+				Inheritable: uint32(i >> 32),
+			},
+		}
+		_, err = t.CopyOut(dataAddr, &data)
+		return 0, nil, err
+
+	default:
+		hdr.Version = linux.HighestCapabilityVersion
+		if _, err := t.CopyOut(hdrAddr, &hdr); err != nil {
+			return 0, nil, err
+		}
+		if dataAddr != 0 {
+			return 0, nil, syserror.EINVAL
+		}
+		return 0, nil, nil
+	}
+}
+
+// Capset implements Linux syscall capset.
+func Capset(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	hdrAddr := args[0].Pointer()
+	dataAddr := args[1].Pointer()
+
+	var hdr linux.CapUserHeader
+	if _, err := t.CopyIn(hdrAddr, &hdr); err != nil {
+		return 0, nil, err
+	}
+	switch hdr.Version {
+	case linux.LINUX_CAPABILITY_VERSION_1:
+		if tid := kernel.ThreadID(hdr.Pid); tid != 0 && tid != t.ThreadID() {
+			return 0, nil, syserror.EPERM
+		}
+		var data linux.CapUserData
+		if _, err := t.CopyIn(dataAddr, &data); err != nil {
+			return 0, nil, err
+		}
+		p := auth.CapabilitySet(data.Permitted) & auth.AllCapabilities
+		i := auth.CapabilitySet(data.Inheritable) & auth.AllCapabilities
+		e := auth.CapabilitySet(data.Effective) & auth.AllCapabilities
+		return 0, nil, t.SetCapabilitySets(p, i, e)
+
+	case linux.LINUX_CAPABILITY_VERSION_2, linux.LINUX_CAPABILITY_VERSION_3:
+		if tid := kernel.ThreadID(hdr.Pid); tid != 0 && tid != t.ThreadID() {
+			return 0, nil, syserror.EPERM
+		}
+		var data [2]linux.CapUserData
+		if _, err := t.CopyIn(dataAddr, &data); err != nil {
+			return 0, nil, err
+		}
+		p := (auth.CapabilitySet(data[0].Permitted) | (auth.CapabilitySet(data[1].Permitted) << 32)) & auth.AllCapabilities
+		i := (auth.CapabilitySet(data[0].Inheritable) | (auth.CapabilitySet(data[1].Inheritable) << 32)) & auth.AllCapabilities
+		e := (auth.CapabilitySet(data[0].Effective) | (auth.CapabilitySet(data[1].Effective) << 32)) & auth.AllCapabilities
+		return 0, nil, t.SetCapabilitySets(p, i, e)
+
+	default:
+		hdr.Version = linux.HighestCapabilityVersion
+		if _, err := t.CopyOut(hdrAddr, &hdr); err != nil {
+			return 0, nil, err
+		}
+		return 0, nil, syserror.EINVAL
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/sys_epoll.go b/pkg/sentry/syscalls/linux/sys_epoll.go
new file mode 100644
index 000000000..e69dfc77a
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_epoll.go
@@ -0,0 +1,171 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/epoll"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/syscalls"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// EpollCreate1 implements the epoll_create1(2) linux syscall.
+func EpollCreate1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	flags := args[0].Int()
+	if flags & ^syscall.EPOLL_CLOEXEC != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	closeOnExec := flags&syscall.EPOLL_CLOEXEC != 0
+	fd, err := syscalls.CreateEpoll(t, closeOnExec)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// EpollCreate implements the epoll_create(2) linux syscall.
+func EpollCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	size := args[0].Int()
+
+	if size <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	fd, err := syscalls.CreateEpoll(t, false)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// EpollCtl implements the epoll_ctl(2) linux syscall.
+func EpollCtl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	epfd := kdefs.FD(args[0].Int())
+	op := args[1].Int()
+	fd := kdefs.FD(args[2].Int())
+	eventAddr := args[3].Pointer()
+
+	// Capture the event state if needed.
+	flags := epoll.EntryFlags(0)
+	mask := waiter.EventMask(0)
+	var data [2]int32
+	if op != syscall.EPOLL_CTL_DEL {
+		var e syscall.EpollEvent
+		if _, err := t.CopyIn(eventAddr, &e); err != nil {
+			return 0, nil, err
+		}
+
+		if e.Events&syscall.EPOLLONESHOT != 0 {
+			flags |= epoll.OneShot
+		}
+
+		// syscall.EPOLLET is incorrectly generated as a negative number
+		// in Go, see https://github.com/golang/go/issues/5328 for
+		// details.
+		if e.Events&-syscall.EPOLLET != 0 {
+			flags |= epoll.EdgeTriggered
+		}
+
+		mask = waiter.EventMask(e.Events)
+		data[0] = e.Fd
+		data[1] = e.Pad
+	}
+
+	// Perform the requested operations.
+	switch op {
+	case syscall.EPOLL_CTL_ADD:
+		// See fs/eventpoll.c.
+		mask |= waiter.EventHUp | waiter.EventErr
+		return 0, nil, syscalls.AddEpoll(t, epfd, fd, flags, mask, data)
+	case syscall.EPOLL_CTL_DEL:
+		return 0, nil, syscalls.RemoveEpoll(t, epfd, fd)
+	case syscall.EPOLL_CTL_MOD:
+		// Same as EPOLL_CTL_ADD.
+		mask |= waiter.EventHUp | waiter.EventErr
+		return 0, nil, syscalls.UpdateEpoll(t, epfd, fd, flags, mask, data)
+	default:
+		return 0, nil, syserror.EINVAL
+	}
+}
+
+// copyOutEvents copies epoll events from the kernel to user memory.
+func copyOutEvents(t *kernel.Task, addr usermem.Addr, e []epoll.Event) error {
+	const itemLen = 12
+	if _, ok := addr.AddLength(uint64(len(e)) * itemLen); !ok {
+		return syserror.EFAULT
+	}
+
+	b := t.CopyScratchBuffer(itemLen)
+	for i := range e {
+		usermem.ByteOrder.PutUint32(b[0:], e[i].Events)
+		usermem.ByteOrder.PutUint32(b[4:], uint32(e[i].Data[0]))
+		usermem.ByteOrder.PutUint32(b[8:], uint32(e[i].Data[1]))
+		if _, err := t.CopyOutBytes(addr, b); err != nil {
+			return err
+		}
+		addr += itemLen
+	}
+
+	return nil
+}
+
+// EpollWait implements the epoll_wait(2) linux syscall.
+func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	epfd := kdefs.FD(args[0].Int())
+	eventsAddr := args[1].Pointer()
+	maxEvents := int(args[2].Int())
+	timeout := int(args[3].Int())
+
+	r, err := syscalls.WaitEpoll(t, epfd, maxEvents, timeout)
+	if err != nil {
+		return 0, nil, syserror.ConvertIntr(err, syserror.EINTR)
+	}
+
+	if len(r) != 0 {
+		if err := copyOutEvents(t, eventsAddr, r); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	return uintptr(len(r)), nil, nil
+}
+
+// EpollPwait implements the epoll_pwait(2) linux syscall.
+func EpollPwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	maskAddr := args[4].Pointer()
+	maskSize := uint(args[5].Uint())
+
+	if maskAddr != 0 {
+		mask, err := copyInSigSet(t, maskAddr, maskSize)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		oldmask := t.SignalMask()
+		t.SetSignalMask(mask)
+		t.SetSavedSignalMask(oldmask)
+	}
+
+	return EpollWait(t, args)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_eventfd.go b/pkg/sentry/syscalls/linux/sys_eventfd.go
new file mode 100644
index 000000000..60fe5a133
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_eventfd.go
@@ -0,0 +1,65 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/eventfd"
+)
+
+const (
+	// EFD_SEMAPHORE is a flag used in syscall eventfd(2) and eventfd2(2). Please
+	// see its man page for more information.
+	EFD_SEMAPHORE = 1
+	EFD_NONBLOCK  = 0x800
+	EFD_CLOEXEC   = 0x80000
+)
+
+// Eventfd2 implements linux syscall eventfd2(2).
+func Eventfd2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	initVal := args[0].Int()
+	flags := uint(args[1].Uint())
+	allOps := uint(EFD_SEMAPHORE | EFD_NONBLOCK | EFD_CLOEXEC)
+
+	if flags & ^allOps != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	event := eventfd.New(t, uint64(initVal), flags&EFD_SEMAPHORE != 0)
+	event.SetFlags(fs.SettableFileFlags{
+		NonBlocking: flags&EFD_NONBLOCK != 0,
+	})
+	defer event.DecRef()
+
+	fd, err := t.FDMap().NewFDFrom(0, event, kernel.FDFlags{
+		CloseOnExec: flags&EFD_CLOEXEC != 0,
+	},
+		t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// Eventfd implements linux syscall eventfd(2).
+func Eventfd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	args[1].Value = 0
+	return Eventfd2(t, args)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go
new file mode 100644
index 000000000..a2dbba7e0
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_file.go
@@ -0,0 +1,1942 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"io"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/lock"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// fileOpAt performs an operation on the second last component in the path.
+func fileOpAt(t *kernel.Task, dirFD kdefs.FD, path string, fn func(root *fs.Dirent, d *fs.Dirent, name string) error) error {
+	// Extract the last component.
+	dir, name := fs.SplitLast(path)
+	if dir == "/" {
+		// Common case: we are accessing a file in the root.
+		root := t.FSContext().RootDirectory()
+		err := fn(root, root, name)
+		root.DecRef()
+		return err
+	} else if dir == "." && dirFD == linux.AT_FDCWD {
+		// Common case: we are accessing a file relative to the current
+		// working directory; skip the look-up.
+		wd := t.FSContext().WorkingDirectory()
+		root := t.FSContext().RootDirectory()
+		err := fn(root, wd, name)
+		wd.DecRef()
+		root.DecRef()
+		return err
+	}
+
+	return fileOpOn(t, dirFD, dir, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return fn(root, d, name)
+	})
+}
+
+// fileOpOn performs an operation on the last entry of the path.
+func fileOpOn(t *kernel.Task, dirFD kdefs.FD, path string, resolve bool, fn func(root *fs.Dirent, d *fs.Dirent) error) error {
+	var (
+		d   *fs.Dirent // The file.
+		wd  *fs.Dirent // The working directory (if required.)
+		rel *fs.Dirent // The relative directory for search (if required.)
+		f   *fs.File   // The file corresponding to dirFD (if required.)
+		err error
+	)
+
+	// Extract the working directory (maybe).
+	if len(path) > 0 && path[0] == '/' {
+		// Absolute path; rel can be nil.
+	} else if dirFD == linux.AT_FDCWD {
+		// Need to reference the working directory.
+		wd = t.FSContext().WorkingDirectory()
+		rel = wd
+	} else {
+		// Need to extract the given FD.
+		f = t.FDMap().GetFile(dirFD)
+		if f == nil {
+			return syserror.EBADF
+		}
+		rel = f.Dirent
+		if !fs.IsDir(rel.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+	}
+
+	// Grab the root (always required.)
+	root := t.FSContext().RootDirectory()
+
+	// Lookup the node.
+	if resolve {
+		d, err = t.MountNamespace().FindInode(t, root, rel, path, linux.MaxSymlinkTraversals)
+	} else {
+		d, err = t.MountNamespace().FindLink(t, root, rel, path, linux.MaxSymlinkTraversals)
+	}
+	root.DecRef()
+	if wd != nil {
+		wd.DecRef()
+	}
+	if f != nil {
+		f.DecRef()
+	}
+	if err != nil {
+		return err
+	}
+
+	err = fn(root, d)
+	d.DecRef()
+	return err
+}
+
+// copyInPath copies a path in.
+func copyInPath(t *kernel.Task, addr usermem.Addr, allowEmpty bool) (path string, dirPath bool, err error) {
+	path, err = t.CopyInString(addr, syscall.PathMax)
+	if err != nil {
+		return "", false, err
+	}
+	if path == "" && !allowEmpty {
+		return "", false, syserror.ENOENT
+	}
+
+	// If the path ends with a /, then checks must be enforced in various
+	// ways in the different callers. We pass this back to the caller.
+	path, dirPath = fs.TrimTrailingSlashes(path)
+
+	return path, dirPath, nil
+}
+
+func openAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, flags uint) (fd uintptr, err error) {
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, err
+	}
+
+	err = fileOpOn(t, dirFD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		// First check a few things about the filesystem before trying to get the file
+		// reference.
+		//
+		// It's required that Check does not try to open files not that aren't backed by
+		// this dirent (e.g. pipes and sockets) because this would result in opening these
+		// files an extra time just to check permissions.
+		if err := d.Inode.CheckPermission(t, flagsToPermissions(flags)); err != nil {
+			return err
+		}
+
+		fileFlags := linuxToFlags(flags)
+		isDir := fs.IsDir(d.Inode.StableAttr)
+
+		// If O_DIRECTORY is set, but the file is not a directory, then fail.
+		if fileFlags.Directory && !isDir {
+			return syserror.ENOTDIR
+		}
+
+		// If it's a directory, then make sure.
+		if dirPath && !isDir {
+			return syserror.ENOTDIR
+		}
+
+		// Don't allow directories to be opened writable.
+		if isDir && fileFlags.Write {
+			return syserror.EISDIR
+		}
+
+		file, err := d.Inode.GetFile(t, d, fileFlags)
+		if err != nil {
+			return syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+		}
+		defer file.DecRef()
+
+		// Success.
+		fdFlags := kernel.FDFlags{CloseOnExec: flags&syscall.O_CLOEXEC != 0}
+		newFD, err := t.FDMap().NewFDFrom(0, file, fdFlags, t.ThreadGroup().Limits())
+		if err != nil {
+			return err
+		}
+
+		// Set return result in frame.
+		fd = uintptr(newFD)
+
+		// Generate notification for opened file.
+		d.InotifyEvent(linux.IN_OPEN, 0)
+
+		return nil
+	})
+	return fd, err // Use result in frame.
+}
+
+func mknodAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, mode linux.FileMode) error {
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+	if dirPath {
+		return syserror.ENOENT
+	}
+
+	return fileOpAt(t, dirFD, path, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Do we have the appropriate permissions on the parent?
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+			return err
+		}
+
+		// Attempt a creation.
+		perms := fs.FilePermsFromMode(mode &^ linux.FileMode(t.FSContext().Umask()))
+
+		switch mode.FileType() {
+		case 0:
+			// "Zero file type is equivalent to type S_IFREG." - mknod(2)
+			fallthrough
+		case linux.ModeRegular:
+			// We are not going to return the file, so the actual
+			// flags used don't matter, but they cannot be empty or
+			// Create will complain.
+			flags := fs.FileFlags{Read: true, Write: true}
+			file, err := d.Create(t, root, name, flags, perms)
+			if err != nil {
+				return err
+			}
+			file.DecRef()
+			return nil
+
+		case linux.ModeNamedPipe:
+			return d.CreateFifo(t, root, name, perms)
+
+		case linux.ModeSocket:
+			// While it is possible create a unix domain socket file on linux
+			// using mknod(2), in practice this is pretty useless from an
+			// application. Linux internally uses mknod() to create the socket
+			// node during bind(2), but we implement bind(2) independently. If
+			// an application explicitly creates a socket node using mknod(),
+			// you can't seem to bind() or connect() to the resulting socket.
+			//
+			// Instead of emulating this seemingly useless behaviour, we'll
+			// indicate that the filesystem doesn't support the creation of
+			// sockets.
+			return syserror.EOPNOTSUPP
+
+		case linux.ModeCharacterDevice:
+			fallthrough
+		case linux.ModeBlockDevice:
+			// TODO: We don't support creating block or character
+			// devices at the moment.
+			//
+			// When we start supporting block and character devices, we'll
+			// need to check for CAP_MKNOD here.
+			return syserror.EPERM
+
+		default:
+			// "EINVAL - mode requested creation of something other than a
+			// regular file, device special file, FIFO or socket." - mknod(2)
+			return syserror.EINVAL
+		}
+	})
+}
+
+// Mknod implements the linux syscall mknod(2).
+func Mknod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	path := args[0].Pointer()
+	mode := linux.FileMode(args[1].ModeT())
+	// We don't need this argument until we support creation of device nodes.
+	_ = args[2].Uint() // dev
+
+	return 0, nil, mknodAt(t, linux.AT_FDCWD, path, mode)
+}
+
+// Mknodat implements the linux syscall mknodat(2).
+func Mknodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	path := args[1].Pointer()
+	mode := linux.FileMode(args[2].ModeT())
+	// We don't need this argument until we support creation of device nodes.
+	_ = args[3].Uint() // dev
+
+	return 0, nil, mknodAt(t, dirFD, path, mode)
+}
+
+func createAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, flags uint, mode linux.FileMode) (fd uintptr, err error) {
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, err
+	}
+	if dirPath {
+		return 0, syserror.ENOENT
+	}
+
+	err = fileOpAt(t, dirFD, path, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Does this file exist already?
+		targetDirent, err := t.MountNamespace().FindInode(t, root, d, name, linux.MaxSymlinkTraversals)
+		var newFile *fs.File
+		switch err {
+		case nil:
+			// The file existed.
+			defer targetDirent.DecRef()
+
+			// Check if we wanted to create.
+			if flags&syscall.O_EXCL != 0 {
+				return syserror.EEXIST
+			}
+
+			// Like sys_open, check for a few things about the
+			// filesystem before trying to get a reference to the
+			// fs.File. The same constraints on Check apply.
+			if err := targetDirent.Inode.CheckPermission(t, flagsToPermissions(flags)); err != nil {
+				return err
+			}
+
+			// Should we truncate the file?
+			if flags&syscall.O_TRUNC != 0 {
+				if err := targetDirent.Inode.Truncate(t, targetDirent, 0); err != nil {
+					return err
+				}
+			}
+
+			// Create a new fs.File.
+			newFile, err = targetDirent.Inode.GetFile(t, targetDirent, linuxToFlags(flags))
+			if err != nil {
+				return syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+			}
+			defer newFile.DecRef()
+		case syserror.EACCES:
+			// Permission denied while walking to the file.
+			return err
+		default:
+			// Do we have write permissions on the parent?
+			if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+				return err
+			}
+
+			// Attempt a creation.
+			perms := fs.FilePermsFromMode(mode &^ linux.FileMode(t.FSContext().Umask()))
+			newFile, err = d.Create(t, root, name, linuxToFlags(flags), perms)
+			if err != nil {
+				// No luck, bail.
+				return err
+			}
+			defer newFile.DecRef()
+			targetDirent = newFile.Dirent
+		}
+
+		// Success.
+		fdFlags := kernel.FDFlags{CloseOnExec: flags&syscall.O_CLOEXEC != 0}
+		newFD, err := t.FDMap().NewFDFrom(0, newFile, fdFlags, t.ThreadGroup().Limits())
+		if err != nil {
+			return err
+		}
+
+		// Set result in frame.
+		fd = uintptr(newFD)
+
+		// Queue the open inotify event. The creation event is
+		// automatically queued when the dirent is targetDirent. The
+		// open events are implemented at the syscall layer so we need
+		// to manually queue one here.
+		targetDirent.InotifyEvent(linux.IN_OPEN, 0)
+
+		return nil
+	})
+	return fd, err // Use result in frame.
+}
+
+// Open implements linux syscall open(2).
+func Open(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	flags := uint(args[1].Uint())
+	if flags&syscall.O_CREAT != 0 {
+		mode := linux.FileMode(args[2].ModeT())
+		n, err := createAt(t, linux.AT_FDCWD, addr, flags, mode)
+		return n, nil, err
+	}
+	n, err := openAt(t, linux.AT_FDCWD, addr, flags)
+	return n, nil, err
+}
+
+// Openat implements linux syscall openat(2).
+func Openat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	flags := uint(args[2].Uint())
+	if flags&syscall.O_CREAT != 0 {
+		mode := linux.FileMode(args[3].ModeT())
+		n, err := createAt(t, dirFD, addr, flags, mode)
+		return n, nil, err
+	}
+	n, err := openAt(t, dirFD, addr, flags)
+	return n, nil, err
+}
+
+// Creat implements linux syscall creat(2).
+func Creat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	mode := linux.FileMode(args[1].ModeT())
+	n, err := createAt(t, linux.AT_FDCWD, addr, syscall.O_WRONLY|syscall.O_TRUNC, mode)
+	return n, nil, err
+}
+
+// accessContext is a context that overrides the credentials used, but
+// otherwise carries the same values as the embedded context.
+//
+// accessContext should only be used for access(2).
+type accessContext struct {
+	context.Context
+	creds auth.Credentials
+}
+
+// Value implements context.Context.
+func (ac accessContext) Value(key interface{}) interface{} {
+	switch key {
+	case auth.CtxCredentials:
+		return &ac.creds
+	default:
+		return ac.Context.Value(key)
+	}
+}
+
+func accessAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, resolve bool, mode uint) error {
+	const rOK = 4
+	const wOK = 2
+	const xOK = 1
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	// Sanity check the mode.
+	if mode&^(rOK|wOK|xOK) != 0 {
+		return syserror.EINVAL
+	}
+
+	return fileOpOn(t, dirFD, path, resolve, func(root *fs.Dirent, d *fs.Dirent) error {
+		// access(2) and faccessat(2) check permissions using real
+		// UID/GID, not effective UID/GID.
+		//
+		// "access() needs to use the real uid/gid, not the effective
+		// uid/gid. We do this by temporarily clearing all FS-related
+		// capabilities and switching the fsuid/fsgid around to the
+		// real ones." -fs/open.c:faccessat
+		creds := t.Credentials()
+		creds.EffectiveKUID = creds.RealKUID
+		creds.EffectiveKGID = creds.RealKGID
+		if creds.EffectiveKUID.In(creds.UserNamespace) == auth.RootUID {
+			creds.EffectiveCaps = creds.PermittedCaps
+		} else {
+			creds.EffectiveCaps = 0
+		}
+
+		ctx := &accessContext{
+			Context: t,
+			creds:   creds,
+		}
+
+		if err := d.Inode.CheckPermission(ctx, fs.PermMask{
+			Read:    mode&rOK != 0,
+			Write:   mode&wOK != 0,
+			Execute: mode&xOK != 0,
+		}); err != nil {
+			return err
+		}
+		return nil
+	})
+}
+
+// Access implements linux syscall access(2).
+func Access(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	mode := args[1].ModeT()
+
+	return 0, nil, accessAt(t, linux.AT_FDCWD, addr, true, mode)
+}
+
+// Faccessat implements linux syscall faccessat(2).
+func Faccessat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	mode := args[2].ModeT()
+	flags := args[3].Int()
+
+	return 0, nil, accessAt(t, dirFD, addr, flags&linux.AT_SYMLINK_NOFOLLOW == 0, mode)
+}
+
+// Ioctl implements linux syscall ioctl(2).
+func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	request := int(args[1].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Shared flags between file and socket.
+	switch request {
+	case linux.FIONCLEX:
+		t.FDMap().SetFlags(fd, kernel.FDFlags{
+			CloseOnExec: false,
+		})
+		return 0, nil, nil
+	case linux.FIOCLEX:
+		t.FDMap().SetFlags(fd, kernel.FDFlags{
+			CloseOnExec: true,
+		})
+		return 0, nil, nil
+
+	case linux.FIONBIO:
+		var set int32
+		if _, err := t.CopyIn(args[2].Pointer(), &set); err != nil {
+			return 0, nil, err
+		}
+		flags := file.Flags()
+		if set != 0 {
+			flags.NonBlocking = true
+		} else {
+			flags.NonBlocking = false
+		}
+		file.SetFlags(flags.Settable())
+		return 0, nil, nil
+
+	default:
+		ret, err := file.FileOperations.Ioctl(t, t.MemoryManager(), args)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		return ret, nil, nil
+	}
+}
+
+// Getcwd implements the linux syscall getcwd(2).
+func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	size := args[1].SizeT()
+	cwd := t.FSContext().WorkingDirectory()
+	defer cwd.DecRef()
+	root := t.FSContext().RootDirectory()
+	defer root.DecRef()
+
+	// Get our fullname from the root and preprend unreachable if the root was
+	// unreachable from our current dirent this is the same behavior as on linux.
+	s, reachable := cwd.FullName(root)
+	if !reachable {
+		s = "(unreachable)" + s
+	}
+
+	// Note this is >= because we need a terminator.
+	if uint(len(s)) >= size {
+		return 0, nil, syserror.ERANGE
+	}
+
+	// Copy out the path name for the node.
+	bytes, err := t.CopyOutBytes(addr, []byte(s))
+	if err != nil {
+		return 0, nil, err
+	}
+
+	// Top it off with a terminator.
+	_, err = t.CopyOut(addr+usermem.Addr(bytes), []byte("\x00"))
+	return uintptr(bytes + 1), nil, err
+}
+
+// Chroot implements the linux syscall chroot(2).
+func Chroot(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	if !t.HasCapability(linux.CAP_SYS_CHROOT) {
+		return 0, nil, syserror.EPERM
+	}
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		// Is it a directory?
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Does it have execute permissions?
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Execute: true}); err != nil {
+			return err
+		}
+
+		t.FSContext().SetRootDirectory(d)
+		return nil
+	})
+}
+
+// Chdir implements the linux syscall chdir(2).
+func Chdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		// Is it a directory?
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Does it have execute permissions?
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Execute: true}); err != nil {
+			return err
+		}
+
+		t.FSContext().SetWorkingDirectory(d)
+		return nil
+	})
+}
+
+// Fchdir implements the linux syscall fchdir(2).
+func Fchdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Is it a directory?
+	if !fs.IsDir(file.Dirent.Inode.StableAttr) {
+		return 0, nil, syserror.ENOTDIR
+	}
+
+	// Does it have execute permissions?
+	if err := file.Dirent.Inode.CheckPermission(t, fs.PermMask{Execute: true}); err != nil {
+		return 0, nil, err
+	}
+
+	t.FSContext().SetWorkingDirectory(file.Dirent)
+	return 0, nil, nil
+}
+
+// Close implements linux syscall close(2).
+func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file, ok := t.FDMap().Remove(fd)
+	if !ok {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	err := file.Flush(t)
+	return 0, nil, handleIOError(t, false /* partial */, err, syscall.EINTR, "close", file)
+}
+
+// Dup implements linux syscall dup(2).
+func Dup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	newfd, err := t.FDMap().NewFDFrom(0, file, kernel.FDFlags{}, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, syserror.EMFILE
+	}
+	return uintptr(newfd), nil, nil
+}
+
+// Dup2 implements linux syscall dup2(2).
+func Dup2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldfd := kdefs.FD(args[0].Int())
+	newfd := kdefs.FD(args[1].Int())
+
+	// If oldfd is a valid file descriptor, and newfd has the same value as oldfd,
+	// then dup2() does nothing, and returns newfd.
+	if oldfd == newfd {
+		oldFile := t.FDMap().GetFile(oldfd)
+		if oldFile == nil {
+			return 0, nil, syserror.EBADF
+		}
+		defer oldFile.DecRef()
+
+		return uintptr(newfd), nil, nil
+	}
+
+	// Zero out flags arg to be used by Dup3.
+	args[2].Value = 0
+	return Dup3(t, args)
+}
+
+// Dup3 implements linux syscall dup3(2).
+func Dup3(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldfd := kdefs.FD(args[0].Int())
+	newfd := kdefs.FD(args[1].Int())
+	flags := args[2].Uint()
+
+	if oldfd == newfd {
+		return 0, nil, syserror.EINVAL
+	}
+
+	oldFile := t.FDMap().GetFile(oldfd)
+	if oldFile == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer oldFile.DecRef()
+
+	err := t.FDMap().NewFDAt(newfd, oldFile, kernel.FDFlags{CloseOnExec: flags&syscall.O_CLOEXEC != 0}, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(newfd), nil, nil
+}
+
+// Fcntl implements linux syscall fcntl(2).
+func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	cmd := args[1].Int()
+
+	file, flags := t.FDMap().GetDescriptor(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	switch cmd {
+	case syscall.F_DUPFD, syscall.F_DUPFD_CLOEXEC:
+		from := kdefs.FD(args[2].Int())
+		fdFlags := kernel.FDFlags{CloseOnExec: cmd == syscall.F_DUPFD_CLOEXEC}
+		fd, err := t.FDMap().NewFDFrom(from, file, fdFlags, t.ThreadGroup().Limits())
+		if err != nil {
+			return 0, nil, err
+		}
+		return uintptr(fd), nil, nil
+	case syscall.F_GETFD:
+		return uintptr(fdFlagsToLinux(flags)), nil, nil
+	case syscall.F_SETFD:
+		flags := args[2].Uint()
+		t.FDMap().SetFlags(fd, kernel.FDFlags{
+			CloseOnExec: flags&syscall.FD_CLOEXEC != 0,
+		})
+	case syscall.F_GETFL:
+		return uintptr(flagsToLinux(file.Flags())), nil, nil
+	case syscall.F_SETFL:
+		flags := uint(args[2].Uint())
+		file.SetFlags(linuxToSettableFlags(flags))
+	case syscall.F_SETLK, syscall.F_SETLKW:
+		// In Linux the file system can choose to provide lock operations for an inode.
+		// Normally pipe and socket types lack lock operations. We diverge and use a heavy
+		// hammer by only allowing locks on files and directories.
+		if !fs.IsFile(file.Dirent.Inode.StableAttr) && !fs.IsDir(file.Dirent.Inode.StableAttr) {
+			return 0, nil, syserror.EBADF
+		}
+
+		// Copy in the lock request.
+		flockAddr := args[2].Pointer()
+		var flock syscall.Flock_t
+		if _, err := t.CopyIn(flockAddr, &flock); err != nil {
+			return 0, nil, err
+		}
+
+		// Compute the lock whence.
+		var sw fs.SeekWhence
+		switch flock.Whence {
+		case 0:
+			sw = fs.SeekSet
+		case 1:
+			sw = fs.SeekCurrent
+		case 2:
+			sw = fs.SeekEnd
+		default:
+			return 0, nil, syserror.EINVAL
+		}
+
+		// Compute the lock offset.
+		var off int64
+		switch sw {
+		case fs.SeekSet:
+			off = 0
+		case fs.SeekCurrent:
+			// Note that Linux does not hold any mutexes while retrieving the file offset,
+			// see fs/locks.c:flock_to_posix_lock and fs/locks.c:fcntl_setlk.
+			off = file.Offset()
+		case fs.SeekEnd:
+			uattr, err := file.Dirent.Inode.UnstableAttr(t)
+			if err != nil {
+				return 0, nil, err
+			}
+			off = uattr.Size
+		default:
+			return 0, nil, syserror.EINVAL
+		}
+
+		// Compute the lock range.
+		rng, err := lock.ComputeRange(flock.Start, flock.Len, off)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		// The lock uid is that of the Task's FDMap.
+		lockUniqueID := lock.UniqueID(t.FDMap().ID())
+
+		// These locks don't block; execute the non-blocking operation using the inode's lock
+		// context directly.
+		switch flock.Type {
+		case syscall.F_RDLCK:
+			if !file.Flags().Read {
+				return 0, nil, syserror.EBADF
+			}
+			if cmd == syscall.F_SETLK {
+				// Non-blocking lock, provide a nil lock.Blocker.
+				if !file.Dirent.Inode.LockCtx.Posix.LockRegion(lockUniqueID, lock.ReadLock, rng, nil) {
+					return 0, nil, syserror.EAGAIN
+				}
+			} else {
+				// Blocking lock, pass in the task to satisfy the lock.Blocker interface.
+				if !file.Dirent.Inode.LockCtx.Posix.LockRegion(lockUniqueID, lock.ReadLock, rng, t) {
+					return 0, nil, syserror.EINTR
+				}
+			}
+			return 0, nil, nil
+		case syscall.F_WRLCK:
+			if !file.Flags().Write {
+				return 0, nil, syserror.EBADF
+			}
+			if cmd == syscall.F_SETLK {
+				// Non-blocking lock, provide a nil lock.Blocker.
+				if !file.Dirent.Inode.LockCtx.Posix.LockRegion(lockUniqueID, lock.WriteLock, rng, nil) {
+					return 0, nil, syserror.EAGAIN
+				}
+			} else {
+				// Blocking lock, pass in the task to satisfy the lock.Blocker interface.
+				if !file.Dirent.Inode.LockCtx.Posix.LockRegion(lockUniqueID, lock.WriteLock, rng, t) {
+					return 0, nil, syserror.EINTR
+				}
+			}
+			return 0, nil, nil
+		case syscall.F_UNLCK:
+			file.Dirent.Inode.LockCtx.Posix.UnlockRegion(lockUniqueID, rng)
+			return 0, nil, nil
+		default:
+			return 0, nil, syserror.EINVAL
+		}
+	default:
+		// Everything else is not yet supported.
+		return 0, nil, syserror.EINVAL
+	}
+	return 0, nil, nil
+}
+
+const (
+	_FADV_NORMAL     = 0
+	_FADV_RANDOM     = 1
+	_FADV_SEQUENTIAL = 2
+	_FADV_WILLNEED   = 3
+	_FADV_DONTNEED   = 4
+	_FADV_NOREUSE    = 5
+)
+
+// Fadvise64 implements linux syscall fadvise64(2).
+// This implementation currently ignores the provided advice.
+func Fadvise64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	offset := args[1].Int64()
+	length := args[2].Uint()
+	advice := args[3].Int()
+
+	if offset < 0 || length < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	switch advice {
+	case _FADV_NORMAL:
+	case _FADV_RANDOM:
+	case _FADV_SEQUENTIAL:
+	case _FADV_WILLNEED:
+	case _FADV_DONTNEED:
+	case _FADV_NOREUSE:
+	default:
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Sure, whatever.
+	return 0, nil, nil
+}
+
+func mkdirAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, mode linux.FileMode) error {
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	return fileOpAt(t, dirFD, path, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Does this directory exist already?
+		f, err := t.MountNamespace().FindInode(t, root, d, name, linux.MaxSymlinkTraversals)
+		switch err {
+		case nil:
+			// The directory existed.
+			defer f.DecRef()
+			return syserror.EEXIST
+		case syserror.EACCES:
+			// Permission denied while walking to the directory.
+			return err
+		default:
+			// Do we have write permissions on the parent?
+			if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+				return err
+			}
+
+			// Create the directory.
+			perms := fs.FilePermsFromMode(mode &^ linux.FileMode(t.FSContext().Umask()))
+			return d.CreateDirectory(t, root, name, perms)
+		}
+	})
+}
+
+// Mkdir implements linux syscall mkdir(2).
+func Mkdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	mode := linux.FileMode(args[1].ModeT())
+
+	return 0, nil, mkdirAt(t, linux.AT_FDCWD, addr, mode)
+}
+
+// Mkdirat implements linux syscall mkdirat(2).
+func Mkdirat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	mode := linux.FileMode(args[2].ModeT())
+
+	return 0, nil, mkdirAt(t, dirFD, addr, mode)
+}
+
+func rmdirAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr) error {
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	// Special case: rmdir rejects anything with '.' as last component.
+	// This would be handled by the busy check for the current working
+	// directory, but this is how it's done.
+	if (len(path) == 1 && path == ".") || (len(path) > 1 && path[len(path)-2:] == "/.") {
+		return syserror.EINVAL
+	}
+
+	return fileOpAt(t, dirFD, path, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		if err := fs.MayDelete(t, root, d, name); err != nil {
+			return err
+		}
+
+		return d.RemoveDirectory(t, root, name)
+	})
+}
+
+// Rmdir implements linux syscall rmdir(2).
+func Rmdir(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	return 0, nil, rmdirAt(t, linux.AT_FDCWD, addr)
+}
+
+func symlinkAt(t *kernel.Task, dirFD kdefs.FD, newAddr usermem.Addr, oldAddr usermem.Addr) error {
+	newPath, dirPath, err := copyInPath(t, newAddr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+	if dirPath {
+		return syserror.ENOENT
+	}
+
+	// The oldPath is copied in verbatim. This is because the symlink
+	// will include all details, including trailing slashes.
+	oldPath, err := t.CopyInString(oldAddr, syscall.PathMax)
+	if err != nil {
+		return err
+	}
+	if oldPath == "" {
+		return syserror.ENOENT
+	}
+
+	return fileOpAt(t, dirFD, newPath, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Make sure we have write permissions on the parent directory.
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+			return err
+		}
+		return d.CreateLink(t, root, oldPath, name)
+	})
+}
+
+// Symlink implements linux syscall symlink(2).
+func Symlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldAddr := args[0].Pointer()
+	newAddr := args[1].Pointer()
+
+	return 0, nil, symlinkAt(t, linux.AT_FDCWD, newAddr, oldAddr)
+}
+
+// Symlinkat implements linux syscall symlinkat(2).
+func Symlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldAddr := args[0].Pointer()
+	dirFD := kdefs.FD(args[1].Int())
+	newAddr := args[2].Pointer()
+
+	return 0, nil, symlinkAt(t, dirFD, newAddr, oldAddr)
+}
+
+// mayLinkAt determines whether t can create a hard link to target.
+//
+// This corresponds to Linux's fs/namei.c:may_linkat.
+func mayLinkAt(t *kernel.Task, target *fs.Inode) error {
+	// Technically Linux is more restrictive in 3.11.10 (requires CAP_FOWNER in
+	// root user namespace); this is from the later f2ca379642d7 "namei: permit
+	// linking with CAP_FOWNER in userns".
+	if !target.CheckOwnership(t) {
+		return syserror.EPERM
+	}
+
+	// Check that the target is not a directory and that permissions are okay.
+	if fs.IsDir(target.StableAttr) || target.CheckPermission(t, fs.PermMask{Read: true, Write: true}) != nil {
+		return syserror.EPERM
+	}
+
+	return nil
+}
+
+// linkAt creates a hard link to the target specified by oldDirFD and oldAddr,
+// specified by newDirFD and newAddr.  If resolve is true, then the symlinks
+// will be followed when evaluating the target.
+func linkAt(t *kernel.Task, oldDirFD kdefs.FD, oldAddr usermem.Addr, newDirFD kdefs.FD, newAddr usermem.Addr, resolve, allowEmpty bool) error {
+	oldPath, _, err := copyInPath(t, oldAddr, allowEmpty)
+	if err != nil {
+		return err
+	}
+	newPath, dirPath, err := copyInPath(t, newAddr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+	if dirPath {
+		return syserror.ENOENT
+	}
+
+	if allowEmpty && oldPath == "" {
+		target := t.FDMap().GetFile(oldDirFD)
+		if target == nil {
+			return syserror.EBADF
+		}
+		defer target.DecRef()
+		if err := mayLinkAt(t, target.Dirent.Inode); err != nil {
+			return err
+		}
+
+		// Resolve the target directory.
+		return fileOpAt(t, newDirFD, newPath, func(root *fs.Dirent, newParent *fs.Dirent, newName string) error {
+			if !fs.IsDir(newParent.Inode.StableAttr) {
+				return syserror.ENOTDIR
+			}
+
+			// Make sure we have write permissions on the parent directory.
+			if err := newParent.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+				return err
+			}
+			return newParent.CreateHardLink(t, root, target.Dirent, newName)
+		})
+	}
+
+	// Resolve oldDirFD and oldAddr to a dirent.  The "resolve" argument
+	// only applies to this name.
+	return fileOpOn(t, oldDirFD, oldPath, resolve, func(root *fs.Dirent, target *fs.Dirent) error {
+		if err := mayLinkAt(t, target.Inode); err != nil {
+			return err
+		}
+
+		// Next resolve newDirFD and newAddr to the parent dirent and name.
+		return fileOpAt(t, newDirFD, newPath, func(root *fs.Dirent, newParent *fs.Dirent, newName string) error {
+			if !fs.IsDir(newParent.Inode.StableAttr) {
+				return syserror.ENOTDIR
+			}
+
+			// Make sure we have write permissions on the parent directory.
+			if err := newParent.Inode.CheckPermission(t, fs.PermMask{Write: true, Execute: true}); err != nil {
+				return err
+			}
+			return newParent.CreateHardLink(t, root, target, newName)
+		})
+	})
+}
+
+// Link implements linux syscall link(2).
+func Link(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldAddr := args[0].Pointer()
+	newAddr := args[1].Pointer()
+
+	// man link(2):
+	// POSIX.1-2001 says that link() should dereference oldpath if it is a
+	// symbolic link. However, since kernel 2.0, Linux does not do so: if
+	// oldpath is a symbolic link, then newpath is created as a (hard) link
+	// to the same symbolic link file (i.e., newpath becomes a symbolic
+	// link to the same file that oldpath refers to).
+	resolve := false
+	return 0, nil, linkAt(t, linux.AT_FDCWD, oldAddr, linux.AT_FDCWD, newAddr, resolve, false /* allowEmpty */)
+}
+
+// Linkat implements linux syscall linkat(2).
+func Linkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldDirFD := kdefs.FD(args[0].Int())
+	oldAddr := args[1].Pointer()
+	newDirFD := kdefs.FD(args[2].Int())
+	newAddr := args[3].Pointer()
+
+	// man linkat(2):
+	// By default, linkat(), does not dereference oldpath if it is a
+	// symbolic link (like link(2)). Since Linux 2.6.18, the flag
+	// AT_SYMLINK_FOLLOW can be specified in flags to cause oldpath to be
+	// dereferenced if it is a symbolic link.
+	flags := args[4].Int()
+	resolve := flags&linux.AT_SYMLINK_FOLLOW == linux.AT_SYMLINK_FOLLOW
+	allowEmpty := flags&linux.AT_EMPTY_PATH == linux.AT_EMPTY_PATH
+
+	if allowEmpty && !t.HasCapabilityIn(linux.CAP_DAC_READ_SEARCH, t.UserNamespace().Root()) {
+		return 0, nil, syserror.ENOENT
+	}
+
+	return 0, nil, linkAt(t, oldDirFD, oldAddr, newDirFD, newAddr, resolve, allowEmpty)
+}
+
+func readlinkAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, bufAddr usermem.Addr, size uint) (copied uintptr, err error) {
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, err
+	}
+	if dirPath {
+		return 0, syserror.ENOENT
+	}
+
+	err = fileOpOn(t, dirFD, path, false /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		// Check for Read permission.
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Read: true}); err != nil {
+			return err
+		}
+
+		s, err := d.Inode.Readlink(t)
+		if err == syserror.ENOLINK {
+			return syserror.EINVAL
+		}
+		if err != nil {
+			return err
+		}
+
+		buffer := []byte(s)
+		if uint(len(buffer)) > size {
+			buffer = buffer[:size]
+		}
+
+		n, err := t.CopyOutBytes(bufAddr, buffer)
+
+		// Update frame return value.
+		copied = uintptr(n)
+
+		return err
+	})
+	return copied, err // Return frame value.
+}
+
+// Readlink implements linux syscall readlink(2).
+func Readlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	bufAddr := args[1].Pointer()
+	size := args[2].SizeT()
+
+	n, err := readlinkAt(t, linux.AT_FDCWD, addr, bufAddr, size)
+	return n, nil, err
+}
+
+// Readlinkat implements linux syscall readlinkat(2).
+func Readlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	bufAddr := args[2].Pointer()
+	size := args[3].SizeT()
+
+	n, err := readlinkAt(t, dirFD, addr, bufAddr, size)
+	return n, nil, err
+}
+
+func unlinkAt(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr) error {
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+	if dirPath {
+		return syserror.ENOENT
+	}
+
+	return fileOpAt(t, dirFD, path, func(root *fs.Dirent, d *fs.Dirent, name string) error {
+		if !fs.IsDir(d.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		if err := fs.MayDelete(t, root, d, name); err != nil {
+			return err
+		}
+
+		return d.Remove(t, root, name)
+	})
+}
+
+// Unlink implements linux syscall unlink(2).
+func Unlink(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	return 0, nil, unlinkAt(t, linux.AT_FDCWD, addr)
+}
+
+// Unlinkat implements linux syscall unlinkat(2).
+func Unlinkat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	flags := args[2].Uint()
+	if flags&linux.AT_REMOVEDIR != 0 {
+		return 0, nil, rmdirAt(t, dirFD, addr)
+	}
+	return 0, nil, unlinkAt(t, dirFD, addr)
+}
+
+// Truncate implements linux syscall truncate(2).
+func Truncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	length := args[1].Int64()
+
+	if length < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+	if dirPath {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if uint64(length) >= t.ThreadGroup().Limits().Get(limits.FileSize).Cur {
+		t.SendSignal(&arch.SignalInfo{
+			Signo: int32(syscall.SIGXFSZ),
+			Code:  arch.SignalInfoUser,
+		})
+		return 0, nil, syserror.EFBIG
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		if fs.IsDir(d.Inode.StableAttr) {
+			return syserror.EISDIR
+		}
+		if !fs.IsFile(d.Inode.StableAttr) {
+			return syserror.EINVAL
+		}
+
+		// Reject truncation if the access permissions do not allow truncation.
+		// This is different from the behavior of sys_ftruncate, see below.
+		if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true}); err != nil {
+			return err
+		}
+
+		if err := d.Inode.Truncate(t, d, length); err != nil {
+			return err
+		}
+
+		// File length modified, generate notification.
+		d.InotifyEvent(linux.IN_MODIFY, 0)
+
+		return nil
+	})
+}
+
+// Ftruncate implements linux syscall ftruncate(2).
+func Ftruncate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	length := args[1].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Reject truncation if the file flags do not permit this operation.
+	// This is different from truncate(2) above.
+	if !file.Flags().Write {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Note that this is different from truncate(2) above, where a
+	// directory returns EISDIR.
+	if !fs.IsFile(file.Dirent.Inode.StableAttr) {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if length < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if uint64(length) >= t.ThreadGroup().Limits().Get(limits.FileSize).Cur {
+		t.SendSignal(&arch.SignalInfo{
+			Signo: int32(syscall.SIGXFSZ),
+			Code:  arch.SignalInfoUser,
+		})
+		return 0, nil, syserror.EFBIG
+	}
+
+	if err := file.Dirent.Inode.Truncate(t, file.Dirent, length); err != nil {
+		return 0, nil, err
+	}
+
+	// File length modified, generate notification.
+	file.Dirent.InotifyEvent(linux.IN_MODIFY, 0)
+
+	return 0, nil, nil
+}
+
+// Umask implements linux syscall umask(2).
+func Umask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	mask := args[0].ModeT()
+	mask = t.FSContext().SwapUmask(mask & 0777)
+	return uintptr(mask), nil, nil
+}
+
+// Change ownership of a file.
+//
+// uid and gid may be -1, in which case they will not be changed.
+func chown(t *kernel.Task, d *fs.Dirent, uid auth.UID, gid auth.GID) error {
+	owner := fs.FileOwner{
+		UID: auth.NoID,
+		GID: auth.NoID,
+	}
+
+	uattr, err := d.Inode.UnstableAttr(t)
+	if err != nil {
+		return err
+	}
+	c := t.Credentials()
+	hasCap := d.Inode.CheckCapability(t, linux.CAP_CHOWN)
+	isOwner := uattr.Owner.UID == c.EffectiveKUID
+	if uid.Ok() {
+		kuid := c.UserNamespace.MapToKUID(uid)
+		// Valid UID must be supplied if UID is to be changed.
+		if !kuid.Ok() {
+			return syserror.EINVAL
+		}
+
+		// "Only a privileged process (CAP_CHOWN) may change the owner
+		// of a file." -chown(2)
+		//
+		// Linux also allows chown if you own the file and are
+		// explicitly not changing its UID.
+		isNoop := uattr.Owner.UID == kuid
+		if !(hasCap || (isOwner && isNoop)) {
+			return syserror.EPERM
+		}
+
+		owner.UID = kuid
+	}
+	if gid.Ok() {
+		kgid := c.UserNamespace.MapToKGID(gid)
+		// Valid GID must be supplied if GID is to be changed.
+		if !kgid.Ok() {
+			return syserror.EINVAL
+		}
+
+		// "The owner of a file may change the group of the file to any
+		// group of which that owner is a member. A privileged process
+		// (CAP_CHOWN) may change the group arbitrarily." -chown(2)
+		isNoop := uattr.Owner.GID == kgid
+		isMemberGroup := c.InGroup(kgid)
+		if !(hasCap || (isOwner && (isNoop || isMemberGroup))) {
+			return syserror.EPERM
+		}
+
+		owner.GID = kgid
+	}
+
+	// FIXME: This is racy; the inode's owner may have changed in
+	// the meantime. (Linux holds i_mutex while calling
+	// fs/attr.c:notify_change() => inode_operations::setattr =>
+	// inode_change_ok().)
+	if err := d.Inode.SetOwner(t, d, owner); err != nil {
+		return err
+	}
+
+	// When the owner or group are changed by an unprivileged user,
+	// chown(2) also clears the set-user-ID and set-group-ID bits, but
+	// we do not support them.
+	return nil
+}
+
+func chownAt(t *kernel.Task, fd kdefs.FD, addr usermem.Addr, resolve, allowEmpty bool, uid auth.UID, gid auth.GID) error {
+	path, _, err := copyInPath(t, addr, allowEmpty)
+	if err != nil {
+		return err
+	}
+
+	if path == "" {
+		// Annoying. What's wrong with fchown?
+		file := t.FDMap().GetFile(fd)
+		if file == nil {
+			return syserror.EBADF
+		}
+		defer file.DecRef()
+
+		return chown(t, file.Dirent, uid, gid)
+	}
+
+	return fileOpOn(t, fd, path, resolve, func(root *fs.Dirent, d *fs.Dirent) error {
+		return chown(t, d, uid, gid)
+	})
+}
+
+// Chown implements linux syscall chown(2).
+func Chown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	uid := auth.UID(args[1].Uint())
+	gid := auth.GID(args[2].Uint())
+
+	return 0, nil, chownAt(t, linux.AT_FDCWD, addr, true /* resolve */, false /* allowEmpty */, uid, gid)
+}
+
+// Lchown implements linux syscall lchown(2).
+func Lchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	uid := auth.UID(args[1].Uint())
+	gid := auth.GID(args[2].Uint())
+
+	return 0, nil, chownAt(t, linux.AT_FDCWD, addr, false /* resolve */, false /* allowEmpty */, uid, gid)
+}
+
+// Fchown implements linux syscall fchown(2).
+func Fchown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	uid := auth.UID(args[1].Uint())
+	gid := auth.GID(args[2].Uint())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	return 0, nil, chown(t, file.Dirent, uid, gid)
+}
+
+// Fchownat implements Linux syscall fchownat(2).
+func Fchownat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	uid := auth.UID(args[2].Uint())
+	gid := auth.GID(args[3].Uint())
+	flags := args[4].Int()
+
+	if flags&^(linux.AT_EMPTY_PATH|linux.AT_SYMLINK_NOFOLLOW) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	return 0, nil, chownAt(t, dirFD, addr, flags&linux.AT_SYMLINK_NOFOLLOW == 0, flags&linux.AT_EMPTY_PATH != 0, uid, gid)
+}
+
+func chmod(t *kernel.Task, d *fs.Dirent, mode linux.FileMode) error {
+	// Must own file to change mode.
+	if !d.Inode.CheckOwnership(t) {
+		return syserror.EPERM
+	}
+
+	p := fs.FilePermsFromMode(mode)
+	if !d.Inode.SetPermissions(t, d, p) {
+		return syserror.EPERM
+	}
+
+	// File attribute changed, generate notification.
+	d.InotifyEvent(linux.IN_ATTRIB, 0)
+
+	return nil
+}
+
+func chmodAt(t *kernel.Task, fd kdefs.FD, addr usermem.Addr, mode linux.FileMode) error {
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	return fileOpOn(t, fd, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return chmod(t, d, mode)
+	})
+}
+
+// Chmod implements linux syscall chmod(2).
+func Chmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	mode := linux.FileMode(args[1].ModeT())
+
+	return 0, nil, chmodAt(t, linux.AT_FDCWD, addr, mode)
+}
+
+// Fchmod implements linux syscall fchmod(2).
+func Fchmod(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	mode := linux.FileMode(args[1].ModeT())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	return 0, nil, chmod(t, file.Dirent, mode)
+}
+
+// Fchmodat implements linux syscall fchmodat(2).
+func Fchmodat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	mode := linux.FileMode(args[2].ModeT())
+
+	return 0, nil, chmodAt(t, fd, addr, mode)
+}
+
+// defaultSetToSystemTimeSpec returns a TimeSpec that will set ATime and MTime
+// to the system time.
+func defaultSetToSystemTimeSpec() fs.TimeSpec {
+	return fs.TimeSpec{
+		ATimeSetSystemTime: true,
+		MTimeSetSystemTime: true,
+	}
+}
+
+func utimes(t *kernel.Task, dirFD kdefs.FD, addr usermem.Addr, ts fs.TimeSpec, resolve bool) error {
+	setTimestamp := func(root *fs.Dirent, d *fs.Dirent) error {
+		// Does the task own the file?
+		if !d.Inode.CheckOwnership(t) {
+			// Trying to set a specific time? Must be owner.
+			if (ts.ATimeOmit || !ts.ATimeSetSystemTime) && (ts.MTimeOmit || !ts.MTimeSetSystemTime) {
+				return syserror.EPERM
+			}
+
+			// Trying to set to current system time? Must have write access.
+			if err := d.Inode.CheckPermission(t, fs.PermMask{Write: true}); err != nil {
+				return err
+			}
+		}
+
+		return d.Inode.SetTimestamps(t, d, ts)
+	}
+
+	// From utimes.c:
+	// "If filename is NULL and dfd refers to an open file, then operate on
+	// the file.  Otherwise look up filename, possibly using dfd as a
+	// starting point."
+	if addr == 0 && dirFD != linux.AT_FDCWD {
+		if !resolve {
+			// Linux returns EINVAL in this case. See utimes.c.
+			return syserror.EINVAL
+		}
+		f := t.FDMap().GetFile(dirFD)
+		if f == nil {
+			return syserror.EBADF
+		}
+		defer f.DecRef()
+
+		root := t.FSContext().RootDirectory()
+		defer root.DecRef()
+
+		return setTimestamp(root, f.Dirent)
+	}
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	return fileOpOn(t, dirFD, path, resolve, setTimestamp)
+}
+
+// Utime implements linux syscall utime(2).
+func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	filenameAddr := args[0].Pointer()
+	timesAddr := args[1].Pointer()
+
+	// No timesAddr argument will be interpreted as current system time.
+	ts := defaultSetToSystemTimeSpec()
+	if timesAddr != 0 {
+		var times syscall.Utimbuf
+		if _, err := t.CopyIn(timesAddr, &times); err != nil {
+			return 0, nil, err
+		}
+		ts = fs.TimeSpec{
+			ATime: ktime.FromSeconds(times.Actime),
+			MTime: ktime.FromSeconds(times.Modtime),
+		}
+	}
+	return 0, nil, utimes(t, linux.AT_FDCWD, filenameAddr, ts, true)
+}
+
+// Utimes implements linux syscall utimes(2).
+func Utimes(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	filenameAddr := args[0].Pointer()
+	timesAddr := args[1].Pointer()
+
+	// No timesAddr argument will be interpreted as current system time.
+	ts := defaultSetToSystemTimeSpec()
+	if timesAddr != 0 {
+		var times [2]linux.Timeval
+		if _, err := t.CopyIn(timesAddr, &times); err != nil {
+			return 0, nil, err
+		}
+		ts = fs.TimeSpec{
+			ATime: ktime.FromTimeval(times[0]),
+			MTime: ktime.FromTimeval(times[1]),
+		}
+	}
+	return 0, nil, utimes(t, linux.AT_FDCWD, filenameAddr, ts, true)
+}
+
+// timespecIsValid checks that the timespec is valid for use in utimensat.
+func timespecIsValid(ts linux.Timespec) bool {
+	// Nsec must be UTIME_OMIT, UTIME_NOW, or less than 10^9.
+	return ts.Nsec == linux.UTIME_OMIT || ts.Nsec == linux.UTIME_NOW || ts.Nsec < 1e9
+}
+
+// Utimensat implements linux syscall utimensat(2).
+func Utimensat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	pathnameAddr := args[1].Pointer()
+	timesAddr := args[2].Pointer()
+	flags := args[3].Int()
+
+	// No timesAddr argument will be interpreted as current system time.
+	ts := defaultSetToSystemTimeSpec()
+	if timesAddr != 0 {
+		var times [2]linux.Timespec
+		if _, err := t.CopyIn(timesAddr, &times); err != nil {
+			return 0, nil, err
+		}
+		if !timespecIsValid(times[0]) || !timespecIsValid(times[1]) {
+			return 0, nil, syserror.EINVAL
+		}
+
+		// If both are UTIME_OMIT, this is a noop.
+		if times[0].Nsec == linux.UTIME_OMIT && times[1].Nsec == linux.UTIME_OMIT {
+			return 0, nil, nil
+		}
+
+		ts = fs.TimeSpec{
+			ATime:              ktime.FromTimespec(times[0]),
+			ATimeOmit:          times[0].Nsec == linux.UTIME_OMIT,
+			ATimeSetSystemTime: times[0].Nsec == linux.UTIME_NOW,
+			MTime:              ktime.FromTimespec(times[1]),
+			MTimeOmit:          times[1].Nsec == linux.UTIME_OMIT,
+			MTimeSetSystemTime: times[0].Nsec == linux.UTIME_NOW,
+		}
+	}
+	return 0, nil, utimes(t, dirFD, pathnameAddr, ts, flags&linux.AT_SYMLINK_NOFOLLOW == 0)
+}
+
+// Futimesat implements linux syscall futimesat(2).
+func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	dirFD := kdefs.FD(args[0].Int())
+	pathnameAddr := args[1].Pointer()
+	timesAddr := args[2].Pointer()
+
+	// No timesAddr argument will be interpreted as current system time.
+	ts := defaultSetToSystemTimeSpec()
+	if timesAddr != 0 {
+		var times [2]linux.Timeval
+		if _, err := t.CopyIn(timesAddr, &times); err != nil {
+			return 0, nil, err
+		}
+		if times[0].Usec >= 1e6 || times[0].Usec < 0 ||
+			times[1].Usec >= 1e6 || times[1].Usec < 0 {
+			return 0, nil, syserror.EINVAL
+		}
+
+		ts = fs.TimeSpec{
+			ATime: ktime.FromTimeval(times[0]),
+			MTime: ktime.FromTimeval(times[1]),
+		}
+	}
+	return 0, nil, utimes(t, dirFD, pathnameAddr, ts, true)
+}
+
+func renameAt(t *kernel.Task, oldDirFD kdefs.FD, oldAddr usermem.Addr, newDirFD kdefs.FD, newAddr usermem.Addr) error {
+	newPath, _, err := copyInPath(t, newAddr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+	oldPath, _, err := copyInPath(t, oldAddr, false /* allowEmpty */)
+	if err != nil {
+		return err
+	}
+
+	return fileOpAt(t, oldDirFD, oldPath, func(root *fs.Dirent, oldParent *fs.Dirent, oldName string) error {
+		if !fs.IsDir(oldParent.Inode.StableAttr) {
+			return syserror.ENOTDIR
+		}
+
+		// Root cannot be renamed to anything.
+		//
+		// TODO: This catches the case when the rename
+		// argument is exactly "/", but we should return EBUSY when
+		// renaming any mount point, or when the argument is not
+		// exactly "/" but still resolves to the root, like "/.." or
+		// "/bin/..".
+		if oldParent == root && oldName == "." {
+			return syscall.EBUSY
+		}
+		return fileOpAt(t, newDirFD, newPath, func(root *fs.Dirent, newParent *fs.Dirent, newName string) error {
+			if !fs.IsDir(newParent.Inode.StableAttr) {
+				return syserror.ENOTDIR
+			}
+
+			// Nothing can be renamed to root.
+			//
+			// TODO: Same as above.
+			if newParent == root && newName == "." {
+				return syscall.EBUSY
+			}
+			return fs.Rename(t, root, oldParent, oldName, newParent, newName)
+		})
+	})
+}
+
+// Rename implements linux syscall rename(2).
+func Rename(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldPathAddr := args[0].Pointer()
+	newPathAddr := args[1].Pointer()
+	return 0, nil, renameAt(t, linux.AT_FDCWD, oldPathAddr, linux.AT_FDCWD, newPathAddr)
+}
+
+// Renameat implements linux syscall renameat(2).
+func Renameat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldDirFD := kdefs.FD(args[0].Int())
+	oldPathAddr := args[1].Pointer()
+	newDirFD := kdefs.FD(args[2].Int())
+	newPathAddr := args[3].Pointer()
+	return 0, nil, renameAt(t, oldDirFD, oldPathAddr, newDirFD, newPathAddr)
+}
+
+// Fallocate implements linux system call fallocate(2).
+// (well, not really, but at least we return the expected error codes)
+func Fallocate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	offset := args[2].Int64()
+	length := args[3].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	if offset < 0 || length <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	return 0, nil, syserror.EOPNOTSUPP
+}
+
+// Flock implements linux syscall flock(2).
+func Flock(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	operation := args[1].Int()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		// flock(2): EBADF fd is not an open file descriptor.
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	nonblocking := operation&linux.LOCK_NB != 0
+	operation &^= linux.LOCK_NB
+
+	// flock(2):
+	// Locks created by flock() are associated with an open file table entry. This means that
+	// duplicate file descriptors (created by, for example, fork(2) or dup(2)) refer to the
+	// same lock, and this lock may be modified or released using any of these descriptors. Furthermore,
+	// the lock is released either by an explicit LOCK_UN operation on any of these duplicate
+	// descriptors, or when all such descriptors have been closed.
+	//
+	// If a process uses open(2) (or similar) to obtain more than one descriptor for the same file,
+	// these descriptors are treated independently by flock(). An attempt to lock the file using
+	// one of these file descriptors may be denied by a lock that the calling process has already placed via
+	// another descriptor.
+	//
+	// We use the File UniqueID as the lock UniqueID because it needs to reference the same lock across dup(2)
+	// and fork(2).
+	lockUniqueID := lock.UniqueID(file.UniqueID)
+
+	// A BSD style lock spans the entire file.
+	rng := lock.LockRange{
+		Start: 0,
+		End:   lock.LockEOF,
+	}
+
+	switch operation {
+	case linux.LOCK_EX:
+		if nonblocking {
+			// Since we're nonblocking we pass a nil lock.Blocker implementation.
+			if !file.Dirent.Inode.LockCtx.BSD.LockRegion(lockUniqueID, lock.WriteLock, rng, nil) {
+				return 0, nil, syserror.EWOULDBLOCK
+			}
+		} else {
+			// Because we're blocking we will pass the task to satisfy the lock.Blocker interface.
+			if !file.Dirent.Inode.LockCtx.BSD.LockRegion(lockUniqueID, lock.WriteLock, rng, t) {
+				return 0, nil, syserror.EINTR
+			}
+		}
+	case linux.LOCK_SH:
+		if nonblocking {
+			// Since we're nonblocking we pass a nil lock.Blocker implementation.
+			if !file.Dirent.Inode.LockCtx.BSD.LockRegion(lockUniqueID, lock.ReadLock, rng, nil) {
+				return 0, nil, syserror.EWOULDBLOCK
+			}
+		} else {
+			// Because we're blocking we will pass the task to satisfy the lock.Blocker interface.
+			if !file.Dirent.Inode.LockCtx.BSD.LockRegion(lockUniqueID, lock.ReadLock, rng, t) {
+				return 0, nil, syserror.EINTR
+			}
+		}
+	case linux.LOCK_UN:
+		file.Dirent.Inode.LockCtx.BSD.UnlockRegion(lockUniqueID, rng)
+	default:
+		// flock(2): EINVAL operation is invalid.
+		return 0, nil, syserror.EINVAL
+	}
+
+	return 0, nil, nil
+}
+
+// Sendfile implements linux system call sendfile(2).
+func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	outFD := kdefs.FD(args[0].Int())
+	inFD := kdefs.FD(args[1].Int())
+	offsetAddr := args[2].Pointer()
+	count := int64(args[3].SizeT())
+
+	// Don't send a negative number of bytes.
+	if count < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Get files.
+	outFile := t.FDMap().GetFile(outFD)
+	if outFile == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer outFile.DecRef()
+
+	inFile := t.FDMap().GetFile(inFD)
+	if inFile == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer inFile.DecRef()
+
+	// Verify that the outfile is writable.
+	outFlags := outFile.Flags()
+	if !outFlags.Write {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Verify that the outfile Append flag is not set.
+	if outFlags.Append {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Verify that we have a regular infile.
+	// http://elixir.free-electrons.com/linux/latest/source/fs/splice.c#L933
+	if !fs.IsRegular(inFile.Dirent.Inode.StableAttr) {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Verify that the infile is readable.
+	if !inFile.Flags().Read {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Setup for sending data.
+	var offset uint64
+	var n int64
+	var err error
+	w := &fs.FileWriter{t, outFile}
+	hasOffset := offsetAddr != 0
+	// If we have a provided offset.
+	if hasOffset {
+		// Copy in the offset.
+		if _, err := t.CopyIn(offsetAddr, &offset); err != nil {
+			return 0, nil, err
+		}
+		// Send data using Preadv.
+		r := io.NewSectionReader(&fs.FileReader{t, inFile}, int64(offset), count)
+		n, err = io.Copy(w, r)
+		// Copy out the new offset.
+		if _, err := t.CopyOut(offsetAddr, n+int64(offset)); err != nil {
+			return 0, nil, err
+		}
+		// If we don't have a provided offset.
+	} else {
+		// Send data using readv.
+		r := &io.LimitedReader{R: &fs.FileReader{t, inFile}, N: count}
+		n, err = io.Copy(w, r)
+	}
+
+	// We can only pass a single file to handleIOError, so pick inFile
+	// arbitrarily.
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "sendfile", inFile)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_futex.go b/pkg/sentry/syscalls/linux/sys_futex.go
new file mode 100644
index 000000000..57762d058
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_futex.go
@@ -0,0 +1,319 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// futexChecker is a futex.Checker that uses a Task's MemoryManager.
+type futexChecker struct {
+	t *kernel.Task
+}
+
+// Check checks if the address contains the given value, and returns
+// syserror.EAGAIN if it doesn't. See Checker interface in futex package
+// for more information.
+func (f futexChecker) Check(addr uintptr, val uint32) error {
+	in := f.t.CopyScratchBuffer(4)
+	_, err := f.t.CopyInBytes(usermem.Addr(addr), in)
+	if err != nil {
+		return err
+	}
+	nval := usermem.ByteOrder.Uint32(in)
+	if val != nval {
+		return syserror.EAGAIN
+	}
+	return nil
+}
+
+func (f futexChecker) atomicOp(addr uintptr, op func(uint32) uint32) (uint32, error) {
+	in := f.t.CopyScratchBuffer(4)
+	_, err := f.t.CopyInBytes(usermem.Addr(addr), in)
+	if err != nil {
+		return 0, err
+	}
+	o := usermem.ByteOrder.Uint32(in)
+	mm := f.t.MemoryManager()
+	for {
+		n := op(o)
+		r, err := mm.CompareAndSwapUint32(f.t, usermem.Addr(addr), o, n, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+		if err != nil {
+			return 0, err
+		}
+
+		if r == o {
+			return o, nil
+		}
+		o = r
+	}
+}
+
+// Op performs an operation on addr and returns a result based on the operation.
+func (f futexChecker) Op(addr uintptr, opIn uint32) (bool, error) {
+	op := (opIn >> 28) & 0xf
+	cmp := (opIn >> 24) & 0xf
+	opArg := (opIn >> 12) & 0xfff
+	cmpArg := opIn & 0xfff
+
+	if op&linux.FUTEX_OP_OPARG_SHIFT != 0 {
+		opArg = 1 << opArg
+		op &^= linux.FUTEX_OP_OPARG_SHIFT // clear flag
+	}
+
+	var oldVal uint32
+	var err error
+	switch op {
+	case linux.FUTEX_OP_SET:
+		oldVal, err = f.t.MemoryManager().SwapUint32(f.t, usermem.Addr(addr), opArg, usermem.IOOpts{
+			AddressSpaceActive: true,
+		})
+	case linux.FUTEX_OP_ADD:
+		oldVal, err = f.atomicOp(addr, func(a uint32) uint32 {
+			return a + opArg
+		})
+	case linux.FUTEX_OP_OR:
+		oldVal, err = f.atomicOp(addr, func(a uint32) uint32 {
+			return a | opArg
+		})
+	case linux.FUTEX_OP_ANDN:
+		oldVal, err = f.atomicOp(addr, func(a uint32) uint32 {
+			return a & ^opArg
+		})
+	case linux.FUTEX_OP_XOR:
+		oldVal, err = f.atomicOp(addr, func(a uint32) uint32 {
+			return a ^ opArg
+		})
+	default:
+		return false, syserror.ENOSYS
+	}
+	if err != nil {
+		return false, err
+	}
+
+	switch cmp {
+	case linux.FUTEX_OP_CMP_EQ:
+		return oldVal == cmpArg, nil
+	case linux.FUTEX_OP_CMP_NE:
+		return oldVal != cmpArg, nil
+	case linux.FUTEX_OP_CMP_LT:
+		return oldVal < cmpArg, nil
+	case linux.FUTEX_OP_CMP_LE:
+		return oldVal <= cmpArg, nil
+	case linux.FUTEX_OP_CMP_GT:
+		return oldVal > cmpArg, nil
+	case linux.FUTEX_OP_CMP_GE:
+		return oldVal >= cmpArg, nil
+	default:
+		return false, syserror.ENOSYS
+	}
+}
+
+// futexWaitRestartBlock encapsulates the state required to restart futex(2)
+// via restart_syscall(2).
+type futexWaitRestartBlock struct {
+	duration time.Duration
+
+	// addr stored as uint64 since uintptr is not save-able.
+	addr uint64
+
+	val  uint32
+	mask uint32
+}
+
+// Restart implements kernel.SyscallRestartBlock.Restart.
+func (f *futexWaitRestartBlock) Restart(t *kernel.Task) (uintptr, error) {
+	return futexWaitDuration(t, f.duration, false, uintptr(f.addr), f.val, f.mask)
+}
+
+// futexWaitAbsolute performs a FUTEX_WAIT_BITSET, blocking until the wait is
+// complete.
+//
+// The wait blocks forever if forever is true, otherwise it blocks until ts.
+//
+// If blocking is interrupted, the syscall is restarted with the original
+// arguments.
+func futexWaitAbsolute(t *kernel.Task, clockRealtime bool, ts linux.Timespec, forever bool, addr uintptr, val, mask uint32) (uintptr, error) {
+	w := t.FutexWaiter()
+	err := t.Futex().WaitPrepare(w, futexChecker{t}, addr, val, mask)
+	if err != nil {
+		return 0, err
+	}
+
+	if forever {
+		err = t.Block(w.C)
+	} else if clockRealtime {
+		notifier, tchan := ktime.NewChannelNotifier()
+		timer := ktime.NewTimer(t.Kernel().RealtimeClock(), notifier)
+		timer.Swap(ktime.Setting{
+			Enabled: true,
+			Next:    ktime.FromTimespec(ts),
+		})
+		err = t.BlockWithTimer(w.C, tchan)
+		timer.Destroy()
+	} else {
+		err = t.BlockWithDeadline(w.C, true, ktime.FromTimespec(ts))
+	}
+
+	t.Futex().WaitComplete(w)
+	return 0, syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+}
+
+// futexWaitDuration performs a FUTEX_WAIT, blocking until the wait is
+// complete.
+//
+// The wait blocks forever if forever is true, otherwise is blocks for
+// duration.
+//
+// If blocking is interrupted, forever determines how to restart the
+// syscall. If forever is true, the syscall is restarted with the original
+// arguments. If forever is false, duration is a relative timeout and the
+// syscall is restarted with the remaining timeout.
+func futexWaitDuration(t *kernel.Task, duration time.Duration, forever bool, addr uintptr, val, mask uint32) (uintptr, error) {
+	w := t.FutexWaiter()
+	err := t.Futex().WaitPrepare(w, futexChecker{t}, addr, val, mask)
+	if err != nil {
+		return 0, err
+	}
+
+	remaining, err := t.BlockWithTimeout(w.C, !forever, duration)
+	t.Futex().WaitComplete(w)
+	if err == nil {
+		return 0, nil
+	}
+
+	// The wait was unsuccessful for some reason other than interruption. Simply
+	// forward the error.
+	if err != syserror.ErrInterrupted {
+		return 0, err
+	}
+
+	// The wait was interrupted and we need to restart. Decide how.
+
+	// The wait duration was absolute, restart with the original arguments.
+	if forever {
+		return 0, kernel.ERESTARTSYS
+	}
+
+	// The wait duration was relative, restart with the remaining duration.
+	t.SetSyscallRestartBlock(&futexWaitRestartBlock{
+		duration: remaining,
+		addr:     uint64(addr),
+		val:      val,
+		mask:     mask,
+	})
+	return 0, kernel.ERESTART_RESTARTBLOCK
+}
+
+// Futex implements linux syscall futex(2).
+// It provides a method for a program to wait for a value at a given address to
+// change, and a method to wake up anyone waiting on a particular address.
+func Futex(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	uaddr := args[0].Pointer()
+	futexOp := args[1].Int()
+	val := int(args[2].Int())
+	nreq := int(args[3].Int())
+	timeout := args[3].Pointer()
+	uaddr2 := args[4].Pointer()
+	val3 := args[5].Int()
+
+	addr := uintptr(uaddr)
+	naddr := uintptr(uaddr2)
+	cmd := futexOp &^ (linux.FUTEX_PRIVATE_FLAG | linux.FUTEX_CLOCK_REALTIME)
+	clockRealtime := (futexOp & linux.FUTEX_CLOCK_REALTIME) == linux.FUTEX_CLOCK_REALTIME
+	mask := uint32(val3)
+
+	switch cmd {
+	case linux.FUTEX_WAIT, linux.FUTEX_WAIT_BITSET:
+		// WAIT{_BITSET} wait forever if the timeout isn't passed.
+		forever := timeout == 0
+
+		var timespec linux.Timespec
+		if !forever {
+			var err error
+			timespec, err = copyTimespecIn(t, timeout)
+			if err != nil {
+				return 0, nil, err
+			}
+		}
+
+		switch cmd {
+		case linux.FUTEX_WAIT:
+			// WAIT uses a relative timeout.
+			mask = ^uint32(0)
+			var timeoutDur time.Duration
+			if !forever {
+				timeoutDur = time.Duration(timespec.ToNsecCapped()) * time.Nanosecond
+			}
+			n, err := futexWaitDuration(t, timeoutDur, forever, addr, uint32(val), mask)
+			return n, nil, err
+
+		case linux.FUTEX_WAIT_BITSET:
+			// WAIT_BITSET uses an absolute timeout which is either
+			// CLOCK_MONOTONIC or CLOCK_REALTIME.
+			if mask == 0 {
+				return 0, nil, syserror.EINVAL
+			}
+			n, err := futexWaitAbsolute(t, clockRealtime, timespec, forever, addr, uint32(val), mask)
+			return n, nil, err
+		default:
+			panic("unreachable")
+		}
+
+	case linux.FUTEX_WAKE:
+		mask = ^uint32(0)
+		fallthrough
+
+	case linux.FUTEX_WAKE_BITSET:
+		if mask == 0 {
+			return 0, nil, syserror.EINVAL
+		}
+		n, err := t.Futex().Wake(addr, mask, val)
+		return uintptr(n), nil, err
+
+	case linux.FUTEX_REQUEUE:
+		n, err := t.Futex().Requeue(addr, naddr, val, nreq)
+		return uintptr(n), nil, err
+
+	case linux.FUTEX_CMP_REQUEUE:
+		// 'val3' contains the value to be checked at 'addr' and
+		// 'val' is the number of waiters that should be woken up.
+		nval := uint32(val3)
+		n, err := t.Futex().RequeueCmp(futexChecker{t}, addr, nval, naddr, val, nreq)
+		return uintptr(n), nil, err
+
+	case linux.FUTEX_WAKE_OP:
+		op := uint32(val3)
+		n, err := t.Futex().WakeOp(futexChecker{t}, addr, naddr, val, nreq, op)
+		return uintptr(n), nil, err
+
+	case linux.FUTEX_LOCK_PI, linux.FUTEX_UNLOCK_PI, linux.FUTEX_TRYLOCK_PI, linux.FUTEX_WAIT_REQUEUE_PI, linux.FUTEX_CMP_REQUEUE_PI:
+		// We don't support any priority inversion futexes.
+		return 0, nil, syserror.ENOSYS
+
+	default:
+		// We don't even know about this command.
+		return 0, nil, syserror.ENOSYS
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/sys_getdents.go b/pkg/sentry/syscalls/linux/sys_getdents.go
new file mode 100644
index 000000000..178714b07
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_getdents.go
@@ -0,0 +1,269 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"bytes"
+	"io"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Getdents implements linux syscall getdents(2) for 64bit systems.
+func Getdents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := int(args[2].Uint())
+
+	minSize := int(smallestDirent(t.Arch()))
+	if size < minSize {
+		// size is smaller than smallest possible dirent.
+		return 0, nil, syserror.EINVAL
+	}
+
+	n, err := getdents(t, fd, addr, size, (*dirent).Serialize)
+	return n, nil, err
+}
+
+// Getdents64 implements linux syscall getdents64(2).
+func Getdents64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := int(args[2].Uint())
+
+	minSize := int(smallestDirent64(t.Arch()))
+	if size < minSize {
+		// size is smaller than smallest possible dirent.
+		return 0, nil, syserror.EINVAL
+	}
+
+	n, err := getdents(t, fd, addr, size, (*dirent).Serialize64)
+	return n, nil, err
+}
+
+// getdents implements the core of getdents(2)/getdents64(2).
+// f is the syscall implementation dirent serialization function.
+func getdents(t *kernel.Task, fd kdefs.FD, addr usermem.Addr, size int, f func(*dirent, io.Writer) (int, error)) (uintptr, error) {
+	dir := t.FDMap().GetFile(fd)
+	if dir == nil {
+		return 0, syserror.EBADF
+	}
+	defer dir.DecRef()
+
+	w := &usermem.IOReadWriter{
+		Ctx:  t,
+		IO:   t.MemoryManager(),
+		Addr: addr,
+		Opts: usermem.IOOpts{
+			AddressSpaceActive: true,
+		},
+	}
+
+	ds := newDirentSerializer(f, w, t.Arch(), size)
+	rerr := dir.Readdir(t, ds)
+
+	switch err := handleIOError(t, ds.Written() > 0, rerr, kernel.ERESTARTSYS, "getdents", dir); err {
+	case nil:
+		dir.Dirent.InotifyEvent(syscall.IN_ACCESS, 0)
+		return uintptr(ds.Written()), nil
+	case io.EOF:
+		return 0, nil
+	default:
+		return 0, err
+	}
+}
+
+// oldDirentHdr is a fixed sized header matching the fixed size
+// fields found in the old linux dirent struct.
+type oldDirentHdr struct {
+	Ino    uint64
+	Off    uint64
+	Reclen uint16
+}
+
+// direntHdr is a fixed sized header matching the fixed size
+// fields found in the new linux dirent struct.
+type direntHdr struct {
+	OldHdr oldDirentHdr
+	Typ    uint8
+}
+
+// dirent contains the data pointed to by a new linux dirent struct.
+type dirent struct {
+	Hdr  direntHdr
+	Name []byte
+}
+
+// newDirent returns a dirent from an fs.InodeOperationsInfo.
+func newDirent(width uint, name string, attr fs.DentAttr, offset uint64) *dirent {
+	d := &dirent{
+		Hdr: direntHdr{
+			OldHdr: oldDirentHdr{
+				Ino: attr.InodeID,
+				Off: offset,
+			},
+			Typ: toType(attr.Type),
+		},
+		Name: []byte(name),
+	}
+	d.Hdr.OldHdr.Reclen = d.padRec(int(width))
+	return d
+}
+
+// smallestDirent returns the size of the smallest possible dirent using
+// the old linux dirent format.
+func smallestDirent(a arch.Context) uint {
+	d := dirent{}
+	return uint(binary.Size(d.Hdr.OldHdr)) + a.Width() + 1
+}
+
+// smallestDirent64 returns the size of the smallest possible dirent using
+// the new linux dirent format.
+func smallestDirent64(a arch.Context) uint {
+	d := dirent{}
+	return uint(binary.Size(d.Hdr)) + a.Width()
+}
+
+// toType converts an fs.InodeOperationsInfo to a linux dirent typ field.
+func toType(nodeType fs.InodeType) uint8 {
+	switch nodeType {
+	case fs.RegularFile, fs.SpecialFile:
+		return syscall.DT_REG
+	case fs.Symlink:
+		return syscall.DT_LNK
+	case fs.Directory:
+		return syscall.DT_DIR
+	case fs.Pipe:
+		return syscall.DT_FIFO
+	case fs.CharacterDevice:
+		return syscall.DT_CHR
+	case fs.BlockDevice:
+		return syscall.DT_BLK
+	case fs.Socket:
+		return syscall.DT_SOCK
+	default:
+		return syscall.DT_UNKNOWN
+	}
+}
+
+// padRec pads the name field until the rec length is a multiple of the width,
+// which must be a power of 2. It returns the padded rec length.
+func (d *dirent) padRec(width int) uint16 {
+	a := int(binary.Size(d.Hdr)) + len(d.Name)
+	r := (a + width) &^ (width - 1)
+	padding := r - a
+	d.Name = append(d.Name, make([]byte, padding)...)
+	return uint16(r)
+}
+
+// Serialize64 serializes a Dirent struct to a byte slice, keeping the new
+// linux dirent format. Returns the number of bytes serialized or an error.
+func (d *dirent) Serialize64(w io.Writer) (int, error) {
+	n1, err := w.Write(binary.Marshal(nil, usermem.ByteOrder, d.Hdr))
+	if err != nil {
+		return 0, err
+	}
+	n2, err := w.Write(d.Name)
+	if err != nil {
+		return 0, err
+	}
+	return n1 + n2, nil
+}
+
+// Serialize serializes a Dirent struct to a byte slice, using the old linux
+// dirent format.
+// Returns the number of bytes serialized or an error.
+func (d *dirent) Serialize(w io.Writer) (int, error) {
+	n1, err := w.Write(binary.Marshal(nil, usermem.ByteOrder, d.Hdr.OldHdr))
+	if err != nil {
+		return 0, err
+	}
+	n2, err := w.Write(d.Name)
+	if err != nil {
+		return 0, err
+	}
+	n3, err := w.Write([]byte{d.Hdr.Typ})
+	if err != nil {
+		return 0, err
+	}
+	return n1 + n2 + n3, nil
+}
+
+// direntSerializer implements fs.InodeOperationsInfoSerializer, serializing dirents to an
+// io.Writer.
+type direntSerializer struct {
+	serialize func(*dirent, io.Writer) (int, error)
+	w         io.Writer
+	// width is the arch native value width.
+	width uint
+	// offset is the current dirent offset.
+	offset uint64
+	// written is the total bytes serialized.
+	written int
+	// size is the size of the buffer to serialize into.
+	size int
+}
+
+func newDirentSerializer(f func(d *dirent, w io.Writer) (int, error), w io.Writer, ac arch.Context, size int) *direntSerializer {
+	return &direntSerializer{
+		serialize: f,
+		w:         w,
+		width:     ac.Width(),
+		size:      size,
+	}
+}
+
+// CopyOut implements fs.InodeOperationsInfoSerializer.CopyOut.
+// It serializes and writes the fs.DentAttr to the direntSerializer io.Writer.
+func (ds *direntSerializer) CopyOut(name string, attr fs.DentAttr) error {
+	ds.offset++
+
+	d := newDirent(ds.width, name, attr, ds.offset)
+
+	// Serialize dirent into a temp buffer.
+	var b bytes.Buffer
+	n, err := ds.serialize(d, &b)
+	if err != nil {
+		ds.offset--
+		return err
+	}
+
+	// Check that we have enough room remaining to write the dirent.
+	if n > (ds.size - ds.written) {
+		ds.offset--
+		return io.EOF
+	}
+
+	// Write out the temp buffer.
+	if _, err := b.WriteTo(ds.w); err != nil {
+		ds.offset--
+		return err
+	}
+
+	ds.written += n
+	return nil
+}
+
+// Written returns the total number of bytes written.
+func (ds *direntSerializer) Written() int {
+	return ds.written
+}
diff --git a/pkg/sentry/syscalls/linux/sys_identity.go b/pkg/sentry/syscalls/linux/sys_identity.go
new file mode 100644
index 000000000..4fd0ed794
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_identity.go
@@ -0,0 +1,180 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	// As NGROUPS_MAX in include/uapi/linux/limits.h.
+	maxNGroups = 65536
+)
+
+// Getuid implements the Linux syscall getuid.
+func Getuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	c := t.Credentials()
+	ruid := c.RealKUID.In(c.UserNamespace).OrOverflow()
+	return uintptr(ruid), nil, nil
+}
+
+// Geteuid implements the Linux syscall geteuid.
+func Geteuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	c := t.Credentials()
+	euid := c.EffectiveKUID.In(c.UserNamespace).OrOverflow()
+	return uintptr(euid), nil, nil
+}
+
+// Getresuid implements the Linux syscall getresuid.
+func Getresuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	ruidAddr := args[0].Pointer()
+	euidAddr := args[1].Pointer()
+	suidAddr := args[2].Pointer()
+	c := t.Credentials()
+	ruid := c.RealKUID.In(c.UserNamespace).OrOverflow()
+	euid := c.EffectiveKUID.In(c.UserNamespace).OrOverflow()
+	suid := c.SavedKUID.In(c.UserNamespace).OrOverflow()
+	if _, err := t.CopyOut(ruidAddr, ruid); err != nil {
+		return 0, nil, err
+	}
+	if _, err := t.CopyOut(euidAddr, euid); err != nil {
+		return 0, nil, err
+	}
+	if _, err := t.CopyOut(suidAddr, suid); err != nil {
+		return 0, nil, err
+	}
+	return 0, nil, nil
+}
+
+// Getgid implements the Linux syscall getgid.
+func Getgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	c := t.Credentials()
+	rgid := c.RealKGID.In(c.UserNamespace).OrOverflow()
+	return uintptr(rgid), nil, nil
+}
+
+// Getegid implements the Linux syscall getegid.
+func Getegid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	c := t.Credentials()
+	egid := c.EffectiveKGID.In(c.UserNamespace).OrOverflow()
+	return uintptr(egid), nil, nil
+}
+
+// Getresgid implements the Linux syscall getresgid.
+func Getresgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	rgidAddr := args[0].Pointer()
+	egidAddr := args[1].Pointer()
+	sgidAddr := args[2].Pointer()
+	c := t.Credentials()
+	rgid := c.RealKGID.In(c.UserNamespace).OrOverflow()
+	egid := c.EffectiveKGID.In(c.UserNamespace).OrOverflow()
+	sgid := c.SavedKGID.In(c.UserNamespace).OrOverflow()
+	if _, err := t.CopyOut(rgidAddr, rgid); err != nil {
+		return 0, nil, err
+	}
+	if _, err := t.CopyOut(egidAddr, egid); err != nil {
+		return 0, nil, err
+	}
+	if _, err := t.CopyOut(sgidAddr, sgid); err != nil {
+		return 0, nil, err
+	}
+	return 0, nil, nil
+}
+
+// Setuid implements the Linux syscall setuid.
+func Setuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	uid := auth.UID(args[0].Int())
+	return 0, nil, t.SetUID(uid)
+}
+
+// Setreuid implements the Linux syscall setreuid.
+func Setreuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	ruid := auth.UID(args[0].Int())
+	euid := auth.UID(args[1].Int())
+	return 0, nil, t.SetREUID(ruid, euid)
+}
+
+// Setresuid implements the Linux syscall setreuid.
+func Setresuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	ruid := auth.UID(args[0].Int())
+	euid := auth.UID(args[1].Int())
+	suid := auth.UID(args[2].Int())
+	return 0, nil, t.SetRESUID(ruid, euid, suid)
+}
+
+// Setgid implements the Linux syscall setgid.
+func Setgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	gid := auth.GID(args[0].Int())
+	return 0, nil, t.SetGID(gid)
+}
+
+// Setregid implements the Linux syscall setregid.
+func Setregid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	rgid := auth.GID(args[0].Int())
+	egid := auth.GID(args[1].Int())
+	return 0, nil, t.SetREGID(rgid, egid)
+}
+
+// Setresgid implements the Linux syscall setregid.
+func Setresgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	rgid := auth.GID(args[0].Int())
+	egid := auth.GID(args[1].Int())
+	sgid := auth.GID(args[2].Int())
+	return 0, nil, t.SetRESGID(rgid, egid, sgid)
+}
+
+// Getgroups implements the Linux syscall getgroups.
+func Getgroups(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	size := int(args[0].Int())
+	if size < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	kgids := t.Credentials().ExtraKGIDs
+	// "If size is zero, list is not modified, but the total number of
+	// supplementary group IDs for the process is returned." - getgroups(2)
+	if size == 0 {
+		return uintptr(len(kgids)), nil, nil
+	}
+	if size < len(kgids) {
+		return 0, nil, syserror.EINVAL
+	}
+	gids := make([]auth.GID, len(kgids))
+	for i, kgid := range kgids {
+		gids[i] = kgid.In(t.UserNamespace()).OrOverflow()
+	}
+	if _, err := t.CopyOut(args[1].Pointer(), gids); err != nil {
+		return 0, nil, err
+	}
+	return uintptr(len(gids)), nil, nil
+}
+
+// Setgroups implements the Linux syscall setgroups.
+func Setgroups(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	size := args[0].Int()
+	if size < 0 || size > maxNGroups {
+		return 0, nil, syserror.EINVAL
+	}
+	if size == 0 {
+		return 0, nil, t.SetExtraGIDs(nil)
+	}
+	gids := make([]auth.GID, size)
+	if _, err := t.CopyIn(args[1].Pointer(), &gids); err != nil {
+		return 0, nil, err
+	}
+	return 0, nil, t.SetExtraGIDs(gids)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_inotify.go b/pkg/sentry/syscalls/linux/sys_inotify.go
new file mode 100644
index 000000000..725204dff
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_inotify.go
@@ -0,0 +1,135 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/anon"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+)
+
+const allFlags = int(linux.IN_NONBLOCK | linux.IN_CLOEXEC)
+
+// InotifyInit1 implements the inotify_init1() syscalls.
+func InotifyInit1(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	flags := int(args[0].Int())
+
+	if flags&^allFlags != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	dirent := fs.NewDirent(anon.NewInode(t), "inotify")
+	fileFlags := fs.FileFlags{
+		Read:        true,
+		Write:       true,
+		NonBlocking: flags&linux.IN_NONBLOCK != 0,
+	}
+	n := fs.NewFile(t, dirent, fileFlags, fs.NewInotify(t))
+	defer n.DecRef()
+
+	fd, err := t.FDMap().NewFDFrom(0, n, kernel.FDFlags{
+		CloseOnExec: flags&linux.IN_CLOEXEC != 0,
+	}, t.ThreadGroup().Limits())
+
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// InotifyInit implements the inotify_init() syscalls.
+func InotifyInit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	args[0].Value = 0
+	return InotifyInit1(t, args)
+}
+
+// fdToInotify resolves an fd to an inotify object. If successful, the file will
+// have an extra ref and the caller is responsible for releasing the ref.
+func fdToInotify(t *kernel.Task, fd kdefs.FD) (*fs.Inotify, *fs.File, error) {
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		// Invalid fd.
+		return nil, nil, syscall.EBADF
+	}
+
+	ino, ok := file.FileOperations.(*fs.Inotify)
+	if !ok {
+		// Not an inotify fd.
+		file.DecRef()
+		return nil, nil, syscall.EINVAL
+	}
+
+	return ino, file, nil
+}
+
+// InotifyAddWatch implements the inotify_add_watch() syscall.
+func InotifyAddWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	mask := args[2].Uint()
+
+	// "IN_DONT_FOLLOW: Don't dereference pathname if it is a symbolic link."
+	//  -- inotify(7)
+	resolve := mask&linux.IN_DONT_FOLLOW == 0
+
+	// "EINVAL: The given event mask contains no valid events."
+	// -- inotify_add_watch(2)
+	if validBits := mask & linux.ALL_INOTIFY_BITS; validBits == 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	ino, file, err := fdToInotify(t, fd)
+	if err != nil {
+		return 0, nil, err
+	}
+	defer file.DecRef()
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	err = fileOpOn(t, linux.AT_FDCWD, path, resolve, func(root *fs.Dirent, dirent *fs.Dirent) error {
+		// "IN_ONLYDIR: Only watch pathname if it is a directory." -- inotify(7)
+		if onlyDir := mask&linux.IN_ONLYDIR != 0; onlyDir && !fs.IsDir(dirent.Inode.StableAttr) {
+			return syscall.ENOTDIR
+		}
+
+		// Copy out to the return frame.
+		fd = kdefs.FD(ino.AddWatch(dirent, mask))
+
+		return nil
+	})
+	return uintptr(fd), nil, err // Return from the existing value.
+}
+
+// InotifyRmWatch implements the inotify_rm_watch() syscall.
+func InotifyRmWatch(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	wd := args[1].Int()
+
+	ino, file, err := fdToInotify(t, fd)
+	if err != nil {
+		return 0, nil, err
+	}
+	defer file.DecRef()
+	return 0, nil, ino.RmWatch(wd)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_lseek.go b/pkg/sentry/syscalls/linux/sys_lseek.go
new file mode 100644
index 000000000..97b51ba7c
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_lseek.go
@@ -0,0 +1,55 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Lseek implements linux syscall lseek(2).
+func Lseek(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	offset := args[1].Int64()
+	whence := args[2].Int()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	var sw fs.SeekWhence
+	switch whence {
+	case 0:
+		sw = fs.SeekSet
+	case 1:
+		sw = fs.SeekCurrent
+	case 2:
+		sw = fs.SeekEnd
+	default:
+		return 0, nil, syserror.EINVAL
+	}
+
+	offset, serr := file.Seek(t, sw, offset)
+	err := handleIOError(t, false /* partialResult */, serr, kernel.ERESTARTSYS, "lseek", file)
+	if err != nil {
+		return 0, nil, err
+	}
+	return uintptr(offset), nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go
new file mode 100644
index 000000000..2c7d41de0
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_mmap.go
@@ -0,0 +1,435 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"bytes"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/mm"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Brk implements linux syscall brk(2).
+func Brk(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr, _ := t.MemoryManager().Brk(t, args[0].Pointer())
+	// "However, the actual Linux system call returns the new program break on
+	// success. On failure, the system call returns the current break." -
+	// brk(2)
+	return uintptr(addr), nil, nil
+}
+
+// Mmap implements linux syscall mmap(2).
+func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	prot := args[2].Int()
+	flags := args[3].Int()
+	fd := kdefs.FD(args[4].Int())
+	fixed := flags&linux.MAP_FIXED != 0
+	private := flags&linux.MAP_PRIVATE != 0
+	shared := flags&linux.MAP_SHARED != 0
+	anon := flags&linux.MAP_ANONYMOUS != 0
+
+	// Require exactly one of MAP_PRIVATE and MAP_SHARED.
+	if private == shared {
+		return 0, nil, syserror.EINVAL
+	}
+
+	opts := memmap.MMapOpts{
+		Length:  args[1].Uint64(),
+		Offset:  args[5].Uint64(),
+		Addr:    args[0].Pointer(),
+		Fixed:   fixed,
+		Unmap:   fixed,
+		Private: private,
+		Perms: usermem.AccessType{
+			Read:    linux.PROT_READ&prot != 0,
+			Write:   linux.PROT_WRITE&prot != 0,
+			Execute: linux.PROT_EXEC&prot != 0,
+		},
+		MaxPerms:  usermem.AnyAccess,
+		GrowsDown: linux.MAP_GROWSDOWN&flags != 0,
+		Precommit: linux.MAP_POPULATE&flags != 0,
+	}
+	defer func() {
+		if opts.MappingIdentity != nil {
+			opts.MappingIdentity.DecRef()
+		}
+	}()
+
+	if !anon {
+		// Convert the passed FD to a file reference.
+		file := t.FDMap().GetFile(fd)
+		if file == nil {
+			return 0, nil, syserror.EBADF
+		}
+		defer file.DecRef()
+
+		flags := file.Flags()
+		// mmap unconditionally requires that the FD is readable.
+		if !flags.Read {
+			return 0, nil, syserror.EACCES
+		}
+		// MAP_SHARED requires that the FD be writable for PROT_WRITE.
+		if shared && !flags.Write {
+			opts.MaxPerms.Write = false
+		}
+
+		if err := file.ConfigureMMap(t, &opts); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	rv, err := t.MemoryManager().MMap(t, opts)
+	return uintptr(rv), nil, err
+}
+
+// Munmap implements linux syscall munmap(2).
+func Munmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, t.MemoryManager().MUnmap(t, args[0].Pointer(), args[1].Uint64())
+}
+
+// Mremap implements linux syscall mremap(2).
+func Mremap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	oldAddr := args[0].Pointer()
+	oldSize := args[1].Uint64()
+	newSize := args[2].Uint64()
+	flags := args[3].Uint64()
+	newAddr := args[4].Pointer()
+
+	if flags&^(linux.MREMAP_MAYMOVE|linux.MREMAP_FIXED) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	mayMove := flags&linux.MREMAP_MAYMOVE != 0
+	fixed := flags&linux.MREMAP_FIXED != 0
+	var moveMode mm.MRemapMoveMode
+	switch {
+	case !mayMove && !fixed:
+		moveMode = mm.MRemapNoMove
+	case mayMove && !fixed:
+		moveMode = mm.MRemapMayMove
+	case mayMove && fixed:
+		moveMode = mm.MRemapMustMove
+	case !mayMove && fixed:
+		// "If MREMAP_FIXED is specified, then MREMAP_MAYMOVE must also be
+		// specified." - mremap(2)
+		return 0, nil, syserror.EINVAL
+	}
+
+	rv, err := t.MemoryManager().MRemap(t, oldAddr, oldSize, newSize, mm.MRemapOpts{
+		Move:    moveMode,
+		NewAddr: newAddr,
+	})
+	return uintptr(rv), nil, err
+}
+
+// Mprotect implements linux syscall mprotect(2).
+func Mprotect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	length := args[1].Uint64()
+	prot := args[2].Int()
+	err := t.MemoryManager().MProtect(args[0].Pointer(), length, usermem.AccessType{
+		Read:    linux.PROT_READ&prot != 0,
+		Write:   linux.PROT_WRITE&prot != 0,
+		Execute: linux.PROT_EXEC&prot != 0,
+	}, linux.PROT_GROWSDOWN&prot != 0)
+	return 0, nil, err
+}
+
+// Madvise implements linux syscall madvise(2).
+func Madvise(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	length := uint64(args[1].SizeT())
+	adv := args[2].Int()
+
+	// "The Linux implementation requires that the address addr be
+	// page-aligned, and allows length to be zero." - madvise(2)
+	if addr.RoundDown() != addr {
+		return 0, nil, syserror.EINVAL
+	}
+	if length == 0 {
+		return 0, nil, nil
+	}
+	// Not explicitly stated: length need not be page-aligned.
+	lenAddr, ok := usermem.Addr(length).RoundUp()
+	if !ok {
+		return 0, nil, syserror.EINVAL
+	}
+	length = uint64(lenAddr)
+
+	switch adv {
+	case linux.MADV_DONTNEED:
+		return 0, nil, t.MemoryManager().Decommit(addr, length)
+	case linux.MADV_HUGEPAGE, linux.MADV_NOHUGEPAGE:
+		fallthrough
+	case linux.MADV_MERGEABLE, linux.MADV_UNMERGEABLE:
+		fallthrough
+	case linux.MADV_NORMAL, linux.MADV_RANDOM, linux.MADV_SEQUENTIAL, linux.MADV_WILLNEED:
+		// Do nothing, we totally ignore the suggestions above.
+		return 0, nil, nil
+	case linux.MADV_REMOVE, linux.MADV_DOFORK, linux.MADV_DONTFORK:
+		// These "suggestions" have application-visible side effects, so we
+		// have to indicate that we don't support them.
+		return 0, nil, syserror.ENOSYS
+	case linux.MADV_HWPOISON:
+		// Only privileged processes are allowed to poison pages.
+		return 0, nil, syserror.EPERM
+	default:
+		// If adv is not a valid value tell the caller.
+		return 0, nil, syserror.EINVAL
+	}
+}
+
+func copyOutIfNotNull(t *kernel.Task, ptr usermem.Addr, val interface{}) (int, error) {
+	if ptr != 0 {
+		return t.CopyOut(ptr, val)
+	}
+	return 0, nil
+}
+
+// GetMempolicy implements the syscall get_mempolicy(2).
+func GetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	mode := args[0].Pointer()
+	nodemask := args[1].Pointer()
+	maxnode := args[2].Uint()
+	addr := args[3].Pointer()
+	flags := args[4].Uint()
+
+	memsAllowed := flags&linux.MPOL_F_MEMS_ALLOWED != 0
+	nodeFlag := flags&linux.MPOL_F_NODE != 0
+	addrFlag := flags&linux.MPOL_F_ADDR != 0
+
+	// TODO: Once sysfs is implemented, report a single numa node in
+	// /sys/devices/system/node.
+	if nodemask != 0 && maxnode < 1 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// 'addr' provided iff 'addrFlag' set.
+	if addrFlag == (addr == 0) {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Default policy for the thread.
+	if flags == 0 {
+		policy, nodemaskVal := t.NumaPolicy()
+		if _, err := copyOutIfNotNull(t, mode, policy); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+		if _, err := copyOutIfNotNull(t, nodemask, nodemaskVal); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+		return 0, nil, nil
+	}
+
+	// Report all nodes available to caller.
+	if memsAllowed {
+		// MPOL_F_NODE and MPOL_F_ADDR not allowed with MPOL_F_MEMS_ALLOWED.
+		if nodeFlag || addrFlag {
+			return 0, nil, syserror.EINVAL
+		}
+
+		// Report a single numa node.
+		if _, err := copyOutIfNotNull(t, nodemask, uint32(0x1)); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+		return 0, nil, nil
+	}
+
+	if addrFlag {
+		if nodeFlag {
+			// Return the id for the node where 'addr' resides, via 'mode'.
+			//
+			// The real get_mempolicy(2) allocates the page referenced by 'addr'
+			// by simulating a read, if it is unallocated before the call. It
+			// then returns the node the page is allocated on through the mode
+			// pointer.
+			b := t.CopyScratchBuffer(1)
+			_, err := t.CopyInBytes(addr, b)
+			if err != nil {
+				return 0, nil, syserror.EFAULT
+			}
+			if _, err := copyOutIfNotNull(t, mode, int32(0)); err != nil {
+				return 0, nil, syserror.EFAULT
+			}
+		} else {
+			storedPolicy, _ := t.NumaPolicy()
+			// Return the policy governing the memory referenced by 'addr'.
+			if _, err := copyOutIfNotNull(t, mode, int32(storedPolicy)); err != nil {
+				return 0, nil, syserror.EFAULT
+			}
+		}
+		return 0, nil, nil
+	}
+
+	storedPolicy, _ := t.NumaPolicy()
+	if nodeFlag && (storedPolicy&^linux.MPOL_MODE_FLAGS == linux.MPOL_INTERLEAVE) {
+		// Policy for current thread is to interleave memory between
+		// nodes. Return the next node we'll allocate on. Since we only have a
+		// single node, this is always node 0.
+		if _, err := copyOutIfNotNull(t, mode, int32(0)); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+		return 0, nil, nil
+	}
+
+	return 0, nil, syserror.EINVAL
+}
+
+func allowedNodesMask() uint32 {
+	const maxNodes = 1
+	return ^uint32((1 << maxNodes) - 1)
+}
+
+// SetMempolicy implements the syscall set_mempolicy(2).
+func SetMempolicy(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	modeWithFlags := args[0].Int()
+	nodemask := args[1].Pointer()
+	maxnode := args[2].Uint()
+
+	if maxnode < 1 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if modeWithFlags&linux.MPOL_MODE_FLAGS == linux.MPOL_MODE_FLAGS {
+		// Can't specify multiple modes simultaneously. Must also contain a
+		// valid mode, which we check below.
+		return 0, nil, syserror.EINVAL
+	}
+
+	mode := modeWithFlags &^ linux.MPOL_MODE_FLAGS
+	if mode < 0 || mode >= linux.MPOL_MAX {
+		return 0, nil, syserror.EINVAL
+	}
+
+	var nodemaskVal uint32
+	if _, err := t.CopyIn(nodemask, &nodemaskVal); err != nil {
+		return 0, nil, syserror.EFAULT
+	}
+
+	// When setting MPOL_INTERLEAVE, nodemask must not be empty.
+	if mode == linux.MPOL_INTERLEAVE && nodemaskVal == 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if nodemaskVal&allowedNodesMask() != 0 {
+		// Invalid node specified.
+		return 0, nil, syserror.EINVAL
+	}
+
+	t.SetNumaPolicy(int32(modeWithFlags), nodemaskVal)
+
+	return 0, nil, nil
+}
+
+// Mincore implements the syscall mincore(2).
+func Mincore(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	length := args[1].SizeT()
+	vec := args[2].Pointer()
+
+	if addr != addr.RoundDown() {
+		return 0, nil, syserror.EINVAL
+	}
+	// "The length argument need not be a multiple of the page size, but since
+	// residency information is returned for whole pages, length is effectively
+	// rounded up to the next multiple of the page size." - mincore(2)
+	la, ok := usermem.Addr(length).RoundUp()
+	if !ok {
+		return 0, nil, syserror.ENOMEM
+	}
+	ar, ok := addr.ToRange(uint64(la))
+	if !ok {
+		return 0, nil, syserror.ENOMEM
+	}
+
+	// Pretend that all mapped pages are "resident in core".
+	mapped := t.MemoryManager().VirtualMemorySizeRange(ar)
+	// "ENOMEM: addr to addr + length contained unmapped memory."
+	if mapped != uint64(la) {
+		return 0, nil, syserror.ENOMEM
+	}
+	resident := bytes.Repeat([]byte{1}, int(mapped/usermem.PageSize))
+	_, err := t.CopyOut(vec, resident)
+	return 0, nil, err
+}
+
+// Msync implements Linux syscall msync(2).
+func Msync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	length := args[1].SizeT()
+	flags := args[2].Int()
+
+	if addr != addr.RoundDown() {
+		return 0, nil, syserror.EINVAL
+	}
+	if length == 0 {
+		return 0, nil, nil
+	}
+	la, ok := usermem.Addr(length).RoundUp()
+	if !ok {
+		return 0, nil, syserror.ENOMEM
+	}
+	// "The flags argument should specify exactly one of MS_ASYNC and MS_SYNC,
+	// and may additionally include the MS_INVALIDATE bit. ... However, Linux
+	// permits a call to msync() that specifies neither of these flags, with
+	// semantics that are (currently) equivalent to specifying MS_ASYNC." -
+	// msync(2)
+	if flags&^(linux.MS_ASYNC|linux.MS_SYNC|linux.MS_INVALIDATE) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	sync := flags&linux.MS_SYNC != 0
+	if sync && flags&linux.MS_ASYNC != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// MS_INVALIDATE "asks to invalidate other mappings of the same file (so
+	// that they can be updated with the fresh values just written)". This is a
+	// no-op given that shared memory exists. However, MS_INVALIDATE can also
+	// be used to detect mlocks: "EBUSY: MS_INVALIDATE was specified in flags,
+	// and a memory lock exists for the specified address range." Given that
+	// mlock is stubbed out, it's unsafe to pass MS_INVALIDATE silently since
+	// some user program could be using it for synchronization.
+	if flags&linux.MS_INVALIDATE != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	// MS_SYNC "requests an update and waits for it to complete."
+	if sync {
+		err := t.MemoryManager().Sync(t, addr, uint64(la))
+		// Sync calls fsync, the same interrupt conversion rules apply, see
+		// mm/msync.c, fsync POSIX.1-2008.
+		return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+	}
+	// MS_ASYNC "specifies that an update be scheduled, but the call returns
+	// immediately". As long as dirty pages are tracked and eventually written
+	// back, this is a no-op. (Correspondingly: "Since Linux 2.6.19, MS_ASYNC
+	// is in fact a no-op, since the kernel properly tracks dirty pages and
+	// flushes them to storage as necessary.")
+	//
+	// However: "ENOMEM: The indicated memory (or part of it) was not mapped."
+	// This applies even for MS_ASYNC.
+	ar, ok := addr.ToRange(uint64(la))
+	if !ok {
+		return 0, nil, syserror.ENOMEM
+	}
+	mapped := t.MemoryManager().VirtualMemorySizeRange(ar)
+	if mapped != uint64(la) {
+		return 0, nil, syserror.ENOMEM
+	}
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_mount.go b/pkg/sentry/syscalls/linux/sys_mount.go
new file mode 100644
index 000000000..d70b79e4f
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_mount.go
@@ -0,0 +1,140 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Mount implements Linux syscall mount(2).
+func Mount(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	sourceAddr := args[0].Pointer()
+	targetAddr := args[1].Pointer()
+	typeAddr := args[2].Pointer()
+	flags := args[3].Uint64()
+	dataAddr := args[4].Pointer()
+
+	fsType, err := t.CopyInString(typeAddr, usermem.PageSize)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	sourcePath, _, err := copyInPath(t, sourceAddr, true /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	targetPath, _, err := copyInPath(t, targetAddr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	// In Linux, a full page is always copied in regardless of null
+	// character placement, and the address is passed to each file system.
+	// Most file systems always treat this data as a string, though, and so
+	// do all of the ones we implement.
+	data, err := t.CopyInString(dataAddr, usermem.PageSize)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	// Ignore magic value that was required before Linux 2.4.
+	if flags&linux.MS_MGC_MSK == linux.MS_MGC_VAL {
+		flags = flags &^ linux.MS_MGC_MSK
+	}
+
+	// Must have CAP_SYS_ADMIN in the mount namespace's associated user
+	// namespace.
+	if !t.HasCapabilityIn(linux.CAP_SYS_ADMIN, t.MountNamespace().UserNamespace()) {
+		return 0, nil, syserror.EPERM
+	}
+
+	const unsupportedOps = linux.MS_REMOUNT | linux.MS_BIND |
+		linux.MS_SHARED | linux.MS_PRIVATE | linux.MS_SLAVE |
+		linux.MS_UNBINDABLE | linux.MS_MOVE
+
+	// Silently allow MS_NOSUID, since we don't implement set-id bits
+	// anyway.
+	const unsupportedFlags = linux.MS_NODEV | linux.MS_NOEXEC |
+		linux.MS_NODIRATIME | linux.MS_STRICTATIME
+
+	// Linux just allows passing any flags to mount(2) - it won't fail when
+	// unknown or unsupported flags are passed. Since we don't implement
+	// everything, we fail explicitly on flags that are unimplemented.
+	if flags&(unsupportedOps|unsupportedFlags) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	rsys, ok := fs.FindFilesystem(fsType)
+	if !ok {
+		return 0, nil, syserror.ENODEV
+	}
+	if !rsys.AllowUserMount() {
+		return 0, nil, syserror.EPERM
+	}
+
+	var superFlags fs.MountSourceFlags
+	if flags&linux.MS_NOATIME == linux.MS_NOATIME {
+		superFlags.NoAtime = true
+	}
+	if flags&linux.MS_RDONLY == linux.MS_RDONLY {
+		superFlags.ReadOnly = true
+	}
+
+	rootInode, err := rsys.Mount(t, sourcePath, superFlags, data)
+	if err != nil {
+		return 0, nil, syserror.EINVAL
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, targetPath, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return t.MountNamespace().Mount(t, d, rootInode)
+	})
+}
+
+// Umount2 implements Linux syscall umount2(2).
+func Umount2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	flags := args[1].Int()
+
+	const unsupported = linux.MNT_FORCE | linux.MNT_EXPIRE
+	if flags&unsupported != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	// Must have CAP_SYS_ADMIN in the mount namespace's associated user
+	// namespace.
+	//
+	// Currently, this is always the init task's user namespace.
+	if !t.HasCapabilityIn(linux.CAP_SYS_ADMIN, t.MountNamespace().UserNamespace()) {
+		return 0, nil, syserror.EPERM
+	}
+
+	resolve := flags&linux.UMOUNT_NOFOLLOW != linux.UMOUNT_NOFOLLOW
+	detachOnly := flags&linux.MNT_DETACH == linux.MNT_DETACH
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, resolve, func(root *fs.Dirent, d *fs.Dirent) error {
+		return t.MountNamespace().Unmount(t, d, detachOnly)
+	})
+}
diff --git a/pkg/sentry/syscalls/linux/sys_pipe.go b/pkg/sentry/syscalls/linux/sys_pipe.go
new file mode 100644
index 000000000..3efc06a27
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_pipe.go
@@ -0,0 +1,78 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/pipe"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// pipe2 implements the actual system call with flags.
+func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) {
+	if flags&^(syscall.O_NONBLOCK|syscall.O_CLOEXEC) != 0 {
+		return 0, syscall.EINVAL
+	}
+	r, w := pipe.NewConnectedPipe(t, pipe.DefaultPipeSize, usermem.PageSize)
+
+	r.SetFlags(linuxToFlags(flags).Settable())
+	defer r.DecRef()
+
+	w.SetFlags(linuxToFlags(flags).Settable())
+	defer w.DecRef()
+
+	rfd, err := t.FDMap().NewFDFrom(0, r, kernel.FDFlags{
+		CloseOnExec: flags&syscall.O_CLOEXEC != 0},
+		t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, err
+	}
+
+	wfd, err := t.FDMap().NewFDFrom(0, w, kernel.FDFlags{
+		CloseOnExec: flags&syscall.O_CLOEXEC != 0},
+		t.ThreadGroup().Limits())
+	if err != nil {
+		t.FDMap().Remove(rfd)
+		return 0, err
+	}
+
+	if _, err := t.CopyOut(addr, []kdefs.FD{rfd, wfd}); err != nil {
+		t.FDMap().Remove(rfd)
+		t.FDMap().Remove(wfd)
+		return 0, syscall.EFAULT
+	}
+	return 0, nil
+}
+
+// Pipe implements linux syscall pipe(2).
+func Pipe(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	n, err := pipe2(t, addr, 0)
+	return n, nil, err
+}
+
+// Pipe2 implements linux syscall pipe2(2).
+func Pipe2(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	flags := uint(args[1].Uint())
+
+	n, err := pipe2(t, addr, flags)
+	return n, nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_poll.go b/pkg/sentry/syscalls/linux/sys_poll.go
new file mode 100644
index 000000000..d4dbfd285
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_poll.go
@@ -0,0 +1,429 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/syscalls"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// fileCap is the maximum allowable files for poll & select.
+const fileCap = 1024 * 1024
+
+// Masks for "readable", "writable", and "exceptional" events as defined by
+// select(2).
+const (
+	// selectReadEvents is analogous to the Linux kernel's
+	// fs/select.c:POLLIN_SET.
+	selectReadEvents = waiter.EventIn | waiter.EventHUp | waiter.EventErr
+
+	// selectWriteEvents is analogous to the Linux kernel's
+	// fs/select.c:POLLOUT_SET.
+	selectWriteEvents = waiter.EventOut | waiter.EventErr
+
+	// selectExceptEvents is analogous to the Linux kernel's
+	// fs/select.c:POLLEX_SET.
+	selectExceptEvents = waiter.EventPri
+)
+
+func doPoll(t *kernel.Task, pfdAddr usermem.Addr, nfds uint, timeout time.Duration) (time.Duration, uintptr, error) {
+	if uint64(nfds) > t.ThreadGroup().Limits().GetCapped(limits.NumberOfFiles, fileCap) {
+		return timeout, 0, syserror.EINVAL
+	}
+
+	pfd := make([]syscalls.PollFD, nfds)
+	if nfds > 0 {
+		if _, err := t.CopyIn(pfdAddr, &pfd); err != nil {
+			return timeout, 0, err
+		}
+	}
+
+	// Compatibility warning: Linux adds POLLHUP and POLLERR just before
+	// polling, in fs/select.c:do_pollfd(). Since pfd is copied out after
+	// polling, changing event masks here is an application-visible difference.
+	// (Linux also doesn't copy out event masks at all, only revents.)
+	for i := range pfd {
+		pfd[i].Events |= waiter.EventHUp | waiter.EventErr
+	}
+	remainingTimeout, n, err := syscalls.Poll(t, pfd, timeout)
+	err = syserror.ConvertIntr(err, syserror.EINTR)
+
+	// The poll entries are copied out regardless of whether
+	// any are set or not. This aligns with the Linux behavior.
+	if nfds > 0 && err == nil {
+		if _, err := t.CopyOut(pfdAddr, pfd); err != nil {
+			return remainingTimeout, 0, err
+		}
+	}
+
+	return remainingTimeout, n, err
+}
+
+func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Addr, timeout time.Duration) (uintptr, error) {
+	if nfds < 0 || uint64(nfds) > t.ThreadGroup().Limits().GetCapped(limits.NumberOfFiles, fileCap) {
+		return 0, syserror.EINVAL
+	}
+
+	// Capture all the provided input vectors.
+	//
+	// N.B. This only works on little-endian architectures.
+	byteCount := (nfds + 7) / 8
+	bitsInLastPartialByte := uint(nfds % 8)
+	r := make([]byte, byteCount)
+	w := make([]byte, byteCount)
+	e := make([]byte, byteCount)
+
+	if readFDs != 0 {
+		if _, err := t.CopyIn(readFDs, &r); err != nil {
+			return 0, err
+		}
+		// Mask out bits above nfds.
+		if bitsInLastPartialByte != 0 {
+			r[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte
+		}
+	}
+
+	if writeFDs != 0 {
+		if _, err := t.CopyIn(writeFDs, &w); err != nil {
+			return 0, err
+		}
+		if bitsInLastPartialByte != 0 {
+			w[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte
+		}
+	}
+
+	if exceptFDs != 0 {
+		if _, err := t.CopyIn(exceptFDs, &e); err != nil {
+			return 0, err
+		}
+		if bitsInLastPartialByte != 0 {
+			e[byteCount-1] &^= byte(0xff) << bitsInLastPartialByte
+		}
+	}
+
+	// Count how many FDs are actually being requested so that we can build
+	// a PollFD array.
+	fdCount := 0
+	for i := 0; i < byteCount; i++ {
+		v := r[i] | w[i] | e[i]
+		for v != 0 {
+			v &= (v - 1)
+			fdCount++
+		}
+	}
+
+	// Build the PollFD array.
+	pfd := make([]syscalls.PollFD, 0, fdCount)
+	fd := kdefs.FD(0)
+	for i := 0; i < byteCount; i++ {
+		rV, wV, eV := r[i], w[i], e[i]
+		v := rV | wV | eV
+		m := byte(1)
+		for j := 0; j < 8; j++ {
+			if (v & m) != 0 {
+				// Make sure the fd is valid and decrement the reference
+				// immediately to ensure we don't leak. Note, another thread
+				// might be about to close fd. This is racy, but that's
+				// OK. Linux is racy in the same way.
+				file := t.FDMap().GetFile(fd)
+				if file == nil {
+					return 0, syserror.EBADF
+				}
+				file.DecRef()
+
+				mask := waiter.EventMask(0)
+				if (rV & m) != 0 {
+					mask |= selectReadEvents
+				}
+
+				if (wV & m) != 0 {
+					mask |= selectWriteEvents
+				}
+
+				if (eV & m) != 0 {
+					mask |= selectExceptEvents
+				}
+
+				pfd = append(pfd, syscalls.PollFD{
+					FD:     fd,
+					Events: mask,
+				})
+			}
+
+			fd++
+			m <<= 1
+		}
+	}
+
+	// Do the syscall, then count the number of bits set.
+	_, _, err := syscalls.Poll(t, pfd, timeout)
+	if err != nil {
+		return 0, syserror.ConvertIntr(err, syserror.EINTR)
+	}
+
+	// r, w, and e are currently event mask bitsets; unset bits corresponding
+	// to events that *didn't* occur.
+	bitSetCount := uintptr(0)
+	for idx := range pfd {
+		events := pfd[idx].REvents
+		i, j := pfd[idx].FD/8, uint(pfd[idx].FD%8)
+		m := byte(1) << j
+		if r[i]&m != 0 {
+			if (events & selectReadEvents) != 0 {
+				bitSetCount++
+			} else {
+				r[i] &^= m
+			}
+		}
+		if w[i]&m != 0 {
+			if (events & selectWriteEvents) != 0 {
+				bitSetCount++
+			} else {
+				w[i] &^= m
+			}
+		}
+		if e[i]&m != 0 {
+			if (events & selectExceptEvents) != 0 {
+				bitSetCount++
+			} else {
+				e[i] &^= m
+			}
+		}
+	}
+
+	// Copy updated vectors back.
+	if readFDs != 0 {
+		if _, err := t.CopyOut(readFDs, r); err != nil {
+			return 0, err
+		}
+	}
+
+	if writeFDs != 0 {
+		if _, err := t.CopyOut(writeFDs, w); err != nil {
+			return 0, err
+		}
+	}
+
+	if exceptFDs != 0 {
+		if _, err := t.CopyOut(exceptFDs, e); err != nil {
+			return 0, err
+		}
+	}
+
+	return bitSetCount, nil
+}
+
+// timeoutRemaining returns the amount of time remaining for the specified
+// timeout or 0 if it has elapsed.
+//
+// startNs must be from CLOCK_MONOTONIC.
+func timeoutRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration) time.Duration {
+	now := t.Kernel().MonotonicClock().Now()
+	remaining := timeout - now.Sub(startNs)
+	if remaining < 0 {
+		remaining = 0
+	}
+	return remaining
+}
+
+// copyOutTimespecRemaining copies the time remaining in timeout to timespecAddr.
+//
+// startNs must be from CLOCK_MONOTONIC.
+func copyOutTimespecRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timespecAddr usermem.Addr) error {
+	if timeout <= 0 {
+		return nil
+	}
+	remaining := timeoutRemaining(t, startNs, timeout)
+	tsRemaining := linux.NsecToTimespec(remaining.Nanoseconds())
+	return copyTimespecOut(t, timespecAddr, &tsRemaining)
+}
+
+// copyOutTimevalRemaining copies the time remaining in timeout to timevalAddr.
+//
+// startNs must be from CLOCK_MONOTONIC.
+func copyOutTimevalRemaining(t *kernel.Task, startNs ktime.Time, timeout time.Duration, timevalAddr usermem.Addr) error {
+	if timeout <= 0 {
+		return nil
+	}
+	remaining := timeoutRemaining(t, startNs, timeout)
+	tvRemaining := linux.NsecToTimeval(remaining.Nanoseconds())
+	return copyTimevalOut(t, timevalAddr, &tvRemaining)
+}
+
+// pollRestartBlock encapsulates the state required to restart poll(2) via
+// restart_syscall(2).
+type pollRestartBlock struct {
+	pfdAddr usermem.Addr
+	nfds    uint
+	timeout time.Duration
+}
+
+// Restart implements kernel.SyscallRestartBlock.Restart.
+func (p *pollRestartBlock) Restart(t *kernel.Task) (uintptr, error) {
+	return poll(t, p.pfdAddr, p.nfds, p.timeout)
+}
+
+func poll(t *kernel.Task, pfdAddr usermem.Addr, nfds uint, timeout time.Duration) (uintptr, error) {
+	remainingTimeout, n, err := doPoll(t, pfdAddr, nfds, timeout)
+	// On an interrupt poll(2) is restarted with the remaining timeout.
+	if err == syserror.EINTR {
+		t.SetSyscallRestartBlock(&pollRestartBlock{
+			pfdAddr: pfdAddr,
+			nfds:    nfds,
+			timeout: remainingTimeout,
+		})
+		return 0, kernel.ERESTART_RESTARTBLOCK
+	}
+	return n, err
+}
+
+// Poll implements linux syscall poll(2).
+func Poll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pfdAddr := args[0].Pointer()
+	nfds := uint(args[1].Uint()) // poll(2) uses unsigned long.
+	timeout := time.Duration(args[2].Int()) * time.Millisecond
+	n, err := poll(t, pfdAddr, nfds, timeout)
+	return n, nil, err
+}
+
+// Ppoll implements linux syscall ppoll(2).
+func Ppoll(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pfdAddr := args[0].Pointer()
+	nfds := uint(args[1].Uint()) // poll(2) uses unsigned long.
+	timespecAddr := args[2].Pointer()
+	maskAddr := args[3].Pointer()
+	maskSize := uint(args[4].Uint())
+
+	timeout, err := copyTimespecInToDuration(t, timespecAddr)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	var startNs ktime.Time
+	if timeout > 0 {
+		startNs = t.Kernel().MonotonicClock().Now()
+	}
+
+	if maskAddr != 0 {
+		mask, err := copyInSigSet(t, maskAddr, maskSize)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		oldmask := t.SignalMask()
+		t.SetSignalMask(mask)
+		t.SetSavedSignalMask(oldmask)
+	}
+
+	_, n, err := doPoll(t, pfdAddr, nfds, timeout)
+	copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr)
+	// doPoll returns EINTR if interrupted, but ppoll is normally restartable
+	// if interrupted by something other than a signal handled by the
+	// application (i.e. returns ERESTARTNOHAND). However, if
+	// copyOutTimespecRemaining failed, then the restarted ppoll would use the
+	// wrong timeout, so the error should be left as EINTR.
+	//
+	// Note that this means that if err is nil but copyErr is not, copyErr is
+	// ignored. This is consistent with Linux.
+	if err == syserror.EINTR && copyErr == nil {
+		err = kernel.ERESTARTNOHAND
+	}
+	return n, nil, err
+}
+
+// Select implements linux syscall select(2).
+func Select(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	nfds := int(args[0].Int()) // select(2) uses an int.
+	readFDs := args[1].Pointer()
+	writeFDs := args[2].Pointer()
+	exceptFDs := args[3].Pointer()
+	timevalAddr := args[4].Pointer()
+
+	// Use a negative Duration to indicate "no timeout".
+	timeout := time.Duration(-1)
+	if timevalAddr != 0 {
+		timeval, err := copyTimevalIn(t, timevalAddr)
+		if err != nil {
+			return 0, nil, err
+		}
+		if timeval.Sec < 0 || timeval.Usec < 0 {
+			return 0, nil, syserror.EINVAL
+		}
+		timeout = time.Duration(timeval.ToNsecCapped())
+	}
+	startNs := t.Kernel().MonotonicClock().Now()
+	n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout)
+	copyErr := copyOutTimevalRemaining(t, startNs, timeout, timevalAddr)
+	// See comment in Ppoll.
+	if err == syserror.EINTR && copyErr == nil {
+		err = kernel.ERESTARTNOHAND
+	}
+	return n, nil, err
+}
+
+// Pselect implements linux syscall pselect(2).
+func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	nfds := int(args[0].Int()) // select(2) uses an int.
+	readFDs := args[1].Pointer()
+	writeFDs := args[2].Pointer()
+	exceptFDs := args[3].Pointer()
+	timespecAddr := args[4].Pointer()
+	maskWithSizeAddr := args[5].Pointer()
+
+	timeout, err := copyTimespecInToDuration(t, timespecAddr)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	var startNs ktime.Time
+	if timeout > 0 {
+		startNs = t.Kernel().MonotonicClock().Now()
+	}
+
+	if maskWithSizeAddr != 0 {
+		maskAddr, size, err := copyInSigSetWithSize(t, maskWithSizeAddr)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		if maskAddr != 0 {
+			mask, err := copyInSigSet(t, maskAddr, size)
+			if err != nil {
+				return 0, nil, err
+			}
+			oldmask := t.SignalMask()
+			t.SetSignalMask(mask)
+			t.SetSavedSignalMask(oldmask)
+		}
+	}
+
+	n, err := doSelect(t, nfds, readFDs, writeFDs, exceptFDs, timeout)
+	copyErr := copyOutTimespecRemaining(t, startNs, timeout, timespecAddr)
+	// See comment in Ppoll.
+	if err == syserror.EINTR && copyErr == nil {
+		err = kernel.ERESTARTNOHAND
+	}
+	return n, nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go
new file mode 100644
index 000000000..2ca7471cf
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_prctl.go
@@ -0,0 +1,188 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/bpf"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// userSockFprog is equivalent to Linux's struct sock_fprog on amd64.
+type userSockFprog struct {
+	// Len is the length of the filter in BPF instructions.
+	Len uint16
+
+	_ [6]byte // padding for alignment
+
+	// Filter is a user pointer to the struct sock_filter array that makes up
+	// the filter program. Filter is a uint64 rather than a usermem.Addr
+	// because usermem.Addr is actually uintptr, which is not a fixed-size
+	// type, and encoding/binary.Read objects to this.
+	Filter uint64
+}
+
+// Prctl implements linux syscall prctl(2).
+// It has a list of subfunctions which operate on the process. The arguments are
+// all based on each subfunction.
+func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	option := args[0].Int()
+
+	switch option {
+	case linux.PR_SET_PDEATHSIG:
+		sig := linux.Signal(args[1].Int())
+		if sig != 0 && !sig.IsValid() {
+			return 0, nil, syscall.EINVAL
+		}
+		t.SetParentDeathSignal(sig)
+		return 0, nil, nil
+
+	case linux.PR_GET_PDEATHSIG:
+		_, err := t.CopyOut(args[1].Pointer(), int32(t.ParentDeathSignal()))
+		return 0, nil, err
+
+	case linux.PR_GET_KEEPCAPS:
+		if t.Credentials().KeepCaps {
+			return 1, nil, nil
+		}
+
+		return 0, nil, nil
+
+	case linux.PR_SET_KEEPCAPS:
+		val := args[1].Int()
+		// prctl(2): arg2 must be either 0 (permitted capabilities are cleared)
+		// or 1 (permitted capabilities are kept).
+		if val == 0 {
+			t.SetKeepCaps(false)
+		} else if val == 1 {
+			t.SetKeepCaps(true)
+		} else {
+			return 0, nil, syscall.EINVAL
+		}
+
+		return 0, nil, nil
+
+	case linux.PR_SET_NAME:
+		addr := args[1].Pointer()
+		name, err := t.CopyInString(addr, linux.TASK_COMM_LEN-1)
+		if err != nil && err != syscall.ENAMETOOLONG {
+			return 0, nil, err
+		}
+		t.SetName(name)
+
+	case linux.PR_GET_NAME:
+		addr := args[1].Pointer()
+		buf := make([]byte, linux.TASK_COMM_LEN)
+		len := copy(buf, t.Name())
+		if len < linux.TASK_COMM_LEN {
+			buf[len] = 0
+			len++
+		}
+		_, err := t.CopyOut(addr, buf[:len])
+		if err != nil {
+			return 0, nil, err
+		}
+
+	case linux.PR_SET_MM:
+		switch args[1].Int() {
+		case linux.PR_SET_MM_EXE_FILE:
+			fd := kdefs.FD(args[2].Int())
+
+			file := t.FDMap().GetFile(fd)
+			if file == nil {
+				return 0, nil, syscall.EBADF
+			}
+			defer file.DecRef()
+
+			// They trying to set exe to a non-file?
+			if !fs.IsFile(file.Dirent.Inode.StableAttr) {
+				return 0, nil, syscall.EBADF
+			}
+
+			// Set the underlying executable.
+			t.MemoryManager().SetExecutable(file.Dirent)
+		default:
+			return 0, nil, syscall.EINVAL
+		}
+
+	case linux.PR_SET_NO_NEW_PRIVS:
+		if args[1].Int() != 1 || args[2].Int() != 0 || args[3].Int() != 0 || args[4].Int() != 0 {
+			return 0, nil, syscall.EINVAL
+		}
+		// no_new_privs is assumed to always be set. See
+		// auth.Credentials.UpdateForExec.
+		return 0, nil, nil
+
+	case linux.PR_GET_NO_NEW_PRIVS:
+		if args[1].Int() != 0 || args[2].Int() != 0 || args[3].Int() != 0 || args[4].Int() != 0 {
+			return 0, nil, syscall.EINVAL
+		}
+		return 1, nil, nil
+
+	case linux.PR_SET_SECCOMP:
+		if args[1].Int() != linux.SECCOMP_MODE_FILTER {
+			// Unsupported mode.
+			return 0, nil, syscall.EINVAL
+		}
+		var fprog userSockFprog
+		if _, err := t.CopyIn(args[2].Pointer(), &fprog); err != nil {
+			return 0, nil, err
+		}
+		filter := make([]linux.BPFInstruction, int(fprog.Len))
+		if _, err := t.CopyIn(usermem.Addr(fprog.Filter), &filter); err != nil {
+			return 0, nil, err
+		}
+		compiledFilter, err := bpf.Compile(filter)
+		if err != nil {
+			t.Debugf("Invalid seccomp-bpf filter: %v", err)
+			return 0, nil, syscall.EINVAL
+		}
+		return 0, nil, t.AppendSyscallFilter(compiledFilter)
+
+	case linux.PR_GET_SECCOMP:
+		return uintptr(t.SeccompMode()), nil, nil
+
+	case linux.PR_CAPBSET_READ:
+		cp := linux.Capability(args[1].Uint64())
+		if !cp.Ok() {
+			return 0, nil, syscall.EINVAL
+		}
+		var rv uintptr
+		if auth.CapabilitySetOf(cp)&t.Credentials().BoundingCaps != 0 {
+			rv = 1
+		}
+		return rv, nil, nil
+
+	case linux.PR_CAPBSET_DROP:
+		cp := linux.Capability(args[1].Uint64())
+		if !cp.Ok() {
+			return 0, nil, syscall.EINVAL
+		}
+		return 0, nil, t.DropBoundingCapability(cp)
+
+	default:
+		t.Warningf("Unsupported prctl %d", option)
+		return 0, nil, syscall.EINVAL
+	}
+
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_random.go b/pkg/sentry/syscalls/linux/sys_random.go
new file mode 100644
index 000000000..2dd59b1c3
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_random.go
@@ -0,0 +1,92 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"crypto/rand"
+	"io"
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	_GRND_NONBLOCK = 0x1
+	_GRND_RANDOM   = 0x2
+)
+
+// GetRandom implements the linux syscall getrandom(2).
+//
+// In a multi-tenant/shared environment, the only valid implementation is to
+// fetch data from the urandom pool, otherwise starvation attacks become
+// possible. The urandom pool is also expected to have plenty of entropy, thus
+// the GRND_RANDOM flag is ignored. The GRND_NONBLOCK flag does not apply, as
+// the pool will already be initialized.
+func GetRandom(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	length := args[1].SizeT()
+	flags := args[2].Int()
+
+	// Flags are checked for validity but otherwise ignored. See above.
+	if flags & ^(_GRND_NONBLOCK|_GRND_RANDOM) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if length > math.MaxInt32 {
+		length = math.MaxInt32
+	}
+	ar, ok := addr.ToRange(uint64(length))
+	if !ok {
+		return 0, nil, syserror.EFAULT
+	}
+
+	// "If the urandom source has been initialized, reads of up to 256 bytes
+	// will always return as many bytes as requested and will not be
+	// interrupted by signals. No such guarantees apply for larger buffer
+	// sizes." - getrandom(2)
+	min := int(length)
+	if min > 256 {
+		min = 256
+	}
+	n, err := t.MemoryManager().CopyOutFrom(t, usermem.AddrRangeSeqOf(ar), safemem.FromIOReader{&randReader{-1, min}}, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if n >= int64(min) {
+		return uintptr(n), nil, nil
+	}
+	return 0, nil, err
+}
+
+// randReader is a io.Reader that handles partial reads from rand.Reader.
+type randReader struct {
+	done int
+	min  int
+}
+
+// Read implements io.Reader.Read.
+func (r *randReader) Read(dst []byte) (int, error) {
+	if r.done >= r.min {
+		return rand.Reader.Read(dst)
+	}
+	min := r.min - r.done
+	if min > len(dst) {
+		min = len(dst)
+	}
+	return io.ReadAtLeast(rand.Reader, dst, min)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_read.go b/pkg/sentry/syscalls/linux/sys_read.go
new file mode 100644
index 000000000..0be2d195a
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_read.go
@@ -0,0 +1,274 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+const (
+	// EventMaskRead contains events that can be triggerd on reads.
+	EventMaskRead = waiter.EventIn | waiter.EventHUp | waiter.EventErr
+)
+
+// Read implements linux syscall read(2).  Note that we try to get a buffer that
+// is exactly the size requested because some applications like qemu expect
+// they can do large reads all at once.  Bug for bug.  Same for other read
+// calls below.
+func Read(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := args[2].SizeT()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the file is readable.
+	if !file.Flags().Read {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Check that the size is legitimate.
+	si := int(size)
+	if si < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Get the destination of the read.
+	dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := readv(t, file, dst)
+	t.IOUsage().AccountReadSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "read", file)
+}
+
+// Pread64 implements linux syscall pread64(2).
+func Pread64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := args[2].SizeT()
+	offset := args[3].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the offset is legitimate.
+	if offset < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Is reading at an offset supported?
+	if !file.Flags().Pread {
+		return 0, nil, syserror.ESPIPE
+	}
+
+	// Check that the file is readable.
+	if !file.Flags().Read {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Check that the size is legitimate.
+	si := int(size)
+	if si < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Get the destination of the read.
+	dst, err := t.SingleIOSequence(addr, si, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := preadv(t, file, dst, offset)
+	t.IOUsage().AccountReadSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "pread64", file)
+}
+
+// Readv implements linux syscall readv(2).
+func Readv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	iovcnt := int(args[2].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the file is readable.
+	if !file.Flags().Read {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Read the iovecs that specify the destination of the read.
+	dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := readv(t, file, dst)
+	t.IOUsage().AccountReadSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "readv", file)
+}
+
+// Preadv implements linux syscall preadv(2).
+func Preadv(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	iovcnt := int(args[2].Int())
+	offset := args[3].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the offset is legitimate.
+	if offset < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Is reading at an offset supported?
+	if !file.Flags().Pread {
+		return 0, nil, syserror.ESPIPE
+	}
+
+	// Check that the file is readable.
+	if !file.Flags().Read {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Read the iovecs that specify the destination of the read.
+	dst, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := preadv(t, file, dst, offset)
+	t.IOUsage().AccountReadSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "preadv", file)
+}
+
+func readv(t *kernel.Task, f *fs.File, dst usermem.IOSequence) (int64, error) {
+	n, err := f.Readv(t, dst)
+	if err != syserror.ErrWouldBlock || f.Flags().NonBlocking {
+		if n > 0 {
+			// Queue notification if we read anything.
+			f.Dirent.InotifyEvent(linux.IN_ACCESS, 0)
+		}
+		return n, err
+	}
+
+	// Register for notifications.
+	w, ch := waiter.NewChannelEntry(nil)
+	f.EventRegister(&w, EventMaskRead)
+
+	total := n
+	for {
+		// Shorten dst to reflect bytes previously read.
+		dst = dst.DropFirst64(n)
+
+		// Issue the request and break out if it completes with anything
+		// other than "would block".
+		n, err = f.Readv(t, dst)
+		total += n
+		if err != syserror.ErrWouldBlock {
+			break
+		}
+
+		// Wait for a notification that we should retry.
+		if err = t.Block(ch); err != nil {
+			break
+		}
+	}
+
+	f.EventUnregister(&w)
+
+	if total > 0 {
+		// Queue notification if we read anything.
+		f.Dirent.InotifyEvent(linux.IN_ACCESS, 0)
+	}
+
+	return total, err
+}
+
+func preadv(t *kernel.Task, f *fs.File, dst usermem.IOSequence, offset int64) (int64, error) {
+	n, err := f.Preadv(t, dst, offset)
+	if err != syserror.ErrWouldBlock || f.Flags().NonBlocking {
+		if n > 0 {
+			// Queue notification if we read anything.
+			f.Dirent.InotifyEvent(linux.IN_ACCESS, 0)
+		}
+		return n, err
+	}
+
+	// Register for notifications.
+	w, ch := waiter.NewChannelEntry(nil)
+	f.EventRegister(&w, EventMaskRead)
+
+	total := n
+	for {
+		// Shorten dst to reflect bytes previously read.
+		dst = dst.DropFirst64(n)
+
+		// Issue the request and break out if it completes with anything
+		// other than "would block".
+		n, err = f.Preadv(t, dst, offset+total)
+		total += n
+		if err != syserror.ErrWouldBlock {
+			break
+		}
+
+		// Wait for a notification that we should retry.
+		if err = t.Block(ch); err != nil {
+			break
+		}
+	}
+
+	f.EventUnregister(&w)
+
+	if total > 0 {
+		// Queue notification if we read anything.
+		f.Dirent.InotifyEvent(linux.IN_ACCESS, 0)
+	}
+
+	return total, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_rlimit.go b/pkg/sentry/syscalls/linux/sys_rlimit.go
new file mode 100644
index 000000000..481e79eaa
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_rlimit.go
@@ -0,0 +1,217 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/limits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// rlimit describes an implementation of 'struct rlimit', which may vary from
+// system-to-system.
+type rlimit interface {
+	// toLimit converts an rlimit to a limits.Limit.
+	toLimit() *limits.Limit
+
+	// fromLimit converts a limits.Limit to an rlimit.
+	fromLimit(lim limits.Limit)
+
+	// copyIn copies an rlimit from the untrusted app to the kernel.
+	copyIn(t *kernel.Task, addr usermem.Addr) error
+
+	// copyOut copies an rlimit from the kernel to the untrusted app.
+	copyOut(t *kernel.Task, addr usermem.Addr) error
+}
+
+// newRlimit returns the appropriate rlimit type for 'struct rlimit' on this system.
+func newRlimit(t *kernel.Task) (rlimit, error) {
+	switch t.Arch().Width() {
+	case 8:
+		// On 64-bit system, struct rlimit and struct rlimit64 are identical.
+		return &rlimit64{}, nil
+	default:
+		return nil, syserror.ENOSYS
+	}
+}
+
+type rlimit64 struct {
+	Cur uint64
+	Max uint64
+}
+
+func (r *rlimit64) toLimit() *limits.Limit {
+	return &limits.Limit{
+		Cur: limits.FromLinux(r.Cur),
+		Max: limits.FromLinux(r.Max),
+	}
+}
+
+func (r *rlimit64) fromLimit(lim limits.Limit) {
+	*r = rlimit64{
+		Cur: limits.ToLinux(lim.Cur),
+		Max: limits.ToLinux(lim.Max),
+	}
+}
+
+func (r *rlimit64) copyIn(t *kernel.Task, addr usermem.Addr) error {
+	_, err := t.CopyIn(addr, r)
+	return err
+}
+
+func (r *rlimit64) copyOut(t *kernel.Task, addr usermem.Addr) error {
+	_, err := t.CopyOut(addr, *r)
+	return err
+}
+
+func makeRlimit64(lim limits.Limit) *rlimit64 {
+	return &rlimit64{Cur: lim.Cur, Max: lim.Max}
+}
+
+// setableLimits is the set of supported setable limits.
+var setableLimits = map[limits.LimitType]struct{}{
+	limits.NumberOfFiles: {},
+	limits.AS:            {},
+	limits.CPU:           {},
+	limits.Data:          {},
+	limits.FileSize:      {},
+	limits.Stack:         {},
+	// These are not enforced, but we include them here to avoid returning
+	// EPERM, since some apps expect them to succeed.
+	limits.Core:         {},
+	limits.ProcessCount: {},
+}
+
+func prlimit64(t *kernel.Task, resource limits.LimitType, newLim *limits.Limit) (limits.Limit, error) {
+	if newLim == nil {
+		return t.ThreadGroup().Limits().Get(resource), nil
+	}
+
+	if _, ok := setableLimits[resource]; !ok {
+		return limits.Limit{}, syserror.EPERM
+	}
+	oldLim, err := t.ThreadGroup().Limits().Set(resource, *newLim)
+	if err != nil {
+		return limits.Limit{}, err
+	}
+
+	if resource == limits.CPU {
+		t.ThreadGroup().SetCPUTimer(newLim)
+	}
+	return oldLim, nil
+}
+
+// Getrlimit implements linux syscall getrlimit(2).
+func Getrlimit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	resource, ok := limits.FromLinuxResource[int(args[0].Int())]
+	if !ok {
+		// Return err; unknown limit.
+		return 0, nil, syserror.EINVAL
+	}
+	addr := args[1].Pointer()
+	rlim, err := newRlimit(t)
+	if err != nil {
+		return 0, nil, err
+	}
+	lim, err := prlimit64(t, resource, nil)
+	if err != nil {
+		return 0, nil, err
+	}
+	rlim.fromLimit(lim)
+	return 0, nil, rlim.copyOut(t, addr)
+}
+
+// Setrlimit implements linux syscall setrlimit(2).
+func Setrlimit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	resource, ok := limits.FromLinuxResource[int(args[0].Int())]
+	if !ok {
+		// Return err; unknown limit.
+		return 0, nil, syserror.EINVAL
+	}
+	addr := args[1].Pointer()
+	rlim, err := newRlimit(t)
+	if err != nil {
+		return 0, nil, err
+	}
+	if err := rlim.copyIn(t, addr); err != nil {
+		return 0, nil, syserror.EFAULT
+	}
+	_, err = prlimit64(t, resource, rlim.toLimit())
+	return 0, nil, err
+}
+
+// Prlimit64 implements linux syscall prlimit64(2).
+func Prlimit64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := kernel.ThreadID(args[0].Int())
+	resource, ok := limits.FromLinuxResource[int(args[1].Int())]
+	if !ok {
+		// Return err; unknown limit.
+		return 0, nil, syserror.EINVAL
+	}
+	newRlimAddr := args[2].Pointer()
+	oldRlimAddr := args[3].Pointer()
+
+	var newLim *limits.Limit
+	if newRlimAddr != 0 {
+		var nrl rlimit64
+		if err := nrl.copyIn(t, newRlimAddr); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+		newLim = nrl.toLimit()
+	}
+
+	if tid < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	ot := t
+	if tid > 0 {
+		if ot = t.PIDNamespace().TaskWithID(tid); ot == nil {
+			return 0, nil, syserror.ESRCH
+		}
+	}
+
+	// "To set or get the resources of a process other than itself, the caller
+	// must have the CAP_SYS_RESOURCE capability, or the real, effective, and
+	// saved set user IDs of the target process must match the real user ID of
+	// the caller and the real, effective, and saved set group IDs of the
+	// target process must match the real group ID of the caller."
+	if !t.HasCapabilityIn(linux.CAP_SYS_RESOURCE, t.PIDNamespace().UserNamespace()) {
+		cred, tcred := t.Credentials(), ot.Credentials()
+		if cred.RealKUID != tcred.RealKUID ||
+			cred.RealKUID != tcred.EffectiveKUID ||
+			cred.RealKUID != tcred.SavedKUID ||
+			cred.RealKGID != tcred.RealKGID ||
+			cred.RealKGID != tcred.EffectiveKGID ||
+			cred.RealKGID != tcred.SavedKGID {
+			return 0, nil, syserror.EPERM
+		}
+	}
+
+	oldLim, err := prlimit64(ot, resource, newLim)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if oldRlimAddr != 0 {
+		if err := makeRlimit64(oldLim).copyOut(t, oldRlimAddr); err != nil {
+			return 0, nil, syserror.EFAULT
+		}
+	}
+
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_rusage.go b/pkg/sentry/syscalls/linux/sys_rusage.go
new file mode 100644
index 000000000..82e42b589
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_rusage.go
@@ -0,0 +1,112 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+func getrusage(t *kernel.Task, which int32) linux.Rusage {
+	var cs usage.CPUStats
+
+	switch which {
+	case linux.RUSAGE_SELF:
+		cs = t.ThreadGroup().CPUStats()
+
+	case linux.RUSAGE_CHILDREN:
+		cs = t.ThreadGroup().JoinedChildCPUStats()
+
+	case linux.RUSAGE_THREAD:
+		cs = t.CPUStats()
+
+	case linux.RUSAGE_BOTH:
+		tg := t.ThreadGroup()
+		cs = tg.CPUStats()
+		cs.Accumulate(tg.JoinedChildCPUStats())
+	}
+
+	return linux.Rusage{
+		UTime:  linux.NsecToTimeval(cs.UserTime.Nanoseconds()),
+		STime:  linux.NsecToTimeval(cs.SysTime.Nanoseconds()),
+		NVCSw:  int64(cs.VoluntarySwitches),
+		MaxRSS: int64(t.MaxRSS(which) / 1024),
+	}
+}
+
+// Getrusage implements linux syscall getrusage(2).
+//	marked "y" are supported now
+//	marked "*" are not used on Linux
+//	marked "p" are pending for support
+//
+//	y    struct timeval ru_utime; /* user CPU time used */
+//	y    struct timeval ru_stime; /* system CPU time used */
+//	p    long   ru_maxrss;        /* maximum resident set size */
+//	*    long   ru_ixrss;         /* integral shared memory size */
+//	*    long   ru_idrss;         /* integral unshared data size */
+//	*    long   ru_isrss;         /* integral unshared stack size */
+//	p    long   ru_minflt;        /* page reclaims (soft page faults) */
+//	p    long   ru_majflt;        /* page faults (hard page faults) */
+//	*    long   ru_nswap;         /* swaps */
+//	p    long   ru_inblock;       /* block input operations */
+//	p    long   ru_oublock;       /* block output operations */
+//	*    long   ru_msgsnd;        /* IPC messages sent */
+//	*    long   ru_msgrcv;        /* IPC messages received */
+//	*    long   ru_nsignals;      /* signals received */
+//	y    long   ru_nvcsw;         /* voluntary context switches */
+//	y    long   ru_nivcsw;        /* involuntary context switches */
+func Getrusage(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	which := args[0].Int()
+	addr := args[1].Pointer()
+
+	if which != linux.RUSAGE_SELF && which != linux.RUSAGE_CHILDREN && which != linux.RUSAGE_THREAD {
+		return 0, nil, syserror.EINVAL
+	}
+
+	ru := getrusage(t, which)
+	_, err := t.CopyOut(addr, &ru)
+	return 0, nil, err
+}
+
+// Times implements linux syscall times(2).
+func Times(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	// Calculate the ticks first, and figure out if any additional work is
+	// necessary. Linux allows for a NULL addr, in which case only the
+	// return value is meaningful. We don't need to do anything else.
+	ticks := uintptr(ktime.NowFromContext(t).Nanoseconds() / linux.ClockTick.Nanoseconds())
+	if addr == 0 {
+		return ticks, nil, nil
+	}
+
+	cs1 := t.ThreadGroup().CPUStats()
+	cs2 := t.ThreadGroup().JoinedChildCPUStats()
+	r := linux.Tms{
+		UTime:  linux.ClockTFromDuration(cs1.UserTime),
+		STime:  linux.ClockTFromDuration(cs1.SysTime),
+		CUTime: linux.ClockTFromDuration(cs2.UserTime),
+		CSTime: linux.ClockTFromDuration(cs2.SysTime),
+	}
+	if _, err := t.CopyOut(addr, &r); err != nil {
+		return 0, nil, err
+	}
+
+	return ticks, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_sched.go b/pkg/sentry/syscalls/linux/sys_sched.go
new file mode 100644
index 000000000..ff9e46077
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_sched.go
@@ -0,0 +1,100 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+const (
+	onlyScheduler = linux.SCHED_NORMAL
+	onlyPriority  = 0
+)
+
+// SchedParam replicates struct sched_param in sched.h.
+type SchedParam struct {
+	schedPriority int64
+}
+
+// SchedGetparam implements linux syscall sched_getparam(2).
+func SchedGetparam(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := args[0].Int()
+	param := args[1].Pointer()
+	if param == 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if pid < 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if pid != 0 && t.PIDNamespace().TaskWithID(kernel.ThreadID(pid)) == nil {
+		return 0, nil, syscall.ESRCH
+	}
+	r := SchedParam{schedPriority: onlyPriority}
+	if _, err := t.CopyOut(param, r); err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, nil
+}
+
+// SchedGetscheduler implements linux syscall sched_getscheduler(2).
+func SchedGetscheduler(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := args[0].Int()
+	if pid < 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if pid != 0 && t.PIDNamespace().TaskWithID(kernel.ThreadID(pid)) == nil {
+		return 0, nil, syscall.ESRCH
+	}
+	return onlyScheduler, nil, nil
+}
+
+// SchedSetscheduler implements linux syscall sched_setscheduler(2).
+func SchedSetscheduler(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := args[0].Int()
+	policy := args[1].Int()
+	param := args[2].Pointer()
+	if pid < 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if policy != onlyScheduler {
+		return 0, nil, syscall.EINVAL
+	}
+	if pid != 0 && t.PIDNamespace().TaskWithID(kernel.ThreadID(pid)) == nil {
+		return 0, nil, syscall.ESRCH
+	}
+	var r SchedParam
+	if _, err := t.CopyIn(param, &r); err != nil {
+		return 0, nil, syscall.EINVAL
+	}
+	if r.schedPriority != onlyPriority {
+		return 0, nil, syscall.EINVAL
+	}
+	return 0, nil, nil
+}
+
+// SchedGetPriorityMax implements linux syscall sched_get_priority_max(2).
+func SchedGetPriorityMax(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return onlyPriority, nil, nil
+}
+
+// SchedGetPriorityMin implements linux syscall sched_get_priority_min(2).
+func SchedGetPriorityMin(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return onlyPriority, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_sem.go b/pkg/sentry/syscalls/linux/sys_sem.go
new file mode 100644
index 000000000..a8983705b
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_sem.go
@@ -0,0 +1,166 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"math"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/auth"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const opsMax = 500 // SEMOPM
+
+// Semget handles: semget(key_t key, int nsems, int semflg)
+func Semget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	key := args[0].Int()
+	nsems := args[1].Int()
+	flag := args[2].Int()
+
+	private := key == linux.IPC_PRIVATE
+	create := flag&linux.IPC_CREAT == linux.IPC_CREAT
+	exclusive := flag&linux.IPC_EXCL == linux.IPC_EXCL
+	mode := linux.FileMode(flag & 0777)
+
+	r := t.IPCNamespace().SemaphoreRegistry()
+	set, err := r.FindOrCreate(t, key, nsems, mode, private, create, exclusive)
+	if err != nil {
+		return 0, nil, err
+	}
+	return uintptr(set.ID), nil, nil
+}
+
+// Semop handles: semop(int semid, struct sembuf *sops, size_t nsops)
+func Semop(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	id := args[0].Int()
+	sembufAddr := args[1].Pointer()
+	nsops := args[2].SizeT()
+
+	r := t.IPCNamespace().SemaphoreRegistry()
+	set := r.FindByID(id)
+	if set == nil {
+		return 0, nil, syserror.EINVAL
+	}
+	if nsops <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+	if nsops > opsMax {
+		return 0, nil, syserror.E2BIG
+	}
+
+	ops := make([]linux.Sembuf, nsops)
+	if _, err := t.CopyIn(sembufAddr, ops); err != nil {
+		return 0, nil, err
+	}
+
+	creds := auth.CredentialsFromContext(t)
+	for {
+		ch, num, err := set.ExecuteOps(t, ops, creds)
+		if ch == nil || err != nil {
+			// We're done (either on success or a failure).
+			return 0, nil, err
+		}
+		if err = t.Block(ch); err != nil {
+			set.AbortWait(num, ch)
+			return 0, nil, err
+		}
+	}
+}
+
+// Semctl handles: semctl(int semid, int semnum, int cmd, ...)
+func Semctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	id := args[0].Int()
+	num := args[1].Int()
+	cmd := args[2].Int()
+
+	switch cmd {
+	case linux.SETVAL:
+		val := args[3].Int()
+		if val > math.MaxInt16 {
+			return 0, nil, syserror.ERANGE
+		}
+		return 0, nil, setVal(t, id, num, int16(val))
+
+	case linux.GETVAL:
+		v, err := getVal(t, id, num)
+		return uintptr(v), nil, err
+
+	case linux.IPC_RMID:
+		return 0, nil, remove(t, id)
+
+	case linux.IPC_SET:
+		arg := args[3].Pointer()
+		s := linux.SemidDS{}
+		if _, err := t.CopyIn(arg, &s); err != nil {
+			return 0, nil, err
+		}
+
+		perms := fs.FilePermsFromMode(linux.FileMode(s.SemPerm.Mode & 0777))
+		return 0, nil, ipcSet(t, id, auth.UID(s.SemPerm.UID), auth.GID(s.SemPerm.GID), perms)
+
+	default:
+		return 0, nil, syserror.EINVAL
+	}
+}
+
+func remove(t *kernel.Task, id int32) error {
+	r := t.IPCNamespace().SemaphoreRegistry()
+	creds := auth.CredentialsFromContext(t)
+	return r.RemoveID(id, creds)
+}
+
+func ipcSet(t *kernel.Task, id int32, uid auth.UID, gid auth.GID, perms fs.FilePermissions) error {
+	r := t.IPCNamespace().SemaphoreRegistry()
+	set := r.FindByID(id)
+	if set == nil {
+		return syserror.EINVAL
+	}
+
+	creds := auth.CredentialsFromContext(t)
+	kuid := creds.UserNamespace.MapToKUID(uid)
+	if !kuid.Ok() {
+		return syserror.EINVAL
+	}
+	kgid := creds.UserNamespace.MapToKGID(gid)
+	if !kgid.Ok() {
+		return syserror.EINVAL
+	}
+	owner := fs.FileOwner{UID: kuid, GID: kgid}
+	return set.Change(t, creds, owner, perms)
+}
+
+func setVal(t *kernel.Task, id int32, num int32, val int16) error {
+	r := t.IPCNamespace().SemaphoreRegistry()
+	set := r.FindByID(id)
+	if set == nil {
+		return syserror.EINVAL
+	}
+	creds := auth.CredentialsFromContext(t)
+	return set.SetVal(t, num, val, creds)
+}
+
+func getVal(t *kernel.Task, id int32, num int32) (int16, error) {
+	r := t.IPCNamespace().SemaphoreRegistry()
+	set := r.FindByID(id)
+	if set == nil {
+		return 0, syserror.EINVAL
+	}
+	creds := auth.CredentialsFromContext(t)
+	return set.GetVal(num, creds)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_signal.go b/pkg/sentry/syscalls/linux/sys_signal.go
new file mode 100644
index 000000000..93b3f531a
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_signal.go
@@ -0,0 +1,553 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"math"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// "For a process to have permission to send a signal it must
+// - either be privileged (CAP_KILL), or
+// - the real or effective user ID of the sending process must be equal to the
+// real or saved set-user-ID of the target process.
+//
+// In the case of SIGCONT it suffices when the sending and receiving processes
+// belong to the same session." - kill(2)
+//
+// Equivalent to kernel/signal.c:check_kill_permission.
+func mayKill(t *kernel.Task, target *kernel.Task, sig linux.Signal) bool {
+	// kernel/signal.c:check_kill_permission also allows a signal if the
+	// sending and receiving tasks share a thread group, which is not
+	// mentioned in kill(2) since kill does not allow task-level
+	// granularity in signal sending.
+	if t.ThreadGroup() == target.ThreadGroup() {
+		return true
+	}
+
+	if t.HasCapabilityIn(linux.CAP_KILL, target.UserNamespace()) {
+		return true
+	}
+
+	creds := t.Credentials()
+	tcreds := target.Credentials()
+	if creds.EffectiveKUID == tcreds.SavedKUID ||
+		creds.EffectiveKUID == tcreds.RealKUID ||
+		creds.RealKUID == tcreds.SavedKUID ||
+		creds.RealKUID == tcreds.RealKUID {
+		return true
+	}
+
+	if sig == linux.SIGCONT && target.ThreadGroup().Session() == t.ThreadGroup().Session() {
+		return true
+	}
+	return false
+}
+
+// Kill implements linux syscall kill(2).
+func Kill(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := kernel.ThreadID(args[0].Int())
+	sig := linux.Signal(args[1].Int())
+
+	switch {
+	case pid > 0:
+		// "If pid is positive, then signal sig is sent to the process with the
+		// ID specified by pid." - kill(2)
+		// This loops to handle races with execve where target dies between
+		// TaskWithID and SendGroupSignal. Compare Linux's
+		// kernel/signal.c:kill_pid_info().
+		for {
+			target := t.PIDNamespace().TaskWithID(pid)
+			if target == nil {
+				return 0, nil, syserror.ESRCH
+			}
+			if !mayKill(t, target, sig) {
+				return 0, nil, syserror.EPERM
+			}
+			info := &arch.SignalInfo{
+				Signo: int32(sig),
+				Code:  arch.SignalInfoUser,
+			}
+			info.SetPid(int32(target.PIDNamespace().IDOfTask(t)))
+			info.SetUid(int32(t.Credentials().RealKUID.In(target.UserNamespace()).OrOverflow()))
+			if err := target.SendGroupSignal(info); err != syserror.ESRCH {
+				return 0, nil, err
+			}
+		}
+	case pid == -1:
+		// "If pid equals -1, then sig is sent to every process for which the
+		// calling process has permission to send signals, except for process 1
+		// (init), but see below. ... POSIX.1-2001 requires that kill(-1,sig)
+		// send sig to all processes that the calling process may send signals
+		// to, except possibly for some implementation-defined system
+		// processes. Linux allows a process to signal itself, but on Linux the
+		// call kill(-1,sig) does not signal the calling process."
+		var (
+			lastErr   error
+			delivered int
+		)
+		for _, tg := range t.PIDNamespace().ThreadGroups() {
+			if tg == t.ThreadGroup() {
+				continue
+			}
+			if t.PIDNamespace().IDOfThreadGroup(tg) == kernel.InitTID {
+				continue
+			}
+
+			// If pid == -1, the returned error is the last non-EPERM error
+			// from any call to group_send_sig_info.
+			if !mayKill(t, tg.Leader(), sig) {
+				continue
+			}
+			// Here and below, whether or not kill returns an error may
+			// depend on the iteration order. We at least implement the
+			// semantics documented by the man page: "On success (at least
+			// one signal was sent), zero is returned."
+			info := &arch.SignalInfo{
+				Signo: int32(sig),
+				Code:  arch.SignalInfoUser,
+			}
+			info.SetPid(int32(tg.PIDNamespace().IDOfTask(t)))
+			info.SetUid(int32(t.Credentials().RealKUID.In(tg.Leader().UserNamespace()).OrOverflow()))
+			err := tg.SendSignal(info)
+			if err == syserror.ESRCH {
+				// ESRCH is ignored because it means the task
+				// exited while we were iterating.  This is a
+				// race which would not normally exist on
+				// Linux, so we suppress it.
+				continue
+			}
+			delivered++
+			if err != nil {
+				lastErr = err
+			}
+		}
+		if delivered > 0 {
+			return 0, nil, lastErr
+		}
+		return 0, nil, syserror.ESRCH
+	default:
+		// "If pid equals 0, then sig is sent to every process in the process
+		// group of the calling process."
+		//
+		// "If pid is less than -1, then sig is sent to every process
+		// in the process group whose ID is -pid."
+		pgid := kernel.ProcessGroupID(-pid)
+		if pgid == 0 {
+			pgid = t.PIDNamespace().IDOfProcessGroup(t.ThreadGroup().ProcessGroup())
+		}
+
+		// If pid != -1 (i.e. signalling a process group), the returned error
+		// is the last error from any call to group_send_sig_info.
+		lastErr := syserror.ESRCH
+		for _, tg := range t.PIDNamespace().ThreadGroups() {
+			if t.PIDNamespace().IDOfProcessGroup(tg.ProcessGroup()) == pgid {
+				if !mayKill(t, tg.Leader(), sig) {
+					lastErr = syserror.EPERM
+					continue
+				}
+
+				info := &arch.SignalInfo{
+					Signo: int32(sig),
+					Code:  arch.SignalInfoUser,
+				}
+				info.SetPid(int32(tg.PIDNamespace().IDOfTask(t)))
+				info.SetUid(int32(t.Credentials().RealKUID.In(tg.Leader().UserNamespace()).OrOverflow()))
+				// See note above regarding ESRCH race above.
+				if err := tg.SendSignal(info); err != syserror.ESRCH {
+					lastErr = err
+				}
+			}
+		}
+
+		return 0, nil, lastErr
+	}
+}
+
+func tkillSigInfo(sender, receiver *kernel.Task, sig linux.Signal) *arch.SignalInfo {
+	info := &arch.SignalInfo{
+		Signo: int32(sig),
+		Code:  arch.SignalInfoTkill,
+	}
+	info.SetPid(int32(receiver.PIDNamespace().IDOfThreadGroup(sender.ThreadGroup())))
+	info.SetUid(int32(sender.Credentials().RealKUID.In(receiver.UserNamespace()).OrOverflow()))
+	return info
+}
+
+// Tkill implements linux syscall tkill(2).
+func Tkill(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := kernel.ThreadID(args[0].Int())
+	sig := linux.Signal(args[1].Int())
+
+	// N.B. Inconsistent with man page, linux actually rejects calls with
+	// tid <=0 by EINVAL. This isn't the same for all signal calls.
+	if tid <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	target := t.PIDNamespace().TaskWithID(tid)
+	if target == nil {
+		return 0, nil, syserror.ESRCH
+	}
+
+	if !mayKill(t, target, sig) {
+		return 0, nil, syserror.EPERM
+	}
+	return 0, nil, target.SendSignal(tkillSigInfo(t, target, sig))
+}
+
+// Tgkill implements linux syscall tgkill(2).
+func Tgkill(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tgid := kernel.ThreadID(args[0].Int())
+	tid := kernel.ThreadID(args[1].Int())
+	sig := linux.Signal(args[2].Int())
+
+	// N.B. Inconsistent with man page, linux actually rejects calls with
+	// tgid/tid <=0 by EINVAL. This isn't the same for all signal calls.
+	if tgid <= 0 || tid <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	targetTG := t.PIDNamespace().ThreadGroupWithID(tgid)
+	target := t.PIDNamespace().TaskWithID(tid)
+	if targetTG == nil || target == nil || target.ThreadGroup() != targetTG {
+		return 0, nil, syserror.ESRCH
+	}
+
+	if !mayKill(t, target, sig) {
+		return 0, nil, syserror.EPERM
+	}
+	return 0, nil, target.SendSignal(tkillSigInfo(t, target, sig))
+}
+
+// RtSigaction implements linux syscall rt_sigaction(2).
+func RtSigaction(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	sig := linux.Signal(args[0].Int())
+	newactarg := args[1].Pointer()
+	oldactarg := args[2].Pointer()
+
+	var newactptr *arch.SignalAct
+	if newactarg != 0 {
+		newact, err := t.CopyInSignalAct(newactarg)
+		if err != nil {
+			return 0, nil, err
+		}
+		newactptr = &newact
+	}
+	oldact, err := t.ThreadGroup().SetSignalAct(sig, newactptr)
+	if err != nil {
+		return 0, nil, err
+	}
+	if oldactarg != 0 {
+		if err := t.CopyOutSignalAct(oldactarg, &oldact); err != nil {
+			return 0, nil, err
+		}
+	}
+	return 0, nil, nil
+}
+
+// Sigreturn implements linux syscall sigreturn(2).
+func Sigreturn(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	ctrl, err := t.SignalReturn(false)
+	return 0, ctrl, err
+}
+
+// RtSigreturn implements linux syscall rt_sigreturn(2).
+func RtSigreturn(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	ctrl, err := t.SignalReturn(true)
+	return 0, ctrl, err
+}
+
+// RtSigprocmask implements linux syscall rt_sigprocmask(2).
+func RtSigprocmask(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	how := args[0].Int()
+	setaddr := args[1].Pointer()
+	oldaddr := args[2].Pointer()
+	sigsetsize := args[3].SizeT()
+
+	if sigsetsize != linux.SignalSetSize {
+		return 0, nil, syserror.EINVAL
+	}
+	oldmask := t.SignalMask()
+	if setaddr != 0 {
+		mask, err := copyInSigSet(t, setaddr, sigsetsize)
+		if err != nil {
+			return 0, nil, err
+		}
+
+		switch how {
+		case linux.SIG_BLOCK:
+			t.SetSignalMask(oldmask | mask)
+		case linux.SIG_UNBLOCK:
+			t.SetSignalMask(oldmask &^ mask)
+		case linux.SIG_SETMASK:
+			t.SetSignalMask(mask)
+		default:
+			return 0, nil, syserror.EINVAL
+		}
+	}
+	if oldaddr != 0 {
+		return 0, nil, copyOutSigSet(t, oldaddr, oldmask)
+	}
+
+	return 0, nil, nil
+}
+
+// Sigaltstack implements linux syscall sigaltstack(2).
+func Sigaltstack(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	setaddr := args[0].Pointer()
+	oldaddr := args[1].Pointer()
+
+	if oldaddr != 0 {
+		alt := t.SignalStack()
+		if t.OnSignalStack(alt) {
+			alt.Flags |= arch.SignalStackFlagOnStack
+		}
+		if err := t.CopyOutSignalStack(oldaddr, &alt); err != nil {
+			return 0, nil, err
+		}
+	}
+	if setaddr != 0 {
+		if t.OnSignalStack(t.SignalStack()) {
+			return 0, nil, syserror.EPERM
+		}
+		alt, err := t.CopyInSignalStack(setaddr)
+		if err != nil {
+			return 0, nil, err
+		}
+		if err := t.SetSignalStack(alt); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	return 0, nil, nil
+}
+
+// Pause implements linux syscall pause(2).
+func Pause(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, syserror.ConvertIntr(t.Block(nil), kernel.ERESTARTNOHAND)
+}
+
+func sigtimedwait(t *kernel.Task, mask linux.SignalSet, timeout time.Duration) (*arch.SignalInfo, error) {
+	// Is it already pending?
+	if info := t.TakeSignal(^mask); info != nil {
+		return info, nil
+	}
+
+	// No signals available immediately and asked not to wait.
+	if timeout == 0 {
+		return nil, syserror.EAGAIN
+	}
+
+	// No signals available yet. Temporarily unblock the ones we are interested
+	// in then wait for either a timeout or a new signal.
+	oldmask := t.SignalMask()
+	t.SetSignalMask(oldmask &^ mask)
+	_, err := t.BlockWithTimeout(nil, true, timeout)
+	t.SetSignalMask(oldmask)
+
+	// How did the wait go?
+	switch err {
+	case syserror.ErrInterrupted:
+		if info := t.TakeSignal(^mask); info != nil {
+			// Got one of the signals we were waiting for.
+			return info, nil
+		}
+		// Got a signal we weren't waiting for.
+		return nil, syserror.EINTR
+	case syserror.ETIMEDOUT:
+		// Timed out and still no signals.
+		return nil, syserror.EAGAIN
+	default:
+		// Some other error? Shouldn't be possible. The event channel
+		// passed to BlockWithTimeout was nil, so the only two ways the
+		// block could've ended are a timeout or an interrupt.
+		panic("unreachable")
+	}
+}
+
+// RtSigpending implements linux syscall rt_sigpending(2).
+func RtSigpending(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	pending := t.PendingSignals()
+	_, err := t.CopyOut(addr, pending)
+	return 0, nil, err
+}
+
+// RtSigtimedwait implements linux syscall rt_sigtimedwait(2).
+func RtSigtimedwait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	sigset := args[0].Pointer()
+	siginfo := args[1].Pointer()
+	timespec := args[2].Pointer()
+	sigsetsize := args[3].SizeT()
+
+	mask, err := copyInSigSet(t, sigset, sigsetsize)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	var timeout time.Duration
+	if timespec != 0 {
+		d, err := copyTimespecIn(t, timespec)
+		if err != nil {
+			return 0, nil, err
+		}
+		if !d.Valid() {
+			return 0, nil, syserror.EINVAL
+		}
+		timeout = time.Duration(d.ToNsecCapped())
+	} else {
+		timeout = time.Duration(math.MaxInt64)
+	}
+
+	si, err := sigtimedwait(t, mask, timeout)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if si != nil {
+		if siginfo != 0 {
+			si.FixSignalCodeForUser()
+			if _, err := t.CopyOut(siginfo, si); err != nil {
+				return 0, nil, err
+			}
+		}
+		return uintptr(si.Signo), nil, nil
+	}
+
+	// sigtimedwait's not supposed to return nil si and err...
+	return 0, nil, nil
+}
+
+// RtSigqueueinfo implements linux syscall rt_sigqueueinfo(2).
+func RtSigqueueinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := kernel.ThreadID(args[0].Int())
+	sig := linux.Signal(args[1].Int())
+	infoAddr := args[2].Pointer()
+
+	// Copy in the info.
+	//
+	// We must ensure that the Signo is set (Linux overrides this in the
+	// same way), and that the code is in the allowed set. This same logic
+	// appears below in RtSigtgqueueinfo and should be kept in sync.
+	var info arch.SignalInfo
+	if _, err := t.CopyIn(infoAddr, &info); err != nil {
+		return 0, nil, err
+	}
+	info.Signo = int32(sig)
+
+	// This must loop to handle the race with execve described in Kill.
+	for {
+		// Deliver to the given task's thread group.
+		target := t.PIDNamespace().TaskWithID(pid)
+		if target == nil {
+			return 0, nil, syserror.ESRCH
+		}
+
+		// If the sender is not the receiver, it can't use si_codes used by the
+		// kernel or SI_TKILL.
+		if (info.Code >= 0 || info.Code == arch.SignalInfoTkill) && target != t {
+			return 0, nil, syserror.EPERM
+		}
+
+		if !mayKill(t, target, sig) {
+			return 0, nil, syserror.EPERM
+		}
+
+		if err := target.SendGroupSignal(&info); err != syserror.ESRCH {
+			return 0, nil, err
+		}
+	}
+}
+
+// RtTgsigqueueinfo implements linux syscall rt_tgsigqueueinfo(2).
+func RtTgsigqueueinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tgid := kernel.ThreadID(args[0].Int())
+	tid := kernel.ThreadID(args[1].Int())
+	sig := linux.Signal(args[2].Int())
+	infoAddr := args[3].Pointer()
+
+	// N.B. Inconsistent with man page, linux actually rejects calls with
+	// tgid/tid <=0 by EINVAL. This isn't the same for all signal calls.
+	if tgid <= 0 || tid <= 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Copy in the info. See RtSigqueueinfo above.
+	var info arch.SignalInfo
+	if _, err := t.CopyIn(infoAddr, &info); err != nil {
+		return 0, nil, err
+	}
+	info.Signo = int32(sig)
+
+	// Deliver to the given task.
+	targetTG := t.PIDNamespace().ThreadGroupWithID(tgid)
+	target := t.PIDNamespace().TaskWithID(tid)
+	if targetTG == nil || target == nil || target.ThreadGroup() != targetTG {
+		return 0, nil, syserror.ESRCH
+	}
+
+	// If the sender is not the receiver, it can't use si_codes used by the
+	// kernel or SI_TKILL.
+	if (info.Code >= 0 || info.Code == arch.SignalInfoTkill) && target != t {
+		return 0, nil, syserror.EPERM
+	}
+
+	if !mayKill(t, target, sig) {
+		return 0, nil, syserror.EPERM
+	}
+	return 0, nil, target.SendSignal(&info)
+}
+
+// RtSigsuspend implements linux syscall rt_sigsuspend(2).
+func RtSigsuspend(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	sigset := args[0].Pointer()
+
+	// Copy in the signal mask.
+	var mask linux.SignalSet
+	if _, err := t.CopyIn(sigset, &mask); err != nil {
+		return 0, nil, err
+	}
+	mask &^= kernel.UnblockableSignals
+
+	// Swap the mask.
+	oldmask := t.SignalMask()
+	t.SetSignalMask(mask)
+	t.SetSavedSignalMask(oldmask)
+
+	// Perform the wait.
+	return 0, nil, syserror.ConvertIntr(t.Block(nil), kernel.ERESTARTNOHAND)
+}
+
+// RestartSyscall implements the linux syscall restart_syscall(2).
+func RestartSyscall(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	if r := t.SyscallRestartBlock(); r != nil {
+		n, err := r.Restart(t)
+		return n, nil, err
+	}
+	// The restart block should never be nil here, but it's possible
+	// ERESTART_RESTARTBLOCK was set by ptrace without the current syscall
+	// setting up a restart block. If ptrace didn't manipulate the return value,
+	// finding a nil restart block is a bug. Linux ensures that the restart
+	// function is never null by (re)initializing it with one that translates
+	// the restart into EINTR. We'll emulate that behaviour.
+	t.Debugf("Restart block missing in restart_syscall(2). Did ptrace inject a return value of ERESTART_RESTARTBLOCK?")
+	return 0, nil, syserror.EINTR
+}
diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go
new file mode 100644
index 000000000..3797c0a5d
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_socket.go
@@ -0,0 +1,1059 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/socket/control"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/tcpip/transport/unix"
+)
+
+// minListenBacklog is the minimum reasonable backlog for listening sockets.
+const minListenBacklog = 8
+
+// maxListenBacklog is the maximum allowed backlog for listening sockets.
+const maxListenBacklog = 1024
+
+// maxAddrLen is the maximum socket address length we're willing to accept.
+const maxAddrLen = 200
+
+// maxOptLen is the maximum sockopt parameter length we're willing to accept.
+const maxOptLen = 1024
+
+// maxControlLen is the maximum length of the msghdr.msg_control buffer we're
+// willing to accept. Note that this limit is smaller than Linux, which allows
+// buffers upto INT_MAX.
+const maxControlLen = 10 * 1024 * 1024
+
+// nameLenOffset is the offset from the start of the MessageHeader64 struct to
+// the NameLen field.
+const nameLenOffset = 8
+
+// controlLenOffset is the offset form the start of the MessageHeader64 struct
+// to the ControlLen field.
+const controlLenOffset = 40
+
+// messageHeader64Len is the length of a MessageHeader64 struct.
+var messageHeader64Len = uint64(binary.Size(MessageHeader64{}))
+
+// multipleMessageHeader64Len is the length of a multipeMessageHeader64 struct.
+var multipleMessageHeader64Len = uint64(binary.Size(multipleMessageHeader64{}))
+
+// MessageHeader64 is the 64-bit representation of the msghdr struct used in
+// the recvmsg and sendmsg syscalls.
+type MessageHeader64 struct {
+	// Name is the optional pointer to a network address buffer.
+	Name uint64
+
+	// NameLen is the length of the buffer pointed to by Name.
+	NameLen uint32
+	_       uint32
+
+	// Iov is a pointer to an array of io vectors that describe the memory
+	// locations involved in the io operation.
+	Iov uint64
+
+	// IovLen is the length of the array pointed to by Iov.
+	IovLen uint64
+
+	// Control is the optional pointer to ancillary control data.
+	Control uint64
+
+	// ControlLen is the length of the data pointed to by Control.
+	ControlLen uint64
+
+	// Flags on the sent/received message.
+	Flags int32
+	_     int32
+}
+
+// multipleMessageHeader64 is the 64-bit representation of the mmsghdr struct used in
+// the recvmmsg and sendmmsg syscalls.
+type multipleMessageHeader64 struct {
+	msgHdr MessageHeader64
+	msgLen uint32
+	_      int32
+}
+
+// CopyInMessageHeader64 copies a message header from user to kernel memory.
+func CopyInMessageHeader64(t *kernel.Task, addr usermem.Addr, msg *MessageHeader64) error {
+	b := t.CopyScratchBuffer(52)
+	if _, err := t.CopyInBytes(addr, b); err != nil {
+		return err
+	}
+
+	msg.Name = usermem.ByteOrder.Uint64(b[0:])
+	msg.NameLen = usermem.ByteOrder.Uint32(b[8:])
+	msg.Iov = usermem.ByteOrder.Uint64(b[16:])
+	msg.IovLen = usermem.ByteOrder.Uint64(b[24:])
+	msg.Control = usermem.ByteOrder.Uint64(b[32:])
+	msg.ControlLen = usermem.ByteOrder.Uint64(b[40:])
+	msg.Flags = int32(usermem.ByteOrder.Uint32(b[48:]))
+
+	return nil
+}
+
+// CaptureAddress allocates memory for and copies a socket address structure
+// from the untrusted address space range.
+func CaptureAddress(t *kernel.Task, addr usermem.Addr, addrlen uint32) ([]byte, error) {
+	if addrlen > maxAddrLen {
+		return nil, syscall.EINVAL
+	}
+
+	addrBuf := make([]byte, addrlen)
+	if _, err := t.CopyInBytes(addr, addrBuf); err != nil {
+		return nil, err
+	}
+
+	return addrBuf, nil
+}
+
+// writeAddress writes a sockaddr structure and its length to an output buffer
+// in the unstrusted address space range. If the address is bigger than the
+// buffer, it is truncated.
+func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr usermem.Addr, addrLenPtr usermem.Addr) error {
+	// Get the buffer length.
+	var bufLen uint32
+	if _, err := t.CopyIn(addrLenPtr, &bufLen); err != nil {
+		return err
+	}
+
+	if int32(bufLen) < 0 {
+		return syscall.EINVAL
+	}
+
+	// Write the length unconditionally.
+	if _, err := t.CopyOut(addrLenPtr, addrLen); err != nil {
+		return err
+	}
+
+	if addr == nil {
+		return nil
+	}
+
+	if bufLen > addrLen {
+		bufLen = addrLen
+	}
+
+	// Copy as much of the address as will fit in the buffer.
+	encodedAddr := binary.Marshal(nil, usermem.ByteOrder, addr)
+	if bufLen > uint32(len(encodedAddr)) {
+		bufLen = uint32(len(encodedAddr))
+	}
+	_, err := t.CopyOutBytes(addrPtr, encodedAddr[:int(bufLen)])
+	return err
+}
+
+// Socket implements the linux syscall socket(2).
+func Socket(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	domain := int(args[0].Int())
+	stype := args[1].Int()
+	protocol := int(args[2].Int())
+
+	// Check and initialize the flags.
+	if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Create the new socket.
+	s, e := socket.New(t, domain, unix.SockType(stype&0xf), protocol)
+	if e != nil {
+		return 0, nil, e.ToError()
+	}
+	s.SetFlags(fs.SettableFileFlags{
+		NonBlocking: stype&linux.SOCK_NONBLOCK != 0,
+	})
+	defer s.DecRef()
+
+	fd, err := t.FDMap().NewFDFrom(0, s, kernel.FDFlags{
+		CloseOnExec: stype&linux.SOCK_CLOEXEC != 0,
+	}, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// SocketPair implements the linux syscall socketpair(2).
+func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	domain := int(args[0].Int())
+	stype := args[1].Int()
+	protocol := int(args[2].Int())
+	socks := args[3].Pointer()
+
+	// Check and initialize the flags.
+	if stype & ^(0xf|linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	fileFlags := fs.SettableFileFlags{
+		NonBlocking: stype&linux.SOCK_NONBLOCK != 0,
+	}
+	fdFlags := kernel.FDFlags{
+		CloseOnExec: stype&linux.SOCK_CLOEXEC != 0,
+	}
+
+	// Create the socket pair.
+	s1, s2, e := socket.Pair(t, domain, unix.SockType(stype&0xf), protocol)
+	if e != nil {
+		return 0, nil, e.ToError()
+	}
+	s1.SetFlags(fileFlags)
+	s2.SetFlags(fileFlags)
+	defer s1.DecRef()
+	defer s2.DecRef()
+
+	// Create the FDs for the sockets.
+	fd1, err := t.FDMap().NewFDFrom(0, s1, fdFlags, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, err
+	}
+	fd2, err := t.FDMap().NewFDFrom(0, s2, fdFlags, t.ThreadGroup().Limits())
+	if err != nil {
+		t.FDMap().Remove(fd1)
+		return 0, nil, err
+	}
+
+	// Copy the file descriptors out.
+	if _, err := t.CopyOut(socks, []int32{int32(fd1), int32(fd2)}); err != nil {
+		t.FDMap().Remove(fd1)
+		t.FDMap().Remove(fd2)
+		return 0, nil, err
+	}
+
+	return 0, nil, nil
+}
+
+// Connect implements the linux syscall connect(2).
+func Connect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Uint()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Capture address and call syscall implementation.
+	a, err := CaptureAddress(t, addr, addrlen)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	blocking := !file.Flags().NonBlocking
+	return 0, nil, syserror.ConvertIntr(s.Connect(t, a, blocking).ToError(), kernel.ERESTARTSYS)
+}
+
+// accept is the implementation of the accept syscall. It is called by accept
+// and accept4 syscall handlers.
+func accept(t *kernel.Task, fd kdefs.FD, addr usermem.Addr, addrLen usermem.Addr, flags int) (uintptr, error) {
+	// Check that no unsupported flags are passed in.
+	if flags & ^(linux.SOCK_NONBLOCK|linux.SOCK_CLOEXEC) != 0 {
+		return 0, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, syscall.ENOTSOCK
+	}
+
+	// Call the syscall implementation for this socket, then copy the
+	// output address if one is specified.
+	blocking := !file.Flags().NonBlocking
+
+	peerRequested := addrLen != 0
+	nfd, peer, peerLen, e := s.Accept(t, peerRequested, flags, blocking)
+	if e != nil {
+		return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS)
+	}
+	if peerRequested {
+		// NOTE: Linux does not give you an error if it can't
+		// write the data back out so neither do we.
+		if err := writeAddress(t, peer, peerLen, addr, addrLen); err == syscall.EINVAL {
+			return 0, err
+		}
+	}
+	return uintptr(nfd), nil
+}
+
+// Accept4 implements the linux syscall accept4(2).
+func Accept4(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Pointer()
+	flags := int(args[3].Int())
+
+	n, err := accept(t, fd, addr, addrlen, flags)
+	return n, nil, err
+}
+
+// Accept implements the linux syscall accept(2).
+func Accept(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Pointer()
+
+	n, err := accept(t, fd, addr, addrlen, 0)
+	return n, nil, err
+}
+
+// Bind implements the linux syscall bind(2).
+func Bind(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Uint()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Capture address and call syscall implementation.
+	a, err := CaptureAddress(t, addr, addrlen)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, s.Bind(t, a).ToError()
+}
+
+// Listen implements the linux syscall listen(2).
+func Listen(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	backlog := args[1].Int()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Per Linux, the backlog is silently capped to reasonable values.
+	if backlog <= 0 {
+		backlog = minListenBacklog
+	}
+	if backlog > maxListenBacklog {
+		backlog = maxListenBacklog
+	}
+
+	return 0, nil, s.Listen(t, int(backlog)).ToError()
+}
+
+// Shutdown implements the linux syscall shutdown(2).
+func Shutdown(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	how := args[1].Int()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Validate how, then call syscall implementation.
+	switch how {
+	case linux.SHUT_RD, linux.SHUT_WR, linux.SHUT_RDWR:
+	default:
+		return 0, nil, syscall.EINVAL
+	}
+
+	return 0, nil, s.Shutdown(t, int(how)).ToError()
+}
+
+// GetSockOpt implements the linux syscall getsockopt(2).
+func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	level := args[1].Int()
+	name := args[2].Int()
+	optValAddr := args[3].Pointer()
+	optLenAddr := args[4].Pointer()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Read the length if present. Reject negative values.
+	optLen := int32(0)
+	if optLenAddr != 0 {
+		if _, err := t.CopyIn(optLenAddr, &optLen); err != nil {
+			return 0, nil, err
+		}
+
+		if optLen < 0 {
+			return 0, nil, syscall.EINVAL
+		}
+	}
+
+	// Call syscall implementation then copy both value and value len out.
+	v, e := s.GetSockOpt(t, int(level), int(name), int(optLen))
+	if e != nil {
+		return 0, nil, e.ToError()
+	}
+
+	if optLenAddr != 0 {
+		vLen := int32(binary.Size(v))
+		if _, err := t.CopyOut(optLenAddr, vLen); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	if v != nil {
+		if _, err := t.CopyOut(optValAddr, v); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	return 0, nil, nil
+}
+
+// SetSockOpt implements the linux syscall setsockopt(2).
+//
+// Note that unlike Linux, enabling SO_PASSCRED does not autobind the socket.
+func SetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	level := args[1].Int()
+	name := args[2].Int()
+	optValAddr := args[3].Pointer()
+	optLen := args[4].Int()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	if optLen <= 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if optLen > maxOptLen {
+		return 0, nil, syscall.EINVAL
+	}
+	buf := make([]byte, optLen)
+	if _, err := t.CopyIn(optValAddr, &buf); err != nil {
+		return 0, nil, err
+	}
+
+	// Call syscall implementation.
+	if err := s.SetSockOpt(t, int(level), int(name), buf); err != nil {
+		return 0, nil, err.ToError()
+	}
+
+	return 0, nil, nil
+}
+
+// GetSockName implements the linux syscall getsockname(2).
+func GetSockName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Pointer()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Get the socket name and copy it to the caller.
+	v, vl, err := s.GetSockName(t)
+	if err != nil {
+		return 0, nil, err.ToError()
+	}
+
+	return 0, nil, writeAddress(t, v, vl, addr, addrlen)
+}
+
+// GetPeerName implements the linux syscall getpeername(2).
+func GetPeerName(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	addrlen := args[2].Pointer()
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Get the socket peer name and copy it to the caller.
+	v, vl, err := s.GetPeerName(t)
+	if err != nil {
+		return 0, nil, err.ToError()
+	}
+
+	return 0, nil, writeAddress(t, v, vl, addr, addrlen)
+}
+
+// RecvMsg implements the linux syscall recvmsg(2).
+func RecvMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	msgPtr := args[1].Pointer()
+	flags := args[2].Int()
+
+	if t.Arch().Width() != 8 {
+		// We only handle 64-bit for now.
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Reject flags that we don't handle yet.
+	if flags & ^(linux.MSG_DONTWAIT|linux.MSG_NOSIGNAL|linux.MSG_PEEK|linux.MSG_TRUNC|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	n, err := recvSingleMsg(t, s, msgPtr, flags, false, ktime.Time{})
+	return n, nil, err
+}
+
+// RecvMMsg implements the linux syscall recvmmsg(2).
+func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	msgPtr := args[1].Pointer()
+	vlen := args[2].Uint()
+	flags := args[3].Int()
+	toPtr := args[4].Pointer()
+
+	if t.Arch().Width() != 8 {
+		// We only handle 64-bit for now.
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Reject flags that we don't handle yet.
+	if flags & ^(linux.MSG_DONTWAIT|linux.MSG_NOSIGNAL|linux.MSG_TRUNC|linux.MSG_CMSG_CLOEXEC|linux.MSG_ERRQUEUE) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	var haveDeadline bool
+	var deadline ktime.Time
+	if toPtr != 0 {
+		ts, err := copyTimespecIn(t, toPtr)
+		if err != nil {
+			return 0, nil, err
+		}
+		if !ts.Valid() {
+			return 0, nil, syscall.EINVAL
+		}
+		deadline = t.Kernel().MonotonicClock().Now().Add(ts.ToDuration())
+		haveDeadline = true
+	}
+
+	if !haveDeadline {
+		dl := s.RecvTimeout()
+		if dl != 0 {
+			deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond)
+			haveDeadline = true
+		}
+	}
+
+	var count uint32
+	var err error
+	for i := uint64(0); i < uint64(vlen); i++ {
+		mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len)
+		if !ok {
+			return 0, nil, syscall.EFAULT
+		}
+		var n uintptr
+		if n, err = recvSingleMsg(t, s, mp, flags, haveDeadline, deadline); err != nil {
+			break
+		}
+
+		// Copy the received length to the caller.
+		lp, ok := mp.AddLength(messageHeader64Len)
+		if !ok {
+			return 0, nil, syscall.EFAULT
+		}
+		if _, err = t.CopyOut(lp, uint32(n)); err != nil {
+			break
+		}
+		count++
+	}
+
+	if count == 0 {
+		return 0, nil, err
+	}
+	return uintptr(count), nil, nil
+}
+
+func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags int32, haveDeadline bool, deadline ktime.Time) (uintptr, error) {
+	// Capture the message header and io vectors.
+	var msg MessageHeader64
+	if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil {
+		return 0, err
+	}
+
+	if msg.IovLen > linux.UIO_MAXIOV {
+		return 0, syscall.EMSGSIZE
+	}
+	dst, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, err
+	}
+
+	// FIXME: Pretend we have an empty error queue.
+	if flags&linux.MSG_ERRQUEUE != 0 {
+		return 0, syscall.EAGAIN
+	}
+
+	// Fast path when no control message nor name buffers are provided.
+	if msg.ControlLen == 0 && msg.NameLen == 0 {
+		n, _, _, _, err := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, false, 0)
+		if err != nil {
+			return 0, syserror.ConvertIntr(err.ToError(), kernel.ERESTARTSYS)
+		}
+		return uintptr(n), nil
+	}
+
+	if msg.ControlLen > maxControlLen {
+		return 0, syscall.ENOBUFS
+	}
+	n, sender, senderLen, cms, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, msg.NameLen != 0, msg.ControlLen)
+	if e != nil {
+		return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS)
+	}
+	defer cms.Release()
+
+	controlData := make([]byte, 0, msg.ControlLen)
+
+	if cr, ok := s.(unix.Credentialer); ok && cr.Passcred() {
+		creds, _ := cms.Credentials.(control.SCMCredentials)
+		controlData = control.PackCredentials(t, creds, controlData)
+	}
+
+	if cms.Rights != nil {
+		controlData = control.PackRights(t, cms.Rights.(control.SCMRights), flags&linux.MSG_CMSG_CLOEXEC != 0, controlData)
+	}
+
+	// Copy the address to the caller.
+	if msg.NameLen != 0 {
+		if err := writeAddress(t, sender, senderLen, usermem.Addr(msg.Name), usermem.Addr(msgPtr+nameLenOffset)); err != nil {
+			return 0, err
+		}
+	}
+
+	// Copy the control data to the caller.
+	if _, err := t.CopyOut(msgPtr+controlLenOffset, uint64(len(controlData))); err != nil {
+		return 0, err
+	}
+	if len(controlData) > 0 {
+		if _, err := t.CopyOut(usermem.Addr(msg.Control), controlData); err != nil {
+			return 0, err
+		}
+	}
+
+	return uintptr(n), nil
+}
+
+// recvFrom is the implementation of the recvfrom syscall. It is called by
+// recvfrom and recv syscall handlers.
+func recvFrom(t *kernel.Task, fd kdefs.FD, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLenPtr usermem.Addr) (uintptr, error) {
+	if int(bufLen) < 0 {
+		return 0, syscall.EINVAL
+	}
+
+	// Reject flags that we don't handle yet.
+	if flags & ^(linux.MSG_DONTWAIT|linux.MSG_NOSIGNAL|linux.MSG_PEEK|linux.MSG_TRUNC) != 0 {
+		return 0, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, syscall.ENOTSOCK
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	dst, err := t.SingleIOSequence(bufPtr, int(bufLen), usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, err
+	}
+
+	var haveDeadline bool
+	var deadline ktime.Time
+
+	if dl := s.RecvTimeout(); dl != 0 {
+		deadline = t.Kernel().MonotonicClock().Now().Add(time.Duration(dl) * time.Nanosecond)
+		haveDeadline = true
+	}
+
+	n, sender, senderLen, cm, e := s.RecvMsg(t, dst, int(flags), haveDeadline, deadline, nameLenPtr != 0, 0)
+	cm.Release()
+	if e != nil {
+		return 0, syserror.ConvertIntr(e.ToError(), kernel.ERESTARTSYS)
+	}
+
+	// Copy the address to the caller.
+	if nameLenPtr != 0 {
+		if err := writeAddress(t, sender, senderLen, namePtr, nameLenPtr); err != nil {
+			return 0, err
+		}
+	}
+
+	return uintptr(n), nil
+}
+
+// RecvFrom implements the linux syscall recvfrom(2).
+func RecvFrom(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	bufPtr := args[1].Pointer()
+	bufLen := args[2].Uint64()
+	flags := args[3].Int()
+	namePtr := args[4].Pointer()
+	nameLenPtr := args[5].Pointer()
+
+	n, err := recvFrom(t, fd, bufPtr, bufLen, flags, namePtr, nameLenPtr)
+	return n, nil, err
+}
+
+// SendMsg implements the linux syscall sendmsg(2).
+func SendMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	msgPtr := args[1].Pointer()
+	flags := args[2].Int()
+
+	if t.Arch().Width() != 8 {
+		// We only handle 64-bit for now.
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Reject flags that we don't handle yet.
+	if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	n, err := sendSingleMsg(t, s, file, msgPtr, flags)
+	return n, nil, err
+}
+
+// SendMMsg implements the linux syscall sendmmsg(2).
+func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	msgPtr := args[1].Pointer()
+	vlen := args[2].Uint()
+	flags := args[3].Int()
+
+	if t.Arch().Width() != 8 {
+		// We only handle 64-bit for now.
+		return 0, nil, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, nil, syscall.ENOTSOCK
+	}
+
+	// Reject flags that we don't handle yet.
+	if flags & ^(linux.MSG_DONTWAIT|linux.MSG_EOR|linux.MSG_MORE|linux.MSG_NOSIGNAL) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	var count uint32
+	var err error
+	for i := uint64(0); i < uint64(vlen); i++ {
+		mp, ok := msgPtr.AddLength(i * multipleMessageHeader64Len)
+		if !ok {
+			return 0, nil, syscall.EFAULT
+		}
+		var n uintptr
+		if n, err = sendSingleMsg(t, s, file, mp, flags); err != nil {
+			break
+		}
+
+		// Copy the received length to the caller.
+		lp, ok := mp.AddLength(messageHeader64Len)
+		if !ok {
+			return 0, nil, syscall.EFAULT
+		}
+		if _, err = t.CopyOut(lp, uint32(n)); err != nil {
+			break
+		}
+		count++
+	}
+
+	if count == 0 {
+		return 0, nil, err
+	}
+	return uintptr(count), nil, nil
+}
+
+func sendSingleMsg(t *kernel.Task, s socket.Socket, file *fs.File, msgPtr usermem.Addr, flags int32) (uintptr, error) {
+	// Capture the message header.
+	var msg MessageHeader64
+	if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil {
+		return 0, err
+	}
+
+	var controlData []byte
+	if msg.ControlLen > 0 {
+		// Put an upper bound to prevent large allocations.
+		if msg.ControlLen > maxControlLen {
+			return 0, syscall.ENOBUFS
+		}
+		controlData = make([]byte, msg.ControlLen)
+		if _, err := t.CopyIn(usermem.Addr(msg.Control), &controlData); err != nil {
+			return 0, err
+		}
+	}
+
+	// Read the destination address if one is specified.
+	var to []byte
+	if msg.NameLen != 0 {
+		var err error
+		to, err = CaptureAddress(t, usermem.Addr(msg.Name), msg.NameLen)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	// Read data then call the sendmsg implementation.
+	if msg.IovLen > linux.UIO_MAXIOV {
+		return 0, syscall.EMSGSIZE
+	}
+	src, err := t.IovecsIOSequence(usermem.Addr(msg.Iov), int(msg.IovLen), usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, err
+	}
+
+	controlMessages, err := control.Parse(t, s, controlData)
+	if err != nil {
+		return 0, err
+	}
+
+	// Call the syscall implementation.
+	n, e := s.SendMsg(t, src, to, int(flags), controlMessages)
+	err = handleIOError(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendmsg", file)
+	if err != nil {
+		controlMessages.Release()
+	}
+	return uintptr(n), err
+}
+
+// sendTo is the implementation of the sendto syscall. It is called by sendto
+// and send syscall handlers.
+func sendTo(t *kernel.Task, fd kdefs.FD, bufPtr usermem.Addr, bufLen uint64, flags int32, namePtr usermem.Addr, nameLen uint32) (uintptr, error) {
+	bl := int(bufLen)
+	if bl < 0 {
+		return 0, syscall.EINVAL
+	}
+
+	// Get socket from the file descriptor.
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, syscall.EBADF
+	}
+	defer file.DecRef()
+
+	// Extract the socket.
+	s, ok := file.FileOperations.(socket.Socket)
+	if !ok {
+		return 0, syscall.ENOTSOCK
+	}
+
+	if file.Flags().NonBlocking {
+		flags |= linux.MSG_DONTWAIT
+	}
+
+	// Read the destination address if one is specified.
+	var to []byte
+	var err error
+	if namePtr != 0 {
+		to, err = CaptureAddress(t, namePtr, nameLen)
+		if err != nil {
+			return 0, err
+		}
+	}
+
+	src, err := t.SingleIOSequence(bufPtr, bl, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, err
+	}
+
+	// Call the syscall implementation.
+	n, e := s.SendMsg(t, src, to, int(flags), control.New(t, s, nil))
+	return uintptr(n), handleIOError(t, n != 0, e.ToError(), kernel.ERESTARTSYS, "sendto", file)
+}
+
+// SendTo implements the linux syscall sendto(2).
+func SendTo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	bufPtr := args[1].Pointer()
+	bufLen := args[2].Uint64()
+	flags := args[3].Int()
+	namePtr := args[4].Pointer()
+	nameLen := args[5].Uint()
+
+	n, err := sendTo(t, fd, bufPtr, bufLen, flags, namePtr, nameLen)
+	return n, nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go
new file mode 100644
index 000000000..6e21b34fd
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_stat.go
@@ -0,0 +1,209 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Stat implements linux syscall stat(2).
+func Stat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	statAddr := args[1].Pointer()
+
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return stat(t, d, dirPath, statAddr)
+	})
+}
+
+// Fstatat implements linux syscall newfstatat, i.e. fstatat(2).
+func Fstatat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	statAddr := args[2].Pointer()
+	flags := args[3].Int()
+
+	path, dirPath, err := copyInPath(t, addr, flags&linux.AT_EMPTY_PATH != 0)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if path == "" {
+		// Annoying. What's wrong with fstat?
+		file := t.FDMap().GetFile(fd)
+		if file == nil {
+			return 0, nil, syserror.EBADF
+		}
+		defer file.DecRef()
+
+		return 0, nil, stat(t, file.Dirent, false, statAddr)
+	}
+
+	return 0, nil, fileOpOn(t, fd, path, flags&linux.AT_SYMLINK_NOFOLLOW == 0, func(root *fs.Dirent, d *fs.Dirent) error {
+		return stat(t, d, dirPath, statAddr)
+	})
+}
+
+// Lstat implements linux syscall lstat(2).
+func Lstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	statAddr := args[1].Pointer()
+
+	path, dirPath, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, false /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return stat(t, d, dirPath, statAddr)
+	})
+}
+
+// Fstat implements linux syscall fstat(2).
+func Fstat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	statAddr := args[1].Pointer()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	return 0, nil, stat(t, file.Dirent, false /* dirPath */, statAddr)
+}
+
+// stat implements stat from the given *fs.Dirent.
+func stat(t *kernel.Task, d *fs.Dirent, dirPath bool, statAddr usermem.Addr) error {
+	if dirPath && !fs.IsDir(d.Inode.StableAttr) {
+		return syserror.ENOTDIR
+	}
+	uattr, err := d.Inode.UnstableAttr(t)
+	if err != nil {
+		return err
+	}
+
+	var mode uint32
+	switch d.Inode.StableAttr.Type {
+	case fs.RegularFile, fs.SpecialFile:
+		mode |= linux.ModeRegular
+	case fs.Symlink:
+		mode |= linux.ModeSymlink
+	case fs.Directory, fs.SpecialDirectory:
+		mode |= linux.ModeDirectory
+	case fs.Pipe:
+		mode |= linux.ModeNamedPipe
+	case fs.CharacterDevice:
+		mode |= linux.ModeCharacterDevice
+	case fs.BlockDevice:
+		mode |= linux.ModeBlockDevice
+	case fs.Socket:
+		mode |= linux.ModeSocket
+	}
+
+	_, err = t.CopyOut(statAddr, linux.Stat{
+		Dev:     uint64(d.Inode.StableAttr.DeviceID),
+		Rdev:    uint64(linux.MakeDeviceID(d.Inode.StableAttr.DeviceFileMajor, d.Inode.StableAttr.DeviceFileMinor)),
+		Ino:     uint64(d.Inode.StableAttr.InodeID),
+		Nlink:   uattr.Links,
+		Mode:    mode | uint32(uattr.Perms.LinuxMode()),
+		UID:     uint32(uattr.Owner.UID.In(t.UserNamespace()).OrOverflow()),
+		GID:     uint32(uattr.Owner.GID.In(t.UserNamespace()).OrOverflow()),
+		Size:    uattr.Size,
+		Blksize: d.Inode.StableAttr.BlockSize,
+		Blocks:  uattr.Usage / 512,
+		ATime:   uattr.AccessTime.Timespec(),
+		MTime:   uattr.ModificationTime.Timespec(),
+		CTime:   uattr.StatusChangeTime.Timespec(),
+	})
+	return err
+}
+
+// Statfs implements linux syscall statfs(2).
+func Statfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	statfsAddr := args[1].Pointer()
+
+	path, _, err := copyInPath(t, addr, false /* allowEmpty */)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return 0, nil, fileOpOn(t, linux.AT_FDCWD, path, true /* resolve */, func(root *fs.Dirent, d *fs.Dirent) error {
+		return statfsImpl(t, d, statfsAddr)
+	})
+}
+
+// Fstatfs implements linux syscall fstatfs(2).
+func Fstatfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	statfsAddr := args[1].Pointer()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	return 0, nil, statfsImpl(t, file.Dirent, statfsAddr)
+}
+
+// statfsImpl implements the linux syscall statfs and fstatfs based on a Dirent,
+// copying the statfs structure out to addr on success, otherwise an error is
+// returned.
+func statfsImpl(t *kernel.Task, d *fs.Dirent, addr usermem.Addr) error {
+	info, err := d.Inode.StatFS(t)
+	if err != nil {
+		return err
+	}
+	// Construct the statfs structure and copy it out.
+	statfs := linux.Statfs{
+		Type: info.Type,
+		// Treat block size and fragment size as the same, as
+		// most consumers of this structure will expect one
+		// or the other to be filled in.
+		BlockSize: d.Inode.StableAttr.BlockSize,
+		Blocks:    info.TotalBlocks,
+		// We don't have the concept of reserved blocks, so
+		// report blocks free the same as available blocks.
+		// This is a normal thing for filesystems, to do, see
+		// udf, hugetlbfs, tmpfs, among others.
+		BlocksFree:      info.FreeBlocks,
+		BlocksAvailable: info.FreeBlocks,
+		Files:           info.TotalFiles,
+		FilesFree:       info.FreeFiles,
+		// Same as Linux for simple_statfs, see fs/libfs.c.
+		NameLength:   syscall.PathMax,
+		FragmentSize: d.Inode.StableAttr.BlockSize,
+		// Leave other fields 0 like simple_statfs does.
+	}
+	if _, err := t.CopyOut(addr, &statfs); err != nil {
+		return err
+	}
+	return nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_sync.go b/pkg/sentry/syscalls/linux/sys_sync.go
new file mode 100644
index 000000000..902d210db
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_sync.go
@@ -0,0 +1,75 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Sync implements linux system call sync(2).
+func Sync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	t.MountNamespace().SyncAll(t)
+	// Sync is always successful.
+	return 0, nil, nil
+}
+
+// Syncfs implements linux system call syncfs(2).
+func Syncfs(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Use "sync-the-world" for now, it's guaranteed that fd is at least
+	// on the root filesystem.
+	return Sync(t, args)
+}
+
+// Fsync implements linux syscall fsync(2).
+func Fsync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	err := file.Fsync(t, 0, fs.FileMaxOffset, fs.SyncAll)
+	return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+}
+
+// Fdatasync implements linux syscall fdatasync(2).
+//
+// At the moment, it just calls Fsync, which is a big hammer, but correct.
+func Fdatasync(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	err := file.Fsync(t, 0, fs.FileMaxOffset, fs.SyncData)
+	return 0, nil, syserror.ConvertIntr(err, kernel.ERESTARTSYS)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_sysinfo.go b/pkg/sentry/syscalls/linux/sys_sysinfo.go
new file mode 100644
index 000000000..bd0ffcd5c
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_sysinfo.go
@@ -0,0 +1,42 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usage"
+)
+
+// Sysinfo implements the sysinfo syscall as described in man 2 sysinfo.
+func Sysinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	mem := t.Kernel().Platform.Memory()
+	mem.UpdateUsage()
+	_, totalUsage := usage.MemoryAccounting.Copy()
+	totalSize := usage.TotalMemory(mem.TotalSize(), totalUsage)
+
+	// Only a subset of the fields in sysinfo_t make sense to return.
+	si := linux.Sysinfo{
+		Procs:    uint16(len(t.PIDNamespace().Tasks())),
+		Uptime:   t.Kernel().MonotonicClock().Now().Seconds(),
+		TotalRAM: totalSize,
+		FreeRAM:  totalSize - totalUsage,
+	}
+	_, err := t.CopyOut(addr, si)
+	return 0, nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_syslog.go b/pkg/sentry/syscalls/linux/sys_syslog.go
new file mode 100644
index 000000000..792040c81
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_syslog.go
@@ -0,0 +1,61 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	_SYSLOG_ACTION_READ_ALL    = 3
+	_SYSLOG_ACTION_SIZE_BUFFER = 10
+)
+
+// logBufLen is the default syslog buffer size on Linux.
+const logBufLen = 1 << 17
+
+// Syslog implements part of Linux syscall syslog.
+//
+// Only the unpriviledged commands are implemented, allowing applications to
+// read a fun dmesg.
+func Syslog(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	command := args[0].Int()
+	buf := args[1].Pointer()
+	size := int(args[2].Int())
+
+	switch command {
+	case _SYSLOG_ACTION_READ_ALL:
+		if size < 0 {
+			return 0, nil, syserror.EINVAL
+		}
+		if size > logBufLen {
+			size = logBufLen
+		}
+
+		log := t.Kernel().Syslog().Log()
+		if len(log) > size {
+			log = log[:size]
+		}
+
+		n, err := t.CopyOutBytes(buf, log)
+		return uintptr(n), nil, err
+	case _SYSLOG_ACTION_SIZE_BUFFER:
+		return logBufLen, nil, nil
+	default:
+		return 0, nil, syserror.ENOSYS
+	}
+}
diff --git a/pkg/sentry/syscalls/linux/sys_thread.go b/pkg/sentry/syscalls/linux/sys_thread.go
new file mode 100644
index 000000000..0adbf160f
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_thread.go
@@ -0,0 +1,704 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/sched"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const (
+	// ExecMaxTotalSize is the maximum length of all argv and envv entries.
+	//
+	// N.B. The behavior here is different than Linux. Linux provides a limit on
+	// individual arguments of 32 pages, and an aggregate limit of at least 32 pages
+	// but otherwise bounded by min(stack size / 4, 8 MB * 3 / 4). We don't implement
+	// any behavior based on the stack size, and instead provide a fixed hard-limit of
+	// 2 MB (which should work well given that 8 MB stack limits are common).
+	ExecMaxTotalSize = 2 * 1024 * 1024
+
+	// ExecMaxElemSize is the maximum length of a single argv or envv entry.
+	ExecMaxElemSize = 32 * usermem.PageSize
+
+	// exitSignalMask is the signal mask to be sent at exit. Same as CSIGNAL in linux.
+	exitSignalMask = 0xff
+)
+
+// Possible values for the idtype argument to waitid(2), defined in Linux's
+// include/uapi/linux/wait.h.
+const (
+	_P_ALL  = 0
+	_P_PID  = 1
+	_P_PGID = 2
+)
+
+// Getppid implements linux syscall getppid(2).
+func Getppid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	parent := t.Parent()
+	if parent == nil {
+		return 0, nil, nil
+	}
+	return uintptr(t.PIDNamespace().IDOfThreadGroup(parent.ThreadGroup())), nil, nil
+}
+
+// Getpid implements linux syscall getpid(2).
+func Getpid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return uintptr(t.ThreadGroup().ID()), nil, nil
+}
+
+// Gettid implements linux syscall gettid(2).
+func Gettid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return uintptr(t.ThreadID()), nil, nil
+}
+
+// Execve implements linux syscall execve(2).
+func Execve(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	filenameAddr := args[0].Pointer()
+	argvAddr := args[1].Pointer()
+	envvAddr := args[2].Pointer()
+
+	// Extract our arguments.
+	filename, err := t.CopyInString(filenameAddr, syscall.PathMax)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	var argv, envv []string
+	if argvAddr != 0 {
+		var err error
+		argv, err = t.CopyInVector(argvAddr, ExecMaxElemSize, ExecMaxTotalSize)
+		if err != nil {
+			return 0, nil, err
+		}
+	}
+	if envvAddr != 0 {
+		var err error
+		envv, err = t.CopyInVector(envvAddr, ExecMaxElemSize, ExecMaxTotalSize)
+		if err != nil {
+			return 0, nil, err
+		}
+	}
+
+	root := t.FSContext().RootDirectory()
+	defer root.DecRef()
+	wd := t.FSContext().WorkingDirectory()
+	defer wd.DecRef()
+
+	// Load the new TaskContext.
+	tc, err := t.Kernel().LoadTaskImage(t, t.MountNamespace(), root, wd, linux.MaxSymlinkTraversals, filename, argv, envv, t.Arch().FeatureSet())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	ctrl, err := t.Execve(tc)
+	return 0, ctrl, err
+}
+
+// Exit implements linux syscall exit(2).
+func Exit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	status := int(args[0].Int())
+	t.PrepareExit(kernel.ExitStatus{Code: status})
+	return 0, kernel.CtrlDoExit, nil
+}
+
+// ExitGroup implements linux syscall exit_group(2).
+func ExitGroup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	status := int(args[0].Int())
+	t.PrepareGroupExit(kernel.ExitStatus{Code: status})
+	return 0, kernel.CtrlDoExit, nil
+}
+
+// clone is used by Clone, Fork, and VFork.
+func clone(t *kernel.Task, flags int, stack usermem.Addr, parentTID usermem.Addr, childTID usermem.Addr, tls usermem.Addr) (uintptr, *kernel.SyscallControl, error) {
+	opts := kernel.CloneOptions{
+		SharingOptions: kernel.SharingOptions{
+			NewAddressSpace:     flags&syscall.CLONE_VM == 0,
+			NewSignalHandlers:   flags&syscall.CLONE_SIGHAND == 0,
+			NewThreadGroup:      flags&syscall.CLONE_THREAD == 0,
+			TerminationSignal:   linux.Signal(flags & exitSignalMask),
+			NewPIDNamespace:     flags&syscall.CLONE_NEWPID == syscall.CLONE_NEWPID,
+			NewUserNamespace:    flags&syscall.CLONE_NEWUSER == syscall.CLONE_NEWUSER,
+			NewNetworkNamespace: flags&syscall.CLONE_NEWNET == syscall.CLONE_NEWNET,
+			NewFiles:            flags&syscall.CLONE_FILES == 0,
+			NewFSContext:        flags&syscall.CLONE_FS == 0,
+			NewUTSNamespace:     flags&syscall.CLONE_NEWUTS == syscall.CLONE_NEWUTS,
+			NewIPCNamespace:     flags&syscall.CLONE_NEWIPC == syscall.CLONE_NEWIPC,
+		},
+		Stack:         stack,
+		SetTLS:        flags&syscall.CLONE_SETTLS == syscall.CLONE_SETTLS,
+		TLS:           tls,
+		ChildClearTID: flags&syscall.CLONE_CHILD_CLEARTID == syscall.CLONE_CHILD_CLEARTID,
+		ChildSetTID:   flags&syscall.CLONE_CHILD_SETTID == syscall.CLONE_CHILD_SETTID,
+		ChildTID:      childTID,
+		ParentSetTID:  flags&syscall.CLONE_PARENT_SETTID == syscall.CLONE_PARENT_SETTID,
+		ParentTID:     parentTID,
+		Vfork:         flags&syscall.CLONE_VFORK == syscall.CLONE_VFORK,
+		Untraced:      flags&syscall.CLONE_UNTRACED == syscall.CLONE_UNTRACED,
+		InheritTracer: flags&syscall.CLONE_PTRACE == syscall.CLONE_PTRACE,
+	}
+	ntid, ctrl, err := t.Clone(&opts)
+	return uintptr(ntid), ctrl, err
+}
+
+// Clone implements linux syscall clone(2).
+// sys_clone has so many flavors. We implement the default one in the
+// current linux 3.11 x86_64:
+//    sys_clone(clone_flags, newsp, parent_tidptr, child_tidptr, tls_val)
+func Clone(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	flags := int(args[0].Int())
+	stack := args[1].Pointer()
+	parentTID := args[2].Pointer()
+	childTID := args[3].Pointer()
+	tls := args[4].Pointer()
+	return clone(t, flags, stack, parentTID, childTID, tls)
+}
+
+// Fork implements Linux syscall fork(2).
+func Fork(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	// "A call to fork() is equivalent to a call to clone(2) specifying flags
+	// as just SIGCHLD." - fork(2)
+	return clone(t, int(syscall.SIGCHLD), 0, 0, 0, 0)
+}
+
+// Vfork implements Linux syscall vfork(2).
+func Vfork(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	// """
+	// A call to vfork() is equivalent to calling clone(2) with flags specified as:
+	//
+	//     CLONE_VM | CLONE_VFORK | SIGCHLD
+	// """ - vfork(2)
+	return clone(t, syscall.CLONE_VM|syscall.CLONE_VFORK|int(syscall.SIGCHLD), 0, 0, 0, 0)
+}
+
+// wait4 waits for the given child process to exit.
+func wait4(t *kernel.Task, pid int, statusAddr usermem.Addr, options int, rusageAddr usermem.Addr) (uintptr, error) {
+	if options&^(syscall.WNOHANG|syscall.WUNTRACED|syscall.WCONTINUED|syscall.WALL|syscall.WCLONE) != 0 {
+		return 0, syscall.EINVAL
+	}
+	wopts := kernel.WaitOptions{
+		Events:       kernel.EventExit | kernel.EventTraceeStop,
+		ConsumeEvent: true,
+	}
+	// There are four cases to consider:
+	//
+	// pid < -1    any child process whose process group ID is equal to the absolute value of pid
+	// pid == -1   any child process
+	// pid == 0    any child process whose process group ID is equal to that of the calling process
+	// pid > 0     the child whose process ID is equal to the value of pid
+	switch {
+	case pid < -1:
+		wopts.SpecificPGID = kernel.ProcessGroupID(-pid)
+	case pid == -1:
+		// Any process is the default.
+	case pid == 0:
+		wopts.SpecificPGID = t.PIDNamespace().IDOfProcessGroup(t.ThreadGroup().ProcessGroup())
+	default:
+		wopts.SpecificTID = kernel.ThreadID(pid)
+	}
+
+	switch options & (syscall.WCLONE | syscall.WALL) {
+	case 0:
+		wopts.NonCloneTasks = true
+	case syscall.WCLONE:
+		wopts.CloneTasks = true
+	case syscall.WALL:
+		wopts.NonCloneTasks = true
+		wopts.CloneTasks = true
+	default:
+		return 0, syscall.EINVAL
+	}
+	if options&syscall.WUNTRACED != 0 {
+		wopts.Events |= kernel.EventChildGroupStop
+	}
+	if options&syscall.WCONTINUED != 0 {
+		wopts.Events |= kernel.EventGroupContinue
+	}
+	if options&syscall.WNOHANG == 0 {
+		wopts.BlockInterruptErr = kernel.ERESTARTSYS
+	}
+
+	wr, err := t.Wait(&wopts)
+	if err != nil {
+		if err == kernel.ErrNoWaitableEvent {
+			return 0, nil
+		}
+		return 0, err
+	}
+	if statusAddr != 0 {
+		if _, err := t.CopyOut(statusAddr, wr.Status); err != nil {
+			return 0, err
+		}
+	}
+	if rusageAddr != 0 {
+		ru := getrusage(wr.Task, linux.RUSAGE_BOTH)
+		if _, err := t.CopyOut(rusageAddr, &ru); err != nil {
+			return 0, err
+		}
+	}
+	return uintptr(wr.TID), nil
+}
+
+// Wait4 implements linux syscall wait4(2).
+func Wait4(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := int(args[0].Int())
+	statusAddr := args[1].Pointer()
+	options := int(args[2].Uint())
+	rusageAddr := args[3].Pointer()
+
+	n, err := wait4(t, pid, statusAddr, options, rusageAddr)
+	return n, nil, err
+}
+
+// WaitPid implements linux syscall waitpid(2).
+func WaitPid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	pid := int(args[0].Int())
+	statusAddr := args[1].Pointer()
+	options := int(args[2].Uint())
+
+	n, err := wait4(t, pid, statusAddr, options, 0)
+	return n, nil, err
+}
+
+// Waitid implements linux syscall waitid(2).
+func Waitid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	idtype := args[0].Int()
+	id := args[1].Int()
+	infop := args[2].Pointer()
+	options := int(args[3].Uint())
+	rusageAddr := args[4].Pointer()
+
+	if options&^(syscall.WNOHANG|syscall.WEXITED|syscall.WSTOPPED|syscall.WCONTINUED|syscall.WNOWAIT) != 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	if options&(syscall.WEXITED|syscall.WSTOPPED|syscall.WCONTINUED) == 0 {
+		return 0, nil, syscall.EINVAL
+	}
+	wopts := kernel.WaitOptions{
+		NonCloneTasks: true,
+		Events:        kernel.EventTraceeStop,
+		ConsumeEvent:  options&syscall.WNOWAIT == 0,
+	}
+	switch idtype {
+	case _P_ALL:
+	case _P_PID:
+		wopts.SpecificTID = kernel.ThreadID(id)
+	case _P_PGID:
+		wopts.SpecificPGID = kernel.ProcessGroupID(id)
+	default:
+		return 0, nil, syscall.EINVAL
+	}
+	if options&syscall.WEXITED != 0 {
+		wopts.Events |= kernel.EventExit
+	}
+	if options&syscall.WSTOPPED != 0 {
+		wopts.Events |= kernel.EventChildGroupStop
+	}
+	if options&syscall.WCONTINUED != 0 {
+		wopts.Events |= kernel.EventGroupContinue
+	}
+	if options&syscall.WNOHANG == 0 {
+		wopts.BlockInterruptErr = kernel.ERESTARTSYS
+	}
+
+	wr, err := t.Wait(&wopts)
+	if err != nil {
+		if err == kernel.ErrNoWaitableEvent {
+			err = nil
+			// "If WNOHANG was specified in options and there were no children
+			// in a waitable state, then waitid() returns 0 immediately and the
+			// state of the siginfo_t structure pointed to by infop is
+			// unspecified." - waitid(2). But Linux's waitid actually zeroes
+			// out the fields it would set for a successful waitid in this case
+			// as well.
+			if infop != 0 {
+				var si arch.SignalInfo
+				_, err = t.CopyOut(infop, &si)
+			}
+		}
+		return 0, nil, err
+	}
+	if rusageAddr != 0 {
+		ru := getrusage(wr.Task, linux.RUSAGE_BOTH)
+		if _, err := t.CopyOut(rusageAddr, &ru); err != nil {
+			return 0, nil, err
+		}
+	}
+	if infop == 0 {
+		return 0, nil, nil
+	}
+	si := arch.SignalInfo{
+		Signo: int32(syscall.SIGCHLD),
+	}
+	si.SetPid(int32(wr.TID))
+	si.SetUid(int32(wr.UID))
+	// TODO: convert kernel.ExitStatus to functions and make
+	// WaitResult.Status a linux.WaitStatus
+	s := syscall.WaitStatus(wr.Status)
+	switch {
+	case s.Exited():
+		si.Code = arch.CLD_EXITED
+		si.SetStatus(int32(s.ExitStatus()))
+	case s.Signaled():
+		si.Code = arch.CLD_KILLED
+		si.SetStatus(int32(s.Signal()))
+	case s.CoreDump():
+		si.Code = arch.CLD_DUMPED
+		si.SetStatus(int32(s.Signal()))
+	case s.Stopped():
+		if wr.Event == kernel.EventTraceeStop {
+			si.Code = arch.CLD_TRAPPED
+			si.SetStatus(int32(s.TrapCause()))
+		} else {
+			si.Code = arch.CLD_STOPPED
+			si.SetStatus(int32(s.StopSignal()))
+		}
+	case s.Continued():
+		si.Code = arch.CLD_CONTINUED
+		si.SetStatus(int32(syscall.SIGCONT))
+	default:
+		t.Warningf("waitid got incomprehensible wait status %d", s)
+	}
+	_, err = t.CopyOut(infop, &si)
+	return 0, nil, err
+}
+
+// SetTidAddress implements linux syscall set_tid_address(2).
+func SetTidAddress(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	// Always succeed, return caller's tid.
+	t.SetClearTID(addr)
+	return uintptr(t.ThreadID()), nil, nil
+}
+
+// Unshare implements linux syscall unshare(2).
+func Unshare(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	flags := args[0].Int()
+	opts := kernel.SharingOptions{
+		NewAddressSpace:     flags&syscall.CLONE_VM == syscall.CLONE_VM,
+		NewSignalHandlers:   flags&syscall.CLONE_SIGHAND == syscall.CLONE_SIGHAND,
+		NewThreadGroup:      flags&syscall.CLONE_THREAD == syscall.CLONE_THREAD,
+		NewPIDNamespace:     flags&syscall.CLONE_NEWPID == syscall.CLONE_NEWPID,
+		NewUserNamespace:    flags&syscall.CLONE_NEWUSER == syscall.CLONE_NEWUSER,
+		NewNetworkNamespace: flags&syscall.CLONE_NEWNET == syscall.CLONE_NEWNET,
+		NewFiles:            flags&syscall.CLONE_FILES == syscall.CLONE_FILES,
+		NewFSContext:        flags&syscall.CLONE_FS == syscall.CLONE_FS,
+		NewUTSNamespace:     flags&syscall.CLONE_NEWUTS == syscall.CLONE_NEWUTS,
+		NewIPCNamespace:     flags&syscall.CLONE_NEWIPC == syscall.CLONE_NEWIPC,
+	}
+	// "CLONE_NEWPID automatically implies CLONE_THREAD as well." - unshare(2)
+	if opts.NewPIDNamespace {
+		opts.NewThreadGroup = true
+	}
+	// "... specifying CLONE_NEWUSER automatically implies CLONE_THREAD. Since
+	// Linux 3.9, CLONE_NEWUSER also automatically implies CLONE_FS."
+	if opts.NewUserNamespace {
+		opts.NewThreadGroup = true
+		opts.NewFSContext = true
+	}
+	return 0, nil, t.Unshare(&opts)
+}
+
+// SchedYield implements linux syscall sched_yield(2).
+func SchedYield(t *kernel.Task, _ arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	t.Yield()
+	return 0, nil, nil
+}
+
+// SchedSetaffinity implements linux syscall sched_setaffinity(2).
+func SchedSetaffinity(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := args[0].Int()
+	size := args[1].SizeT()
+	maskAddr := args[2].Pointer()
+
+	var task *kernel.Task
+	if tid == 0 {
+		task = t
+	} else {
+		task = t.PIDNamespace().TaskWithID(kernel.ThreadID(tid))
+		if task == nil {
+			return 0, nil, syserror.ESRCH
+		}
+	}
+
+	mask := sched.NewCPUSet(t.Kernel().ApplicationCores())
+	if size > mask.Size() {
+		size = mask.Size()
+	}
+	if _, err := t.CopyInBytes(maskAddr, mask[:size]); err != nil {
+		return 0, nil, err
+	}
+	return 0, nil, task.SetCPUMask(mask)
+}
+
+// SchedGetaffinity implements linux syscall sched_getaffinity(2).
+func SchedGetaffinity(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := args[0].Int()
+	size := args[1].SizeT()
+	maskAddr := args[2].Pointer()
+
+	// This limitation is because linux stores the cpumask
+	// in an array of "unsigned long" so the buffer needs to
+	// be a multiple of the word size.
+	if size&(t.Arch().Width()-1) > 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	var task *kernel.Task
+	if tid == 0 {
+		task = t
+	} else {
+		task = t.PIDNamespace().TaskWithID(kernel.ThreadID(tid))
+		if task == nil {
+			return 0, nil, syserror.ESRCH
+		}
+	}
+
+	mask := task.CPUMask()
+	// The buffer needs to be big enough to hold a cpumask with
+	// all possible cpus.
+	if size < mask.Size() {
+		return 0, nil, syserror.EINVAL
+	}
+	_, err := t.CopyOutBytes(maskAddr, mask)
+
+	// NOTE: The syscall interface is slightly different than the glibc
+	// interface. The raw sched_getaffinity syscall returns the number of
+	// bytes used to represent a cpu mask.
+	return uintptr(mask.Size()), nil, err
+}
+
+// Getcpu implements linux syscall getcpu(2).
+func Getcpu(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	cpu := args[0].Pointer()
+	node := args[1].Pointer()
+	// third argument to this system call is nowadays unused.
+
+	if cpu != 0 {
+		buf := t.CopyScratchBuffer(4)
+		usermem.ByteOrder.PutUint32(buf, uint32(t.CPU()))
+		if _, err := t.CopyOutBytes(cpu, buf); err != nil {
+			return 0, nil, err
+		}
+	}
+	// We always return node 0.
+	if node != 0 {
+		if _, err := t.MemoryManager().ZeroOut(t, node, 4, usermem.IOOpts{
+			AddressSpaceActive: true,
+		}); err != nil {
+			return 0, nil, err
+		}
+	}
+	return 0, nil, nil
+}
+
+// Setpgid implements the linux syscall setpgid(2).
+func Setpgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	// Note that throughout this function, pgid is interpreted with respect
+	// to t's namespace, not with respect to the selected ThreadGroup's
+	// namespace (which may be different).
+	pid := kernel.ThreadID(args[0].Int())
+	pgid := kernel.ProcessGroupID(args[1].Int())
+
+	// "If pid is zero, then the process ID of the calling process is used."
+	tg := t.ThreadGroup()
+	if pid != 0 {
+		ot := t.PIDNamespace().TaskWithID(pid)
+		if ot == nil {
+			return 0, nil, syserror.ESRCH
+		}
+		tg = ot.ThreadGroup()
+		if tg.Leader() != ot {
+			return 0, nil, syserror.EINVAL
+		}
+
+		// Setpgid only operates on child threadgroups.
+		if tg != t.ThreadGroup() && (tg.Leader().Parent() == nil || tg.Leader().Parent().ThreadGroup() != t.ThreadGroup()) {
+			return 0, nil, syserror.ESRCH
+		}
+	}
+
+	// "If pgid is zero, then the PGID of the process specified by pid is made
+	// the same as its process ID."
+	defaultPGID := kernel.ProcessGroupID(t.PIDNamespace().IDOfThreadGroup(tg))
+	if pgid == 0 {
+		pgid = defaultPGID
+	} else if pgid < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// If the pgid is the same as the group, then create a new one. Otherwise,
+	// we attempt to join an existing process group.
+	if pgid == defaultPGID {
+		// For convenience, errors line up with Linux syscall API.
+		if err := tg.CreateProcessGroup(); err != nil {
+			// Is the process group already as expected? If so,
+			// just return success. This is the same behavior as
+			// Linux.
+			if t.PIDNamespace().IDOfProcessGroup(tg.ProcessGroup()) == defaultPGID {
+				return 0, nil, nil
+			}
+			return 0, nil, err
+		}
+	} else {
+		// Same as CreateProcessGroup, above.
+		if err := tg.JoinProcessGroup(t.PIDNamespace(), pgid, tg != t.ThreadGroup()); err != nil {
+			// See above.
+			if t.PIDNamespace().IDOfProcessGroup(tg.ProcessGroup()) == pgid {
+				return 0, nil, nil
+			}
+			return 0, nil, err
+		}
+	}
+
+	// Success.
+	return 0, nil, nil
+}
+
+// Getpgrp implements the linux syscall getpgrp(2).
+func Getpgrp(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return uintptr(t.PIDNamespace().IDOfProcessGroup(t.ThreadGroup().ProcessGroup())), nil, nil
+}
+
+// Getpgid implements the linux syscall getpgid(2).
+func Getpgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := kernel.ThreadID(args[0].Int())
+	if tid == 0 {
+		return Getpgrp(t, args)
+	}
+
+	target := t.PIDNamespace().TaskWithID(tid)
+	if target == nil {
+		return 0, nil, syserror.ESRCH
+	}
+
+	return uintptr(t.PIDNamespace().IDOfProcessGroup(target.ThreadGroup().ProcessGroup())), nil, nil
+}
+
+// Setsid implements the linux syscall setsid(2).
+func Setsid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, t.ThreadGroup().CreateSession()
+}
+
+// Getsid implements the linux syscall getsid(2).
+func Getsid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tid := kernel.ThreadID(args[0].Int())
+	if tid == 0 {
+		return uintptr(t.PIDNamespace().IDOfSession(t.ThreadGroup().Session())), nil, nil
+	}
+
+	target := t.PIDNamespace().TaskWithID(tid)
+	if target == nil {
+		return 0, nil, syserror.ESRCH
+	}
+
+	return uintptr(t.PIDNamespace().IDOfSession(target.ThreadGroup().Session())), nil, nil
+}
+
+// Getpriority pretends to implement the linux syscall getpriority(2).
+//
+// This is a stub; real priorities require a full scheduler.
+func Getpriority(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	which := args[0].Int()
+	who := kernel.ThreadID(args[1].Int())
+
+	switch which {
+	case syscall.PRIO_PROCESS:
+		// Look for who, return ESRCH if not found.
+		var task *kernel.Task
+		if who == 0 {
+			task = t
+		} else {
+			task = t.PIDNamespace().TaskWithID(who)
+		}
+
+		if task == nil {
+			return 0, nil, syscall.ESRCH
+		}
+
+		// From kernel/sys.c:getpriority:
+		// "To avoid negative return values, 'getpriority()'
+		// will not return the normal nice-value, but a negated
+		// value that has been offset by 20"
+		return uintptr(20 - task.Niceness()), nil, nil
+	case syscall.PRIO_USER:
+		fallthrough
+	case syscall.PRIO_PGRP:
+		// PRIO_USER and PRIO_PGRP have no further implementation yet.
+		return 0, nil, nil
+	default:
+		return 0, nil, syscall.EINVAL
+	}
+}
+
+// Setpriority pretends to implement the linux syscall setpriority(2).
+//
+// This is a stub; real priorities require a full scheduler.
+func Setpriority(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	which := args[0].Int()
+	who := kernel.ThreadID(args[1].Int())
+	niceval := int(args[2].Int())
+
+	// In the kernel's implementation, values outside the range
+	// of [-20, 19] are truncated to these minimum and maximum
+	// values.
+	if niceval < -20 /* min niceval */ {
+		niceval = -20
+	} else if niceval > 19 /* max niceval */ {
+		niceval = 19
+	}
+
+	switch which {
+	case syscall.PRIO_PROCESS:
+		// Look for who, return ESRCH if not found.
+		var task *kernel.Task
+		if who == 0 {
+			task = t
+		} else {
+			task = t.PIDNamespace().TaskWithID(who)
+		}
+
+		if task == nil {
+			return 0, nil, syscall.ESRCH
+		}
+
+		task.SetNiceness(niceval)
+	case syscall.PRIO_USER:
+		fallthrough
+	case syscall.PRIO_PGRP:
+		// PRIO_USER and PRIO_PGRP have no further implementation yet.
+		return 0, nil, nil
+	default:
+		return 0, nil, syscall.EINVAL
+	}
+
+	return 0, nil, nil
+}
+
+// Ptrace implements linux system call ptrace(2).
+func Ptrace(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	req := args[0].Int64()
+	pid := kernel.ThreadID(args[1].Int())
+	addr := args[2].Pointer()
+	data := args[3].Pointer()
+
+	return 0, nil, t.Ptrace(req, pid, addr, data)
+}
diff --git a/pkg/sentry/syscalls/linux/sys_time.go b/pkg/sentry/syscalls/linux/sys_time.go
new file mode 100644
index 000000000..dcee694b2
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_time.go
@@ -0,0 +1,338 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// The most significant 29 bits hold either a pid or a file descriptor.
+func pidOfClockID(c int32) kernel.ThreadID {
+	return kernel.ThreadID(^(c >> 3))
+}
+
+// whichCPUClock returns one of CPUCLOCK_PERF, CPUCLOCK_VIRT, CPUCLOCK_SCHED or
+// CLOCK_FD.
+func whichCPUClock(c int32) int32 {
+	return c & linux.CPUCLOCK_CLOCK_MASK
+}
+
+// isCPUClockPerThread returns true if the CPUCLOCK_PERTHREAD bit is set in the
+// clock id.
+func isCPUClockPerThread(c int32) bool {
+	return c&linux.CPUCLOCK_PERTHREAD_MASK != 0
+}
+
+// isValidCPUClock returns checks that the cpu clock id is valid.
+func isValidCPUClock(c int32) bool {
+	// Bits 0, 1, and 2 cannot all be set.
+	if c&7 == 7 {
+		return false
+	}
+	if whichCPUClock(c) >= linux.CPUCLOCK_MAX {
+		return false
+	}
+	return true
+}
+
+// targetTask returns the kernel.Task for the given clock id.
+func targetTask(t *kernel.Task, c int32) *kernel.Task {
+	pid := pidOfClockID(c)
+	if pid == 0 {
+		return t
+	}
+	return t.PIDNamespace().TaskWithID(pid)
+}
+
+// ClockGetres implements linux syscall clock_getres(2).
+func ClockGetres(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	clockID := int32(args[0].Int())
+	addr := args[1].Pointer()
+	r := linux.Timespec{
+		Sec:  0,
+		Nsec: 1,
+	}
+
+	if _, err := getClock(t, clockID); err != nil {
+		return 0, nil, syserror.EINVAL
+	}
+
+	if addr == 0 {
+		// Don't need to copy out.
+		return 0, nil, nil
+	}
+
+	return 0, nil, copyTimespecOut(t, addr, &r)
+}
+
+type cpuClocker interface {
+	UserCPUClock() ktime.Clock
+	CPUClock() ktime.Clock
+}
+
+func getClock(t *kernel.Task, clockID int32) (ktime.Clock, error) {
+	if clockID < 0 {
+		if !isValidCPUClock(clockID) {
+			return nil, syserror.EINVAL
+		}
+
+		targetTask := targetTask(t, clockID)
+		if targetTask == nil {
+			return nil, syserror.EINVAL
+		}
+
+		var target cpuClocker
+		if isCPUClockPerThread(clockID) {
+			target = targetTask
+		} else {
+			target = targetTask.ThreadGroup()
+		}
+
+		switch whichCPUClock(clockID) {
+		case linux.CPUCLOCK_VIRT:
+			return target.UserCPUClock(), nil
+		case linux.CPUCLOCK_PROF, linux.CPUCLOCK_SCHED:
+			// CPUCLOCK_SCHED is approximated by CPUCLOCK_PROF.
+			return target.CPUClock(), nil
+		default:
+			return nil, syserror.EINVAL
+		}
+	}
+
+	switch clockID {
+	case linux.CLOCK_REALTIME, linux.CLOCK_REALTIME_COARSE:
+		return t.Kernel().RealtimeClock(), nil
+	case linux.CLOCK_MONOTONIC, linux.CLOCK_MONOTONIC_COARSE, linux.CLOCK_MONOTONIC_RAW:
+		// CLOCK_MONOTONIC approximates CLOCK_MONOTONIC_RAW.
+		return t.Kernel().MonotonicClock(), nil
+	case linux.CLOCK_PROCESS_CPUTIME_ID:
+		return t.ThreadGroup().CPUClock(), nil
+	case linux.CLOCK_THREAD_CPUTIME_ID:
+		return t.CPUClock(), nil
+	default:
+		return nil, syserror.EINVAL
+	}
+}
+
+// ClockGettime implements linux syscall clock_gettime(2).
+func ClockGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	clockID := int32(args[0].Int())
+	addr := args[1].Pointer()
+
+	c, err := getClock(t, clockID)
+	if err != nil {
+		return 0, nil, err
+	}
+	ts := c.Now().Timespec()
+	return 0, nil, copyTimespecOut(t, addr, &ts)
+}
+
+// ClockSettime implements linux syscall clock_settime(2).
+func ClockSettime(*kernel.Task, arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	return 0, nil, syserror.EPERM
+}
+
+// Time implements linux syscall time(2).
+func Time(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+
+	r := t.Kernel().RealtimeClock().Now().TimeT()
+	if addr == usermem.Addr(0) {
+		return uintptr(r), nil, nil
+	}
+
+	if _, err := t.CopyOut(addr, r); err != nil {
+		return 0, nil, err
+	}
+	return uintptr(r), nil, nil
+}
+
+// clockNanosleepRestartBlock encapsulates the state required to restart
+// clock_nanosleep(2) via restart_syscall(2).
+type clockNanosleepRestartBlock struct {
+	c        ktime.Clock
+	duration time.Duration
+	rem      usermem.Addr
+}
+
+// Restart implements kernel.SyscallRestartBlock.Restart.
+func (n *clockNanosleepRestartBlock) Restart(t *kernel.Task) (uintptr, error) {
+	return 0, clockNanosleepFor(t, n.c, n.duration, n.rem)
+}
+
+// clockNanosleepUntil blocks until a specified time.
+//
+// If blocking is interrupted, the syscall is restarted with the original
+// arguments.
+func clockNanosleepUntil(t *kernel.Task, c ktime.Clock, ts linux.Timespec) error {
+	notifier, tchan := ktime.NewChannelNotifier()
+	timer := ktime.NewTimer(c, notifier)
+
+	// Turn on the timer.
+	timer.Swap(ktime.Setting{
+		Period:  0,
+		Enabled: true,
+		Next:    ktime.FromTimespec(ts),
+	})
+
+	err := t.BlockWithTimer(nil, tchan)
+
+	timer.Destroy()
+
+	// Did we just block until the timeout happened?
+	if err == syserror.ETIMEDOUT {
+		return nil
+	}
+
+	return syserror.ConvertIntr(err, kernel.ERESTARTNOHAND)
+}
+
+// clockNanosleepFor blocks for a specified duration.
+//
+// If blocking is interrupted, the syscall is restarted with the remaining
+// duration timeout.
+func clockNanosleepFor(t *kernel.Task, c ktime.Clock, dur time.Duration, rem usermem.Addr) error {
+	timer, start, tchan := ktime.After(c, dur)
+
+	err := t.BlockWithTimer(nil, tchan)
+
+	after := c.Now()
+
+	timer.Destroy()
+
+	var remaining time.Duration
+	// Did we just block for the entire duration?
+	if err == syserror.ETIMEDOUT {
+		remaining = 0
+	} else {
+		remaining = dur - after.Sub(start)
+		if remaining < 0 {
+			remaining = time.Duration(0)
+		}
+	}
+
+	// Copy out remaining time.
+	if err != nil && rem != usermem.Addr(0) {
+		timeleft := linux.NsecToTimespec(remaining.Nanoseconds())
+		if err := copyTimespecOut(t, rem, &timeleft); err != nil {
+			return err
+		}
+	}
+
+	// Did we just block for the entire duration?
+	if err == syserror.ETIMEDOUT {
+		return nil
+	}
+
+	// If interrupted, arrange for a restart with the remaining duration.
+	if err == syserror.ErrInterrupted {
+		t.SetSyscallRestartBlock(&clockNanosleepRestartBlock{
+			c:        c,
+			duration: remaining,
+			rem:      rem,
+		})
+		return kernel.ERESTART_RESTARTBLOCK
+	}
+
+	return err
+}
+
+// Nanosleep implements linux syscall Nanosleep(2).
+func Nanosleep(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	addr := args[0].Pointer()
+	rem := args[1].Pointer()
+
+	ts, err := copyTimespecIn(t, addr)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if !ts.Valid() {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Just like linux, we cap the timeout with the max number that int64 can
+	// represent which is roughly 292 years.
+	dur := time.Duration(ts.ToNsecCapped()) * time.Nanosecond
+	return 0, nil, clockNanosleepFor(t, t.Kernel().MonotonicClock(), dur, rem)
+}
+
+// ClockNanosleep implements linux syscall clock_nanosleep(2).
+func ClockNanosleep(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	clockID := int32(args[0].Int())
+	flags := args[1].Int()
+	addr := args[2].Pointer()
+	rem := args[3].Pointer()
+
+	req, err := copyTimespecIn(t, addr)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if !req.Valid() {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Only allow clock constants also allowed by Linux.
+	if clockID > 0 {
+		if clockID != linux.CLOCK_REALTIME &&
+			clockID != linux.CLOCK_MONOTONIC &&
+			clockID != linux.CLOCK_PROCESS_CPUTIME_ID {
+			return 0, nil, syserror.EINVAL
+		}
+	}
+
+	c, err := getClock(t, clockID)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	if flags&linux.TIMER_ABSTIME != 0 {
+		return 0, nil, clockNanosleepUntil(t, c, req)
+	}
+
+	dur := time.Duration(req.ToNsecCapped()) * time.Nanosecond
+	return 0, nil, clockNanosleepFor(t, c, dur, rem)
+}
+
+// Gettimeofday implements linux syscall gettimeofday(2).
+func Gettimeofday(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	tv := args[0].Pointer()
+	tz := args[1].Pointer()
+
+	if tv != usermem.Addr(0) {
+		nowTv := t.Kernel().RealtimeClock().Now().Timeval()
+		if err := copyTimevalOut(t, tv, &nowTv); err != nil {
+			return 0, nil, err
+		}
+	}
+
+	if tz != usermem.Addr(0) {
+		// Ask the time package for the timezone.
+		_, offset := time.Now().Zone()
+		// This int32 array mimics linux's struct timezone.
+		timezone := [2]int32{-int32(offset) / 60, 0}
+		_, err := t.CopyOut(tz, timezone)
+		return 0, nil, err
+	}
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_timer.go b/pkg/sentry/syscalls/linux/sys_timer.go
new file mode 100644
index 000000000..4ed077626
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_timer.go
@@ -0,0 +1,168 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+)
+
+// ItimerType denotes the type of interval timer.
+type ItimerType int
+
+// Interval timer types from <sys/time.h>.
+const (
+	// ItimerReal equals to ITIMER_REAL.
+	ItimerReal ItimerType = iota
+	// ItimerVirtual equals to ITIMER_VIRTUAL.
+	ItimerVirtual
+	// ItimerProf equals to ITIMER_PROF.
+	ItimerProf
+)
+
+const nsecPerSec = int64(time.Second)
+
+// copyItimerValIn copies an ItimerVal from the untrusted app range to the
+// kernel.  The ItimerVal may be either 32 or 64 bits.
+// A NULL address is allowed because because Linux allows
+// setitimer(which, NULL, &old_value) which disables the timer.
+// There is a KERN_WARN message saying this misfeature will be removed.
+// However, that hasn't happened as of 3.19, so we continue to support it.
+func copyItimerValIn(t *kernel.Task, addr usermem.Addr) (linux.ItimerVal, error) {
+	if addr == usermem.Addr(0) {
+		return linux.ItimerVal{}, nil
+	}
+
+	switch t.Arch().Width() {
+	case 8:
+		// Native size, just copy directly.
+		var itv linux.ItimerVal
+		if _, err := t.CopyIn(addr, &itv); err != nil {
+			return linux.ItimerVal{}, err
+		}
+
+		return itv, nil
+	default:
+		return linux.ItimerVal{}, syscall.ENOSYS
+	}
+}
+
+// copyItimerValOut copies an ItimerVal to the untrusted app range.
+// The ItimerVal may be either 32 or 64 bits.
+// A NULL address is allowed, in which case no copy takes place
+func copyItimerValOut(t *kernel.Task, addr usermem.Addr, itv *linux.ItimerVal) error {
+	if addr == usermem.Addr(0) {
+		return nil
+	}
+
+	switch t.Arch().Width() {
+	case 8:
+		// Native size, just copy directly.
+		_, err := t.CopyOut(addr, itv)
+		return err
+	default:
+		return syscall.ENOSYS
+	}
+}
+
+func findTimer(t *kernel.Task, w ItimerType) (*ktime.Timer, error) {
+	switch w {
+	case ItimerReal:
+		return t.ThreadGroup().Timer().RealTimer, nil
+	case ItimerVirtual:
+		return t.ThreadGroup().Timer().VirtualTimer, nil
+	case ItimerProf:
+		return t.ThreadGroup().Timer().ProfTimer, nil
+	default:
+		return nil, syscall.EINVAL
+	}
+}
+
+// Getitimer implements linux syscall getitimer(2).
+func Getitimer(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := ItimerType(args[0].Int())
+	val := args[1].Pointer()
+
+	timer, err := findTimer(t, timerID)
+	if err != nil {
+		return 0, nil, err
+	}
+	value, interval := ktime.SpecFromSetting(timer.Get())
+	olditv := linux.ItimerVal{
+		Value:    linux.DurationToTimeval(value),
+		Interval: linux.DurationToTimeval(interval),
+	}
+
+	return 0, nil, copyItimerValOut(t, val, &olditv)
+}
+
+// Setitimer implements linux syscall setitimer(2).
+func Setitimer(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	timerID := ItimerType(args[0].Int())
+	newVal := args[1].Pointer()
+	oldVal := args[2].Pointer()
+
+	timer, err := findTimer(t, timerID)
+	if err != nil {
+		return 0, nil, err
+	}
+
+	itv, err := copyItimerValIn(t, newVal)
+	if err != nil {
+		return 0, nil, err
+	}
+	// Just like linux, we cap the timer value and interval with the max
+	// number that int64 can represent which is roughly 292 years.
+	s, err := ktime.SettingFromSpec(itv.Value.ToDuration(),
+		itv.Interval.ToDuration(), timer.Clock())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	valueNS, intervalNS := ktime.SpecFromSetting(timer.Swap(s))
+	olditv := linux.ItimerVal{
+		Value:    linux.DurationToTimeval(valueNS),
+		Interval: linux.DurationToTimeval(intervalNS),
+	}
+
+	return 0, nil, copyItimerValOut(t, oldVal, &olditv)
+}
+
+// Alarm implements linux syscall alarm(2).
+func Alarm(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	duration := time.Duration(args[0].Uint()) * time.Second
+
+	timer := t.ThreadGroup().Timer().RealTimer
+	s, err := ktime.SettingFromSpec(duration, 0, timer.Clock())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	value, _ := ktime.SpecFromSetting(timer.Swap(s))
+	sec := int64(value) / nsecPerSec
+	nsec := int64(value) % nsecPerSec
+	// We can't return 0 if we have an alarm pending ...
+	if (sec == 0 && nsec > 0) || nsec >= nsecPerSec/2 {
+		sec++
+	}
+
+	return uintptr(sec), nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_timerfd.go b/pkg/sentry/syscalls/linux/sys_timerfd.go
new file mode 100644
index 000000000..cb81d42b9
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_timerfd.go
@@ -0,0 +1,135 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs/timerfd"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// TimerfdCreate implements Linux syscall timerfd_create(2).
+func TimerfdCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	clockID := args[0].Int()
+	flags := args[1].Int()
+
+	if flags&^(linux.TFD_CLOEXEC|linux.TFD_NONBLOCK) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	var c ktime.Clock
+	switch clockID {
+	case linux.CLOCK_REALTIME:
+		c = t.Kernel().RealtimeClock()
+	case linux.CLOCK_MONOTONIC:
+		c = t.Kernel().MonotonicClock()
+	default:
+		return 0, nil, syserror.EINVAL
+	}
+	f := timerfd.NewFile(t, c)
+	defer f.DecRef()
+	f.SetFlags(fs.SettableFileFlags{
+		NonBlocking: flags&linux.TFD_NONBLOCK != 0,
+	})
+
+	fd, err := t.FDMap().NewFDFrom(0, f, kernel.FDFlags{
+		CloseOnExec: flags&linux.TFD_CLOEXEC != 0,
+	}, t.ThreadGroup().Limits())
+	if err != nil {
+		return 0, nil, err
+	}
+
+	return uintptr(fd), nil, nil
+}
+
+// TimerfdSettime implements Linux syscall timerfd_settime(2).
+func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	flags := args[1].Int()
+	newValAddr := args[2].Pointer()
+	oldValAddr := args[3].Pointer()
+
+	if flags&^(linux.TFD_TIMER_ABSTIME) != 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	f := t.FDMap().GetFile(fd)
+	if f == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer f.DecRef()
+
+	tf, ok := f.FileOperations.(*timerfd.TimerOperations)
+	if !ok {
+		return 0, nil, syserror.EINVAL
+	}
+
+	var newVal linux.Itimerspec
+	if _, err := t.CopyIn(newValAddr, &newVal); err != nil {
+		return 0, nil, err
+	}
+	var s ktime.Setting
+	var err error
+	if flags&linux.TFD_TIMER_ABSTIME != 0 {
+		s, err = ktime.SettingFromAbsSpec(ktime.FromTimespec(newVal.Value),
+			newVal.Interval.ToDuration())
+	} else {
+		s, err = ktime.SettingFromSpec(newVal.Value.ToDuration(),
+			newVal.Interval.ToDuration(), tf.Clock())
+	}
+	if err != nil {
+		return 0, nil, err
+	}
+	valueNS, intervalNS := ktime.SpecFromSetting(tf.SetTime(s))
+	if oldValAddr == 0 {
+		return 0, nil, nil
+	}
+	oldVal := linux.Itimerspec{
+		Interval: linux.DurationToTimespec(intervalNS),
+		Value:    linux.DurationToTimespec(valueNS),
+	}
+	_, err = t.CopyOut(oldValAddr, &oldVal)
+	return 0, nil, err
+}
+
+// TimerfdGettime implements Linux syscall timerfd_gettime(2).
+func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	curValAddr := args[1].Pointer()
+
+	f := t.FDMap().GetFile(fd)
+	if f == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer f.DecRef()
+
+	tf, ok := f.FileOperations.(*timerfd.TimerOperations)
+	if !ok {
+		return 0, nil, syserror.EINVAL
+	}
+
+	valueNS, intervalNS := ktime.SpecFromSetting(tf.GetTime())
+	curVal := linux.Itimerspec{
+		Interval: linux.DurationToTimespec(intervalNS),
+		Value:    linux.DurationToTimespec(valueNS),
+	}
+	_, err := t.CopyOut(curValAddr, &curVal)
+	return 0, nil, err
+}
diff --git a/pkg/sentry/syscalls/linux/sys_tls.go b/pkg/sentry/syscalls/linux/sys_tls.go
new file mode 100644
index 000000000..1047364b3
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_tls.go
@@ -0,0 +1,48 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+//+build amd64
+
+package linux
+
+import (
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+)
+
+// ArchPrctl implements linux syscall arch_prctl(2).
+// It sets architecture-specific process or thread state for t.
+func ArchPrctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	switch args[0].Int() {
+	case linux.ARCH_GET_FS:
+		addr := args[1].Pointer()
+		_, err := t.CopyOut(addr, &t.Arch().StateData().Regs.Fs_base)
+		if err != nil {
+			return 0, nil, err
+		}
+
+	case linux.ARCH_SET_FS:
+		regs := &t.Arch().StateData().Regs
+		regs.Fs = 0
+		regs.Fs_base = args[1].Uint64()
+
+	default:
+		return 0, nil, syscall.EINVAL
+	}
+
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_utsname.go b/pkg/sentry/syscalls/linux/sys_utsname.go
new file mode 100644
index 000000000..899116374
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_utsname.go
@@ -0,0 +1,89 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Uname implements linux syscall uname.
+func Uname(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	version := t.SyscallTable().Version
+
+	uts := t.UTSNamespace()
+
+	// Fill in structure fields.
+	var u linux.UtsName
+	copy(u.Sysname[:], version.Sysname)
+	copy(u.Nodename[:], uts.HostName())
+	copy(u.Release[:], version.Release)
+	copy(u.Version[:], version.Version)
+	copy(u.Machine[:], "x86_64") // +build tag above.
+	copy(u.Domainname[:], uts.DomainName())
+
+	// Copy out the result.
+	va := args[0].Pointer()
+	_, err := t.CopyOut(va, u)
+	return 0, nil, err
+}
+
+// Setdomainname implements Linux syscall setdomainname.
+func Setdomainname(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	nameAddr := args[0].Pointer()
+	size := args[1].Int()
+
+	utsns := t.UTSNamespace()
+	if !t.HasCapabilityIn(linux.CAP_SYS_ADMIN, utsns.UserNamespace()) {
+		return 0, nil, syserror.EPERM
+	}
+	if size < 0 || size > linux.UTSLen {
+		return 0, nil, syserror.EINVAL
+	}
+
+	name, err := t.CopyInString(nameAddr, int(size))
+	if err != nil {
+		return 0, nil, err
+	}
+
+	utsns.SetDomainName(name)
+	return 0, nil, nil
+}
+
+// Sethostname implements Linux syscall sethostname.
+func Sethostname(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	nameAddr := args[0].Pointer()
+	size := args[1].Int()
+
+	utsns := t.UTSNamespace()
+	if !t.HasCapabilityIn(linux.CAP_SYS_ADMIN, utsns.UserNamespace()) {
+		return 0, nil, syserror.EPERM
+	}
+	if size < 0 || size > linux.UTSLen {
+		return 0, nil, syserror.EINVAL
+	}
+
+	name, err := t.CopyInString(nameAddr, int(size))
+	if err != nil {
+		return 0, nil, err
+	}
+
+	utsns.SetHostName(name)
+	return 0, nil, nil
+}
diff --git a/pkg/sentry/syscalls/linux/sys_write.go b/pkg/sentry/syscalls/linux/sys_write.go
new file mode 100644
index 000000000..caa7b01ea
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/sys_write.go
@@ -0,0 +1,274 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+const (
+	// EventMaskWrite contains events that can be triggered on writes.
+	//
+	// Note that EventHUp is not going to happen for pipes but may for
+	// implementations of poll on some sockets, see net/core/datagram.c.
+	EventMaskWrite = waiter.EventOut | waiter.EventHUp | waiter.EventErr
+)
+
+// Write implements linux syscall write(2).
+func Write(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := args[2].SizeT()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the file is writable.
+	if !file.Flags().Write {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Check that the size is legitimate.
+	si := int(size)
+	if si < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Get the source of the write.
+	src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := writev(t, file, src)
+	t.IOUsage().AccountWriteSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "write", file)
+}
+
+// Pwrite64 implements linux syscall pwrite64(2).
+func Pwrite64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	size := args[2].SizeT()
+	offset := args[3].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the offset is legitimate.
+	if offset < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Is writing at an offset supported?
+	if !file.Flags().Pwrite {
+		return 0, nil, syserror.ESPIPE
+	}
+
+	// Check that the file is writable.
+	if !file.Flags().Write {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Check that the size is legitimate.
+	si := int(size)
+	if si < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Get the source of the write.
+	src, err := t.SingleIOSequence(addr, si, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := pwritev(t, file, src, offset)
+	t.IOUsage().AccountWriteSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "pwrite64", file)
+}
+
+// Writev implements linux syscall writev(2).
+func Writev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	iovcnt := int(args[2].Int())
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the file is writable.
+	if !file.Flags().Write {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Read the iovecs that specify the source of the write.
+	src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := writev(t, file, src)
+	t.IOUsage().AccountWriteSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "writev", file)
+}
+
+// Pwritev implements linux syscall pwritev(2).
+func Pwritev(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+	fd := kdefs.FD(args[0].Int())
+	addr := args[1].Pointer()
+	iovcnt := int(args[2].Int())
+	offset := args[3].Int64()
+
+	file := t.FDMap().GetFile(fd)
+	if file == nil {
+		return 0, nil, syserror.EBADF
+	}
+	defer file.DecRef()
+
+	// Check that the offset is legitimate.
+	if offset < 0 {
+		return 0, nil, syserror.EINVAL
+	}
+
+	// Is writing at an offset supported?
+	if !file.Flags().Pwrite {
+		return 0, nil, syserror.ESPIPE
+	}
+
+	// Check that the file is writable.
+	if !file.Flags().Write {
+		return 0, nil, syserror.EBADF
+	}
+
+	// Read the iovecs that specify the source of the write.
+	src, err := t.IovecsIOSequence(addr, iovcnt, usermem.IOOpts{
+		AddressSpaceActive: true,
+	})
+	if err != nil {
+		return 0, nil, err
+	}
+
+	n, err := pwritev(t, file, src, offset)
+	t.IOUsage().AccountWriteSyscall(n)
+	return uintptr(n), nil, handleIOError(t, n != 0, err, kernel.ERESTARTSYS, "pwritev", file)
+}
+
+func writev(t *kernel.Task, f *fs.File, src usermem.IOSequence) (int64, error) {
+	n, err := f.Writev(t, src)
+	if err != syserror.ErrWouldBlock || f.Flags().NonBlocking {
+		if n > 0 {
+			// Queue notification if we wrote anything.
+			f.Dirent.InotifyEvent(linux.IN_MODIFY, 0)
+		}
+		return n, err
+	}
+
+	// Register for notifications.
+	w, ch := waiter.NewChannelEntry(nil)
+	f.EventRegister(&w, EventMaskWrite)
+
+	total := n
+	for {
+		// Shorten src to reflect bytes previously written.
+		src = src.DropFirst64(n)
+
+		// Issue the request and break out if it completes with
+		// anything other than "would block".
+		n, err = f.Writev(t, src)
+		total += n
+		if err != syserror.ErrWouldBlock {
+			break
+		}
+
+		// Wait for a notification that we should retry.
+		if err = t.Block(ch); err != nil {
+			break
+		}
+	}
+
+	f.EventUnregister(&w)
+
+	if total > 0 {
+		// Queue notification if we wrote anything.
+		f.Dirent.InotifyEvent(linux.IN_MODIFY, 0)
+	}
+
+	return total, err
+}
+
+func pwritev(t *kernel.Task, f *fs.File, src usermem.IOSequence, offset int64) (int64, error) {
+	n, err := f.Pwritev(t, src, offset)
+	if err != syserror.ErrWouldBlock || f.Flags().NonBlocking {
+		if n > 0 {
+			// Queue notification if we wrote anything.
+			f.Dirent.InotifyEvent(linux.IN_MODIFY, 0)
+		}
+		return n, err
+	}
+
+	// Register for notifications.
+	w, ch := waiter.NewChannelEntry(nil)
+	f.EventRegister(&w, EventMaskWrite)
+
+	total := n
+	for {
+		// Shorten src to reflect bytes previously written.
+		src = src.DropFirst64(n)
+
+		// Issue the request and break out if it completes with
+		// anything other than "would block".
+		n, err = f.Pwritev(t, src, offset+total)
+		total += n
+		if err != syserror.ErrWouldBlock {
+			break
+		}
+
+		// Wait for a notification that we should retry.
+		if err = t.Block(ch); err != nil {
+			break
+		}
+	}
+
+	f.EventUnregister(&w)
+
+	if total > 0 {
+		// Queue notification if we wrote anything.
+		f.Dirent.InotifyEvent(linux.IN_MODIFY, 0)
+	}
+
+	return total, err
+}
diff --git a/pkg/sentry/syscalls/linux/timespec.go b/pkg/sentry/syscalls/linux/timespec.go
new file mode 100644
index 000000000..e865c6fc0
--- /dev/null
+++ b/pkg/sentry/syscalls/linux/timespec.go
@@ -0,0 +1,112 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package linux
+
+import (
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// copyTimespecIn copies a Timespec from the untrusted app range to the kernel.
+func copyTimespecIn(t *kernel.Task, addr usermem.Addr) (linux.Timespec, error) {
+	switch t.Arch().Width() {
+	case 8:
+		ts := linux.Timespec{}
+		in := t.CopyScratchBuffer(16)
+		_, err := t.CopyInBytes(addr, in)
+		if err != nil {
+			return ts, err
+		}
+		ts.Sec = int64(usermem.ByteOrder.Uint64(in[0:]))
+		ts.Nsec = int64(usermem.ByteOrder.Uint64(in[8:]))
+		return ts, nil
+	default:
+		return linux.Timespec{}, syserror.ENOSYS
+	}
+}
+
+// copyTimespecOut copies a Timespec to the untrusted app range.
+func copyTimespecOut(t *kernel.Task, addr usermem.Addr, ts *linux.Timespec) error {
+	switch t.Arch().Width() {
+	case 8:
+		out := t.CopyScratchBuffer(16)
+		usermem.ByteOrder.PutUint64(out[0:], uint64(ts.Sec))
+		usermem.ByteOrder.PutUint64(out[8:], uint64(ts.Nsec))
+		_, err := t.CopyOutBytes(addr, out)
+		return err
+	default:
+		return syserror.ENOSYS
+	}
+}
+
+// copyTimevalIn copies a Timeval from the untrusted app range to the kernel.
+func copyTimevalIn(t *kernel.Task, addr usermem.Addr) (linux.Timeval, error) {
+	switch t.Arch().Width() {
+	case 8:
+		tv := linux.Timeval{}
+		in := t.CopyScratchBuffer(16)
+		_, err := t.CopyInBytes(addr, in)
+		if err != nil {
+			return tv, err
+		}
+		tv.Sec = int64(usermem.ByteOrder.Uint64(in[0:]))
+		tv.Usec = int64(usermem.ByteOrder.Uint64(in[8:]))
+		return tv, nil
+	default:
+		return linux.Timeval{}, syscall.ENOSYS
+	}
+}
+
+// copyTimevalOut copies a Timeval to the untrusted app range.
+func copyTimevalOut(t *kernel.Task, addr usermem.Addr, tv *linux.Timeval) error {
+	switch t.Arch().Width() {
+	case 8:
+		out := t.CopyScratchBuffer(16)
+		usermem.ByteOrder.PutUint64(out[0:], uint64(tv.Sec))
+		usermem.ByteOrder.PutUint64(out[8:], uint64(tv.Usec))
+		_, err := t.CopyOutBytes(addr, out)
+		return err
+	default:
+		return syscall.ENOSYS
+	}
+}
+
+// copyTimespecInToDuration copies a Timespec from the untrusted app range,
+// validates it and converts it to a Duration.
+//
+// If the Timespec is larger than what can be represented in a Duration, the
+// returned value is the maximum that Duration will allow.
+//
+// If timespecAddr is NULL, the returned value is negative.
+func copyTimespecInToDuration(t *kernel.Task, timespecAddr usermem.Addr) (time.Duration, error) {
+	// Use a negative Duration to indicate "no timeout".
+	timeout := time.Duration(-1)
+	if timespecAddr != 0 {
+		timespec, err := copyTimespecIn(t, timespecAddr)
+		if err != nil {
+			return 0, err
+		}
+		if !timespec.Valid() {
+			return 0, syscall.EINVAL
+		}
+		timeout = time.Duration(timespec.ToNsecCapped())
+	}
+	return timeout, nil
+}
diff --git a/pkg/sentry/syscalls/polling.go b/pkg/sentry/syscalls/polling.go
new file mode 100644
index 000000000..fd90184ef
--- /dev/null
+++ b/pkg/sentry/syscalls/polling.go
@@ -0,0 +1,137 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package syscalls
+
+import (
+	"syscall"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/fs"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel/kdefs"
+	"gvisor.googlesource.com/gvisor/pkg/waiter"
+)
+
+// PollFD describes a pollable FD.
+type PollFD struct {
+	FD      kdefs.FD
+	Events  waiter.EventMask
+	REvents waiter.EventMask
+}
+
+// pollState tracks the associated file descriptor and waiter of a PollFD.
+type pollState struct {
+	file   *fs.File
+	waiter waiter.Entry
+}
+
+// initReadiness gets the current ready mask for the file represented by the FD
+// stored in pfd.FD. If a channel is passed in, the waiter entry in "state" is
+// used to register with the file for event notifications, and a reference to
+// the file is stored in "state".
+func (pfd *PollFD) initReadiness(t *kernel.Task, state *pollState, ch chan struct{}) {
+	if pfd.FD < 0 {
+		pfd.REvents = 0
+		return
+	}
+
+	file := t.FDMap().GetFile(pfd.FD)
+	if file == nil {
+		pfd.REvents = waiter.EventNVal
+		return
+	}
+
+	if ch == nil {
+		defer file.DecRef()
+	} else {
+		state.file = file
+		state.waiter, _ = waiter.NewChannelEntry(ch)
+		file.EventRegister(&state.waiter, pfd.Events)
+	}
+
+	pfd.REvents = file.Readiness(pfd.Events) & pfd.Events
+}
+
+// releaseState releases all the pollState in "state".
+func releaseState(state []pollState) {
+	for i := range state {
+		if state[i].file != nil {
+			state[i].file.EventUnregister(&state[i].waiter)
+			state[i].file.DecRef()
+		}
+	}
+}
+
+// Poll polls the PollFDs in "pfd" with a bounded time specified in "timeout"
+// when "timeout" is greater than zero.
+//
+// Poll returns the remaining timeout, which is always 0 on a timeout; and 0 or
+// positive if interrupted by a signal.
+func Poll(t *kernel.Task, pfd []PollFD, timeout time.Duration) (time.Duration, uintptr, error) {
+	var ch chan struct{}
+	if timeout != 0 {
+		ch = make(chan struct{}, 1)
+	}
+
+	// Register for event notification in the files involved if we may
+	// block (timeout not zero). Once we find a file that has a non-zero
+	// result, we stop registering for events but still go through all files
+	// to get their ready masks.
+	state := make([]pollState, len(pfd))
+	defer releaseState(state)
+	n := uintptr(0)
+	for i := range pfd {
+		pfd[i].initReadiness(t, &state[i], ch)
+		if pfd[i].REvents != 0 {
+			n++
+			ch = nil
+		}
+	}
+
+	if timeout == 0 {
+		return timeout, n, nil
+	}
+
+	forever := timeout < 0
+
+	for n == 0 {
+		var err error
+		// Wait for a notification.
+		timeout, err = t.BlockWithTimeout(ch, !forever, timeout)
+		if err != nil {
+			if err == syscall.ETIMEDOUT {
+				err = nil
+			}
+			return timeout, 0, err
+		}
+
+		// We got notified, count how many files are ready. If none,
+		// then this was a spurious notification, and we just go back
+		// to sleep with the remaining timeout.
+		for i := range state {
+			if state[i].file == nil {
+				continue
+			}
+
+			ready := state[i].file.Readiness(pfd[i].Events) & pfd[i].Events
+			if ready != 0 {
+				pfd[i].REvents = ready
+				n++
+			}
+		}
+	}
+
+	return timeout, n, nil
+}
diff --git a/pkg/sentry/syscalls/syscalls.go b/pkg/sentry/syscalls/syscalls.go
new file mode 100644
index 000000000..1176f858d
--- /dev/null
+++ b/pkg/sentry/syscalls/syscalls.go
@@ -0,0 +1,72 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package syscalls is the interface from the application to the kernel.
+// Traditionally, syscalls is the interface that is used by applications to
+// request services from the kernel of a operating system. We provide a
+// user-mode kernel that needs to handle those requests coming from unmodified
+// applications. Therefore, we still use the term "syscalls" to denote this
+// interface.
+//
+// Note that the stubs in this package may merely provide the interface, not
+// the actual implementation. It just makes writing syscall stubs
+// straightforward.
+package syscalls
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/eventchannel"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/arch"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	uspb "gvisor.googlesource.com/gvisor/pkg/sentry/syscalls/unimplemented_syscall_go_proto"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// Error returns a syscall handler that will always give the passed error.
+func Error(err error) kernel.SyscallFn {
+	return func(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+		return 0, nil, err
+	}
+}
+
+// ErrorWithEvent gives a syscall function that sends an unimplemented
+// syscall event via the event channel and returns the passed error.
+func ErrorWithEvent(err error) kernel.SyscallFn {
+	return func(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+		UnimplementedEvent(t)
+		return 0, nil, err
+	}
+}
+
+// CapError gives a syscall function that checks for capability c.  If the task
+// has the capability, it returns ENOSYS, otherwise EPERM. To unprivileged
+// tasks, it will seem like there is an implementation.
+func CapError(c linux.Capability) kernel.SyscallFn {
+	return func(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) {
+		if !t.HasCapability(c) {
+			return 0, nil, syserror.EPERM
+		}
+		UnimplementedEvent(t)
+		return 0, nil, syserror.ENOSYS
+	}
+}
+
+// UnimplementedEvent emits an UnimplementedSyscall event via the event
+// channel.
+func UnimplementedEvent(t *kernel.Task) {
+	eventchannel.Emit(&uspb.UnimplementedSyscall{
+		Tid:       int32(t.ThreadID()),
+		Registers: t.Arch().StateData().Proto(),
+	})
+}
diff --git a/pkg/sentry/syscalls/unimplemented_syscall.proto b/pkg/sentry/syscalls/unimplemented_syscall.proto
new file mode 100644
index 000000000..d6febf5b1
--- /dev/null
+++ b/pkg/sentry/syscalls/unimplemented_syscall.proto
@@ -0,0 +1,27 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+syntax = "proto3";
+
+package gvisor;
+
+import "pkg/sentry/arch/registers.proto";
+
+message UnimplementedSyscall {
+  // Task ID.
+  int32 tid = 1;
+
+  // Registers at the time of the call.
+  Registers registers = 2;
+}
diff --git a/pkg/sentry/time/BUILD b/pkg/sentry/time/BUILD
new file mode 100644
index 000000000..cbcd699d5
--- /dev/null
+++ b/pkg/sentry/time/BUILD
@@ -0,0 +1,48 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+
+go_template_instance(
+    name = "seqatomic_parameters",
+    out = "seqatomic_parameters.go",
+    package = "time",
+    suffix = "Parameters",
+    template = "//pkg/sync:generic_seqatomic",
+    types = {
+        "Value": "Parameters",
+    },
+)
+
+go_library(
+    name = "time",
+    srcs = [
+        "calibrated_clock.go",
+        "clock_id.go",
+        "clocks.go",
+        "muldiv_amd64.s",
+        "parameters.go",
+        "sampler.go",
+        "sampler_unsafe.go",
+        "seqatomic_parameters.go",
+        "tsc_amd64.s",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/time",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/log",
+        "//pkg/metric",
+        "//pkg/sync",
+        "//pkg/syserror",
+    ],
+)
+
+go_test(
+    name = "time_test",
+    srcs = [
+        "calibrated_clock_test.go",
+        "parameters_test.go",
+        "sampler_test.go",
+    ],
+    embed = [":time"],
+)
diff --git a/pkg/sentry/time/calibrated_clock.go b/pkg/sentry/time/calibrated_clock.go
new file mode 100644
index 000000000..cbb95e2d7
--- /dev/null
+++ b/pkg/sentry/time/calibrated_clock.go
@@ -0,0 +1,269 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package time provides a calibrated clock synchronized to a system reference
+// clock.
+package time
+
+import (
+	"sync"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/metric"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// fallbackMetric tracks failed updates. It is not sync, as it is not critical
+// that all occurrences are captured and CalibratedClock may fallback many
+// times.
+var fallbackMetric = metric.MustCreateNewUint64Metric("/time/fallback", false /* sync */, "Incremented when a clock falls back to system calls due to a failed update")
+
+// CalibratedClock implements a clock that tracks a reference clock.
+//
+// Users should call Update at regular intervals of around approxUpdateInterval
+// to ensure that the clock does not drift significantly from the reference
+// clock.
+type CalibratedClock struct {
+	// mu protects the fields below.
+	// TODO: consider a sequence counter for read locking.
+	mu sync.RWMutex
+
+	// ref sample the reference clock that this clock is calibrated
+	// against.
+	ref *sampler
+
+	// ready indicates that the fields below are ready for use calculating
+	// time.
+	ready bool
+
+	// params are the current timekeeping parameters.
+	params Parameters
+
+	// errorNS is the estimated clock error in nanoseconds.
+	errorNS ReferenceNS
+}
+
+// NewCalibratedClock creates a CalibratedClock that tracks the given ClockID.
+func NewCalibratedClock(c ClockID) *CalibratedClock {
+	return &CalibratedClock{
+		ref: newSampler(c),
+	}
+}
+
+// Debugf logs at debug level.
+func (c *CalibratedClock) Debugf(format string, v ...interface{}) {
+	if log.IsLogging(log.Debug) {
+		args := []interface{}{c.ref.clockID}
+		args = append(args, v...)
+		log.Debugf("CalibratedClock(%v): "+format, args...)
+	}
+}
+
+// Infof logs at debug level.
+func (c *CalibratedClock) Infof(format string, v ...interface{}) {
+	if log.IsLogging(log.Info) {
+		args := []interface{}{c.ref.clockID}
+		args = append(args, v...)
+		log.Infof("CalibratedClock(%v): "+format, args...)
+	}
+}
+
+// Warningf logs at debug level.
+func (c *CalibratedClock) Warningf(format string, v ...interface{}) {
+	if log.IsLogging(log.Warning) {
+		args := []interface{}{c.ref.clockID}
+		args = append(args, v...)
+		log.Warningf("CalibratedClock(%v): "+format, args...)
+	}
+}
+
+// reset forces the clock to restart the calibration process, logging the
+// passed message.
+func (c *CalibratedClock) reset(str string, v ...interface{}) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	c.resetLocked(str, v...)
+}
+
+// resetLocked is equivalent to reset with c.mu already held for writing.
+func (c *CalibratedClock) resetLocked(str string, v ...interface{}) {
+	c.Warningf(str+" Resetting clock; time may jump.", v...)
+	c.ready = false
+	c.ref.Reset()
+	fallbackMetric.Increment()
+}
+
+// updateParams updates the timekeeping parameters based on the passed
+// parameters.
+//
+// actual is the actual estimated timekeeping parameters. The stored parameters
+// may need to be adjusted slightly from these values to compensate for error.
+//
+// Preconditions: c.mu must be held for writing.
+func (c *CalibratedClock) updateParams(actual Parameters) {
+	if !c.ready {
+		// At initial calibration there is nothing to correct.
+		c.params = actual
+		c.ready = true
+
+		c.Infof("ready")
+
+		return
+	}
+
+	// Otherwise, adjust the params to correct for errors.
+	newParams, errorNS, err := errorAdjust(c.params, actual, actual.BaseCycles)
+	if err != nil {
+		// Something is very wrong. Reset and try again from the
+		// beginning.
+		c.resetLocked("Unable to update params: %v.", err)
+		return
+	}
+	logErrorAdjustment(c.ref.clockID, errorNS, c.params, newParams)
+
+	if errorNS.Magnitude() >= MaxClockError {
+		// We should never get such extreme error, something is very
+		// wrong. Reset everything and start again.
+		//
+		// N.B. logErrorAdjustment will have already logged the error
+		// at warning level.
+		//
+		// TODO: We could allow Realtime clock jumps here.
+		c.resetLocked("Extreme clock error.")
+		return
+	}
+
+	c.params = newParams
+	c.errorNS = errorNS
+}
+
+// Update runs the update step of the clock, updating its synchronization with
+// the reference clock.
+//
+// Update returns timekeeping and true with the new timekeeping parameters if
+// the clock is calibrated. Update should be called regularly to prevent the
+// clock from getting significantly out of sync from the reference clock.
+//
+// The returned timekeeping parameters are invalidated on the next call to
+// Update.
+func (c *CalibratedClock) Update() (Parameters, bool) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+
+	if err := c.ref.Sample(); err != nil {
+		c.resetLocked("Unable to update calibrated clock: %v.", err)
+		return Parameters{}, false
+	}
+
+	oldest, newest, ok := c.ref.Range()
+	if !ok {
+		// Not ready yet.
+		return Parameters{}, false
+	}
+
+	minCount := uint64(newest.before - oldest.after)
+	maxCount := uint64(newest.after - oldest.before)
+	refInterval := uint64(newest.ref - oldest.ref)
+
+	// freq hz = count / (interval ns) * (nsPerS ns) / (1 s)
+	nsPerS := uint64(time.Second.Nanoseconds())
+
+	minHz, ok := muldiv64(minCount, nsPerS, refInterval)
+	if !ok {
+		c.resetLocked("Unable to update calibrated clock: (%v - %v) * %v / %v overflows.", newest.before, oldest.after, nsPerS, refInterval)
+		return Parameters{}, false
+	}
+
+	maxHz, ok := muldiv64(maxCount, nsPerS, refInterval)
+	if !ok {
+		c.resetLocked("Unable to update calibrated clock: (%v - %v) * %v / %v overflows.", newest.after, oldest.before, nsPerS, refInterval)
+		return Parameters{}, false
+	}
+
+	c.updateParams(Parameters{
+		Frequency:  (minHz + maxHz) / 2,
+		BaseRef:    newest.ref,
+		BaseCycles: newest.after,
+	})
+
+	return c.params, true
+}
+
+// GetTime returns the current time based on the clock calibration.
+func (c *CalibratedClock) GetTime() (int64, error) {
+	c.mu.RLock()
+
+	if !c.ready {
+		// Fallback to a syscall.
+		now, err := c.ref.Syscall()
+		c.mu.RUnlock()
+		return int64(now), err
+	}
+
+	now := c.ref.Cycles()
+	v, ok := c.params.ComputeTime(now)
+	if !ok {
+		// Something is seriously wrong with the clock. Try
+		// again with syscalls.
+		c.resetLocked("Time computation overflowed. params = %+v, now = %v.", c.params, now)
+		now, err := c.ref.Syscall()
+		c.mu.RUnlock()
+		return int64(now), err
+	}
+
+	c.mu.RUnlock()
+	return v, nil
+}
+
+// CalibratedClocks contains calibrated monotonic and realtime clocks.
+//
+// TODO: We know that Linux runs the monotonic and realtime clocks at
+// the same rate, so rather than tracking both individually, we could do one
+// calibration for both clocks.
+type CalibratedClocks struct {
+	// monotonic is the clock tracking the system monotonic clock.
+	monotonic *CalibratedClock
+
+	// realtime is the realtime equivalent of monotonic.
+	realtime *CalibratedClock
+}
+
+// NewCalibratedClocks creates a CalibratedClocks.
+func NewCalibratedClocks() *CalibratedClocks {
+	return &CalibratedClocks{
+		monotonic: NewCalibratedClock(Monotonic),
+		realtime:  NewCalibratedClock(Realtime),
+	}
+}
+
+// Update implements Clocks.Update.
+func (c *CalibratedClocks) Update() (Parameters, bool, Parameters, bool) {
+	monotonicParams, monotonicOk := c.monotonic.Update()
+	realtimeParams, realtimeOk := c.realtime.Update()
+
+	return monotonicParams, monotonicOk, realtimeParams, realtimeOk
+}
+
+// GetTime implements Clocks.GetTime.
+func (c *CalibratedClocks) GetTime(id ClockID) (int64, error) {
+	switch id {
+	case Monotonic:
+		return c.monotonic.GetTime()
+	case Realtime:
+		return c.realtime.GetTime()
+	default:
+		return 0, syserror.EINVAL
+	}
+}
diff --git a/pkg/sentry/time/calibrated_clock_test.go b/pkg/sentry/time/calibrated_clock_test.go
new file mode 100644
index 000000000..8b6dd5592
--- /dev/null
+++ b/pkg/sentry/time/calibrated_clock_test.go
@@ -0,0 +1,186 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"testing"
+	"time"
+)
+
+// newTestCalibratedClock returns a CalibratedClock that collects samples from
+// the given sample list and cycle counts from the given cycle list.
+func newTestCalibratedClock(samples []sample, cycles []TSCValue) *CalibratedClock {
+	return &CalibratedClock{
+		ref: newTestSampler(samples, cycles),
+	}
+}
+
+func TestConstantFrequency(t *testing.T) {
+	// Perfectly constant frequency.
+	samples := []sample{
+		{before: 100000, after: 100000 + defaultOverheadCycles, ref: 100},
+		{before: 200000, after: 200000 + defaultOverheadCycles, ref: 200},
+		{before: 300000, after: 300000 + defaultOverheadCycles, ref: 300},
+		{before: 400000, after: 400000 + defaultOverheadCycles, ref: 400},
+		{before: 500000, after: 500000 + defaultOverheadCycles, ref: 500},
+		{before: 600000, after: 600000 + defaultOverheadCycles, ref: 600},
+		{before: 700000, after: 700000 + defaultOverheadCycles, ref: 700},
+	}
+
+	c := newTestCalibratedClock(samples, nil)
+
+	// Update from all samples.
+	for range samples {
+		c.Update()
+	}
+
+	c.mu.RLock()
+	if !c.ready {
+		c.mu.RUnlock()
+		t.Fatalf("clock not ready")
+	}
+	// A bit after the last sample.
+	now, ok := c.params.ComputeTime(750000)
+	c.mu.RUnlock()
+	if !ok {
+		t.Fatalf("ComputeTime ok got %v want true", ok)
+	}
+
+	t.Logf("now: %v", now)
+
+	// Time should be between the current sample and where we'd expect the
+	// next sample.
+	if now < 700 || now > 800 {
+		t.Errorf("now got %v want > 700 && < 800", now)
+	}
+}
+
+func TestErrorCorrection(t *testing.T) {
+	testCases := []struct {
+		name               string
+		samples            [5]sample
+		projectedTimeStart int64
+		projectedTimeEnd   int64
+	}{
+		// Initial calibration should be ~1MHz for each of these, and
+		// the reference clock changes in samples[2].
+		{
+			name: "slow-down",
+			samples: [5]sample{
+				{before: 1000000, after: 1000001, ref: ReferenceNS(1 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 2000000, after: 2000001, ref: ReferenceNS(2 * ApproxUpdateInterval.Nanoseconds())},
+				// Reference clock has slowed down, causing 100ms of error.
+				{before: 3010000, after: 3010001, ref: ReferenceNS(3 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 4020000, after: 4020001, ref: ReferenceNS(4 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 5030000, after: 5030001, ref: ReferenceNS(5 * ApproxUpdateInterval.Nanoseconds())},
+			},
+			projectedTimeStart: 3005 * time.Millisecond.Nanoseconds(),
+			projectedTimeEnd:   3015 * time.Millisecond.Nanoseconds(),
+		},
+		{
+			name: "speed-up",
+			samples: [5]sample{
+				{before: 1000000, after: 1000001, ref: ReferenceNS(1 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 2000000, after: 2000001, ref: ReferenceNS(2 * ApproxUpdateInterval.Nanoseconds())},
+				// Reference clock has sped up, causing 100ms of error.
+				{before: 2990000, after: 2990001, ref: ReferenceNS(3 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 3980000, after: 3980001, ref: ReferenceNS(4 * ApproxUpdateInterval.Nanoseconds())},
+				{before: 4970000, after: 4970001, ref: ReferenceNS(5 * ApproxUpdateInterval.Nanoseconds())},
+			},
+			projectedTimeStart: 2985 * time.Millisecond.Nanoseconds(),
+			projectedTimeEnd:   2995 * time.Millisecond.Nanoseconds(),
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			c := newTestCalibratedClock(tc.samples[:], nil)
+
+			// Initial calibration takes two updates.
+			_, ok := c.Update()
+			if ok {
+				t.Fatalf("Update ready too early")
+			}
+
+			params, ok := c.Update()
+			if !ok {
+				t.Fatalf("Update not ready")
+			}
+
+			// Initial calibration is ~1MHz.
+			hz := params.Frequency
+			if hz < 990000 || hz > 1010000 {
+				t.Fatalf("Frequency got %v want > 990kHz && < 1010kHz", hz)
+			}
+
+			// Project time at the next update. Given the 1MHz
+			// calibration, it is expected to be ~3.1s/2.9s, not
+			// the actual 3s.
+			//
+			// N.B. the next update time is the "after" time above.
+			projected, ok := params.ComputeTime(tc.samples[2].after)
+			if !ok {
+				t.Fatalf("ComputeTime ok got %v want true", ok)
+			}
+			if projected < tc.projectedTimeStart || projected > tc.projectedTimeEnd {
+				t.Fatalf("ComputeTime(%v) got %v want > %v && < %v", tc.samples[2].after, projected, tc.projectedTimeStart, tc.projectedTimeEnd)
+			}
+
+			// Update again to see the changed reference clock.
+			params, ok = c.Update()
+			if !ok {
+				t.Fatalf("Update not ready")
+			}
+
+			// We now know that TSC = tc.samples[2].after -> 3s,
+			// but with the previous params indicated that TSC
+			// tc.samples[2].after -> 3.5s/2.5s. We can't allow the
+			// clock to go backwards, and having the clock jump
+			// forwards is undesirable. There should be a smooth
+			// transition that corrects the clock error over time.
+			// Check that the clock is continuous at TSC =
+			// tc.samples[2].after.
+			newProjected, ok := params.ComputeTime(tc.samples[2].after)
+			if !ok {
+				t.Fatalf("ComputeTime ok got %v want true", ok)
+			}
+			if newProjected != projected {
+				t.Errorf("Discontinuous time; ComputeTime(%v) got %v want %v", tc.samples[2].after, newProjected, projected)
+			}
+
+			// As the reference clock stablizes, ensure that the clock error
+			// decreases.
+			initialErr := c.errorNS
+			t.Logf("initial error: %v ns", initialErr)
+
+			_, ok = c.Update()
+			if !ok {
+				t.Fatalf("Update not ready")
+			}
+			if c.errorNS.Magnitude() > initialErr.Magnitude() {
+				t.Errorf("errorNS increased, got %v want |%v| <= |%v|", c.errorNS, c.errorNS, initialErr)
+			}
+
+			_, ok = c.Update()
+			if !ok {
+				t.Fatalf("Update not ready")
+			}
+			if c.errorNS.Magnitude() > initialErr.Magnitude() {
+				t.Errorf("errorNS increased, got %v want |%v| <= |%v|", c.errorNS, c.errorNS, initialErr)
+			}
+
+			t.Logf("final error: %v ns", c.errorNS)
+		})
+	}
+}
diff --git a/pkg/sentry/time/clock_id.go b/pkg/sentry/time/clock_id.go
new file mode 100644
index 000000000..500102e58
--- /dev/null
+++ b/pkg/sentry/time/clock_id.go
@@ -0,0 +1,40 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"strconv"
+)
+
+// ClockID is a Linux clock identifier.
+type ClockID int32
+
+// These are the supported Linux clock identifiers.
+const (
+	Realtime ClockID = iota
+	Monotonic
+)
+
+// String implements fmt.Stringer.String.
+func (c ClockID) String() string {
+	switch c {
+	case Realtime:
+		return "Realtime"
+	case Monotonic:
+		return "Monotonic"
+	default:
+		return strconv.Itoa(int(c))
+	}
+}
diff --git a/pkg/sentry/time/clocks.go b/pkg/sentry/time/clocks.go
new file mode 100644
index 000000000..9925b407d
--- /dev/null
+++ b/pkg/sentry/time/clocks.go
@@ -0,0 +1,31 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+// Clocks represents a clock source that contains both a monotonic and realtime
+// clock.
+type Clocks interface {
+	// Update performs an update step, keeping the clocks in sync with the
+	// reference host clocks, and returning the new timekeeping parameters.
+	//
+	// Update should be called at approximately ApproxUpdateInterval.
+	Update() (monotonicParams Parameters, monotonicOk bool, realtimeParam Parameters, realtimeOk bool)
+
+	// GetTime returns the current time in nanoseconds for the given clock.
+	//
+	// Clocks implementations must support at least Monotonic and
+	// Realtime.
+	GetTime(c ClockID) (int64, error)
+}
diff --git a/pkg/sentry/time/muldiv_amd64.s b/pkg/sentry/time/muldiv_amd64.s
new file mode 100644
index 000000000..291940b1d
--- /dev/null
+++ b/pkg/sentry/time/muldiv_amd64.s
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+// Documentation is available in parameters.go.
+//
+// func muldiv64(value, multiplier, divisor uint64) (uint64, bool)
+TEXT ·muldiv64(SB),NOSPLIT,$0-33
+	MOVQ value+0(FP), AX
+	MOVQ multiplier+8(FP), BX
+	MOVQ divisor+16(FP), CX
+
+	// Multiply AX*BX and store result in DX:AX.
+	MULQ BX
+
+	// If divisor <= (value*multiplier) / 2^64, then the division will overflow.
+	//
+	// (value*multiplier) / 2^64 is DX:AX >> 64, or simply DX.
+	CMPQ CX, DX
+	JLE overflow
+
+	// Divide DX:AX by CX.
+	DIVQ CX
+
+	MOVQ AX, result+24(FP)
+	MOVB $1, ok+32(FP)
+	RET
+
+overflow:
+	MOVQ $0, result+24(FP)
+	MOVB $0, ok+32(FP)
+	RET
diff --git a/pkg/sentry/time/parameters.go b/pkg/sentry/time/parameters.go
new file mode 100644
index 000000000..594b4874b
--- /dev/null
+++ b/pkg/sentry/time/parameters.go
@@ -0,0 +1,239 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"fmt"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+)
+
+const (
+	// ApproxUpdateInterval is the approximate interval that parameters
+	// should be updated at.
+	//
+	// Error correction assumes that the next update will occur after this
+	// much time.
+	//
+	// If an update occurs before ApproxUpdateInterval passes, it has no
+	// adverse effect on error correction behavior.
+	//
+	// If an update occurs after ApproxUpdateInterval passes, the clock
+	// will overshoot its error correction target and begin accumulating
+	// error in the other direction.
+	//
+	// If updates occur after more than 2*ApproxUpdateInterval passes, the
+	// clock becomes unstable, accumulating more error than it had
+	// originally. Repeated updates after more than 2*ApproxUpdateInterval
+	// will cause unbounded increases in error.
+	//
+	// These statements assume that the host clock does not change. Actual
+	// error will depend upon host clock changes.
+	//
+	// TODO: make error correction more robust to delayed
+	// updates.
+	ApproxUpdateInterval = 1 * time.Second
+
+	// MaxClockError is the maximum amount of error that the clocks will
+	// try to correct.
+	//
+	// This limit:
+	//
+	//  * Puts a limit on cases of otherwise unbounded increases in error.
+	//
+	//  * Avoids unreasonably large frequency adjustments required to
+	//    correct large errors over a single update interval.
+	MaxClockError = ReferenceNS(ApproxUpdateInterval) / 4
+)
+
+// Parameters are the timekeeping parameters needed to compute the current
+// time.
+type Parameters struct {
+	// BaseCycles was the TSC counter value when the time was BaseRef.
+	BaseCycles TSCValue
+
+	// BaseRef is the reference clock time in nanoseconds corresponding to
+	// BaseCycles.
+	BaseRef ReferenceNS
+
+	// Frequency is the frequency of the cycle clock in Hertz.
+	Frequency uint64
+}
+
+// muldiv64 multiplies two 64-bit numbers, then divides the result by another
+// 64-bit number.
+//
+// It requires that the result fit in 64 bits, but doesn't require that
+// intermediate values do; in particular, the result of the multiplication may
+// require 128 bits.
+//
+// It returns !ok if divisor is zero or the result does not fit in 64 bits.
+func muldiv64(value, multiplier, divisor uint64) (uint64, bool)
+
+// ComputeTime calculates the current time from a "now" TSC value.
+//
+// time = ref + (now - base) / f
+func (p Parameters) ComputeTime(nowCycles TSCValue) (int64, bool) {
+	diffCycles := nowCycles - p.BaseCycles
+	if diffCycles < 0 {
+		log.Warningf("now cycles %v < base cycles %v", nowCycles, p.BaseCycles)
+		diffCycles = 0
+	}
+
+	// Overflow "won't ever happen". If diffCycles is the max value
+	// (2^63 - 1), then to overflow,
+	//
+	// frequency <= ((2^63 - 1) * 10^9) / 2^64 = 500Mhz
+	//
+	// A TSC running at 2GHz takes 201 years to reach 2^63-1. 805 years at
+	// 500MHz.
+	diffNS, ok := muldiv64(uint64(diffCycles), uint64(time.Second.Nanoseconds()), p.Frequency)
+	return int64(uint64(p.BaseRef) + diffNS), ok
+}
+
+// errorAdjust returns a new Parameters struct "adjusted" that satisfies:
+//
+// 1. adjusted.ComputeTime(now) = prevParams.ComputeTime(now)
+//   * i.e., the current time does not jump.
+//
+// 2. adjusted.ComputeTime(TSC at next update) = newParams.ComputeTime(TSC at next update)
+//   * i.e., Any error between prevParams and newParams will be corrected over
+//     the course of the next update period.
+//
+// errorAdjust also returns the current clock error.
+//
+// Preconditions:
+// * newParams.BaseCycles >= prevParams.BaseCycles; i.e., TSC must not go
+//   backwards.
+// * newParams.BaseCycles <= now; i.e., the new parameters be computed at or
+//   before now.
+func errorAdjust(prevParams Parameters, newParams Parameters, now TSCValue) (Parameters, ReferenceNS, error) {
+	if newParams.BaseCycles < prevParams.BaseCycles {
+		// Oh dear! Something is very wrong.
+		return Parameters{}, 0, fmt.Errorf("TSC went backwards in updated clock params: %v < %v", newParams.BaseCycles, prevParams.BaseCycles)
+	}
+	if newParams.BaseCycles > now {
+		return Parameters{}, 0, fmt.Errorf("parameters contain base cycles later than now: %v > %v", newParams.BaseCycles, now)
+	}
+
+	intervalNS := int64(ApproxUpdateInterval.Nanoseconds())
+	nsPerSec := uint64(time.Second.Nanoseconds())
+
+	// Current time as computed by prevParams.
+	oldNowNS, ok := prevParams.ComputeTime(now)
+	if !ok {
+		return Parameters{}, 0, fmt.Errorf("old now time computation overflowed. params = %+v, now = %v", prevParams, now)
+	}
+
+	// We expect the update ticker to run based on this clock (i.e., it has
+	// been using prevParams and will use the returned adjusted
+	// parameters). Hence it will decide to fire intervalNS from the
+	// current (oldNowNS) "now".
+	nextNS := oldNowNS + intervalNS
+
+	if nextNS <= int64(newParams.BaseRef) {
+		// The next update time already passed before the new
+		// parameters were created! We definitely can't correct the
+		// error by then.
+		return Parameters{}, 0, fmt.Errorf("unable to correct error in single period. oldNowNS = %v, nextNS = %v, p = %v", oldNowNS, nextNS, newParams)
+	}
+
+	// For what TSC value next will newParams.ComputeTime(next) = nextNS?
+	//
+	// Solve ComputeTime for next:
+	//
+	// next = newParams.Frequency * (nextNS - newParams.BaseRef) + newParams.BaseCycles
+	c, ok := muldiv64(newParams.Frequency, uint64(nextNS-int64(newParams.BaseRef)), nsPerSec)
+	if !ok {
+		return Parameters{}, 0, fmt.Errorf("%v * (%v - %v) / %v overflows", newParams.Frequency, nextNS, newParams.BaseRef, nsPerSec)
+	}
+
+	cycles := TSCValue(c)
+	next := cycles + newParams.BaseCycles
+
+	if next <= now {
+		// The next update time already passed now with the new
+		// parameters! We can't correct the error in a single period.
+		return Parameters{}, 0, fmt.Errorf("unable to correct error in single period. oldNowNS = %v, nextNS = %v, now = %v, next = %v", oldNowNS, nextNS, now, next)
+	}
+
+	// We want to solve for parameters that satisfy:
+	//
+	// adjusted.ComputeTime(now) = oldNowNS
+	//
+	// adjusted.ComputeTime(next) = nextNS
+	//
+	// i.e., the current time does not change, but by the time we reach
+	// next we reach the same time as newParams.
+
+	// We choose to keep BaseCycles fixed.
+	adjusted := Parameters{
+		BaseCycles: newParams.BaseCycles,
+	}
+
+	// We want a slope such that time goes from oldNowNS to nextNS when
+	// we reach next.
+	//
+	// In other words, cycles should increase by next - now in the next
+	// interval.
+
+	cycles = next - now
+	ns := intervalNS
+
+	// adjusted.Frequency = cycles / ns
+	adjusted.Frequency, ok = muldiv64(uint64(cycles), nsPerSec, uint64(ns))
+	if !ok {
+		return Parameters{}, 0, fmt.Errorf("(%v - %v) * %v / %v overflows", next, now, nsPerSec, ns)
+	}
+
+	// Now choose a base reference such that the current time remains the
+	// same. Note that this is just ComputeTime, solving for BaseRef:
+	//
+	// oldNowNS = BaseRef + (now - BaseCycles) / Frequency
+	// BaseRef = oldNowNS - (now - BaseCycles) / Frequency
+	diffNS, ok := muldiv64(uint64(now-adjusted.BaseCycles), nsPerSec, adjusted.Frequency)
+	if !ok {
+		return Parameters{}, 0, fmt.Errorf("(%v - %v) * %v / %v overflows", now, adjusted.BaseCycles, nsPerSec, adjusted.Frequency)
+	}
+
+	adjusted.BaseRef = ReferenceNS(oldNowNS - int64(diffNS))
+
+	// The error is the difference between the current time and what the
+	// new parameters say the current time should be.
+	newNowNS, ok := newParams.ComputeTime(now)
+	if !ok {
+		return Parameters{}, 0, fmt.Errorf("new now time computation overflowed. params = %+v, now = %v", newParams, now)
+	}
+
+	errorNS := ReferenceNS(oldNowNS - newNowNS)
+
+	return adjusted, errorNS, nil
+}
+
+// logErrorAdjustment logs the clock error and associated error correction
+// frequency adjustment.
+//
+// The log level is determined by the error severity.
+func logErrorAdjustment(clock ClockID, errorNS ReferenceNS, orig, adjusted Parameters) {
+	fn := log.Debugf
+	if int64(errorNS.Magnitude()) > time.Millisecond.Nanoseconds() {
+		fn = log.Warningf
+	} else if int64(errorNS.Magnitude()) > 10*time.Microsecond.Nanoseconds() {
+		fn = log.Infof
+	}
+
+	fn("Clock(%v): error: %v ns, adjusted frequency from %v Hz to %v Hz", clock, errorNS, orig.Frequency, adjusted.Frequency)
+}
diff --git a/pkg/sentry/time/parameters_test.go b/pkg/sentry/time/parameters_test.go
new file mode 100644
index 000000000..7394fc5ee
--- /dev/null
+++ b/pkg/sentry/time/parameters_test.go
@@ -0,0 +1,486 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"math"
+	"testing"
+	"time"
+)
+
+func TestParametersComputeTime(t *testing.T) {
+	testCases := []struct {
+		name   string
+		params Parameters
+		now    TSCValue
+		want   int64
+	}{
+		{
+			// Now is the same as the base cycles.
+			name: "base-cycles",
+			params: Parameters{
+				BaseCycles: 10000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now:  10000,
+			want: 5000 * time.Millisecond.Nanoseconds(),
+		},
+		{
+			// Now is the behind the base cycles. Time is frozen.
+			name: "backwards",
+			params: Parameters{
+				BaseCycles: 10000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now:  9000,
+			want: 5000 * time.Millisecond.Nanoseconds(),
+		},
+		{
+			// Now is ahead of the base cycles.
+			name: "ahead",
+			params: Parameters{
+				BaseCycles: 10000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now:  15000,
+			want: 5500 * time.Millisecond.Nanoseconds(),
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			got, ok := tc.params.ComputeTime(tc.now)
+			if !ok {
+				t.Errorf("ComputeTime ok got %v want true", got)
+			}
+			if got != tc.want {
+				t.Errorf("ComputeTime got %+v want %+v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestParametersErrorAdjust(t *testing.T) {
+	testCases := []struct {
+		name      string
+		oldParams Parameters
+		now       TSCValue
+		newParams Parameters
+		want      Parameters
+		errorNS   ReferenceNS
+		wantErr   bool
+	}{
+		{
+			// newParams are perfectly continuous with oldParams
+			// and don't need adjustment.
+			name: "continuous",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 50000,
+			newParams: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			want: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+		},
+		{
+			// Same as "continuous", but with now ahead of
+			// newParams.BaseCycles. The result is the same as
+			// there is no error to correct.
+			name: "continuous-nowdiff",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 60000,
+			newParams: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			want: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+		},
+		{
+			// errorAdjust bails out if the TSC goes backwards.
+			name: "tsc-backwards",
+			oldParams: Parameters{
+				BaseCycles: 10000,
+				BaseRef:    ReferenceNS(1000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now: 9000,
+			newParams: Parameters{
+				BaseCycles: 9000,
+				BaseRef:    ReferenceNS(1100 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			wantErr: true,
+		},
+		{
+			// errorAdjust bails out if new params are from after now.
+			name: "params-after-now",
+			oldParams: Parameters{
+				BaseCycles: 10000,
+				BaseRef:    ReferenceNS(1000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now: 11000,
+			newParams: Parameters{
+				BaseCycles: 12000,
+				BaseRef:    ReferenceNS(1200 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			wantErr: true,
+		},
+		{
+			// Host clock sped up.
+			name: "speed-up",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 45000,
+			// Host frequency changed to 9000 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 45000,
+				// From oldParams, we think ref = 4.5s at cycles = 45000.
+				BaseRef:   ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency: 9000,
+			},
+			want: Parameters{
+				BaseCycles: 45000,
+				BaseRef:    ReferenceNS(4500 * time.Millisecond.Nanoseconds()),
+				// We must decrease the new frequency by 50% to
+				// correct 0.5s of error in 1s
+				// (ApproxUpdateInterval).
+				Frequency: 4500,
+			},
+			errorNS: ReferenceNS(-500 * time.Millisecond.Nanoseconds()),
+		},
+		{
+			// Host clock sped up, with now ahead of newParams.
+			name: "speed-up-nowdiff",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 50000,
+			// Host frequency changed to 9000 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 45000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  9000,
+			},
+			// nextRef = 6000ms
+			// nextCycles = 9000 * (6000ms - 5000ms) + 45000
+			// nextCycles = 9000 * (1s) + 45000
+			// nextCycles = 54000
+			// f = (54000 - 50000) / 1s = 4000
+			//
+			// ref = 5000ms - (50000 - 45000) / 4000
+			// ref = 3.75s
+			want: Parameters{
+				BaseCycles: 45000,
+				BaseRef:    ReferenceNS(3750 * time.Millisecond.Nanoseconds()),
+				Frequency:  4000,
+			},
+			// oldNow = 50000 * 10000 = 5s
+			// newNow = (50000 - 45000) / 9000 + 5s = 5.555s
+			errorNS: ReferenceNS((5000*time.Millisecond - 5555555555).Nanoseconds()),
+		},
+		{
+			// Host clock sped up. The new parameters are so far
+			// ahead that the next update time already passed.
+			name: "speed-up-uncorrectable-baseref",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 50000,
+			// Host frequency changed to 5000 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 45000,
+				BaseRef:    ReferenceNS(9000 * time.Millisecond.Nanoseconds()),
+				Frequency:  5000,
+			},
+			// The next update should be at 10s, but newParams
+			// already passed 6s.  Thus it is impossible to correct
+			// the clock by then.
+			wantErr: true,
+		},
+		{
+			// Host clock sped up. The new parameters are moving so
+			// fast that the next update should be before now.
+			name: "speed-up-uncorrectable-frequency",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 55000,
+			// Host frequency changed to 7500 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 45000,
+				BaseRef:    ReferenceNS(6000 * time.Millisecond.Nanoseconds()),
+				Frequency:  7500,
+			},
+			// The next update should be at 6.5s, but newParams are
+			// so far ahead and fast that they reach 6.5s at cycle
+			// 48750, which before now! Thus it is impossible to
+			// correct the clock by then.
+			wantErr: true,
+		},
+		{
+			// Host clock slowed down.
+			name: "slow-down",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 55000,
+			// Host frequency changed to 11000 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 55000,
+				// From oldParams, we think ref = 5.5s at cycles = 55000.
+				BaseRef:   ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency: 11000,
+			},
+			want: Parameters{
+				BaseCycles: 55000,
+				BaseRef:    ReferenceNS(5500 * time.Millisecond.Nanoseconds()),
+				// We must increase the new frequency by 50% to
+				// correct 0.5s of error in 1s
+				// (ApproxUpdateInterval).
+				Frequency: 16500,
+			},
+			errorNS: ReferenceNS(500 * time.Millisecond.Nanoseconds()),
+		},
+		{
+			// Host clock slowed down, with now ahead of newParams.
+			name: "slow-down-nowdiff",
+			oldParams: Parameters{
+				BaseCycles: 0,
+				BaseRef:    0,
+				Frequency:  10000,
+			},
+			now: 60000,
+			// Host frequency changed to 11000 immediately after
+			// oldParams was returned.
+			newParams: Parameters{
+				BaseCycles: 55000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  11000,
+			},
+			// nextRef = 7000ms
+			// nextCycles = 11000 * (7000ms - 5000ms) + 55000
+			// nextCycles = 11000 * (2000ms) + 55000
+			// nextCycles = 77000
+			// f = (77000 - 60000) / 1s = 17000
+			//
+			// ref = 6000ms - (60000 - 55000) / 17000
+			// ref = 5705882353ns
+			want: Parameters{
+				BaseCycles: 55000,
+				BaseRef:    ReferenceNS(5705882353),
+				Frequency:  17000,
+			},
+			// oldNow = 60000 * 10000 = 6s
+			// newNow = (60000 - 55000) / 11000 + 5s = 5.4545s
+			errorNS: ReferenceNS((6*time.Second - 5454545454).Nanoseconds()),
+		},
+		{
+			// Host time went backwards.
+			name: "time-backwards",
+			oldParams: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now: 60000,
+			newParams: Parameters{
+				BaseCycles: 60000,
+				// From oldParams, we think ref = 6s at cycles = 60000.
+				BaseRef:   ReferenceNS(4000 * time.Millisecond.Nanoseconds()),
+				Frequency: 10000,
+			},
+			want: Parameters{
+				BaseCycles: 60000,
+				BaseRef:    ReferenceNS(6000 * time.Millisecond.Nanoseconds()),
+				// We must increase the frequency by 200% to
+				// correct 2s of error in 1s
+				// (ApproxUpdateInterval).
+				Frequency: 30000,
+			},
+			errorNS: ReferenceNS(2000 * time.Millisecond.Nanoseconds()),
+		},
+		{
+			// Host time went backwards, with now ahead of newParams.
+			name: "time-backwards-nowdiff",
+			oldParams: Parameters{
+				BaseCycles: 50000,
+				BaseRef:    ReferenceNS(5000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			now: 65000,
+			// nextRef = 7500ms
+			// nextCycles = 10000 * (7500ms - 4000ms) + 60000
+			// nextCycles = 10000 * (3500ms) + 60000
+			// nextCycles = 95000
+			// f = (95000 - 65000) / 1s = 30000
+			//
+			// ref = 6500ms - (65000 - 60000) / 30000
+			// ref = 6333333333ns
+			newParams: Parameters{
+				BaseCycles: 60000,
+				BaseRef:    ReferenceNS(4000 * time.Millisecond.Nanoseconds()),
+				Frequency:  10000,
+			},
+			want: Parameters{
+				BaseCycles: 60000,
+				BaseRef:    ReferenceNS(6333333334),
+				Frequency:  30000,
+			},
+			// oldNow = 65000 * 10000 = 6.5s
+			// newNow = (65000 - 60000) / 10000 + 4s = 4.5s
+			errorNS: ReferenceNS(2000 * time.Millisecond.Nanoseconds()),
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			got, errorNS, err := errorAdjust(tc.oldParams, tc.newParams, tc.now)
+			if err != nil && !tc.wantErr {
+				t.Errorf("err got %v want nil", err)
+			} else if err == nil && tc.wantErr {
+				t.Errorf("err got nil want non-nil")
+			}
+
+			if got != tc.want {
+				t.Errorf("Parameters got %+v want %+v", got, tc.want)
+			}
+			if errorNS != tc.errorNS {
+				t.Errorf("errorNS got %v want %v", errorNS, tc.errorNS)
+			}
+		})
+	}
+}
+
+func testMuldiv(t *testing.T, v uint64) {
+	for i := uint64(1); i <= 1000000; i++ {
+		mult := uint64(1000000000)
+		div := i * mult
+		res, ok := muldiv64(v, mult, div)
+		if !ok {
+			t.Errorf("Result of %v * %v / %v ok got false want true", v, mult, div)
+		}
+		if want := v / i; res != want {
+			t.Errorf("Bad result of %v * %v / %v: got %v, want %v", v, mult, div, res, want)
+		}
+	}
+}
+
+func TestMulDiv(t *testing.T) {
+	testMuldiv(t, math.MaxUint64)
+	for i := int64(-10); i <= 10; i++ {
+		testMuldiv(t, uint64(i))
+	}
+}
+
+func TestMulDivZero(t *testing.T) {
+	if r, ok := muldiv64(2, 4, 0); ok {
+		t.Errorf("muldiv64(2, 4, 0) got %d, ok want !ok", r)
+	}
+
+	if r, ok := muldiv64(0, 0, 0); ok {
+		t.Errorf("muldiv64(0, 0, 0) got %d, ok want !ok", r)
+	}
+}
+
+func TestMulDivOverflow(t *testing.T) {
+	testCases := []struct {
+		name string
+		val  uint64
+		mult uint64
+		div  uint64
+		ok   bool
+		ret  uint64
+	}{
+		{
+			name: "2^62",
+			val:  1 << 63,
+			mult: 4,
+			div:  8,
+			ok:   true,
+			ret:  1 << 62,
+		},
+		{
+			name: "2^64-1",
+			val:  0xffffffffffffffff,
+			mult: 1,
+			div:  1,
+			ok:   true,
+			ret:  0xffffffffffffffff,
+		},
+		{
+			name: "2^64",
+			val:  1 << 63,
+			mult: 4,
+			div:  2,
+			ok:   false,
+		},
+		{
+			name: "2^125",
+			val:  1 << 63,
+			mult: 1 << 63,
+			div:  2,
+			ok:   false,
+		},
+	}
+
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			r, ok := muldiv64(tc.val, tc.mult, tc.div)
+			if ok != tc.ok {
+				t.Errorf("ok got %v want %v", ok, tc.ok)
+			}
+			if tc.ok && r != tc.ret {
+				t.Errorf("ret got %v want %v", r, tc.ret)
+			}
+		})
+	}
+}
diff --git a/pkg/sentry/time/sampler.go b/pkg/sentry/time/sampler.go
new file mode 100644
index 000000000..cf581b5fa
--- /dev/null
+++ b/pkg/sentry/time/sampler.go
@@ -0,0 +1,225 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"errors"
+
+	"gvisor.googlesource.com/gvisor/pkg/log"
+)
+
+const (
+	// defaultOverheadTSC is the default estimated syscall overhead in TSC cycles.
+	// It is further refined as syscalls are made.
+	defaultOverheadCycles = 1 * 1000
+
+	// maxOverheadCycles is the maximum allowed syscall overhead in TSC cycles.
+	maxOverheadCycles = 100 * defaultOverheadCycles
+
+	// maxSampleLoops is the maximum number of times to try to get a clock sample
+	// under the expected overhead.
+	maxSampleLoops = 5
+
+	// maxSamples is the maximum number of samples to collect.
+	maxSamples = 10
+)
+
+// errOverheadTooHigh is returned from sampler.Sample if the syscall
+// overhead is too high.
+var errOverheadTooHigh = errors.New("time syscall overhead exceeds maximum")
+
+// TSCValue is a value from the TSC.
+type TSCValue int64
+
+// Rdtsc reads the TSC.
+//
+// Intel SDM, Vol 3, Ch 17.15:
+// "The RDTSC instruction reads the time-stamp counter and is guaranteed to
+// return a monotonically increasing unique value whenever executed, except for
+// a 64-bit counter wraparound. Intel guarantees that the time-stamp counter
+// will not wraparound within 10 years after being reset."
+//
+// We use int64, so we have 5 years before wrap-around.
+func Rdtsc() TSCValue
+
+// ReferenceNS are nanoseconds in the reference clock domain.
+// int64 gives us ~290 years before this overflows.
+type ReferenceNS int64
+
+// Magnitude returns the absolute value of r.
+func (r ReferenceNS) Magnitude() ReferenceNS {
+	if r < 0 {
+		return -r
+	}
+	return r
+}
+
+// cycleClock is a TSC-based cycle clock.
+type cycleClock interface {
+	// Cycles returns a count value from the TSC.
+	Cycles() TSCValue
+}
+
+// tscCycleClock is a cycleClock that uses the real TSC.
+type tscCycleClock struct{}
+
+// Cycles implements cycleClock.Cycles.
+func (tscCycleClock) Cycles() TSCValue {
+	return Rdtsc()
+}
+
+// sample contains a sample from the reference clock, with TSC values from
+// before and after the reference clock value was captured.
+type sample struct {
+	before TSCValue
+	after  TSCValue
+	ref    ReferenceNS
+}
+
+// Overhead returns the sample overhead in TSC cycles.
+func (s *sample) Overhead() TSCValue {
+	return s.after - s.before
+}
+
+// referenceClocks collects individual samples from a reference clock ID and
+// TSC.
+type referenceClocks interface {
+	cycleClock
+
+	// Sample returns a single sample from the reference clock ID.
+	Sample(c ClockID) (sample, error)
+}
+
+// sampler collects samples from a reference system clock, minimizing
+// the overhead in each sample.
+type sampler struct {
+	// clockID is the reference clock ID (e.g., CLOCK_MONOTONIC).
+	clockID ClockID
+
+	// clocks provides raw samples.
+	clocks referenceClocks
+
+	// overhead is the estimated sample overhead in TSC cycles.
+	overhead TSCValue
+
+	// samples is a ring buffer of the latest samples collected.
+	samples []sample
+}
+
+// newSampler creates a sampler for clockID.
+func newSampler(c ClockID) *sampler {
+	return &sampler{
+		clockID:  c,
+		clocks:   syscallTSCReferenceClocks{},
+		overhead: defaultOverheadCycles,
+	}
+}
+
+// Reset discards previously collected clock samples.
+func (s *sampler) Reset() {
+	s.overhead = defaultOverheadCycles
+	s.samples = []sample{}
+}
+
+// lowOverheadSample returns a reference clock sample with minimized syscall overhead.
+func (s *sampler) lowOverheadSample() (sample, error) {
+	for {
+		for i := 0; i < maxSampleLoops; i++ {
+			samp, err := s.clocks.Sample(s.clockID)
+			if err != nil {
+				return sample{}, err
+			}
+
+			if samp.before > samp.after {
+				log.Warningf("TSC went backwards: %v > %v", samp.before, samp.after)
+				continue
+			}
+
+			if samp.Overhead() <= s.overhead {
+				return samp, nil
+			}
+		}
+
+		// Couldn't get a sample with the current overhead. Increase it.
+		newOverhead := 2 * s.overhead
+		if newOverhead > maxOverheadCycles {
+			// We'll give it one more shot with the max overhead.
+
+			if s.overhead == maxOverheadCycles {
+				return sample{}, errOverheadTooHigh
+			}
+
+			newOverhead = maxOverheadCycles
+		}
+
+		s.overhead = newOverhead
+		log.Debugf("Time: Adjusting syscall overhead up to %v", s.overhead)
+	}
+}
+
+// Sample collects a reference clock sample.
+func (s *sampler) Sample() error {
+	sample, err := s.lowOverheadSample()
+	if err != nil {
+		return err
+	}
+
+	s.samples = append(s.samples, sample)
+	if len(s.samples) > maxSamples {
+		s.samples = s.samples[1:]
+	}
+
+	// If the 4 most recent samples all have an overhead less than half the
+	// expected overhead, adjust downwards.
+	if len(s.samples) < 4 {
+		return nil
+	}
+
+	for _, sample := range s.samples[len(s.samples)-4:] {
+		if sample.Overhead() > s.overhead/2 {
+			return nil
+		}
+	}
+
+	s.overhead -= s.overhead / 8
+	log.Debugf("Time: Adjusting syscall overhead down to %v", s.overhead)
+
+	return nil
+}
+
+// Syscall returns the current raw reference time without storing TSC
+// samples.
+func (s *sampler) Syscall() (ReferenceNS, error) {
+	sample, err := s.clocks.Sample(s.clockID)
+	if err != nil {
+		return 0, err
+	}
+
+	return sample.ref, nil
+}
+
+// Cycles returns a raw TSC value.
+func (s *sampler) Cycles() TSCValue {
+	return s.clocks.Cycles()
+}
+
+// Range returns the widest range of clock samples available.
+func (s *sampler) Range() (sample, sample, bool) {
+	if len(s.samples) < 2 {
+		return sample{}, sample{}, false
+	}
+
+	return s.samples[0], s.samples[len(s.samples)-1], true
+}
diff --git a/pkg/sentry/time/sampler_test.go b/pkg/sentry/time/sampler_test.go
new file mode 100644
index 000000000..caf7e5c53
--- /dev/null
+++ b/pkg/sentry/time/sampler_test.go
@@ -0,0 +1,183 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"errors"
+	"testing"
+)
+
+// errNoSamples is returned when testReferenceClocks runs out of samples.
+var errNoSamples = errors.New("no samples available")
+
+// testReferenceClocks returns a preset list of samples and cycle counts.
+type testReferenceClocks struct {
+	samples []sample
+	cycles  []TSCValue
+}
+
+// Sample implements referenceClocks.Sample, returning the next sample in the list.
+func (t *testReferenceClocks) Sample(_ ClockID) (sample, error) {
+	if len(t.samples) == 0 {
+		return sample{}, errNoSamples
+	}
+
+	s := t.samples[0]
+	if len(t.samples) == 1 {
+		t.samples = nil
+	} else {
+		t.samples = t.samples[1:]
+	}
+
+	return s, nil
+}
+
+// Cycles implements referenceClocks.Cycles, returning the next TSCValue in the list.
+func (t *testReferenceClocks) Cycles() TSCValue {
+	if len(t.cycles) == 0 {
+		return 0
+	}
+
+	c := t.cycles[0]
+	if len(t.cycles) == 1 {
+		t.cycles = nil
+	} else {
+		t.cycles = t.cycles[1:]
+	}
+
+	return c
+}
+
+// newTestSampler returns a sampler that collects samples from
+// the given sample list and cycle counts from the given cycle list.
+func newTestSampler(samples []sample, cycles []TSCValue) *sampler {
+	return &sampler{
+		clocks: &testReferenceClocks{
+			samples: samples,
+			cycles:  cycles,
+		},
+		overhead: defaultOverheadCycles,
+	}
+}
+
+// generateSamples generates n samples with the given overhead.
+func generateSamples(n int, overhead TSCValue) []sample {
+	samples := []sample{{before: 1000000, after: 1000000 + overhead, ref: 100}}
+	for i := 0; i < n-1; i++ {
+		prev := samples[len(samples)-1]
+		samples = append(samples, sample{
+			before: prev.before + 1000000,
+			after:  prev.after + 1000000,
+			ref:    prev.ref + 100,
+		})
+	}
+	return samples
+}
+
+// TestSample ensures that samples can be collected.
+func TestSample(t *testing.T) {
+	testCases := []struct {
+		name    string
+		samples []sample
+		err     error
+	}{
+		{
+			name: "basic",
+			samples: []sample{
+				{before: 100000, after: 100000 + defaultOverheadCycles, ref: 100},
+			},
+			err: nil,
+		},
+		{
+			// Sample with backwards TSC ignored.
+			// referenceClock should retry and get errNoSamples.
+			name: "backwards-tsc-ignored",
+			samples: []sample{
+				{before: 100000, after: 90000, ref: 100},
+			},
+			err: errNoSamples,
+		},
+		{
+			// Sample far above overhead skipped.
+			// referenceClock should retry and get errNoSamples.
+			name: "reject-overhead",
+			samples: []sample{
+				{before: 100000, after: 100000 + 5*defaultOverheadCycles, ref: 100},
+			},
+			err: errNoSamples,
+		},
+		{
+			// Maximum overhead allowed is bounded.
+			name: "over-max-overhead",
+			// Generate a bunch of samples. The reference clock
+			// needs a while to ramp up its expected overhead.
+			samples: generateSamples(100, 2*maxOverheadCycles),
+			err:     errOverheadTooHigh,
+		},
+		{
+			// Overhead at maximum overhead is allowed.
+			name: "max-overhead",
+			// Generate a bunch of samples. The reference clock
+			// needs a while to ramp up its expected overhead.
+			samples: generateSamples(100, maxOverheadCycles),
+			err:     nil,
+		},
+	}
+	for _, tc := range testCases {
+		t.Run(tc.name, func(t *testing.T) {
+			s := newTestSampler(tc.samples, nil)
+			err := s.Sample()
+			if err != tc.err {
+				t.Errorf("Sample err got %v want %v", err, tc.err)
+			}
+		})
+	}
+}
+
+// TestOutliersIgnored tests that referenceClock ignores samples with very high
+// overhead.
+func TestOutliersIgnored(t *testing.T) {
+	s := newTestSampler([]sample{
+		{before: 100000, after: 100000 + defaultOverheadCycles, ref: 100},
+		{before: 200000, after: 200000 + defaultOverheadCycles, ref: 200},
+		{before: 300000, after: 300000 + defaultOverheadCycles, ref: 300},
+		{before: 400000, after: 400000 + defaultOverheadCycles, ref: 400},
+		{before: 500000, after: 500000 + 5*defaultOverheadCycles, ref: 500}, // Ignored
+		{before: 600000, after: 600000 + defaultOverheadCycles, ref: 600},
+		{before: 700000, after: 700000 + defaultOverheadCycles, ref: 700},
+	}, nil)
+
+	// Collect 5 samples.
+	for i := 0; i < 5; i++ {
+		err := s.Sample()
+		if err != nil {
+			t.Fatalf("Unexpected error while sampling: %v", err)
+		}
+	}
+
+	oldest, newest, ok := s.Range()
+	if !ok {
+		t.Fatalf("Range not ok")
+	}
+
+	if oldest.ref != 100 {
+		t.Errorf("oldest.ref got %v want %v", oldest.ref, 100)
+	}
+
+	// We skipped the high-overhead sample.
+	if newest.ref != 600 {
+		t.Errorf("newest.ref got %v want %v", newest.ref, 600)
+	}
+}
diff --git a/pkg/sentry/time/sampler_unsafe.go b/pkg/sentry/time/sampler_unsafe.go
new file mode 100644
index 000000000..7ea19d387
--- /dev/null
+++ b/pkg/sentry/time/sampler_unsafe.go
@@ -0,0 +1,56 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package time
+
+import (
+	"syscall"
+	"unsafe"
+)
+
+// syscallTSCReferenceClocks is the standard referenceClocks, collecting
+// samples using CLOCK_GETTIME and RDTSC.
+type syscallTSCReferenceClocks struct {
+	tscCycleClock
+}
+
+// Sample implements sampler.Sample.
+func (syscallTSCReferenceClocks) Sample(c ClockID) (sample, error) {
+	var s sample
+
+	s.before = Rdtsc()
+
+	// Don't call clockGettime to avoid a call which may call morestack.
+	var ts syscall.Timespec
+	_, _, e := syscall.RawSyscall(syscall.SYS_CLOCK_GETTIME, uintptr(c), uintptr(unsafe.Pointer(&ts)), 0)
+	if e != 0 {
+		return sample{}, e
+	}
+
+	s.after = Rdtsc()
+	s.ref = ReferenceNS(ts.Nano())
+
+	return s, nil
+}
+
+// clockGettime calls SYS_CLOCK_GETTIME, returning time in nanoseconds.
+func clockGettime(c ClockID) (ReferenceNS, error) {
+	var ts syscall.Timespec
+	_, _, e := syscall.RawSyscall(syscall.SYS_CLOCK_GETTIME, uintptr(c), uintptr(unsafe.Pointer(&ts)), 0)
+	if e != 0 {
+		return 0, e
+	}
+
+	return ReferenceNS(ts.Nano()), nil
+}
diff --git a/pkg/sentry/time/tsc_amd64.s b/pkg/sentry/time/tsc_amd64.s
new file mode 100644
index 000000000..4cc604392
--- /dev/null
+++ b/pkg/sentry/time/tsc_amd64.s
@@ -0,0 +1,27 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "textflag.h"
+
+TEXT ·Rdtsc(SB),NOSPLIT,$0-8
+	// N.B. We need LFENCE on Intel, AMD is more complicated.
+	// Modern AMD CPUs with modern kernels make LFENCE behave like it does
+	// on Intel with MSR_F10H_DECFG_LFENCE_SERIALIZE_BIT. MFENCE is
+	// otherwise needed on AMD.
+	LFENCE
+	RDTSC
+	SHLQ	$32, DX
+	ADDQ	DX, AX
+	MOVQ	AX, ret+0(FP)
+	RET
diff --git a/pkg/sentry/uniqueid/BUILD b/pkg/sentry/uniqueid/BUILD
new file mode 100644
index 000000000..c8ab03c3d
--- /dev/null
+++ b/pkg/sentry/uniqueid/BUILD
@@ -0,0 +1,11 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "uniqueid",
+    srcs = ["context.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/uniqueid",
+    visibility = ["//pkg/sentry:internal"],
+    deps = ["//pkg/sentry/context"],
+)
diff --git a/pkg/sentry/uniqueid/context.go b/pkg/sentry/uniqueid/context.go
new file mode 100644
index 000000000..eeb8c4286
--- /dev/null
+++ b/pkg/sentry/uniqueid/context.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package uniqueid defines context.Context keys for obtaining system-wide
+// unique identifiers.
+package uniqueid
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// contextID is the kernel package's type for context.Context.Value keys.
+type contextID int
+
+const (
+	// CtxGlobalUniqueID is a Context.Value key for a system-wide
+	// unique identifier.
+	CtxGlobalUniqueID contextID = iota
+
+	// CtxInotifyCookie is a Context.Value key for a unique inotify
+	// event cookie.
+	CtxInotifyCookie
+)
+
+// GlobalFromContext returns a system-wide unique identifier from ctx.
+func GlobalFromContext(ctx context.Context) uint64 {
+	return ctx.Value(CtxGlobalUniqueID).(uint64)
+}
+
+// InotifyCookie generates a unique inotify event cookie from ctx.
+func InotifyCookie(ctx context.Context) uint32 {
+	return ctx.Value(CtxInotifyCookie).(uint32)
+}
diff --git a/pkg/sentry/usage/BUILD b/pkg/sentry/usage/BUILD
new file mode 100644
index 000000000..a0fe0aa07
--- /dev/null
+++ b/pkg/sentry/usage/BUILD
@@ -0,0 +1,38 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "usage_state",
+    srcs = [
+        "cpu.go",
+        "io.go",
+        "memory.go",
+    ],
+    out = "usage_state.go",
+    package = "usage",
+)
+
+go_library(
+    name = "usage",
+    srcs = [
+        "cpu.go",
+        "io.go",
+        "memory.go",
+        "memory_unsafe.go",
+        "usage.go",
+        "usage_state.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/usage",
+    visibility = [
+        "//pkg/sentry:internal",
+    ],
+    deps = [
+        "//pkg/bits",
+        "//pkg/log",
+        "//pkg/sentry/memutil",
+        "//pkg/state",
+        "@org_golang_x_sys//unix:go_default_library",
+    ],
+)
diff --git a/pkg/sentry/usage/cpu.go b/pkg/sentry/usage/cpu.go
new file mode 100644
index 000000000..1c2cc90e1
--- /dev/null
+++ b/pkg/sentry/usage/cpu.go
@@ -0,0 +1,44 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usage
+
+import (
+	"time"
+)
+
+// CPUStats contains the subset of struct rusage fields that relate to CPU
+// scheduling.
+type CPUStats struct {
+	// UserTime is the amount of time spent executing application code.
+	UserTime time.Duration
+
+	// SysTime is the amount of time spent executing sentry code.
+	SysTime time.Duration
+
+	// VoluntarySwitches is the number of times control has been voluntarily
+	// ceded due to blocking, etc.
+	VoluntarySwitches uint64
+
+	// InvoluntarySwitches (struct rusage::ru_nivcsw) is unsupported, since
+	// "preemptive" scheduling is managed by the Go runtime, which doesn't
+	// provide this information.
+}
+
+// Accumulate adds s2 to s.
+func (s *CPUStats) Accumulate(s2 CPUStats) {
+	s.UserTime += s2.UserTime
+	s.SysTime += s2.SysTime
+	s.VoluntarySwitches += s2.VoluntarySwitches
+}
diff --git a/pkg/sentry/usage/io.go b/pkg/sentry/usage/io.go
new file mode 100644
index 000000000..a05053c32
--- /dev/null
+++ b/pkg/sentry/usage/io.go
@@ -0,0 +1,88 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usage
+
+import (
+	"sync/atomic"
+)
+
+// IO contains I/O-related statistics.
+type IO struct {
+	// CharsRead is the number of bytes read by read syscalls.
+	CharsRead uint64
+
+	// CharsWritten is the number of bytes written by write syscalls.
+	CharsWritten uint64
+
+	// ReadSyscalls is the number of read syscalls.
+	ReadSyscalls uint64
+
+	// WriteSyscalls is the number of write syscalls.
+	WriteSyscalls uint64
+
+	// The following counter is only meaningful when Sentry has internal
+	// pagecache.
+
+	// BytesRead is the number of bytes actually read into pagecache.
+	BytesRead uint64
+
+	// BytesWritten is the number of bytes actually written from pagecache.
+	BytesWritten uint64
+
+	// BytesWriteCancelled is the number of bytes not written out due to
+	// truncation.
+	BytesWriteCancelled uint64
+}
+
+// AccountReadSyscall does the accounting for a read syscall.
+func (i *IO) AccountReadSyscall(bytes int64) {
+	atomic.AddUint64(&i.ReadSyscalls, 1)
+	if bytes > 0 {
+		atomic.AddUint64(&i.CharsRead, uint64(bytes))
+	}
+}
+
+// AccountWriteSyscall does the accounting for a write syscall.
+func (i *IO) AccountWriteSyscall(bytes int64) {
+	atomic.AddUint64(&i.WriteSyscalls, 1)
+	if bytes > 0 {
+		atomic.AddUint64(&i.CharsWritten, uint64(bytes))
+	}
+}
+
+// AccountReadIO does the accounting for a read IO into the file system.
+func (i *IO) AccountReadIO(bytes int64) {
+	if bytes > 0 {
+		atomic.AddUint64(&i.BytesRead, uint64(bytes))
+	}
+}
+
+// AccountWriteIO does the accounting for a write IO into the file system.
+func (i *IO) AccountWriteIO(bytes int64) {
+	if bytes > 0 {
+		atomic.AddUint64(&i.BytesWritten, uint64(bytes))
+	}
+}
+
+// Accumulate adds up io usages.
+func (i *IO) Accumulate(io *IO) {
+	atomic.AddUint64(&i.CharsRead, atomic.LoadUint64(&io.CharsRead))
+	atomic.AddUint64(&i.CharsWritten, atomic.LoadUint64(&io.CharsWritten))
+	atomic.AddUint64(&i.ReadSyscalls, atomic.LoadUint64(&io.ReadSyscalls))
+	atomic.AddUint64(&i.WriteSyscalls, atomic.LoadUint64(&io.WriteSyscalls))
+	atomic.AddUint64(&i.BytesRead, atomic.LoadUint64(&io.BytesRead))
+	atomic.AddUint64(&i.BytesWritten, atomic.LoadUint64(&io.BytesWritten))
+	atomic.AddUint64(&i.BytesWriteCancelled, atomic.LoadUint64(&io.BytesWriteCancelled))
+}
diff --git a/pkg/sentry/usage/memory.go b/pkg/sentry/usage/memory.go
new file mode 100644
index 000000000..5d1b3a595
--- /dev/null
+++ b/pkg/sentry/usage/memory.go
@@ -0,0 +1,282 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usage
+
+import (
+	"fmt"
+	"os"
+	"sync"
+	"sync/atomic"
+	"syscall"
+
+	"gvisor.googlesource.com/gvisor/pkg/bits"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/memutil"
+)
+
+// MemoryKind represents a type of memory used by the application.
+//
+// For efficiency reasons, it is assumed that the Memory implementation is
+// responsible for specific stats (documented below), and those may be reported
+// in aggregate independently. See the platform.Memory interface as well as the
+// control.Usage.Collect method for more information.
+type MemoryKind int
+
+const (
+	// System represents miscellaneous system memory. This may include
+	// memory that is in the process of being reclaimed, system caches,
+	// page tables, swap, etc.
+	//
+	// This memory kind is backed by platform memory.
+	System MemoryKind = iota
+
+	// Anonymous represents anonymous application memory.
+	//
+	// This memory kind is backed by platform memory.
+	Anonymous
+
+	// PageCache represents memory allocated to back sandbox-visible files that
+	// do not have a local fd. The contents of these files are buffered in
+	// memory to support application mmaps.
+	//
+	// This memory kind is backed by platform memory.
+	PageCache
+
+	// Tmpfs represents memory used by the sandbox-visible tmpfs.
+	//
+	// This memory kind is backed by platform memory.
+	Tmpfs
+
+	// Ramdiskfs represents memory used by the ramdiskfs.
+	//
+	// This memory kind is backed by platform memory.
+	Ramdiskfs
+
+	// Mapped represents memory related to files which have a local fd on the
+	// host, and thus can be directly mapped. Typically these are files backed
+	// by gofers with donated-fd support. Note that this value may not track the
+	// exact amount of memory used by mapping on the host, because we don't have
+	// any visibility into the host kernel memory management. In particular,
+	// once we map some part of a host file, the host kernel is free to
+	// abitrarily populate/decommit the pages, which it may do for various
+	// reasons (ex. host memory reclaim, NUMA balancing).
+	//
+	// This memory kind is backed by the host pagecache, via host mmaps.
+	Mapped
+)
+
+// MemoryStats tracks application memory usage in bytes. All fields correspond to the
+// memory category with the same name. This object is thread-safe if accessed
+// through the provided methods. The public fields may be safely accessed
+// directly on a copy of the object obtained from Memory.Copy().
+type MemoryStats struct {
+	System    uint64
+	Anonymous uint64
+	PageCache uint64
+	Tmpfs     uint64
+	// Lazily updated based on the value in RTMapped.
+	Mapped    uint64
+	Ramdiskfs uint64
+}
+
+// RTMemoryStats contains the memory usage values that need to be directly
+// exposed through a shared memory file for real-time access. These are
+// categories not backed by platform memory. For details about how this works,
+// see the memory accounting docs.
+//
+// N.B. Please keep the struct in sync with the API. Noteably, changes to this
+// struct requires a version bump and addition of compatibility logic in the
+// control server. As a special-case, adding fields without re-ordering existing
+// ones do not require a version bump because the mapped page we use is
+// initially zeroed. Any added field will be ignored by an older API and will be
+// zero if read by a newer API.
+type RTMemoryStats struct {
+	RTMapped uint64
+}
+
+// MemoryLocked is Memory with access methods.
+type MemoryLocked struct {
+	mu sync.RWMutex
+	// MemoryStats records the memory stats.
+	MemoryStats
+	// RTMemoryStats records the memory stats that need to be exposed through
+	// shared page.
+	*RTMemoryStats
+	// File is the backing file storing the memory stats.
+	File *os.File
+}
+
+func newMemoryLocked() MemoryLocked {
+	name := "memory-usage"
+	fd, err := memutil.CreateMemFD(name, 0)
+	if err != nil {
+		panic("error creating usage file: " + err.Error())
+	}
+	file := os.NewFile(uintptr(fd), name)
+	if err := file.Truncate(int64(RTMemoryStatsSize)); err != nil {
+		panic("error truncating usage file: " + err.Error())
+	}
+	// Note: We rely on the returned page being initially zeroed. This will
+	// always be the case for a newly mapped page from /dev/shm. If we obtain
+	// the shared memory through some other means in the future, we may have to
+	// explicitly zero the page.
+	mmap, err := syscall.Mmap(int(file.Fd()), 0, int(RTMemoryStatsSize), syscall.PROT_READ|syscall.PROT_WRITE, syscall.MAP_SHARED)
+	if err != nil {
+		panic("error mapping usage file: " + err.Error())
+	}
+
+	return MemoryLocked{
+		File:          file,
+		RTMemoryStats: RTMemoryStatsPointer(mmap),
+	}
+}
+
+// MemoryAccounting is the global memory stats.
+//
+// There is no need to save or restore the global memory accounting object,
+// because individual frame kinds are saved and charged only when they become
+// resident.
+var MemoryAccounting = newMemoryLocked()
+
+func (m *MemoryLocked) incLocked(val uint64, kind MemoryKind) {
+	switch kind {
+	case System:
+		atomic.AddUint64(&m.System, val)
+	case Anonymous:
+		atomic.AddUint64(&m.Anonymous, val)
+	case PageCache:
+		atomic.AddUint64(&m.PageCache, val)
+	case Mapped:
+		atomic.AddUint64(&m.RTMapped, val)
+	case Tmpfs:
+		atomic.AddUint64(&m.Tmpfs, val)
+	case Ramdiskfs:
+		atomic.AddUint64(&m.Ramdiskfs, val)
+	default:
+		panic(fmt.Sprintf("invalid memory kind: %v", kind))
+	}
+}
+
+// Inc adds an additional usage of 'val' bytes to memory category 'kind'.
+//
+// This method is thread-safe.
+func (m *MemoryLocked) Inc(val uint64, kind MemoryKind) {
+	m.mu.RLock()
+	m.incLocked(val, kind)
+	m.mu.RUnlock()
+}
+
+func (m *MemoryLocked) decLocked(val uint64, kind MemoryKind) {
+	switch kind {
+	case System:
+		atomic.AddUint64(&m.System, ^(val - 1))
+	case Anonymous:
+		atomic.AddUint64(&m.Anonymous, ^(val - 1))
+	case PageCache:
+		atomic.AddUint64(&m.PageCache, ^(val - 1))
+	case Mapped:
+		atomic.AddUint64(&m.RTMapped, ^(val - 1))
+	case Tmpfs:
+		atomic.AddUint64(&m.Tmpfs, ^(val - 1))
+	case Ramdiskfs:
+		atomic.AddUint64(&m.Ramdiskfs, ^(val - 1))
+	default:
+		panic(fmt.Sprintf("invalid memory kind: %v", kind))
+	}
+}
+
+// Dec remove a usage of 'val' bytes from memory category 'kind'.
+//
+// This method is thread-safe.
+func (m *MemoryLocked) Dec(val uint64, kind MemoryKind) {
+	m.mu.RLock()
+	m.decLocked(val, kind)
+	m.mu.RUnlock()
+}
+
+// Move moves a usage of 'val' bytes from 'from' to 'to'.
+//
+// This method is thread-safe.
+func (m *MemoryLocked) Move(val uint64, to MemoryKind, from MemoryKind) {
+	m.mu.RLock()
+	// Just call decLocked and incLocked directly. We held the RLock to
+	// protect against concurrent callers to Total().
+	m.decLocked(val, from)
+	m.incLocked(val, to)
+	m.mu.RUnlock()
+}
+
+// totalLocked returns a total usage.
+//
+// Precondition: must be called when locked.
+func (m *MemoryLocked) totalLocked() (total uint64) {
+	total += atomic.LoadUint64(&m.System)
+	total += atomic.LoadUint64(&m.Anonymous)
+	total += atomic.LoadUint64(&m.PageCache)
+	total += atomic.LoadUint64(&m.RTMapped)
+	total += atomic.LoadUint64(&m.Tmpfs)
+	total += atomic.LoadUint64(&m.Ramdiskfs)
+	return
+}
+
+// Total returns a total memory usage.
+//
+// This method is thread-safe.
+func (m *MemoryLocked) Total() uint64 {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	return m.totalLocked()
+}
+
+// Copy returns a copy of the structure with a total.
+//
+// This method is thread-safe.
+func (m *MemoryLocked) Copy() (MemoryStats, uint64) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+	ms := m.MemoryStats
+	ms.Mapped = m.RTMapped
+	return ms, m.totalLocked()
+}
+
+// MinimumTotalMemoryBytes is the minimum reported total system memory.
+var MinimumTotalMemoryBytes uint64 = 2 << 30 // 2 GB
+
+// TotalMemory returns the "total usable memory" available.
+//
+// This number doesn't really have a true value so it's based on the following
+// inputs and further bounded to be above some minimum guaranteed value (2GB),
+// additionally ensuring that total memory reported is always less than used.
+//
+// memSize should be the platform.Memory size reported by platform.Memory.TotalSize()
+// used is the total memory reported by MemoryLocked.Total()
+func TotalMemory(memSize, used uint64) uint64 {
+	if memSize < MinimumTotalMemoryBytes {
+		memSize = MinimumTotalMemoryBytes
+	}
+	if memSize < used {
+		memSize = used
+		// Bump totalSize to the next largest power of 2, if one exists, so
+		// that MemFree isn't 0.
+		if msb := bits.MostSignificantOne64(memSize); msb < 63 {
+			memSize = uint64(1) << (uint(msb) + 1)
+		}
+	}
+	return memSize
+}
+
+// IncrementalMappedAccounting controls whether host mapped memory is accounted
+// incrementally during map translation. This may be modified during early
+// initialization, and is read-only afterward.
+var IncrementalMappedAccounting = false
diff --git a/pkg/sentry/usage/memory_unsafe.go b/pkg/sentry/usage/memory_unsafe.go
new file mode 100644
index 000000000..f990a7750
--- /dev/null
+++ b/pkg/sentry/usage/memory_unsafe.go
@@ -0,0 +1,27 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usage
+
+import (
+	"unsafe"
+)
+
+// RTMemoryStatsSize is the size of the RTMemoryStats struct.
+var RTMemoryStatsSize = unsafe.Sizeof(RTMemoryStats{})
+
+// RTMemoryStatsPointer casts the address of the byte slice into a RTMemoryStats pointer.
+func RTMemoryStatsPointer(b []byte) *RTMemoryStats {
+	return (*RTMemoryStats)(unsafe.Pointer(&b[0]))
+}
diff --git a/pkg/sentry/usage/usage.go b/pkg/sentry/usage/usage.go
new file mode 100644
index 000000000..3b3118659
--- /dev/null
+++ b/pkg/sentry/usage/usage.go
@@ -0,0 +1,16 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package usage provides representations of resource usage.
+package usage
diff --git a/pkg/sentry/usermem/BUILD b/pkg/sentry/usermem/BUILD
new file mode 100644
index 000000000..36c0760dd
--- /dev/null
+++ b/pkg/sentry/usermem/BUILD
@@ -0,0 +1,70 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library", "go_test")
+load("//tools/go_generics:defs.bzl", "go_template_instance")
+load("//tools/go_stateify:defs.bzl", "go_stateify")
+
+go_stateify(
+    name = "usermem_state",
+    srcs = [
+        "access_type.go",
+        "addr.go",
+        "addr_range.go",
+        "addr_range_seq_unsafe.go",
+    ],
+    out = "usermem_state.go",
+    package = "usermem",
+)
+
+go_template_instance(
+    name = "addr_range",
+    out = "addr_range.go",
+    package = "usermem",
+    prefix = "Addr",
+    template = "//pkg/segment:generic_range",
+    types = {
+        "T": "Addr",
+    },
+)
+
+go_library(
+    name = "usermem",
+    srcs = [
+        "access_type.go",
+        "addr.go",
+        "addr_range.go",
+        "addr_range_seq_unsafe.go",
+        "bytes_io.go",
+        "bytes_io_unsafe.go",
+        "usermem.go",
+        "usermem_state.go",
+        "usermem_x86.go",
+    ],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/usermem",
+    visibility = ["//pkg/sentry:internal"],
+    deps = [
+        "//pkg/atomicbitops",
+        "//pkg/binary",
+        "//pkg/log",
+        "//pkg/sentry/context",
+        "//pkg/sentry/safemem",
+        "//pkg/state",
+        "//pkg/syserror",
+        "//pkg/tcpip/buffer",
+    ],
+)
+
+go_test(
+    name = "usermem_test",
+    size = "small",
+    srcs = [
+        "addr_range_seq_test.go",
+        "usermem_test.go",
+    ],
+    embed = [":usermem"],
+    deps = [
+        "//pkg/sentry/context",
+        "//pkg/sentry/safemem",
+        "//pkg/syserror",
+    ],
+)
diff --git a/pkg/sentry/usermem/README.md b/pkg/sentry/usermem/README.md
new file mode 100644
index 000000000..2ebd3bcc1
--- /dev/null
+++ b/pkg/sentry/usermem/README.md
@@ -0,0 +1,31 @@
+This package defines primitives for sentry access to application memory.
+
+Major types:
+
+- The `IO` interface represents a virtual address space and provides I/O methods
+  on that address space. `IO` is the lowest-level primitive. The primary
+  implementation of the `IO` interface is `mm.MemoryManager`.
+
+- `IOSequence` represents a collection of individually-contiguous address ranges
+  in a `IO` that is operated on sequentially, analogous to Linux's `struct
+  iov_iter`.
+
+Major usage patterns:
+
+- Access to a task's virtual memory, subject to the application's memory
+  protections and while running on that task's goroutine, from a context that is
+  at or above the level of the `kernel` package (e.g. most syscall
+  implementations in `syscalls/linux`); use the `kernel.Task.Copy*` wrappers
+  defined in `kernel/task_usermem.go`.
+
+- Access to a task's virtual memory, from a context that is at or above the
+  level of the `kernel` package, but where any of the above constraints does not
+  hold (e.g. `PTRACE_POKEDATA`, which ignores application memory protections);
+  obtain the task's `mm.MemoryManager` by calling `kernel.Task.MemoryManager`,
+  and call its `IO` methods directly.
+
+- Access to a task's virtual memory, from a context that is below the level of
+  the `kernel` package (e.g. filesystem I/O); clients must pass I/O arguments
+  from higher layers, usually in the form of an `IOSequence`. The
+  `kernel.Task.SingleIOSequence` and `kernel.Task.IovecsIOSequence` functions in
+  `kernel/task_usermem.go` are convenience functions for doing so.
diff --git a/pkg/sentry/usermem/access_type.go b/pkg/sentry/usermem/access_type.go
new file mode 100644
index 000000000..7eabecf30
--- /dev/null
+++ b/pkg/sentry/usermem/access_type.go
@@ -0,0 +1,117 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"syscall"
+)
+
+// AccessType specifies memory access types. This is used for
+// setting mapping permissions, as well as communicating faults.
+type AccessType struct {
+	// Read is read access.
+	Read bool
+
+	// Write is write access.
+	Write bool
+
+	// Execute is executable access.
+	Execute bool
+}
+
+// String returns a pretty representation of access. This looks like the
+// familiar r-x, rw-, etc. and can be relied on as such.
+func (a AccessType) String() string {
+	bits := [3]byte{'-', '-', '-'}
+	if a.Read {
+		bits[0] = 'r'
+	}
+	if a.Write {
+		bits[1] = 'w'
+	}
+	if a.Execute {
+		bits[2] = 'x'
+	}
+	return string(bits[:])
+}
+
+// Any returns true iff at least one of Read, Write or Execute is true.
+func (a AccessType) Any() bool {
+	return a.Read || a.Write || a.Execute
+}
+
+// Prot returns the system prot (syscall.PROT_READ, etc.) for this access.
+func (a AccessType) Prot() int {
+	var prot int
+	if a.Read {
+		prot |= syscall.PROT_READ
+	}
+	if a.Write {
+		prot |= syscall.PROT_WRITE
+	}
+	if a.Execute {
+		prot |= syscall.PROT_EXEC
+	}
+	return prot
+}
+
+// SupersetOf returns true iff the access types in a are a superset of the
+// access types in other.
+func (a AccessType) SupersetOf(other AccessType) bool {
+	if !a.Read && other.Read {
+		return false
+	}
+	if !a.Write && other.Write {
+		return false
+	}
+	if !a.Execute && other.Execute {
+		return false
+	}
+	return true
+}
+
+// Intersect returns the access types set in both a and other.
+func (a AccessType) Intersect(other AccessType) AccessType {
+	return AccessType{
+		Read:    a.Read && other.Read,
+		Write:   a.Write && other.Write,
+		Execute: a.Execute && other.Execute,
+	}
+}
+
+// Effective returns the set of effective access types allowed by a, even if
+// some types are not explicitly allowed.
+func (a AccessType) Effective() AccessType {
+	// In Linux, Write and Execute access generally imply Read access. See
+	// mm/mmap.c:protection_map.
+	//
+	// The notable exception is get_user_pages, which only checks against
+	// the original vma flags. That said, most user memory accesses do not
+	// use GUP.
+	if a.Write || a.Execute {
+		a.Read = true
+	}
+	return a
+}
+
+// Convenient access types.
+var (
+	NoAccess  = AccessType{}
+	Read      = AccessType{Read: true}
+	Write     = AccessType{Write: true}
+	Execute   = AccessType{Execute: true}
+	ReadWrite = AccessType{Read: true, Write: true}
+	AnyAccess = AccessType{Read: true, Write: true, Execute: true}
+)
diff --git a/pkg/sentry/usermem/addr.go b/pkg/sentry/usermem/addr.go
new file mode 100644
index 000000000..d175fdc74
--- /dev/null
+++ b/pkg/sentry/usermem/addr.go
@@ -0,0 +1,106 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"fmt"
+)
+
+// Addr represents a generic virtual address.
+type Addr uintptr
+
+// AddLength adds the given length to start and returns the result. ok is true
+// iff adding the length did not overflow the range of Addr.
+//
+// Note: This function is usually used to get the end of an address range
+// defined by its start address and length. Since the resulting end is
+// exclusive, end == 0 is technically valid, and corresponds to a range that
+// extends to the end of the address space, but ok will be false. This isn't
+// expected to ever come up in practice.
+func (v Addr) AddLength(length uint64) (end Addr, ok bool) {
+	end = v + Addr(length)
+	// The second half of the following check is needed in case uintptr is
+	// smaller than 64 bits.
+	ok = end >= v && length <= uint64(^Addr(0))
+	return
+}
+
+// RoundDown returns the address rounded down to the nearest page boundary.
+func (v Addr) RoundDown() Addr {
+	return v & ^Addr(PageSize-1)
+}
+
+// RoundUp returns the address rounded up to the nearest page boundary. ok is
+// true iff rounding up did not wrap around.
+func (v Addr) RoundUp() (addr Addr, ok bool) {
+	addr = Addr(v + PageSize - 1).RoundDown()
+	ok = addr >= v
+	return
+}
+
+// MustRoundUp is equivalent to RoundUp, but panics if rounding up wraps
+// around.
+func (v Addr) MustRoundUp() Addr {
+	addr, ok := v.RoundUp()
+	if !ok {
+		panic(fmt.Sprintf("usermem.Addr(%d).RoundUp() wraps", v))
+	}
+	return addr
+}
+
+// HugeRoundDown returns the address rounded down to the nearest huge page
+// boundary.
+func (v Addr) HugeRoundDown() Addr {
+	return v & ^Addr(HugePageSize-1)
+}
+
+// HugeRoundUp returns the address rounded up to the nearest huge page boundary.
+// ok is true iff rounding up did not wrap around.
+func (v Addr) HugeRoundUp() (addr Addr, ok bool) {
+	addr = Addr(v + HugePageSize - 1).HugeRoundDown()
+	ok = addr >= v
+	return
+}
+
+// PageOffset returns the offset of v into the current page.
+func (v Addr) PageOffset() uint64 {
+	return uint64(v & Addr(PageSize-1))
+}
+
+// IsPageAligned returns true if v.PageOffset() == 0.
+func (v Addr) IsPageAligned() bool {
+	return v.PageOffset() == 0
+}
+
+// AddrRange is a range of Addrs.
+//
+// type AddrRange <generated by go_generics>
+
+// ToRange returns [v, v+length).
+func (v Addr) ToRange(length uint64) (AddrRange, bool) {
+	end, ok := v.AddLength(length)
+	return AddrRange{v, end}, ok
+}
+
+// IsPageAligned returns true if ar.Start.IsPageAligned() and
+// ar.End.IsPageAligned().
+func (ar AddrRange) IsPageAligned() bool {
+	return ar.Start.IsPageAligned() && ar.End.IsPageAligned()
+}
+
+// String implements fmt.Stringer.String.
+func (ar AddrRange) String() string {
+	return fmt.Sprintf("[%#x, %#x)", ar.Start, ar.End)
+}
diff --git a/pkg/sentry/usermem/addr_range_seq_test.go b/pkg/sentry/usermem/addr_range_seq_test.go
new file mode 100644
index 000000000..cf9d785ed
--- /dev/null
+++ b/pkg/sentry/usermem/addr_range_seq_test.go
@@ -0,0 +1,197 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"testing"
+)
+
+var addrRangeSeqTests = []struct {
+	desc   string
+	ranges []AddrRange
+}{
+	{
+		desc: "Empty sequence",
+	},
+	{
+		desc: "Single empty AddrRange",
+		ranges: []AddrRange{
+			{0x10, 0x10},
+		},
+	},
+	{
+		desc: "Single non-empty AddrRange of length 1",
+		ranges: []AddrRange{
+			{0x10, 0x11},
+		},
+	},
+	{
+		desc: "Single non-empty AddrRange of length 2",
+		ranges: []AddrRange{
+			{0x10, 0x12},
+		},
+	},
+	{
+		desc: "Multiple non-empty AddrRanges",
+		ranges: []AddrRange{
+			{0x10, 0x11},
+			{0x20, 0x22},
+		},
+	},
+	{
+		desc: "Multiple AddrRanges including empty AddrRanges",
+		ranges: []AddrRange{
+			{0x10, 0x10},
+			{0x20, 0x20},
+			{0x30, 0x33},
+			{0x40, 0x44},
+			{0x50, 0x50},
+			{0x60, 0x60},
+			{0x70, 0x77},
+			{0x80, 0x88},
+			{0x90, 0x90},
+			{0xa0, 0xa0},
+		},
+	},
+}
+
+func testAddrRangeSeqEqualityWithTailIteration(t *testing.T, ars AddrRangeSeq, wantRanges []AddrRange) {
+	var wantLen int64
+	for _, ar := range wantRanges {
+		wantLen += int64(ar.Length())
+	}
+
+	var i int
+	for !ars.IsEmpty() {
+		if gotLen := ars.NumBytes(); gotLen != wantLen {
+			t.Errorf("Iteration %d: %v.NumBytes(): got %d, wanted %d", i, ars, gotLen, wantLen)
+		}
+		if gotN, wantN := ars.NumRanges(), len(wantRanges)-i; gotN != wantN {
+			t.Errorf("Iteration %d: %v.NumRanges(): got %d, wanted %d", i, ars, gotN, wantN)
+		}
+		got := ars.Head()
+		if i >= len(wantRanges) {
+			t.Errorf("Iteration %d: %v.Head(): got %s, wanted <end of sequence>", i, ars, got)
+		} else if want := wantRanges[i]; got != want {
+			t.Errorf("Iteration %d: %v.Head(): got %s, wanted %s", i, ars, got, want)
+		}
+		ars = ars.Tail()
+		wantLen -= int64(got.Length())
+		i++
+	}
+	if gotLen := ars.NumBytes(); gotLen != 0 || wantLen != 0 {
+		t.Errorf("Iteration %d: %v.NumBytes(): got %d, wanted %d (which should be 0)", i, ars, gotLen, wantLen)
+	}
+	if gotN := ars.NumRanges(); gotN != 0 {
+		t.Errorf("Iteration %d: %v.NumRanges(): got %d, wanted 0", i, ars, gotN)
+	}
+}
+
+func TestAddrRangeSeqTailIteration(t *testing.T) {
+	for _, test := range addrRangeSeqTests {
+		t.Run(test.desc, func(t *testing.T) {
+			testAddrRangeSeqEqualityWithTailIteration(t, AddrRangeSeqFromSlice(test.ranges), test.ranges)
+		})
+	}
+}
+
+func TestAddrRangeSeqDropFirstEmpty(t *testing.T) {
+	var ars AddrRangeSeq
+	if got, want := ars.DropFirst(1), ars; got != want {
+		t.Errorf("%v.DropFirst(1): got %v, wanted %v", ars, got, want)
+	}
+}
+
+func TestAddrRangeSeqDropSingleByteIteration(t *testing.T) {
+	// Tests AddrRangeSeq iteration using Head/DropFirst, simulating
+	// I/O-per-AddrRange.
+	for _, test := range addrRangeSeqTests {
+		t.Run(test.desc, func(t *testing.T) {
+			// Figure out what AddrRanges we expect to see.
+			var wantLen int64
+			var wantRanges []AddrRange
+			for _, ar := range test.ranges {
+				wantLen += int64(ar.Length())
+				wantRanges = append(wantRanges, ar)
+				if ar.Length() == 0 {
+					// We "do" 0 bytes of I/O and then call DropFirst(0),
+					// advancing to the next AddrRange.
+					continue
+				}
+				// Otherwise we "do" 1 byte of I/O and then call DropFirst(1),
+				// advancing the AddrRange by 1 byte, or to the next AddrRange
+				// if this one is exhausted.
+				for ar.Start++; ar.Length() != 0; ar.Start++ {
+					wantRanges = append(wantRanges, ar)
+				}
+			}
+			t.Logf("Expected AddrRanges: %s (%d bytes)", wantRanges, wantLen)
+
+			ars := AddrRangeSeqFromSlice(test.ranges)
+			var i int
+			for !ars.IsEmpty() {
+				if gotLen := ars.NumBytes(); gotLen != wantLen {
+					t.Errorf("Iteration %d: %v.NumBytes(): got %d, wanted %d", i, ars, gotLen, wantLen)
+				}
+				got := ars.Head()
+				if i >= len(wantRanges) {
+					t.Errorf("Iteration %d: %v.Head(): got %s, wanted <end of sequence>", i, ars, got)
+				} else if want := wantRanges[i]; got != want {
+					t.Errorf("Iteration %d: %v.Head(): got %s, wanted %s", i, ars, got, want)
+				}
+				if got.Length() == 0 {
+					ars = ars.DropFirst(0)
+				} else {
+					ars = ars.DropFirst(1)
+					wantLen--
+				}
+				i++
+			}
+			if gotLen := ars.NumBytes(); gotLen != 0 || wantLen != 0 {
+				t.Errorf("Iteration %d: %v.NumBytes(): got %d, wanted %d (which should be 0)", i, ars, gotLen, wantLen)
+			}
+		})
+	}
+}
+
+func TestAddrRangeSeqTakeFirstEmpty(t *testing.T) {
+	var ars AddrRangeSeq
+	if got, want := ars.TakeFirst(1), ars; got != want {
+		t.Errorf("%v.TakeFirst(1): got %v, wanted %v", ars, got, want)
+	}
+}
+
+func TestAddrRangeSeqTakeFirst(t *testing.T) {
+	ranges := []AddrRange{
+		{0x10, 0x11},
+		{0x20, 0x22},
+		{0x30, 0x30},
+		{0x40, 0x44},
+		{0x50, 0x55},
+		{0x60, 0x60},
+		{0x70, 0x77},
+	}
+	ars := AddrRangeSeqFromSlice(ranges).TakeFirst(5)
+	want := []AddrRange{
+		{0x10, 0x11}, // +1 byte (total 1 byte), not truncated
+		{0x20, 0x22}, // +2 bytes (total 3 bytes), not truncated
+		{0x30, 0x30}, // +0 bytes (total 3 bytes), no change
+		{0x40, 0x42}, // +2 bytes (total 5 bytes), partially truncated
+		{0x50, 0x50}, // +0 bytes (total 5 bytes), fully truncated
+		{0x60, 0x60}, // +0 bytes (total 5 bytes), "fully truncated" (no change)
+		{0x70, 0x70}, // +0 bytes (total 5 bytes), fully truncated
+	}
+	testAddrRangeSeqEqualityWithTailIteration(t, ars, want)
+}
diff --git a/pkg/sentry/usermem/addr_range_seq_unsafe.go b/pkg/sentry/usermem/addr_range_seq_unsafe.go
new file mode 100644
index 000000000..13b2998b3
--- /dev/null
+++ b/pkg/sentry/usermem/addr_range_seq_unsafe.go
@@ -0,0 +1,277 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"bytes"
+	"fmt"
+	"reflect"
+	"unsafe"
+)
+
+// An AddrRangeSeq represents a sequence of AddrRanges.
+//
+// AddrRangeSeqs are immutable and may be copied by value. The zero value of
+// AddrRangeSeq represents an empty sequence.
+//
+// An AddrRangeSeq may contain AddrRanges with a length of 0. This is necessary
+// since zero-length AddrRanges are significant to MM bounds checks.
+type AddrRangeSeq struct {
+	// If length is 0, then the AddrRangeSeq represents no AddrRanges.
+	// Invariants: data == 0; offset == 0; limit == 0.
+	//
+	// If length is 1, then the AddrRangeSeq represents the single
+	// AddrRange{offset, offset+limit}. Invariants: data == 0.
+	//
+	// Otherwise, length >= 2, and the AddrRangeSeq represents the `length`
+	// AddrRanges in the array of AddrRanges starting at address `data`,
+	// starting at `offset` bytes into the first AddrRange and limited to the
+	// following `limit` bytes. (AddrRanges after `limit` are still iterated,
+	// but are truncated to a length of 0.) Invariants: data != 0; offset <=
+	// data[0].Length(); limit > 0; offset+limit <= the combined length of all
+	// AddrRanges in the array.
+	data   unsafe.Pointer
+	length int
+	offset Addr
+	limit  Addr
+}
+
+// AddrRangeSeqOf returns an AddrRangeSeq representing the single AddrRange ar.
+func AddrRangeSeqOf(ar AddrRange) AddrRangeSeq {
+	return AddrRangeSeq{
+		length: 1,
+		offset: ar.Start,
+		limit:  ar.Length(),
+	}
+}
+
+// AddrRangeSeqFromSlice returns an AddrRangeSeq representing all AddrRanges in
+// slice.
+//
+// Whether the returned AddrRangeSeq shares memory with slice is unspecified;
+// clients should avoid mutating slices passed to AddrRangeSeqFromSlice.
+//
+// Preconditions: The combined length of all AddrRanges in slice <=
+// math.MaxInt64.
+func AddrRangeSeqFromSlice(slice []AddrRange) AddrRangeSeq {
+	var limit int64
+	for _, ar := range slice {
+		len64 := int64(ar.Length())
+		if len64 < 0 {
+			panic(fmt.Sprintf("Length of AddrRange %v overflows int64", ar))
+		}
+		sum := limit + len64
+		if sum < limit {
+			panic(fmt.Sprintf("Total length of AddrRanges %v overflows int64", slice))
+		}
+		limit = sum
+	}
+	return addrRangeSeqFromSliceLimited(slice, limit)
+}
+
+// Preconditions: The combined length of all AddrRanges in slice <= limit.
+// limit >= 0. If len(slice) != 0, then limit > 0.
+func addrRangeSeqFromSliceLimited(slice []AddrRange, limit int64) AddrRangeSeq {
+	switch len(slice) {
+	case 0:
+		return AddrRangeSeq{}
+	case 1:
+		return AddrRangeSeq{
+			length: 1,
+			offset: slice[0].Start,
+			limit:  Addr(limit),
+		}
+	default:
+		return AddrRangeSeq{
+			data:   unsafe.Pointer(&slice[0]),
+			length: len(slice),
+			limit:  Addr(limit),
+		}
+	}
+}
+
+// IsEmpty returns true if ars.NumRanges() == 0.
+//
+// Note that since AddrRangeSeq may contain AddrRanges with a length of zero,
+// an AddrRange representing 0 bytes (AddrRangeSeq.NumBytes() == 0) is not
+// necessarily empty.
+func (ars AddrRangeSeq) IsEmpty() bool {
+	return ars.length == 0
+}
+
+// NumRanges returns the number of AddrRanges in ars.
+func (ars AddrRangeSeq) NumRanges() int {
+	return ars.length
+}
+
+// NumBytes returns the number of bytes represented by ars.
+func (ars AddrRangeSeq) NumBytes() int64 {
+	return int64(ars.limit)
+}
+
+// Head returns the first AddrRange in ars.
+//
+// Preconditions: !ars.IsEmpty().
+func (ars AddrRangeSeq) Head() AddrRange {
+	if ars.length == 0 {
+		panic("empty AddrRangeSeq")
+	}
+	if ars.length == 1 {
+		return AddrRange{ars.offset, ars.offset + ars.limit}
+	}
+	ar := *(*AddrRange)(ars.data)
+	ar.Start += ars.offset
+	if ar.Length() > ars.limit {
+		ar.End = ar.Start + ars.limit
+	}
+	return ar
+}
+
+// Tail returns an AddrRangeSeq consisting of all AddrRanges in ars after the
+// first.
+//
+// Preconditions: !ars.IsEmpty().
+func (ars AddrRangeSeq) Tail() AddrRangeSeq {
+	if ars.length == 0 {
+		panic("empty AddrRangeSeq")
+	}
+	if ars.length == 1 {
+		return AddrRangeSeq{}
+	}
+	return ars.externalTail()
+}
+
+// Preconditions: ars.length >= 2.
+func (ars AddrRangeSeq) externalTail() AddrRangeSeq {
+	headLen := (*AddrRange)(ars.data).Length() - ars.offset
+	var tailLimit int64
+	if ars.limit > headLen {
+		tailLimit = int64(ars.limit - headLen)
+	}
+	var extSlice []AddrRange
+	extSliceHdr := (*reflect.SliceHeader)(unsafe.Pointer(&extSlice))
+	extSliceHdr.Data = uintptr(ars.data)
+	extSliceHdr.Len = ars.length
+	extSliceHdr.Cap = ars.length
+	return addrRangeSeqFromSliceLimited(extSlice[1:], tailLimit)
+}
+
+// DropFirst returns an AddrRangeSeq equivalent to ars, but with the first n
+// bytes omitted. If n > ars.NumBytes(), DropFirst returns an empty
+// AddrRangeSeq.
+//
+// If !ars.IsEmpty() and ars.Head().Length() == 0, DropFirst will always omit
+// at least ars.Head(), even if n == 0. This guarantees that the basic pattern
+// of:
+//
+//     for !ars.IsEmpty() {
+//       n, err = doIOWith(ars.Head())
+//       if err != nil {
+//         return err
+//       }
+//       ars = ars.DropFirst(n)
+//     }
+//
+// works even in the presence of zero-length AddrRanges.
+//
+// Preconditions: n >= 0.
+func (ars AddrRangeSeq) DropFirst(n int) AddrRangeSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return ars.DropFirst64(int64(n))
+}
+
+// DropFirst64 is equivalent to DropFirst but takes an int64.
+func (ars AddrRangeSeq) DropFirst64(n int64) AddrRangeSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	if Addr(n) > ars.limit {
+		return AddrRangeSeq{}
+	}
+	// Handle initial empty AddrRange.
+	switch ars.length {
+	case 0:
+		return AddrRangeSeq{}
+	case 1:
+		if ars.limit == 0 {
+			return AddrRangeSeq{}
+		}
+	default:
+		if rawHeadLen := (*AddrRange)(ars.data).Length(); ars.offset == rawHeadLen {
+			ars = ars.externalTail()
+		}
+	}
+	for n != 0 {
+		// Calling ars.Head() here is surprisingly expensive, so inline getting
+		// the head's length.
+		var headLen Addr
+		if ars.length == 1 {
+			headLen = ars.limit
+		} else {
+			headLen = (*AddrRange)(ars.data).Length() - ars.offset
+		}
+		if Addr(n) < headLen {
+			// Dropping ends partway through the head AddrRange.
+			ars.offset += Addr(n)
+			ars.limit -= Addr(n)
+			return ars
+		}
+		n -= int64(headLen)
+		ars = ars.Tail()
+	}
+	return ars
+}
+
+// TakeFirst returns an AddrRangeSeq equivalent to ars, but iterating at most n
+// bytes. TakeFirst never removes AddrRanges from ars; AddrRanges beyond the
+// first n bytes are reduced to a length of zero, but will still be iterated.
+//
+// Preconditions: n >= 0.
+func (ars AddrRangeSeq) TakeFirst(n int) AddrRangeSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	return ars.TakeFirst64(int64(n))
+}
+
+// TakeFirst64 is equivalent to TakeFirst but takes an int64.
+func (ars AddrRangeSeq) TakeFirst64(n int64) AddrRangeSeq {
+	if n < 0 {
+		panic(fmt.Sprintf("invalid n: %d", n))
+	}
+	if ars.limit > Addr(n) {
+		ars.limit = Addr(n)
+	}
+	return ars
+}
+
+// String implements fmt.Stringer.String.
+func (ars AddrRangeSeq) String() string {
+	// This is deliberately chosen to be the same as fmt's automatic stringer
+	// for []AddrRange.
+	var buf bytes.Buffer
+	buf.WriteByte('[')
+	var sep string
+	for !ars.IsEmpty() {
+		buf.WriteString(sep)
+		sep = " "
+		buf.WriteString(ars.Head().String())
+		ars = ars.Tail()
+	}
+	buf.WriteByte(']')
+	return buf.String()
+}
diff --git a/pkg/sentry/usermem/bytes_io.go b/pkg/sentry/usermem/bytes_io.go
new file mode 100644
index 000000000..01a746404
--- /dev/null
+++ b/pkg/sentry/usermem/bytes_io.go
@@ -0,0 +1,126 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+const maxInt = int(^uint(0) >> 1)
+
+// BytesIO implements IO using a byte slice. Addresses are interpreted as
+// offsets into the slice. Reads and writes beyond the end of the slice return
+// EFAULT.
+type BytesIO struct {
+	Bytes []byte
+}
+
+// CopyOut implements IO.CopyOut.
+func (b *BytesIO) CopyOut(ctx context.Context, addr Addr, src []byte, opts IOOpts) (int, error) {
+	rngN, rngErr := b.rangeCheck(addr, len(src))
+	if rngN == 0 {
+		return 0, rngErr
+	}
+	return copy(b.Bytes[int(addr):], src[:rngN]), rngErr
+}
+
+// CopyIn implements IO.CopyIn.
+func (b *BytesIO) CopyIn(ctx context.Context, addr Addr, dst []byte, opts IOOpts) (int, error) {
+	rngN, rngErr := b.rangeCheck(addr, len(dst))
+	if rngN == 0 {
+		return 0, rngErr
+	}
+	return copy(dst[:rngN], b.Bytes[int(addr):]), rngErr
+}
+
+// ZeroOut implements IO.ZeroOut.
+func (b *BytesIO) ZeroOut(ctx context.Context, addr Addr, toZero int64, opts IOOpts) (int64, error) {
+	if toZero > int64(maxInt) {
+		return 0, syserror.EINVAL
+	}
+	rngN, rngErr := b.rangeCheck(addr, int(toZero))
+	if rngN == 0 {
+		return 0, rngErr
+	}
+	zeroSlice := b.Bytes[int(addr) : int(addr)+rngN]
+	for i := range zeroSlice {
+		zeroSlice[i] = 0
+	}
+	return int64(rngN), rngErr
+}
+
+// CopyOutFrom implements IO.CopyOutFrom.
+func (b *BytesIO) CopyOutFrom(ctx context.Context, ars AddrRangeSeq, src safemem.Reader, opts IOOpts) (int64, error) {
+	dsts, rngErr := b.blocksFromAddrRanges(ars)
+	n, err := src.ReadToBlocks(dsts)
+	if err != nil {
+		return int64(n), err
+	}
+	return int64(n), rngErr
+}
+
+// CopyInTo implements IO.CopyInTo.
+func (b *BytesIO) CopyInTo(ctx context.Context, ars AddrRangeSeq, dst safemem.Writer, opts IOOpts) (int64, error) {
+	srcs, rngErr := b.blocksFromAddrRanges(ars)
+	n, err := dst.WriteFromBlocks(srcs)
+	if err != nil {
+		return int64(n), err
+	}
+	return int64(n), rngErr
+}
+
+func (b *BytesIO) rangeCheck(addr Addr, length int) (int, error) {
+	if length == 0 {
+		return 0, nil
+	}
+	if length < 0 {
+		return 0, syserror.EINVAL
+	}
+	max := Addr(len(b.Bytes))
+	if addr >= max {
+		return 0, syserror.EFAULT
+	}
+	end, ok := addr.AddLength(uint64(length))
+	if !ok || end > max {
+		return int(max - addr), syserror.EFAULT
+	}
+	return length, nil
+}
+
+func (b *BytesIO) blocksFromAddrRanges(ars AddrRangeSeq) (safemem.BlockSeq, error) {
+	blocks := make([]safemem.Block, 0, ars.NumRanges())
+	for !ars.IsEmpty() {
+		ar := ars.Head()
+		n, err := b.rangeCheck(ar.Start, int(ar.Length()))
+		if n != 0 {
+			blocks = append(blocks, safemem.BlockFromSafeSlice(b.Bytes[int(ar.Start):int(ar.Start)+n]))
+		}
+		if err != nil {
+			return safemem.BlockSeqFromSlice(blocks), err
+		}
+		ars = ars.Tail()
+	}
+	return safemem.BlockSeqFromSlice(blocks), nil
+}
+
+// BytesIOSequence returns an IOSequence representing the given byte slice.
+func BytesIOSequence(buf []byte) IOSequence {
+	return IOSequence{
+		IO:    &BytesIO{buf},
+		Addrs: AddrRangeSeqOf(AddrRange{0, Addr(len(buf))}),
+	}
+}
diff --git a/pkg/sentry/usermem/bytes_io_unsafe.go b/pkg/sentry/usermem/bytes_io_unsafe.go
new file mode 100644
index 000000000..efd71fcbc
--- /dev/null
+++ b/pkg/sentry/usermem/bytes_io_unsafe.go
@@ -0,0 +1,39 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"sync/atomic"
+	"unsafe"
+
+	"gvisor.googlesource.com/gvisor/pkg/atomicbitops"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+)
+
+// SwapUint32 implements IO.SwapUint32.
+func (b *BytesIO) SwapUint32(ctx context.Context, addr Addr, new uint32, opts IOOpts) (uint32, error) {
+	if _, rngErr := b.rangeCheck(addr, 4); rngErr != nil {
+		return 0, rngErr
+	}
+	return atomic.SwapUint32((*uint32)(unsafe.Pointer(&b.Bytes[int(addr)])), new), nil
+}
+
+// CompareAndSwapUint32 implements IO.CompareAndSwapUint32.
+func (b *BytesIO) CompareAndSwapUint32(ctx context.Context, addr Addr, old, new uint32, opts IOOpts) (uint32, error) {
+	if _, rngErr := b.rangeCheck(addr, 4); rngErr != nil {
+		return 0, rngErr
+	}
+	return atomicbitops.CompareAndSwapUint32((*uint32)(unsafe.Pointer(&b.Bytes[int(addr)])), old, new), nil
+}
diff --git a/pkg/sentry/usermem/usermem.go b/pkg/sentry/usermem/usermem.go
new file mode 100644
index 000000000..5d8a1c558
--- /dev/null
+++ b/pkg/sentry/usermem/usermem.go
@@ -0,0 +1,572 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package usermem governs access to user memory.
+package usermem
+
+import (
+	"errors"
+	"io"
+	"strconv"
+
+	"gvisor.googlesource.com/gvisor/pkg/binary"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// IO provides access to the contents of a virtual memory space.
+//
+// FIXME: Implementations of IO cannot expect ctx to contain any
+// meaningful data.
+type IO interface {
+	// CopyOut copies len(src) bytes from src to the memory mapped at addr. It
+	// returns the number of bytes copied. If the number of bytes copied is <
+	// len(src), it returns a non-nil error explaining why.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order.
+	CopyOut(ctx context.Context, addr Addr, src []byte, opts IOOpts) (int, error)
+
+	// CopyIn copies len(dst) bytes from the memory mapped at addr to dst.
+	// It returns the number of bytes copied. If the number of bytes copied is
+	// < len(dst), it returns a non-nil error explaining why.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order.
+	CopyIn(ctx context.Context, addr Addr, dst []byte, opts IOOpts) (int, error)
+
+	// ZeroOut sets toZero bytes to 0, starting at addr. It returns the number
+	// of bytes zeroed. If the number of bytes zeroed is < toZero, it returns a
+	// non-nil error explaining why.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order. toZero >= 0.
+	ZeroOut(ctx context.Context, addr Addr, toZero int64, opts IOOpts) (int64, error)
+
+	// CopyOutFrom copies ars.NumBytes() bytes from src to the memory mapped at
+	// ars. It returns the number of bytes copied, which may be less than the
+	// number of bytes read from src if copying fails. CopyOutFrom may return a
+	// partial copy without an error iff src.ReadToBlocks returns a partial
+	// read without an error.
+	//
+	// CopyOutFrom calls src.ReadToBlocks at most once.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order. src.ReadToBlocks must not block
+	// on mm.MemoryManager.activeMu or any preceding locks in the lock order.
+	CopyOutFrom(ctx context.Context, ars AddrRangeSeq, src safemem.Reader, opts IOOpts) (int64, error)
+
+	// CopyInTo copies ars.NumBytes() bytes from the memory mapped at ars to
+	// dst. It returns the number of bytes copied. CopyInTo may return a
+	// partial copy without an error iff dst.WriteFromBlocks returns a partial
+	// write without an error.
+	//
+	// CopyInTo calls dst.WriteFromBlocks at most once.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order. dst.WriteFromBlocks must not
+	// block on mm.MemoryManager.activeMu or any preceding locks in the lock
+	// order.
+	CopyInTo(ctx context.Context, ars AddrRangeSeq, dst safemem.Writer, opts IOOpts) (int64, error)
+
+	// TODO: The requirement that CopyOutFrom/CopyInTo call src/dst
+	// at most once, which is unnecessary in most cases, forces implementations
+	// to gather safemem.Blocks into a single slice to pass to src/dst. Add
+	// CopyOutFromIter/CopyInToIter, which relaxes this restriction, to avoid
+	// this allocation.
+
+	// SwapUint32 atomically sets the uint32 value at addr to new and
+	// returns the previous value.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order. addr must be aligned to a 4-byte
+	// boundary.
+	SwapUint32(ctx context.Context, addr Addr, new uint32, opts IOOpts) (uint32, error)
+
+	// CompareAndSwapUint32 atomically compares the uint32 value at addr to
+	// old; if they are equal, the value in memory is replaced by new. In
+	// either case, the previous value stored in memory is returned.
+	//
+	// Preconditions: The caller must not hold mm.MemoryManager.mappingMu or
+	// any following locks in the lock order. addr must be aligned to a 4-byte
+	// boundary.
+	CompareAndSwapUint32(ctx context.Context, addr Addr, old, new uint32, opts IOOpts) (uint32, error)
+}
+
+// IOOpts contains options applicable to all IO methods.
+type IOOpts struct {
+	// If IgnorePermissions is true, application-defined memory protections set
+	// by mmap(2) or mprotect(2) will be ignored. (Memory protections required
+	// by the target of the mapping are never ignored.)
+	IgnorePermissions bool
+
+	// If AddressSpaceActive is true, the IO implementation may assume that it
+	// has an active AddressSpace and can therefore use AddressSpace copying
+	// without performing activation. See mm/io.go for details.
+	AddressSpaceActive bool
+}
+
+// IOReadWriter is an io.ReadWriter that reads from / writes to addresses
+// starting at addr in IO. The preconditions that apply to IO.CopyIn and
+// IO.CopyOut also apply to IOReadWriter.Read and IOReadWriter.Write
+// respectively.
+type IOReadWriter struct {
+	Ctx  context.Context
+	IO   IO
+	Addr Addr
+	Opts IOOpts
+}
+
+// Read implements io.Reader.Read.
+//
+// Note that an address space does not have an "end of file", so Read can only
+// return io.EOF if IO.CopyIn returns io.EOF. Attempts to read unmapped or
+// unreadable memory, or beyond the end of the address space, should return
+// EFAULT.
+func (rw *IOReadWriter) Read(dst []byte) (int, error) {
+	n, err := rw.IO.CopyIn(rw.Ctx, rw.Addr, dst, rw.Opts)
+	end, ok := rw.Addr.AddLength(uint64(n))
+	if ok {
+		rw.Addr = end
+	} else {
+		// Disallow wraparound.
+		rw.Addr = ^Addr(0)
+		if err != nil {
+			err = syserror.EFAULT
+		}
+	}
+	return n, err
+}
+
+// Writer implements io.Writer.Write.
+func (rw *IOReadWriter) Write(src []byte) (int, error) {
+	n, err := rw.IO.CopyOut(rw.Ctx, rw.Addr, src, rw.Opts)
+	end, ok := rw.Addr.AddLength(uint64(n))
+	if ok {
+		rw.Addr = end
+	} else {
+		// Disallow wraparound.
+		rw.Addr = ^Addr(0)
+		if err != nil {
+			err = syserror.EFAULT
+		}
+	}
+	return n, err
+}
+
+// CopyObjectOut copies a fixed-size value or slice of fixed-size values from
+// src to the memory mapped at addr in uio. It returns the number of bytes
+// copied.
+//
+// CopyObjectOut must use reflection to encode src; performance-sensitive
+// clients should do encoding manually and use uio.CopyOut directly.
+//
+// Preconditions: As for IO.CopyOut.
+func CopyObjectOut(ctx context.Context, uio IO, addr Addr, src interface{}, opts IOOpts) (int, error) {
+	w := &IOReadWriter{
+		Ctx:  ctx,
+		IO:   uio,
+		Addr: addr,
+		Opts: opts,
+	}
+	return w.Write(binary.Marshal(nil, ByteOrder, src))
+}
+
+// CopyObjectIn copies a fixed-size value or slice of fixed-size values from
+// the memory mapped at addr in uio to dst. It returns the number of bytes
+// copied.
+//
+// CopyObjectIn must use reflection to decode dst; performance-sensitive
+// clients should use uio.CopyIn directly and do decoding manually.
+//
+// Preconditions: As for IO.CopyIn.
+func CopyObjectIn(ctx context.Context, uio IO, addr Addr, dst interface{}, opts IOOpts) (int, error) {
+	r := &IOReadWriter{
+		Ctx:  ctx,
+		IO:   uio,
+		Addr: addr,
+		Opts: opts,
+	}
+	buf := make([]byte, binary.Size(dst))
+	if _, err := io.ReadFull(r, buf); err != nil {
+		return 0, err
+	}
+	binary.Unmarshal(buf, ByteOrder, dst)
+	return int(r.Addr - addr), nil
+}
+
+// copyStringIncrement is the maximum number of bytes that are copied from
+// virtual memory at a time by CopyStringIn.
+const copyStringIncrement = 64
+
+// CopyStringIn copies a NUL-terminated string of unknown length from the
+// memory mapped at addr in uio and returns it as a string (not including the
+// trailing NUL). If the length of the string, including the terminating NUL,
+// would exceed maxlen, CopyStringIn returns the string truncated to maxlen and
+// ENAMETOOLONG.
+//
+// Preconditions: As for IO.CopyFromUser. maxlen >= 0.
+func CopyStringIn(ctx context.Context, uio IO, addr Addr, maxlen int, opts IOOpts) (string, error) {
+	buf := make([]byte, maxlen)
+	var done int
+	for done < maxlen {
+		start, ok := addr.AddLength(uint64(done))
+		if !ok {
+			// Last page of kernel memory. The application can't use this
+			// anyway.
+			return string(buf[:done]), syserror.EFAULT
+		}
+		// Read up to copyStringIncrement bytes at a time.
+		readlen := copyStringIncrement
+		if readlen > maxlen-done {
+			readlen = maxlen - done
+		}
+		end, ok := start.AddLength(uint64(readlen))
+		if !ok {
+			return string(buf[:done]), syserror.EFAULT
+		}
+		// Shorten the read to avoid crossing page boundaries, since faulting
+		// in a page unnecessarily is expensive. This also ensures that partial
+		// copies up to the end of application-mappable memory succeed.
+		if start.RoundDown() != end.RoundDown() {
+			end = end.RoundDown()
+		}
+		n, err := uio.CopyIn(ctx, start, buf[done:done+int(end-start)], opts)
+		// Look for the terminating zero byte, which may have occurred before
+		// hitting err.
+		for i, c := range buf[done : done+n] {
+			if c == 0 {
+				return string(buf[:done+i]), nil
+			}
+		}
+		done += n
+		if err != nil {
+			return string(buf[:done]), err
+		}
+	}
+	return string(buf), syserror.ENAMETOOLONG
+}
+
+// CopyOutVec copies bytes from src to the memory mapped at ars in uio. The
+// maximum number of bytes copied is ars.NumBytes() or len(src), whichever is
+// less. CopyOutVec returns the number of bytes copied; if this is less than
+// the maximum, it returns a non-nil error explaining why.
+//
+// Preconditions: As for IO.CopyOut.
+func CopyOutVec(ctx context.Context, uio IO, ars AddrRangeSeq, src []byte, opts IOOpts) (int, error) {
+	var done int
+	for !ars.IsEmpty() && done < len(src) {
+		ar := ars.Head()
+		cplen := len(src) - done
+		if Addr(cplen) >= ar.Length() {
+			cplen = int(ar.Length())
+		}
+		n, err := uio.CopyOut(ctx, ar.Start, src[done:done+cplen], opts)
+		done += n
+		if err != nil {
+			return done, err
+		}
+		ars = ars.DropFirst(n)
+	}
+	return done, nil
+}
+
+// CopyInVec copies bytes from the memory mapped at ars in uio to dst. The
+// maximum number of bytes copied is ars.NumBytes() or len(dst), whichever is
+// less. CopyInVec returns the number of bytes copied; if this is less than the
+// maximum, it returns a non-nil error explaining why.
+//
+// Preconditions: As for IO.CopyIn.
+func CopyInVec(ctx context.Context, uio IO, ars AddrRangeSeq, dst []byte, opts IOOpts) (int, error) {
+	var done int
+	for !ars.IsEmpty() && done < len(dst) {
+		ar := ars.Head()
+		cplen := len(dst) - done
+		if Addr(cplen) >= ar.Length() {
+			cplen = int(ar.Length())
+		}
+		n, err := uio.CopyIn(ctx, ar.Start, dst[done:done+cplen], opts)
+		done += n
+		if err != nil {
+			return done, err
+		}
+		ars = ars.DropFirst(n)
+	}
+	return done, nil
+}
+
+// ZeroOutVec writes zeroes to the memory mapped at ars in uio. The maximum
+// number of bytes written is ars.NumBytes() or toZero, whichever is less.
+// ZeroOutVec returns the number of bytes written; if this is less than the
+// maximum, it returns a non-nil error explaining why.
+//
+// Preconditions: As for IO.ZeroOut.
+func ZeroOutVec(ctx context.Context, uio IO, ars AddrRangeSeq, toZero int64, opts IOOpts) (int64, error) {
+	var done int64
+	for !ars.IsEmpty() && done < toZero {
+		ar := ars.Head()
+		cplen := toZero - done
+		if Addr(cplen) >= ar.Length() {
+			cplen = int64(ar.Length())
+		}
+		n, err := uio.ZeroOut(ctx, ar.Start, cplen, opts)
+		done += n
+		if err != nil {
+			return done, err
+		}
+		ars = ars.DropFirst64(n)
+	}
+	return done, nil
+}
+
+func isASCIIWhitespace(b byte) bool {
+	// Compare Linux include/linux/ctype.h, lib/ctype.c.
+	//  9 => horizontal tab '\t'
+	// 10 => line feed '\n'
+	// 11 => vertical tab '\v'
+	// 12 => form feed '\c'
+	// 13 => carriage return '\r'
+	return b == ' ' || (b >= 9 && b <= 13)
+}
+
+// CopyInt32StringsInVec copies up to len(dsts) whitespace-separated decimal
+// strings from the memory mapped at ars in uio and converts them to int32
+// values in dsts. It returns the number of bytes read.
+//
+// CopyInt32StringsInVec shares the following properties with Linux's
+// kernel/sysctl.c:proc_dointvec(write=1):
+//
+// - If any read value overflows the range of int32, or any invalid characters
+// are encountered during the read, CopyInt32StringsInVec returns EINVAL.
+//
+// - If, upon reaching the end of ars, fewer than len(dsts) values have been
+// read, CopyInt32StringsInVec returns no error if at least 1 value was read
+// and EINVAL otherwise.
+//
+// - Trailing whitespace after the last successfully read value is counted in
+// the number of bytes read.
+//
+// Unlike proc_dointvec():
+//
+// - CopyInt32StringsInVec does not implicitly limit ars.NumBytes() to
+// PageSize-1; callers that require this must do so explicitly.
+//
+// - CopyInt32StringsInVec returns EINVAL if ars.NumBytes() == 0.
+//
+// Preconditions: As for CopyInVec.
+func CopyInt32StringsInVec(ctx context.Context, uio IO, ars AddrRangeSeq, dsts []int32, opts IOOpts) (int64, error) {
+	if len(dsts) == 0 {
+		return 0, nil
+	}
+
+	buf := make([]byte, ars.NumBytes())
+	n, cperr := CopyInVec(ctx, uio, ars, buf, opts)
+	buf = buf[:n]
+
+	var i, j int
+	for ; j < len(dsts); j++ {
+		// Skip leading whitespace.
+		for i < len(buf) && isASCIIWhitespace(buf[i]) {
+			i++
+		}
+		if i == len(buf) {
+			break
+		}
+
+		// Find the end of the value to be parsed (next whitespace or end of string).
+		nextI := i + 1
+		for nextI < len(buf) && !isASCIIWhitespace(buf[nextI]) {
+			nextI++
+		}
+
+		// Parse a single value.
+		val, err := strconv.ParseInt(string(buf[i:nextI]), 10, 32)
+		if err != nil {
+			return int64(i), syserror.EINVAL
+		}
+		dsts[j] = int32(val)
+
+		i = nextI
+	}
+
+	// Skip trailing whitespace.
+	for i < len(buf) && isASCIIWhitespace(buf[i]) {
+		i++
+	}
+
+	if cperr != nil {
+		return int64(i), cperr
+	}
+	if j == 0 {
+		return int64(i), syserror.EINVAL
+	}
+	return int64(i), nil
+}
+
+// CopyInt32StringInVec is equivalent to CopyInt32StringsInVec, but copies at
+// most one int32.
+func CopyInt32StringInVec(ctx context.Context, uio IO, ars AddrRangeSeq, dst *int32, opts IOOpts) (int64, error) {
+	dsts := [1]int32{*dst}
+	n, err := CopyInt32StringsInVec(ctx, uio, ars, dsts[:], opts)
+	*dst = dsts[0]
+	return n, err
+}
+
+// IOSequence holds arguments to IO methods.
+type IOSequence struct {
+	IO    IO
+	Addrs AddrRangeSeq
+	Opts  IOOpts
+}
+
+// NumBytes returns s.Addrs.NumBytes().
+//
+// Note that NumBytes() may return 0 even if !s.Addrs.IsEmpty(), since
+// s.Addrs may contain a non-zero number of zero-length AddrRanges.
+// Many clients of
+// IOSequence currently do something like:
+//
+//     if ioseq.NumBytes() == 0 {
+//       return 0, nil
+//     }
+//     if f.availableBytes == 0 {
+//       return 0, syserror.ErrWouldBlock
+//     }
+//     return ioseq.CopyOutFrom(..., reader)
+//
+// In such cases, using s.Addrs.IsEmpty() will cause them to have the wrong
+// behavior for zero-length I/O. However, using s.NumBytes() == 0 instead means
+// that we will return success for zero-length I/O in cases where Linux would
+// return EFAULT due to a failed access_ok() check, so in the long term we
+// should move checks for ErrWouldBlock etc. into the body of
+// reader.ReadToBlocks and use s.Addrs.IsEmpty() instead.
+func (s IOSequence) NumBytes() int64 {
+	return s.Addrs.NumBytes()
+}
+
+// DropFirst returns a copy of s with s.Addrs.DropFirst(n).
+//
+// Preconditions: As for AddrRangeSeq.DropFirst.
+func (s IOSequence) DropFirst(n int) IOSequence {
+	return IOSequence{s.IO, s.Addrs.DropFirst(n), s.Opts}
+}
+
+// DropFirst64 returns a copy of s with s.Addrs.DropFirst64(n).
+//
+// Preconditions: As for AddrRangeSeq.DropFirst64.
+func (s IOSequence) DropFirst64(n int64) IOSequence {
+	return IOSequence{s.IO, s.Addrs.DropFirst64(n), s.Opts}
+}
+
+// TakeFirst returns a copy of s with s.Addrs.TakeFirst(n).
+//
+// Preconditions: As for AddrRangeSeq.TakeFirst.
+func (s IOSequence) TakeFirst(n int) IOSequence {
+	return IOSequence{s.IO, s.Addrs.TakeFirst(n), s.Opts}
+}
+
+// TakeFirst64 returns a copy of s with s.Addrs.TakeFirst64(n).
+//
+// Preconditions: As for AddrRangeSeq.TakeFirst64.
+func (s IOSequence) TakeFirst64(n int64) IOSequence {
+	return IOSequence{s.IO, s.Addrs.TakeFirst64(n), s.Opts}
+}
+
+// CopyOut invokes CopyOutVec over s.Addrs.
+//
+// As with CopyOutVec, if s.NumBytes() < len(src), the copy will be truncated
+// to s.NumBytes(), and a nil error will be returned.
+//
+// Preconditions: As for CopyOutVec.
+func (s IOSequence) CopyOut(ctx context.Context, src []byte) (int, error) {
+	return CopyOutVec(ctx, s.IO, s.Addrs, src, s.Opts)
+}
+
+// CopyIn invokes CopyInVec over s.Addrs.
+//
+// As with CopyInVec, if s.NumBytes() < len(dst), the copy will be truncated to
+// s.NumBytes(), and a nil error will be returned.
+//
+// Preconditions: As for CopyInVec.
+func (s IOSequence) CopyIn(ctx context.Context, dst []byte) (int, error) {
+	return CopyInVec(ctx, s.IO, s.Addrs, dst, s.Opts)
+}
+
+// ZeroOut invokes ZeroOutVec over s.Addrs.
+//
+// As with ZeroOutVec, if s.NumBytes() < toZero, the write will be truncated
+// to s.NumBytes(), and a nil error will be returned.
+//
+// Preconditions: As for ZeroOutVec.
+func (s IOSequence) ZeroOut(ctx context.Context, toZero int64) (int64, error) {
+	return ZeroOutVec(ctx, s.IO, s.Addrs, toZero, s.Opts)
+}
+
+// CopyOutFrom invokes s.CopyOutFrom over s.Addrs.
+//
+// Preconditions: As for IO.CopyOutFrom.
+func (s IOSequence) CopyOutFrom(ctx context.Context, src safemem.Reader) (int64, error) {
+	return s.IO.CopyOutFrom(ctx, s.Addrs, src, s.Opts)
+}
+
+// CopyInTo invokes s.CopyInTo over s.Addrs.
+//
+// Preconditions: As for IO.CopyInTo.
+func (s IOSequence) CopyInTo(ctx context.Context, dst safemem.Writer) (int64, error) {
+	return s.IO.CopyInTo(ctx, s.Addrs, dst, s.Opts)
+}
+
+// Reader returns an io.Reader that reads from s. Reads beyond the end of s
+// return io.EOF. The preconditions that apply to s.CopyIn also apply to the
+// returned io.Reader.Read.
+func (s IOSequence) Reader(ctx context.Context) io.Reader {
+	return &ioSequenceReadWriter{ctx, s}
+}
+
+// Writer returns an io.Writer that writes to s. Writes beyond the end of s
+// return ErrEndOfIOSequence. The preconditions that apply to s.CopyOut also
+// apply to the returned io.Writer.Write.
+func (s IOSequence) Writer(ctx context.Context) io.Writer {
+	return &ioSequenceReadWriter{ctx, s}
+}
+
+// ErrEndOfIOSequence is returned by IOSequence.Writer().Write() when
+// attempting to write beyond the end of the IOSequence.
+var ErrEndOfIOSequence = errors.New("write beyond end of IOSequence")
+
+type ioSequenceReadWriter struct {
+	ctx context.Context
+	s   IOSequence
+}
+
+// Read implements io.Reader.Read.
+func (rw *ioSequenceReadWriter) Read(dst []byte) (int, error) {
+	n, err := rw.s.CopyIn(rw.ctx, dst)
+	rw.s = rw.s.DropFirst(n)
+	if err == nil && rw.s.NumBytes() == 0 {
+		err = io.EOF
+	}
+	return n, err
+}
+
+// Write implements io.Writer.Write.
+func (rw *ioSequenceReadWriter) Write(src []byte) (int, error) {
+	n, err := rw.s.CopyOut(rw.ctx, src)
+	rw.s = rw.s.DropFirst(n)
+	if err == nil && n < len(src) {
+		err = ErrEndOfIOSequence
+	}
+	return n, err
+}
diff --git a/pkg/sentry/usermem/usermem_test.go b/pkg/sentry/usermem/usermem_test.go
new file mode 100644
index 000000000..563560da8
--- /dev/null
+++ b/pkg/sentry/usermem/usermem_test.go
@@ -0,0 +1,411 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+package usermem
+
+import (
+	"bytes"
+	"encoding/binary"
+	"fmt"
+	"reflect"
+	"strings"
+	"testing"
+
+	"gvisor.googlesource.com/gvisor/pkg/sentry/context"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
+)
+
+// newContext returns a context.Context that we can use in these tests (we
+// can't use contexttest because it depends on usermem).
+func newContext() context.Context {
+	return context.Background()
+}
+
+func newBytesIOString(s string) *BytesIO {
+	return &BytesIO{[]byte(s)}
+}
+
+func TestBytesIOCopyOutSuccess(t *testing.T) {
+	b := newBytesIOString("ABCDE")
+	n, err := b.CopyOut(newContext(), 1, []byte("foo"), IOOpts{})
+	if wantN := 3; n != wantN || err != nil {
+		t.Errorf("CopyOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := b.Bytes, []byte("AfooE"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyOutFailure(t *testing.T) {
+	b := newBytesIOString("ABC")
+	n, err := b.CopyOut(newContext(), 1, []byte("foo"), IOOpts{})
+	if wantN, wantErr := 2, syserror.EFAULT; n != wantN || err != wantErr {
+		t.Errorf("CopyOut: got (%v, %v), wanted (%v, %v)", n, err, wantN, wantErr)
+	}
+	if got, want := b.Bytes, []byte("Afo"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyInSuccess(t *testing.T) {
+	b := newBytesIOString("AfooE")
+	var dst [3]byte
+	n, err := b.CopyIn(newContext(), 1, dst[:], IOOpts{})
+	if wantN := 3; n != wantN || err != nil {
+		t.Errorf("CopyIn: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst[:], []byte("foo"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyInFailure(t *testing.T) {
+	b := newBytesIOString("Afo")
+	var dst [3]byte
+	n, err := b.CopyIn(newContext(), 1, dst[:], IOOpts{})
+	if wantN, wantErr := 2, syserror.EFAULT; n != wantN || err != wantErr {
+		t.Errorf("CopyIn: got (%v, %v), wanted (%v, %v)", n, err, wantN, wantErr)
+	}
+	if got, want := dst[:], []byte("fo\x00"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOZeroOutSuccess(t *testing.T) {
+	b := newBytesIOString("ABCD")
+	n, err := b.ZeroOut(newContext(), 1, 2, IOOpts{})
+	if wantN := int64(2); n != wantN || err != nil {
+		t.Errorf("ZeroOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := b.Bytes, []byte("A\x00\x00D"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOZeroOutFailure(t *testing.T) {
+	b := newBytesIOString("ABC")
+	n, err := b.ZeroOut(newContext(), 1, 3, IOOpts{})
+	if wantN, wantErr := int64(2), syserror.EFAULT; n != wantN || err != wantErr {
+		t.Errorf("ZeroOut: got (%v, %v), wanted (%v, %v)", n, err, wantN, wantErr)
+	}
+	if got, want := b.Bytes, []byte("A\x00\x00"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyOutFromSuccess(t *testing.T) {
+	b := newBytesIOString("ABCDEFGH")
+	n, err := b.CopyOutFrom(newContext(), AddrRangeSeqFromSlice([]AddrRange{
+		{Start: 4, End: 7},
+		{Start: 1, End: 4},
+	}), safemem.FromIOReader{bytes.NewBufferString("barfoo")}, IOOpts{})
+	if wantN := int64(6); n != wantN || err != nil {
+		t.Errorf("CopyOutFrom: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := b.Bytes, []byte("AfoobarH"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyOutFromFailure(t *testing.T) {
+	b := newBytesIOString("ABCDE")
+	n, err := b.CopyOutFrom(newContext(), AddrRangeSeqFromSlice([]AddrRange{
+		{Start: 1, End: 4},
+		{Start: 4, End: 7},
+	}), safemem.FromIOReader{bytes.NewBufferString("foobar")}, IOOpts{})
+	if wantN, wantErr := int64(4), syserror.EFAULT; n != wantN || err != wantErr {
+		t.Errorf("CopyOutFrom: got (%v, %v), wanted (%v, %v)", n, err, wantN, wantErr)
+	}
+	if got, want := b.Bytes, []byte("Afoob"); !bytes.Equal(got, want) {
+		t.Errorf("Bytes: got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyInToSuccess(t *testing.T) {
+	b := newBytesIOString("AfoobarH")
+	var dst bytes.Buffer
+	n, err := b.CopyInTo(newContext(), AddrRangeSeqFromSlice([]AddrRange{
+		{Start: 4, End: 7},
+		{Start: 1, End: 4},
+	}), safemem.FromIOWriter{&dst}, IOOpts{})
+	if wantN := int64(6); n != wantN || err != nil {
+		t.Errorf("CopyInTo: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst.Bytes(), []byte("barfoo"); !bytes.Equal(got, want) {
+		t.Errorf("dst.Bytes(): got %q, wanted %q", got, want)
+	}
+}
+
+func TestBytesIOCopyInToFailure(t *testing.T) {
+	b := newBytesIOString("Afoob")
+	var dst bytes.Buffer
+	n, err := b.CopyInTo(newContext(), AddrRangeSeqFromSlice([]AddrRange{
+		{Start: 1, End: 4},
+		{Start: 4, End: 7},
+	}), safemem.FromIOWriter{&dst}, IOOpts{})
+	if wantN, wantErr := int64(4), syserror.EFAULT; n != wantN || err != wantErr {
+		t.Errorf("CopyOutFrom: got (%v, %v), wanted (%v, %v)", n, err, wantN, wantErr)
+	}
+	if got, want := dst.Bytes(), []byte("foob"); !bytes.Equal(got, want) {
+		t.Errorf("dst.Bytes(): got %q, wanted %q", got, want)
+	}
+}
+
+type testStruct struct {
+	Int8   int8
+	Uint8  uint8
+	Int16  int16
+	Uint16 uint16
+	Int32  int32
+	Uint32 uint32
+	Int64  int64
+	Uint64 uint64
+}
+
+func TestCopyObject(t *testing.T) {
+	wantObj := testStruct{1, 2, 3, 4, 5, 6, 7, 8}
+	wantN := binary.Size(wantObj)
+	b := &BytesIO{make([]byte, wantN)}
+	ctx := newContext()
+	if n, err := CopyObjectOut(ctx, b, 0, &wantObj, IOOpts{}); n != wantN || err != nil {
+		t.Fatalf("CopyObjectOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	var gotObj testStruct
+	if n, err := CopyObjectIn(ctx, b, 0, &gotObj, IOOpts{}); n != wantN || err != nil {
+		t.Errorf("CopyObjectIn: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if gotObj != wantObj {
+		t.Errorf("CopyObject round trip: got %+v, wanted %+v", gotObj, wantObj)
+	}
+}
+
+func TestCopyStringInShort(t *testing.T) {
+	want := strings.Repeat("A", copyStringIncrement-2)
+	mem := want + "\x00"
+	if got, err := CopyStringIn(newContext(), newBytesIOString(mem), 0, 2*copyStringIncrement, IOOpts{}); got != want || err != nil {
+		t.Errorf("CopyStringIn: got (%q, %v), wanted (%q, nil)", got, err, want)
+	}
+}
+
+func TestCopyStringInLong(t *testing.T) {
+	want := strings.Repeat("A", copyStringIncrement+1)
+	mem := want + "\x00"
+	if got, err := CopyStringIn(newContext(), newBytesIOString(mem), 0, 2*copyStringIncrement, IOOpts{}); got != want || err != nil {
+		t.Errorf("CopyStringIn: got (%q, %v), wanted (%q, nil)", got, err, want)
+	}
+}
+
+func TestCopyStringInNoTerminatingZeroByte(t *testing.T) {
+	want := strings.Repeat("A", copyStringIncrement-1)
+	got, err := CopyStringIn(newContext(), newBytesIOString(want), 0, 2*copyStringIncrement, IOOpts{})
+	if wantErr := syserror.EFAULT; got != want || err != wantErr {
+		t.Errorf("CopyStringIn: got (%q, %v), wanted (%q, %v)", got, err, want, wantErr)
+	}
+}
+
+func TestCopyStringInTruncatedByMaxlen(t *testing.T) {
+	got, err := CopyStringIn(newContext(), newBytesIOString(strings.Repeat("A", 10)), 0, 5, IOOpts{})
+	if want, wantErr := strings.Repeat("A", 5), syserror.ENAMETOOLONG; got != want || err != wantErr {
+		t.Errorf("CopyStringIn: got (%q, %v), wanted (%q, %v)", got, err, want, wantErr)
+	}
+}
+
+func TestCopyInt32StringsInVec(t *testing.T) {
+	for _, test := range []struct {
+		str     string
+		n       int
+		initial []int32
+		final   []int32
+	}{
+		{
+			str:     "100 200",
+			n:       len("100 200"),
+			initial: []int32{1, 2},
+			final:   []int32{100, 200},
+		},
+		{
+			// Fewer values ok
+			str:     "100",
+			n:       len("100"),
+			initial: []int32{1, 2},
+			final:   []int32{100, 2},
+		},
+		{
+			// Extra values ok
+			str:     "100 200 300",
+			n:       len("100 200 "),
+			initial: []int32{1, 2},
+			final:   []int32{100, 200},
+		},
+		{
+			// Leading and trailing whitespace ok
+			str:     " 100\t200\n",
+			n:       len(" 100\t200\n"),
+			initial: []int32{1, 2},
+			final:   []int32{100, 200},
+		},
+	} {
+		t.Run(fmt.Sprintf("%q", test.str), func(t *testing.T) {
+			src := BytesIOSequence([]byte(test.str))
+			dsts := append([]int32(nil), test.initial...)
+			if n, err := CopyInt32StringsInVec(newContext(), src.IO, src.Addrs, dsts, src.Opts); n != int64(test.n) || err != nil {
+				t.Errorf("CopyInt32StringsInVec: got (%d, %v), wanted (%d, nil)", n, err, test.n)
+			}
+			if !reflect.DeepEqual(dsts, test.final) {
+				t.Errorf("dsts: got %v, wanted %v", dsts, test.final)
+			}
+		})
+	}
+}
+
+func TestCopyInt32StringsInVecRequiresOneValidValue(t *testing.T) {
+	for _, s := range []string{"", "\n", "a123"} {
+		t.Run(fmt.Sprintf("%q", s), func(t *testing.T) {
+			src := BytesIOSequence([]byte(s))
+			initial := []int32{1, 2}
+			dsts := append([]int32(nil), initial...)
+			if n, err := CopyInt32StringsInVec(newContext(), src.IO, src.Addrs, dsts, src.Opts); err != syserror.EINVAL {
+				t.Errorf("CopyInt32StringsInVec: got (%d, %v), wanted (_, %v)", n, err, syserror.EINVAL)
+			}
+			if !reflect.DeepEqual(dsts, initial) {
+				t.Errorf("dsts: got %v, wanted %v", dsts, initial)
+			}
+		})
+	}
+}
+
+func TestIOSequenceCopyOut(t *testing.T) {
+	buf := []byte("ABCD")
+	s := BytesIOSequence(buf)
+
+	// CopyOut limited by len(src).
+	n, err := s.CopyOut(newContext(), []byte("fo"))
+	if wantN := 2; n != wantN || err != nil {
+		t.Errorf("CopyOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("foCD"); !bytes.Equal(buf, want) {
+		t.Errorf("buf: got %q, wanted %q", buf, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(2); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	// CopyOut limited by s.NumBytes().
+	n, err = s.CopyOut(newContext(), []byte("obar"))
+	if wantN := 2; n != wantN || err != nil {
+		t.Errorf("CopyOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("foob"); !bytes.Equal(buf, want) {
+		t.Errorf("buf: got %q, wanted %q", buf, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(0); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+}
+
+func TestIOSequenceCopyIn(t *testing.T) {
+	s := BytesIOSequence([]byte("foob"))
+	dst := []byte("ABCDEF")
+
+	// CopyIn limited by len(dst).
+	n, err := s.CopyIn(newContext(), dst[:2])
+	if wantN := 2; n != wantN || err != nil {
+		t.Errorf("CopyIn: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("foCDEF"); !bytes.Equal(dst, want) {
+		t.Errorf("dst: got %q, wanted %q", dst, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(2); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	// CopyIn limited by s.Remaining().
+	n, err = s.CopyIn(newContext(), dst[2:])
+	if wantN := 2; n != wantN || err != nil {
+		t.Errorf("CopyIn: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("foobEF"); !bytes.Equal(dst, want) {
+		t.Errorf("dst: got %q, wanted %q", dst, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(0); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+}
+
+func TestIOSequenceZeroOut(t *testing.T) {
+	buf := []byte("ABCD")
+	s := BytesIOSequence(buf)
+
+	// ZeroOut limited by toZero.
+	n, err := s.ZeroOut(newContext(), 2)
+	if wantN := int64(2); n != wantN || err != nil {
+		t.Errorf("ZeroOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("\x00\x00CD"); !bytes.Equal(buf, want) {
+		t.Errorf("buf: got %q, wanted %q", buf, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(2); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	// ZeroOut limited by s.NumBytes().
+	n, err = s.ZeroOut(newContext(), 4)
+	if wantN := int64(2); n != wantN || err != nil {
+		t.Errorf("CopyOut: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if want := []byte("\x00\x00\x00\x00"); !bytes.Equal(buf, want) {
+		t.Errorf("buf: got %q, wanted %q", buf, want)
+	}
+	s = s.DropFirst(2)
+	if got, want := s.NumBytes(), int64(0); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+}
+
+func TestIOSequenceTakeFirst(t *testing.T) {
+	s := BytesIOSequence([]byte("foobar"))
+	if got, want := s.NumBytes(), int64(6); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	s = s.TakeFirst(3)
+	if got, want := s.NumBytes(), int64(3); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	// TakeFirst(n) where n > s.NumBytes() is a no-op.
+	s = s.TakeFirst(9)
+	if got, want := s.NumBytes(), int64(3); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+
+	var dst [3]byte
+	n, err := s.CopyIn(newContext(), dst[:])
+	if wantN := 3; n != wantN || err != nil {
+		t.Errorf("CopyIn: got (%v, %v), wanted (%v, nil)", n, err, wantN)
+	}
+	if got, want := dst[:], []byte("foo"); !bytes.Equal(got, want) {
+		t.Errorf("dst: got %q, wanted %q", got, want)
+	}
+	s = s.DropFirst(3)
+	if got, want := s.NumBytes(), int64(0); got != want {
+		t.Errorf("NumBytes: got %v, wanted %v", got, want)
+	}
+}
diff --git a/pkg/sentry/usermem/usermem_x86.go b/pkg/sentry/usermem/usermem_x86.go
new file mode 100644
index 000000000..2484b0d82
--- /dev/null
+++ b/pkg/sentry/usermem/usermem_x86.go
@@ -0,0 +1,38 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// +build amd64 i386
+
+package usermem
+
+import "encoding/binary"
+
+const (
+	// PageSize is the system page size.
+	PageSize = 1 << PageShift
+
+	// HugePageSize is the system huge page size.
+	HugePageSize = 1 << HugePageShift
+
+	// PageShift is the binary log of the system page size.
+	PageShift = 12
+
+	// HugePageShift is the binary log of the system huge page size.
+	HugePageShift = 21
+)
+
+var (
+	// ByteOrder is the native byte order (little endian).
+	ByteOrder = binary.LittleEndian
+)
diff --git a/pkg/sentry/watchdog/BUILD b/pkg/sentry/watchdog/BUILD
new file mode 100644
index 000000000..28fae4490
--- /dev/null
+++ b/pkg/sentry/watchdog/BUILD
@@ -0,0 +1,17 @@
+package(licenses = ["notice"])  # Apache 2.0
+
+load("@io_bazel_rules_go//go:def.bzl", "go_library")
+
+go_library(
+    name = "watchdog",
+    srcs = ["watchdog.go"],
+    importpath = "gvisor.googlesource.com/gvisor/pkg/sentry/watchdog",
+    visibility = ["//:sandbox"],
+    deps = [
+        "//pkg/abi/linux",
+        "//pkg/log",
+        "//pkg/metric",
+        "//pkg/sentry/kernel",
+        "//pkg/sentry/kernel/time",
+    ],
+)
diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go
new file mode 100644
index 000000000..5b620693d
--- /dev/null
+++ b/pkg/sentry/watchdog/watchdog.go
@@ -0,0 +1,279 @@
+// Copyright 2018 Google Inc.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Package watchdog is responsible for monitoring the sentry for tasks that may
+// potentially be stuck or looping inderterminally causing hard to debug hungs in
+// the untrusted app.
+//
+// It works by periodically querying all tasks to check whether they are in user
+// mode (RunUser), kernel mode (RunSys), or blocked in the kernel (OffCPU). Tasks
+// that have been running in kernel mode for a long time in the same syscall
+// without blocking are considered stuck and are reported.
+//
+// When a stuck task is detected, the watchdog can take one of the following actions:
+//		1. LogWarning: Logs a warning message followed by a stack dump of all goroutines.
+//			 If a tasks continues to be stuck, the message will repeat every minute, unless
+//			 a new stuck task is detected
+//		2. Panic: same as above, followed by panic()
+//
+package watchdog
+
+import (
+	"bytes"
+	"fmt"
+	"sync"
+	"time"
+
+	"gvisor.googlesource.com/gvisor/pkg/abi/linux"
+	"gvisor.googlesource.com/gvisor/pkg/log"
+	"gvisor.googlesource.com/gvisor/pkg/metric"
+	"gvisor.googlesource.com/gvisor/pkg/sentry/kernel"
+	ktime "gvisor.googlesource.com/gvisor/pkg/sentry/kernel/time"
+)
+
+// DefaultTimeout is a resonable timeout value for most applications.
+const DefaultTimeout = 3 * time.Minute
+
+// descheduleThreshold is the amount of time scheduling needs to be off before the entire wait period
+// is discounted from task's last update time. It's set high enough that small scheduling delays won't
+// trigger it.
+const descheduleThreshold = 1 * time.Second
+
+var stuckTasks = metric.MustCreateNewUint64Metric("/watchdog/stuck_tasks_detected", true /* sync */, "Cumulative count of stuck tasks detected")
+
+// Amount of time to wait before dumping the stack to the log again when the same task(s) remains stuck.
+var stackDumpSameTaskPeriod = time.Minute
+
+// Action defines what action to take when a stuck task is detected.
+type Action int
+
+const (
+	// LogWarning logs warning message followed by stack trace.
+	LogWarning Action = iota
+	// Panic will do the same logging as LogWarning and panic().
+	Panic
+)
+
+// String returns Action's string representation.
+func (a Action) String() string {
+	switch a {
+	case LogWarning:
+		return "LogWarning"
+	case Panic:
+		return "Panic"
+	default:
+		panic(fmt.Sprintf("Invalid action: %d", a))
+	}
+}
+
+// Watchdog is the main watchdog class. It controls a goroutine that periodically
+// analyses all tasks and reports if any of them appear to be stuck.
+type Watchdog struct {
+	// period indicates how often to check all tasks. It's calculated based on
+	// 'taskTimeout'.
+	period time.Duration
+
+	// taskTimeout is the amount of time to allow a task to execute the same syscall
+	// without blocking before it's declared stuck.
+	taskTimeout time.Duration
+
+	// timeoutAction indicates what action to take when a stuck tasks is detected.
+	timeoutAction Action
+
+	// k is where the tasks come from.
+	k *kernel.Kernel
+
+	// stop is used to notify to watchdog should stop.
+	stop chan struct{}
+
+	// done is used to notify when the watchdog has stopped.
+	done chan struct{}
+
+	// offenders map contains all tasks that are currently stuck.
+	offenders map[*kernel.Task]*offender
+
+	// lastStackDump tracks the last time a stack dump was generated to prevent
+	// spamming the log.
+	lastStackDump time.Time
+
+	// lastRun is set to the last time the watchdog executed a monitoring loop.
+	lastRun ktime.Time
+
+	// mu protects the fields below.
+	mu sync.Mutex
+
+	// started is true if the watchdog has been started before.
+	started bool
+}
+
+type offender struct {
+	lastUpdateTime ktime.Time
+}
+
+// New creates a new watchdog.
+func New(k *kernel.Kernel, taskTimeout time.Duration, a Action) *Watchdog {
+	// 4 is arbitrary, just don't want to prolong 'taskTimeout' too much.
+	period := taskTimeout / 4
+	return &Watchdog{
+		k:             k,
+		period:        period,
+		taskTimeout:   taskTimeout,
+		timeoutAction: a,
+		offenders:     make(map[*kernel.Task]*offender),
+		stop:          make(chan struct{}),
+		done:          make(chan struct{}),
+	}
+}
+
+// Start starts the watchdog.
+func (w *Watchdog) Start() {
+	if w.taskTimeout == 0 {
+		log.Infof("Watchdog disabled")
+		return
+	}
+
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if w.started {
+		return
+	}
+
+	w.lastRun = w.k.MonotonicClock().Now()
+
+	log.Infof("Starting watchdog, period: %v, timeout: %v, action: %v", w.period, w.taskTimeout, w.timeoutAction)
+	go w.loop() // S/R-SAFE: watchdog is stopped during save and restarted after restore.
+	w.started = true
+}
+
+// Stop requests the watchdog to stop and wait for it.
+func (w *Watchdog) Stop() {
+	if w.taskTimeout == 0 {
+		return
+	}
+
+	w.mu.Lock()
+	defer w.mu.Unlock()
+	if !w.started {
+		return
+	}
+	log.Infof("Stopping watchdog")
+	w.stop <- struct{}{}
+	<-w.done
+	w.started = false
+	log.Infof("Watchdog stopped")
+}
+
+// loop is the main watchdog routine. It only returns when 'Stop()' is called.
+func (w *Watchdog) loop() {
+	// Loop until someone stops it.
+	for {
+		select {
+		case <-w.stop:
+			w.done <- struct{}{}
+			return
+		case <-time.After(w.period):
+			w.runTurn()
+		}
+	}
+}
+
+// runTurn runs a single pass over all tasks and reports anything it finds.
+func (w *Watchdog) runTurn() {
+	tasks := w.k.TaskSet().Root.Tasks()
+
+	newOffenders := make(map[*kernel.Task]*offender)
+	newTaskFound := false
+	now := ktime.FromNanoseconds(int64(w.k.CPUClockNow() * uint64(linux.ClockTick)))
+
+	// The process may be running with low CPU limit making tasks appear stuck because
+	// are starved of CPU cycles. An estimate is that Tasks could have been starved
+	// since the last time the watchdog run. If the watchdog detects that scheduling
+	// is off, it will discount the entire duration since last run from 'lastUpdateTime'.
+	discount := time.Duration(0)
+	if now.Sub(w.lastRun.Add(w.period)) > descheduleThreshold {
+		discount = now.Sub(w.lastRun)
+	}
+	w.lastRun = now
+
+	log.Infof("Watchdog starting loop, tasks: %d, discount: %v", len(tasks), discount)
+	for _, t := range tasks {
+		tsched := t.TaskGoroutineSchedInfo()
+
+		// An offender is a task running inside the kernel for longer than the specified timeout.
+		if tsched.State == kernel.TaskGoroutineRunningSys {
+			lastUpdateTime := ktime.FromNanoseconds(int64(tsched.Timestamp * uint64(linux.ClockTick)))
+			elapsed := now.Sub(lastUpdateTime) - discount
+			if elapsed > w.taskTimeout {
+				tc, ok := w.offenders[t]
+				if !ok {
+					// New stuck task detected.
+					//
+					// TODO: Tasks blocked doing IO may be considered stuck in kernel.
+					tc = &offender{lastUpdateTime: lastUpdateTime}
+					stuckTasks.Increment()
+					newTaskFound = true
+				}
+				newOffenders[t] = tc
+			}
+		}
+	}
+	if len(newOffenders) > 0 {
+		w.report(newOffenders, newTaskFound, now)
+	}
+
+	// Remember which tasks have been reported.
+	w.offenders = newOffenders
+}
+
+// report takes appropriate action when a stuck task is detected.
+func (w *Watchdog) report(offenders map[*kernel.Task]*offender, newTaskFound bool, now ktime.Time) {
+	var buf bytes.Buffer
+	buf.WriteString(fmt.Sprintf("Sentry detected %d stuck task(s):\n", len(offenders)))
+	for t, o := range offenders {
+		tid := w.k.TaskSet().Root.IDOfTask(t)
+		buf.WriteString(fmt.Sprintf("\tTask tid: %v (%#x), entered RunSys state %v ago.\n", tid, uint64(tid), now.Sub(o.lastUpdateTime)))
+	}
+	buf.WriteString("Search for '(*Task).run(0x..., 0x<tid>)' in the stack dump to find the offending goroutine")
+
+	switch w.timeoutAction {
+	case LogWarning:
+		// Dump stack only if a new task is detected or if it sometime has passed since
+		// the last time a stack dump was generated.
+		if !newTaskFound && time.Since(w.lastStackDump) < stackDumpSameTaskPeriod {
+			buf.WriteString("\n...[stack dump skipped]...")
+			log.Warningf(buf.String())
+		} else {
+			log.TracebackAll(buf.String())
+			w.lastStackDump = time.Now()
+		}
+
+	case Panic:
+		// Panic will skip over running tasks, which is likely the culprit here. So manually
+		// dump all stacks before panic'ing.
+		log.TracebackAll(buf.String())
+
+		// Attempt to flush metrics, timeout and move on in case metrics are stuck as well.
+		metricsEmitted := make(chan struct{}, 1)
+		go func() { // S/R-SAFE: watchdog is stopped during save and restarted after restore.
+			// Flush metrics before killing process.
+			metric.EmitMetricUpdate()
+			metricsEmitted <- struct{}{}
+		}()
+		select {
+		case <-metricsEmitted:
+		case <-time.After(1 * time.Second):
+		}
+		panic("Sentry detected stuck task(s). See stack trace and message above for more details")
+	}
+}