diff options
Diffstat (limited to 'pkg/sentry')
436 files changed, 17389 insertions, 6571 deletions
diff --git a/pkg/sentry/arch/BUILD b/pkg/sentry/arch/BUILD index 901e0f320..4af4d6e84 100644 --- a/pkg/sentry/arch/BUILD +++ b/pkg/sentry/arch/BUILD @@ -22,6 +22,7 @@ go_library( "signal_info.go", "signal_stack.go", "stack.go", + "stack_unsafe.go", "syscalls_amd64.go", "syscalls_arm64.go", ], @@ -33,11 +34,12 @@ go_library( "//pkg/context", "//pkg/cpuid", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/sentry/limits", "//pkg/sync", "//pkg/syserror", "//pkg/usermem", - "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/arch/arch.go b/pkg/sentry/arch/arch.go index a903d031c..d75d665ae 100644 --- a/pkg/sentry/arch/arch.go +++ b/pkg/sentry/arch/arch.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/usermem" ) @@ -72,12 +73,12 @@ type Context interface { // 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{} + Native(val uintptr) marshal.Marshallable // 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 + Value(val marshal.Marshallable) uintptr // Width returns the number of bytes for a native value. Width() uint @@ -205,7 +206,7 @@ type Context interface { // equivalent of arch_ptrace(): // PtracePeekUser implements ptrace(PTRACE_PEEKUSR). - PtracePeekUser(addr uintptr) (interface{}, error) + PtracePeekUser(addr uintptr) (marshal.Marshallable, error) // PtracePokeUser implements ptrace(PTRACE_POKEUSR). PtracePokeUser(addr, data uintptr) error diff --git a/pkg/sentry/arch/arch_aarch64.go b/pkg/sentry/arch/arch_aarch64.go index 0f433ee79..fd73751e7 100644 --- a/pkg/sentry/arch/arch_aarch64.go +++ b/pkg/sentry/arch/arch_aarch64.go @@ -154,6 +154,7 @@ func (s State) Proto() *rpb.Registers { Sp: s.Regs.Sp, Pc: s.Regs.Pc, Pstate: s.Regs.Pstate, + Tls: s.Regs.TPIDR_EL0, } return &rpb.Registers{Arch: &rpb.Registers_Arm64{Arm64: regs}} } @@ -232,6 +233,7 @@ func (s *State) RegisterMap() (map[string]uintptr, error) { "Sp": uintptr(s.Regs.Sp), "Pc": uintptr(s.Regs.Pc), "Pstate": uintptr(s.Regs.Pstate), + "Tls": uintptr(s.Regs.TPIDR_EL0), }, nil } diff --git a/pkg/sentry/arch/arch_amd64.go b/pkg/sentry/arch/arch_amd64.go index 1c3e3c14c..c7d3a206d 100644 --- a/pkg/sentry/arch/arch_amd64.go +++ b/pkg/sentry/arch/arch_amd64.go @@ -23,6 +23,8 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/usermem" ) @@ -179,14 +181,14 @@ func (c *context64) SetOldRSeqInterruptedIP(value uintptr) { } // Native returns the native type for the given val. -func (c *context64) Native(val uintptr) interface{} { - v := uint64(val) +func (c *context64) Native(val uintptr) marshal.Marshallable { + v := primitive.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)) +func (c *context64) Value(val marshal.Marshallable) uintptr { + return uintptr(*val.(*primitive.Uint64)) } // Width returns the byte width of this architecture. @@ -293,7 +295,7 @@ func (c *context64) PIELoadAddress(l MmapLayout) usermem.Addr { const userStructSize = 928 // PtracePeekUser implements Context.PtracePeekUser. -func (c *context64) PtracePeekUser(addr uintptr) (interface{}, error) { +func (c *context64) PtracePeekUser(addr uintptr) (marshal.Marshallable, error) { if addr&7 != 0 || addr >= userStructSize { return nil, syscall.EIO } diff --git a/pkg/sentry/arch/arch_arm64.go b/pkg/sentry/arch/arch_arm64.go index 550741d8c..680d23a9f 100644 --- a/pkg/sentry/arch/arch_arm64.go +++ b/pkg/sentry/arch/arch_arm64.go @@ -22,6 +22,8 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/cpuid" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/usermem" ) @@ -163,14 +165,14 @@ func (c *context64) SetOldRSeqInterruptedIP(value uintptr) { } // Native returns the native type for the given val. -func (c *context64) Native(val uintptr) interface{} { - v := uint64(val) +func (c *context64) Native(val uintptr) marshal.Marshallable { + v := primitive.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)) +func (c *context64) Value(val marshal.Marshallable) uintptr { + return uintptr(*val.(*primitive.Uint64)) } // Width returns the byte width of this architecture. @@ -274,7 +276,7 @@ func (c *context64) PIELoadAddress(l MmapLayout) usermem.Addr { } // PtracePeekUser implements Context.PtracePeekUser. -func (c *context64) PtracePeekUser(addr uintptr) (interface{}, error) { +func (c *context64) PtracePeekUser(addr uintptr) (marshal.Marshallable, error) { // TODO(gvisor.dev/issue/1239): Full ptrace supporting for Arm64. return c.Native(0), nil } diff --git a/pkg/sentry/arch/registers.proto b/pkg/sentry/arch/registers.proto index 60c027aab..2727ba08a 100644 --- a/pkg/sentry/arch/registers.proto +++ b/pkg/sentry/arch/registers.proto @@ -83,6 +83,7 @@ message ARM64Registers { uint64 sp = 32; uint64 pc = 33; uint64 pstate = 34; + uint64 tls = 35; } message Registers { oneof arch { diff --git a/pkg/sentry/arch/signal_act.go b/pkg/sentry/arch/signal_act.go index 32173aa20..d3e2324a8 100644 --- a/pkg/sentry/arch/signal_act.go +++ b/pkg/sentry/arch/signal_act.go @@ -14,7 +14,7 @@ package arch -import "gvisor.dev/gvisor/tools/go_marshal/marshal" +import "gvisor.dev/gvisor/pkg/marshal" // Special values for SignalAct.Handler. const ( diff --git a/pkg/sentry/arch/signal_amd64.go b/pkg/sentry/arch/signal_amd64.go index 6fb756f0e..72e07a988 100644 --- a/pkg/sentry/arch/signal_amd64.go +++ b/pkg/sentry/arch/signal_amd64.go @@ -17,17 +17,19 @@ package arch import ( - "encoding/binary" "math" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/usermem" ) // SignalContext64 is equivalent to struct sigcontext, the type passed as the // second argument to signal handlers set by signal(2). +// +// +marshal type SignalContext64 struct { R8 uint64 R9 uint64 @@ -68,6 +70,8 @@ const ( ) // UContext64 is equivalent to ucontext_t on 64-bit x86. +// +// +marshal type UContext64 struct { Flags uint64 Link uint64 @@ -172,12 +176,7 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt // "... 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") - } + ucSize := uc.SizeBytes() // 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 @@ -195,18 +194,18 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt info.FixSignalCodeForUser() // Set up the stack frame. - infoAddr, err := st.Push(info) - if err != nil { + if _, err := info.CopyOut(st, StackBottomMagic); err != nil { return err } - ucAddr, err := st.Push(uc) - if err != nil { + infoAddr := st.Bottom + if _, err := uc.CopyOut(st, StackBottomMagic); err != nil { return err } + ucAddr := st.Bottom 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 { + if _, err := primitive.CopyUint64Out(st, StackBottomMagic, act.Restorer); err != nil { return err } } else { @@ -240,11 +239,11 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { // Copy out the stack frame. var uc UContext64 - if _, err := st.Pop(&uc); err != nil { + if _, err := uc.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } var info SignalInfo - if _, err := st.Pop(&info); err != nil { + if _, err := info.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } diff --git a/pkg/sentry/arch/signal_arm64.go b/pkg/sentry/arch/signal_arm64.go index 642c79dda..7fde5d34e 100644 --- a/pkg/sentry/arch/signal_arm64.go +++ b/pkg/sentry/arch/signal_arm64.go @@ -12,10 +12,11 @@ // See the License for the specific language governing permissions and // limitations under the License. +// +build arm64 + package arch import ( - "encoding/binary" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" @@ -25,6 +26,8 @@ import ( // SignalContext64 is equivalent to struct sigcontext, the type passed as the // second argument to signal handlers set by signal(2). +// +// +marshal type SignalContext64 struct { FaultAddr uint64 Regs [31]uint64 @@ -36,6 +39,7 @@ type SignalContext64 struct { Reserved [3568]uint8 } +// +marshal type aarch64Ctx struct { Magic uint32 Size uint32 @@ -43,6 +47,8 @@ type aarch64Ctx struct { // FpsimdContext is equivalent to struct fpsimd_context on arm64 // (arch/arm64/include/uapi/asm/sigcontext.h). +// +// +marshal type FpsimdContext struct { Head aarch64Ctx Fpsr uint32 @@ -51,13 +57,15 @@ type FpsimdContext struct { } // UContext64 is equivalent to ucontext on arm64(arch/arm64/include/uapi/asm/ucontext.h). +// +// +marshal type UContext64 struct { Flags uint64 Link uint64 Stack SignalStack Sigset linux.SignalSet // glibc uses a 1024-bit sigset_t - _pad [(1024 - 64) / 8]byte + _pad [120]byte // (1024 - 64) / 8 = 120 // sigcontext must be aligned to 16-byte _pad2 [8]byte // last for future expansion @@ -94,11 +102,7 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt }, Sigset: sigset, } - - ucSize := binary.Size(uc) - if ucSize < 0 { - panic("can't get size of UContext64") - } + ucSize := uc.SizeBytes() // frameSize = ucSize + sizeof(siginfo). // sizeof(siginfo) == 128. @@ -119,14 +123,14 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt info.FixSignalCodeForUser() // Set up the stack frame. - infoAddr, err := st.Push(info) - if err != nil { + if _, err := info.CopyOut(st, StackBottomMagic); err != nil { return err } - ucAddr, err := st.Push(uc) - if err != nil { + infoAddr := st.Bottom + if _, err := uc.CopyOut(st, StackBottomMagic); err != nil { return err } + ucAddr := st.Bottom // Set up registers. c.Regs.Sp = uint64(st.Bottom) @@ -147,11 +151,11 @@ func (c *context64) SignalSetup(st *Stack, act *SignalAct, info *SignalInfo, alt func (c *context64) SignalRestore(st *Stack, rt bool) (linux.SignalSet, SignalStack, error) { // Copy out the stack frame. var uc UContext64 - if _, err := st.Pop(&uc); err != nil { + if _, err := uc.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } var info SignalInfo - if _, err := st.Pop(&info); err != nil { + if _, err := info.CopyIn(st, StackBottomMagic); err != nil { return 0, SignalStack{}, err } diff --git a/pkg/sentry/arch/signal_stack.go b/pkg/sentry/arch/signal_stack.go index 0fa738a1d..a1eae98f9 100644 --- a/pkg/sentry/arch/signal_stack.go +++ b/pkg/sentry/arch/signal_stack.go @@ -17,8 +17,8 @@ package arch import ( + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) const ( diff --git a/pkg/sentry/arch/stack.go b/pkg/sentry/arch/stack.go index 1108fa0bd..5f06c751d 100644 --- a/pkg/sentry/arch/stack.go +++ b/pkg/sentry/arch/stack.go @@ -15,14 +15,16 @@ package arch import ( - "encoding/binary" - "fmt" - "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/usermem" ) -// Stack is a simple wrapper around a usermem.IO and an address. +// Stack is a simple wrapper around a usermem.IO and an address. Stack +// implements marshal.CopyContext, and marshallable values can be pushed or +// popped from the stack through the marshal.Marshallable interface. +// +// Stack is not thread-safe. type Stack struct { // Our arch info. // We use this for automatic Native conversion of usermem.Addrs during @@ -34,105 +36,60 @@ type Stack struct { // 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 - } + // Scratch buffer used for marshalling to avoid having to repeatedly + // allocate scratch memory. + scratchBuf []byte +} - if term != nil { - _, err := s.Push(term) - if err != nil { - return 0, err - } - } +// scratchBufLen is the default length of Stack.scratchBuf. The +// largest structs the stack regularly serializes are arch.SignalInfo +// and arch.UContext64. We'll set the default size as the larger of +// the two, arch.UContext64. +var scratchBufLen = (*UContext64)(nil).SizeBytes() - c := binary.Size(norm) - if c < 0 { - return 0, fmt.Errorf("bad binary.Size for %T", v) - } - n, err := usermem.CopyObjectOut(context.Background(), s.IO, s.Bottom-usermem.Addr(c), norm, usermem.IOOpts{}) - if err != nil || c != n { - return 0, err - } +// CopyScratchBuffer implements marshal.CopyContext.CopyScratchBuffer. +func (s *Stack) CopyScratchBuffer(size int) []byte { + if len(s.scratchBuf) < size { + s.scratchBuf = make([]byte, size) + } + return s.scratchBuf[:size] +} +// StackBottomMagic is the special address callers must past to all stack +// marshalling operations to cause the src/dst address to be computed based on +// the current end of the stack. +const StackBottomMagic = ^usermem.Addr(0) // usermem.Addr(-1) + +// CopyOutBytes implements marshal.CopyContext.CopyOutBytes. CopyOutBytes +// computes an appropriate address based on the current end of the +// stack. Callers use the sentinel address StackBottomMagic to marshal methods +// to indicate this. +func (s *Stack) CopyOutBytes(sentinel usermem.Addr, b []byte) (int, error) { + if sentinel != StackBottomMagic { + panic("Attempted to copy out to stack with absolute address") + } + c := len(b) + n, err := s.IO.CopyOut(context.Background(), s.Bottom-usermem.Addr(c), b, usermem.IOOpts{}) + if err == nil && n == c { s.Bottom -= usermem.Addr(n) } - - return s.Bottom, nil + return n, err } -// 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)) - n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, value, usermem.IOOpts{}) - *vaddr = usermem.Addr(s.Arch.Value(value)) - } else { - n, err = usermem.CopyObjectIn(context.Background(), s.IO, s.Bottom, v, usermem.IOOpts{}) - } - if err != nil { - return 0, err - } - +// CopyInBytes implements marshal.CopyContext.CopyInBytes. CopyInBytes computes +// an appropriate address based on the current end of the stack. Callers must +// use the sentinel address StackBottomMagic to marshal methods to indicate +// this. +func (s *Stack) CopyInBytes(sentinel usermem.Addr, b []byte) (int, error) { + if sentinel != StackBottomMagic { + panic("Attempted to copy in from stack with absolute address") + } + n, err := s.IO.CopyIn(context.Background(), s.Bottom, b, usermem.IOOpts{}) + if err == nil { s.Bottom += usermem.Addr(n) } - - return s.Bottom, nil + return n, err } // Align aligns the stack to the given offset. @@ -142,6 +99,22 @@ func (s *Stack) Align(offset int) { } } +// PushNullTerminatedByteSlice writes bs to the stack, followed by an extra null +// byte at the end. On error, the contents of the stack and the bottom cursor +// are undefined. +func (s *Stack) PushNullTerminatedByteSlice(bs []byte) (int, error) { + // Note: Stack grows up, so write the terminal null byte first. + nNull, err := primitive.CopyUint8Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + n, err := primitive.CopyByteSliceOut(s, StackBottomMagic, bs) + if err != nil { + return 0, err + } + return n + nNull, nil +} + // StackLayout describes the location of the arguments and environment on the // stack. type StackLayout struct { @@ -177,11 +150,10 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) 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 { + if _, err := s.PushNullTerminatedByteSlice([]byte(env[i])); err != nil { return StackLayout{}, err } - envAddrs[i] = addr + envAddrs[i] = s.Bottom } l.EnvvStart = s.Bottom @@ -189,11 +161,10 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) 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 { + if _, err := s.PushNullTerminatedByteSlice([]byte(args[i])); err != nil { return StackLayout{}, err } - argAddrs[i] = addr + argAddrs[i] = s.Bottom } l.ArgvStart = s.Bottom @@ -222,26 +193,26 @@ func (s *Stack) Load(args []string, env []string, aux Auxv) (StackLayout, error) auxv = append(auxv, usermem.Addr(a.Key), a.Value) } auxv = append(auxv, usermem.Addr(0)) - _, err := s.Push(auxv) + _, err := s.pushAddrSliceAndTerminator(auxv) if err != nil { return StackLayout{}, err } // Push environment. - _, err = s.Push(envAddrs) + _, err = s.pushAddrSliceAndTerminator(envAddrs) if err != nil { return StackLayout{}, err } // Push args. - _, err = s.Push(argAddrs) + _, err = s.pushAddrSliceAndTerminator(argAddrs) if err != nil { return StackLayout{}, err } // Push arg count. - _, err = s.Push(usermem.Addr(len(args))) - if err != nil { + lenP := s.Arch.Native(uintptr(len(args))) + if _, err = lenP.CopyOut(s, StackBottomMagic); err != nil { return StackLayout{}, err } diff --git a/pkg/sentry/arch/stack_unsafe.go b/pkg/sentry/arch/stack_unsafe.go new file mode 100644 index 000000000..a90d297ee --- /dev/null +++ b/pkg/sentry/arch/stack_unsafe.go @@ -0,0 +1,69 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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" + "runtime" + "unsafe" + + "gvisor.dev/gvisor/pkg/marshal/primitive" + "gvisor.dev/gvisor/pkg/usermem" +) + +// pushAddrSliceAndTerminator copies a slices of addresses to the stack, and +// also pushes an extra null address element at the end of the slice. +// +// Internally, we unsafely transmute the slice type from the arch-dependent +// []usermem.Addr type, to a slice of fixed-sized ints so that we can pass it to +// go-marshal. +// +// On error, the contents of the stack and the bottom cursor are undefined. +func (s *Stack) pushAddrSliceAndTerminator(src []usermem.Addr) (int, error) { + // Note: Stack grows upwards, so push the terminator first. + srcHdr := (*reflect.SliceHeader)(unsafe.Pointer(&src)) + switch s.Arch.Width() { + case 8: + nNull, err := primitive.CopyUint64Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + var dst []uint64 + dstHdr := (*reflect.SliceHeader)(unsafe.Pointer(&dst)) + dstHdr.Data = srcHdr.Data + dstHdr.Len = srcHdr.Len + dstHdr.Cap = srcHdr.Cap + n, err := primitive.CopyUint64SliceOut(s, StackBottomMagic, dst) + // Ensures src doesn't get GCed until we're done using it through dst. + runtime.KeepAlive(src) + return n + nNull, err + case 4: + nNull, err := primitive.CopyUint32Out(s, StackBottomMagic, 0) + if err != nil { + return 0, err + } + var dst []uint32 + dstHdr := (*reflect.SliceHeader)(unsafe.Pointer(&dst)) + dstHdr.Data = srcHdr.Data + dstHdr.Len = srcHdr.Len + dstHdr.Cap = srcHdr.Cap + n, err := primitive.CopyUint32SliceOut(s, StackBottomMagic, dst) + // Ensure src doesn't get GCed until we're done using it through dst. + runtime.KeepAlive(src) + return n + nNull, err + default: + panic("Unsupported arch width") + } +} diff --git a/pkg/sentry/contexttest/contexttest.go b/pkg/sentry/contexttest/contexttest.go index 8e5658c7a..dfd195a23 100644 --- a/pkg/sentry/contexttest/contexttest.go +++ b/pkg/sentry/contexttest/contexttest.go @@ -144,27 +144,7 @@ func (t *TestContext) MemoryFile() *pgalloc.MemoryFile { // 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) - } + return auth.ContextWithCredentials(Context(tb), auth.NewRootCredentials(auth.NewRootUserNamespace())) } // WithLimitSet returns a copy of ctx carrying l. diff --git a/pkg/sentry/control/BUILD b/pkg/sentry/control/BUILD index 2c5d14be5..deaf5fa23 100644 --- a/pkg/sentry/control/BUILD +++ b/pkg/sentry/control/BUILD @@ -35,7 +35,6 @@ go_library( "//pkg/sync", "//pkg/tcpip/link/sniffer", "//pkg/urpc", - "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/control/proc.go b/pkg/sentry/control/proc.go index dfa936563..1d88db12f 100644 --- a/pkg/sentry/control/proc.go +++ b/pkg/sentry/control/proc.go @@ -23,8 +23,8 @@ import ( "text/tabwriter" "time" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/fd" "gvisor.dev/gvisor/pkg/sentry/fdimport" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/host" @@ -183,9 +183,9 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI if initArgs.MountNamespaceVFS2 == nil { // Set initArgs so that 'ctx' returns the namespace. // - // MountNamespaceVFS2 adds a reference to the namespace, which is - // transferred to the new process. + // Add a reference to the namespace, which is transferred to the new process. initArgs.MountNamespaceVFS2 = proc.Kernel.GlobalInit().Leader().MountNamespaceVFS2() + initArgs.MountNamespaceVFS2.IncRef() } } else { if initArgs.MountNamespace == nil { @@ -203,27 +203,17 @@ func (proc *Proc) execAsync(args *ExecArgs) (*kernel.ThreadGroup, kernel.ThreadI } initArgs.Filename = resolved - fds := make([]int, len(args.FilePayload.Files)) - for i, file := range args.FilePayload.Files { - if kernel.VFS2Enabled { - // Need to dup to remove ownership from os.File. - dup, err := unix.Dup(int(file.Fd())) - if err != nil { - return nil, 0, nil, nil, fmt.Errorf("duplicating payload files: %w", err) - } - fds[i] = dup - } else { - // VFS1 dups the file on import. - fds[i] = int(file.Fd()) - } + fds, err := fd.NewFromFiles(args.Files) + if err != nil { + return nil, 0, nil, nil, fmt.Errorf("duplicating payload files: %w", err) } + defer func() { + for _, fd := range fds { + _ = fd.Close() + } + }() ttyFile, ttyFileVFS2, err := fdimport.Import(ctx, fdTable, args.StdioIsPty, fds) if err != nil { - if kernel.VFS2Enabled { - for _, fd := range fds { - unix.Close(fd) - } - } return nil, 0, nil, nil, err } diff --git a/pkg/sentry/control/state.go b/pkg/sentry/control/state.go index 41feeffe3..d800f2c85 100644 --- a/pkg/sentry/control/state.go +++ b/pkg/sentry/control/state.go @@ -69,5 +69,5 @@ func (s *State) Save(o *SaveOpts, _ *struct{}) error { s.Kernel.Kill(kernel.ExitStatus{}) }, } - return saveOpts.Save(s.Kernel, s.Watchdog) + return saveOpts.Save(s.Kernel.SupervisorContext(), s.Kernel, s.Watchdog) } diff --git a/pkg/sentry/devices/memdev/BUILD b/pkg/sentry/devices/memdev/BUILD index abe58f818..4c8604d58 100644 --- a/pkg/sentry/devices/memdev/BUILD +++ b/pkg/sentry/devices/memdev/BUILD @@ -18,9 +18,10 @@ go_library( "//pkg/rand", "//pkg/safemem", "//pkg/sentry/fsimpl/devtmpfs", + "//pkg/sentry/fsimpl/tmpfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", - "//pkg/sentry/mm", - "//pkg/sentry/pgalloc", "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/usermem", diff --git a/pkg/sentry/devices/memdev/full.go b/pkg/sentry/devices/memdev/full.go index 511179e31..fece3e762 100644 --- a/pkg/sentry/devices/memdev/full.go +++ b/pkg/sentry/devices/memdev/full.go @@ -24,6 +24,8 @@ import ( const fullDevMinor = 7 // fullDevice implements vfs.Device for /dev/full. +// +// +stateify savable type fullDevice struct{} // Open implements vfs.Device.Open. @@ -38,6 +40,8 @@ func (fullDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // fullFD implements vfs.FileDescriptionImpl for /dev/full. +// +// +stateify savable type fullFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/null.go b/pkg/sentry/devices/memdev/null.go index 4918dbeeb..ff5837747 100644 --- a/pkg/sentry/devices/memdev/null.go +++ b/pkg/sentry/devices/memdev/null.go @@ -25,6 +25,8 @@ import ( const nullDevMinor = 3 // nullDevice implements vfs.Device for /dev/null. +// +// +stateify savable type nullDevice struct{} // Open implements vfs.Device.Open. @@ -39,6 +41,8 @@ func (nullDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // nullFD implements vfs.FileDescriptionImpl for /dev/null. +// +// +stateify savable type nullFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/random.go b/pkg/sentry/devices/memdev/random.go index 5e7fe0280..ac943e3ba 100644 --- a/pkg/sentry/devices/memdev/random.go +++ b/pkg/sentry/devices/memdev/random.go @@ -30,6 +30,8 @@ const ( ) // randomDevice implements vfs.Device for /dev/random and /dev/urandom. +// +// +stateify savable type randomDevice struct{} // Open implements vfs.Device.Open. @@ -44,6 +46,8 @@ func (randomDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, } // randomFD implements vfs.FileDescriptionImpl for /dev/random. +// +// +stateify savable type randomFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl diff --git a/pkg/sentry/devices/memdev/zero.go b/pkg/sentry/devices/memdev/zero.go index 2e631a252..1929e41cd 100644 --- a/pkg/sentry/devices/memdev/zero.go +++ b/pkg/sentry/devices/memdev/zero.go @@ -16,9 +16,10 @@ package memdev import ( "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/mm" - "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/usermem" ) @@ -26,6 +27,8 @@ import ( const zeroDevMinor = 5 // zeroDevice implements vfs.Device for /dev/zero. +// +// +stateify savable type zeroDevice struct{} // Open implements vfs.Device.Open. @@ -40,6 +43,8 @@ func (zeroDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // zeroFD implements vfs.FileDescriptionImpl for /dev/zero. +// +// +stateify savable type zeroFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -79,11 +84,22 @@ func (fd *zeroFD) Seek(ctx context.Context, offset int64, whence int32) (int64, // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. func (fd *zeroFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { - m, err := mm.NewSharedAnonMappable(opts.Length, pgalloc.MemoryFileProviderFromContext(ctx)) + if opts.Private || !opts.MaxPerms.Write { + // This mapping will never permit writing to the "underlying file" (in + // Linux terms, it isn't VM_SHARED), so implement it as an anonymous + // mapping, but back it with fd; this is what Linux does, and is + // actually application-visible because the resulting VMA will show up + // in /proc/[pid]/maps with fd.vfsfd.VirtualDentry()'s path rather than + // "/dev/zero (deleted)". + opts.Offset = 0 + opts.MappingIdentity = &fd.vfsfd + opts.MappingIdentity.IncRef() + return nil + } + tmpfsFD, err := tmpfs.NewZeroFile(ctx, auth.CredentialsFromContext(ctx), kernel.KernelFromContext(ctx).ShmMount(), opts.Length) if err != nil { return err } - opts.MappingIdentity = m - opts.Mappable = m - return nil + defer tmpfsFD.DecRef(ctx) + return tmpfsFD.ConfigureMMap(ctx, opts) } diff --git a/pkg/sentry/devices/ttydev/ttydev.go b/pkg/sentry/devices/ttydev/ttydev.go index 664e54498..a287c65ca 100644 --- a/pkg/sentry/devices/ttydev/ttydev.go +++ b/pkg/sentry/devices/ttydev/ttydev.go @@ -30,6 +30,8 @@ const ( ) // ttyDevice implements vfs.Device for /dev/tty. +// +// +stateify savable type ttyDevice struct{} // Open implements vfs.Device.Open. diff --git a/pkg/sentry/devices/tundev/BUILD b/pkg/sentry/devices/tundev/BUILD index 71c59287c..14a8bf9cd 100644 --- a/pkg/sentry/devices/tundev/BUILD +++ b/pkg/sentry/devices/tundev/BUILD @@ -17,6 +17,7 @@ go_library( "//pkg/sentry/vfs", "//pkg/syserror", "//pkg/tcpip/link/tun", + "//pkg/tcpip/network/arp", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/devices/tundev/tundev.go b/pkg/sentry/devices/tundev/tundev.go index a40625e19..ff5d49fbd 100644 --- a/pkg/sentry/devices/tundev/tundev.go +++ b/pkg/sentry/devices/tundev/tundev.go @@ -16,6 +16,8 @@ package tundev import ( + "fmt" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -26,6 +28,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip/link/tun" + "gvisor.dev/gvisor/pkg/tcpip/network/arp" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -36,6 +39,8 @@ const ( ) // tunDevice implements vfs.Device for /dev/net/tun. +// +// +stateify savable type tunDevice struct{} // Open implements vfs.Device.Open. @@ -50,6 +55,8 @@ func (tunDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, opt } // tunFD implements vfs.FileDescriptionImpl for /dev/net/tun. +// +// +stateify savable type tunFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -64,12 +71,13 @@ func (fd *tunFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArg request := args[1].Uint() data := args[2].Pointer() + t := kernel.TaskFromContext(ctx) + if t == nil { + panic("Ioctl should be called from a task context") + } + switch request { case linux.TUNSETIFF: - t := kernel.TaskFromContext(ctx) - if t == nil { - panic("Ioctl should be called from a task context") - } if !t.HasCapability(linux.CAP_NET_ADMIN) { return 0, syserror.EPERM } @@ -79,13 +87,20 @@ func (fd *tunFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArg } var req linux.IFReq - if _, err := usermem.CopyObjectIn(ctx, uio, data, &req, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := req.CopyIn(t, data); err != nil { return 0, err } flags := usermem.ByteOrder.Uint16(req.Data[:]) - return 0, fd.device.SetIff(stack.Stack, req.Name(), flags) + created, err := fd.device.SetIff(stack.Stack, req.Name(), flags) + if err == nil && created { + // Always start with an ARP address for interfaces so they can handle ARP + // packets. + nicID := fd.device.NICID() + if err := stack.Stack.AddAddress(nicID, arp.ProtocolNumber, arp.ProtocolAddress); err != nil { + panic(fmt.Sprintf("failed to add ARP address after creating new TUN/TAP interface with ID = %d", nicID)) + } + } + return 0, err case linux.TUNGETIFF: var req linux.IFReq @@ -97,9 +112,7 @@ func (fd *tunFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArg flags := fd.device.Flags() | linux.IFF_NOFILTER usermem.ByteOrder.PutUint16(req.Data[:], flags) - _, err := usermem.CopyObjectOut(ctx, uio, data, &req, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := req.CopyOut(t, data) return 0, err default: diff --git a/pkg/sentry/fdimport/BUILD b/pkg/sentry/fdimport/BUILD index 5e41ceb4e..6b4f8b0ed 100644 --- a/pkg/sentry/fdimport/BUILD +++ b/pkg/sentry/fdimport/BUILD @@ -10,6 +10,7 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/context", + "//pkg/fd", "//pkg/sentry/fs", "//pkg/sentry/fs/host", "//pkg/sentry/fsimpl/host", diff --git a/pkg/sentry/fdimport/fdimport.go b/pkg/sentry/fdimport/fdimport.go index 1b7cb94c0..314661475 100644 --- a/pkg/sentry/fdimport/fdimport.go +++ b/pkg/sentry/fdimport/fdimport.go @@ -18,6 +18,7 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fd" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/host" hostvfs2 "gvisor.dev/gvisor/pkg/sentry/fsimpl/host" @@ -27,8 +28,9 @@ import ( // Import imports a slice of FDs into the given FDTable. If console is true, // sets up TTY for the first 3 FDs in the slice representing stdin, stdout, -// stderr. Upon success, Import takes ownership of all FDs. -func Import(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []int) (*host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) { +// stderr. Used FDs are either closed or released. It's safe for the caller to +// close any remaining files upon return. +func Import(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []*fd.FD) (*host.TTYFileOperations, *hostvfs2.TTYFileDescription, error) { if kernel.VFS2Enabled { ttyFile, err := importVFS2(ctx, fdTable, console, fds) return nil, ttyFile, err @@ -37,7 +39,7 @@ func Import(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []in return ttyFile, nil, err } -func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []int) (*host.TTYFileOperations, error) { +func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds []*fd.FD) (*host.TTYFileOperations, error) { var ttyFile *fs.File for appFD, hostFD := range fds { var appFile *fs.File @@ -46,11 +48,12 @@ func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds [] // Import the file as a host TTY file. if ttyFile == nil { var err error - appFile, err = host.ImportFile(ctx, hostFD, true /* isTTY */) + appFile, err = host.ImportFile(ctx, hostFD.FD(), true /* isTTY */) if err != nil { return nil, err } defer appFile.DecRef(ctx) + _ = hostFD.Close() // FD is dup'd i ImportFile. // Remember this in the TTY file, as we will // use it for the other stdio FDs. @@ -65,11 +68,12 @@ func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds [] } else { // Import the file as a regular host file. var err error - appFile, err = host.ImportFile(ctx, hostFD, false /* isTTY */) + appFile, err = host.ImportFile(ctx, hostFD.FD(), false /* isTTY */) if err != nil { return nil, err } defer appFile.DecRef(ctx) + _ = hostFD.Close() // FD is dup'd i ImportFile. } // Add the file to the FD map. @@ -84,7 +88,7 @@ func importFS(ctx context.Context, fdTable *kernel.FDTable, console bool, fds [] return ttyFile.FileOperations.(*host.TTYFileOperations), nil } -func importVFS2(ctx context.Context, fdTable *kernel.FDTable, console bool, stdioFDs []int) (*hostvfs2.TTYFileDescription, error) { +func importVFS2(ctx context.Context, fdTable *kernel.FDTable, console bool, stdioFDs []*fd.FD) (*hostvfs2.TTYFileDescription, error) { k := kernel.KernelFromContext(ctx) if k == nil { return nil, fmt.Errorf("cannot find kernel from context") @@ -98,11 +102,12 @@ func importVFS2(ctx context.Context, fdTable *kernel.FDTable, console bool, stdi // Import the file as a host TTY file. if ttyFile == nil { var err error - appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD, true /* isTTY */) + appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD.FD(), true /* isTTY */) if err != nil { return nil, err } defer appFile.DecRef(ctx) + hostFD.Release() // FD is transfered to host FD. // Remember this in the TTY file, as we will use it for the other stdio // FDs. @@ -115,11 +120,12 @@ func importVFS2(ctx context.Context, fdTable *kernel.FDTable, console bool, stdi } } else { var err error - appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD, false /* isTTY */) + appFile, err = hostvfs2.ImportFD(ctx, k.HostMount(), hostFD.FD(), false /* isTTY */) if err != nil { return nil, err } defer appFile.DecRef(ctx) + hostFD.Release() // FD is transfered to host FD. } if err := fdTable.NewFDAtVFS2(ctx, int32(appFD), appFile, kernel.FDFlags{}); err != nil { diff --git a/pkg/sentry/fs/copy_up.go b/pkg/sentry/fs/copy_up.go index 735452b07..ff2fe6712 100644 --- a/pkg/sentry/fs/copy_up.go +++ b/pkg/sentry/fs/copy_up.go @@ -107,8 +107,7 @@ func copyUp(ctx context.Context, d *Dirent) error { // 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. +// 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 @@ -183,12 +182,12 @@ func doCopyUp(ctx context.Context, d *Dirent) error { // 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. -// - next.Inode.overlay.lower.StableAttr.Type must be RegularFile, Directory, +// * parent.Inode.overlay.upper must be non-nil. +// * next.Inode.overlay.copyMu must be locked writable. +// * next.Inode.overlay.lower must be non-nil. +// * next.Inode.overlay.lower.StableAttr.Type must be RegularFile, Directory, // or Symlink. -// - upper filesystem must support setting file ownership and timestamps. +// * 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) diff --git a/pkg/sentry/fs/dev/BUILD b/pkg/sentry/fs/dev/BUILD index 9379a4d7b..6b7b451b8 100644 --- a/pkg/sentry/fs/dev/BUILD +++ b/pkg/sentry/fs/dev/BUILD @@ -34,6 +34,7 @@ go_library( "//pkg/sentry/socket/netstack", "//pkg/syserror", "//pkg/tcpip/link/tun", + "//pkg/tcpip/network/arp", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/dev/net_tun.go b/pkg/sentry/fs/dev/net_tun.go index ec474e554..19ffdec47 100644 --- a/pkg/sentry/fs/dev/net_tun.go +++ b/pkg/sentry/fs/dev/net_tun.go @@ -15,6 +15,8 @@ package dev import ( + "fmt" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -25,6 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/netstack" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip/link/tun" + "gvisor.dev/gvisor/pkg/tcpip/network/arp" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -60,7 +63,7 @@ func newNetTunDevice(ctx context.Context, owner fs.FileOwner, mode linux.FileMod } // GetFile implements fs.InodeOperations.GetFile. -func (iops *netTunInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { +func (*netTunInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { return fs.NewFile(ctx, d, flags, &netTunFileOperations{}), nil } @@ -80,21 +83,22 @@ type netTunFileOperations struct { var _ fs.FileOperations = (*netTunFileOperations)(nil) // Release implements fs.FileOperations.Release. -func (fops *netTunFileOperations) Release(ctx context.Context) { - fops.device.Release(ctx) +func (n *netTunFileOperations) Release(ctx context.Context) { + n.device.Release(ctx) } // Ioctl implements fs.FileOperations.Ioctl. -func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (n *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { request := args[1].Uint() data := args[2].Pointer() + t := kernel.TaskFromContext(ctx) + if t == nil { + panic("Ioctl should be called from a task context") + } + switch request { case linux.TUNSETIFF: - t := kernel.TaskFromContext(ctx) - if t == nil { - panic("Ioctl should be called from a task context") - } if !t.HasCapability(linux.CAP_NET_ADMIN) { return 0, syserror.EPERM } @@ -104,27 +108,32 @@ func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io u } var req linux.IFReq - if _, err := usermem.CopyObjectIn(ctx, io, data, &req, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := req.CopyIn(t, data); err != nil { return 0, err } flags := usermem.ByteOrder.Uint16(req.Data[:]) - return 0, fops.device.SetIff(stack.Stack, req.Name(), flags) + created, err := n.device.SetIff(stack.Stack, req.Name(), flags) + if err == nil && created { + // Always start with an ARP address for interfaces so they can handle ARP + // packets. + nicID := n.device.NICID() + if err := stack.Stack.AddAddress(nicID, arp.ProtocolNumber, arp.ProtocolAddress); err != nil { + panic(fmt.Sprintf("failed to add ARP address after creating new TUN/TAP interface with ID = %d", nicID)) + } + } + return 0, err case linux.TUNGETIFF: var req linux.IFReq - copy(req.IFName[:], fops.device.Name()) + copy(req.IFName[:], n.device.Name()) // Linux adds IFF_NOFILTER (the same value as IFF_NO_PI unfortunately) when // there is no sk_filter. See __tun_chr_ioctl() in net/drivers/tun.c. - flags := fops.device.Flags() | linux.IFF_NOFILTER + flags := n.device.Flags() | linux.IFF_NOFILTER usermem.ByteOrder.PutUint16(req.Data[:], flags) - _, err := usermem.CopyObjectOut(ctx, io, data, &req, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := req.CopyOut(t, data) return 0, err default: @@ -133,41 +142,41 @@ func (fops *netTunFileOperations) Ioctl(ctx context.Context, file *fs.File, io u } // Write implements fs.FileOperations.Write. -func (fops *netTunFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) { +func (n *netTunFileOperations) Write(ctx context.Context, file *fs.File, src usermem.IOSequence, offset int64) (int64, error) { data := make([]byte, src.NumBytes()) if _, err := src.CopyIn(ctx, data); err != nil { return 0, err } - return fops.device.Write(data) + return n.device.Write(data) } // Read implements fs.FileOperations.Read. -func (fops *netTunFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { - data, err := fops.device.Read() +func (n *netTunFileOperations) Read(ctx context.Context, file *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + data, err := n.device.Read() if err != nil { return 0, err } - n, err := dst.CopyOut(ctx, data) - if n > 0 && n < len(data) { + bytesCopied, err := dst.CopyOut(ctx, data) + if bytesCopied > 0 && bytesCopied < len(data) { // Not an error for partial copying. Packet truncated. err = nil } - return int64(n), err + return int64(bytesCopied), err } // Readiness implements watier.Waitable.Readiness. -func (fops *netTunFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { - return fops.device.Readiness(mask) +func (n *netTunFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { + return n.device.Readiness(mask) } // EventRegister implements watier.Waitable.EventRegister. -func (fops *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { - fops.device.EventRegister(e, mask) +func (n *netTunFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + n.device.EventRegister(e, mask) } // EventUnregister implements watier.Waitable.EventUnregister. -func (fops *netTunFileOperations) EventUnregister(e *waiter.Entry) { - fops.device.EventUnregister(e) +func (n *netTunFileOperations) EventUnregister(e *waiter.Entry) { + n.device.EventUnregister(e) } // isNetTunSupported returns whether /dev/net/tun device is supported for s. diff --git a/pkg/sentry/fs/dirent.go b/pkg/sentry/fs/dirent.go index a2f751068..00c526b03 100644 --- a/pkg/sentry/fs/dirent.go +++ b/pkg/sentry/fs/dirent.go @@ -413,9 +413,9 @@ func (d *Dirent) descendantOf(p *Dirent) bool { // Inode.Lookup, otherwise walk will keep d.mu locked. // // Preconditions: -// - renameMu must be held for reading. -// - d.mu must be held. -// - name must must not contain "/"s. +// * renameMu must be held for reading. +// * 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 @@ -577,9 +577,9 @@ func (d *Dirent) Walk(ctx context.Context, root *Dirent, name string) (*Dirent, // exists returns true if name exists in relation to d. // // Preconditions: -// - renameMu must be held for reading. -// - d.mu must be held. -// - name must must not contain "/"s. +// * renameMu must be held for reading. +// * d.mu must be held. +// * name must must not contain "/"s. func (d *Dirent) exists(ctx context.Context, root *Dirent, name string) bool { child, err := d.walk(ctx, root, name, false /* may unlock */) if err != nil { diff --git a/pkg/sentry/fs/file_operations.go b/pkg/sentry/fs/file_operations.go index 305c0f840..6ec721022 100644 --- a/pkg/sentry/fs/file_operations.go +++ b/pkg/sentry/fs/file_operations.go @@ -159,8 +159,9 @@ type FileOperations interface { // 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. - // Must only be called from a task goroutine. + // Preconditions: + // * The AddressSpace (if any) that io refers to is activated. + // * Must only be called from a task goroutine. Ioctl(ctx context.Context, file *File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) } diff --git a/pkg/sentry/fs/fsutil/file_range_set.go b/pkg/sentry/fs/fsutil/file_range_set.go index bbafebf03..1dc409d38 100644 --- a/pkg/sentry/fs/fsutil/file_range_set.go +++ b/pkg/sentry/fs/fsutil/file_range_set.go @@ -70,7 +70,9 @@ func (seg FileRangeIterator) FileRange() memmap.FileRange { // FileRangeOf returns the FileRange mapped by mr. // -// Preconditions: seg.Range().IsSupersetOf(mr). mr.Length() != 0. +// Preconditions: +// * seg.Range().IsSupersetOf(mr). +// * mr.Length() != 0. func (seg FileRangeIterator) FileRangeOf(mr memmap.MappableRange) memmap.FileRange { frstart := seg.Value() + (mr.Start - seg.Start()) return memmap.FileRange{frstart, frstart + mr.Length()} @@ -82,15 +84,18 @@ func (seg FileRangeIterator) FileRangeOf(mr memmap.MappableRange) memmap.FileRan // 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. +// invalid. fileSize is an upper bound on the file's size; bytes after fileSize +// will be zeroed without calling readAt. // // 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, mf *pgalloc.MemoryFile, kind usage.MemoryKind, readAt func(ctx context.Context, dsts safemem.BlockSeq, offset uint64) (uint64, error)) error { +// 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, fileSize uint64, mf *pgalloc.MemoryFile, 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 { @@ -103,7 +108,21 @@ func (frs *FileRangeSet) Fill(ctx context.Context, required, optional memmap.Map fr, err := mf.AllocateAndFill(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) + n, err := func() (uint64, error) { + off := gr.Start + done + if off >= fileSize { + return 0, io.EOF + } + if off+dsts.NumBytes() > fileSize { + rd := fileSize - off + n, err := readAt(ctx, dsts.TakeFirst64(rd), off) + if n == rd && err == nil { + return n, io.EOF + } + return n, err + } + return readAt(ctx, dsts, off) + }() done += n dsts = dsts.DropFirst64(n) if err != nil { diff --git a/pkg/sentry/fs/fsutil/host_file_mapper.go b/pkg/sentry/fs/fsutil/host_file_mapper.go index ef0113b52..4468f5dd2 100644 --- a/pkg/sentry/fs/fsutil/host_file_mapper.go +++ b/pkg/sentry/fs/fsutil/host_file_mapper.go @@ -70,6 +70,13 @@ func (f *HostFileMapper) Init() { f.mappings = make(map[uint64]mapping) } +// IsInited returns true if f.Init() has been called. This is used when +// restoring a checkpoint that contains a HostFileMapper that may or may not +// have been initialized. +func (f *HostFileMapper) IsInited() bool { + return f.refs != nil +} + // NewHostFileMapper returns an initialized HostFileMapper allocated on the // heap with no references or cached mappings. func NewHostFileMapper() *HostFileMapper { @@ -80,7 +87,9 @@ func NewHostFileMapper() *HostFileMapper { // IncRefOn increments the reference count on all offsets in mr. // -// Preconditions: mr.Length() != 0. mr.Start and mr.End must be page-aligned. +// 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() @@ -97,7 +106,9 @@ func (f *HostFileMapper) IncRefOn(mr memmap.MappableRange) { // DecRefOn decrements the reference count on all offsets in mr. // -// Preconditions: mr.Length() != 0. mr.Start and mr.End must be page-aligned. +// 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() @@ -204,7 +215,9 @@ func (f *HostFileMapper) UnmapAll() { } } -// Preconditions: f.mapsMu must be locked. f.mappings[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 diff --git a/pkg/sentry/fs/fsutil/inode_cached.go b/pkg/sentry/fs/fsutil/inode_cached.go index fe8b0b6ac..82eda3e43 100644 --- a/pkg/sentry/fs/fsutil/inode_cached.go +++ b/pkg/sentry/fs/fsutil/inode_cached.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/fs" - "gvisor.dev/gvisor/pkg/sentry/kernel/time" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -444,7 +443,7 @@ func (c *CachingInodeOperations) TouchAccessTime(ctx context.Context, inode *fs. // time. // // Preconditions: c.attrMu is locked for writing. -func (c *CachingInodeOperations) touchAccessTimeLocked(now time.Time) { +func (c *CachingInodeOperations) touchAccessTimeLocked(now ktime.Time) { c.attr.AccessTime = now c.dirtyAttr.AccessTime = true } @@ -461,7 +460,7 @@ func (c *CachingInodeOperations) TouchModificationAndStatusChangeTime(ctx contex // and status change times in-place to the current time. // // Preconditions: c.attrMu is locked for writing. -func (c *CachingInodeOperations) touchModificationAndStatusChangeTimeLocked(now time.Time) { +func (c *CachingInodeOperations) touchModificationAndStatusChangeTimeLocked(now ktime.Time) { c.attr.ModificationTime = now c.dirtyAttr.ModificationTime = true c.attr.StatusChangeTime = now @@ -480,7 +479,7 @@ func (c *CachingInodeOperations) TouchStatusChangeTime(ctx context.Context) { // in-place to the current time. // // Preconditions: c.attrMu is locked for writing. -func (c *CachingInodeOperations) touchStatusChangeTimeLocked(now time.Time) { +func (c *CachingInodeOperations) touchStatusChangeTimeLocked(now ktime.Time) { c.attr.StatusChangeTime = now c.dirtyAttr.StatusChangeTime = true } @@ -645,7 +644,7 @@ func (rw *inodeReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { End: fs.OffsetPageEnd(int64(gapMR.End)), } optMR := gap.Range() - err := rw.c.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mem, usage.PageCache, rw.c.backingFile.ReadToBlocksAt) + err := rw.c.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), uint64(rw.c.attr.Size), mem, usage.PageCache, rw.c.backingFile.ReadToBlocksAt) mem.MarkEvictable(rw.c, pgalloc.EvictableRange{optMR.Start, optMR.End}) seg, gap = rw.c.cache.Find(uint64(rw.offset)) if !seg.Ok() { @@ -672,9 +671,6 @@ func (rw *inodeReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { // Continue. seg, gap = gap.NextSegment(), FileRangeGapIterator{} } - - default: - break } } unlock() @@ -684,7 +680,9 @@ func (rw *inodeReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { // maybeGrowFile grows the file's size if data has been written past the old // size. // -// Preconditions: rw.c.attrMu and rw.c.dataMu bust be locked. +// Preconditions: +// * rw.c.attrMu must be locked. +// * rw.c.dataMu must be locked. func (rw *inodeReadWriter) maybeGrowFile() { // If the write ends beyond the file's previous size, it causes the // file to grow. @@ -766,9 +764,6 @@ func (rw *inodeReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error // Continue. seg, gap = gap.NextSegment(), FileRangeGapIterator{} - - default: - break } } rw.maybeGrowFile() @@ -875,7 +870,7 @@ func (c *CachingInodeOperations) Translate(ctx context.Context, required, option } mf := c.mfp.MemoryFile() - cerr := c.cache.Fill(ctx, required, maxFillRange(required, optional), mf, usage.PageCache, c.backingFile.ReadToBlocksAt) + cerr := c.cache.Fill(ctx, required, maxFillRange(required, optional), uint64(c.attr.Size), mf, usage.PageCache, c.backingFile.ReadToBlocksAt) var ts []memmap.Translation var translatedEnd uint64 diff --git a/pkg/sentry/fs/g3doc/fuse.md b/pkg/sentry/fs/g3doc/fuse.md index 2ca84dd74..05e043583 100644 --- a/pkg/sentry/fs/g3doc/fuse.md +++ b/pkg/sentry/fs/g3doc/fuse.md @@ -79,7 +79,7 @@ ops can be implemented in parallel. - Implement `/dev/fuse` - a character device used to establish an FD for communication between the sentry and the server daemon. -- Implement basic FUSE ops like `FUSE_INIT`, `FUSE_DESTROY`. +- Implement basic FUSE ops like `FUSE_INIT`. #### Read-only mount with basic file operations @@ -95,6 +95,103 @@ ops can be implemented in parallel. - Implement the remaining FUSE ops and decide if we can omit rarely used operations like ioctl. +### Design Details + +#### Lifecycle for a FUSE Request + +- User invokes a syscall +- Sentry prepares corresponding request + - If FUSE device is available + - Write the request in binary + - If FUSE device is full + - Kernel task blocked until available +- Sentry notifies the readers of fuse device that it's ready for read +- FUSE daemon reads the request and processes it +- Sentry waits until a reply is written to the FUSE device + - but returns directly for async requests +- FUSE daemon writes to the fuse device +- Sentry processes the reply + - For sync requests, unblock blocked kernel task + - For async requests, execute pre-specified callback if any +- Sentry returns the syscall to the user + +#### Channels and Queues for Requests in Different Stages + +`connection.initializedChan` + +- a channel that the requests issued before connection initialization blocks + on. + +`fd.queue` + +- a queue of requests that haven’t been read by the FUSE daemon yet. + +`fd.completions` + +- a map of the requests that have been prepared but not yet received a + response, including the ones on the `fd.queue`. + +`fd.waitQueue` + +- a queue of waiters that is waiting for the fuse device fd to be available, + such as the FUSE daemon. + +`fd.fullQueueCh` + +- a channel that the kernel task will be blocked on when the fd is not + available. + +#### Basic I/O Implementation + +Currently we have implemented basic functionalities of read and write for our +FUSE. We describe the design and ways to improve it here: + +##### Basic FUSE Read + +The vfs2 expects implementations of `vfs.FileDescriptionImpl.Read()` and +`vfs.FileDescriptionImpl.PRead()`. When a syscall is made, it will eventually +reach our implementation of those interface functions located at +`pkg/sentry/fsimpl/fuse/regular_file.go` for regular files. + +After validation checks of the input, sentry sends `FUSE_READ` requests to the +FUSE daemon. The FUSE daemon returns data after the `fuse_out_header` as the +responses. For the first version, we create a copy in kernel memory of those +data. They are represented as a byte slice in the marshalled struct. This +happens as a common process for all the FUSE responses at this moment at +`pkg/sentry/fsimpl/fuse/dev.go:writeLocked()`. We then directly copy from this +intermediate buffer to the input buffer provided by the read syscall. + +There is an extra requirement for FUSE: When mounting the FUSE fs, the mounter +or the FUSE daemon can specify a `max_read` or a `max_pages` parameter. They are +the upperbound of the bytes to read in each `FUSE_READ` request. We implemented +the code to handle the fragmented reads. + +To improve the performance: ideally we should have buffer cache to copy those +data from the responses of FUSE daemon into, as is also the design of several +other existing file system implementations for sentry, instead of a single-use +temporary buffer. Directly mapping the memory of one process to another could +also boost the performance, but to keep them isolated, we did not choose to do +so. + +##### Basic FUSE Write + +The vfs2 invokes implementations of `vfs.FileDescriptionImpl.Write()` and +`vfs.FileDescriptionImpl.PWrite()` on the regular file descriptor of FUSE when a +user makes write(2) and pwrite(2) syscall. + +For valid writes, sentry sends the bytes to write after a `FUSE_WRITE` header +(can be regarded as a request with 2 payloads) to the FUSE daemon. For the first +version, we allocate a buffer inside kernel memory to store the bytes from the +user, and copy directly from that buffer to the memory of FUSE daemon. This +happens at `pkg/sentry/fsimpl/fuse/dev.go:readLocked()` + +The parameters `max_write` and `max_pages` restrict the number of bytes in one +`FUSE_WRITE`. There are code handling fragmented writes in current +implementation. + +To have better performance: the extra copy created to store the bytes to write +can be replaced by the buffer cache as well. + # Appendix ## FUSE Protocol diff --git a/pkg/sentry/fs/host/BUILD b/pkg/sentry/fs/host/BUILD index d41d23a43..1368014c4 100644 --- a/pkg/sentry/fs/host/BUILD +++ b/pkg/sentry/fs/host/BUILD @@ -32,6 +32,7 @@ go_library( "//pkg/fdnotifier", "//pkg/iovec", "//pkg/log", + "//pkg/marshal/primitive", "//pkg/refs", "//pkg/safemem", "//pkg/secio", diff --git a/pkg/sentry/fs/host/socket_unsafe.go b/pkg/sentry/fs/host/socket_unsafe.go index 5d4f312cf..c8231e0aa 100644 --- a/pkg/sentry/fs/host/socket_unsafe.go +++ b/pkg/sentry/fs/host/socket_unsafe.go @@ -65,10 +65,10 @@ func fdReadVec(fd int, bufs [][]byte, control []byte, peek bool, maxlen int64) ( controlTrunc = msg.Flags&syscall.MSG_CTRUNC == syscall.MSG_CTRUNC if n > length { - return length, n, msg.Controllen, controlTrunc, err + return length, n, msg.Controllen, controlTrunc, nil } - return n, n, msg.Controllen, controlTrunc, err + return n, n, msg.Controllen, controlTrunc, nil } // fdWriteVec sends from bufs to fd. diff --git a/pkg/sentry/fs/host/tty.go b/pkg/sentry/fs/host/tty.go index e29ae00f2..1183727ab 100644 --- a/pkg/sentry/fs/host/tty.go +++ b/pkg/sentry/fs/host/tty.go @@ -17,6 +17,7 @@ package host import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -53,7 +54,7 @@ type TTYFileOperations struct { func newTTYFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags, iops *inodeOperations) *fs.File { return fs.NewFile(ctx, dirent, flags, &TTYFileOperations{ fileOperations: fileOperations{iops: iops}, - termios: linux.DefaultSlaveTermios, + termios: linux.DefaultReplicaTermios, }) } @@ -123,6 +124,11 @@ func (t *TTYFileOperations) Release(ctx context.Context) { // Ioctl implements fs.FileOperations.Ioctl. func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + return 0, syserror.ENOTTY + } + // Ignore arg[0]. This is the real FD: fd := t.fileOperations.iops.fileState.FD() ioctl := args[1].Uint64() @@ -132,9 +138,7 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO if err != nil { return 0, err } - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), termios, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err = termios.CopyOut(task, args[2].Pointer()) return 0, err case linux.TCSETS, linux.TCSETSW, linux.TCSETSF: @@ -146,9 +150,7 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO } var termios linux.Termios - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &termios, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := termios.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } err := ioctlSetTermios(fd, ioctl, &termios) @@ -173,10 +175,8 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO // Map the ProcessGroup into a ProcessGroupID in the task's PID // namespace. - pgID := pidns.IDOfProcessGroup(t.fgProcessGroup) - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ - AddressSpaceActive: true, - }) + pgID := primitive.Int32(pidns.IDOfProcessGroup(t.fgProcessGroup)) + _, err := pgID.CopyOut(task, args[2].Pointer()) return 0, err case linux.TIOCSPGRP: @@ -184,11 +184,6 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO // Equivalent to tcsetpgrp(fd, *argp). // Set the foreground process group ID of this terminal. - task := kernel.TaskFromContext(ctx) - if task == nil { - return 0, syserror.ENOTTY - } - t.mu.Lock() defer t.mu.Unlock() @@ -208,12 +203,11 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO return 0, syserror.ENOTTY } - var pgID kernel.ProcessGroupID - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + var pgIDP primitive.Int32 + if _, err := pgIDP.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } + pgID := kernel.ProcessGroupID(pgIDP) // pgID must be non-negative. if pgID < 0 { @@ -242,9 +236,7 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO if err != nil { return 0, err } - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), winsize, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err = winsize.CopyOut(task, args[2].Pointer()) return 0, err case linux.TIOCSWINSZ: @@ -255,9 +247,7 @@ func (t *TTYFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO // background ones) can set the winsize. var winsize linux.Winsize - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &winsize, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := winsize.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } err := ioctlSetWinsize(fd, &winsize) diff --git a/pkg/sentry/fs/inode.go b/pkg/sentry/fs/inode.go index b79cd9877..004910453 100644 --- a/pkg/sentry/fs/inode.go +++ b/pkg/sentry/fs/inode.go @@ -270,7 +270,7 @@ func (i *Inode) GetXattr(ctx context.Context, name string, size uint64) (string, // SetXattr calls i.InodeOperations.SetXattr with i as the Inode. func (i *Inode) SetXattr(ctx context.Context, d *Dirent, name, value string, flags uint32) error { if i.overlay != nil { - return overlaySetxattr(ctx, i.overlay, d, name, value, flags) + return overlaySetXattr(ctx, i.overlay, d, name, value, flags) } return i.InodeOperations.SetXattr(ctx, i, name, value, flags) } diff --git a/pkg/sentry/fs/inode_overlay.go b/pkg/sentry/fs/inode_overlay.go index dc2e353d9..b16ab08ba 100644 --- a/pkg/sentry/fs/inode_overlay.go +++ b/pkg/sentry/fs/inode_overlay.go @@ -16,7 +16,6 @@ package fs import ( "fmt" - "strings" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -539,7 +538,7 @@ func overlayGetXattr(ctx context.Context, o *overlayEntry, name string, size uin // 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) { + if isXattrOverlay(name) { return "", syserror.ENODATA } @@ -553,9 +552,9 @@ func overlayGetXattr(ctx context.Context, o *overlayEntry, name string, size uin return s, err } -func overlaySetxattr(ctx context.Context, o *overlayEntry, d *Dirent, name, value string, flags uint32) error { +func overlaySetXattr(ctx context.Context, o *overlayEntry, d *Dirent, name, value string, flags uint32) error { // Don't allow changes to overlay xattrs through a setxattr syscall. - if strings.HasPrefix(XattrOverlayPrefix, name) { + if isXattrOverlay(name) { return syserror.EPERM } @@ -578,7 +577,7 @@ func overlayListXattr(ctx context.Context, o *overlayEntry, size uint64) (map[st for name := range names { // Same as overlayGetXattr, we shouldn't forward along // overlay attributes. - if strings.HasPrefix(XattrOverlayPrefix, name) { + if isXattrOverlay(name) { delete(names, name) } } @@ -587,7 +586,7 @@ func overlayListXattr(ctx context.Context, o *overlayEntry, size uint64) (map[st func overlayRemoveXattr(ctx context.Context, o *overlayEntry, d *Dirent, name string) error { // Don't allow changes to overlay xattrs through a removexattr syscall. - if strings.HasPrefix(XattrOverlayPrefix, name) { + if isXattrOverlay(name) { return syserror.EPERM } diff --git a/pkg/sentry/fs/overlay.go b/pkg/sentry/fs/overlay.go index 35013a21b..01a1235b8 100644 --- a/pkg/sentry/fs/overlay.go +++ b/pkg/sentry/fs/overlay.go @@ -86,13 +86,12 @@ func isXattrOverlay(name string) bool { // NewOverlayRoot produces the root of an overlay. // // Preconditions: -// -// - upper and lower must be non-nil. -// - upper must not be an overlay. -// - lower should not expose character devices, pipes, or sockets, because +// * upper and lower must be non-nil. +// * upper must not be an overlay. +// * lower should not expose character devices, pipes, or sockets, because // copying up these types of files is not supported. -// - lower must not require that file objects be revalidated. -// - lower must not have dynamic file/directory content. +// * lower must not require that file objects be revalidated. +// * 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 a %v, not a directory", upper.StableAttr.Type) @@ -117,12 +116,11 @@ func NewOverlayRoot(ctx context.Context, upper *Inode, lower *Inode, flags Mount // NewOverlayRootFile produces the root of an overlay that points to a file. // // Preconditions: -// -// - lower must be non-nil. -// - lower should not expose character devices, pipes, or sockets, because +// * lower must be non-nil. +// * lower should not expose character devices, pipes, or sockets, because // copying up these types of files is not supported. Neither it can be a dir. -// - lower must not require that file objects be revalidated. -// - lower must not have dynamic file/directory content. +// * lower must not require that file objects be revalidated. +// * lower must not have dynamic file/directory content. func NewOverlayRootFile(ctx context.Context, upperMS *MountSource, lower *Inode, flags MountSourceFlags) (*Inode, error) { if !IsRegular(lower.StableAttr) { return nil, fmt.Errorf("lower Inode is not a regular file") diff --git a/pkg/sentry/fs/proc/BUILD b/pkg/sentry/fs/proc/BUILD index 77c2c5c0e..b8b2281a8 100644 --- a/pkg/sentry/fs/proc/BUILD +++ b/pkg/sentry/fs/proc/BUILD @@ -50,6 +50,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/tcpip/header", + "//pkg/tcpip/network/ipv4", "//pkg/usermem", "//pkg/waiter", ], diff --git a/pkg/sentry/fs/proc/sys_net.go b/pkg/sentry/fs/proc/sys_net.go index 8615b60f0..e555672ad 100644 --- a/pkg/sentry/fs/proc/sys_net.go +++ b/pkg/sentry/fs/proc/sys_net.go @@ -26,6 +26,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/fs/ramfs" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -54,7 +55,7 @@ type tcpMemInode struct { // size stores the tcp buffer size during save, and sets the buffer // size in netstack in restore. We must save/restore this here, since - // netstack itself is stateless. + // a netstack instance is created on restore. size inet.TCPBufferSize // mu protects against concurrent reads/writes to files based on this @@ -258,6 +259,9 @@ func (f *tcpSackFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSeque if src.NumBytes() == 0 { return 0, nil } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. src = src.TakeFirst(usermem.PageSize - 1) var v int32 @@ -383,11 +387,125 @@ func (p *proc) newSysNetCore(ctx context.Context, msrc *fs.MountSource, s inet.S return newProcInode(ctx, d, msrc, fs.SpecialDirectory, nil) } +// ipForwarding implements fs.InodeOperations. +// +// ipForwarding is used to enable/disable packet forwarding of netstack. +// +// +stateify savable +type ipForwarding struct { + fsutil.SimpleFileInode + + stack inet.Stack `state:"wait"` + + // enabled stores the IPv4 forwarding state on save. + // We must save/restore this here, since a netstack instance + // is created on restore. + enabled *bool +} + +func newIPForwardingInode(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { + ipf := &ipForwarding{ + SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, fs.RootOwner, fs.FilePermsFromMode(0444), linux.PROC_SUPER_MAGIC), + stack: s, + } + sattr := fs.StableAttr{ + DeviceID: device.ProcDevice.DeviceID(), + InodeID: device.ProcDevice.NextIno(), + BlockSize: usermem.PageSize, + Type: fs.SpecialFile, + } + return fs.NewInode(ctx, ipf, msrc, sattr) +} + +// Truncate implements fs.InodeOperations.Truncate. Truncate is called when +// O_TRUNC is specified for any kind of existing Dirent but is not called via +// (f)truncate for proc files. +func (*ipForwarding) Truncate(context.Context, *fs.Inode, int64) error { + return nil +} + +// +stateify savable +type ipForwardingFile struct { + fsutil.FileGenericSeek `state:"nosave"` + fsutil.FileNoIoctl `state:"nosave"` + fsutil.FileNoMMap `state:"nosave"` + fsutil.FileNoSplice `state:"nosave"` + fsutil.FileNoopFlush `state:"nosave"` + fsutil.FileNoopFsync `state:"nosave"` + fsutil.FileNoopRelease `state:"nosave"` + fsutil.FileNotDirReaddir `state:"nosave"` + fsutil.FileUseInodeUnstableAttr `state:"nosave"` + waiter.AlwaysReady `state:"nosave"` + + ipf *ipForwarding + + stack inet.Stack `state:"wait"` +} + +// GetFile implements fs.InodeOperations.GetFile. +func (ipf *ipForwarding) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + flags.Pread = true + flags.Pwrite = true + return fs.NewFile(ctx, dirent, flags, &ipForwardingFile{ + stack: ipf.stack, + ipf: ipf, + }), nil +} + +// Read implements fs.FileOperations.Read. +func (f *ipForwardingFile) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + return 0, io.EOF + } + + if f.ipf.enabled == nil { + enabled := f.stack.Forwarding(ipv4.ProtocolNumber) + f.ipf.enabled = &enabled + } + + val := "0\n" + if *f.ipf.enabled { + // Technically, this is not quite compatible with Linux. Linux + // stores these as an integer, so if you write "2" into + // ip_forward, you should get 2 back. + val = "1\n" + } + n, err := dst.CopyOut(ctx, []byte(val)) + return int64(n), err +} + +// Write implements fs.FileOperations.Write. +// +// Offset is ignored, multiple writes are not supported. +func (f *ipForwardingFile) Write(ctx context.Context, _ *fs.File, src usermem.IOSequence, offset int64) (int64, error) { + if src.NumBytes() == 0 { + return 0, nil + } + + // Only consider size of one memory page for input for performance reasons. + // We are only reading if it's zero or not anyway. + 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 + } + if f.ipf.enabled == nil { + f.ipf.enabled = new(bool) + } + *f.ipf.enabled = v != 0 + return n, f.stack.SetForwarding(ipv4.ProtocolNumber, *f.ipf.enabled) +} + func (p *proc) newSysNetIPv4Dir(ctx context.Context, msrc *fs.MountSource, s inet.Stack) *fs.Inode { contents := map[string]*fs.Inode{ // Add tcp_sack. "tcp_sack": newTCPSackInode(ctx, msrc, s), + // Add ip_forward. + "ip_forward": newIPForwardingInode(ctx, msrc, s), + // The following files are simple stubs until they are // implemented in netstack, most of these files are // configuration related. We use the value closest to the diff --git a/pkg/sentry/fs/proc/sys_net_state.go b/pkg/sentry/fs/proc/sys_net_state.go index 6eba709c6..4cb4741af 100644 --- a/pkg/sentry/fs/proc/sys_net_state.go +++ b/pkg/sentry/fs/proc/sys_net_state.go @@ -14,7 +14,11 @@ package proc -import "fmt" +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" +) // beforeSave is invoked by stateify. func (t *tcpMemInode) beforeSave() { @@ -40,3 +44,12 @@ func (s *tcpSack) afterLoad() { } } } + +// afterLoad is invoked by stateify. +func (ipf *ipForwarding) afterLoad() { + if ipf.enabled != nil { + if err := ipf.stack.SetForwarding(ipv4.ProtocolNumber, *ipf.enabled); err != nil { + panic(fmt.Sprintf("failed to set IPv4 forwarding [%v]: %v", *ipf.enabled, err)) + } + } +} diff --git a/pkg/sentry/fs/proc/sys_net_test.go b/pkg/sentry/fs/proc/sys_net_test.go index 355e83d47..6ef5738e7 100644 --- a/pkg/sentry/fs/proc/sys_net_test.go +++ b/pkg/sentry/fs/proc/sys_net_test.go @@ -123,3 +123,76 @@ func TestConfigureRecvBufferSize(t *testing.T) { } } } + +// TestIPForwarding tests the implementation of +// /proc/sys/net/ipv4/ip_forwarding +func TestIPForwarding(t *testing.T) { + ctx := context.Background() + s := inet.NewTestStack() + + var cases = []struct { + comment string + initial bool + str string + final bool + }{ + { + comment: `Forwarding is disabled; write 1 and enable forwarding`, + initial: false, + str: "1", + final: true, + }, + { + comment: `Forwarding is disabled; write 0 and disable forwarding`, + initial: false, + str: "0", + final: false, + }, + { + comment: `Forwarding is enabled; write 1 and enable forwarding`, + initial: true, + str: "1", + final: true, + }, + { + comment: `Forwarding is enabled; write 0 and disable forwarding`, + initial: true, + str: "0", + final: false, + }, + { + comment: `Forwarding is disabled; write 2404 and enable forwarding`, + initial: false, + str: "2404", + final: true, + }, + { + comment: `Forwarding is enabled; write 2404 and enable forwarding`, + initial: true, + str: "2404", + final: true, + }, + } + for _, c := range cases { + t.Run(c.comment, func(t *testing.T) { + s.IPForwarding = c.initial + ipf := &ipForwarding{stack: s} + file := &ipForwardingFile{ + stack: s, + ipf: ipf, + } + + // Write the values. + src := usermem.BytesIOSequence([]byte(c.str)) + if n, err := file.Write(ctx, nil, src, 0); n != int64(len(c.str)) || err != nil { + t.Errorf("file.Write(ctx, nil, %q, 0) = (%d, %v); want (%d, nil)", c.str, n, err, len(c.str)) + } + + // Read the values from the stack and check them. + if got, want := s.IPForwarding, c.final; got != want { + t.Errorf("s.IPForwarding incorrect; got: %v, want: %v", got, want) + } + + }) + } +} diff --git a/pkg/sentry/fs/proc/task.go b/pkg/sentry/fs/proc/task.go index 9cf7f2a62..22d658acf 100644 --- a/pkg/sentry/fs/proc/task.go +++ b/pkg/sentry/fs/proc/task.go @@ -84,6 +84,7 @@ func (p *proc) newTaskDir(t *kernel.Task, msrc *fs.MountSource, isThreadGroup bo "auxv": newAuxvec(t, msrc), "cmdline": newExecArgInode(t, msrc, cmdlineExecArg), "comm": newComm(t, msrc), + "cwd": newCwd(t, msrc), "environ": newExecArgInode(t, msrc, environExecArg), "exe": newExe(t, msrc), "fd": newFdDir(t, msrc), @@ -300,6 +301,49 @@ func (e *exe) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { return exec.PathnameWithDeleted(ctx), nil } +// cwd is an fs.InodeOperations symlink for the /proc/PID/cwd file. +// +// +stateify savable +type cwd struct { + ramfs.Symlink + + t *kernel.Task +} + +func newCwd(t *kernel.Task, msrc *fs.MountSource) *fs.Inode { + cwdSymlink := &cwd{ + Symlink: *ramfs.NewSymlink(t, fs.RootOwner, ""), + t: t, + } + return newProcInode(t, cwdSymlink, msrc, fs.Symlink, t) +} + +// Readlink implements fs.InodeOperations. +func (e *cwd) Readlink(ctx context.Context, inode *fs.Inode) (string, error) { + if !kernel.ContextCanTrace(ctx, e.t, false) { + return "", syserror.EACCES + } + if err := checkTaskState(e.t); err != nil { + return "", err + } + cwd := e.t.FSContext().WorkingDirectory() + if cwd == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer cwd.DecRef(ctx) + + root := fs.RootFromContext(ctx) + if root == nil { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + name, _ := cwd.FullName(root) + return name, nil +} + // namespaceSymlink represents a symlink in the namespacefs, such as the files // in /proc/<pid>/ns. // @@ -604,7 +648,7 @@ func (s *statusData) ReadSeqFileData(ctx context.Context, h seqfile.SeqHandle) ( var vss, rss, data uint64 s.t.WithMuLocked(func(t *kernel.Task) { if fdTable := t.FDTable(); fdTable != nil { - fds = fdTable.Size() + fds = fdTable.CurrentMaxFDs() } if mm := t.MemoryManager(); mm != nil { vss = mm.VirtualMemorySize() diff --git a/pkg/sentry/fs/tmpfs/inode_file.go b/pkg/sentry/fs/tmpfs/inode_file.go index 1dc75291d..fc0498f17 100644 --- a/pkg/sentry/fs/tmpfs/inode_file.go +++ b/pkg/sentry/fs/tmpfs/inode_file.go @@ -613,7 +613,7 @@ func (f *fileInodeOperations) Translate(ctx context.Context, required, optional } mf := f.kernel.MemoryFile() - cerr := f.data.Fill(ctx, required, optional, mf, f.memUsage, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { + cerr := f.data.Fill(ctx, required, optional, uint64(f.attr.Size), mf, 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 }) diff --git a/pkg/sentry/fs/tmpfs/tmpfs.go b/pkg/sentry/fs/tmpfs/tmpfs.go index b095312fe..998b697ca 100644 --- a/pkg/sentry/fs/tmpfs/tmpfs.go +++ b/pkg/sentry/fs/tmpfs/tmpfs.go @@ -16,6 +16,8 @@ package tmpfs import ( + "math" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -32,9 +34,15 @@ import ( var fsInfo = fs.Info{ Type: linux.TMPFS_MAGIC, + // tmpfs currently does not support configurable size limits. In Linux, + // such a tmpfs mount will return f_blocks == f_bfree == f_bavail == 0 from + // statfs(2). However, many applications treat this as having a size limit + // of 0. To work around this, claim to have a very large but non-zero size, + // chosen to ensure that BlockSize * Blocks does not overflow int64 (which + // applications may also handle incorrectly). // TODO(b/29637826): allow configuring a tmpfs size and enforce it. - TotalBlocks: 0, - FreeBlocks: 0, + TotalBlocks: math.MaxInt64 / usermem.PageSize, + FreeBlocks: math.MaxInt64 / usermem.PageSize, } // rename implements fs.InodeOperations.Rename for tmpfs nodes. diff --git a/pkg/sentry/fs/tty/BUILD b/pkg/sentry/fs/tty/BUILD index 5cb0e0417..e6d0eb359 100644 --- a/pkg/sentry/fs/tty/BUILD +++ b/pkg/sentry/fs/tty/BUILD @@ -10,13 +10,14 @@ go_library( "line_discipline.go", "master.go", "queue.go", - "slave.go", + "replica.go", "terminal.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/marshal/primitive", "//pkg/refs", "//pkg/safemem", "//pkg/sentry/arch", diff --git a/pkg/sentry/fs/tty/dir.go b/pkg/sentry/fs/tty/dir.go index 463f6189e..c2da80bc2 100644 --- a/pkg/sentry/fs/tty/dir.go +++ b/pkg/sentry/fs/tty/dir.go @@ -37,14 +37,14 @@ import ( // 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 +// the replica Inode in the this directory for discovery via Lookup/Readdir. The +// replica inode is unregistered when the master file is Released, as the replica // is no longer discoverable at that point. // // References on the underlying Terminal are held by masterFileOperations and -// slaveInodeOperations. +// replicaInodeOperations. // -// masterInodeOperations and slaveInodeOperations hold a pointer to +// masterInodeOperations and replicaInodeOperations hold a pointer to // dirInodeOperations, which is reference counted by the refcount their // corresponding Dirents hold on their parent (this directory). // @@ -76,16 +76,16 @@ type dirInodeOperations struct { // master is the master PTY inode. master *fs.Inode - // slaves contains the slave inodes reachable from the directory. + // replicas contains the replica inodes reachable from the directory. // - // A new slave is added by allocateTerminal and is removed by + // A new replica 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 + // A reference is held on every replica in the map. + replicas map[uint32]*fs.Inode // dentryMap is a SortedDentryMap used to implement Readdir containing - // the master and all entries in slaves. + // the master and all entries in replicas. dentryMap *fs.SortedDentryMap // next is the next pty index to use. @@ -101,7 +101,7 @@ func newDir(ctx context.Context, m *fs.MountSource) *fs.Inode { d := &dirInodeOperations{ InodeSimpleAttributes: fsutil.NewInodeSimpleAttributes(ctx, fs.RootOwner, fs.FilePermsFromMode(0555), linux.DEVPTS_SUPER_MAGIC), msrc: m, - slaves: make(map[uint32]*fs.Inode), + replicas: make(map[uint32]*fs.Inode), dentryMap: fs.NewSortedDentryMap(nil), } // Linux devpts uses a default mode of 0000 for ptmx which can be @@ -133,7 +133,7 @@ func (d *dirInodeOperations) Release(ctx context.Context) { defer d.mu.Unlock() d.master.DecRef(ctx) - if len(d.slaves) != 0 { + if len(d.replicas) != 0 { panic(fmt.Sprintf("devpts directory still contains active terminals: %+v", d)) } } @@ -149,14 +149,14 @@ func (d *dirInodeOperations) Lookup(ctx context.Context, dir *fs.Inode, name str return fs.NewDirent(ctx, d.master, name), nil } - // Slave number? + // Replica number? n, err := strconv.ParseUint(name, 10, 32) if err != nil { // Not found. return nil, syserror.ENOENT } - s, ok := d.slaves[uint32(n)] + s, ok := d.replicas[uint32(n)] if !ok { return nil, syserror.ENOENT } @@ -236,7 +236,7 @@ func (d *dirInodeOperations) allocateTerminal(ctx context.Context) (*Terminal, e return nil, syserror.ENOMEM } - if _, ok := d.slaves[n]; ok { + if _, ok := d.replicas[n]; ok { panic(fmt.Sprintf("pty index collision; index %d already exists", n)) } @@ -244,19 +244,19 @@ func (d *dirInodeOperations) allocateTerminal(ctx context.Context) (*Terminal, e d.next++ // The reference returned by newTerminal is returned to the caller. - // Take another for the slave inode. + // Take another for the replica 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)) + replica := newReplicaInode(ctx, d, t, fs.FileOwner{uid, gid}, fs.FilePermsFromMode(0666)) - d.slaves[n] = slave + d.replicas[n] = replica d.dentryMap.Add(strconv.FormatUint(uint64(n), 10), fs.DentAttr{ - Type: slave.StableAttr.Type, - InodeID: slave.StableAttr.InodeID, + Type: replica.StableAttr.Type, + InodeID: replica.StableAttr.InodeID, }) return t, nil @@ -267,18 +267,18 @@ func (d *dirInodeOperations) masterClose(ctx context.Context, 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. + // The replica end disappears from the directory when the master end is + // closed, even if the replica 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] + s, ok := d.replicas[t.n] if !ok { panic(fmt.Sprintf("Terminal %+v doesn't exist in %+v?", t, d)) } s.DecRef(ctx) - delete(d.slaves, t.n) + delete(d.replicas, t.n) d.dentryMap.Remove(strconv.FormatUint(uint64(t.n), 10)) } diff --git a/pkg/sentry/fs/tty/fs.go b/pkg/sentry/fs/tty/fs.go index 2d4d44bf3..13f4901db 100644 --- a/pkg/sentry/fs/tty/fs.go +++ b/pkg/sentry/fs/tty/fs.go @@ -79,8 +79,8 @@ type superOperations struct{} // // 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 +// Replica entries are dropped from dir when their master is closed, so an +// existing replica Dirent in the tree is not sufficient to guarantee that it // still exists on the filesystem. func (superOperations) Revalidate(context.Context, string, *fs.Inode, *fs.Inode) bool { return true diff --git a/pkg/sentry/fs/tty/line_discipline.go b/pkg/sentry/fs/tty/line_discipline.go index 2e9dd2d55..b34f4a0eb 100644 --- a/pkg/sentry/fs/tty/line_discipline.go +++ b/pkg/sentry/fs/tty/line_discipline.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -43,7 +44,7 @@ const ( ) // lineDiscipline dictates how input and output are handled between the -// pseudoterminal (pty) master and slave. It can be configured to alter I/O, +// pseudoterminal (pty) master and replica. 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: // @@ -54,8 +55,8 @@ const ( // // 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 +// terminal master to replica (the input queue) and one from replica 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: // @@ -64,7 +65,7 @@ const ( // | (inputQueueWrite) +-------------+ (inputQueueRead) | // | | // | v -// masterFD slaveFD +// masterFD replicaFD // ^ | // | | // | output to terminal +--------------+ output from process | @@ -103,8 +104,8 @@ type lineDiscipline struct { // masterWaiter is used to wait on the master end of the TTY. masterWaiter waiter.Queue `state:"zerovalue"` - // slaveWaiter is used to wait on the slave end of the TTY. - slaveWaiter waiter.Queue `state:"zerovalue"` + // replicaWaiter is used to wait on the replica end of the TTY. + replicaWaiter waiter.Queue `state:"zerovalue"` } func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { @@ -115,27 +116,23 @@ func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { } // getTermios gets the linux.Termios for the tty. -func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (l *lineDiscipline) getTermios(task *kernel.Task, args arch.SyscallArguments) (uintptr, error) { l.termiosMu.RLock() defer l.termiosMu.RUnlock() // 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, - }) + _, err := t.CopyOut(task, args[2].Pointer()) 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) { +func (l *lineDiscipline) setTermios(task *kernel.Task, args arch.SyscallArguments) (uintptr, error) { l.termiosMu.Lock() defer l.termiosMu.Unlock() oldCanonEnabled := l.termios.LEnabled(linux.ICANON) // 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, - }) + _, err := t.CopyIn(task, args[2].Pointer()) l.termios.FromTermios(t) // If canonical mode is turned off, move bytes from inQueue's wait @@ -146,27 +143,23 @@ func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arc l.inQueue.pushWaitBufLocked(l) l.inQueue.readable = true l.inQueue.mu.Unlock() - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) } return 0, err } -func (l *lineDiscipline) windowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (l *lineDiscipline) windowSize(t *kernel.Task, args arch.SyscallArguments) error { l.sizeMu.Lock() defer l.sizeMu.Unlock() - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), l.size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := l.size.CopyOut(t, args[2].Pointer()) return err } -func (l *lineDiscipline) setWindowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (l *lineDiscipline) setWindowSize(t *kernel.Task, args arch.SyscallArguments) error { l.sizeMu.Lock() defer l.sizeMu.Unlock() - _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &l.size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := l.size.CopyIn(t, args[2].Pointer()) return err } @@ -176,14 +169,14 @@ func (l *lineDiscipline) masterReadiness() waiter.EventMask { return l.inQueue.writeReadiness(&linux.MasterTermios) | l.outQueue.readReadiness(&linux.MasterTermios) } -func (l *lineDiscipline) slaveReadiness() waiter.EventMask { +func (l *lineDiscipline) replicaReadiness() waiter.EventMask { l.termiosMu.RLock() defer l.termiosMu.RUnlock() return l.outQueue.writeReadiness(&l.termios) | l.inQueue.readReadiness(&l.termios) } -func (l *lineDiscipline) inputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { - return l.inQueue.readableSize(ctx, io, args) +func (l *lineDiscipline) inputQueueReadSize(t *kernel.Task, args arch.SyscallArguments) error { + return l.inQueue.readableSize(t, args) } func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { @@ -196,7 +189,7 @@ func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSeque if n > 0 { l.masterWaiter.Notify(waiter.EventOut) if pushed { - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) } return n, nil } @@ -211,14 +204,14 @@ func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequ return 0, err } if n > 0 { - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) return n, nil } return 0, syserror.ErrWouldBlock } -func (l *lineDiscipline) outputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { - return l.outQueue.readableSize(ctx, io, args) +func (l *lineDiscipline) outputQueueReadSize(t *kernel.Task, args arch.SyscallArguments) error { + return l.outQueue.readableSize(t, args) } func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { @@ -229,7 +222,7 @@ func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequ return 0, err } if n > 0 { - l.slaveWaiter.Notify(waiter.EventOut) + l.replicaWaiter.Notify(waiter.EventOut) if pushed { l.masterWaiter.Notify(waiter.EventIn) } diff --git a/pkg/sentry/fs/tty/master.go b/pkg/sentry/fs/tty/master.go index e00746017..b91184b1b 100644 --- a/pkg/sentry/fs/tty/master.go +++ b/pkg/sentry/fs/tty/master.go @@ -17,9 +17,11 @@ package tty import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/unimpl" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -152,46 +154,51 @@ func (mf *masterFileOperations) Write(ctx context.Context, _ *fs.File, src userm // Ioctl implements fs.FileOperations.Ioctl. func (mf *masterFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // ioctl(2) may only be called from a task goroutine. + return 0, syserror.ENOTTY + } + switch cmd := args[1].Uint(); cmd { case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ // Get the number of bytes in the output queue read buffer. - return 0, mf.t.ld.outputQueueReadSize(ctx, io, args) + return 0, mf.t.ld.outputQueueReadSize(t, args) 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) + // of the replica end. + return mf.t.ld.getTermios(t, 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) + // of the replica end. + return mf.t.ld.setTermios(t, args) case linux.TCSETSW: // TODO(b/29356795): This should drain the output queue first. - return mf.t.ld.setTermios(ctx, io, args) + return mf.t.ld.setTermios(t, args) case linux.TIOCGPTN: - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mf.t.n), usermem.IOOpts{ - AddressSpaceActive: true, - }) + nP := primitive.Uint32(mf.t.n) + _, err := nP.CopyOut(t, args[2].Pointer()) return 0, err case linux.TIOCSPTLCK: // TODO(b/29356795): Implement pty locking. For now just pretend we do. return 0, nil case linux.TIOCGWINSZ: - return 0, mf.t.ld.windowSize(ctx, io, args) + return 0, mf.t.ld.windowSize(t, args) case linux.TIOCSWINSZ: - return 0, mf.t.ld.setWindowSize(ctx, io, args) + return 0, mf.t.ld.setWindowSize(t, args) case linux.TIOCSCTTY: // Make the given terminal the controlling terminal of the // calling process. - return 0, mf.t.setControllingTTY(ctx, io, args, true /* isMaster */) + return 0, mf.t.setControllingTTY(ctx, args, true /* isMaster */) case linux.TIOCNOTTY: // Release this process's controlling terminal. - return 0, mf.t.releaseControllingTTY(ctx, io, args, true /* isMaster */) + return 0, mf.t.releaseControllingTTY(ctx, args, true /* isMaster */) case linux.TIOCGPGRP: // Get the foreground process group. - return mf.t.foregroundProcessGroup(ctx, io, args, true /* isMaster */) + return mf.t.foregroundProcessGroup(ctx, args, true /* isMaster */) case linux.TIOCSPGRP: // Set the foreground process group. - return mf.t.setForegroundProcessGroup(ctx, io, args, true /* isMaster */) + return mf.t.setForegroundProcessGroup(ctx, args, true /* isMaster */) default: maybeEmitUnimplementedEvent(ctx, cmd) return 0, syserror.ENOTTY diff --git a/pkg/sentry/fs/tty/queue.go b/pkg/sentry/fs/tty/queue.go index ceabb9b1e..79975d812 100644 --- a/pkg/sentry/fs/tty/queue.go +++ b/pkg/sentry/fs/tty/queue.go @@ -17,8 +17,10 @@ package tty import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -32,7 +34,7 @@ import ( const waitBufMaxBytes = 131072 // queue represents one of the input or output queues between a pty master and -// slave. Bytes written to a queue are added to the read buffer until it is +// replica. Bytes written to a queue are added to the read buffer until it is // full, at which point they are written to the wait buffer. Bytes are // processed (i.e. undergo termios transformations) as they are added to the // read buffer. The read buffer is readable when its length is nonzero and @@ -85,17 +87,15 @@ func (q *queue) writeReadiness(t *linux.KernelTermios) waiter.EventMask { } // readableSize writes the number of readable bytes to userspace. -func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (q *queue) readableSize(t *kernel.Task, args arch.SyscallArguments) error { q.mu.Lock() defer q.mu.Unlock() - var size int32 + size := primitive.Int32(0) if q.readable { - size = int32(len(q.readBuf)) + size = primitive.Int32(len(q.readBuf)) } - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := size.CopyOut(t, args[2].Pointer()) return err } @@ -104,8 +104,7 @@ func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.Sysca // as whether the read caused more readable data to become available (whether // data was pushed from the wait buffer to the read buffer). // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipline) (int64, bool, error) { q.mu.Lock() defer q.mu.Unlock() @@ -145,8 +144,7 @@ func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipl // write writes to q from userspace. // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscipline) (int64, error) { q.mu.Lock() defer q.mu.Unlock() @@ -188,8 +186,7 @@ func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscip // writeBytes writes to q from b. // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) writeBytes(b []byte, l *lineDiscipline) { q.mu.Lock() defer q.mu.Unlock() diff --git a/pkg/sentry/fs/tty/slave.go b/pkg/sentry/fs/tty/replica.go index 7c7292687..385d230fb 100644 --- a/pkg/sentry/fs/tty/slave.go +++ b/pkg/sentry/fs/tty/replica.go @@ -17,9 +17,11 @@ package tty import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" @@ -27,11 +29,11 @@ import ( // LINT.IfChange -// slaveInodeOperations are the fs.InodeOperations for the slave end of the +// replicaInodeOperations are the fs.InodeOperations for the replica end of the // Terminal (pts file). // // +stateify savable -type slaveInodeOperations struct { +type replicaInodeOperations struct { fsutil.SimpleFileInode // d is the containing dir. @@ -41,13 +43,13 @@ type slaveInodeOperations struct { t *Terminal } -var _ fs.InodeOperations = (*slaveInodeOperations)(nil) +var _ fs.InodeOperations = (*replicaInodeOperations)(nil) -// newSlaveInode creates an fs.Inode for the slave end of a terminal. +// newReplicaInode creates an fs.Inode for the replica 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{ +// newReplicaInode takes ownership of t. +func newReplicaInode(ctx context.Context, d *dirInodeOperations, t *Terminal, owner fs.FileOwner, p fs.FilePermissions) *fs.Inode { + iops := &replicaInodeOperations{ SimpleFileInode: *fsutil.NewSimpleFileInode(ctx, owner, p, linux.DEVPTS_SUPER_MAGIC), d: d, t: t, @@ -64,18 +66,18 @@ func newSlaveInode(ctx context.Context, d *dirInodeOperations, t *Terminal, owne Type: fs.CharacterDevice, // See fs/devpts/inode.c:devpts_fill_super. BlockSize: 1024, - DeviceFileMajor: linux.UNIX98_PTY_SLAVE_MAJOR, + DeviceFileMajor: linux.UNIX98_PTY_REPLICA_MAJOR, DeviceFileMinor: t.n, }) } // Release implements fs.InodeOperations.Release. -func (si *slaveInodeOperations) Release(ctx context.Context) { +func (si *replicaInodeOperations) Release(ctx context.Context) { si.t.DecRef(ctx) } // Truncate implements fs.InodeOperations.Truncate. -func (*slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { +func (*replicaInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { return nil } @@ -83,14 +85,15 @@ func (*slaveInodeOperations) Truncate(context.Context, *fs.Inode, int64) error { // // 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 +func (si *replicaInodeOperations) GetFile(ctx context.Context, d *fs.Dirent, flags fs.FileFlags) (*fs.File, error) { + return fs.NewFile(ctx, d, flags, &replicaFileOperations{si: si}), nil } -// slaveFileOperations are the fs.FileOperations for the slave end of a terminal. +// replicaFileOperations are the fs.FileOperations for the replica end of a +// terminal. // // +stateify savable -type slaveFileOperations struct { +type replicaFileOperations struct { fsutil.FilePipeSeek `state:"nosave"` fsutil.FileNotDirReaddir `state:"nosave"` fsutil.FileNoFsync `state:"nosave"` @@ -100,79 +103,84 @@ type slaveFileOperations struct { fsutil.FileUseInodeUnstableAttr `state:"nosave"` // si is the inode operations. - si *slaveInodeOperations + si *replicaInodeOperations } -var _ fs.FileOperations = (*slaveFileOperations)(nil) +var _ fs.FileOperations = (*replicaFileOperations)(nil) // Release implements fs.FileOperations.Release. -func (sf *slaveFileOperations) Release(context.Context) { +func (sf *replicaFileOperations) Release(context.Context) { } // EventRegister implements waiter.Waitable.EventRegister. -func (sf *slaveFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { - sf.si.t.ld.slaveWaiter.EventRegister(e, mask) +func (sf *replicaFileOperations) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + sf.si.t.ld.replicaWaiter.EventRegister(e, mask) } // EventUnregister implements waiter.Waitable.EventUnregister. -func (sf *slaveFileOperations) EventUnregister(e *waiter.Entry) { - sf.si.t.ld.slaveWaiter.EventUnregister(e) +func (sf *replicaFileOperations) EventUnregister(e *waiter.Entry) { + sf.si.t.ld.replicaWaiter.EventUnregister(e) } // Readiness implements waiter.Waitable.Readiness. -func (sf *slaveFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { - return sf.si.t.ld.slaveReadiness() +func (sf *replicaFileOperations) Readiness(mask waiter.EventMask) waiter.EventMask { + return sf.si.t.ld.replicaReadiness() } // Read implements fs.FileOperations.Read. -func (sf *slaveFileOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOSequence, _ int64) (int64, error) { +func (sf *replicaFileOperations) 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) { +func (sf *replicaFileOperations) 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, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (sf *replicaFileOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // ioctl(2) may only be called from a task goroutine. + return 0, syserror.ENOTTY + } + switch cmd := args[1].Uint(); cmd { case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ // Get the number of bytes in the input queue read buffer. - return 0, sf.si.t.ld.inputQueueReadSize(ctx, io, args) + return 0, sf.si.t.ld.inputQueueReadSize(t, args) case linux.TCGETS: - return sf.si.t.ld.getTermios(ctx, io, args) + return sf.si.t.ld.getTermios(t, args) case linux.TCSETS: - return sf.si.t.ld.setTermios(ctx, io, args) + return sf.si.t.ld.setTermios(t, args) case linux.TCSETSW: // TODO(b/29356795): This should drain the output queue first. - return sf.si.t.ld.setTermios(ctx, io, args) + return sf.si.t.ld.setTermios(t, args) case linux.TIOCGPTN: - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sf.si.t.n), usermem.IOOpts{ - AddressSpaceActive: true, - }) + nP := primitive.Uint32(sf.si.t.n) + _, err := nP.CopyOut(t, args[2].Pointer()) return 0, err case linux.TIOCGWINSZ: - return 0, sf.si.t.ld.windowSize(ctx, io, args) + return 0, sf.si.t.ld.windowSize(t, args) case linux.TIOCSWINSZ: - return 0, sf.si.t.ld.setWindowSize(ctx, io, args) + return 0, sf.si.t.ld.setWindowSize(t, args) case linux.TIOCSCTTY: // Make the given terminal the controlling terminal of the // calling process. - return 0, sf.si.t.setControllingTTY(ctx, io, args, false /* isMaster */) + return 0, sf.si.t.setControllingTTY(ctx, args, false /* isMaster */) case linux.TIOCNOTTY: // Release this process's controlling terminal. - return 0, sf.si.t.releaseControllingTTY(ctx, io, args, false /* isMaster */) + return 0, sf.si.t.releaseControllingTTY(ctx, args, false /* isMaster */) case linux.TIOCGPGRP: // Get the foreground process group. - return sf.si.t.foregroundProcessGroup(ctx, io, args, false /* isMaster */) + return sf.si.t.foregroundProcessGroup(ctx, args, false /* isMaster */) case linux.TIOCSPGRP: // Set the foreground process group. - return sf.si.t.setForegroundProcessGroup(ctx, io, args, false /* isMaster */) + return sf.si.t.setForegroundProcessGroup(ctx, args, false /* isMaster */) default: maybeEmitUnimplementedEvent(ctx, cmd) return 0, syserror.ENOTTY } } -// LINT.ThenChange(../../fsimpl/devpts/slave.go) +// LINT.ThenChange(../../fsimpl/devpts/replica.go) diff --git a/pkg/sentry/fs/tty/terminal.go b/pkg/sentry/fs/tty/terminal.go index ddcccf4da..4f431d74d 100644 --- a/pkg/sentry/fs/tty/terminal.go +++ b/pkg/sentry/fs/tty/terminal.go @@ -17,10 +17,10 @@ package tty import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) // LINT.IfChange @@ -44,19 +44,19 @@ type Terminal struct { // this terminal. This field is immutable. masterKTTY *kernel.TTY - // slaveKTTY contains the controlling process of the slave end of this + // replicaKTTY contains the controlling process of the replica end of this // terminal. This field is immutable. - slaveKTTY *kernel.TTY + replicaKTTY *kernel.TTY } func newTerminal(ctx context.Context, d *dirInodeOperations, n uint32) *Terminal { - termios := linux.DefaultSlaveTermios + termios := linux.DefaultReplicaTermios t := Terminal{ - d: d, - n: n, - ld: newLineDiscipline(termios), - masterKTTY: &kernel.TTY{Index: n}, - slaveKTTY: &kernel.TTY{Index: n}, + d: d, + n: n, + ld: newLineDiscipline(termios), + masterKTTY: &kernel.TTY{Index: n}, + replicaKTTY: &kernel.TTY{Index: n}, } t.EnableLeakCheck("tty.Terminal") return &t @@ -64,7 +64,7 @@ func newTerminal(ctx context.Context, d *dirInodeOperations, n uint32) *Terminal // setControllingTTY makes tm the controlling terminal of the calling thread // group. -func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { +func (tm *Terminal) setControllingTTY(ctx context.Context, args arch.SyscallArguments, isMaster bool) error { task := kernel.TaskFromContext(ctx) if task == nil { panic("setControllingTTY must be called from a task context") @@ -75,7 +75,7 @@ func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args a // releaseControllingTTY removes tm as the controlling terminal of the calling // thread group. -func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { +func (tm *Terminal) releaseControllingTTY(ctx context.Context, args arch.SyscallArguments, isMaster bool) error { task := kernel.TaskFromContext(ctx) if task == nil { panic("releaseControllingTTY must be called from a task context") @@ -85,7 +85,7 @@ func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, ar } // foregroundProcessGroup gets the process group ID of tm's foreground process. -func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { +func (tm *Terminal) foregroundProcessGroup(ctx context.Context, args arch.SyscallArguments, isMaster bool) (uintptr, error) { task := kernel.TaskFromContext(ctx) if task == nil { panic("foregroundProcessGroup must be called from a task context") @@ -97,24 +97,21 @@ func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, a } // Write it out to *arg. - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(ret), usermem.IOOpts{ - AddressSpaceActive: true, - }) + retP := primitive.Int32(ret) + _, err = retP.CopyOut(task, args[2].Pointer()) return 0, err } // foregroundProcessGroup sets tm's foreground process. -func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { +func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, args arch.SyscallArguments, isMaster bool) (uintptr, error) { task := kernel.TaskFromContext(ctx) if task == nil { panic("setForegroundProcessGroup must be called from a task context") } // Read in the process group ID. - var pgid int32 - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgid, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + var pgid primitive.Int32 + if _, err := pgid.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } @@ -126,7 +123,7 @@ func (tm *Terminal) tty(isMaster bool) *kernel.TTY { if isMaster { return tm.masterKTTY } - return tm.slaveKTTY + return tm.replicaKTTY } // LINT.ThenChange(../../fsimpl/devpts/terminal.go) diff --git a/pkg/sentry/fs/tty/tty_test.go b/pkg/sentry/fs/tty/tty_test.go index 2cbc05678..49edee83d 100644 --- a/pkg/sentry/fs/tty/tty_test.go +++ b/pkg/sentry/fs/tty/tty_test.go @@ -22,8 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -func TestSimpleMasterToSlave(t *testing.T) { - ld := newLineDiscipline(linux.DefaultSlaveTermios) +func TestSimpleMasterToReplica(t *testing.T) { + ld := newLineDiscipline(linux.DefaultReplicaTermios) ctx := contexttest.Context(t) inBytes := []byte("hello, tty\n") src := usermem.BytesIOSequence(inBytes) diff --git a/pkg/sentry/fs/user/path.go b/pkg/sentry/fs/user/path.go index 2f5a43b84..124bc95ed 100644 --- a/pkg/sentry/fs/user/path.go +++ b/pkg/sentry/fs/user/path.go @@ -121,6 +121,7 @@ func resolve(ctx context.Context, mns *fs.MountNamespace, paths []string, name s func resolveVFS2(ctx context.Context, creds *auth.Credentials, mns *vfs.MountNamespace, paths []string, name string) (string, error) { root := mns.Root() + root.IncRef() defer root.DecRef(ctx) for _, p := range paths { if !path.IsAbs(p) { diff --git a/pkg/sentry/fs/user/user.go b/pkg/sentry/fs/user/user.go index 936fd3932..1f8684dc6 100644 --- a/pkg/sentry/fs/user/user.go +++ b/pkg/sentry/fs/user/user.go @@ -105,6 +105,7 @@ func getExecUserHomeVFS2(ctx context.Context, mns *vfs.MountNamespace, uid auth. const defaultHome = "/" root := mns.Root() + root.IncRef() defer root.DecRef(ctx) creds := auth.CredentialsFromContext(ctx) diff --git a/pkg/sentry/fsbridge/vfs.go b/pkg/sentry/fsbridge/vfs.go index 323506d33..be0900030 100644 --- a/pkg/sentry/fsbridge/vfs.go +++ b/pkg/sentry/fsbridge/vfs.go @@ -122,7 +122,7 @@ func NewVFSLookup(mntns *vfs.MountNamespace, root, workingDir vfs.VirtualDentry) // remainingTraversals is not configurable in VFS2, all callers are using the // default anyways. func (l *vfsLookup) OpenPath(ctx context.Context, pathname string, opts vfs.OpenOptions, _ *uint, resolveFinal bool) (File, error) { - vfsObj := l.mntns.Root().Mount().Filesystem().VirtualFilesystem() + vfsObj := l.root.Mount().Filesystem().VirtualFilesystem() creds := auth.CredentialsFromContext(ctx) path := fspath.Parse(pathname) pop := &vfs.PathOperation{ diff --git a/pkg/sentry/fsimpl/devpts/BUILD b/pkg/sentry/fsimpl/devpts/BUILD index 93512c9b6..6af3c3781 100644 --- a/pkg/sentry/fsimpl/devpts/BUILD +++ b/pkg/sentry/fsimpl/devpts/BUILD @@ -1,7 +1,19 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "root_inode_refs", + out = "root_inode_refs.go", + package = "devpts", + prefix = "rootInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "rootInode", + }, +) + go_library( name = "devpts", srcs = [ @@ -9,15 +21,22 @@ go_library( "line_discipline.go", "master.go", "queue.go", - "slave.go", + "replica.go", + "root_inode_refs.go", "terminal.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", + "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/arch", + "//pkg/sentry/fs", "//pkg/sentry/fs/lock", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", diff --git a/pkg/sentry/fsimpl/devpts/devpts.go b/pkg/sentry/fsimpl/devpts/devpts.go index 7169e91af..9185877f6 100644 --- a/pkg/sentry/fsimpl/devpts/devpts.go +++ b/pkg/sentry/fsimpl/devpts/devpts.go @@ -35,29 +35,56 @@ import ( const Name = "devpts" // FilesystemType implements vfs.FilesystemType. -type FilesystemType struct{} +// +// +stateify savable +type FilesystemType struct { + initOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1663): not yet supported. + initErr error + + // fs backs all mounts of this FilesystemType. root is fs' root. fs and root + // are immutable. + fs *vfs.Filesystem + root *vfs.Dentry +} // Name implements vfs.FilesystemType.Name. -func (FilesystemType) Name() string { +func (*FilesystemType) Name() string { return Name } -var _ vfs.FilesystemType = (*FilesystemType)(nil) - // GetFilesystem implements vfs.FilesystemType.GetFilesystem. -func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { +func (fstype *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { // No data allowed. if opts.Data != "" { return nil, nil, syserror.EINVAL } - fs, root, err := fstype.newFilesystem(vfsObj, creds) - if err != nil { - return nil, nil, err + fstype.initOnce.Do(func() { + fs, root, err := fstype.newFilesystem(ctx, vfsObj, creds) + if err != nil { + fstype.initErr = err + return + } + fstype.fs = fs.VFSFilesystem() + fstype.root = root.VFSDentry() + }) + if fstype.initErr != nil { + return nil, nil, fstype.initErr + } + fstype.fs.IncRef() + fstype.root.IncRef() + return fstype.fs, fstype.root, nil +} + +// Release implements vfs.FilesystemType.Release. +func (fstype *FilesystemType) Release(ctx context.Context) { + if fstype.fs != nil { + fstype.root.DecRef(ctx) + fstype.fs.DecRef(ctx) } - return fs.Filesystem.VFSFilesystem(), root.VFSDentry(), nil } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -66,7 +93,7 @@ type filesystem struct { // newFilesystem creates a new devpts filesystem with root directory and ptmx // master inode. It returns the filesystem and root Dentry. -func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*filesystem, *kernfs.Dentry, error) { +func (fstype *FilesystemType) newFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials) (*filesystem, *kernfs.Dentry, error) { devMinor, err := vfsObj.GetAnonBlockDevMinor() if err != nil { return nil, nil, err @@ -79,27 +106,29 @@ func (fstype FilesystemType) newFilesystem(vfsObj *vfs.VirtualFilesystem, creds // Construct the root directory. This is always inode id 1. root := &rootInode{ - slaves: make(map[uint32]*slaveInode), + replicas: make(map[uint32]*replicaInode), } - root.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 1, linux.ModeDirectory|0555) + root.InodeAttrs.Init(ctx, creds, linux.UNNAMED_MAJOR, devMinor, 1, linux.ModeDirectory|0555) root.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - root.dentry.Init(root) + root.EnableLeakCheck() + + var rootD kernfs.Dentry + rootD.Init(&fs.Filesystem, root) // Construct the pts master inode and dentry. Linux always uses inode // id 2 for ptmx. See fs/devpts/inode.c:mknod_ptmx. master := &masterInode{ root: root, } - master.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, devMinor, 2, linux.ModeCharacterDevice|0666) - master.dentry.Init(master) + master.InodeAttrs.Init(ctx, creds, linux.UNNAMED_MAJOR, devMinor, 2, linux.ModeCharacterDevice|0666) // Add the master as a child of the root. - links := root.OrderedChildren.Populate(&root.dentry, map[string]*kernfs.Dentry{ - "ptmx": &master.dentry, + links := root.OrderedChildren.Populate(map[string]kernfs.Inode{ + "ptmx": master, }) root.IncLinks(links) - return fs, &root.dentry, nil + return fs, &rootD, nil } // Release implements vfs.FilesystemImpl.Release. @@ -109,29 +138,28 @@ func (fs *filesystem) Release(ctx context.Context) { } // rootInode is the root directory inode for the devpts mounts. +// +// +stateify savable type rootInode struct { - kernfs.AlwaysValid + implStatFS + kernfs.InodeAlwaysValid kernfs.InodeAttrs kernfs.InodeDirectoryNoNewChildren kernfs.InodeNotSymlink + kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid. kernfs.OrderedChildren + rootInodeRefs locks vfs.FileLocks - // Keep a reference to this inode's dentry. - dentry kernfs.Dentry - // master is the master pty inode. Immutable. master *masterInode - // root is the root directory inode for this filesystem. Immutable. - root *rootInode - // mu protects the fields below. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` - // slaves maps pty ids to slave inodes. - slaves map[uint32]*slaveInode + // replicas maps pty ids to replica inodes. + replicas map[uint32]*replicaInode // nextIdx is the next pty index to use. Must be accessed atomically. // @@ -142,7 +170,7 @@ type rootInode struct { var _ kernfs.Inode = (*rootInode)(nil) // allocateTerminal creates a new Terminal and installs a pts node for it. -func (i *rootInode) allocateTerminal(creds *auth.Credentials) (*Terminal, error) { +func (i *rootInode) allocateTerminal(ctx context.Context, creds *auth.Credentials) (*Terminal, error) { i.mu.Lock() defer i.mu.Unlock() if i.nextIdx == math.MaxUint32 { @@ -151,41 +179,46 @@ func (i *rootInode) allocateTerminal(creds *auth.Credentials) (*Terminal, error) idx := i.nextIdx i.nextIdx++ - // Sanity check that slave with idx does not exist. - if _, ok := i.slaves[idx]; ok { + // Sanity check that replica with idx does not exist. + if _, ok := i.replicas[idx]; ok { panic(fmt.Sprintf("pty index collision; index %d already exists", idx)) } - // Create the new terminal and slave. + // Create the new terminal and replica. t := newTerminal(idx) - slave := &slaveInode{ + replica := &replicaInode{ root: i, t: t, } // Linux always uses pty index + 3 as the inode id. See // fs/devpts/inode.c:devpts_pty_new(). - slave.InodeAttrs.Init(creds, i.InodeAttrs.DevMajor(), i.InodeAttrs.DevMinor(), uint64(idx+3), linux.ModeCharacterDevice|0600) - slave.dentry.Init(slave) - i.slaves[idx] = slave + replica.InodeAttrs.Init(ctx, creds, i.InodeAttrs.DevMajor(), i.InodeAttrs.DevMinor(), uint64(idx+3), linux.ModeCharacterDevice|0600) + i.replicas[idx] = replica return t, nil } // masterClose is called when the master end of t is closed. -func (i *rootInode) masterClose(t *Terminal) { +func (i *rootInode) masterClose(ctx context.Context, t *Terminal) { i.mu.Lock() defer i.mu.Unlock() - // Sanity check that slave with idx exists. - if _, ok := i.slaves[t.n]; !ok { + // Sanity check that replica with idx exists. + ri, ok := i.replicas[t.n] + if !ok { panic(fmt.Sprintf("pty with index %d does not exist", t.n)) } - delete(i.slaves, t.n) + + // Drop the ref on replica inode taken during rootInode.allocateTerminal. + ri.DecRef(ctx) + delete(i.replicas, t.n) } // Open implements kernfs.Inode.Open. -func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndStaticEntries, + }) if err != nil { return nil, err } @@ -193,27 +226,34 @@ func (i *rootInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D } // Lookup implements kernfs.Inode.Lookup. -func (i *rootInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (i *rootInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { + // Check if a static entry was looked up. + if d, err := i.OrderedChildren.Lookup(ctx, name); err == nil { + return d, nil + } + + // Not a static entry. idx, err := strconv.ParseUint(name, 10, 32) if err != nil { return nil, syserror.ENOENT } i.mu.Lock() defer i.mu.Unlock() - if si, ok := i.slaves[uint32(idx)]; ok { - si.dentry.IncRef() - return si.dentry.VFSDentry(), nil + if ri, ok := i.replicas[uint32(idx)]; ok { + ri.IncRef() // This ref is passed to the dentry upon creation via Init. + return ri, nil } return nil, syserror.ENOENT } // IterDirents implements kernfs.Inode.IterDirents. -func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +func (i *rootInode) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { i.mu.Lock() defer i.mu.Unlock() - ids := make([]int, 0, len(i.slaves)) - for id := range i.slaves { + i.InodeAttrs.TouchAtime(ctx, mnt) + ids := make([]int, 0, len(i.replicas)) + for id := range i.replicas { ids = append(ids, int(id)) } sort.Ints(ids) @@ -221,7 +261,7 @@ func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, dirent := vfs.Dirent{ Name: strconv.FormatUint(uint64(id), 10), Type: linux.DT_CHR, - Ino: i.slaves[uint32(id)].InodeAttrs.Ino(), + Ino: i.replicas[uint32(id)].InodeAttrs.Ino(), NextOff: offset + 1, } if err := cb.Handle(dirent); err != nil { @@ -231,3 +271,16 @@ func (i *rootInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, } return offset, nil } + +// DecRef implements kernfs.Inode.DecRef. +func (i *rootInode) DecRef(ctx context.Context) { + i.rootInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + +// +stateify savable +type implStatFS struct{} + +// StatFS implements kernfs.Inode.StatFS. +func (*implStatFS) StatFS(context.Context, *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.DEVPTS_SUPER_MAGIC), nil +} diff --git a/pkg/sentry/fsimpl/devpts/devpts_test.go b/pkg/sentry/fsimpl/devpts/devpts_test.go index b7c149047..448390cfe 100644 --- a/pkg/sentry/fsimpl/devpts/devpts_test.go +++ b/pkg/sentry/fsimpl/devpts/devpts_test.go @@ -22,8 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -func TestSimpleMasterToSlave(t *testing.T) { - ld := newLineDiscipline(linux.DefaultSlaveTermios) +func TestSimpleMasterToReplica(t *testing.T) { + ld := newLineDiscipline(linux.DefaultReplicaTermios) ctx := contexttest.Context(t) inBytes := []byte("hello, tty\n") src := usermem.BytesIOSequence(inBytes) diff --git a/pkg/sentry/fsimpl/devpts/line_discipline.go b/pkg/sentry/fsimpl/devpts/line_discipline.go index f7bc325d1..ae95fdd08 100644 --- a/pkg/sentry/fsimpl/devpts/line_discipline.go +++ b/pkg/sentry/fsimpl/devpts/line_discipline.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -41,7 +42,7 @@ const ( ) // lineDiscipline dictates how input and output are handled between the -// pseudoterminal (pty) master and slave. It can be configured to alter I/O, +// pseudoterminal (pty) master and replica. 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: // @@ -52,8 +53,8 @@ const ( // // 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 +// terminal master to replica (the input queue) and one from replica 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: // @@ -62,7 +63,7 @@ const ( // | (inputQueueWrite) +-------------+ (inputQueueRead) | // | | // | v -// masterFD slaveFD +// masterFD replicaFD // ^ | // | | // | output to terminal +--------------+ output from process | @@ -99,10 +100,10 @@ type lineDiscipline struct { column int // masterWaiter is used to wait on the master end of the TTY. - masterWaiter waiter.Queue `state:"zerovalue"` + masterWaiter waiter.Queue - // slaveWaiter is used to wait on the slave end of the TTY. - slaveWaiter waiter.Queue `state:"zerovalue"` + // replicaWaiter is used to wait on the replica end of the TTY. + replicaWaiter waiter.Queue } func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { @@ -113,27 +114,23 @@ func newLineDiscipline(termios linux.KernelTermios) *lineDiscipline { } // getTermios gets the linux.Termios for the tty. -func (l *lineDiscipline) getTermios(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { +func (l *lineDiscipline) getTermios(task *kernel.Task, args arch.SyscallArguments) (uintptr, error) { l.termiosMu.RLock() defer l.termiosMu.RUnlock() // 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, - }) + _, err := t.CopyOut(task, args[2].Pointer()) 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) { +func (l *lineDiscipline) setTermios(task *kernel.Task, args arch.SyscallArguments) (uintptr, error) { l.termiosMu.Lock() defer l.termiosMu.Unlock() oldCanonEnabled := l.termios.LEnabled(linux.ICANON) // 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, - }) + _, err := t.CopyIn(task, args[2].Pointer()) l.termios.FromTermios(t) // If canonical mode is turned off, move bytes from inQueue's wait @@ -144,27 +141,23 @@ func (l *lineDiscipline) setTermios(ctx context.Context, io usermem.IO, args arc l.inQueue.pushWaitBufLocked(l) l.inQueue.readable = true l.inQueue.mu.Unlock() - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) } return 0, err } -func (l *lineDiscipline) windowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (l *lineDiscipline) windowSize(t *kernel.Task, args arch.SyscallArguments) error { l.sizeMu.Lock() defer l.sizeMu.Unlock() - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), l.size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := l.size.CopyOut(t, args[2].Pointer()) return err } -func (l *lineDiscipline) setWindowSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (l *lineDiscipline) setWindowSize(t *kernel.Task, args arch.SyscallArguments) error { l.sizeMu.Lock() defer l.sizeMu.Unlock() - _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &l.size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := l.size.CopyIn(t, args[2].Pointer()) return err } @@ -174,14 +167,14 @@ func (l *lineDiscipline) masterReadiness() waiter.EventMask { return l.inQueue.writeReadiness(&linux.MasterTermios) | l.outQueue.readReadiness(&linux.MasterTermios) } -func (l *lineDiscipline) slaveReadiness() waiter.EventMask { +func (l *lineDiscipline) replicaReadiness() waiter.EventMask { l.termiosMu.RLock() defer l.termiosMu.RUnlock() return l.outQueue.writeReadiness(&l.termios) | l.inQueue.readReadiness(&l.termios) } -func (l *lineDiscipline) inputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { - return l.inQueue.readableSize(ctx, io, args) +func (l *lineDiscipline) inputQueueReadSize(t *kernel.Task, io usermem.IO, args arch.SyscallArguments) error { + return l.inQueue.readableSize(t, io, args) } func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { @@ -194,7 +187,7 @@ func (l *lineDiscipline) inputQueueRead(ctx context.Context, dst usermem.IOSeque if n > 0 { l.masterWaiter.Notify(waiter.EventOut) if pushed { - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) } return n, nil } @@ -209,14 +202,14 @@ func (l *lineDiscipline) inputQueueWrite(ctx context.Context, src usermem.IOSequ return 0, err } if n > 0 { - l.slaveWaiter.Notify(waiter.EventIn) + l.replicaWaiter.Notify(waiter.EventIn) return n, nil } return 0, syserror.ErrWouldBlock } -func (l *lineDiscipline) outputQueueReadSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { - return l.outQueue.readableSize(ctx, io, args) +func (l *lineDiscipline) outputQueueReadSize(t *kernel.Task, io usermem.IO, args arch.SyscallArguments) error { + return l.outQueue.readableSize(t, io, args) } func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequence) (int64, error) { @@ -227,7 +220,7 @@ func (l *lineDiscipline) outputQueueRead(ctx context.Context, dst usermem.IOSequ return 0, err } if n > 0 { - l.slaveWaiter.Notify(waiter.EventOut) + l.replicaWaiter.Notify(waiter.EventOut) if pushed { l.masterWaiter.Notify(waiter.EventIn) } diff --git a/pkg/sentry/fsimpl/devpts/master.go b/pkg/sentry/fsimpl/devpts/master.go index 3bb397f71..e91fa26a4 100644 --- a/pkg/sentry/fsimpl/devpts/master.go +++ b/pkg/sentry/fsimpl/devpts/master.go @@ -17,9 +17,11 @@ package devpts import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/unimpl" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -29,7 +31,10 @@ import ( ) // masterInode is the inode for the master end of the Terminal. +// +// +stateify savable type masterInode struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeNotDirectory @@ -37,9 +42,6 @@ type masterInode struct { locks vfs.FileLocks - // Keep a reference to this inode's dentry. - dentry kernfs.Dentry - // root is the devpts root inode. root *rootInode } @@ -47,20 +49,18 @@ type masterInode struct { var _ kernfs.Inode = (*masterInode)(nil) // Open implements kernfs.Inode.Open. -func (mi *masterInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - t, err := mi.root.allocateTerminal(rp.Credentials()) +func (mi *masterInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + t, err := mi.root.allocateTerminal(ctx, rp.Credentials()) if err != nil { return nil, err } - mi.IncRef() fd := &masterFileDescription{ inode: mi, t: t, } fd.LockFD.Init(&mi.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { - mi.DecRef(ctx) + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -86,6 +86,7 @@ func (mi *masterInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds return mi.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) } +// +stateify savable type masterFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -99,8 +100,7 @@ var _ vfs.FileDescriptionImpl = (*masterFileDescription)(nil) // Release implements vfs.FileDescriptionImpl.Release. func (mfd *masterFileDescription) Release(ctx context.Context) { - mfd.inode.root.masterClose(mfd.t) - mfd.inode.DecRef(ctx) + mfd.inode.root.masterClose(ctx, mfd.t) } // EventRegister implements waiter.Waitable.EventRegister. @@ -130,46 +130,51 @@ func (mfd *masterFileDescription) Write(ctx context.Context, src usermem.IOSeque // Ioctl implements vfs.FileDescriptionImpl.Ioctl. func (mfd *masterFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // ioctl(2) may only be called from a task goroutine. + return 0, syserror.ENOTTY + } + switch cmd := args[1].Uint(); cmd { case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ // Get the number of bytes in the output queue read buffer. - return 0, mfd.t.ld.outputQueueReadSize(ctx, io, args) + return 0, mfd.t.ld.outputQueueReadSize(t, io, args) case linux.TCGETS: // N.B. TCGETS on the master actually returns the configuration - // of the slave end. - return mfd.t.ld.getTermios(ctx, io, args) + // of the replica end. + return mfd.t.ld.getTermios(t, args) case linux.TCSETS: // N.B. TCSETS on the master actually affects the configuration - // of the slave end. - return mfd.t.ld.setTermios(ctx, io, args) + // of the replica end. + return mfd.t.ld.setTermios(t, args) case linux.TCSETSW: // TODO(b/29356795): This should drain the output queue first. - return mfd.t.ld.setTermios(ctx, io, args) + return mfd.t.ld.setTermios(t, args) case linux.TIOCGPTN: - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(mfd.t.n), usermem.IOOpts{ - AddressSpaceActive: true, - }) + nP := primitive.Uint32(mfd.t.n) + _, err := nP.CopyOut(t, args[2].Pointer()) return 0, err case linux.TIOCSPTLCK: // TODO(b/29356795): Implement pty locking. For now just pretend we do. return 0, nil case linux.TIOCGWINSZ: - return 0, mfd.t.ld.windowSize(ctx, io, args) + return 0, mfd.t.ld.windowSize(t, args) case linux.TIOCSWINSZ: - return 0, mfd.t.ld.setWindowSize(ctx, io, args) + return 0, mfd.t.ld.setWindowSize(t, args) case linux.TIOCSCTTY: // Make the given terminal the controlling terminal of the // calling process. - return 0, mfd.t.setControllingTTY(ctx, io, args, true /* isMaster */) + return 0, mfd.t.setControllingTTY(ctx, args, true /* isMaster */) case linux.TIOCNOTTY: // Release this process's controlling terminal. - return 0, mfd.t.releaseControllingTTY(ctx, io, args, true /* isMaster */) + return 0, mfd.t.releaseControllingTTY(ctx, args, true /* isMaster */) case linux.TIOCGPGRP: // Get the foreground process group. - return mfd.t.foregroundProcessGroup(ctx, io, args, true /* isMaster */) + return mfd.t.foregroundProcessGroup(ctx, args, true /* isMaster */) case linux.TIOCSPGRP: // Set the foreground process group. - return mfd.t.setForegroundProcessGroup(ctx, io, args, true /* isMaster */) + return mfd.t.setForegroundProcessGroup(ctx, args, true /* isMaster */) default: maybeEmitUnimplementedEvent(ctx, cmd) return 0, syserror.ENOTTY diff --git a/pkg/sentry/fsimpl/devpts/queue.go b/pkg/sentry/fsimpl/devpts/queue.go index dffb4232c..55bff3e60 100644 --- a/pkg/sentry/fsimpl/devpts/queue.go +++ b/pkg/sentry/fsimpl/devpts/queue.go @@ -17,8 +17,10 @@ package devpts import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -30,7 +32,7 @@ import ( const waitBufMaxBytes = 131072 // queue represents one of the input or output queues between a pty master and -// slave. Bytes written to a queue are added to the read buffer until it is +// replica. Bytes written to a queue are added to the read buffer until it is // full, at which point they are written to the wait buffer. Bytes are // processed (i.e. undergo termios transformations) as they are added to the // read buffer. The read buffer is readable when its length is nonzero and @@ -83,17 +85,15 @@ func (q *queue) writeReadiness(t *linux.KernelTermios) waiter.EventMask { } // readableSize writes the number of readable bytes to userspace. -func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.SyscallArguments) error { +func (q *queue) readableSize(t *kernel.Task, io usermem.IO, args arch.SyscallArguments) error { q.mu.Lock() defer q.mu.Unlock() - var size int32 + size := primitive.Int32(0) if q.readable { - size = int32(len(q.readBuf)) + size = primitive.Int32(len(q.readBuf)) } - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), size, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err := size.CopyOut(t, args[2].Pointer()) return err } @@ -102,8 +102,7 @@ func (q *queue) readableSize(ctx context.Context, io usermem.IO, args arch.Sysca // as whether the read caused more readable data to become available (whether // data was pushed from the wait buffer to the read buffer). // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipline) (int64, bool, error) { q.mu.Lock() defer q.mu.Unlock() @@ -143,8 +142,7 @@ func (q *queue) read(ctx context.Context, dst usermem.IOSequence, l *lineDiscipl // write writes to q from userspace. // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscipline) (int64, error) { q.mu.Lock() defer q.mu.Unlock() @@ -186,8 +184,7 @@ func (q *queue) write(ctx context.Context, src usermem.IOSequence, l *lineDiscip // writeBytes writes to q from b. // -// Preconditions: -// * l.termiosMu must be held for reading. +// Preconditions: l.termiosMu must be held for reading. func (q *queue) writeBytes(b []byte, l *lineDiscipline) { q.mu.Lock() defer q.mu.Unlock() diff --git a/pkg/sentry/fsimpl/devpts/replica.go b/pkg/sentry/fsimpl/devpts/replica.go new file mode 100644 index 000000000..70c68cf0a --- /dev/null +++ b/pkg/sentry/fsimpl/devpts/replica.go @@ -0,0 +1,201 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devpts + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" + "gvisor.dev/gvisor/pkg/sentry/arch" + fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +// replicaInode is the inode for the replica end of the Terminal. +// +// +stateify savable +type replicaInode struct { + implStatFS + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + + locks vfs.FileLocks + + // root is the devpts root inode. + root *rootInode + + // t is the connected Terminal. + t *Terminal +} + +var _ kernfs.Inode = (*replicaInode)(nil) + +// Open implements kernfs.Inode.Open. +func (ri *replicaInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd := &replicaFileDescription{ + inode: ri, + } + fd.LockFD.Init(&ri.locks) + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return &fd.vfsfd, nil + +} + +// Valid implements kernfs.Inode.Valid. +func (ri *replicaInode) Valid(context.Context) bool { + // Return valid if the replica still exists. + ri.root.mu.Lock() + defer ri.root.mu.Unlock() + _, ok := ri.root.replicas[ri.t.n] + return ok +} + +// Stat implements kernfs.Inode.Stat. +func (ri *replicaInode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + statx, err := ri.InodeAttrs.Stat(ctx, vfsfs, opts) + if err != nil { + return linux.Statx{}, err + } + statx.Blksize = 1024 + statx.RdevMajor = linux.UNIX98_PTY_REPLICA_MAJOR + statx.RdevMinor = ri.t.n + return statx, nil +} + +// SetStat implements kernfs.Inode.SetStat +func (ri *replicaInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + if opts.Stat.Mask&linux.STATX_SIZE != 0 { + return syserror.EINVAL + } + return ri.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) +} + +// +stateify savable +type replicaFileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.LockFD + + inode *replicaInode +} + +var _ vfs.FileDescriptionImpl = (*replicaFileDescription)(nil) + +// Release implements fs.FileOperations.Release. +func (rfd *replicaFileDescription) Release(ctx context.Context) {} + +// EventRegister implements waiter.Waitable.EventRegister. +func (rfd *replicaFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + rfd.inode.t.ld.replicaWaiter.EventRegister(e, mask) +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (rfd *replicaFileDescription) EventUnregister(e *waiter.Entry) { + rfd.inode.t.ld.replicaWaiter.EventUnregister(e) +} + +// Readiness implements waiter.Waitable.Readiness. +func (rfd *replicaFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { + return rfd.inode.t.ld.replicaReadiness() +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (rfd *replicaFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { + return rfd.inode.t.ld.inputQueueRead(ctx, dst) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (rfd *replicaFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { + return rfd.inode.t.ld.outputQueueWrite(ctx, src) +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (rfd *replicaFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + // ioctl(2) may only be called from a task goroutine. + return 0, syserror.ENOTTY + } + + switch cmd := args[1].Uint(); cmd { + case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ + // Get the number of bytes in the input queue read buffer. + return 0, rfd.inode.t.ld.inputQueueReadSize(t, io, args) + case linux.TCGETS: + return rfd.inode.t.ld.getTermios(t, args) + case linux.TCSETS: + return rfd.inode.t.ld.setTermios(t, args) + case linux.TCSETSW: + // TODO(b/29356795): This should drain the output queue first. + return rfd.inode.t.ld.setTermios(t, args) + case linux.TIOCGPTN: + nP := primitive.Uint32(rfd.inode.t.n) + _, err := nP.CopyOut(t, args[2].Pointer()) + return 0, err + case linux.TIOCGWINSZ: + return 0, rfd.inode.t.ld.windowSize(t, args) + case linux.TIOCSWINSZ: + return 0, rfd.inode.t.ld.setWindowSize(t, args) + case linux.TIOCSCTTY: + // Make the given terminal the controlling terminal of the + // calling process. + return 0, rfd.inode.t.setControllingTTY(ctx, args, false /* isMaster */) + case linux.TIOCNOTTY: + // Release this process's controlling terminal. + return 0, rfd.inode.t.releaseControllingTTY(ctx, args, false /* isMaster */) + case linux.TIOCGPGRP: + // Get the foreground process group. + return rfd.inode.t.foregroundProcessGroup(ctx, args, false /* isMaster */) + case linux.TIOCSPGRP: + // Set the foreground process group. + return rfd.inode.t.setForegroundProcessGroup(ctx, args, false /* isMaster */) + default: + maybeEmitUnimplementedEvent(ctx, cmd) + return 0, syserror.ENOTTY + } +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (rfd *replicaFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := rfd.vfsfd.VirtualDentry().Mount().Filesystem() + return rfd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (rfd *replicaFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := rfd.vfsfd.VirtualDentry().Mount().Filesystem() + return rfd.inode.Stat(ctx, fs, opts) +} + +// LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. +func (rfd *replicaFileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { + return rfd.Locks().LockPOSIX(ctx, &rfd.vfsfd, uid, t, start, length, whence, block) +} + +// UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. +func (rfd *replicaFileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { + return rfd.Locks().UnlockPOSIX(ctx, &rfd.vfsfd, uid, start, length, whence) +} diff --git a/pkg/sentry/fsimpl/devpts/slave.go b/pkg/sentry/fsimpl/devpts/slave.go deleted file mode 100644 index 32e4e1908..000000000 --- a/pkg/sentry/fsimpl/devpts/slave.go +++ /dev/null @@ -1,197 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package devpts - -import ( - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/arch" - fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" - "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/pkg/waiter" -) - -// slaveInode is the inode for the slave end of the Terminal. -type slaveInode struct { - kernfs.InodeAttrs - kernfs.InodeNoopRefCount - kernfs.InodeNotDirectory - kernfs.InodeNotSymlink - - locks vfs.FileLocks - - // Keep a reference to this inode's dentry. - dentry kernfs.Dentry - - // root is the devpts root inode. - root *rootInode - - // t is the connected Terminal. - t *Terminal -} - -var _ kernfs.Inode = (*slaveInode)(nil) - -// Open implements kernfs.Inode.Open. -func (si *slaveInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - si.IncRef() - fd := &slaveFileDescription{ - inode: si, - } - fd.LockFD.Init(&si.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { - si.DecRef(ctx) - return nil, err - } - return &fd.vfsfd, nil - -} - -// Valid implements kernfs.Inode.Valid. -func (si *slaveInode) Valid(context.Context) bool { - // Return valid if the slave still exists. - si.root.mu.Lock() - defer si.root.mu.Unlock() - _, ok := si.root.slaves[si.t.n] - return ok -} - -// Stat implements kernfs.Inode.Stat. -func (si *slaveInode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { - statx, err := si.InodeAttrs.Stat(ctx, vfsfs, opts) - if err != nil { - return linux.Statx{}, err - } - statx.Blksize = 1024 - statx.RdevMajor = linux.UNIX98_PTY_SLAVE_MAJOR - statx.RdevMinor = si.t.n - return statx, nil -} - -// SetStat implements kernfs.Inode.SetStat -func (si *slaveInode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { - if opts.Stat.Mask&linux.STATX_SIZE != 0 { - return syserror.EINVAL - } - return si.InodeAttrs.SetStat(ctx, vfsfs, creds, opts) -} - -type slaveFileDescription struct { - vfsfd vfs.FileDescription - vfs.FileDescriptionDefaultImpl - vfs.LockFD - - inode *slaveInode -} - -var _ vfs.FileDescriptionImpl = (*slaveFileDescription)(nil) - -// Release implements fs.FileOperations.Release. -func (sfd *slaveFileDescription) Release(ctx context.Context) { - sfd.inode.DecRef(ctx) -} - -// EventRegister implements waiter.Waitable.EventRegister. -func (sfd *slaveFileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { - sfd.inode.t.ld.slaveWaiter.EventRegister(e, mask) -} - -// EventUnregister implements waiter.Waitable.EventUnregister. -func (sfd *slaveFileDescription) EventUnregister(e *waiter.Entry) { - sfd.inode.t.ld.slaveWaiter.EventUnregister(e) -} - -// Readiness implements waiter.Waitable.Readiness. -func (sfd *slaveFileDescription) Readiness(mask waiter.EventMask) waiter.EventMask { - return sfd.inode.t.ld.slaveReadiness() -} - -// Read implements vfs.FileDescriptionImpl.Read. -func (sfd *slaveFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { - return sfd.inode.t.ld.inputQueueRead(ctx, dst) -} - -// Write implements vfs.FileDescriptionImpl.Write. -func (sfd *slaveFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { - return sfd.inode.t.ld.outputQueueWrite(ctx, src) -} - -// Ioctl implements vfs.FileDescriptionImpl.Ioctl. -func (sfd *slaveFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { - switch cmd := args[1].Uint(); cmd { - case linux.FIONREAD: // linux.FIONREAD == linux.TIOCINQ - // Get the number of bytes in the input queue read buffer. - return 0, sfd.inode.t.ld.inputQueueReadSize(ctx, io, args) - case linux.TCGETS: - return sfd.inode.t.ld.getTermios(ctx, io, args) - case linux.TCSETS: - return sfd.inode.t.ld.setTermios(ctx, io, args) - case linux.TCSETSW: - // TODO(b/29356795): This should drain the output queue first. - return sfd.inode.t.ld.setTermios(ctx, io, args) - case linux.TIOCGPTN: - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), uint32(sfd.inode.t.n), usermem.IOOpts{ - AddressSpaceActive: true, - }) - return 0, err - case linux.TIOCGWINSZ: - return 0, sfd.inode.t.ld.windowSize(ctx, io, args) - case linux.TIOCSWINSZ: - return 0, sfd.inode.t.ld.setWindowSize(ctx, io, args) - case linux.TIOCSCTTY: - // Make the given terminal the controlling terminal of the - // calling process. - return 0, sfd.inode.t.setControllingTTY(ctx, io, args, false /* isMaster */) - case linux.TIOCNOTTY: - // Release this process's controlling terminal. - return 0, sfd.inode.t.releaseControllingTTY(ctx, io, args, false /* isMaster */) - case linux.TIOCGPGRP: - // Get the foreground process group. - return sfd.inode.t.foregroundProcessGroup(ctx, io, args, false /* isMaster */) - case linux.TIOCSPGRP: - // Set the foreground process group. - return sfd.inode.t.setForegroundProcessGroup(ctx, io, args, false /* isMaster */) - default: - maybeEmitUnimplementedEvent(ctx, cmd) - return 0, syserror.ENOTTY - } -} - -// SetStat implements vfs.FileDescriptionImpl.SetStat. -func (sfd *slaveFileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - creds := auth.CredentialsFromContext(ctx) - fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() - return sfd.inode.SetStat(ctx, fs, creds, opts) -} - -// Stat implements vfs.FileDescriptionImpl.Stat. -func (sfd *slaveFileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { - fs := sfd.vfsfd.VirtualDentry().Mount().Filesystem() - return sfd.inode.Stat(ctx, fs, opts) -} - -// LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. -func (sfd *slaveFileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { - return sfd.Locks().LockPOSIX(ctx, &sfd.vfsfd, uid, t, start, length, whence, block) -} - -// UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. -func (sfd *slaveFileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { - return sfd.Locks().UnlockPOSIX(ctx, &sfd.vfsfd, uid, start, length, whence) -} diff --git a/pkg/sentry/fsimpl/devpts/terminal.go b/pkg/sentry/fsimpl/devpts/terminal.go index 7d2781c54..510bd6d89 100644 --- a/pkg/sentry/fsimpl/devpts/terminal.go +++ b/pkg/sentry/fsimpl/devpts/terminal.go @@ -17,9 +17,9 @@ package devpts import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/usermem" ) // Terminal is a pseudoterminal. @@ -36,25 +36,25 @@ type Terminal struct { // this terminal. This field is immutable. masterKTTY *kernel.TTY - // slaveKTTY contains the controlling process of the slave end of this + // replicaKTTY contains the controlling process of the replica end of this // terminal. This field is immutable. - slaveKTTY *kernel.TTY + replicaKTTY *kernel.TTY } func newTerminal(n uint32) *Terminal { - termios := linux.DefaultSlaveTermios + termios := linux.DefaultReplicaTermios t := Terminal{ - n: n, - ld: newLineDiscipline(termios), - masterKTTY: &kernel.TTY{Index: n}, - slaveKTTY: &kernel.TTY{Index: n}, + n: n, + ld: newLineDiscipline(termios), + masterKTTY: &kernel.TTY{Index: n}, + replicaKTTY: &kernel.TTY{Index: n}, } return &t } // setControllingTTY makes tm the controlling terminal of the calling thread // group. -func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { +func (tm *Terminal) setControllingTTY(ctx context.Context, args arch.SyscallArguments, isMaster bool) error { task := kernel.TaskFromContext(ctx) if task == nil { panic("setControllingTTY must be called from a task context") @@ -65,7 +65,7 @@ func (tm *Terminal) setControllingTTY(ctx context.Context, io usermem.IO, args a // releaseControllingTTY removes tm as the controlling terminal of the calling // thread group. -func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) error { +func (tm *Terminal) releaseControllingTTY(ctx context.Context, args arch.SyscallArguments, isMaster bool) error { task := kernel.TaskFromContext(ctx) if task == nil { panic("releaseControllingTTY must be called from a task context") @@ -75,7 +75,7 @@ func (tm *Terminal) releaseControllingTTY(ctx context.Context, io usermem.IO, ar } // foregroundProcessGroup gets the process group ID of tm's foreground process. -func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { +func (tm *Terminal) foregroundProcessGroup(ctx context.Context, args arch.SyscallArguments, isMaster bool) (uintptr, error) { task := kernel.TaskFromContext(ctx) if task == nil { panic("foregroundProcessGroup must be called from a task context") @@ -87,24 +87,21 @@ func (tm *Terminal) foregroundProcessGroup(ctx context.Context, io usermem.IO, a } // Write it out to *arg. - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(ret), usermem.IOOpts{ - AddressSpaceActive: true, - }) + retP := primitive.Int32(ret) + _, err = retP.CopyOut(task, args[2].Pointer()) return 0, err } // foregroundProcessGroup sets tm's foreground process. -func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, io usermem.IO, args arch.SyscallArguments, isMaster bool) (uintptr, error) { +func (tm *Terminal) setForegroundProcessGroup(ctx context.Context, args arch.SyscallArguments, isMaster bool) (uintptr, error) { task := kernel.TaskFromContext(ctx) if task == nil { panic("setForegroundProcessGroup must be called from a task context") } // Read in the process group ID. - var pgid int32 - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgid, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + var pgid primitive.Int32 + if _, err := pgid.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } @@ -116,5 +113,5 @@ func (tm *Terminal) tty(isMaster bool) *kernel.TTY { if isMaster { return tm.masterKTTY } - return tm.slaveKTTY + return tm.replicaKTTY } diff --git a/pkg/sentry/fsimpl/devtmpfs/BUILD b/pkg/sentry/fsimpl/devtmpfs/BUILD index aa0c2ad8c..e49a04c1b 100644 --- a/pkg/sentry/fsimpl/devtmpfs/BUILD +++ b/pkg/sentry/fsimpl/devtmpfs/BUILD @@ -4,7 +4,10 @@ licenses(["notice"]) go_library( name = "devtmpfs", - srcs = ["devtmpfs.go"], + srcs = [ + "devtmpfs.go", + "save_restore.go", + ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", @@ -24,6 +27,7 @@ go_test( library = ":devtmpfs", deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/fspath", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/tmpfs", diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go index 2ed5fa8a9..e6fe0fc0d 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs.go @@ -18,6 +18,7 @@ package devtmpfs import ( "fmt" + "path" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -32,8 +33,10 @@ import ( const Name = "devtmpfs" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct { - initOnce sync.Once + initOnce sync.Once `state:"nosave"` // FIXME(gvisor.dev/issue/1664): not yet supported. initErr error // fs is the tmpfs filesystem that backs all mounts of this FilesystemType. @@ -68,6 +71,15 @@ func (fst *FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virtua return fst.fs, fst.root, nil } +// Release implements vfs.FilesystemType.Release. +func (fst *FilesystemType) Release(ctx context.Context) { + if fst.fs != nil { + // Release the original reference obtained when creating the filesystem. + fst.root.DecRef(ctx) + fst.fs.DecRef(ctx) + } +} + // Accessor allows devices to create device special files in devtmpfs. type Accessor struct { vfsObj *vfs.VirtualFilesystem @@ -79,14 +91,17 @@ type Accessor struct { // NewAccessor returns an Accessor that supports creation of device special // files in the devtmpfs instance registered with name fsTypeName in vfsObj. func NewAccessor(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, fsTypeName string) (*Accessor, error) { - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "devtmpfs" /* source */, fsTypeName, &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "devtmpfs" /* source */, fsTypeName, &vfs.MountOptions{}) if err != nil { return nil, err } + // Pass a reference on root to the Accessor. + root := mntns.Root() + root.IncRef() return &Accessor{ vfsObj: vfsObj, mntns: mntns, - root: mntns.Root(), + root: root, creds: creds, }, nil } @@ -150,13 +165,11 @@ func (a *Accessor) CreateDeviceFile(ctx context.Context, pathname string, kind v // Create any parent directories. See // devtmpfs.c:handle_create()=>path_create(). - for it := fspath.Parse(pathname).Begin; it.NextOk(); it = it.Next() { - pop := a.pathOperationAt(it.String()) - if err := a.vfsObj.MkdirAt(actx, a.creds, pop, &vfs.MkdirOptions{ - Mode: 0755, - }); err != nil { - return fmt.Errorf("failed to create directory %q: %v", it.String(), err) - } + parent := path.Dir(pathname) + if err := a.vfsObj.MkdirAllAt(ctx, parent, a.root, a.creds, &vfs.MkdirOptions{ + Mode: 0755, + }); err != nil { + return fmt.Errorf("failed to create device parent directory %q: %v", parent, err) } // NOTE: Linux's devtmpfs refuses to automatically delete files it didn't diff --git a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go index 747867cca..e058eda7a 100644 --- a/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go +++ b/pkg/sentry/fsimpl/devtmpfs/devtmpfs_test.go @@ -15,9 +15,11 @@ package devtmpfs import ( + "path" "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" @@ -25,10 +27,13 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) -func TestDevtmpfs(t *testing.T) { +const devPath = "/dev" + +func setupDevtmpfs(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesystem, vfs.VirtualDentry, func()) { + t.Helper() + ctx := contexttest.Context(t) creds := auth.CredentialsFromContext(ctx) - vfsObj := &vfs.VirtualFilesystem{} if err := vfsObj.Init(ctx); err != nil { t.Fatalf("VFS init: %v", err) @@ -43,14 +48,12 @@ func TestDevtmpfs(t *testing.T) { }) // Create a test mount namespace with devtmpfs mounted at "/dev". - const devPath = "/dev" - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "tmpfs" /* source */, "tmpfs" /* fsTypeName */, &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "tmpfs" /* source */, "tmpfs" /* fsTypeName */, &vfs.MountOptions{}) if err != nil { t.Fatalf("failed to create tmpfs root mount: %v", err) } - defer mntns.DecRef(ctx) root := mntns.Root() - defer root.DecRef(ctx) + root.IncRef() devpop := vfs.PathOperation{ Root: root, Start: root, @@ -61,10 +64,20 @@ func TestDevtmpfs(t *testing.T) { }); err != nil { t.Fatalf("failed to create mount point: %v", err) } - if err := vfsObj.MountAt(ctx, creds, "devtmpfs" /* source */, &devpop, "devtmpfs" /* fsTypeName */, &vfs.MountOptions{}); err != nil { + if _, err := vfsObj.MountAt(ctx, creds, "devtmpfs" /* source */, &devpop, "devtmpfs" /* fsTypeName */, &vfs.MountOptions{}); err != nil { t.Fatalf("failed to mount devtmpfs: %v", err) } + return ctx, creds, vfsObj, root, func() { + root.DecRef(ctx) + mntns.DecRef(ctx) + } +} + +func TestUserspaceInit(t *testing.T) { + ctx, creds, vfsObj, root, cleanup := setupDevtmpfs(t) + defer cleanup() + a, err := NewAccessor(ctx, vfsObj, creds, "devtmpfs") if err != nil { t.Fatalf("failed to create devtmpfs.Accessor: %v", err) @@ -75,48 +88,143 @@ func TestDevtmpfs(t *testing.T) { if err := a.UserspaceInit(ctx); err != nil { t.Fatalf("failed to userspace-initialize devtmpfs: %v", err) } + // Created files should be visible in the test mount namespace. - abspath := devPath + "/fd" - target, err := vfsObj.ReadlinkAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(abspath), - }) - if want := "/proc/self/fd"; err != nil || target != want { - t.Fatalf("readlink(%q): got (%q, %v), wanted (%q, nil)", abspath, target, err, want) + links := []struct { + source string + target string + }{ + { + source: "fd", + target: "/proc/self/fd", + }, + { + source: "stdin", + target: "/proc/self/fd/0", + }, + { + source: "stdout", + target: "/proc/self/fd/1", + }, + { + source: "stderr", + target: "/proc/self/fd/2", + }, + { + source: "ptmx", + target: "pts/ptmx", + }, } - // Create a dummy device special file using a devtmpfs.Accessor. - const ( - pathInDev = "dummy" - kind = vfs.CharDevice - major = 12 - minor = 34 - perms = 0600 - wantMode = linux.S_IFCHR | perms - ) - if err := a.CreateDeviceFile(ctx, pathInDev, kind, major, minor, perms); err != nil { - t.Fatalf("failed to create device file: %v", err) + for _, link := range links { + abspath := path.Join(devPath, link.source) + if gotTarget, err := vfsObj.ReadlinkAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(abspath), + }); err != nil || gotTarget != link.target { + t.Errorf("readlink(%q): got (%q, %v), wanted (%q, nil)", abspath, gotTarget, err, link.target) + } } - // The device special file should be visible in the test mount namespace. - abspath = devPath + "/" + pathInDev - stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ - Root: root, - Start: root, - Path: fspath.Parse(abspath), - }, &vfs.StatOptions{ - Mask: linux.STATX_TYPE | linux.STATX_MODE, - }) - if err != nil { - t.Fatalf("failed to stat device file at %q: %v", abspath, err) + + dirs := []string{"shm", "pts"} + for _, dir := range dirs { + abspath := path.Join(devPath, dir) + statx, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(abspath), + }, &vfs.StatOptions{ + Mask: linux.STATX_MODE, + }) + if err != nil { + t.Errorf("stat(%q): got error %v ", abspath, err) + continue + } + if want := uint16(0755) | linux.S_IFDIR; statx.Mode != want { + t.Errorf("stat(%q): got mode %x, want %x", abspath, statx.Mode, want) + } } - if stat.Mode != wantMode { - t.Errorf("device file mode: got %v, wanted %v", stat.Mode, wantMode) +} + +func TestCreateDeviceFile(t *testing.T) { + ctx, creds, vfsObj, root, cleanup := setupDevtmpfs(t) + defer cleanup() + + a, err := NewAccessor(ctx, vfsObj, creds, "devtmpfs") + if err != nil { + t.Fatalf("failed to create devtmpfs.Accessor: %v", err) } - if stat.RdevMajor != major { - t.Errorf("major device number: got %v, wanted %v", stat.RdevMajor, major) + defer a.Release(ctx) + + devFiles := []struct { + path string + kind vfs.DeviceKind + major uint32 + minor uint32 + perms uint16 + }{ + { + path: "dummy", + kind: vfs.CharDevice, + major: 12, + minor: 34, + perms: 0600, + }, + { + path: "foo/bar", + kind: vfs.BlockDevice, + major: 13, + minor: 35, + perms: 0660, + }, + { + path: "foo/baz", + kind: vfs.CharDevice, + major: 12, + minor: 40, + perms: 0666, + }, + { + path: "a/b/c/d/e", + kind: vfs.BlockDevice, + major: 12, + minor: 34, + perms: 0600, + }, } - if stat.RdevMinor != minor { - t.Errorf("minor device number: got %v, wanted %v", stat.RdevMinor, minor) + + for _, f := range devFiles { + if err := a.CreateDeviceFile(ctx, f.path, f.kind, f.major, f.minor, f.perms); err != nil { + t.Fatalf("failed to create device file: %v", err) + } + // The device special file should be visible in the test mount namespace. + abspath := path.Join(devPath, f.path) + stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(abspath), + }, &vfs.StatOptions{ + Mask: linux.STATX_TYPE | linux.STATX_MODE, + }) + if err != nil { + t.Fatalf("failed to stat device file at %q: %v", abspath, err) + } + if stat.RdevMajor != f.major { + t.Errorf("major device number: got %v, wanted %v", stat.RdevMajor, f.major) + } + if stat.RdevMinor != f.minor { + t.Errorf("minor device number: got %v, wanted %v", stat.RdevMinor, f.minor) + } + wantMode := f.perms + switch f.kind { + case vfs.CharDevice: + wantMode |= linux.S_IFCHR + case vfs.BlockDevice: + wantMode |= linux.S_IFBLK + } + if stat.Mode != wantMode { + t.Errorf("device file mode: got %v, wanted %v", stat.Mode, wantMode) + } } } diff --git a/pkg/sentry/fsimpl/devtmpfs/save_restore.go b/pkg/sentry/fsimpl/devtmpfs/save_restore.go new file mode 100644 index 000000000..28832d850 --- /dev/null +++ b/pkg/sentry/fsimpl/devtmpfs/save_restore.go @@ -0,0 +1,23 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package devtmpfs + +// afterLoad is invoked by stateify. +func (fst *FilesystemType) afterLoad() { + if fst.fs != nil { + // Ensure that we don't create another filesystem. + fst.initOnce.Do(func() {}) + } +} diff --git a/pkg/sentry/fsimpl/eventfd/eventfd.go b/pkg/sentry/fsimpl/eventfd/eventfd.go index 812171fa3..5b29f2358 100644 --- a/pkg/sentry/fsimpl/eventfd/eventfd.go +++ b/pkg/sentry/fsimpl/eventfd/eventfd.go @@ -30,9 +30,11 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// EventFileDescription implements FileDescriptionImpl for file-based event +// EventFileDescription implements vfs.FileDescriptionImpl for file-based event // notification (eventfd). Eventfds are usually internal to the Sentry but in // certain situations they may be converted into a host-backed eventfd. +// +// +stateify savable type EventFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -41,7 +43,7 @@ type EventFileDescription struct { // queue is used to notify interested parties when the event object // becomes readable or writable. - queue waiter.Queue `state:"zerovalue"` + queue waiter.Queue // mu protects the fields below. mu sync.Mutex `state:"nosave"` @@ -106,7 +108,7 @@ func (efd *EventFileDescription) HostFD() (int, error) { return efd.hostfd, nil } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (efd *EventFileDescription) Release(context.Context) { efd.mu.Lock() defer efd.mu.Unlock() @@ -119,7 +121,7 @@ func (efd *EventFileDescription) Release(context.Context) { } } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { if dst.NumBytes() < 8 { return 0, syscall.EINVAL @@ -130,7 +132,7 @@ func (efd *EventFileDescription) Read(ctx context.Context, dst usermem.IOSequenc return 8, nil } -// Write implements FileDescriptionImpl.Write. +// Write implements vfs.FileDescriptionImpl.Write. func (efd *EventFileDescription) Write(ctx context.Context, src usermem.IOSequence, _ vfs.WriteOptions) (int64, error) { if src.NumBytes() < 8 { return 0, syscall.EINVAL diff --git a/pkg/sentry/fsimpl/ext/BUILD b/pkg/sentry/fsimpl/ext/BUILD index abc610ef3..7b1eec3da 100644 --- a/pkg/sentry/fsimpl/ext/BUILD +++ b/pkg/sentry/fsimpl/ext/BUILD @@ -51,6 +51,8 @@ go_library( "//pkg/fd", "//pkg/fspath", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs", @@ -86,9 +88,9 @@ go_test( library = ":ext", deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/fspath", + "//pkg/marshal/primitive", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/ext/disklayout", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go index 8f7d5a9bb..2ee7cc7ac 100644 --- a/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go +++ b/pkg/sentry/fsimpl/ext/benchmark/benchmark_test.go @@ -59,13 +59,18 @@ func setUp(b *testing.B, imagePath string) (context.Context, *vfs.VirtualFilesys vfsObj.MustRegisterFilesystemType("extfs", ext.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, imagePath, "extfs", &vfs.GetFilesystemOptions{InternalData: int(f.Fd())}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, imagePath, "extfs", &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: int(f.Fd()), + }, + }) if err != nil { f.Close() return nil, nil, nil, nil, err } root := mntns.Root() + root.IncRef() tearDown := func() { root.DecRef(ctx) @@ -90,7 +95,7 @@ func mount(b *testing.B, imagePath string, vfsfs *vfs.VirtualFilesystem, pop *vf ctx := contexttest.Context(b) creds := auth.CredentialsFromContext(ctx) - if err := vfsfs.MountAt(ctx, creds, imagePath, pop, "extfs", &vfs.MountOptions{ + if _, err := vfsfs.MountAt(ctx, creds, imagePath, pop, "extfs", &vfs.MountOptions{ GetFilesystemOptions: vfs.GetFilesystemOptions{ InternalData: int(f.Fd()), }, diff --git a/pkg/sentry/fsimpl/ext/block_map_file.go b/pkg/sentry/fsimpl/ext/block_map_file.go index 8bb104ff0..1165234f9 100644 --- a/pkg/sentry/fsimpl/ext/block_map_file.go +++ b/pkg/sentry/fsimpl/ext/block_map_file.go @@ -18,7 +18,7 @@ import ( "io" "math" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/syserror" ) @@ -34,19 +34,19 @@ type blockMapFile struct { // directBlks are the direct blocks numbers. The physical blocks pointed by // these holds file data. Contains file blocks 0 to 11. - directBlks [numDirectBlks]uint32 + directBlks [numDirectBlks]primitive.Uint32 // indirectBlk is the physical block which contains (blkSize/4) direct block // numbers (as uint32 integers). - indirectBlk uint32 + indirectBlk primitive.Uint32 // doubleIndirectBlk is the physical block which contains (blkSize/4) indirect // block numbers (as uint32 integers). - doubleIndirectBlk uint32 + doubleIndirectBlk primitive.Uint32 // tripleIndirectBlk is the physical block which contains (blkSize/4) doubly // indirect block numbers (as uint32 integers). - tripleIndirectBlk uint32 + tripleIndirectBlk primitive.Uint32 // coverage at (i)th index indicates the amount of file data a node at // height (i) covers. Height 0 is the direct block. @@ -68,10 +68,12 @@ func newBlockMapFile(args inodeArgs) (*blockMapFile, error) { } blkMap := file.regFile.inode.diskInode.Data() - binary.Unmarshal(blkMap[:numDirectBlks*4], binary.LittleEndian, &file.directBlks) - binary.Unmarshal(blkMap[numDirectBlks*4:(numDirectBlks+1)*4], binary.LittleEndian, &file.indirectBlk) - binary.Unmarshal(blkMap[(numDirectBlks+1)*4:(numDirectBlks+2)*4], binary.LittleEndian, &file.doubleIndirectBlk) - binary.Unmarshal(blkMap[(numDirectBlks+2)*4:(numDirectBlks+3)*4], binary.LittleEndian, &file.tripleIndirectBlk) + for i := 0; i < numDirectBlks; i++ { + file.directBlks[i].UnmarshalBytes(blkMap[i*4 : (i+1)*4]) + } + file.indirectBlk.UnmarshalBytes(blkMap[numDirectBlks*4 : (numDirectBlks+1)*4]) + file.doubleIndirectBlk.UnmarshalBytes(blkMap[(numDirectBlks+1)*4 : (numDirectBlks+2)*4]) + file.tripleIndirectBlk.UnmarshalBytes(blkMap[(numDirectBlks+2)*4 : (numDirectBlks+3)*4]) return file, nil } @@ -117,16 +119,16 @@ func (f *blockMapFile) ReadAt(dst []byte, off int64) (int, error) { switch { case offset < dirBlksEnd: // Direct block. - curR, err = f.read(f.directBlks[offset/f.regFile.inode.blkSize], offset%f.regFile.inode.blkSize, 0, dst[read:]) + curR, err = f.read(uint32(f.directBlks[offset/f.regFile.inode.blkSize]), offset%f.regFile.inode.blkSize, 0, dst[read:]) case offset < indirBlkEnd: // Indirect block. - curR, err = f.read(f.indirectBlk, offset-dirBlksEnd, 1, dst[read:]) + curR, err = f.read(uint32(f.indirectBlk), offset-dirBlksEnd, 1, dst[read:]) case offset < doubIndirBlkEnd: // Doubly indirect block. - curR, err = f.read(f.doubleIndirectBlk, offset-indirBlkEnd, 2, dst[read:]) + curR, err = f.read(uint32(f.doubleIndirectBlk), offset-indirBlkEnd, 2, dst[read:]) default: // Triply indirect block. - curR, err = f.read(f.tripleIndirectBlk, offset-doubIndirBlkEnd, 3, dst[read:]) + curR, err = f.read(uint32(f.tripleIndirectBlk), offset-doubIndirBlkEnd, 3, dst[read:]) } read += curR @@ -174,13 +176,13 @@ func (f *blockMapFile) read(curPhyBlk uint32, relFileOff uint64, height uint, ds read := 0 curChildOff := relFileOff % childCov for i := startIdx; i < endIdx; i++ { - var childPhyBlk uint32 + var childPhyBlk primitive.Uint32 err := readFromDisk(f.regFile.inode.fs.dev, curPhyBlkOff+int64(i*4), &childPhyBlk) if err != nil { return read, err } - n, err := f.read(childPhyBlk, curChildOff, height-1, dst[read:]) + n, err := f.read(uint32(childPhyBlk), curChildOff, height-1, dst[read:]) read += n if err != nil { return read, err diff --git a/pkg/sentry/fsimpl/ext/block_map_test.go b/pkg/sentry/fsimpl/ext/block_map_test.go index 6fa84e7aa..ed98b482e 100644 --- a/pkg/sentry/fsimpl/ext/block_map_test.go +++ b/pkg/sentry/fsimpl/ext/block_map_test.go @@ -20,7 +20,7 @@ import ( "testing" "github.com/google/go-cmp/cmp" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" ) @@ -87,29 +87,33 @@ func blockMapSetUp(t *testing.T) (*blockMapFile, []byte) { mockDisk := make([]byte, mockBMDiskSize) var fileData []byte blkNums := newBlkNumGen() - var data []byte + off := 0 + data := make([]byte, (numDirectBlks+3)*(*primitive.Uint32)(nil).SizeBytes()) // Write the direct blocks. for i := 0; i < numDirectBlks; i++ { - curBlkNum := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, curBlkNum) - fileData = append(fileData, writeFileDataToBlock(mockDisk, curBlkNum, 0, blkNums)...) + curBlkNum := primitive.Uint32(blkNums.next()) + curBlkNum.MarshalBytes(data[off:]) + off += curBlkNum.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(curBlkNum), 0, blkNums)...) } // Write to indirect block. - indirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, indirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, indirectBlk, 1, blkNums)...) - - // Write to indirect block. - doublyIndirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, doublyIndirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, doublyIndirectBlk, 2, blkNums)...) - - // Write to indirect block. - triplyIndirectBlk := blkNums.next() - data = binary.Marshal(data, binary.LittleEndian, triplyIndirectBlk) - fileData = append(fileData, writeFileDataToBlock(mockDisk, triplyIndirectBlk, 3, blkNums)...) + indirectBlk := primitive.Uint32(blkNums.next()) + indirectBlk.MarshalBytes(data[off:]) + off += indirectBlk.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(indirectBlk), 1, blkNums)...) + + // Write to double indirect block. + doublyIndirectBlk := primitive.Uint32(blkNums.next()) + doublyIndirectBlk.MarshalBytes(data[off:]) + off += doublyIndirectBlk.SizeBytes() + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(doublyIndirectBlk), 2, blkNums)...) + + // Write to triple indirect block. + triplyIndirectBlk := primitive.Uint32(blkNums.next()) + triplyIndirectBlk.MarshalBytes(data[off:]) + fileData = append(fileData, writeFileDataToBlock(mockDisk, uint32(triplyIndirectBlk), 3, blkNums)...) args := inodeArgs{ fs: &filesystem{ @@ -142,9 +146,9 @@ func writeFileDataToBlock(disk []byte, blkNum uint32, height uint, blkNums *blkN var fileData []byte for off := blkNum * mockBMBlkSize; off < (blkNum+1)*mockBMBlkSize; off += 4 { - curBlkNum := blkNums.next() - copy(disk[off:off+4], binary.Marshal(nil, binary.LittleEndian, curBlkNum)) - fileData = append(fileData, writeFileDataToBlock(disk, curBlkNum, height-1, blkNums)...) + curBlkNum := primitive.Uint32(blkNums.next()) + curBlkNum.MarshalBytes(disk[off : off+4]) + fileData = append(fileData, writeFileDataToBlock(disk, uint32(curBlkNum), height-1, blkNums)...) } return fileData } diff --git a/pkg/sentry/fsimpl/ext/dentry.go b/pkg/sentry/fsimpl/ext/dentry.go index 7a1b4219f..9bfed883a 100644 --- a/pkg/sentry/fsimpl/ext/dentry.go +++ b/pkg/sentry/fsimpl/ext/dentry.go @@ -20,6 +20,8 @@ import ( ) // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry diff --git a/pkg/sentry/fsimpl/ext/directory.go b/pkg/sentry/fsimpl/ext/directory.go index 0fc01668d..0ad79b381 100644 --- a/pkg/sentry/fsimpl/ext/directory.go +++ b/pkg/sentry/fsimpl/ext/directory.go @@ -16,7 +16,6 @@ package ext import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -28,6 +27,8 @@ import ( ) // directory represents a directory inode. It holds the childList in memory. +// +// +stateify savable type directory struct { inode inode @@ -39,7 +40,7 @@ type directory struct { // Lock Order (outermost locks must be taken first): // directory.mu // filesystem.mu - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // childList is a list containing (1) child dirents and (2) fake dirents // (with diskDirent == nil) that represent the iteration position of @@ -98,7 +99,7 @@ func newDirectory(args inodeArgs, newDirent bool) (*directory, error) { } else { curDirent.diskDirent = &disklayout.DirentOld{} } - binary.Unmarshal(buf, binary.LittleEndian, curDirent.diskDirent) + curDirent.diskDirent.UnmarshalBytes(buf) if curDirent.diskDirent.Inode() != 0 && len(curDirent.diskDirent.FileName()) != 0 { // Inode number and name length fields being set to 0 is used to indicate @@ -120,6 +121,8 @@ func (i *inode) isDir() bool { } // dirent is the directory.childList node. +// +// +stateify savable type dirent struct { diskDirent disklayout.Dirent @@ -129,6 +132,8 @@ type dirent struct { // directoryFD represents a directory file description. It implements // vfs.FileDescriptionImpl. +// +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl diff --git a/pkg/sentry/fsimpl/ext/disklayout/BUILD b/pkg/sentry/fsimpl/ext/disklayout/BUILD index 9bd9c76c0..d98a05dd8 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/BUILD +++ b/pkg/sentry/fsimpl/ext/disklayout/BUILD @@ -22,10 +22,11 @@ go_library( "superblock_old.go", "test_utils.go", ], + marshal = True, visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", + "//pkg/marshal", "//pkg/sentry/fs", "//pkg/sentry/kernel/auth", "//pkg/sentry/kernel/time", diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group.go b/pkg/sentry/fsimpl/ext/disklayout/block_group.go index ad6f4fef8..0d56ae9da 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + // BlockGroup represents a Linux ext block group descriptor. An ext file system // is split into a series of block groups. This provides an access layer to // information needed to access and use a block group. @@ -30,6 +34,8 @@ package disklayout // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/globals.html#block-group-descriptors. type BlockGroup interface { + marshal.Marshallable + // InodeTable returns the absolute block number of the block containing the // inode table. This points to an array of Inode structs. Inode tables are // statically allocated at mkfs time. The superblock records the number of diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go index 3e16c76db..a35fa22a0 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_32.go @@ -17,6 +17,8 @@ package disklayout // BlockGroup32Bit emulates the first half of struct ext4_group_desc in // fs/ext4/ext4.h. It is the block group descriptor struct for ext2, ext3 and // 32-bit ext4 filesystems. It implements BlockGroup interface. +// +// +marshal type BlockGroup32Bit struct { BlockBitmapLo uint32 InodeBitmapLo uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go index 9a809197a..d54d1d345 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_64.go @@ -18,6 +18,8 @@ package disklayout // It is the block group descriptor struct for 64-bit ext4 filesystems. // It implements BlockGroup interface. It is an extension of the 32-bit // version of BlockGroup. +// +// +marshal type BlockGroup64Bit struct { // We embed the 32-bit struct here because 64-bit version is just an extension // of the 32-bit version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go b/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go index 0ef4294c0..e4ce484e4 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/block_group_test.go @@ -21,6 +21,8 @@ import ( // TestBlockGroupSize tests that the block group descriptor structs are of the // correct size. func TestBlockGroupSize(t *testing.T) { - assertSize(t, BlockGroup32Bit{}, 32) - assertSize(t, BlockGroup64Bit{}, 64) + var bgSmall BlockGroup32Bit + assertSize(t, &bgSmall, 32) + var bgBig BlockGroup64Bit + assertSize(t, &bgBig, 64) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent.go b/pkg/sentry/fsimpl/ext/disklayout/dirent.go index 417b6cf65..568c8cb4c 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent.go @@ -15,6 +15,7 @@ package disklayout import ( + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fs" ) @@ -51,6 +52,8 @@ var ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#linear-classic-directories. type Dirent interface { + marshal.Marshallable + // Inode returns the absolute inode number of the underlying inode. // Inode number 0 signifies an unused dirent. Inode() uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go index 29ae4a5c2..51f9c2946 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_new.go @@ -29,12 +29,14 @@ import ( // Note: This struct can be of variable size on disk. The one described below // is of maximum size and the FileName beyond NameLength bytes might contain // garbage. +// +// +marshal type DirentNew struct { InodeNumber uint32 RecordLength uint16 NameLength uint8 FileTypeRaw uint8 - FileNameRaw [MaxFileName]byte + FileNameRaw [MaxFileName]byte `marshal:"unaligned"` } // Compiles only if DirentNew implements Dirent. diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go index 6fff12a6e..d4b19e086 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_old.go @@ -22,11 +22,13 @@ import "gvisor.dev/gvisor/pkg/sentry/fs" // Note: This struct can be of variable size on disk. The one described below // is of maximum size and the FileName beyond NameLength bytes might contain // garbage. +// +// +marshal type DirentOld struct { InodeNumber uint32 RecordLength uint16 NameLength uint16 - FileNameRaw [MaxFileName]byte + FileNameRaw [MaxFileName]byte `marshal:"unaligned"` } // Compiles only if DirentOld implements Dirent. diff --git a/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go b/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go index 934919f8a..3486864dc 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/dirent_test.go @@ -21,6 +21,8 @@ import ( // TestDirentSize tests that the dirent structs are of the correct // size. func TestDirentSize(t *testing.T) { - assertSize(t, DirentOld{}, uintptr(DirentSize)) - assertSize(t, DirentNew{}, uintptr(DirentSize)) + var dOld DirentOld + assertSize(t, &dOld, DirentSize) + var dNew DirentNew + assertSize(t, &dNew, DirentSize) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/disklayout.go b/pkg/sentry/fsimpl/ext/disklayout/disklayout.go index bdf4e2132..0834e9ba8 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/disklayout.go +++ b/pkg/sentry/fsimpl/ext/disklayout/disklayout.go @@ -36,8 +36,6 @@ // escape analysis on an unknown implementation at compile time. // // Notes: -// - All fields in these structs are exported because binary.Read would -// panic otherwise. // - All structures on disk are in little-endian order. Only jbd2 (journal) // structures are in big-endian order. // - All OS dependent fields in these structures will be interpretted using diff --git a/pkg/sentry/fsimpl/ext/disklayout/extent.go b/pkg/sentry/fsimpl/ext/disklayout/extent.go index 4110649ab..b13999bfc 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/extent.go +++ b/pkg/sentry/fsimpl/ext/disklayout/extent.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + // Extents were introduced in ext4 and provide huge performance gains in terms // data locality and reduced metadata block usage. Extents are organized in // extent trees. The root node is contained in inode.BlocksRaw. @@ -64,6 +68,8 @@ type ExtentNode struct { // ExtentEntry represents an extent tree node entry. The entry can either be // an ExtentIdx or Extent itself. This exists to simplify navigation logic. type ExtentEntry interface { + marshal.Marshallable + // FileBlock returns the first file block number covered by this entry. FileBlock() uint32 @@ -75,6 +81,8 @@ type ExtentEntry interface { // tree node begins with this and is followed by `NumEntries` number of: // - Extent if `Depth` == 0 // - ExtentIdx otherwise +// +// +marshal type ExtentHeader struct { // Magic in the extent magic number, must be 0xf30a. Magic uint16 @@ -96,6 +104,8 @@ type ExtentHeader struct { // internal nodes. Sorted in ascending order based on FirstFileBlock since // Linux does a binary search on this. This points to a block containing the // child node. +// +// +marshal type ExtentIdx struct { FirstFileBlock uint32 ChildBlockLo uint32 @@ -121,6 +131,8 @@ func (ei *ExtentIdx) PhysicalBlock() uint64 { // nodes. Sorted in ascending order based on FirstFileBlock since Linux does a // binary search on this. This points to an array of data blocks containing the // file data. It covers `Length` data blocks starting from `StartBlock`. +// +// +marshal type Extent struct { FirstFileBlock uint32 Length uint16 diff --git a/pkg/sentry/fsimpl/ext/disklayout/extent_test.go b/pkg/sentry/fsimpl/ext/disklayout/extent_test.go index 8762b90db..c96002e19 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/extent_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/extent_test.go @@ -21,7 +21,10 @@ import ( // TestExtentSize tests that the extent structs are of the correct // size. func TestExtentSize(t *testing.T) { - assertSize(t, ExtentHeader{}, ExtentHeaderSize) - assertSize(t, ExtentIdx{}, ExtentEntrySize) - assertSize(t, Extent{}, ExtentEntrySize) + var h ExtentHeader + assertSize(t, &h, ExtentHeaderSize) + var i ExtentIdx + assertSize(t, &i, ExtentEntrySize) + var e Extent + assertSize(t, &e, ExtentEntrySize) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode.go b/pkg/sentry/fsimpl/ext/disklayout/inode.go index 88ae913f5..ef25040a9 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode.go @@ -16,6 +16,7 @@ package disklayout import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/time" ) @@ -38,6 +39,8 @@ const ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#index-nodes. type Inode interface { + marshal.Marshallable + // Mode returns the linux file mode which is majorly used to extract // information like: // - File permissions (read/write/execute by user/group/others). diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_new.go b/pkg/sentry/fsimpl/ext/disklayout/inode_new.go index 8f9f574ce..a4503f5cf 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_new.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_new.go @@ -27,6 +27,8 @@ import "gvisor.dev/gvisor/pkg/sentry/kernel/time" // are used to provide nanoscond precision. Hence, these timestamps will now // overflow in May 2446. // See https://www.kernel.org/doc/html/latest/filesystems/ext4/dynamic.html#inode-timestamps. +// +// +marshal type InodeNew struct { InodeOld diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_old.go b/pkg/sentry/fsimpl/ext/disklayout/inode_old.go index db25b11b6..e6b28babf 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_old.go @@ -30,6 +30,8 @@ const ( // // All fields representing time are in seconds since the epoch. Which means that // they will overflow in January 2038. +// +// +marshal type InodeOld struct { ModeRaw uint16 UIDLo uint16 diff --git a/pkg/sentry/fsimpl/ext/disklayout/inode_test.go b/pkg/sentry/fsimpl/ext/disklayout/inode_test.go index dd03ee50e..90744e956 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/inode_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/inode_test.go @@ -24,10 +24,12 @@ import ( // TestInodeSize tests that the inode structs are of the correct size. func TestInodeSize(t *testing.T) { - assertSize(t, InodeOld{}, OldInodeSize) + var iOld InodeOld + assertSize(t, &iOld, OldInodeSize) // This was updated from 156 bytes to 160 bytes in Oct 2015. - assertSize(t, InodeNew{}, 160) + var iNew InodeNew + assertSize(t, &iNew, 160) } // TestTimestampSeconds tests that the seconds part of [a/c/m] timestamps in diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock.go b/pkg/sentry/fsimpl/ext/disklayout/superblock.go index 8bb327006..70948ebe9 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock.go @@ -14,6 +14,10 @@ package disklayout +import ( + "gvisor.dev/gvisor/pkg/marshal" +) + const ( // SbOffset is the absolute offset at which the superblock is placed. SbOffset = 1024 @@ -38,6 +42,8 @@ const ( // // See https://www.kernel.org/doc/html/latest/filesystems/ext4/globals.html#super-block. type SuperBlock interface { + marshal.Marshallable + // InodesCount returns the total number of inodes in this filesystem. InodesCount() uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go index 53e515fd3..4dc6080fb 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_32.go @@ -17,6 +17,8 @@ package disklayout // SuperBlock32Bit implements SuperBlock and represents the 32-bit version of // the ext4_super_block struct in fs/ext4/ext4.h. Should be used only if // RevLevel = DynamicRev and 64-bit feature is disabled. +// +// +marshal type SuperBlock32Bit struct { // We embed the old superblock struct here because the 32-bit version is just // an extension of the old version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go index 7c1053fb4..2c9039327 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_64.go @@ -19,6 +19,8 @@ package disklayout // 1024 bytes (smallest possible block size) and hence the superblock always // fits in no more than one data block. Should only be used when the 64-bit // feature is set. +// +// +marshal type SuperBlock64Bit struct { // We embed the 32-bit struct here because 64-bit version is just an extension // of the 32-bit version. diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go index 9221e0251..e4709f23c 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_old.go @@ -16,6 +16,8 @@ package disklayout // SuperBlockOld implements SuperBlock and represents the old version of the // superblock struct. Should be used only if RevLevel = OldRev. +// +// +marshal type SuperBlockOld struct { InodesCountRaw uint32 BlocksCountLo uint32 diff --git a/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go b/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go index 463b5ba21..b734b6987 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go +++ b/pkg/sentry/fsimpl/ext/disklayout/superblock_test.go @@ -21,7 +21,10 @@ import ( // TestSuperBlockSize tests that the superblock structs are of the correct // size. func TestSuperBlockSize(t *testing.T) { - assertSize(t, SuperBlockOld{}, 84) - assertSize(t, SuperBlock32Bit{}, 336) - assertSize(t, SuperBlock64Bit{}, 1024) + var sbOld SuperBlockOld + assertSize(t, &sbOld, 84) + var sb32 SuperBlock32Bit + assertSize(t, &sb32, 336) + var sb64 SuperBlock64Bit + assertSize(t, &sb64, 1024) } diff --git a/pkg/sentry/fsimpl/ext/disklayout/test_utils.go b/pkg/sentry/fsimpl/ext/disklayout/test_utils.go index 9c63f04c0..a4bc08411 100644 --- a/pkg/sentry/fsimpl/ext/disklayout/test_utils.go +++ b/pkg/sentry/fsimpl/ext/disklayout/test_utils.go @@ -18,13 +18,13 @@ import ( "reflect" "testing" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" ) -func assertSize(t *testing.T, v interface{}, want uintptr) { +func assertSize(t *testing.T, v marshal.Marshallable, want int) { t.Helper() - if got := binary.Size(v); got != want { + if got := v.SizeBytes(); got != want { t.Errorf("struct %s should be exactly %d bytes but is %d bytes", reflect.TypeOf(v).Name(), want, got) } } diff --git a/pkg/sentry/fsimpl/ext/ext.go b/pkg/sentry/fsimpl/ext/ext.go index 08ffc2834..38fb7962b 100644 --- a/pkg/sentry/fsimpl/ext/ext.go +++ b/pkg/sentry/fsimpl/ext/ext.go @@ -34,11 +34,10 @@ import ( const Name = "ext" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} -// Compiles only if FilesystemType implements vfs.FilesystemType. -var _ vfs.FilesystemType = (*FilesystemType)(nil) - // getDeviceFd returns an io.ReaderAt to the underlying device. // Currently there are two ways of mounting an ext(2/3/4) fs: // 1. Specify a mount with our internal special MountType in the OCI spec. @@ -99,6 +98,9 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { // TODO(b/134676337): Ensure that the user is mounting readonly. If not, diff --git a/pkg/sentry/fsimpl/ext/ext_test.go b/pkg/sentry/fsimpl/ext/ext_test.go index 2dbaee287..d9fd4590c 100644 --- a/pkg/sentry/fsimpl/ext/ext_test.go +++ b/pkg/sentry/fsimpl/ext/ext_test.go @@ -71,13 +71,18 @@ func setUp(t *testing.T, imagePath string) (context.Context, *vfs.VirtualFilesys vfsObj.MustRegisterFilesystemType("extfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, localImagePath, "extfs", &vfs.GetFilesystemOptions{InternalData: int(f.Fd())}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, localImagePath, "extfs", &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: int(f.Fd()), + }, + }) if err != nil { f.Close() return nil, nil, nil, nil, err } root := mntns.Root() + root.IncRef() tearDown := func() { root.DecRef(ctx) diff --git a/pkg/sentry/fsimpl/ext/extent_file.go b/pkg/sentry/fsimpl/ext/extent_file.go index c36225a7c..778460107 100644 --- a/pkg/sentry/fsimpl/ext/extent_file.go +++ b/pkg/sentry/fsimpl/ext/extent_file.go @@ -18,12 +18,13 @@ import ( "io" "sort" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" "gvisor.dev/gvisor/pkg/syserror" ) // extentFile is a type of regular file which uses extents to store file data. +// +// +stateify savable type extentFile struct { regFile regularFile @@ -58,7 +59,7 @@ func newExtentFile(args inodeArgs) (*extentFile, error) { func (f *extentFile) buildExtTree() error { rootNodeData := f.regFile.inode.diskInode.Data() - binary.Unmarshal(rootNodeData[:disklayout.ExtentHeaderSize], binary.LittleEndian, &f.root.Header) + f.root.Header.UnmarshalBytes(rootNodeData[:disklayout.ExtentHeaderSize]) // Root node can not have more than 4 entries: 60 bytes = 1 header + 4 entries. if f.root.Header.NumEntries > 4 { @@ -77,7 +78,7 @@ func (f *extentFile) buildExtTree() error { // Internal node. curEntry = &disklayout.ExtentIdx{} } - binary.Unmarshal(rootNodeData[off:off+disklayout.ExtentEntrySize], binary.LittleEndian, curEntry) + curEntry.UnmarshalBytes(rootNodeData[off : off+disklayout.ExtentEntrySize]) f.root.Entries[i].Entry = curEntry } diff --git a/pkg/sentry/fsimpl/ext/extent_test.go b/pkg/sentry/fsimpl/ext/extent_test.go index cd10d46ee..985f76ac0 100644 --- a/pkg/sentry/fsimpl/ext/extent_test.go +++ b/pkg/sentry/fsimpl/ext/extent_test.go @@ -21,7 +21,6 @@ import ( "github.com/google/go-cmp/cmp" "github.com/google/go-cmp/cmp/cmpopts" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" ) @@ -202,13 +201,14 @@ func extentTreeSetUp(t *testing.T, root *disklayout.ExtentNode) (*extentFile, [] // writeTree writes the tree represented by `root` to the inode and disk. It // also writes random file data on disk. func writeTree(in *inode, disk []byte, root *disklayout.ExtentNode, mockExtentBlkSize uint64) []byte { - rootData := binary.Marshal(nil, binary.LittleEndian, root.Header) + rootData := in.diskInode.Data() + root.Header.MarshalBytes(rootData) + off := root.Header.SizeBytes() for _, ep := range root.Entries { - rootData = binary.Marshal(rootData, binary.LittleEndian, ep.Entry) + ep.Entry.MarshalBytes(rootData[off:]) + off += ep.Entry.SizeBytes() } - copy(in.diskInode.Data(), rootData) - var fileData []byte for _, ep := range root.Entries { if root.Header.Height == 0 { @@ -223,13 +223,14 @@ func writeTree(in *inode, disk []byte, root *disklayout.ExtentNode, mockExtentBl // writeTreeToDisk is the recursive step for writeTree which writes the tree // on the disk only. Also writes random file data on disk. func writeTreeToDisk(disk []byte, curNode disklayout.ExtentEntryPair) []byte { - nodeData := binary.Marshal(nil, binary.LittleEndian, curNode.Node.Header) + nodeData := disk[curNode.Entry.PhysicalBlock()*mockExtentBlkSize:] + curNode.Node.Header.MarshalBytes(nodeData) + off := curNode.Node.Header.SizeBytes() for _, ep := range curNode.Node.Entries { - nodeData = binary.Marshal(nodeData, binary.LittleEndian, ep.Entry) + ep.Entry.MarshalBytes(nodeData[off:]) + off += ep.Entry.SizeBytes() } - copy(disk[curNode.Entry.PhysicalBlock()*mockExtentBlkSize:], nodeData) - var fileData []byte for _, ep := range curNode.Node.Entries { if curNode.Node.Header.Height == 0 { diff --git a/pkg/sentry/fsimpl/ext/filesystem.go b/pkg/sentry/fsimpl/ext/filesystem.go index c714ddf73..917f1873d 100644 --- a/pkg/sentry/fsimpl/ext/filesystem.go +++ b/pkg/sentry/fsimpl/ext/filesystem.go @@ -38,11 +38,13 @@ var ( ) // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem // mu serializes changes to the Dentry tree. - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` // dev represents the underlying fs device. It does not require protection // because io.ReaderAt permits concurrent read calls to it. It translates to @@ -81,9 +83,9 @@ var _ vfs.FilesystemImpl = (*filesystem)(nil) // stepLocked is loosely analogous to fs/namei.c:walk_component(). // // Preconditions: -// - filesystem.mu must be locked (for writing if write param is true). -// - !rp.Done(). -// - inode == vfsd.Impl().(*Dentry).inode. +// * filesystem.mu must be locked (for writing if write param is true). +// * !rp.Done(). +// * inode == vfsd.Impl().(*Dentry).inode. func stepLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, inode *inode, write bool) (*vfs.Dentry, *inode, error) { if !inode.isDir() { return nil, nil, syserror.ENOTDIR @@ -166,7 +168,7 @@ func stepLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, in // walkLocked is loosely analogous to Linux's fs/namei.c:path_lookupat(). // // Preconditions: -// - filesystem.mu must be locked (for writing if write param is true). +// * filesystem.mu must be locked (for writing if write param is true). func walkLocked(ctx context.Context, rp *vfs.ResolvingPath, write bool) (*vfs.Dentry, *inode, error) { vfsd := rp.Start() inode := vfsd.Impl().(*dentry).inode @@ -194,8 +196,8 @@ func walkLocked(ctx context.Context, rp *vfs.ResolvingPath, write bool) (*vfs.De // walkParentLocked is loosely analogous to Linux's fs/namei.c:path_parentat(). // // Preconditions: -// - filesystem.mu must be locked (for writing if write param is true). -// - !rp.Done(). +// * filesystem.mu must be locked (for writing if write param is true). +// * !rp.Done(). func walkParentLocked(ctx context.Context, rp *vfs.ResolvingPath, write bool) (*vfs.Dentry, *inode, error) { vfsd := rp.Start() inode := vfsd.Impl().(*dentry).inode @@ -490,7 +492,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { _, inode, err := fs.walk(ctx, rp, false) if err != nil { @@ -504,8 +506,8 @@ func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath return nil, syserror.ECONNREFUSED } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { _, _, err := fs.walk(ctx, rp, false) if err != nil { return nil, err @@ -513,8 +515,8 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, si return nil, syserror.ENOTSUP } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { _, _, err := fs.walk(ctx, rp, false) if err != nil { return "", err @@ -522,8 +524,8 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt return "", syserror.ENOTSUP } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { _, _, err := fs.walk(ctx, rp, false) if err != nil { return err @@ -531,8 +533,8 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt return syserror.ENOTSUP } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { _, _, err := fs.walk(ctx, rp, false) if err != nil { return err diff --git a/pkg/sentry/fsimpl/ext/inode.go b/pkg/sentry/fsimpl/ext/inode.go index 30636cf66..9009ba3c7 100644 --- a/pkg/sentry/fsimpl/ext/inode.go +++ b/pkg/sentry/fsimpl/ext/inode.go @@ -37,6 +37,8 @@ import ( // |-- regular-- // |-- extent file // |-- block map file +// +// +stateify savable type inode struct { // refs is a reference count. refs is accessed using atomic memory operations. refs int64 diff --git a/pkg/sentry/fsimpl/ext/regular_file.go b/pkg/sentry/fsimpl/ext/regular_file.go index e73e740d6..4a5539b37 100644 --- a/pkg/sentry/fsimpl/ext/regular_file.go +++ b/pkg/sentry/fsimpl/ext/regular_file.go @@ -31,6 +31,8 @@ import ( // regularFile represents a regular file's inode. This too follows the // inheritance pattern prevelant in the vfs layer described in // pkg/sentry/vfs/README.md. +// +// +stateify savable type regularFile struct { inode inode @@ -67,6 +69,8 @@ func (in *inode) isRegular() bool { // directoryFD represents a directory file description. It implements // vfs.FileDescriptionImpl. +// +// +stateify savable type regularFileFD struct { fileDescription vfs.LockFD @@ -75,7 +79,7 @@ type regularFileFD struct { off int64 // offMu serializes operations that may mutate off. - offMu sync.Mutex + offMu sync.Mutex `state:"nosave"` } // Release implements vfs.FileDescriptionImpl.Release. diff --git a/pkg/sentry/fsimpl/ext/symlink.go b/pkg/sentry/fsimpl/ext/symlink.go index 2fd0d1fa8..5e2bcc837 100644 --- a/pkg/sentry/fsimpl/ext/symlink.go +++ b/pkg/sentry/fsimpl/ext/symlink.go @@ -23,6 +23,8 @@ import ( ) // symlink represents a symlink inode. +// +// +stateify savable type symlink struct { inode inode target string // immutable @@ -61,9 +63,11 @@ func (in *inode) isSymlink() bool { return ok } -// symlinkFD represents a symlink file description and implements implements +// symlinkFD represents a symlink file description and implements // vfs.FileDescriptionImpl. which may only be used if open options contains // O_PATH. For this reason most of the functions return EBADF. +// +// +stateify savable type symlinkFD struct { fileDescription vfs.NoLockFD diff --git a/pkg/sentry/fsimpl/ext/utils.go b/pkg/sentry/fsimpl/ext/utils.go index d8b728f8c..58ef7b9b8 100644 --- a/pkg/sentry/fsimpl/ext/utils.go +++ b/pkg/sentry/fsimpl/ext/utils.go @@ -17,21 +17,21 @@ package ext import ( "io" - "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fsimpl/ext/disklayout" "gvisor.dev/gvisor/pkg/syserror" ) // readFromDisk performs a binary read from disk into the given struct from // the absolute offset provided. -func readFromDisk(dev io.ReaderAt, abOff int64, v interface{}) error { - n := binary.Size(v) +func readFromDisk(dev io.ReaderAt, abOff int64, v marshal.Marshallable) error { + n := v.SizeBytes() buf := make([]byte, n) if read, _ := dev.ReadAt(buf, abOff); read < int(n) { return syserror.EIO } - binary.Unmarshal(buf, binary.LittleEndian, v) + v.UnmarshalBytes(buf) return nil } diff --git a/pkg/sentry/fsimpl/fuse/BUILD b/pkg/sentry/fsimpl/fuse/BUILD index 999111deb..2158b1bbc 100644 --- a/pkg/sentry/fsimpl/fuse/BUILD +++ b/pkg/sentry/fsimpl/fuse/BUILD @@ -15,21 +15,42 @@ go_template_instance( }, ) +go_template_instance( + name = "inode_refs", + out = "inode_refs.go", + package = "fuse", + prefix = "inode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "inode", + }, +) + go_library( name = "fuse", srcs = [ "connection.go", + "connection_control.go", "dev.go", + "directory.go", + "file.go", "fusefs.go", - "init.go", + "inode_refs.go", + "read_write.go", "register.go", + "regular_file.go", "request_list.go", + "request_response.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", "//pkg/log", + "//pkg/marshal", + "//pkg/refs", + "//pkg/refsvfs2", + "//pkg/safemem", "//pkg/sentry/fsimpl/devtmpfs", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", @@ -39,7 +60,6 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", "@org_golang_x_sys//unix:go_default_library", ], ) @@ -47,10 +67,15 @@ go_library( go_test( name = "fuse_test", size = "small", - srcs = ["dev_test.go"], + srcs = [ + "connection_test.go", + "dev_test.go", + "utils_test.go", + ], library = ":fuse", deps = [ "//pkg/abi/linux", + "//pkg/marshal", "//pkg/sentry/fsimpl/testutil", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", @@ -58,6 +83,5 @@ go_test( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/fsimpl/fuse/connection.go b/pkg/sentry/fsimpl/fuse/connection.go index 6df2728ab..8ccda1264 100644 --- a/pkg/sentry/fsimpl/fuse/connection.go +++ b/pkg/sentry/fsimpl/fuse/connection.go @@ -15,31 +15,17 @@ package fuse import ( - "errors" - "fmt" "sync" - "sync/atomic" - "syscall" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/kernel" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) -// maxActiveRequestsDefault is the default setting controlling the upper bound -// on the number of active requests at any given time. -const maxActiveRequestsDefault = 10000 - -// Ordinary requests have even IDs, while interrupts IDs are odd. -// Used to increment the unique ID for each FUSE request. -var reqIDStep uint64 = 2 - const ( // fuseDefaultMaxBackground is the default value for MaxBackground. fuseDefaultMaxBackground = 12 @@ -52,43 +38,39 @@ const ( fuseDefaultMaxPagesPerReq = 32 ) -// Request represents a FUSE operation request that hasn't been sent to the -// server yet. +// connection is the struct by which the sentry communicates with the FUSE server daemon. // -// +stateify savable -type Request struct { - requestEntry - - id linux.FUSEOpID - hdr *linux.FUSEHeaderIn - data []byte -} - -// Response represents an actual response from the server, including the -// response payload. +// Lock order: +// - conn.fd.mu +// - conn.mu +// - conn.asyncMu // // +stateify savable -type Response struct { - opcode linux.FUSEOpcode - hdr linux.FUSEHeaderOut - data []byte -} - -// connection is the struct by which the sentry communicates with the FUSE server daemon. type connection struct { fd *DeviceFD + // mu protects access to struct memebers. + mu sync.Mutex `state:"nosave"` + + // attributeVersion is the version of connection's attributes. + attributeVersion uint64 + + // We target FUSE 7.23. // The following FUSE_INIT flags are currently unsupported by this implementation: - // - FUSE_ATOMIC_O_TRUNC: requires open(..., O_TRUNC) // - FUSE_EXPORT_SUPPORT - // - FUSE_HANDLE_KILLPRIV // - FUSE_POSIX_LOCKS: requires POSIX locks // - FUSE_FLOCK_LOCKS: requires POSIX locks // - FUSE_AUTO_INVAL_DATA: requires page caching eviction - // - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction // - FUSE_DO_READDIRPLUS/FUSE_READDIRPLUS_AUTO: requires FUSE_READDIRPLUS implementation // - FUSE_ASYNC_DIO - // - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler + // - FUSE_PARALLEL_DIROPS (7.25) + // - FUSE_HANDLE_KILLPRIV (7.26) + // - FUSE_POSIX_ACL: affects defaultPermissions, posixACL, xattr handler (7.26) + // - FUSE_ABORT_ERROR (7.27) + // - FUSE_CACHE_SYMLINKS (7.28) + // - FUSE_NO_OPENDIR_SUPPORT (7.29) + // - FUSE_EXPLICIT_INVAL_DATA: requires page caching eviction (7.30) + // - FUSE_MAP_ALIGNMENT (7.31) // initialized after receiving FUSE_INIT reply. // Until it's set, suspend sending FUSE requests. @@ -96,11 +78,7 @@ type connection struct { initialized int32 // initializedChan is used to block requests before initialization. - initializedChan chan struct{} - - // blocked when there are too many outstading backgrounds requests (NumBackground == MaxBackground). - // TODO(gvisor.dev/issue/3185): update the numBackground accordingly; use a channel to block. - blocked bool + initializedChan chan struct{} `state:".(bool)"` // connected (connection established) when a new FUSE file system is created. // Set to false when: @@ -109,48 +87,55 @@ type connection struct { // device release. connected bool - // aborted via sysfs. - // TODO(gvisor.dev/issue/3185): abort all queued requests. - aborted bool - // connInitError if FUSE_INIT encountered error (major version mismatch). // Only set in INIT. connInitError bool // connInitSuccess if FUSE_INIT is successful. // Only set in INIT. - // Used for destory. + // Used for destory (not yet implemented). connInitSuccess bool - // TODO(gvisor.dev/issue/3185): All the queue logic are working in progress. - - // NumberBackground is the number of requests in the background. - numBackground uint16 + // aborted via sysfs, and will send ECONNABORTED to read after disconnection (instead of ENODEV). + // Set only if abortErr is true and via fuse control fs (not yet implemented). + // TODO(gvisor.dev/issue/3525): set this to true when user aborts. + aborted bool - // congestionThreshold for NumBackground. - // Negotiated in FUSE_INIT. - congestionThreshold uint16 + // numWating is the number of requests waiting to be + // sent to FUSE device or being processed by FUSE daemon. + numWaiting uint32 - // maxBackground is the maximum number of NumBackground. - // Block connection when it is reached. - // Negotiated in FUSE_INIT. - maxBackground uint16 + // Terminology note: + // + // - `asyncNumMax` is the `MaxBackground` in the FUSE_INIT_IN struct. + // + // - `asyncCongestionThreshold` is the `CongestionThreshold` in the FUSE_INIT_IN struct. + // + // We call the "background" requests in unix term as async requests. + // The "async requests" in unix term is our async requests that expect a reply, + // i.e. `!request.noReply` - // numActiveBackground is the number of requests in background and has being marked as active. - numActiveBackground uint16 + // asyncMu protects the async request fields. + asyncMu sync.Mutex `state:"nosave"` - // numWating is the number of requests waiting for completion. - numWaiting uint32 + // asyncNum is the number of async requests. + // Protected by asyncMu. + asyncNum uint16 - // TODO(gvisor.dev/issue/3185): BgQueue - // some queue for background queued requests. + // asyncCongestionThreshold the number of async requests. + // Negotiated in FUSE_INIT as "CongestionThreshold". + // TODO(gvisor.dev/issue/3529): add congestion control. + // Protected by asyncMu. + asyncCongestionThreshold uint16 - // bgLock protects: - // MaxBackground, CongestionThreshold, NumBackground, - // NumActiveBackground, BgQueue, Blocked. - bgLock sync.Mutex + // asyncNumMax is the maximum number of asyncNum. + // Connection blocks the async requests when it is reached. + // Negotiated in FUSE_INIT as "MaxBackground". + // Protected by asyncMu. + asyncNumMax uint16 // maxRead is the maximum size of a read buffer in in bytes. + // Initialized from a fuse fs parameter. maxRead uint32 // maxWrite is the maximum size of a write buffer in bytes. @@ -165,23 +150,20 @@ type connection struct { // Negotiated and only set in INIT. minor uint32 - // asyncRead if read pages asynchronously. + // atomicOTrunc is true when FUSE does not send a separate SETATTR request + // before open with O_TRUNC flag. // Negotiated and only set in INIT. - asyncRead bool + atomicOTrunc bool - // abortErr is true if kernel need to return an unique read error after abort. + // asyncRead if read pages asynchronously. // Negotiated and only set in INIT. - abortErr bool + asyncRead bool // writebackCache is true for write-back cache policy, // false for write-through policy. // Negotiated and only set in INIT. writebackCache bool - // cacheSymlinks if filesystem needs to cache READLINK responses in page cache. - // Negotiated and only set in INIT. - cacheSymlinks bool - // bigWrites if doing multi-page cached writes. // Negotiated and only set in INIT. bigWrites bool @@ -189,116 +171,86 @@ type connection struct { // dontMask if filestestem does not apply umask to creation modes. // Negotiated in INIT. dontMask bool + + // noOpen if FUSE server doesn't support open operation. + // This flag only influence performance, not correctness of the program. + noOpen bool +} + +func (conn *connection) saveInitializedChan() bool { + select { + case <-conn.initializedChan: + return true // Closed. + default: + return false // Not closed. + } +} + +func (conn *connection) loadInitializedChan(closed bool) { + conn.initializedChan = make(chan struct{}, 1) + if closed { + close(conn.initializedChan) + } } // newFUSEConnection creates a FUSE connection to fd. -func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, maxInFlightRequests uint64) (*connection, error) { +func newFUSEConnection(_ context.Context, fd *vfs.FileDescription, opts *filesystemOptions) (*connection, error) { // Mark the device as ready so it can be used. /dev/fuse can only be used if the FD was used to // mount a FUSE filesystem. fuseFD := fd.Impl().(*DeviceFD) - fuseFD.mounted = true // Create the writeBuf for the header to be stored in. hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) fuseFD.writeBuf = make([]byte, hdrLen) fuseFD.completions = make(map[linux.FUSEOpID]*futureResponse) - fuseFD.fullQueueCh = make(chan struct{}, maxInFlightRequests) + fuseFD.fullQueueCh = make(chan struct{}, opts.maxActiveRequests) fuseFD.writeCursor = 0 return &connection{ - fd: fuseFD, - maxBackground: fuseDefaultMaxBackground, - congestionThreshold: fuseDefaultCongestionThreshold, - maxPages: fuseDefaultMaxPagesPerReq, - initializedChan: make(chan struct{}), - connected: true, - }, nil -} - -// SetInitialized atomically sets the connection as initialized. -func (conn *connection) SetInitialized() { - // Unblock the requests sent before INIT. - close(conn.initializedChan) - - // Close the channel first to avoid the non-atomic situation - // where conn.initialized is true but there are - // tasks being blocked on the channel. - // And it prevents the newer tasks from gaining - // unnecessary higher chance to be issued before the blocked one. - - atomic.StoreInt32(&(conn.initialized), int32(1)) -} - -// IsInitialized atomically check if the connection is initialized. -// pairs with SetInitialized(). -func (conn *connection) Initialized() bool { - return atomic.LoadInt32(&(conn.initialized)) != 0 -} - -// NewRequest creates a new request that can be sent to the FUSE server. -func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) { - conn.fd.mu.Lock() - defer conn.fd.mu.Unlock() - conn.fd.nextOpID += linux.FUSEOpID(reqIDStep) - - hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes() - hdr := linux.FUSEHeaderIn{ - Len: uint32(hdrLen + payload.SizeBytes()), - Opcode: opcode, - Unique: conn.fd.nextOpID, - NodeID: ino, - UID: uint32(creds.EffectiveKUID), - GID: uint32(creds.EffectiveKGID), - PID: pid, - } - - buf := make([]byte, hdr.Len) - hdr.MarshalUnsafe(buf[:hdrLen]) - payload.MarshalUnsafe(buf[hdrLen:]) - - return &Request{ - id: hdr.Unique, - hdr: &hdr, - data: buf, + fd: fuseFD, + asyncNumMax: fuseDefaultMaxBackground, + asyncCongestionThreshold: fuseDefaultCongestionThreshold, + maxRead: opts.maxRead, + maxPages: fuseDefaultMaxPagesPerReq, + initializedChan: make(chan struct{}), + connected: true, }, nil } -// Call makes a request to the server and blocks the invoking task until a -// server responds with a response. Task should never be nil. -// Requests will not be sent before the connection is initialized. -// For async tasks, use CallAsync(). -func (conn *connection) Call(t *kernel.Task, r *Request) (*Response, error) { - // Block requests sent before connection is initalized. - if !conn.Initialized() { - if err := t.Block(conn.initializedChan); err != nil { - return nil, err - } - } - - return conn.call(t, r) +// CallAsync makes an async (aka background) request. +// It's a simple wrapper around Call(). +func (conn *connection) CallAsync(t *kernel.Task, r *Request) error { + r.async = true + _, err := conn.Call(t, r) + return err } -// CallAsync makes an async (aka background) request. -// Those requests either do not expect a response (e.g. release) or -// the response should be handled by others (e.g. init). -// Return immediately unless the connection is blocked (before initialization). -// Async call example: init, release, forget, aio, interrupt. +// Call makes a request to the server. +// Block before the connection is initialized. // When the Request is FUSE_INIT, it will not be blocked before initialization. -func (conn *connection) CallAsync(t *kernel.Task, r *Request) error { +// Task should never be nil. +// +// For a sync request, it blocks the invoking task until +// a server responds with a response. +// +// For an async request (that do not expect a response immediately), +// it returns directly unless being blocked either before initialization +// or when there are too many async requests ongoing. +// +// Example for async request: +// init, readahead, write, async read/write, fuse_notify_reply, +// non-sync release, interrupt, forget. +// +// The forget request does not have a reply, +// as documented in include/uapi/linux/fuse.h:FUSE_FORGET. +func (conn *connection) Call(t *kernel.Task, r *Request) (*Response, error) { // Block requests sent before connection is initalized. if !conn.Initialized() && r.hdr.Opcode != linux.FUSE_INIT { if err := t.Block(conn.initializedChan); err != nil { - return err + return nil, err } } - // This should be the only place that invokes call() with a nil task. - _, err := conn.call(nil, r) - return err -} - -// call makes a call without blocking checks. -func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) { if !conn.connected { return nil, syserror.ENOTCONN } @@ -315,31 +267,6 @@ func (conn *connection) call(t *kernel.Task, r *Request) (*Response, error) { return fut.resolve(t) } -// Error returns the error of the FUSE call. -func (r *Response) Error() error { - errno := r.hdr.Error - if errno >= 0 { - return nil - } - - sysErrNo := syscall.Errno(-errno) - return error(sysErrNo) -} - -// UnmarshalPayload unmarshals the response data into m. -func (r *Response) UnmarshalPayload(m marshal.Marshallable) error { - hdrLen := r.hdr.SizeBytes() - haveDataLen := r.hdr.Len - uint32(hdrLen) - wantDataLen := uint32(m.SizeBytes()) - - if haveDataLen < wantDataLen { - return fmt.Errorf("payload too small. Minimum data lenth required: %d, but got data length %d", wantDataLen, haveDataLen) - } - - m.UnmarshalUnsafe(r.data[hdrLen:]) - return nil -} - // callFuture makes a request to the server and returns a future response. // Call resolve() when the response needs to be fulfilled. func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, error) { @@ -358,11 +285,6 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, // if there are always too many ongoing requests all the time. The // supported maxActiveRequests setting should be really high to avoid this. for conn.fd.numActiveRequests == conn.fd.fs.opts.maxActiveRequests { - if t == nil { - // Since there is no task that is waiting. We must error out. - return nil, errors.New("FUSE request queue full") - } - log.Infof("Blocking request %v from being queued. Too many active requests: %v", r.id, conn.fd.numActiveRequests) conn.fd.mu.Unlock() @@ -378,9 +300,19 @@ func (conn *connection) callFuture(t *kernel.Task, r *Request) (*futureResponse, // callFutureLocked makes a request to the server and returns a future response. func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureResponse, error) { + // Check connected again holding conn.mu. + conn.mu.Lock() + if !conn.connected { + conn.mu.Unlock() + // we checked connected before, + // this must be due to aborted connection. + return nil, syserror.ECONNABORTED + } + conn.mu.Unlock() + conn.fd.queue.PushBack(r) - conn.fd.numActiveRequests += 1 - fut := newFutureResponse(r.hdr.Opcode) + conn.fd.numActiveRequests++ + fut := newFutureResponse(r) conn.fd.completions[r.id] = fut // Signal the readers that there is something to read. @@ -388,50 +320,3 @@ func (conn *connection) callFutureLocked(t *kernel.Task, r *Request) (*futureRes return fut, nil } - -// futureResponse represents an in-flight request, that may or may not have -// completed yet. Convert it to a resolved Response by calling Resolve, but note -// that this may block. -// -// +stateify savable -type futureResponse struct { - opcode linux.FUSEOpcode - ch chan struct{} - hdr *linux.FUSEHeaderOut - data []byte -} - -// newFutureResponse creates a future response to a FUSE request. -func newFutureResponse(opcode linux.FUSEOpcode) *futureResponse { - return &futureResponse{ - opcode: opcode, - ch: make(chan struct{}), - } -} - -// resolve blocks the task until the server responds to its corresponding request, -// then returns a resolved response. -func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) { - // If there is no Task associated with this request - then we don't try to resolve - // the response. Instead, the task writing the response (proxy to the server) will - // process the response on our behalf. - if t == nil { - log.Infof("fuse.Response.resolve: Not waiting on a response from server.") - return nil, nil - } - - if err := t.Block(f.ch); err != nil { - return nil, err - } - - return f.getResponse(), nil -} - -// getResponse creates a Response from the data the futureResponse has. -func (f *futureResponse) getResponse() *Response { - return &Response{ - opcode: f.opcode, - hdr: *f.hdr, - data: f.data, - } -} diff --git a/pkg/sentry/fsimpl/fuse/init.go b/pkg/sentry/fsimpl/fuse/connection_control.go index 779c2bd3f..bfde78559 100644 --- a/pkg/sentry/fsimpl/fuse/init.go +++ b/pkg/sentry/fsimpl/fuse/connection_control.go @@ -15,7 +15,11 @@ package fuse import ( + "sync/atomic" + "syscall" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) @@ -29,9 +33,10 @@ const ( // Follow the same behavior as unix fuse implementation. fuseMaxTimeGranNs = 1000000000 - // Minimum value for MaxWrite. + // Minimum value for MaxWrite and MaxRead. // Follow the same behavior as unix fuse implementation. fuseMinMaxWrite = 4096 + fuseMinMaxRead = 4096 // Temporary default value for max readahead, 128kb. fuseDefaultMaxReadahead = 131072 @@ -49,6 +54,26 @@ var ( MaxUserCongestionThreshold uint16 = fuseDefaultCongestionThreshold ) +// SetInitialized atomically sets the connection as initialized. +func (conn *connection) SetInitialized() { + // Unblock the requests sent before INIT. + close(conn.initializedChan) + + // Close the channel first to avoid the non-atomic situation + // where conn.initialized is true but there are + // tasks being blocked on the channel. + // And it prevents the newer tasks from gaining + // unnecessary higher chance to be issued before the blocked one. + + atomic.StoreInt32(&(conn.initialized), int32(1)) +} + +// IsInitialized atomically check if the connection is initialized. +// pairs with SetInitialized(). +func (conn *connection) Initialized() bool { + return atomic.LoadInt32(&(conn.initialized)) != 0 +} + // InitSend sends a FUSE_INIT request. func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error { in := linux.FUSEInitIn{ @@ -70,29 +95,31 @@ func (conn *connection) InitSend(creds *auth.Credentials, pid uint32) error { } // InitRecv receives a FUSE_INIT reply and process it. +// +// Preconditions: conn.asyncMu must not be held if minor verion is newer than 13. func (conn *connection) InitRecv(res *Response, hasSysAdminCap bool) error { if err := res.Error(); err != nil { return err } - var out linux.FUSEInitOut - if err := res.UnmarshalPayload(&out); err != nil { + initRes := fuseInitRes{initLen: res.DataLen()} + if err := res.UnmarshalPayload(&initRes); err != nil { return err } - return conn.initProcessReply(&out, hasSysAdminCap) + return conn.initProcessReply(&initRes.initOut, hasSysAdminCap) } // Process the FUSE_INIT reply from the FUSE server. +// It tries to acquire the conn.asyncMu lock if minor version is newer than 13. func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap bool) error { + // No matter error or not, always set initialzied. + // to unblock the blocked requests. + defer conn.SetInitialized() + // No support for old major fuse versions. if out.Major != linux.FUSE_KERNEL_VERSION { conn.connInitError = true - - // Set the connection as initialized and unblock the blocked requests - // (i.e. return error for them). - conn.SetInitialized() - return nil } @@ -100,29 +127,14 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap conn.connInitSuccess = true conn.minor = out.Minor - // No support for limits before minor version 13. - if out.Minor >= 13 { - conn.bgLock.Lock() - - if out.MaxBackground > 0 { - conn.maxBackground = out.MaxBackground - - if !hasSysAdminCap && - conn.maxBackground > MaxUserBackgroundRequest { - conn.maxBackground = MaxUserBackgroundRequest - } - } - - if out.CongestionThreshold > 0 { - conn.congestionThreshold = out.CongestionThreshold - - if !hasSysAdminCap && - conn.congestionThreshold > MaxUserCongestionThreshold { - conn.congestionThreshold = MaxUserCongestionThreshold - } - } - - conn.bgLock.Unlock() + // No support for negotiating MaxWrite before minor version 5. + if out.Minor >= 5 { + conn.maxWrite = out.MaxWrite + } else { + conn.maxWrite = fuseMinMaxWrite + } + if conn.maxWrite < fuseMinMaxWrite { + conn.maxWrite = fuseMinMaxWrite } // No support for the following flags before minor version 6. @@ -131,8 +143,6 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap conn.bigWrites = out.Flags&linux.FUSE_BIG_WRITES != 0 conn.dontMask = out.Flags&linux.FUSE_DONT_MASK != 0 conn.writebackCache = out.Flags&linux.FUSE_WRITEBACK_CACHE != 0 - conn.cacheSymlinks = out.Flags&linux.FUSE_CACHE_SYMLINKS != 0 - conn.abortErr = out.Flags&linux.FUSE_ABORT_ERROR != 0 // TODO(gvisor.dev/issue/3195): figure out how to use TimeGran (0 < TimeGran <= fuseMaxTimeGranNs). @@ -148,19 +158,90 @@ func (conn *connection) initProcessReply(out *linux.FUSEInitOut, hasSysAdminCap } } - // No support for negotiating MaxWrite before minor version 5. - if out.Minor >= 5 { - conn.maxWrite = out.MaxWrite - } else { - conn.maxWrite = fuseMinMaxWrite + // No support for limits before minor version 13. + if out.Minor >= 13 { + conn.asyncMu.Lock() + + if out.MaxBackground > 0 { + conn.asyncNumMax = out.MaxBackground + + if !hasSysAdminCap && + conn.asyncNumMax > MaxUserBackgroundRequest { + conn.asyncNumMax = MaxUserBackgroundRequest + } + } + + if out.CongestionThreshold > 0 { + conn.asyncCongestionThreshold = out.CongestionThreshold + + if !hasSysAdminCap && + conn.asyncCongestionThreshold > MaxUserCongestionThreshold { + conn.asyncCongestionThreshold = MaxUserCongestionThreshold + } + } + + conn.asyncMu.Unlock() } - if conn.maxWrite < fuseMinMaxWrite { - conn.maxWrite = fuseMinMaxWrite + + return nil +} + +// Abort this FUSE connection. +// It tries to acquire conn.fd.mu, conn.lock, conn.bgLock in order. +// All possible requests waiting or blocking will be aborted. +// +// Preconditions: conn.fd.mu is locked. +func (conn *connection) Abort(ctx context.Context) { + conn.mu.Lock() + conn.asyncMu.Lock() + + if !conn.connected { + conn.asyncMu.Unlock() + conn.mu.Unlock() + conn.fd.mu.Unlock() + return } - // Set connection as initialized and unblock the requests - // issued before init. - conn.SetInitialized() + conn.connected = false - return nil + // Empty the `fd.queue` that holds the requests + // not yet read by the FUSE daemon yet. + // These are a subset of the requests in `fuse.completion` map. + for !conn.fd.queue.Empty() { + req := conn.fd.queue.Front() + conn.fd.queue.Remove(req) + } + + var terminate []linux.FUSEOpID + + // 2. Collect the requests have not been sent to FUSE daemon, + // or have not received a reply. + for unique := range conn.fd.completions { + terminate = append(terminate, unique) + } + + // Release locks to avoid deadlock. + conn.asyncMu.Unlock() + conn.mu.Unlock() + + // 1. The requets blocked before initialization. + // Will reach call() `connected` check and return. + if !conn.Initialized() { + conn.SetInitialized() + } + + // 2. Terminate the requests collected above. + // Set ECONNABORTED error. + // sendError() will remove them from `fd.completion` map. + // Will enter the path of a normally received error. + for _, toTerminate := range terminate { + conn.fd.sendError(ctx, -int32(syscall.ECONNABORTED), toTerminate) + } + + // 3. The requests not yet written to FUSE device. + // Early terminate. + // Will reach callFutureLocked() `connected` check and return. + close(conn.fd.fullQueueCh) + + // TODO(gvisor.dev/issue/3528): Forget all pending forget reqs. } diff --git a/pkg/sentry/fsimpl/fuse/connection_test.go b/pkg/sentry/fsimpl/fuse/connection_test.go new file mode 100644 index 000000000..91d16c1cf --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/connection_test.go @@ -0,0 +1,117 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "math/rand" + "syscall" + "testing" + + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" +) + +// TestConnectionInitBlock tests if initialization +// correctly blocks and unblocks the connection. +// Since it's unfeasible to test kernelTask.Block() in unit test, +// the code in Call() are not tested here. +func TestConnectionInitBlock(t *testing.T) { + s := setup(t) + defer s.Destroy() + + k := kernel.KernelFromContext(s.Ctx) + + conn, _, err := newTestConnection(s, k, maxActiveRequestsDefault) + if err != nil { + t.Fatalf("newTestConnection: %v", err) + } + + select { + case <-conn.initializedChan: + t.Fatalf("initializedChan should be blocking before SetInitialized") + default: + } + + conn.SetInitialized() + + select { + case <-conn.initializedChan: + default: + t.Fatalf("initializedChan should not be blocking after SetInitialized") + } +} + +func TestConnectionAbort(t *testing.T) { + s := setup(t) + defer s.Destroy() + + k := kernel.KernelFromContext(s.Ctx) + creds := auth.CredentialsFromContext(s.Ctx) + task := kernel.TaskFromContext(s.Ctx) + + const numRequests uint64 = 256 + + conn, _, err := newTestConnection(s, k, numRequests) + if err != nil { + t.Fatalf("newTestConnection: %v", err) + } + + testObj := &testPayload{ + data: rand.Uint32(), + } + + var futNormal []*futureResponse + + for i := 0; i < int(numRequests); i++ { + req, err := conn.NewRequest(creds, uint32(i), uint64(i), 0, testObj) + if err != nil { + t.Fatalf("NewRequest creation failed: %v", err) + } + fut, err := conn.callFutureLocked(task, req) + if err != nil { + t.Fatalf("callFutureLocked failed: %v", err) + } + futNormal = append(futNormal, fut) + } + + conn.Abort(s.Ctx) + + // Abort should unblock the initialization channel. + // Note: no test requests are actually blocked on `conn.initializedChan`. + select { + case <-conn.initializedChan: + default: + t.Fatalf("initializedChan should not be blocking after SetInitialized") + } + + // Abort will return ECONNABORTED error to unblocked requests. + for _, fut := range futNormal { + if fut.getResponse().hdr.Error != -int32(syscall.ECONNABORTED) { + t.Fatalf("Incorrect error code received for aborted connection: %v", fut.getResponse().hdr.Error) + } + } + + // After abort, Call() should return directly with ENOTCONN. + req, err := conn.NewRequest(creds, 0, 0, 0, testObj) + if err != nil { + t.Fatalf("NewRequest creation failed: %v", err) + } + _, err = conn.Call(task, req) + if err != syserror.ENOTCONN { + t.Fatalf("Incorrect error code received for Call() after connection aborted") + } + +} diff --git a/pkg/sentry/fsimpl/fuse/dev.go b/pkg/sentry/fsimpl/fuse/dev.go index e522ff9a0..1b86a4b4c 100644 --- a/pkg/sentry/fsimpl/fuse/dev.go +++ b/pkg/sentry/fsimpl/fuse/dev.go @@ -19,7 +19,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -32,6 +31,8 @@ import ( const fuseDevMinor = 229 // fuseDevice implements vfs.Device for /dev/fuse. +// +// +stateify savable type fuseDevice struct{} // Open implements vfs.Device.Open. @@ -50,15 +51,14 @@ func (fuseDevice) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, op } // DeviceFD implements vfs.FileDescriptionImpl for /dev/fuse. +// +// +stateify savable type DeviceFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl vfs.NoLockFD - // mounted specifies whether a FUSE filesystem was mounted using the DeviceFD. - mounted bool - // nextOpID is used to create new requests. nextOpID linux.FUSEOpID @@ -83,7 +83,7 @@ type DeviceFD struct { writeCursorFR *futureResponse // mu protects all the queues, maps, buffers and cursors and nextOpID. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // waitQueue is used to notify interested parties when the device becomes // readable or writable. @@ -92,21 +92,36 @@ type DeviceFD struct { // fullQueueCh is a channel used to synchronize the readers with the writers. // Writers (inbound requests to the filesystem) block if there are too many // unprocessed in-flight requests. - fullQueueCh chan struct{} + fullQueueCh chan struct{} `state:".(int)"` // fs is the FUSE filesystem that this FD is being used for. fs *filesystem } +func (fd *DeviceFD) saveFullQueueCh() int { + return cap(fd.fullQueueCh) +} + +func (fd *DeviceFD) loadFullQueueCh(capacity int) { + fd.fullQueueCh = make(chan struct{}, capacity) +} + // Release implements vfs.FileDescriptionImpl.Release. -func (fd *DeviceFD) Release(context.Context) { - fd.fs.conn.connected = false +func (fd *DeviceFD) Release(ctx context.Context) { + if fd.fs != nil { + fd.fs.conn.mu.Lock() + fd.fs.conn.connected = false + fd.fs.conn.mu.Unlock() + + fd.fs.VFSFilesystem().DecRef(ctx) + fd.fs = nil + } } // PRead implements vfs.FileDescriptionImpl.PRead. func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -116,10 +131,16 @@ func (fd *DeviceFD) PRead(ctx context.Context, dst usermem.IOSequence, offset in // Read implements vfs.FileDescriptionImpl.Read. func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } + // Return ENODEV if the filesystem is umounted. + if fd.fs.umounted { + // TODO(gvisor.dev/issue/3525): return ECONNABORTED if aborted via fuse control fs. + return 0, syserror.ENODEV + } + // We require that any Read done on this filesystem have a sane minimum // read buffer. It must have the capacity for the fixed parts of any request // header (Linux uses the request header and the FUSEWriteIn header for this @@ -143,58 +164,82 @@ func (fd *DeviceFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.R } // readLocked implements the reading of the fuse device while locked with DeviceFD.mu. +// +// Preconditions: dst is large enough for any reasonable request. func (fd *DeviceFD) readLocked(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - if fd.queue.Empty() { - return 0, syserror.ErrWouldBlock - } + var req *Request - var readCursor uint32 - var bytesRead int64 - for { - req := fd.queue.Front() - if dst.NumBytes() < int64(req.hdr.Len) { - // The request is too large. Cannot process it. All requests must be smaller than the - // negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT - // handshake. - errno := -int32(syscall.EIO) - if req.hdr.Opcode == linux.FUSE_SETXATTR { - errno = -int32(syscall.E2BIG) - } + // Find the first valid request. + // For the normal case this loop only execute once. + for !fd.queue.Empty() { + req = fd.queue.Front() - // Return the error to the calling task. - if err := fd.sendError(ctx, errno, req); err != nil { - return 0, err - } + if int64(req.hdr.Len)+int64(len(req.payload)) <= dst.NumBytes() { + break + } - // We're done with this request. - fd.queue.Remove(req) + // The request is too large. Cannot process it. All requests must be smaller than the + // negotiated size as specified by Connection.MaxWrite set as part of the FUSE_INIT + // handshake. + errno := -int32(syscall.EIO) + if req.hdr.Opcode == linux.FUSE_SETXATTR { + errno = -int32(syscall.E2BIG) + } - // Restart the read as this request was invalid. - log.Warningf("fuse.DeviceFD.Read: request found was too large. Restarting read.") - return fd.readLocked(ctx, dst, opts) + // Return the error to the calling task. + if err := fd.sendError(ctx, errno, req.hdr.Unique); err != nil { + return 0, err } - n, err := dst.CopyOut(ctx, req.data[readCursor:]) + // We're done with this request. + fd.queue.Remove(req) + req = nil + } + + if req == nil { + return 0, syserror.ErrWouldBlock + } + + // We already checked the size: dst must be able to fit the whole request. + // Now we write the marshalled header, the payload, + // and the potential additional payload + // to the user memory IOSequence. + + n, err := dst.CopyOut(ctx, req.data) + if err != nil { + return 0, err + } + if n != len(req.data) { + return 0, syserror.EIO + } + + if req.hdr.Opcode == linux.FUSE_WRITE { + written, err := dst.DropFirst(n).CopyOut(ctx, req.payload) if err != nil { return 0, err } - readCursor += uint32(n) - bytesRead += int64(n) - - if readCursor >= req.hdr.Len { - // Fully done with this req, remove it from the queue. - fd.queue.Remove(req) - break + if written != len(req.payload) { + return 0, syserror.EIO } + n += int(written) } - return bytesRead, nil + // Fully done with this req, remove it from the queue. + fd.queue.Remove(req) + + // Remove noReply ones from map of requests expecting a reply. + if req.noReply { + fd.numActiveRequests -= 1 + delete(fd.completions, req.hdr.Unique) + } + + return int64(n), nil } // PWrite implements vfs.FileDescriptionImpl.PWrite. func (fd *DeviceFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -211,10 +256,15 @@ func (fd *DeviceFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs. // writeLocked implements writing to the fuse device while locked with DeviceFD.mu. func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } + // Return ENODEV if the filesystem is umounted. + if fd.fs.umounted { + return 0, syserror.ENODEV + } + var cn, n int64 hdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) @@ -276,7 +326,8 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt fut, ok := fd.completions[hdr.Unique] if !ok { - // Server sent us a response for a request we never sent? + // Server sent us a response for a request we never sent, + // or for which we already received a reply (e.g. aborted), an unlikely event. return 0, syserror.EINVAL } @@ -307,8 +358,23 @@ func (fd *DeviceFD) writeLocked(ctx context.Context, src usermem.IOSequence, opt // Readiness implements vfs.FileDescriptionImpl.Readiness. func (fd *DeviceFD) Readiness(mask waiter.EventMask) waiter.EventMask { + fd.mu.Lock() + defer fd.mu.Unlock() + return fd.readinessLocked(mask) +} + +// readinessLocked implements checking the readiness of the fuse device while +// locked with DeviceFD.mu. +func (fd *DeviceFD) readinessLocked(mask waiter.EventMask) waiter.EventMask { var ready waiter.EventMask - ready |= waiter.EventOut // FD is always writable + + if fd.fs.umounted { + ready |= waiter.EventErr + return ready & mask + } + + // FD is always writable. + ready |= waiter.EventOut if !fd.queue.Empty() { // Have reqs available, FD is readable. ready |= waiter.EventIn @@ -330,7 +396,7 @@ func (fd *DeviceFD) EventUnregister(e *waiter.Entry) { // Seek implements vfs.FileDescriptionImpl.Seek. func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { // Operations on /dev/fuse don't make sense until a FUSE filesystem is mounted. - if !fd.mounted { + if fd.fs == nil { return 0, syserror.EPERM } @@ -338,59 +404,59 @@ func (fd *DeviceFD) Seek(ctx context.Context, offset int64, whence int32) (int64 } // sendResponse sends a response to the waiting task (if any). +// +// Preconditions: fd.mu must be held. func (fd *DeviceFD) sendResponse(ctx context.Context, fut *futureResponse) error { - // See if the running task need to perform some action before returning. - // Since we just finished writing the future, we can be sure that - // getResponse generates a populated response. - if err := fd.noReceiverAction(ctx, fut.getResponse()); err != nil { - return err - } + // Signal the task waiting on a response if any. + defer close(fut.ch) // Signal that the queue is no longer full. select { case fd.fullQueueCh <- struct{}{}: default: } - fd.numActiveRequests -= 1 + fd.numActiveRequests-- + + if fut.async { + return fd.asyncCallBack(ctx, fut.getResponse()) + } - // Signal the task waiting on a response. - close(fut.ch) return nil } -// sendError sends an error response to the waiting task (if any). -func (fd *DeviceFD) sendError(ctx context.Context, errno int32, req *Request) error { +// sendError sends an error response to the waiting task (if any) by calling sendResponse(). +// +// Preconditions: fd.mu must be held. +func (fd *DeviceFD) sendError(ctx context.Context, errno int32, unique linux.FUSEOpID) error { // Return the error to the calling task. outHdrLen := uint32((*linux.FUSEHeaderOut)(nil).SizeBytes()) respHdr := linux.FUSEHeaderOut{ Len: outHdrLen, Error: errno, - Unique: req.hdr.Unique, + Unique: unique, } fut, ok := fd.completions[respHdr.Unique] if !ok { - // Server sent us a response for a request we never sent? + // A response for a request we never sent, + // or for which we already received a reply (e.g. aborted). return syserror.EINVAL } delete(fd.completions, respHdr.Unique) fut.hdr = &respHdr - if err := fd.sendResponse(ctx, fut); err != nil { - return err - } - - return nil + return fd.sendResponse(ctx, fut) } -// noReceiverAction has the calling kernel.Task do some action if its known that no -// receiver is going to be waiting on the future channel. This is to be used by: -// FUSE_INIT. -func (fd *DeviceFD) noReceiverAction(ctx context.Context, r *Response) error { - if r.opcode == linux.FUSE_INIT { +// asyncCallBack executes pre-defined callback function for async requests. +// Currently used by: FUSE_INIT. +func (fd *DeviceFD) asyncCallBack(ctx context.Context, r *Response) error { + switch r.opcode { + case linux.FUSE_INIT: creds := auth.CredentialsFromContext(ctx) rootUserNs := kernel.KernelFromContext(ctx).RootUserNamespace() return fd.fs.conn.InitRecv(r, creds.HasCapabilityIn(linux.CAP_SYS_ADMIN, rootUserNs)) + // TODO(gvisor.dev/issue/3247): support async read: correctly process the response. } return nil diff --git a/pkg/sentry/fsimpl/fuse/dev_test.go b/pkg/sentry/fsimpl/fuse/dev_test.go index 1ffe7ccd2..5986133e9 100644 --- a/pkg/sentry/fsimpl/fuse/dev_test.go +++ b/pkg/sentry/fsimpl/fuse/dev_test.go @@ -16,7 +16,6 @@ package fuse import ( "fmt" - "io" "math/rand" "testing" @@ -28,17 +27,12 @@ import ( "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) // echoTestOpcode is the Opcode used during testing. The server used in tests // will simply echo the payload back with the appropriate headers. const echoTestOpcode linux.FUSEOpcode = 1000 -type testPayload struct { - data uint32 -} - // TestFUSECommunication tests that the communication layer between the Sentry and the // FUSE server daemon works as expected. func TestFUSECommunication(t *testing.T) { @@ -327,102 +321,3 @@ func fuseServerRun(t *testing.T, s *testutil.System, k *kernel.Kernel, fd *vfs.F } } } - -func setup(t *testing.T) *testutil.System { - k, err := testutil.Boot() - if err != nil { - t.Fatalf("Error creating kernel: %v", err) - } - - ctx := k.SupervisorContext() - creds := auth.CredentialsFromContext(ctx) - - k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ - AllowUserList: true, - AllowUserMount: true, - }) - - mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) - if err != nil { - t.Fatalf("NewMountNamespace(): %v", err) - } - - return testutil.NewSystem(ctx, t, k.VFS(), mntns) -} - -// newTestConnection creates a fuse connection that the sentry can communicate with -// and the FD for the server to communicate with. -func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) { - vfsObj := &vfs.VirtualFilesystem{} - fuseDev := &DeviceFD{} - - if err := vfsObj.Init(system.Ctx); err != nil { - return nil, nil, err - } - - vd := vfsObj.NewAnonVirtualDentry("genCountFD") - defer vd.DecRef(system.Ctx) - if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil { - return nil, nil, err - } - - fsopts := filesystemOptions{ - maxActiveRequests: maxActiveRequests, - } - fs, err := NewFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd) - if err != nil { - return nil, nil, err - } - - return fs.conn, &fuseDev.vfsfd, nil -} - -// SizeBytes implements marshal.Marshallable.SizeBytes. -func (t *testPayload) SizeBytes() int { - return 4 -} - -// MarshalBytes implements marshal.Marshallable.MarshalBytes. -func (t *testPayload) MarshalBytes(dst []byte) { - usermem.ByteOrder.PutUint32(dst[:4], t.data) -} - -// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes. -func (t *testPayload) UnmarshalBytes(src []byte) { - *t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])} -} - -// Packed implements marshal.Marshallable.Packed. -func (t *testPayload) Packed() bool { - return true -} - -// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe. -func (t *testPayload) MarshalUnsafe(dst []byte) { - t.MarshalBytes(dst) -} - -// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe. -func (t *testPayload) UnmarshalUnsafe(src []byte) { - t.UnmarshalBytes(src) -} - -// CopyOutN implements marshal.Marshallable.CopyOutN. -func (t *testPayload) CopyOutN(task marshal.Task, addr usermem.Addr, limit int) (int, error) { - panic("not implemented") -} - -// CopyOut implements marshal.Marshallable.CopyOut. -func (t *testPayload) CopyOut(task marshal.Task, addr usermem.Addr) (int, error) { - panic("not implemented") -} - -// CopyIn implements marshal.Marshallable.CopyIn. -func (t *testPayload) CopyIn(task marshal.Task, addr usermem.Addr) (int, error) { - panic("not implemented") -} - -// WriteTo implements io.WriterTo.WriteTo. -func (t *testPayload) WriteTo(w io.Writer) (int64, error) { - panic("not implemented") -} diff --git a/pkg/sentry/fsimpl/fuse/directory.go b/pkg/sentry/fsimpl/fuse/directory.go new file mode 100644 index 000000000..8f220a04b --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/directory.go @@ -0,0 +1,105 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type directoryFD struct { + fileDescription +} + +// Allocate implements directoryFD.Allocate. +func (*directoryFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + return syserror.EISDIR +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (*directoryFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (*directoryFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (*directoryFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (*directoryFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EISDIR +} + +// IterDirents implements vfs.FileDescriptionImpl.IterDirents. +func (dir *directoryFD) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback) error { + fusefs := dir.inode().fs + task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) + + in := linux.FUSEReadIn{ + Fh: dir.Fh, + Offset: uint64(atomic.LoadInt64(&dir.off)), + Size: linux.FUSE_PAGE_SIZE, + Flags: dir.statusFlags(), + } + + // TODO(gVisor.dev/issue/3404): Support FUSE_READDIRPLUS. + req, err := fusefs.conn.NewRequest(creds, uint32(task.ThreadID()), dir.inode().nodeID, linux.FUSE_READDIR, &in) + if err != nil { + return err + } + + res, err := fusefs.conn.Call(task, req) + if err != nil { + return err + } + if err := res.Error(); err != nil { + return err + } + + var out linux.FUSEDirents + if err := res.UnmarshalPayload(&out); err != nil { + return err + } + + for _, fuseDirent := range out.Dirents { + nextOff := int64(fuseDirent.Meta.Off) + dirent := vfs.Dirent{ + Name: fuseDirent.Name, + Type: uint8(fuseDirent.Meta.Type), + Ino: fuseDirent.Meta.Ino, + NextOff: nextOff, + } + + if err := callback.Handle(dirent); err != nil { + return err + } + atomic.StoreInt64(&dir.off, nextOff) + } + + return nil +} diff --git a/pkg/sentry/fsimpl/fuse/file.go b/pkg/sentry/fsimpl/fuse/file.go new file mode 100644 index 000000000..83f2816b7 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/file.go @@ -0,0 +1,133 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fileDescription implements vfs.FileDescriptionImpl for fuse. +type fileDescription struct { + vfsfd vfs.FileDescription + vfs.FileDescriptionDefaultImpl + vfs.DentryMetadataFileDescriptionImpl + vfs.NoLockFD + + // the file handle used in userspace. + Fh uint64 + + // Nonseekable is indicate cannot perform seek on a file. + Nonseekable bool + + // DirectIO suggest fuse to use direct io operation. + DirectIO bool + + // OpenFlag is the flag returned by open. + OpenFlag uint32 + + // off is the file offset. + off int64 +} + +func (fd *fileDescription) dentry() *kernfs.Dentry { + return fd.vfsfd.Dentry().Impl().(*kernfs.Dentry) +} + +func (fd *fileDescription) inode() *inode { + return fd.dentry().Inode().(*inode) +} + +func (fd *fileDescription) filesystem() *vfs.Filesystem { + return fd.vfsfd.VirtualDentry().Mount().Filesystem() +} + +func (fd *fileDescription) statusFlags() uint32 { + return fd.vfsfd.StatusFlags() +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *fileDescription) Release(ctx context.Context) { + // no need to release if FUSE server doesn't implement Open. + conn := fd.inode().fs.conn + if conn.noOpen { + return + } + + in := linux.FUSEReleaseIn{ + Fh: fd.Fh, + Flags: fd.statusFlags(), + } + // TODO(gvisor.dev/issue/3245): add logic when we support file lock owner. + var opcode linux.FUSEOpcode + if fd.inode().Mode().IsDir() { + opcode = linux.FUSE_RELEASEDIR + } else { + opcode = linux.FUSE_RELEASE + } + kernelTask := kernel.TaskFromContext(ctx) + // ignoring errors and FUSE server reply is analogous to Linux's behavior. + req, err := conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), fd.inode().nodeID, opcode, &in) + if err != nil { + // No way to invoke Call() with an errored request. + return + } + // The reply will be ignored since no callback is defined in asyncCallBack(). + conn.CallAsync(kernelTask, req) +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + return 0, nil +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + return 0, nil +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, nil +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return 0, nil +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *fileDescription) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + return 0, nil +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + fs := fd.filesystem() + inode := fd.inode() + return inode.Stat(ctx, fs, opts) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + fs := fd.filesystem() + creds := auth.CredentialsFromContext(ctx) + return fd.inode().setAttr(ctx, fs, creds, opts, true, fd.Fh) +} diff --git a/pkg/sentry/fsimpl/fuse/fusefs.go b/pkg/sentry/fsimpl/fuse/fusefs.go index 83c24ec25..e7ef5998e 100644 --- a/pkg/sentry/fsimpl/fuse/fusefs.go +++ b/pkg/sentry/fsimpl/fuse/fusefs.go @@ -16,24 +16,36 @@ package fuse import ( + "math" "strconv" + "sync" + "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/waiter" ) // Name is the default filesystem name. const Name = "fuse" +// maxActiveRequestsDefault is the default setting controlling the upper bound +// on the number of active requests at any given time. +const maxActiveRequestsDefault = 10000 + // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} +// +stateify savable type filesystemOptions struct { // userID specifies the numeric uid of the mount owner. // This option should not be specified by the filesystem owner. @@ -56,9 +68,16 @@ type filesystemOptions struct { // exist at any time. Any further requests will block when trying to // Call the server. maxActiveRequests uint64 + + // maxRead is the max number of bytes to read, + // specified as "max_read" in fs parameters. + // If not specified by user, use math.MaxUint32 as default value. + maxRead uint32 } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem devMinor uint32 @@ -69,6 +88,9 @@ type filesystem struct { // opts is the options the fusefs is initialized with. opts *filesystemOptions + + // umounted is true if filesystem.Release() has been called. + umounted bool } // Name implements vfs.FilesystemType.Name. @@ -76,6 +98,9 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { devMinor, err := vfsObj.GetAnonBlockDevMinor() @@ -142,14 +167,29 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt // Set the maxInFlightRequests option. fsopts.maxActiveRequests = maxActiveRequestsDefault + if maxReadStr, ok := mopts["max_read"]; ok { + delete(mopts, "max_read") + maxRead, err := strconv.ParseUint(maxReadStr, 10, 32) + if err != nil { + log.Warningf("%s.GetFilesystem: invalid max_read: max_read=%s", fsType.Name(), maxReadStr) + return nil, nil, syserror.EINVAL + } + if maxRead < fuseMinMaxRead { + maxRead = fuseMinMaxRead + } + fsopts.maxRead = uint32(maxRead) + } else { + fsopts.maxRead = math.MaxUint32 + } + // Check for unparsed options. if len(mopts) != 0 { - log.Warningf("%s.GetFilesystem: unknown options: %v", fsType.Name(), mopts) + log.Warningf("%s.GetFilesystem: unsupported or unknown options: %v", fsType.Name(), mopts) return nil, nil, syserror.EINVAL } // Create a new FUSE filesystem. - fs, err := NewFUSEFilesystem(ctx, devMinor, &fsopts, fuseFd) + fs, err := newFUSEFilesystem(ctx, devMinor, &fsopts, fuseFd) if err != nil { log.Warningf("%s.NewFUSEFilesystem: failed with error: %v", fsType.Name(), err) return nil, nil, err @@ -165,26 +205,28 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } // root is the fusefs root directory. - root := fs.newInode(creds, fsopts.rootMode) + root := fs.newRootInode(ctx, creds, fsopts.rootMode) return fs.VFSFilesystem(), root.VFSDentry(), nil } -// NewFUSEFilesystem creates a new FUSE filesystem. -func NewFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOptions, device *vfs.FileDescription) (*filesystem, error) { - fs := &filesystem{ - devMinor: devMinor, - opts: opts, - } - - conn, err := newFUSEConnection(ctx, device, opts.maxActiveRequests) +// newFUSEFilesystem creates a new FUSE filesystem. +func newFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOptions, device *vfs.FileDescription) (*filesystem, error) { + conn, err := newFUSEConnection(ctx, device, opts) if err != nil { log.Warningf("fuse.NewFUSEFilesystem: NewFUSEConnection failed with error: %v", err) return nil, syserror.EINVAL } - fs.conn = conn fuseFD := device.Impl().(*DeviceFD) + + fs := &filesystem{ + devMinor: devMinor, + opts: opts, + conn: conn, + } + + fs.VFSFilesystem().IncRef() fuseFD.fs = fs return fs, nil @@ -192,39 +234,361 @@ func NewFUSEFilesystem(ctx context.Context, devMinor uint32, opts *filesystemOpt // Release implements vfs.FilesystemImpl.Release. func (fs *filesystem) Release(ctx context.Context) { + fs.conn.fd.mu.Lock() + + fs.umounted = true + fs.conn.Abort(ctx) + // Notify all the waiters on this fd. + fs.conn.fd.waitQueue.Notify(waiter.EventIn) + + fs.conn.fd.mu.Unlock() + fs.Filesystem.VFSFilesystem().VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) fs.Filesystem.Release(ctx) } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { + inodeRefs + kernfs.InodeAlwaysValid kernfs.InodeAttrs - kernfs.InodeNoDynamicLookup - kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink kernfs.OrderedChildren + // the owning filesystem. fs is immutable. + fs *filesystem + + // metaDataMu protects the metadata of this inode. + metadataMu sync.Mutex + + nodeID uint64 + locks vfs.FileLocks - dentry kernfs.Dentry + // size of the file. + size uint64 + + // attributeVersion is the version of inode's attributes. + attributeVersion uint64 + + // attributeTime is the remaining vaild time of attributes. + attributeTime uint64 + + // version of the inode. + version uint64 + + // link is result of following a symbolic link. + link string } -func (fs *filesystem) newInode(creds *auth.Credentials, mode linux.FileMode) *kernfs.Dentry { - i := &inode{} - i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) +func (fs *filesystem) newRootInode(ctx context.Context, creds *auth.Credentials, mode linux.FileMode) *kernfs.Dentry { + i := &inode{fs: fs, nodeID: 1} + i.InodeAttrs.Init(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, 1, linux.ModeDirectory|0755) i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - i.dentry.Init(i) + i.EnableLeakCheck() - return &i.dentry + var d kernfs.Dentry + d.Init(&fs.Filesystem, i) + return &d +} + +func (fs *filesystem) newInode(ctx context.Context, nodeID uint64, attr linux.FUSEAttr) kernfs.Inode { + i := &inode{fs: fs, nodeID: nodeID} + creds := auth.Credentials{EffectiveKGID: auth.KGID(attr.UID), EffectiveKUID: auth.KUID(attr.UID)} + i.InodeAttrs.Init(ctx, &creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.FileMode(attr.Mode)) + atomic.StoreUint64(&i.size, attr.Size) + i.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + i.EnableLeakCheck() + return i } // Open implements kernfs.Inode.Open. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + isDir := i.InodeAttrs.Mode().IsDir() + // return error if specified to open directory but inode is not a directory. + if !isDir && opts.Mode.IsDir() { + return nil, syserror.ENOTDIR + } + if opts.Flags&linux.O_LARGEFILE == 0 && atomic.LoadUint64(&i.size) > linux.MAX_NON_LFS { + return nil, syserror.EOVERFLOW + } + + var fd *fileDescription + var fdImpl vfs.FileDescriptionImpl + if isDir { + directoryFD := &directoryFD{} + fd = &(directoryFD.fileDescription) + fdImpl = directoryFD + } else { + regularFD := ®ularFileFD{} + fd = &(regularFD.fileDescription) + fdImpl = regularFD + } + // FOPEN_KEEP_CACHE is the defualt flag for noOpen. + fd.OpenFlag = linux.FOPEN_KEEP_CACHE + + // Only send open request when FUSE server support open or is opening a directory. + if !i.fs.conn.noOpen || isDir { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.Open: couldn't get kernel task from context") + return nil, syserror.EINVAL + } + + // Build the request. + var opcode linux.FUSEOpcode + if isDir { + opcode = linux.FUSE_OPENDIR + } else { + opcode = linux.FUSE_OPEN + } + + in := linux.FUSEOpenIn{Flags: opts.Flags & ^uint32(linux.O_CREAT|linux.O_EXCL|linux.O_NOCTTY)} + if !i.fs.conn.atomicOTrunc { + in.Flags &= ^uint32(linux.O_TRUNC) + } + + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, opcode, &in) + if err != nil { + return nil, err + } + + // Send the request and receive the reply. + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return nil, err + } + if err := res.Error(); err == syserror.ENOSYS && !isDir { + i.fs.conn.noOpen = true + } else if err != nil { + return nil, err + } else { + out := linux.FUSEOpenOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return nil, err + } + + // Process the reply. + fd.OpenFlag = out.OpenFlag + if isDir { + fd.OpenFlag &= ^uint32(linux.FOPEN_DIRECT_IO) + } + + fd.Fh = out.Fh + } + } + + // TODO(gvisor.dev/issue/3234): invalidate mmap after implemented it for FUSE Inode + fd.DirectIO = fd.OpenFlag&linux.FOPEN_DIRECT_IO != 0 + fdOptions := &vfs.FileDescriptionOptions{} + if fd.OpenFlag&linux.FOPEN_NONSEEKABLE != 0 { + fdOptions.DenyPRead = true + fdOptions.DenyPWrite = true + fd.Nonseekable = true + } + + // If we don't send SETATTR before open (which is indicated by atomicOTrunc) + // and O_TRUNC is set, update the inode's version number and clean existing data + // by setting the file size to 0. + if i.fs.conn.atomicOTrunc && opts.Flags&linux.O_TRUNC != 0 { + i.fs.conn.mu.Lock() + i.fs.conn.attributeVersion++ + i.attributeVersion = i.fs.conn.attributeVersion + atomic.StoreUint64(&i.size, 0) + i.fs.conn.mu.Unlock() + i.attributeTime = 0 + } + + if err := fd.vfsfd.Init(fdImpl, opts.Flags, rp.Mount(), d.VFSDentry(), fdOptions); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// Lookup implements kernfs.Inode.Lookup. +func (i *inode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { + in := linux.FUSELookupIn{Name: name} + return i.newEntry(ctx, name, 0, linux.FUSE_LOOKUP, &in) +} + +// Keep implements kernfs.Inode.Keep. +func (i *inode) Keep() bool { + // Return true so that kernfs keeps the new dentry pointing to this + // inode in the dentry tree. This is needed because inodes created via + // Lookup are not temporary. They might refer to existing files on server + // that can be Unlink'd/Rmdir'd. + return true +} + +// IterDirents implements kernfs.Inode.IterDirents. +func (*inode) IterDirents(ctx context.Context, mnt *vfs.Mount, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { + return offset, nil +} + +// NewFile implements kernfs.Inode.NewFile. +func (i *inode) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (kernfs.Inode, error) { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.NewFile: couldn't get kernel task from context", i.nodeID) + return nil, syserror.EINVAL + } + in := linux.FUSECreateIn{ + CreateMeta: linux.FUSECreateMeta{ + Flags: opts.Flags, + Mode: uint32(opts.Mode) | linux.S_IFREG, + Umask: uint32(kernelTask.FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, linux.S_IFREG, linux.FUSE_CREATE, &in) +} + +// NewNode implements kernfs.Inode.NewNode. +func (i *inode) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (kernfs.Inode, error) { + in := linux.FUSEMknodIn{ + MknodMeta: linux.FUSEMknodMeta{ + Mode: uint32(opts.Mode), + Rdev: linux.MakeDeviceID(uint16(opts.DevMajor), opts.DevMinor), + Umask: uint32(kernel.TaskFromContext(ctx).FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, opts.Mode.FileType(), linux.FUSE_MKNOD, &in) +} + +// NewSymlink implements kernfs.Inode.NewSymlink. +func (i *inode) NewSymlink(ctx context.Context, name, target string) (kernfs.Inode, error) { + in := linux.FUSESymLinkIn{ + Name: name, + Target: target, + } + return i.newEntry(ctx, name, linux.S_IFLNK, linux.FUSE_SYMLINK, &in) +} + +// Unlink implements kernfs.Inode.Unlink. +func (i *inode) Unlink(ctx context.Context, name string, child kernfs.Inode) error { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.newEntry: couldn't get kernel task from context", i.nodeID) + return syserror.EINVAL + } + in := linux.FUSEUnlinkIn{Name: name} + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, linux.FUSE_UNLINK, &in) if err != nil { + return err + } + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return err + } + // only return error, discard res. + return res.Error() +} + +// NewDir implements kernfs.Inode.NewDir. +func (i *inode) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (kernfs.Inode, error) { + in := linux.FUSEMkdirIn{ + MkdirMeta: linux.FUSEMkdirMeta{ + Mode: uint32(opts.Mode), + Umask: uint32(kernel.TaskFromContext(ctx).FSContext().Umask()), + }, + Name: name, + } + return i.newEntry(ctx, name, linux.S_IFDIR, linux.FUSE_MKDIR, &in) +} + +// RmDir implements kernfs.Inode.RmDir. +func (i *inode) RmDir(ctx context.Context, name string, child kernfs.Inode) error { + fusefs := i.fs + task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) + + in := linux.FUSERmDirIn{Name: name} + req, err := fusefs.conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_RMDIR, &in) + if err != nil { + return err + } + + res, err := i.fs.conn.Call(task, req) + if err != nil { + return err + } + return res.Error() +} + +// newEntry calls FUSE server for entry creation and allocates corresponding entry according to response. +// Shared by FUSE_MKNOD, FUSE_MKDIR, FUSE_SYMLINK, FUSE_LINK and FUSE_LOOKUP. +func (i *inode) newEntry(ctx context.Context, name string, fileType linux.FileMode, opcode linux.FUSEOpcode, payload marshal.Marshallable) (kernfs.Inode, error) { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.newEntry: couldn't get kernel task from context", i.nodeID) + return nil, syserror.EINVAL + } + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, opcode, payload) + if err != nil { + return nil, err + } + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return nil, err + } + if err := res.Error(); err != nil { + return nil, err + } + out := linux.FUSEEntryOut{} + if err := res.UnmarshalPayload(&out); err != nil { return nil, err } - return fd.VFSFileDescription(), nil + if opcode != linux.FUSE_LOOKUP && ((out.Attr.Mode&linux.S_IFMT)^uint32(fileType) != 0 || out.NodeID == 0 || out.NodeID == linux.FUSE_ROOT_ID) { + return nil, syserror.EIO + } + child := i.fs.newInode(ctx, out.NodeID, out.Attr) + return child, nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (i *inode) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + path, err := i.Readlink(ctx, mnt) + return vfs.VirtualDentry{}, path, err +} + +// Readlink implements kernfs.Inode.Readlink. +func (i *inode) Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { + if i.Mode().FileType()&linux.S_IFLNK == 0 { + return "", syserror.EINVAL + } + if len(i.link) == 0 { + kernelTask := kernel.TaskFromContext(ctx) + if kernelTask == nil { + log.Warningf("fusefs.Inode.Readlink: couldn't get kernel task from context") + return "", syserror.EINVAL + } + req, err := i.fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(kernelTask.ThreadID()), i.nodeID, linux.FUSE_READLINK, &linux.FUSEEmptyIn{}) + if err != nil { + return "", err + } + res, err := i.fs.conn.Call(kernelTask, req) + if err != nil { + return "", err + } + i.link = string(res.data[res.hdr.SizeBytes():]) + if !mnt.Options().ReadOnly { + i.attributeTime = 0 + } + } + return i.link, nil +} + +// getFUSEAttr returns a linux.FUSEAttr of this inode stored in local cache. +// TODO(gvisor.dev/issue/3679): Add support for other fields. +func (i *inode) getFUSEAttr() linux.FUSEAttr { + return linux.FUSEAttr{ + Ino: i.Ino(), + Size: atomic.LoadUint64(&i.size), + Mode: uint32(i.Mode()), + } } // statFromFUSEAttr makes attributes from linux.FUSEAttr to linux.Statx. The @@ -280,45 +644,179 @@ func statFromFUSEAttr(attr linux.FUSEAttr, mask, devMinor uint32) linux.Statx { return stat } -// Stat implements kernfs.Inode.Stat. -func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { - fusefs := fs.Impl().(*filesystem) - conn := fusefs.conn - task, creds := kernel.TaskFromContext(ctx), auth.CredentialsFromContext(ctx) +// getAttr gets the attribute of this inode by issuing a FUSE_GETATTR request +// or read from local cache. It updates the corresponding attributes if +// necessary. +func (i *inode) getAttr(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions, flags uint32, fh uint64) (linux.FUSEAttr, error) { + attributeVersion := atomic.LoadUint64(&i.fs.conn.attributeVersion) + + // TODO(gvisor.dev/issue/3679): send the request only if + // - invalid local cache for fields specified in the opts.Mask + // - forced update + // - i.attributeTime expired + // If local cache is still valid, return local cache. + // Currently we always send a request, + // and we always set the metadata with the new result, + // unless attributeVersion has changed. + + task := kernel.TaskFromContext(ctx) if task == nil { log.Warningf("couldn't get kernel task from context") - return linux.Statx{}, syserror.EINVAL + return linux.FUSEAttr{}, syserror.EINVAL } - var in linux.FUSEGetAttrIn - // We don't set any attribute in the request, because in VFS2 fstat(2) will - // finally be translated into vfs.FilesystemImpl.StatAt() (see - // pkg/sentry/syscalls/linux/vfs2/stat.go), resulting in the same flow - // as stat(2). Thus GetAttrFlags and Fh variable will never be used in VFS2. - req, err := conn.NewRequest(creds, uint32(task.ThreadID()), i.Ino(), linux.FUSE_GETATTR, &in) + creds := auth.CredentialsFromContext(ctx) + + in := linux.FUSEGetAttrIn{ + GetAttrFlags: flags, + Fh: fh, + } + req, err := i.fs.conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_GETATTR, &in) if err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } - res, err := conn.Call(task, req) + res, err := i.fs.conn.Call(task, req) if err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } if err := res.Error(); err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err } var out linux.FUSEGetAttrOut if err := res.UnmarshalPayload(&out); err != nil { - return linux.Statx{}, err + return linux.FUSEAttr{}, err + } + + // Local version is newer, return the local one. + // Skip the update. + if attributeVersion != 0 && atomic.LoadUint64(&i.attributeVersion) > attributeVersion { + return i.getFUSEAttr(), nil } - // Set all metadata into kernfs.InodeAttrs. - if err := i.SetStat(ctx, fs, creds, vfs.SetStatOptions{ - Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, fusefs.devMinor), + // Set the metadata of kernfs.InodeAttrs. + if err := i.InodeAttrs.SetStat(ctx, fs, creds, vfs.SetStatOptions{ + Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, i.fs.devMinor), }); err != nil { + return linux.FUSEAttr{}, err + } + + // Set the size if no error (after SetStat() check). + atomic.StoreUint64(&i.size, out.Attr.Size) + + return out.Attr, nil +} + +// reviseAttr attempts to update the attributes for internal purposes +// by calling getAttr with a pre-specified mask. +// Used by read, write, lseek. +func (i *inode) reviseAttr(ctx context.Context, flags uint32, fh uint64) error { + // Never need atime for internal purposes. + _, err := i.getAttr(ctx, i.fs.VFSFilesystem(), vfs.StatOptions{ + Mask: linux.STATX_BASIC_STATS &^ linux.STATX_ATIME, + }, flags, fh) + return err +} + +// Stat implements kernfs.Inode.Stat. +func (i *inode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { + attr, err := i.getAttr(ctx, fs, opts, 0, 0) + if err != nil { return linux.Statx{}, err } - return statFromFUSEAttr(out.Attr, opts.Mask, fusefs.devMinor), nil + return statFromFUSEAttr(attr, opts.Mask, i.fs.devMinor), nil +} + +// DecRef implements kernfs.Inode.DecRef. +func (i *inode) DecRef(ctx context.Context) { + i.inodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + +// StatFS implements kernfs.Inode.StatFS. +func (i *inode) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, error) { + // TODO(gvisor.dev/issues/3413): Complete the implementation of statfs. + return vfs.GenericStatFS(linux.FUSE_SUPER_MAGIC), nil +} + +// fattrMaskFromStats converts vfs.SetStatOptions.Stat.Mask to linux stats mask +// aligned with the attribute mask defined in include/linux/fs.h. +func fattrMaskFromStats(mask uint32) uint32 { + var fuseAttrMask uint32 + maskMap := map[uint32]uint32{ + linux.STATX_MODE: linux.FATTR_MODE, + linux.STATX_UID: linux.FATTR_UID, + linux.STATX_GID: linux.FATTR_GID, + linux.STATX_SIZE: linux.FATTR_SIZE, + linux.STATX_ATIME: linux.FATTR_ATIME, + linux.STATX_MTIME: linux.FATTR_MTIME, + linux.STATX_CTIME: linux.FATTR_CTIME, + } + for statxMask, fattrMask := range maskMap { + if mask&statxMask != 0 { + fuseAttrMask |= fattrMask + } + } + return fuseAttrMask +} + +// SetStat implements kernfs.Inode.SetStat. +func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { + return i.setAttr(ctx, fs, creds, opts, false, 0) +} + +func (i *inode) setAttr(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions, useFh bool, fh uint64) error { + conn := i.fs.conn + task := kernel.TaskFromContext(ctx) + if task == nil { + log.Warningf("couldn't get kernel task from context") + return syserror.EINVAL + } + + // We should retain the original file type when assigning new mode. + fileType := uint16(i.Mode()) & linux.S_IFMT + fattrMask := fattrMaskFromStats(opts.Stat.Mask) + if useFh { + fattrMask |= linux.FATTR_FH + } + in := linux.FUSESetAttrIn{ + Valid: fattrMask, + Fh: fh, + Size: opts.Stat.Size, + Atime: uint64(opts.Stat.Atime.Sec), + Mtime: uint64(opts.Stat.Mtime.Sec), + Ctime: uint64(opts.Stat.Ctime.Sec), + AtimeNsec: opts.Stat.Atime.Nsec, + MtimeNsec: opts.Stat.Mtime.Nsec, + CtimeNsec: opts.Stat.Ctime.Nsec, + Mode: uint32(fileType | opts.Stat.Mode), + UID: opts.Stat.UID, + GID: opts.Stat.GID, + } + req, err := conn.NewRequest(creds, uint32(task.ThreadID()), i.nodeID, linux.FUSE_SETATTR, &in) + if err != nil { + return err + } + + res, err := conn.Call(task, req) + if err != nil { + return err + } + if err := res.Error(); err != nil { + return err + } + out := linux.FUSEGetAttrOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return err + } + + // Set the metadata of kernfs.InodeAttrs. + if err := i.InodeAttrs.SetStat(ctx, fs, creds, vfs.SetStatOptions{ + Stat: statFromFUSEAttr(out.Attr, linux.STATX_ALL, i.fs.devMinor), + }); err != nil { + return err + } + + return nil } diff --git a/pkg/sentry/fsimpl/fuse/read_write.go b/pkg/sentry/fsimpl/fuse/read_write.go new file mode 100644 index 000000000..2d396e84c --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/read_write.go @@ -0,0 +1,244 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "io" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// ReadInPages sends FUSE_READ requests for the size after round it up to +// a multiple of page size, blocks on it for reply, processes the reply +// and returns the payload (or joined payloads) as a byte slice. +// This is used for the general purpose reading. +// We do not support direct IO (which read the exact number of bytes) +// at this moment. +func (fs *filesystem) ReadInPages(ctx context.Context, fd *regularFileFD, off uint64, size uint32) ([][]byte, uint32, error) { + attributeVersion := atomic.LoadUint64(&fs.conn.attributeVersion) + + t := kernel.TaskFromContext(ctx) + if t == nil { + log.Warningf("fusefs.Read: couldn't get kernel task from context") + return nil, 0, syserror.EINVAL + } + + // Round up to a multiple of page size. + readSize, _ := usermem.PageRoundUp(uint64(size)) + + // One request cannnot exceed either maxRead or maxPages. + maxPages := fs.conn.maxRead >> usermem.PageShift + if maxPages > uint32(fs.conn.maxPages) { + maxPages = uint32(fs.conn.maxPages) + } + + var outs [][]byte + var sizeRead uint32 + + // readSize is a multiple of usermem.PageSize. + // Always request bytes as a multiple of pages. + pagesRead, pagesToRead := uint32(0), uint32(readSize>>usermem.PageShift) + + // Reuse the same struct for unmarshalling to avoid unnecessary memory allocation. + in := linux.FUSEReadIn{ + Fh: fd.Fh, + LockOwner: 0, // TODO(gvisor.dev/issue/3245): file lock + ReadFlags: 0, // TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER + Flags: fd.statusFlags(), + } + + // This loop is intended for fragmented read where the bytes to read is + // larger than either the maxPages or maxRead. + // For the majority of reads with normal size, this loop should only + // execute once. + for pagesRead < pagesToRead { + pagesCanRead := pagesToRead - pagesRead + if pagesCanRead > maxPages { + pagesCanRead = maxPages + } + + in.Offset = off + (uint64(pagesRead) << usermem.PageShift) + in.Size = pagesCanRead << usermem.PageShift + + req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), fd.inode().nodeID, linux.FUSE_READ, &in) + if err != nil { + return nil, 0, err + } + + // TODO(gvisor.dev/issue/3247): support async read. + + res, err := fs.conn.Call(t, req) + if err != nil { + return nil, 0, err + } + if err := res.Error(); err != nil { + return nil, 0, err + } + + // Not enough bytes in response, + // either we reached EOF, + // or the FUSE server sends back a response + // that cannot even fit the hdr. + if len(res.data) <= res.hdr.SizeBytes() { + // We treat both case as EOF here for now + // since there is no reliable way to detect + // the over-short hdr case. + break + } + + // Directly using the slice to avoid extra copy. + out := res.data[res.hdr.SizeBytes():] + + outs = append(outs, out) + sizeRead += uint32(len(out)) + + pagesRead += pagesCanRead + } + + defer fs.ReadCallback(ctx, fd, off, size, sizeRead, attributeVersion) + + // No bytes returned: offset >= EOF. + if len(outs) == 0 { + return nil, 0, io.EOF + } + + return outs, sizeRead, nil +} + +// ReadCallback updates several information after receiving a read response. +// Due to readahead, sizeRead can be larger than size. +func (fs *filesystem) ReadCallback(ctx context.Context, fd *regularFileFD, off uint64, size uint32, sizeRead uint32, attributeVersion uint64) { + // TODO(gvisor.dev/issue/3247): support async read. + // If this is called by an async read, correctly process it. + // May need to update the signature. + + i := fd.inode() + i.InodeAttrs.TouchAtime(ctx, fd.vfsfd.Mount()) + + // Reached EOF. + if sizeRead < size { + // TODO(gvisor.dev/issue/3630): If we have writeback cache, then we need to fill this hole. + // Might need to update the buf to be returned from the Read(). + + // Update existing size. + newSize := off + uint64(sizeRead) + fs.conn.mu.Lock() + if attributeVersion == i.attributeVersion && newSize < atomic.LoadUint64(&i.size) { + fs.conn.attributeVersion++ + i.attributeVersion = i.fs.conn.attributeVersion + atomic.StoreUint64(&i.size, newSize) + } + fs.conn.mu.Unlock() + } +} + +// Write sends FUSE_WRITE requests and return the bytes +// written according to the response. +// +// Preconditions: len(data) == size. +func (fs *filesystem) Write(ctx context.Context, fd *regularFileFD, off uint64, size uint32, data []byte) (uint32, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + log.Warningf("fusefs.Read: couldn't get kernel task from context") + return 0, syserror.EINVAL + } + + // One request cannnot exceed either maxWrite or maxPages. + maxWrite := uint32(fs.conn.maxPages) << usermem.PageShift + if maxWrite > fs.conn.maxWrite { + maxWrite = fs.conn.maxWrite + } + + // Reuse the same struct for unmarshalling to avoid unnecessary memory allocation. + in := linux.FUSEWriteIn{ + Fh: fd.Fh, + // TODO(gvisor.dev/issue/3245): file lock + LockOwner: 0, + // TODO(gvisor.dev/issue/3245): |= linux.FUSE_READ_LOCKOWNER + // TODO(gvisor.dev/issue/3237): |= linux.FUSE_WRITE_CACHE (not added yet) + WriteFlags: 0, + Flags: fd.statusFlags(), + } + + inode := fd.inode() + var written uint32 + + // This loop is intended for fragmented write where the bytes to write is + // larger than either the maxWrite or maxPages or when bigWrites is false. + // Unless a small value for max_write is explicitly used, this loop + // is expected to execute only once for the majority of the writes. + for written < size { + toWrite := size - written + + // Limit the write size to one page. + // Note that the bigWrites flag is obsolete, + // latest libfuse always sets it on. + if !fs.conn.bigWrites && toWrite > usermem.PageSize { + toWrite = usermem.PageSize + } + + // Limit the write size to maxWrite. + if toWrite > maxWrite { + toWrite = maxWrite + } + + in.Offset = off + uint64(written) + in.Size = toWrite + + req, err := fs.conn.NewRequest(auth.CredentialsFromContext(ctx), uint32(t.ThreadID()), inode.nodeID, linux.FUSE_WRITE, &in) + if err != nil { + return 0, err + } + + req.payload = data[written : written+toWrite] + + // TODO(gvisor.dev/issue/3247): support async write. + + res, err := fs.conn.Call(t, req) + if err != nil { + return 0, err + } + if err := res.Error(); err != nil { + return 0, err + } + + out := linux.FUSEWriteOut{} + if err := res.UnmarshalPayload(&out); err != nil { + return 0, err + } + + // Write more than requested? EIO. + if out.Size > toWrite { + return 0, syserror.EIO + } + + written += out.Size + + // Break if short write. Not necessarily an error. + if out.Size != toWrite { + break + } + } + inode.InodeAttrs.TouchCMtime(ctx) + + return written, nil +} diff --git a/pkg/sentry/fsimpl/fuse/regular_file.go b/pkg/sentry/fsimpl/fuse/regular_file.go new file mode 100644 index 000000000..5bdd096c3 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/regular_file.go @@ -0,0 +1,230 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "io" + "math" + "sync" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type regularFileFD struct { + fileDescription + + // off is the file offset. + off int64 + // offMu protects off. + offMu sync.Mutex +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + if offset < 0 { + return 0, syserror.EINVAL + } + + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, syserror.EOPNOTSUPP + } + + size := dst.NumBytes() + if size == 0 { + // Early return if count is 0. + return 0, nil + } else if size > math.MaxUint32 { + // FUSE only supports uint32 for size. + // Overflow. + return 0, syserror.EINVAL + } + + // TODO(gvisor.dev/issue/3678): Add direct IO support. + + inode := fd.inode() + + // Reading beyond EOF, update file size if outdated. + if uint64(offset+size) > atomic.LoadUint64(&inode.size) { + if err := inode.reviseAttr(ctx, linux.FUSE_GETATTR_FH, fd.Fh); err != nil { + return 0, err + } + // If the offset after update is still too large, return error. + if uint64(offset) >= atomic.LoadUint64(&inode.size) { + return 0, io.EOF + } + } + + // Truncate the read with updated file size. + fileSize := atomic.LoadUint64(&inode.size) + if uint64(offset+size) > fileSize { + size = int64(fileSize) - offset + } + + buffers, n, err := inode.fs.ReadInPages(ctx, fd, uint64(offset), uint32(size)) + if err != nil { + return 0, err + } + + // TODO(gvisor.dev/issue/3237): support indirect IO (e.g. caching), + // store the bytes that were read ahead. + + // Update the number of bytes to copy for short read. + if n < uint32(size) { + size = int64(n) + } + + // Copy the bytes read to the dst. + // This loop is intended for fragmented reads. + // For the majority of reads, this loop only execute once. + var copied int64 + for _, buffer := range buffers { + toCopy := int64(len(buffer)) + if copied+toCopy > size { + toCopy = size - copied + } + cp, err := dst.DropFirst64(copied).CopyOut(ctx, buffer[:toCopy]) + if err != nil { + return 0, err + } + if int64(cp) != toCopy { + return 0, syserror.EIO + } + copied += toCopy + } + + return copied, nil +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + fd.offMu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.offMu.Unlock() + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + n, _, err := fd.pwrite(ctx, src, offset, opts) + return n, err +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + fd.offMu.Lock() + n, off, err := fd.pwrite(ctx, src, fd.off, opts) + fd.off = off + fd.offMu.Unlock() + return n, err +} + +// pwrite returns the number of bytes written, final offset and error. The +// final offset should be ignored by PWrite. +func (fd *regularFileFD) pwrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (written, finalOff int64, err error) { + if offset < 0 { + return 0, offset, syserror.EINVAL + } + + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, offset, syserror.EOPNOTSUPP + } + + inode := fd.inode() + inode.metadataMu.Lock() + defer inode.metadataMu.Unlock() + + // If the file is opened with O_APPEND, update offset to file size. + // Note: since our Open() implements the interface of kernfs, + // and kernfs currently does not support O_APPEND, this will never + // be true before we switch out from kernfs. + if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 { + // Locking inode.metadataMu is sufficient for reading size + offset = int64(inode.size) + } + + srclen := src.NumBytes() + + if srclen > math.MaxUint32 { + // FUSE only supports uint32 for size. + // Overflow. + return 0, offset, syserror.EINVAL + } + if end := offset + srclen; end < offset { + // Overflow. + return 0, offset, syserror.EINVAL + } + + srclen, err = vfs.CheckLimit(ctx, offset, srclen) + if err != nil { + return 0, offset, err + } + + if srclen == 0 { + // Return before causing any side effects. + return 0, offset, nil + } + + src = src.TakeFirst64(srclen) + + // TODO(gvisor.dev/issue/3237): Add cache support: + // buffer cache. Ideally we write from src to our buffer cache first. + // The slice passed to fs.Write() should be a slice from buffer cache. + data := make([]byte, srclen) + // Reason for making a copy here: connection.Call() blocks on kerneltask, + // which in turn acquires mm.activeMu lock. Functions like CopyInTo() will + // attemp to acquire the mm.activeMu lock as well -> deadlock. + // We must finish reading from the userspace memory before + // t.Block() deactivates it. + cp, err := src.CopyIn(ctx, data) + if err != nil { + return 0, offset, err + } + if int64(cp) != srclen { + return 0, offset, syserror.EIO + } + + n, err := fd.inode().fs.Write(ctx, fd, uint64(offset), uint32(srclen), data) + if err != nil { + return 0, offset, err + } + + if n == 0 { + // We have checked srclen != 0 previously. + // If err == nil, then it's a short write and we return EIO. + return 0, offset, syserror.EIO + } + + written = int64(n) + finalOff = offset + written + + if finalOff > int64(inode.size) { + atomic.StoreUint64(&inode.size, uint64(finalOff)) + atomic.AddUint64(&inode.fs.conn.attributeVersion, 1) + } + + return +} diff --git a/pkg/sentry/fsimpl/fuse/request_response.go b/pkg/sentry/fsimpl/fuse/request_response.go new file mode 100644 index 000000000..7fa00569b --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/request_response.go @@ -0,0 +1,229 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "fmt" + "syscall" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/usermem" +) + +// fuseInitRes is a variable-length wrapper of linux.FUSEInitOut. The FUSE +// server may implement an older version of FUSE protocol, which contains a +// linux.FUSEInitOut with less attributes. +// +// Dynamically-sized objects cannot be marshalled. +type fuseInitRes struct { + marshal.StubMarshallable + + // initOut contains the response from the FUSE server. + initOut linux.FUSEInitOut + + // initLen is the total length of bytes of the response. + initLen uint32 +} + +// UnmarshalBytes deserializes src to the initOut attribute in a fuseInitRes. +func (r *fuseInitRes) UnmarshalBytes(src []byte) { + out := &r.initOut + + // Introduced before FUSE kernel version 7.13. + out.Major = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.Minor = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.MaxReadahead = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.Flags = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + out.MaxBackground = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + out.CongestionThreshold = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + out.MaxWrite = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + + // Introduced in FUSE kernel version 7.23. + if len(src) >= 4 { + out.TimeGran = uint32(usermem.ByteOrder.Uint32(src[:4])) + src = src[4:] + } + // Introduced in FUSE kernel version 7.28. + if len(src) >= 2 { + out.MaxPages = uint16(usermem.ByteOrder.Uint16(src[:2])) + src = src[2:] + } +} + +// SizeBytes is the size of the payload of the FUSE_INIT response. +func (r *fuseInitRes) SizeBytes() int { + return int(r.initLen) +} + +// Ordinary requests have even IDs, while interrupts IDs are odd. +// Used to increment the unique ID for each FUSE request. +var reqIDStep uint64 = 2 + +// Request represents a FUSE operation request that hasn't been sent to the +// server yet. +// +// +stateify savable +type Request struct { + requestEntry + + id linux.FUSEOpID + hdr *linux.FUSEHeaderIn + data []byte + + // payload for this request: extra bytes to write after + // the data slice. Used by FUSE_WRITE. + payload []byte + + // If this request is async. + async bool + // If we don't care its response. + // Manually set by the caller. + noReply bool +} + +// NewRequest creates a new request that can be sent to the FUSE server. +func (conn *connection) NewRequest(creds *auth.Credentials, pid uint32, ino uint64, opcode linux.FUSEOpcode, payload marshal.Marshallable) (*Request, error) { + conn.fd.mu.Lock() + defer conn.fd.mu.Unlock() + conn.fd.nextOpID += linux.FUSEOpID(reqIDStep) + + hdrLen := (*linux.FUSEHeaderIn)(nil).SizeBytes() + hdr := linux.FUSEHeaderIn{ + Len: uint32(hdrLen + payload.SizeBytes()), + Opcode: opcode, + Unique: conn.fd.nextOpID, + NodeID: ino, + UID: uint32(creds.EffectiveKUID), + GID: uint32(creds.EffectiveKGID), + PID: pid, + } + + buf := make([]byte, hdr.Len) + + // TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again. + hdr.MarshalBytes(buf[:hdrLen]) + payload.MarshalBytes(buf[hdrLen:]) + + return &Request{ + id: hdr.Unique, + hdr: &hdr, + data: buf, + }, nil +} + +// futureResponse represents an in-flight request, that may or may not have +// completed yet. Convert it to a resolved Response by calling Resolve, but note +// that this may block. +// +// +stateify savable +type futureResponse struct { + opcode linux.FUSEOpcode + ch chan struct{} + hdr *linux.FUSEHeaderOut + data []byte + + // If this request is async. + async bool +} + +// newFutureResponse creates a future response to a FUSE request. +func newFutureResponse(req *Request) *futureResponse { + return &futureResponse{ + opcode: req.hdr.Opcode, + ch: make(chan struct{}), + async: req.async, + } +} + +// resolve blocks the task until the server responds to its corresponding request, +// then returns a resolved response. +func (f *futureResponse) resolve(t *kernel.Task) (*Response, error) { + // Return directly for async requests. + if f.async { + return nil, nil + } + + if err := t.Block(f.ch); err != nil { + return nil, err + } + + return f.getResponse(), nil +} + +// getResponse creates a Response from the data the futureResponse has. +func (f *futureResponse) getResponse() *Response { + return &Response{ + opcode: f.opcode, + hdr: *f.hdr, + data: f.data, + } +} + +// Response represents an actual response from the server, including the +// response payload. +// +// +stateify savable +type Response struct { + opcode linux.FUSEOpcode + hdr linux.FUSEHeaderOut + data []byte +} + +// Error returns the error of the FUSE call. +func (r *Response) Error() error { + errno := r.hdr.Error + if errno >= 0 { + return nil + } + + sysErrNo := syscall.Errno(-errno) + return error(sysErrNo) +} + +// DataLen returns the size of the response without the header. +func (r *Response) DataLen() uint32 { + return r.hdr.Len - uint32(r.hdr.SizeBytes()) +} + +// UnmarshalPayload unmarshals the response data into m. +func (r *Response) UnmarshalPayload(m marshal.Marshallable) error { + hdrLen := r.hdr.SizeBytes() + haveDataLen := r.hdr.Len - uint32(hdrLen) + wantDataLen := uint32(m.SizeBytes()) + + if haveDataLen < wantDataLen { + return fmt.Errorf("payload too small. Minimum data lenth required: %d, but got data length %d", wantDataLen, haveDataLen) + } + + // The response data is empty unless there is some payload. And so, doesn't + // need to be unmarshalled. + if r.data == nil { + return nil + } + + // TODO(gVisor.dev/issue/3698): Use the unsafe version once go_marshal is safe to use again. + m.UnmarshalBytes(r.data[hdrLen:]) + return nil +} diff --git a/pkg/sentry/fsimpl/fuse/utils_test.go b/pkg/sentry/fsimpl/fuse/utils_test.go new file mode 100644 index 000000000..e1d9e3365 --- /dev/null +++ b/pkg/sentry/fsimpl/fuse/utils_test.go @@ -0,0 +1,132 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package fuse + +import ( + "io" + "testing" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/usermem" +) + +func setup(t *testing.T) *testutil.System { + k, err := testutil.Boot() + if err != nil { + t.Fatalf("Error creating kernel: %v", err) + } + + ctx := k.SupervisorContext() + creds := auth.CredentialsFromContext(ctx) + + k.VFS().MustRegisterFilesystemType(Name, &FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserList: true, + AllowUserMount: true, + }) + + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) + if err != nil { + t.Fatalf("NewMountNamespace(): %v", err) + } + + return testutil.NewSystem(ctx, t, k.VFS(), mntns) +} + +// newTestConnection creates a fuse connection that the sentry can communicate with +// and the FD for the server to communicate with. +func newTestConnection(system *testutil.System, k *kernel.Kernel, maxActiveRequests uint64) (*connection, *vfs.FileDescription, error) { + vfsObj := &vfs.VirtualFilesystem{} + fuseDev := &DeviceFD{} + + if err := vfsObj.Init(system.Ctx); err != nil { + return nil, nil, err + } + + vd := vfsObj.NewAnonVirtualDentry("genCountFD") + defer vd.DecRef(system.Ctx) + if err := fuseDev.vfsfd.Init(fuseDev, linux.O_RDWR|linux.O_CREAT, vd.Mount(), vd.Dentry(), &vfs.FileDescriptionOptions{}); err != nil { + return nil, nil, err + } + + fsopts := filesystemOptions{ + maxActiveRequests: maxActiveRequests, + } + fs, err := newFUSEFilesystem(system.Ctx, 0, &fsopts, &fuseDev.vfsfd) + if err != nil { + return nil, nil, err + } + + return fs.conn, &fuseDev.vfsfd, nil +} + +type testPayload struct { + marshal.StubMarshallable + data uint32 +} + +// SizeBytes implements marshal.Marshallable.SizeBytes. +func (t *testPayload) SizeBytes() int { + return 4 +} + +// MarshalBytes implements marshal.Marshallable.MarshalBytes. +func (t *testPayload) MarshalBytes(dst []byte) { + usermem.ByteOrder.PutUint32(dst[:4], t.data) +} + +// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes. +func (t *testPayload) UnmarshalBytes(src []byte) { + *t = testPayload{data: usermem.ByteOrder.Uint32(src[:4])} +} + +// Packed implements marshal.Marshallable.Packed. +func (t *testPayload) Packed() bool { + return true +} + +// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe. +func (t *testPayload) MarshalUnsafe(dst []byte) { + t.MarshalBytes(dst) +} + +// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe. +func (t *testPayload) UnmarshalUnsafe(src []byte) { + t.UnmarshalBytes(src) +} + +// CopyOutN implements marshal.Marshallable.CopyOutN. +func (t *testPayload) CopyOutN(task marshal.CopyContext, addr usermem.Addr, limit int) (int, error) { + panic("not implemented") +} + +// CopyOut implements marshal.Marshallable.CopyOut. +func (t *testPayload) CopyOut(task marshal.CopyContext, addr usermem.Addr) (int, error) { + panic("not implemented") +} + +// CopyIn implements marshal.Marshallable.CopyIn. +func (t *testPayload) CopyIn(task marshal.CopyContext, addr usermem.Addr) (int, error) { + panic("not implemented") +} + +// WriteTo implements io.WriterTo.WriteTo. +func (t *testPayload) WriteTo(w io.Writer) (int64, error) { + panic("not implemented") +} diff --git a/pkg/sentry/fsimpl/gofer/BUILD b/pkg/sentry/fsimpl/gofer/BUILD index 16787116f..4c3e9acf8 100644 --- a/pkg/sentry/fsimpl/gofer/BUILD +++ b/pkg/sentry/fsimpl/gofer/BUILD @@ -38,6 +38,7 @@ go_library( "host_named_pipe.go", "p9file.go", "regular_file.go", + "save_restore.go", "socket.go", "special_file.go", "symlink.go", @@ -52,6 +53,8 @@ go_library( "//pkg/fspath", "//pkg/log", "//pkg/p9", + "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/lock", @@ -69,6 +72,7 @@ go_library( "//pkg/sentry/socket/unix/transport", "//pkg/sentry/usage", "//pkg/sentry/vfs", + "//pkg/sync", "//pkg/syserr", "//pkg/syserror", "//pkg/unet", diff --git a/pkg/sentry/fsimpl/gofer/directory.go b/pkg/sentry/fsimpl/gofer/directory.go index 2a8011eb4..e993c8e36 100644 --- a/pkg/sentry/fsimpl/gofer/directory.go +++ b/pkg/sentry/fsimpl/gofer/directory.go @@ -16,16 +16,17 @@ package gofer import ( "fmt" - "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/refsvfs2" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -34,8 +35,11 @@ func (d *dentry) isDir() bool { return d.fileType() == linux.S_IFDIR } -// Preconditions: filesystem.renameMu must be locked. d.dirMu must be locked. -// d.isDir(). child must be a newly-created dentry that has never had a parent. +// Preconditions: +// * filesystem.renameMu must be locked. +// * d.dirMu must be locked. +// * d.isDir(). +// * child must be a newly-created dentry that has never had a parent. func (d *dentry) cacheNewChildLocked(child *dentry, name string) { d.IncRef() // reference held by child on its parent child.parent = d @@ -46,7 +50,9 @@ func (d *dentry) cacheNewChildLocked(child *dentry, name string) { d.children[name] = child } -// Preconditions: d.dirMu must be locked. d.isDir(). +// Preconditions: +// * d.dirMu must be locked. +// * d.isDir(). func (d *dentry) cacheNegativeLookupLocked(name string) { // Don't cache negative lookups if InteropModeShared is in effect (since // this makes remote lookup unavoidable), or if d.isSynthetic() (in which @@ -79,13 +85,15 @@ type createSyntheticOpts struct { // createSyntheticChildLocked creates a synthetic file with the given name // in d. // -// Preconditions: d.dirMu must be locked. d.isDir(). d does not already contain -// a child with the given name. +// Preconditions: +// * d.dirMu must be locked. +// * d.isDir(). +// * d does not already contain a child with the given name. func (d *dentry) createSyntheticChildLocked(opts *createSyntheticOpts) { - d2 := &dentry{ + child := &dentry{ refs: 1, // held by d fs: d.fs, - ino: d.fs.nextSyntheticIno(), + ino: d.fs.nextIno(), mode: uint32(opts.mode), uid: uint32(opts.kuid), gid: uint32(opts.kgid), @@ -93,28 +101,32 @@ func (d *dentry) createSyntheticChildLocked(opts *createSyntheticOpts) { hostFD: -1, nlink: uint32(2), } + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Register(child, "gofer.dentry") + } switch opts.mode.FileType() { case linux.S_IFDIR: // Nothing else needs to be done. case linux.S_IFSOCK: - d2.endpoint = opts.endpoint + child.endpoint = opts.endpoint case linux.S_IFIFO: - d2.pipe = opts.pipe + child.pipe = opts.pipe default: panic(fmt.Sprintf("failed to create synthetic file of unrecognized type: %v", opts.mode.FileType())) } - d2.pf.dentry = d2 - d2.vfsd.Init(d2) + child.pf.dentry = child + child.vfsd.Init(child) - d.cacheNewChildLocked(d2, opts.name) + d.cacheNewChildLocked(child, opts.name) d.syntheticChildren++ } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 dirents []vfs.Dirent } @@ -151,7 +163,9 @@ func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallba return nil } -// Preconditions: d.isDir(). There exists at least one directoryFD representing d. +// Preconditions: +// * d.isDir(). +// * There exists at least one directoryFD representing d. func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { // NOTE(b/135560623): 9P2000.L's readdir does not specify behavior in the // presence of concurrent mutation of an iterated directory, so @@ -225,7 +239,7 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { } dirent := vfs.Dirent{ Name: p9d.Name, - Ino: uint64(inoFromPath(p9d.QID.Path)), + Ino: d.fs.inoFromQIDPath(p9d.QID.Path), NextOff: int64(len(dirents) + 1), } // p9 does not expose 9P2000.U's DMDEVICE, DMNAMEDPIPE, or diff --git a/pkg/sentry/fsimpl/gofer/filesystem.go b/pkg/sentry/fsimpl/gofer/filesystem.go index 9a90351e5..baecb88c4 100644 --- a/pkg/sentry/fsimpl/gofer/filesystem.go +++ b/pkg/sentry/fsimpl/gofer/filesystem.go @@ -35,7 +35,7 @@ import ( // Sync implements vfs.FilesystemImpl.Sync. func (fs *filesystem) Sync(ctx context.Context) error { - // Snapshot current syncable dentries and special files. + // Snapshot current syncable dentries and special file FDs. fs.syncMu.Lock() ds := make([]*dentry, 0, len(fs.syncableDentries)) for d := range fs.syncableDentries { @@ -53,22 +53,28 @@ func (fs *filesystem) Sync(ctx context.Context) error { // regardless. var retErr error - // Sync regular files. + // Sync syncable dentries. for _, d := range ds { - err := d.syncCachedFile(ctx) + err := d.syncCachedFile(ctx, true /* forFilesystemSync */) d.DecRef(ctx) - if err != nil && retErr == nil { - retErr = err + if err != nil { + ctx.Infof("gofer.filesystem.Sync: dentry.syncCachedFile failed: %v", err) + if retErr == nil { + retErr = err + } } } // Sync special files, which may be writable but do not use dentry shared // handles (so they won't be synced by the above). for _, sffd := range sffds { - err := sffd.Sync(ctx) + err := sffd.sync(ctx, true /* forFilesystemSync */) sffd.vfsfd.DecRef(ctx) - if err != nil && retErr == nil { - retErr = err + if err != nil { + ctx.Infof("gofer.filesystem.Sync: specialFileFD.sync failed: %v", err) + if retErr == nil { + retErr = err + } } } @@ -115,9 +121,12 @@ func putDentrySlice(ds *[]*dentry) { // Dentries which may become cached as a result of the traversal are appended // to *ds. // -// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. -// !rp.Done(). If !d.cachedMetadataAuthoritative(), then d's cached metadata -// must be up to date. +// Preconditions: +// * fs.renameMu must be locked. +// * d.dirMu must be locked. +// * !rp.Done(). +// * If !d.cachedMetadataAuthoritative(), then d's cached metadata must be up +// to date. // // Postconditions: The returned dentry's cached metadata is up to date. func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, mayFollowSymlinks bool, ds **[]*dentry) (*dentry, error) { @@ -185,8 +194,11 @@ afterSymlink: // getChildLocked returns a dentry representing the child of parent with the // given name. If no such child exists, getChildLocked returns (nil, nil). // -// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. -// parent.isDir(). name is not "." or "..". +// Preconditions: +// * fs.renameMu must be locked. +// * parent.dirMu must be locked. +// * parent.isDir(). +// * name is not "." or "..". // // Postconditions: If getChildLocked returns a non-nil dentry, its cached // metadata is up to date. @@ -206,7 +218,8 @@ func (fs *filesystem) getChildLocked(ctx context.Context, vfsObj *vfs.VirtualFil return fs.revalidateChildLocked(ctx, vfsObj, parent, name, child, ds) } -// Preconditions: As for getChildLocked. !parent.isSynthetic(). +// Preconditions: Same as getChildLocked, plus: +// * !parent.isSynthetic(). func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *dentry, name string, child *dentry, ds **[]*dentry) (*dentry, error) { if child != nil { // Need to lock child.metadataMu because we might be updating child @@ -222,7 +235,7 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir return nil, err } if child != nil { - if !file.isNil() && inoFromPath(qid.Path) == child.ino { + if !file.isNil() && qid.Path == child.qidPath { // The file at this path hasn't changed. Just update cached metadata. file.close(ctx) child.updateFromP9AttrsLocked(attrMask, &attr) @@ -279,9 +292,11 @@ func (fs *filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // rp.Start().Impl().(*dentry)). It does not check that the returned directory // is searchable by the provider of rp. // -// Preconditions: fs.renameMu must be locked. !rp.Done(). If -// !d.cachedMetadataAuthoritative(), then d's cached metadata must be up to -// date. +// Preconditions: +// * fs.renameMu must be locked. +// * !rp.Done(). +// * If !d.cachedMetadataAuthoritative(), then d's cached metadata must be up +// to date. func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { for !rp.Final() { d.dirMu.Lock() @@ -328,9 +343,10 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, // createInRemoteDir (if the parent directory is a real remote directory) or // createInSyntheticDir (if the parent directory is synthetic) to do so. // -// Preconditions: !rp.Done(). For the final path component in rp, -// !rp.ShouldFollowSymlink(). -func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string) error, createInSyntheticDir func(parent *dentry, name string) error) error { +// Preconditions: +// * !rp.Done(). +// * For the final path component in rp, !rp.ShouldFollowSymlink(). +func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, createInRemoteDir func(parent *dentry, name string, ds **[]*dentry) error, createInSyntheticDir func(parent *dentry, name string) error) error { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(ctx, &ds) @@ -399,7 +415,7 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir // RPC will fail with EEXIST like we would have. If the RPC succeeds, and a // stale dentry exists, the dentry will fail revalidation next time it's // used. - if err := createInRemoteDir(parent, name); err != nil { + if err := createInRemoteDir(parent, name, &ds); err != nil { return err } ev := linux.IN_CREATE @@ -414,7 +430,7 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir } // No cached dentry exists; however, there might still be an existing file // at name. As above, we attempt the file creation RPC anyway. - if err := createInRemoteDir(parent, name); err != nil { + if err := createInRemoteDir(parent, name, &ds); err != nil { return err } if child, ok := parent.children[name]; ok && child == nil { @@ -721,7 +737,7 @@ func (fs *filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa // LinkAt implements vfs.FilesystemImpl.LinkAt. func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs.VirtualDentry) error { - return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, childName string) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, childName string, _ **[]*dentry) error { if rp.Mount() != vd.Mount() { return syserror.EXDEV } @@ -754,7 +770,7 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. // MkdirAt implements vfs.FilesystemImpl.MkdirAt. func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MkdirOptions) error { creds := rp.Credentials() - return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(ctx, rp, true /* dir */, func(parent *dentry, name string, _ **[]*dentry) error { if _, err := parent.file.mkdir(ctx, name, (p9.FileMode)(opts.Mode), (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)); err != nil { if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { return err @@ -789,34 +805,49 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v // MknodAt implements vfs.FilesystemImpl.MknodAt. func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.MknodOptions) error { - return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string, ds **[]*dentry) error { creds := rp.Credentials() _, err := parent.file.mknod(ctx, name, (p9.FileMode)(opts.Mode), opts.DevMajor, opts.DevMinor, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) - // If the gofer does not allow creating a socket or pipe, create a - // synthetic one, i.e. one that is kept entirely in memory. - if err == syserror.EPERM { - switch opts.Mode.FileType() { - case linux.S_IFSOCK: - parent.createSyntheticChildLocked(&createSyntheticOpts{ - name: name, - mode: opts.Mode, - kuid: creds.EffectiveKUID, - kgid: creds.EffectiveKGID, - endpoint: opts.Endpoint, - }) - return nil - case linux.S_IFIFO: - parent.createSyntheticChildLocked(&createSyntheticOpts{ - name: name, - mode: opts.Mode, - kuid: creds.EffectiveKUID, - kgid: creds.EffectiveKGID, - pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize), - }) - return nil - } + if err != syserror.EPERM { + return err } - return err + + // EPERM means that gofer does not allow creating a socket or pipe. Fallback + // to creating a synthetic one, i.e. one that is kept entirely in memory. + + // Check that we're not overriding an existing file with a synthetic one. + _, err = fs.stepLocked(ctx, rp, parent, true, ds) + switch { + case err == nil: + // Step succeeded, another file exists. + return syserror.EEXIST + case err != syserror.ENOENT: + // Unexpected error. + return err + } + + switch opts.Mode.FileType() { + case linux.S_IFSOCK: + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + endpoint: opts.Endpoint, + }) + return nil + case linux.S_IFIFO: + parent.createSyntheticChildLocked(&createSyntheticOpts{ + name: name, + mode: opts.Mode, + kuid: creds.EffectiveKUID, + kgid: creds.EffectiveKGID, + pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize), + }) + return nil + } + // Retain error from gofer if synthetic file cannot be created internally. + return syserror.EPERM }, nil) } @@ -1001,7 +1032,7 @@ func (d *dentry) open(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.Open // step is required even if !d.cachedMetadataAuthoritative() because // d.mappings has to be updated. // d.metadataMu has already been acquired if trunc == true. - d.updateFileSizeLocked(0) + d.updateSizeLocked(0) if d.cachedMetadataAuthoritative() { d.touchCMtimeLocked() @@ -1072,8 +1103,10 @@ retry: return &fd.vfsfd, nil } -// Preconditions: d.fs.renameMu must be locked. d.dirMu must be locked. -// !d.isSynthetic(). +// Preconditions: +// * d.fs.renameMu must be locked. +// * d.dirMu must be locked. +// * !d.isSynthetic(). func (d *dentry) createAndOpenChildLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions, ds **[]*dentry) (*vfs.FileDescription, error) { if err := d.checkPermissions(rp.Credentials(), vfs.MayWrite); err != nil { return nil, err @@ -1284,6 +1317,9 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if !renamed.isDir() { return syserror.EISDIR } + if genericIsAncestorDentry(replaced, renamed) { + return syserror.ENOTEMPTY + } } else { if rp.MustBeDir() || renamed.isDir() { return syserror.ENOTDIR @@ -1334,14 +1370,15 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa // with reference counts and queue oldParent for checkCachingLocked if the // parent isn't actually changing. if oldParent != newParent { + oldParent.decRefLocked() ds = appendDentry(ds, oldParent) newParent.IncRef() if renamed.isSynthetic() { oldParent.syntheticChildren-- newParent.syntheticChildren++ } + renamed.parent = newParent } - renamed.parent = newParent renamed.name = newName if newParent.children == nil { newParent.children = make(map[string]*dentry) @@ -1385,11 +1422,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.setStat(ctx, rp.Credentials(), &opts, rp.Mount()); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.setStat(ctx, rp.Credentials(), &opts, rp.Mount()) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) @@ -1452,7 +1489,7 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, target string) error { - return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string) error { + return fs.doCreateAt(ctx, rp, false /* dir */, func(parent *dentry, name string, _ **[]*dentry) error { creds := rp.Credentials() _, err := parent.file.symlink(ctx, target, name, (p9.UID)(creds.EffectiveKUID), (p9.GID)(creds.EffectiveKGID)) return err @@ -1464,7 +1501,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return fs.unlinkAt(ctx, rp, false /* dir */) } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { var ds *[]*dentry fs.renameMu.RLock() @@ -1481,17 +1518,18 @@ func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath d.IncRef() return &endpoint{ dentry: d, - file: d.file.file, path: opts.Addr, }, nil } - return d.endpoint, nil + if d.endpoint != nil { + return d.endpoint, nil + } } return nil, syserror.ECONNREFUSED } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(ctx, &ds) @@ -1499,11 +1537,11 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, si if err != nil { return nil, err } - return d.listxattr(ctx, rp.Credentials(), size) + return d.listXattr(ctx, rp.Credentials(), size) } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckCaching(ctx, &ds) @@ -1511,11 +1549,11 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt if err != nil { return "", err } - return d.getxattr(ctx, rp.Credentials(), &opts) + return d.getXattr(ctx, rp.Credentials(), &opts) } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { var ds *[]*dentry fs.renameMu.RLock() d, err := fs.resolveLocked(ctx, rp, &ds) @@ -1523,18 +1561,18 @@ func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.setxattr(ctx, rp.Credentials(), &opts); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.setXattr(ctx, rp.Credentials(), &opts) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { var ds *[]*dentry fs.renameMu.RLock() d, err := fs.resolveLocked(ctx, rp, &ds) @@ -1542,11 +1580,11 @@ func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, fs.renameMuRUnlockAndCheckCaching(ctx, &ds) return err } - if err := d.removexattr(ctx, rp.Credentials(), name); err != nil { - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + err = d.removeXattr(ctx, rp.Credentials(), name) + fs.renameMuRUnlockAndCheckCaching(ctx, &ds) + if err != nil { return err } - fs.renameMuRUnlockAndCheckCaching(ctx, &ds) d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil @@ -1558,7 +1596,3 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe defer fs.renameMu.RUnlock() return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*dentry), b) } - -func (fs *filesystem) nextSyntheticIno() inodeNumber { - return inodeNumber(atomic.AddUint64(&fs.syntheticSeq, 1) | syntheticInoMask) -} diff --git a/pkg/sentry/fsimpl/gofer/gofer.go b/pkg/sentry/fsimpl/gofer/gofer.go index 63e589859..80668ebc1 100644 --- a/pkg/sentry/fsimpl/gofer/gofer.go +++ b/pkg/sentry/fsimpl/gofer/gofer.go @@ -26,6 +26,9 @@ // *** "memmap.Mappable locks taken by Translate" below this point // dentry.handleMu // dentry.dataMu +// filesystem.inoMu +// specialFileFD.mu +// specialFileFD.bufMu // // Locking dentry.dirMu in multiple dentries requires that either ancestor // dentries are locked before descendant dentries, or that filesystem.renameMu @@ -36,7 +39,6 @@ import ( "fmt" "strconv" "strings" - "sync" "sync/atomic" "syscall" @@ -44,6 +46,8 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/p9" + refs_vfs1 "gvisor.dev/gvisor/pkg/refs" + "gvisor.dev/gvisor/pkg/refsvfs2" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -53,6 +57,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/unet" "gvisor.dev/gvisor/pkg/usermem" @@ -62,9 +67,13 @@ import ( const Name = "9p" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -77,7 +86,7 @@ type filesystem struct { iopts InternalFilesystemOptions // client is the client used by this filesystem. client is immutable. - client *p9.Client + client *p9.Client `state:"nosave"` // clock is a realtime clock used to set timestamps in file operations. clock ktime.Clock @@ -85,6 +94,9 @@ type filesystem struct { // devMinor is the filesystem's minor device number. devMinor is immutable. devMinor uint32 + // root is the root dentry. root is immutable. + root *dentry + // renameMu serves two purposes: // // - It synchronizes path resolution with renaming initiated by this @@ -95,43 +107,42 @@ type filesystem struct { // reference count (such that it is usable as vfs.ResolvingPath.Start() or // is reachable from its children), or if it is a child dentry (such that // it is reachable from its parent). - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` // cachedDentries contains all dentries with 0 references. (Due to race // conditions, it may also contain dentries with non-zero references.) - // cachedDentriesLen is the number of dentries in cachedDentries. These - // fields are protected by renameMu. + // cachedDentriesLen is the number of dentries in cachedDentries. These fields + // are protected by renameMu. cachedDentries dentryList cachedDentriesLen uint64 - // syncableDentries contains all dentries in this filesystem for which - // !dentry.file.isNil(). specialFileFDs contains all open specialFileFDs. - // These fields are protected by syncMu. - syncMu sync.Mutex + // syncableDentries contains all non-synthetic dentries. specialFileFDs + // contains all open specialFileFDs. These fields are protected by syncMu. + syncMu sync.Mutex `state:"nosave"` syncableDentries map[*dentry]struct{} specialFileFDs map[*specialFileFD]struct{} - // syntheticSeq stores a counter to used to generate unique inodeNumber for - // synthetic dentries. - syntheticSeq uint64 -} + // inoByQIDPath maps previously-observed QID.Paths to inode numbers + // assigned to those paths. inoByQIDPath is not preserved across + // checkpoint/restore because QIDs may be reused between different gofer + // processes, so QIDs may be repeated for different files across + // checkpoint/restore. inoByQIDPath is protected by inoMu. + inoMu sync.Mutex `state:"nosave"` + inoByQIDPath map[uint64]uint64 `state:"nosave"` -// inodeNumber represents inode number reported in Dirent.Ino. For regular -// dentries, it comes from QID.Path from the 9P server. Synthetic dentries -// have have their inodeNumber generated sequentially, with the MSB reserved to -// prevent conflicts with regular dentries. -type inodeNumber uint64 + // lastIno is the last inode number assigned to a file. lastIno is accessed + // using atomic memory operations. + lastIno uint64 -// Reserve MSB for synthetic mounts. -const syntheticInoMask = uint64(1) << 63 + // savedDentryRW records open read/write handles during save/restore. + savedDentryRW map[*dentry]savedDentryRW -func inoFromPath(path uint64) inodeNumber { - if path&syntheticInoMask != 0 { - log.Warningf("Dropping MSB from ino, collision is possible. Original: %d, new: %d", path, path&^syntheticInoMask) - } - return inodeNumber(path &^ syntheticInoMask) + // released is nonzero once filesystem.Release has been called. It is accessed + // with atomic memory operations. + released int32 } +// +stateify savable type filesystemOptions struct { // "Standard" 9P options. fd int @@ -142,8 +153,7 @@ type filesystemOptions struct { msize uint32 version string - // maxCachedDentries is the maximum number of dentries with 0 references - // retained by the client. + // maxCachedDentries is the maximum size of filesystem.cachedDentries. maxCachedDentries uint64 // If forcePageCache is true, host FDs may not be used for application @@ -177,6 +187,8 @@ type filesystemOptions struct { // InteropMode controls the client's interaction with other remote filesystem // users. +// +// +stateify savable type InteropMode uint32 const ( @@ -195,11 +207,7 @@ const ( // and consistent with Linux's semantics (in particular, it is not always // possible for clients to set arbitrary atimes and mtimes depending on the // remote filesystem implementation, and never possible for clients to set - // arbitrary ctimes.) If a dentry containing a client-defined atime or - // mtime is evicted from cache, client timestamps will be sent to the - // remote filesystem on a best-effort basis to attempt to ensure that - // timestamps will be preserved when another dentry representing the same - // file is instantiated. + // arbitrary ctimes.) InteropModeExclusive InteropMode = iota // InteropModeWritethrough is appropriate when there are read-only users of @@ -239,7 +247,13 @@ const ( // InternalFilesystemOptions may be passed as // vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem. +// +// +stateify savable type InternalFilesystemOptions struct { + // If UniqueID is non-empty, it is an opaque string used to reassociate the + // filesystem with a new server FD during restoration from checkpoint. + UniqueID string + // If LeakConnection is true, do not close the connection to the server // when the Filesystem is released. This is necessary for deployments in // which servers can handle only a single client and report failure if that @@ -265,6 +279,9 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { mfp := pgalloc.MemoryFileProviderFromContext(ctx) @@ -276,46 +293,11 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt mopts := vfs.GenericParseMountOptions(opts.Data) var fsopts filesystemOptions - // Check that the transport is "fd". - trans, ok := mopts["trans"] - if !ok { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: transport must be specified as 'trans=fd'") - return nil, nil, syserror.EINVAL - } - delete(mopts, "trans") - if trans != "fd" { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: unsupported transport: trans=%s", trans) - return nil, nil, syserror.EINVAL - } - - // Check that read and write FDs are provided and identical. - rfdstr, ok := mopts["rfdno"] - if !ok { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD must be specified as 'rfdno=<file descriptor>") - return nil, nil, syserror.EINVAL - } - delete(mopts, "rfdno") - rfd, err := strconv.Atoi(rfdstr) + fd, err := getFDFromMountOptionsMap(ctx, mopts) if err != nil { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid read FD: rfdno=%s", rfdstr) - return nil, nil, syserror.EINVAL - } - wfdstr, ok := mopts["wfdno"] - if !ok { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: write FD must be specified as 'wfdno=<file descriptor>") - return nil, nil, syserror.EINVAL - } - delete(mopts, "wfdno") - wfd, err := strconv.Atoi(wfdstr) - if err != nil { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: invalid write FD: wfdno=%s", wfdstr) - return nil, nil, syserror.EINVAL - } - if rfd != wfd { - ctx.Warningf("gofer.FilesystemType.GetFilesystem: read FD (%d) and write FD (%d) must be equal", rfd, wfd) - return nil, nil, syserror.EINVAL + return nil, nil, err } - fsopts.fd = rfd + fsopts.fd = fd // Get the attach name. fsopts.aname = "/" @@ -431,56 +413,43 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } // If !ok, iopts being the zero value is correct. - // Establish a connection with the server. - conn, err := unet.NewSocket(fsopts.fd) + // Construct the filesystem object. + devMinor, err := vfsObj.GetAnonBlockDevMinor() if err != nil { return nil, nil, err } + fs := &filesystem{ + mfp: mfp, + opts: fsopts, + iopts: iopts, + clock: ktime.RealtimeClockFromContext(ctx), + devMinor: devMinor, + syncableDentries: make(map[*dentry]struct{}), + specialFileFDs: make(map[*specialFileFD]struct{}), + inoByQIDPath: make(map[uint64]uint64), + } + fs.vfsfs.Init(vfsObj, &fstype, fs) - // Perform version negotiation with the server. - ctx.UninterruptibleSleepStart(false) - client, err := p9.NewClient(conn, fsopts.msize, fsopts.version) - ctx.UninterruptibleSleepFinish(false) - if err != nil { - conn.Close() + // Connect to the server. + if err := fs.dial(ctx); err != nil { return nil, nil, err } - // Ownership of conn has been transferred to client. // Perform attach to obtain the filesystem root. ctx.UninterruptibleSleepStart(false) - attached, err := client.Attach(fsopts.aname) + attached, err := fs.client.Attach(fsopts.aname) ctx.UninterruptibleSleepFinish(false) if err != nil { - client.Close() + fs.vfsfs.DecRef(ctx) return nil, nil, err } attachFile := p9file{attached} qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) if err != nil { attachFile.close(ctx) - client.Close() - return nil, nil, err - } - - // Construct the filesystem object. - devMinor, err := vfsObj.GetAnonBlockDevMinor() - if err != nil { - attachFile.close(ctx) - client.Close() + fs.vfsfs.DecRef(ctx) return nil, nil, err } - fs := &filesystem{ - mfp: mfp, - opts: fsopts, - iopts: iopts, - client: client, - clock: ktime.RealtimeClockFromContext(ctx), - devMinor: devMinor, - syncableDentries: make(map[*dentry]struct{}), - specialFileFDs: make(map[*specialFileFD]struct{}), - } - fs.vfsfs.Init(vfsObj, &fstype, fs) // Construct the root dentry. root, err := fs.newDentry(ctx, attachFile, qid, attrMask, &attr) @@ -490,25 +459,87 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, err } // Set the root's reference count to 2. One reference is returned to the - // caller, and the other is deliberately leaked to prevent the root from - // being "cached" and subsequently evicted. Its resources will still be - // cleaned up by fs.Release(). + // caller, and the other is held by fs to prevent the root from being "cached" + // and subsequently evicted. root.refs = 2 + fs.root = root return &fs.vfsfs, &root.vfsd, nil } +func getFDFromMountOptionsMap(ctx context.Context, mopts map[string]string) (int, error) { + // Check that the transport is "fd". + trans, ok := mopts["trans"] + if !ok || trans != "fd" { + ctx.Warningf("gofer.getFDFromMountOptionsMap: transport must be specified as 'trans=fd'") + return -1, syserror.EINVAL + } + delete(mopts, "trans") + + // Check that read and write FDs are provided and identical. + rfdstr, ok := mopts["rfdno"] + if !ok { + ctx.Warningf("gofer.getFDFromMountOptionsMap: read FD must be specified as 'rfdno=<file descriptor>'") + return -1, syserror.EINVAL + } + delete(mopts, "rfdno") + rfd, err := strconv.Atoi(rfdstr) + if err != nil { + ctx.Warningf("gofer.getFDFromMountOptionsMap: invalid read FD: rfdno=%s", rfdstr) + return -1, syserror.EINVAL + } + wfdstr, ok := mopts["wfdno"] + if !ok { + ctx.Warningf("gofer.getFDFromMountOptionsMap: write FD must be specified as 'wfdno=<file descriptor>'") + return -1, syserror.EINVAL + } + delete(mopts, "wfdno") + wfd, err := strconv.Atoi(wfdstr) + if err != nil { + ctx.Warningf("gofer.getFDFromMountOptionsMap: invalid write FD: wfdno=%s", wfdstr) + return -1, syserror.EINVAL + } + if rfd != wfd { + ctx.Warningf("gofer.getFDFromMountOptionsMap: read FD (%d) and write FD (%d) must be equal", rfd, wfd) + return -1, syserror.EINVAL + } + return rfd, nil +} + +// Preconditions: fs.client == nil. +func (fs *filesystem) dial(ctx context.Context) error { + // Establish a connection with the server. + conn, err := unet.NewSocket(fs.opts.fd) + if err != nil { + return err + } + + // Perform version negotiation with the server. + ctx.UninterruptibleSleepStart(false) + client, err := p9.NewClient(conn, fs.opts.msize, fs.opts.version) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + conn.Close() + return err + } + // Ownership of conn has been transferred to client. + + fs.client = client + return nil +} + // Release implements vfs.FilesystemImpl.Release. func (fs *filesystem) Release(ctx context.Context) { - mf := fs.mfp.MemoryFile() + atomic.StoreInt32(&fs.released, 1) + mf := fs.mfp.MemoryFile() fs.syncMu.Lock() for d := range fs.syncableDentries { d.handleMu.Lock() d.dataMu.Lock() if h := d.writeHandleLocked(); h.isOpen() { // Write dirty cached data to the remote file. - if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, fs.mfp.MemoryFile(), h.writeFromBlocksAt); err != nil { + if err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, mf, h.writeFromBlocksAt); err != nil { log.Warningf("gofer.filesystem.Release: failed to flush dentry: %v", err) } // TODO(jamieliu): Do we need to flushf/fsync d? @@ -529,6 +560,21 @@ func (fs *filesystem) Release(ctx context.Context) { // fs. fs.syncMu.Unlock() + // If leak checking is enabled, release all outstanding references in the + // filesystem. We deliberately avoid doing this outside of leak checking; we + // have released all external resources above rather than relying on dentry + // destructors. + if refs_vfs1.GetLeakMode() != refs_vfs1.NoLeakChecking { + fs.renameMu.Lock() + fs.root.releaseSyntheticRecursiveLocked(ctx) + fs.evictAllCachedDentriesLocked(ctx) + fs.renameMu.Unlock() + + // An extra reference was held by the filesystem on the root to prevent it from + // being cached/evicted. + fs.root.DecRef(ctx) + } + if !fs.iopts.LeakConnection { // Close the connection to the server. This implicitly clunks all fids. fs.client.Close() @@ -537,7 +583,34 @@ func (fs *filesystem) Release(ctx context.Context) { fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) } +// releaseSyntheticRecursiveLocked traverses the tree with root d and decrements +// the reference count on every synthetic dentry. Synthetic dentries have one +// reference for existence that should be dropped during filesystem.Release. +// +// Precondition: d.fs.renameMu is locked. +func (d *dentry) releaseSyntheticRecursiveLocked(ctx context.Context) { + if d.isSynthetic() { + d.decRefLocked() + d.checkCachingLocked(ctx) + } + if d.isDir() { + var children []*dentry + d.dirMu.Lock() + for _, child := range d.children { + children = append(children, child) + } + d.dirMu.Unlock() + for _, child := range children { + if child != nil { + child.releaseSyntheticRecursiveLocked(ctx) + } + } + } +} + // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -562,12 +635,15 @@ type dentry struct { // filesystem.renameMu. name string + // qidPath is the p9.QID.Path for this file. qidPath is immutable. + qidPath uint64 + // file is the unopened p9.File that backs this dentry. file is immutable. // // If file.isNil(), this dentry represents a synthetic file, i.e. a file // that does not exist on the remote filesystem. As of this writing, the // only files that can be synthetic are sockets, pipes, and directories. - file p9file + file p9file `state:"nosave"` // If deleted is non-zero, the file represented by this dentry has been // deleted. deleted is accessed using atomic memory operations. @@ -579,7 +655,7 @@ type dentry struct { cached bool dentryEntry - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` // If this dentry represents a directory, children contains: // @@ -611,12 +687,12 @@ type dentry struct { // To mutate: // - Lock metadataMu and use atomic operations to update because we might // have atomic readers that don't hold the lock. - metadataMu sync.Mutex - ino inodeNumber // immutable - mode uint32 // type is immutable, perms are mutable - uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic - gid uint32 // auth.KGID, but ... - blockSize uint32 // 0 if unknown + metadataMu sync.Mutex `state:"nosave"` + ino uint64 // immutable + mode uint32 // type is immutable, perms are mutable + uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic + gid uint32 // auth.KGID, but ... + blockSize uint32 // 0 if unknown // Timestamps, all nsecs from the Unix epoch. atime int64 mtime int64 @@ -642,7 +718,7 @@ type dentry struct { // other metadata fields. nlink uint32 - mapsMu sync.Mutex + mapsMu sync.Mutex `state:"nosave"` // If this dentry represents a regular file, mappings tracks mappings of // the file into memmap.MappingSpaces. mappings is protected by mapsMu. @@ -666,12 +742,12 @@ type dentry struct { // either p9.File transitions from closed (isNil() == true) to open // (isNil() == false), it may be mutated with handleMu locked, but cannot // be closed until the dentry is destroyed. - handleMu sync.RWMutex - readFile p9file - writeFile p9file - hostFD int32 + handleMu sync.RWMutex `state:"nosave"` + readFile p9file `state:"nosave"` + writeFile p9file `state:"nosave"` + hostFD int32 `state:"nosave"` - dataMu sync.RWMutex + dataMu sync.RWMutex `state:"nosave"` // If this dentry represents a regular file that is client-cached, cache // maps offsets into the cached file to offsets into @@ -703,6 +779,13 @@ type dentry struct { locks vfs.FileLocks // Inotify watches for this dentry. + // + // Note that inotify may behave unexpectedly in the presence of hard links, + // because dentries corresponding to the same file have separate inotify + // watches when they should share the same set. This is the case because it is + // impossible for us to know for sure whether two dentries correspond to the + // same underlying file (see the gofer filesystem section fo vfs/inotify.md for + // a more in-depth discussion on this matter). watches vfs.Watches } @@ -739,8 +822,9 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma d := &dentry{ fs: fs, + qidPath: qid.Path, file: file, - ino: inoFromPath(qid.Path), + ino: fs.inoFromQIDPath(qid.Path), mode: uint32(attr.Mode), uid: uint32(fs.opts.dfltuid), gid: uint32(fs.opts.dfltgid), @@ -776,6 +860,9 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma d.nlink = uint32(attr.NLink) } d.vfsd.Init(d) + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Register(d, "gofer.dentry") + } fs.syncMu.Lock() fs.syncableDentries[d] = struct{}{} @@ -783,6 +870,21 @@ func (fs *filesystem) newDentry(ctx context.Context, file p9file, qid p9.QID, ma return d, nil } +func (fs *filesystem) inoFromQIDPath(qidPath uint64) uint64 { + fs.inoMu.Lock() + defer fs.inoMu.Unlock() + if ino, ok := fs.inoByQIDPath[qidPath]; ok { + return ino + } + ino := fs.nextIno() + fs.inoByQIDPath[qidPath] = ino + return ino +} + +func (fs *filesystem) nextIno() uint64 { + return atomic.AddUint64(&fs.lastIno, 1) +} + func (d *dentry) isSynthetic() bool { return d.file.isNil() } @@ -830,11 +932,11 @@ func (d *dentry) updateFromP9AttrsLocked(mask p9.AttrMask, attr *p9.Attr) { atomic.StoreUint32(&d.nlink, uint32(attr.NLink)) } if mask.Size { - d.updateFileSizeLocked(attr.Size) + d.updateSizeLocked(attr.Size) } } -// Preconditions: !d.isSynthetic() +// Preconditions: !d.isSynthetic(). func (d *dentry) updateFromGetattr(ctx context.Context) error { // Use d.readFile or d.writeFile, which represent 9P fids that have been // opened, in preference to d.file, which represents a 9P fid that has not. @@ -897,10 +999,10 @@ func (d *dentry) statTo(stat *linux.Statx) { // This is consistent with regularFileFD.Seek(), which treats regular files // as having no holes. stat.Blocks = (stat.Size + 511) / 512 - stat.Atime = statxTimestampFromDentry(atomic.LoadInt64(&d.atime)) - stat.Btime = statxTimestampFromDentry(atomic.LoadInt64(&d.btime)) - stat.Ctime = statxTimestampFromDentry(atomic.LoadInt64(&d.ctime)) - stat.Mtime = statxTimestampFromDentry(atomic.LoadInt64(&d.mtime)) + stat.Atime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&d.atime)) + stat.Btime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&d.btime)) + stat.Ctime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&d.ctime)) + stat.Mtime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&d.mtime)) stat.DevMajor = linux.UNNAMED_MAJOR stat.DevMinor = d.fs.devMinor } @@ -948,10 +1050,10 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // Use client clocks for timestamps. now = d.fs.clock.Now().Nanoseconds() if stat.Mask&linux.STATX_ATIME != 0 && stat.Atime.Nsec == linux.UTIME_NOW { - stat.Atime = statxTimestampFromDentry(now) + stat.Atime = linux.NsecToStatxTimestamp(now) } if stat.Mask&linux.STATX_MTIME != 0 && stat.Mtime.Nsec == linux.UTIME_NOW { - stat.Mtime = statxTimestampFromDentry(now) + stat.Mtime = linux.NsecToStatxTimestamp(now) } } @@ -984,7 +1086,7 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // d.size should be kept up to date, and privatized // copy-on-write mappings of truncated pages need to be // invalidated, even if InteropModeShared is in effect. - d.updateFileSizeLocked(stat.Size) + d.updateSizeLocked(stat.Size) } } if d.fs.opts.interop == InteropModeShared { @@ -1010,19 +1112,42 @@ func (d *dentry) setStat(ctx context.Context, creds *auth.Credentials, opts *vfs // !d.cachedMetadataAuthoritative() then we returned after calling // d.file.setAttr(). For the same reason, now must have been initialized. if stat.Mask&linux.STATX_ATIME != 0 { - atomic.StoreInt64(&d.atime, dentryTimestampFromStatx(stat.Atime)) + atomic.StoreInt64(&d.atime, stat.Atime.ToNsec()) atomic.StoreUint32(&d.atimeDirty, 0) } if stat.Mask&linux.STATX_MTIME != 0 { - atomic.StoreInt64(&d.mtime, dentryTimestampFromStatx(stat.Mtime)) + atomic.StoreInt64(&d.mtime, stat.Mtime.ToNsec()) atomic.StoreUint32(&d.mtimeDirty, 0) } atomic.StoreInt64(&d.ctime, now) return nil } +// doAllocate performs an allocate operation on d. Note that d.metadataMu will +// be held when allocate is called. +func (d *dentry) doAllocate(ctx context.Context, offset, length uint64, allocate func() error) error { + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + + // Allocating a smaller size is a noop. + size := offset + length + if d.cachedMetadataAuthoritative() && size <= d.size { + return nil + } + + err := allocate() + if err != nil { + return err + } + d.updateSizeLocked(size) + if d.cachedMetadataAuthoritative() { + d.touchCMtimeLocked() + } + return nil +} + // Preconditions: d.metadataMu must be locked. -func (d *dentry) updateFileSizeLocked(newSize uint64) { +func (d *dentry) updateSizeLocked(newSize uint64) { d.dataMu.Lock() oldSize := d.size atomic.StoreUint64(&d.size, newSize) @@ -1060,6 +1185,21 @@ func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) } +func (d *dentry) checkXattrPermissions(creds *auth.Credentials, name string, ats vfs.AccessTypes) error { + // We only support xattrs prefixed with "user." (see b/148380782). Currently, + // there is no need to expose any other xattrs through a gofer. + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return syserror.EOPNOTSUPP + } + mode := linux.FileMode(atomic.LoadUint32(&d.mode)) + kuid := auth.KUID(atomic.LoadUint32(&d.uid)) + kgid := auth.KGID(atomic.LoadUint32(&d.gid)) + if err := vfs.GenericCheckPermissions(creds, ats, mode, kuid, kgid); err != nil { + return err + } + return vfs.CheckXattrPermissions(creds, ats, mode, kuid, name) +} + func (d *dentry) mayDelete(creds *auth.Credentials, child *dentry) error { return vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&child.uid))) } @@ -1118,6 +1258,11 @@ func (d *dentry) decRefLocked() { } } +// LeakMessage implements refsvfs2.CheckedObject.LeakMessage. +func (d *dentry) LeakMessage() string { + return fmt.Sprintf("[gofer.dentry %p] reference count of %d instead of -1", d, atomic.LoadInt64(&d.refs)) +} + // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. func (d *dentry) InotifyWithParent(ctx context.Context, events, cookie uint32, et vfs.EventType) { if d.isDir() { @@ -1166,6 +1311,10 @@ func (d *dentry) checkCachingLocked(ctx context.Context) { // resolution, which requires renameMu, so if d.refs is zero then it will // remain zero while we hold renameMu for writing.) refs := atomic.LoadInt64(&d.refs) + if refs == -1 { + // Dentry has already been destroyed. + return + } if refs > 0 { if d.cached { d.fs.cachedDentries.Remove(d) @@ -1174,10 +1323,6 @@ func (d *dentry) checkCachingLocked(ctx context.Context) { } return } - if refs == -1 { - // Dentry has already been destroyed. - return - } // Deleted and invalidated dentries with zero references are no longer // reachable by path resolution and should be dropped immediately. if d.vfsd.IsDead() { @@ -1200,6 +1345,16 @@ func (d *dentry) checkCachingLocked(ctx context.Context) { if d.watches.Size() > 0 { return } + + if atomic.LoadInt32(&d.fs.released) != 0 { + if d.parent != nil { + d.parent.dirMu.Lock() + delete(d.parent.children, d.name) + d.parent.dirMu.Unlock() + } + d.destroyLocked(ctx) + } + // If d is already cached, just move it to the front of the LRU. if d.cached { d.fs.cachedDentries.Remove(d) @@ -1212,33 +1367,48 @@ func (d *dentry) checkCachingLocked(ctx context.Context) { d.fs.cachedDentriesLen++ d.cached = true if d.fs.cachedDentriesLen > d.fs.opts.maxCachedDentries { - victim := d.fs.cachedDentries.Back() - d.fs.cachedDentries.Remove(victim) - d.fs.cachedDentriesLen-- - victim.cached = false - // victim.refs may have become non-zero from an earlier path resolution - // since it was inserted into fs.cachedDentries. - if atomic.LoadInt64(&victim.refs) == 0 { - if victim.parent != nil { - victim.parent.dirMu.Lock() - if !victim.vfsd.IsDead() { - // Note that victim can't be a mount point (in any mount - // namespace), since VFS holds references on mount points. - d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, &victim.vfsd) - delete(victim.parent.children, victim.name) - // We're only deleting the dentry, not the file it - // represents, so we don't need to update - // victimParent.dirents etc. - } - victim.parent.dirMu.Unlock() - } - victim.destroyLocked(ctx) - } + d.fs.evictCachedDentryLocked(ctx) // Whether or not victim was destroyed, we brought fs.cachedDentriesLen // back down to fs.opts.maxCachedDentries, so we don't loop. } } +// Precondition: fs.renameMu must be locked for writing; it may be temporarily +// unlocked. +func (fs *filesystem) evictAllCachedDentriesLocked(ctx context.Context) { + for fs.cachedDentriesLen != 0 { + fs.evictCachedDentryLocked(ctx) + } +} + +// Preconditions: +// * fs.renameMu must be locked for writing; it may be temporarily unlocked. +// * fs.cachedDentriesLen != 0. +func (fs *filesystem) evictCachedDentryLocked(ctx context.Context) { + victim := fs.cachedDentries.Back() + fs.cachedDentries.Remove(victim) + fs.cachedDentriesLen-- + victim.cached = false + // victim.refs may have become non-zero from an earlier path resolution + // since it was inserted into fs.cachedDentries. + if atomic.LoadInt64(&victim.refs) == 0 { + if victim.parent != nil { + victim.parent.dirMu.Lock() + if !victim.vfsd.IsDead() { + // Note that victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount points. + fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, &victim.vfsd) + delete(victim.parent.children, victim.name) + // We're only deleting the dentry, not the file it + // represents, so we don't need to update + // victimParent.dirents etc. + } + victim.parent.dirMu.Unlock() + } + victim.destroyLocked(ctx) + } +} + // destroyLocked destroys the dentry. // // Preconditions: @@ -1293,30 +1463,19 @@ func (d *dentry) destroyLocked(ctx context.Context) { d.handleMu.Unlock() if !d.file.isNil() { - if !d.isDeleted() { - // Write dirty timestamps back to the remote filesystem. - atimeDirty := atomic.LoadUint32(&d.atimeDirty) != 0 - mtimeDirty := atomic.LoadUint32(&d.mtimeDirty) != 0 - if atimeDirty || mtimeDirty { - atime := atomic.LoadInt64(&d.atime) - mtime := atomic.LoadInt64(&d.mtime) - if err := d.file.setAttr(ctx, p9.SetAttrMask{ - ATime: atimeDirty, - ATimeNotSystemTime: atimeDirty, - MTime: mtimeDirty, - MTimeNotSystemTime: mtimeDirty, - }, p9.SetAttr{ - ATimeSeconds: uint64(atime / 1e9), - ATimeNanoSeconds: uint64(atime % 1e9), - MTimeSeconds: uint64(mtime / 1e9), - MTimeNanoSeconds: uint64(mtime % 1e9), - }); err != nil { - log.Warningf("gofer.dentry.destroyLocked: failed to write dirty timestamps back: %v", err) - } - } + // Note that it's possible that d.atimeDirty or d.mtimeDirty are true, + // i.e. client and server timestamps may differ (because e.g. a client + // write was serviced by the page cache, and only written back to the + // remote file later). Ideally, we'd write client timestamps back to + // the remote filesystem so that timestamps for a new dentry + // instantiated for the same file would remain coherent. Unfortunately, + // this turns out to be too expensive in many cases, so for now we + // don't do this. + if err := d.file.close(ctx); err != nil { + log.Warningf("gofer.dentry.destroyLocked: failed to close file: %v", err) } - d.file.close(ctx) d.file = p9file{} + // Remove d from the set of syncable dentries. d.fs.syncMu.Lock() delete(d.fs.syncableDentries, d) @@ -1334,6 +1493,10 @@ func (d *dentry) destroyLocked(ctx context.Context) { panic("gofer.dentry.DecRef() called without holding a reference") } } + + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Unregister(d, "gofer.dentry") + } } func (d *dentry) isDeleted() bool { @@ -1344,9 +1507,7 @@ func (d *dentry) setDeleted() { atomic.StoreUint32(&d.deleted, 1) } -// We only support xattrs prefixed with "user." (see b/148380782). Currently, -// there is no need to expose any other xattrs through a gofer. -func (d *dentry) listxattr(ctx context.Context, creds *auth.Credentials, size uint64) ([]string, error) { +func (d *dentry) listXattr(ctx context.Context, creds *auth.Credentials, size uint64) ([]string, error) { if d.file.isNil() || !d.userXattrSupported() { return nil, nil } @@ -1356,6 +1517,7 @@ func (d *dentry) listxattr(ctx context.Context, creds *auth.Credentials, size ui } xattrs := make([]string, 0, len(xattrMap)) for x := range xattrMap { + // We only support xattrs in the user.* namespace. if strings.HasPrefix(x, linux.XATTR_USER_PREFIX) { xattrs = append(xattrs, x) } @@ -1363,51 +1525,33 @@ func (d *dentry) listxattr(ctx context.Context, creds *auth.Credentials, size ui return xattrs, nil } -func (d *dentry) getxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { +func (d *dentry) getXattr(ctx context.Context, creds *auth.Credentials, opts *vfs.GetXattrOptions) (string, error) { if d.file.isNil() { return "", syserror.ENODATA } - if err := d.checkPermissions(creds, vfs.MayRead); err != nil { + if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayRead); err != nil { return "", err } - if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { - return "", syserror.EOPNOTSUPP - } - if !d.userXattrSupported() { - return "", syserror.ENODATA - } return d.file.getXattr(ctx, opts.Name, opts.Size) } -func (d *dentry) setxattr(ctx context.Context, creds *auth.Credentials, opts *vfs.SetxattrOptions) error { +func (d *dentry) setXattr(ctx context.Context, creds *auth.Credentials, opts *vfs.SetXattrOptions) error { if d.file.isNil() { return syserror.EPERM } - if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayWrite); err != nil { return err } - if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { - return syserror.EOPNOTSUPP - } - if !d.userXattrSupported() { - return syserror.EPERM - } return d.file.setXattr(ctx, opts.Name, opts.Value, opts.Flags) } -func (d *dentry) removexattr(ctx context.Context, creds *auth.Credentials, name string) error { +func (d *dentry) removeXattr(ctx context.Context, creds *auth.Credentials, name string) error { if d.file.isNil() { return syserror.EPERM } - if err := d.checkPermissions(creds, vfs.MayWrite); err != nil { + if err := d.checkXattrPermissions(creds, name, vfs.MayWrite); err != nil { return err } - if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { - return syserror.EOPNOTSUPP - } - if !d.userXattrSupported() { - return syserror.EPERM - } return d.file.removeXattr(ctx, name) } @@ -1418,7 +1562,9 @@ func (d *dentry) userXattrSupported() bool { return filetype == linux.ModeRegular || filetype == linux.ModeDirectory } -// Preconditions: !d.isSynthetic(). d.isRegularFile() || d.isDir(). +// Preconditions: +// * !d.isSynthetic(). +// * d.isRegularFile() || d.isDir(). func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool) error { // O_TRUNC unconditionally requires us to obtain a new handle (opened with // O_TRUNC). @@ -1463,8 +1609,9 @@ func (d *dentry) ensureSharedHandle(ctx context.Context, read, write, trunc bool return err } - if d.hostFD < 0 && openReadable && h.fd >= 0 { - // We have no existing FD; use the new FD for at least reading. + if d.hostFD < 0 && h.fd >= 0 && openReadable && (d.writeFile.isNil() || openWritable) { + // We have no existing FD, and the new FD meets the requirements + // for d.hostFD, so start using it. d.hostFD = h.fd } else if d.hostFD >= 0 && d.writeFile.isNil() && openWritable { // We have an existing read-only FD, but the file has just been @@ -1593,6 +1740,33 @@ func (d *dentry) syncRemoteFileLocked(ctx context.Context) error { return nil } +func (d *dentry) syncCachedFile(ctx context.Context, forFilesystemSync bool) error { + d.handleMu.RLock() + defer d.handleMu.RUnlock() + h := d.writeHandleLocked() + if h.isOpen() { + // Write back dirty pages to the remote file. + d.dataMu.Lock() + err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), h.writeFromBlocksAt) + d.dataMu.Unlock() + if err != nil { + return err + } + } + if err := d.syncRemoteFileLocked(ctx); err != nil { + if !forFilesystemSync { + return err + } + // Only return err if we can reasonably have expected sync to succeed + // (d is a regular file and was opened for writing). + if d.isRegularFile() && h.isOpen() { + return err + } + ctx.Debugf("gofer.dentry.syncCachedFile: syncing non-writable or non-regular-file dentry failed: %v", err) + } + return nil +} + // incLinks increments link count. func (d *dentry) incLinks() { if atomic.LoadUint32(&d.nlink) == 0 { @@ -1613,12 +1787,14 @@ func (d *dentry) decLinks() { // fileDescription is embedded by gofer implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.LockFD - lockLogging sync.Once + lockLogging sync.Once `state:"nosave"` } func (fd *fileDescription) filesystem() *filesystem { @@ -1656,30 +1832,30 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) return nil } -// Listxattr implements vfs.FileDescriptionImpl.Listxattr. -func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { - return fd.dentry().listxattr(ctx, auth.CredentialsFromContext(ctx), size) +// ListXattr implements vfs.FileDescriptionImpl.ListXattr. +func (fd *fileDescription) ListXattr(ctx context.Context, size uint64) ([]string, error) { + return fd.dentry().listXattr(ctx, auth.CredentialsFromContext(ctx), size) } -// Getxattr implements vfs.FileDescriptionImpl.Getxattr. -func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { - return fd.dentry().getxattr(ctx, auth.CredentialsFromContext(ctx), &opts) +// GetXattr implements vfs.FileDescriptionImpl.GetXattr. +func (fd *fileDescription) GetXattr(ctx context.Context, opts vfs.GetXattrOptions) (string, error) { + return fd.dentry().getXattr(ctx, auth.CredentialsFromContext(ctx), &opts) } -// Setxattr implements vfs.FileDescriptionImpl.Setxattr. -func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { +// SetXattr implements vfs.FileDescriptionImpl.SetXattr. +func (fd *fileDescription) SetXattr(ctx context.Context, opts vfs.SetXattrOptions) error { d := fd.dentry() - if err := d.setxattr(ctx, auth.CredentialsFromContext(ctx), &opts); err != nil { + if err := d.setXattr(ctx, auth.CredentialsFromContext(ctx), &opts); err != nil { return err } d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil } -// Removexattr implements vfs.FileDescriptionImpl.Removexattr. -func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { +// RemoveXattr implements vfs.FileDescriptionImpl.RemoveXattr. +func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error { d := fd.dentry() - if err := d.removexattr(ctx, auth.CredentialsFromContext(ctx), name); err != nil { + if err := d.removeXattr(ctx, auth.CredentialsFromContext(ctx), name); err != nil { return err } d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) diff --git a/pkg/sentry/fsimpl/gofer/gofer_test.go b/pkg/sentry/fsimpl/gofer/gofer_test.go index bfe75dfe4..76f08e252 100644 --- a/pkg/sentry/fsimpl/gofer/gofer_test.go +++ b/pkg/sentry/fsimpl/gofer/gofer_test.go @@ -26,12 +26,13 @@ import ( func TestDestroyIdempotent(t *testing.T) { ctx := contexttest.Context(t) fs := filesystem{ - mfp: pgalloc.MemoryFileProviderFromContext(ctx), - syncableDentries: make(map[*dentry]struct{}), + mfp: pgalloc.MemoryFileProviderFromContext(ctx), opts: filesystemOptions{ // Test relies on no dentry being held in the cache. maxCachedDentries: 0, }, + syncableDentries: make(map[*dentry]struct{}), + inoByQIDPath: make(map[uint64]uint64), } attr := &p9.Attr{ diff --git a/pkg/sentry/fsimpl/gofer/handle.go b/pkg/sentry/fsimpl/gofer/handle.go index 104157512..a9ebe1206 100644 --- a/pkg/sentry/fsimpl/gofer/handle.go +++ b/pkg/sentry/fsimpl/gofer/handle.go @@ -25,6 +25,8 @@ import ( // handle represents a remote "open file descriptor", consisting of an opened // fid (p9.File) and optionally a host file descriptor. +// +// These are explicitly not savable. type handle struct { file p9file fd int32 // -1 if unavailable diff --git a/pkg/sentry/fsimpl/gofer/host_named_pipe.go b/pkg/sentry/fsimpl/gofer/host_named_pipe.go index 7294de7d6..c7bf10007 100644 --- a/pkg/sentry/fsimpl/gofer/host_named_pipe.go +++ b/pkg/sentry/fsimpl/gofer/host_named_pipe.go @@ -51,8 +51,24 @@ func blockUntilNonblockingPipeHasWriter(ctx context.Context, fd int32) error { if ok { return nil } - if err := sleepBetweenNamedPipeOpenChecks(ctx); err != nil { - return err + if sleepErr := sleepBetweenNamedPipeOpenChecks(ctx); sleepErr != nil { + // Another application thread may have opened this pipe for + // writing, succeeded because we previously opened the pipe for + // reading, and subsequently interrupted us for checkpointing (e.g. + // this occurs in mknod tests under cooperative save/restore). In + // this case, our open has to succeed for the checkpoint to include + // a readable FD for the pipe, which is in turn necessary to + // restore the other thread's writable FD for the same pipe + // (otherwise it will get ENXIO). So we have to check + // nonblockingPipeHasWriter() once last time. + ok, err := nonblockingPipeHasWriter(fd) + if err != nil { + return err + } + if ok { + return nil + } + return sleepErr } } } diff --git a/pkg/sentry/fsimpl/gofer/p9file.go b/pkg/sentry/fsimpl/gofer/p9file.go index 87f0b877f..21b4a96fe 100644 --- a/pkg/sentry/fsimpl/gofer/p9file.go +++ b/pkg/sentry/fsimpl/gofer/p9file.go @@ -127,6 +127,13 @@ func (f p9file) close(ctx context.Context) error { return err } +func (f p9file) setAttrClose(ctx context.Context, valid p9.SetAttrMask, attr p9.SetAttr) error { + ctx.UninterruptibleSleepStart(false) + err := f.file.SetAttrClose(valid, attr) + ctx.UninterruptibleSleepFinish(false) + return err +} + func (f p9file) open(ctx context.Context, flags p9.OpenFlags) (*fd.FD, p9.QID, uint32, error) { ctx.UninterruptibleSleepStart(false) fdobj, qid, iounit, err := f.file.Open(flags) diff --git a/pkg/sentry/fsimpl/gofer/regular_file.go b/pkg/sentry/fsimpl/gofer/regular_file.go index 7e1cbf065..dc8a890cb 100644 --- a/pkg/sentry/fsimpl/gofer/regular_file.go +++ b/pkg/sentry/fsimpl/gofer/regular_file.go @@ -18,7 +18,6 @@ import ( "fmt" "io" "math" - "sync" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -31,6 +30,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/usage" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -39,11 +39,12 @@ func (d *dentry) isRegularFile() bool { return d.fileType() == linux.S_IFREG } +// +stateify savable type regularFileFD struct { fileDescription // off is the file offset. off is protected by mu. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 } @@ -56,10 +57,16 @@ func (fd *regularFileFD) OnClose(ctx context.Context) error { if !fd.vfsfd.IsWritable() { return nil } - // Skip flushing if writes may be buffered by the client, since (as with - // the VFS1 client) we don't flush buffered writes on close anyway. + // Skip flushing if there are client-buffered writes, since (as with the + // VFS1 client) we don't flush buffered writes on close anyway. d := fd.dentry() - if d.fs.opts.interop == InteropModeExclusive { + if d.fs.opts.interop != InteropModeExclusive { + return nil + } + d.dataMu.RLock() + haveDirtyPages := !d.dirty.IsEmpty() + d.dataMu.RUnlock() + if haveDirtyPages { return nil } d.handleMu.RLock() @@ -73,28 +80,11 @@ func (fd *regularFileFD) OnClose(ctx context.Context) error { // Allocate implements vfs.FileDescriptionImpl.Allocate. func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error { d := fd.dentry() - d.metadataMu.Lock() - defer d.metadataMu.Unlock() - - // Allocating a smaller size is a noop. - size := offset + length - if d.cachedMetadataAuthoritative() && size <= d.size { - return nil - } - - d.handleMu.RLock() - err := d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length) - d.handleMu.RUnlock() - if err != nil { - return err - } - d.dataMu.Lock() - atomic.StoreUint64(&d.size, size) - d.dataMu.Unlock() - if d.cachedMetadataAuthoritative() { - d.touchCMtimeLocked() - } - return nil + return d.doAllocate(ctx, offset, length, func() error { + d.handleMu.RLock() + defer d.handleMu.RUnlock() + return d.writeFile.allocate(ctx, p9.ToAllocateMode(mode), offset, length) + }) } // PRead implements vfs.FileDescriptionImpl.PRead. @@ -117,6 +107,10 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs return 0, io.EOF } + var ( + n int64 + readErr error + ) if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { // Lock d.metadataMu for the rest of the read to prevent d.size from // changing. @@ -127,20 +121,25 @@ func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs if err := d.writeback(ctx, offset, dst.NumBytes()); err != nil { return 0, err } - } - - rw := getDentryReadWriter(ctx, d, offset) - if fd.vfsfd.StatusFlags()&linux.O_DIRECT != 0 { + rw := getDentryReadWriter(ctx, d, offset) // Require the read to go to the remote file. rw.direct = true + n, readErr = dst.CopyOutFrom(ctx, rw) + putDentryReadWriter(rw) + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). + d.touchAtimeLocked(fd.vfsfd.Mount()) + } + } else { + rw := getDentryReadWriter(ctx, d, offset) + n, readErr = dst.CopyOutFrom(ctx, rw) + putDentryReadWriter(rw) + if d.fs.opts.interop != InteropModeShared { + // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). + d.touchAtime(fd.vfsfd.Mount()) + } } - n, err := dst.CopyOutFrom(ctx, rw) - putDentryReadWriter(rw) - if d.fs.opts.interop != InteropModeShared { - // Compare Linux's mm/filemap.c:do_generic_file_read() => file_accessed(). - d.touchAtime(fd.vfsfd.Mount()) - } - return n, err + return n, readErr } // Read implements vfs.FileDescriptionImpl.Read. @@ -396,7 +395,7 @@ func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) End: gapEnd, } optMR := gap.Range() - err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), mf, usage.PageCache, h.readToBlocksAt) + err := rw.d.cache.Fill(rw.ctx, reqMR, maxFillRange(reqMR, optMR), rw.d.size, mf, usage.PageCache, h.readToBlocksAt) mf.MarkEvictable(rw.d, pgalloc.EvictableRange{optMR.Start, optMR.End}) seg, gap = rw.d.cache.Find(rw.off) if !seg.Ok() { @@ -404,10 +403,10 @@ func (rw *dentryReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) rw.d.handleMu.RUnlock() return done, err } - // err might have occurred in part of gap.Range() outside - // gapMR. Forget about it for now; if the error matters and - // persists, we'll run into it again in a later iteration of - // this loop. + // err might have occurred in part of gap.Range() outside gapMR + // (in particular, gap.End() might be beyond EOF). Forget about + // it for now; if the error matters and persists, we'll run + // into it again in a later iteration of this loop. } else { // Read directly from the file. gapDsts := dsts.TakeFirst64(gapMR.Length()) @@ -625,23 +624,7 @@ func regularFileSeekLocked(ctx context.Context, d *dentry, fdOffset, offset int6 // Sync implements vfs.FileDescriptionImpl.Sync. func (fd *regularFileFD) Sync(ctx context.Context) error { - return fd.dentry().syncCachedFile(ctx) -} - -func (d *dentry) syncCachedFile(ctx context.Context) error { - d.handleMu.RLock() - defer d.handleMu.RUnlock() - - if h := d.writeHandleLocked(); h.isOpen() { - d.dataMu.Lock() - // Write dirty cached data to the remote file. - err := fsutil.SyncDirtyAll(ctx, &d.cache, &d.dirty, d.size, d.fs.mfp.MemoryFile(), h.writeFromBlocksAt) - d.dataMu.Unlock() - if err != nil { - return err - } - } - return d.syncRemoteFileLocked(ctx) + return fd.dentry().syncCachedFile(ctx, false /* lowSyncExpectations */) } // ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. @@ -781,7 +764,7 @@ func (d *dentry) Translate(ctx context.Context, required, optional memmap.Mappab mf := d.fs.mfp.MemoryFile() h := d.readHandleLocked() - cerr := d.cache.Fill(ctx, required, maxFillRange(required, optional), mf, usage.PageCache, h.readToBlocksAt) + cerr := d.cache.Fill(ctx, required, maxFillRange(required, optional), d.size, mf, usage.PageCache, h.readToBlocksAt) var ts []memmap.Translation var translatedEnd uint64 @@ -900,6 +883,8 @@ func (d *dentry) Evict(ctx context.Context, er pgalloc.EvictableRange) { // dentryPlatformFile is only used when a host FD representing the remote file // is available (i.e. dentry.hostFD >= 0), and that FD is used for application // memory mappings (i.e. !filesystem.opts.forcePageCache). +// +// +stateify savable type dentryPlatformFile struct { *dentry @@ -912,7 +897,7 @@ type dentryPlatformFile struct { hostFileMapper fsutil.HostFileMapper // hostFileMapperInitOnce is used to lazily initialize hostFileMapper. - hostFileMapperInitOnce sync.Once + hostFileMapperInitOnce sync.Once `state:"nosave"` } // IncRef implements memmap.File.IncRef. diff --git a/pkg/sentry/fsimpl/gofer/save_restore.go b/pkg/sentry/fsimpl/gofer/save_restore.go new file mode 100644 index 000000000..2ea224c43 --- /dev/null +++ b/pkg/sentry/fsimpl/gofer/save_restore.go @@ -0,0 +1,329 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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" + "io" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/p9" + "gvisor.dev/gvisor/pkg/refsvfs2" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +type saveRestoreContextID int + +const ( + // CtxRestoreServerFDMap is a Context.Value key for a map[string]int + // mapping filesystem unique IDs (cf. InternalFilesystemOptions.UniqueID) + // to host FDs. + CtxRestoreServerFDMap saveRestoreContextID = iota +) + +// +stateify savable +type savedDentryRW struct { + read bool + write bool +} + +// PreprareSave implements vfs.FilesystemImplSaveRestoreExtension.PrepareSave. +func (fs *filesystem) PrepareSave(ctx context.Context) error { + if len(fs.iopts.UniqueID) == 0 { + return fmt.Errorf("gofer.filesystem with no UniqueID cannot be saved") + } + + // Purge cached dentries, which may not be reopenable after restore due to + // permission changes. + fs.renameMu.Lock() + fs.evictAllCachedDentriesLocked(ctx) + fs.renameMu.Unlock() + + // Buffer pipe data so that it's available for reading after restore. (This + // is a legacy VFS1 feature.) + fs.syncMu.Lock() + for sffd := range fs.specialFileFDs { + if sffd.dentry().fileType() == linux.S_IFIFO && sffd.vfsfd.IsReadable() { + if err := sffd.savePipeData(ctx); err != nil { + fs.syncMu.Unlock() + return err + } + } + } + fs.syncMu.Unlock() + + // Flush local state to the remote filesystem. + if err := fs.Sync(ctx); err != nil { + return err + } + + fs.savedDentryRW = make(map[*dentry]savedDentryRW) + return fs.root.prepareSaveRecursive(ctx) +} + +// Preconditions: +// * fd represents a pipe. +// * fd is readable. +func (fd *specialFileFD) savePipeData(ctx context.Context) error { + fd.bufMu.Lock() + defer fd.bufMu.Unlock() + var buf [usermem.PageSize]byte + for { + n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:])), ^uint64(0)) + if n != 0 { + fd.buf = append(fd.buf, buf[:n]...) + } + if err != nil { + if err == io.EOF || err == syserror.EAGAIN { + break + } + return err + } + } + if len(fd.buf) != 0 { + atomic.StoreUint32(&fd.haveBuf, 1) + } + return nil +} + +func (d *dentry) prepareSaveRecursive(ctx context.Context) error { + if d.isRegularFile() && !d.cachedMetadataAuthoritative() { + // Get updated metadata for d in case we need to perform metadata + // validation during restore. + if err := d.updateFromGetattr(ctx); err != nil { + return err + } + } + if !d.readFile.isNil() || !d.writeFile.isNil() { + d.fs.savedDentryRW[d] = savedDentryRW{ + read: !d.readFile.isNil(), + write: !d.writeFile.isNil(), + } + } + d.dirMu.Lock() + defer d.dirMu.Unlock() + for _, child := range d.children { + if child != nil { + if err := child.prepareSaveRecursive(ctx); err != nil { + return err + } + } + } + return nil +} + +// beforeSave is invoked by stateify. +func (d *dentry) beforeSave() { + if d.vfsd.IsDead() { + panic(fmt.Sprintf("gofer.dentry(%q).beforeSave: deleted and invalidated dentries can't be restored", genericDebugPathname(d))) + } +} + +// afterLoad is invoked by stateify. +func (d *dentry) afterLoad() { + d.hostFD = -1 + if refsvfs2.LeakCheckEnabled() && atomic.LoadInt64(&d.refs) != -1 { + refsvfs2.Register(d, "gofer.dentry") + } +} + +// afterLoad is invoked by stateify. +func (d *dentryPlatformFile) afterLoad() { + if d.hostFileMapper.IsInited() { + // Ensure that we don't call d.hostFileMapper.Init() again. + d.hostFileMapperInitOnce.Do(func() {}) + } +} + +// afterLoad is invoked by stateify. +func (fd *specialFileFD) afterLoad() { + fd.handle.fd = -1 +} + +// CompleteRestore implements +// vfs.FilesystemImplSaveRestoreExtension.CompleteRestore. +func (fs *filesystem) CompleteRestore(ctx context.Context, opts vfs.CompleteRestoreOptions) error { + fdmapv := ctx.Value(CtxRestoreServerFDMap) + if fdmapv == nil { + return fmt.Errorf("no server FD map available") + } + fdmap := fdmapv.(map[string]int) + fd, ok := fdmap[fs.iopts.UniqueID] + if !ok { + return fmt.Errorf("no server FD available for filesystem with unique ID %q", fs.iopts.UniqueID) + } + fs.opts.fd = fd + if err := fs.dial(ctx); err != nil { + return err + } + fs.inoByQIDPath = make(map[uint64]uint64) + + // Restore the filesystem root. + ctx.UninterruptibleSleepStart(false) + attached, err := fs.client.Attach(fs.opts.aname) + ctx.UninterruptibleSleepFinish(false) + if err != nil { + return err + } + attachFile := p9file{attached} + qid, attrMask, attr, err := attachFile.getAttr(ctx, dentryAttrMask()) + if err != nil { + return err + } + if err := fs.root.restoreFile(ctx, attachFile, qid, attrMask, &attr, &opts); err != nil { + return err + } + + // Restore remaining dentries. + if err := fs.root.restoreDescendantsRecursive(ctx, &opts); err != nil { + return err + } + + // Re-open handles for specialFileFDs. Unlike the initial open + // (dentry.openSpecialFile()), pipes are always opened without blocking; + // non-readable pipe FDs are opened last to ensure that they don't get + // ENXIO if another specialFileFD represents the read end of the same pipe. + // This is consistent with VFS1. + haveWriteOnlyPipes := false + for fd := range fs.specialFileFDs { + if fd.dentry().fileType() == linux.S_IFIFO && !fd.vfsfd.IsReadable() { + haveWriteOnlyPipes = true + continue + } + if err := fd.completeRestore(ctx); err != nil { + return err + } + } + if haveWriteOnlyPipes { + for fd := range fs.specialFileFDs { + if fd.dentry().fileType() == linux.S_IFIFO && !fd.vfsfd.IsReadable() { + if err := fd.completeRestore(ctx); err != nil { + return err + } + } + } + } + + // Discard state only required during restore. + fs.savedDentryRW = nil + + return nil +} + +func (d *dentry) restoreFile(ctx context.Context, file p9file, qid p9.QID, attrMask p9.AttrMask, attr *p9.Attr, opts *vfs.CompleteRestoreOptions) error { + d.file = file + + // Gofers do not preserve QID across checkpoint/restore, so: + // + // - We must assume that the remote filesystem did not change in a way that + // would invalidate dentries, since we can't revalidate dentries by + // checking QIDs. + // + // - We need to associate the new QID.Path with the existing d.ino. + d.qidPath = qid.Path + d.fs.inoMu.Lock() + d.fs.inoByQIDPath[qid.Path] = d.ino + d.fs.inoMu.Unlock() + + // Check metadata stability before updating metadata. + d.metadataMu.Lock() + defer d.metadataMu.Unlock() + if d.isRegularFile() { + if opts.ValidateFileSizes { + if !attrMask.Size { + return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: file size not available", genericDebugPathname(d)) + } + if d.size != attr.Size { + return fmt.Errorf("gofer.dentry(%q).restoreFile: file size validation failed: size changed from %d to %d", genericDebugPathname(d), d.size, attr.Size) + } + } + if opts.ValidateFileModificationTimestamps { + if !attrMask.MTime { + return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime not available", genericDebugPathname(d)) + } + if want := dentryTimestampFromP9(attr.MTimeSeconds, attr.MTimeNanoSeconds); d.mtime != want { + return fmt.Errorf("gofer.dentry(%q).restoreFile: mtime validation failed: mtime changed from %+v to %+v", genericDebugPathname(d), linux.NsecToStatxTimestamp(d.mtime), linux.NsecToStatxTimestamp(want)) + } + } + } + if !d.cachedMetadataAuthoritative() { + d.updateFromP9AttrsLocked(attrMask, attr) + } + + if rw, ok := d.fs.savedDentryRW[d]; ok { + if err := d.ensureSharedHandle(ctx, rw.read, rw.write, false /* trunc */); err != nil { + return err + } + } + + return nil +} + +// Preconditions: d is not synthetic. +func (d *dentry) restoreDescendantsRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error { + for _, child := range d.children { + if child == nil { + continue + } + if _, ok := d.fs.syncableDentries[child]; !ok { + // child is synthetic. + continue + } + if err := child.restoreRecursive(ctx, opts); err != nil { + return err + } + } + return nil +} + +// Preconditions: d is not synthetic (but note that since this function +// restores d.file, d.file.isNil() is always true at this point, so this can +// only be detected by checking filesystem.syncableDentries). d.parent has been +// restored. +func (d *dentry) restoreRecursive(ctx context.Context, opts *vfs.CompleteRestoreOptions) error { + qid, file, attrMask, attr, err := d.parent.file.walkGetAttrOne(ctx, d.name) + if err != nil { + return err + } + if err := d.restoreFile(ctx, file, qid, attrMask, &attr, opts); err != nil { + return err + } + return d.restoreDescendantsRecursive(ctx, opts) +} + +func (fd *specialFileFD) completeRestore(ctx context.Context) error { + d := fd.dentry() + h, err := openHandle(ctx, d.file, fd.vfsfd.IsReadable(), fd.vfsfd.IsWritable(), false /* trunc */) + if err != nil { + return err + } + fd.handle = h + + ftype := d.fileType() + fd.haveQueue = (ftype == linux.S_IFIFO || ftype == linux.S_IFSOCK) && fd.handle.fd >= 0 + if fd.haveQueue { + if err := fdnotifier.AddFD(fd.handle.fd, &fd.queue); err != nil { + return err + } + } + + return nil +} diff --git a/pkg/sentry/fsimpl/gofer/socket.go b/pkg/sentry/fsimpl/gofer/socket.go index 85d2bee72..a21199eac 100644 --- a/pkg/sentry/fsimpl/gofer/socket.go +++ b/pkg/sentry/fsimpl/gofer/socket.go @@ -36,13 +36,12 @@ func (d *dentry) isSocket() bool { // An endpoint's lifetime is the time between when filesystem.BoundEndpointAt() // is called and either BoundEndpoint.BidirectionalConnect or // BoundEndpoint.UnidirectionalConnect is called. +// +// +stateify savable type endpoint struct { // dentry is the filesystem dentry which produced this endpoint. dentry *dentry - // 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 } @@ -114,7 +113,7 @@ func (e *endpoint) UnidirectionalConnect(ctx context.Context) (transport.Connect } func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFlags, queue *waiter.Queue) (*host.SCMConnectedEndpoint, *syserr.Error) { - hostFile, err := e.file.Connect(flags) + hostFile, err := e.dentry.file.connect(ctx, flags) if err != nil { return nil, syserr.ErrConnectionRefused } @@ -129,7 +128,7 @@ func (e *endpoint) newConnectedEndpoint(ctx context.Context, flags p9.ConnectFla c, serr := host.NewSCMEndpoint(ctx, hostFD, queue, e.path) if serr != nil { - log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.file, flags, serr) + log.Warningf("Gofer returned invalid host socket for BidirectionalConnect; file %+v flags %+v: %v", e.dentry.file, flags, serr) return nil, serr } return c, nil diff --git a/pkg/sentry/fsimpl/gofer/special_file.go b/pkg/sentry/fsimpl/gofer/special_file.go index 3c39aa9b7..625400c0b 100644 --- a/pkg/sentry/fsimpl/gofer/special_file.go +++ b/pkg/sentry/fsimpl/gofer/special_file.go @@ -15,15 +15,16 @@ package gofer import ( - "sync" "sync/atomic" "syscall" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/p9" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" @@ -33,11 +34,13 @@ import ( // special files, and (when filesystemOptions.regularFilesUseSpecialFileFD is // in effect) regular files. specialFileFD differs from regularFileFD by using // per-FD handles instead of shared per-dentry handles, and never buffering I/O. +// +// +stateify savable type specialFileFD struct { fileDescription // handle is used for file I/O. handle is immutable. - handle handle + handle handle `state:"nosave"` // isRegularFile is true if this FD represents a regular file which is only // possible when filesystemOptions.regularFilesUseSpecialFileFD is in @@ -51,12 +54,20 @@ type specialFileFD struct { // haveQueue is true if this file description represents a file for which // queue may send I/O readiness events. haveQueue is immutable. - haveQueue bool + haveQueue bool `state:"nosave"` queue waiter.Queue // If seekable is true, off is the file offset. off is protected by mu. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 + + // If haveBuf is non-zero, this FD represents a pipe, and buf contains data + // read from the pipe from previous calls to specialFileFD.savePipeData(). + // haveBuf and buf are protected by bufMu. haveBuf is accessed using atomic + // memory operations. + bufMu sync.Mutex `state:"nosave"` + haveBuf uint32 + buf []byte } func newSpecialFileFD(h handle, mnt *vfs.Mount, d *dentry, locks *vfs.FileLocks, flags uint32) (*specialFileFD, error) { @@ -84,6 +95,9 @@ func newSpecialFileFD(h handle, mnt *vfs.Mount, d *dentry, locks *vfs.FileLocks, } return nil, err } + d.fs.syncMu.Lock() + d.fs.specialFileFDs[fd] = struct{}{} + d.fs.syncMu.Unlock() return fd, nil } @@ -135,6 +149,16 @@ func (fd *specialFileFD) EventUnregister(e *waiter.Entry) { fd.fileDescription.EventUnregister(e) } +func (fd *specialFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + if fd.isRegularFile { + d := fd.dentry() + return d.doAllocate(ctx, offset, length, func() error { + return fd.handle.file.allocate(ctx, p9.ToAllocateMode(mode), offset, length) + }) + } + return fd.FileDescriptionDefaultImpl.Allocate(ctx, mode, offset, length) +} + // PRead implements vfs.FileDescriptionImpl.PRead. func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { if fd.seekable && offset < 0 { @@ -148,26 +172,51 @@ func (fd *specialFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offs return 0, syserror.EOPNOTSUPP } - // Going through dst.CopyOutFrom() holds MM locks around file operations of - // unknown duration. For regularFileFD, doing so is necessary to support - // mmap due to lock ordering; MM locks precede dentry.dataMu. That doesn't - // hold here since specialFileFD doesn't client-cache data. Just buffer the - // read instead. if d := fd.dentry(); d.cachedMetadataAuthoritative() { d.touchAtime(fd.vfsfd.Mount()) } + + bufN := int64(0) + if atomic.LoadUint32(&fd.haveBuf) != 0 { + var err error + fd.bufMu.Lock() + if len(fd.buf) != 0 { + var n int + n, err = dst.CopyOut(ctx, fd.buf) + dst = dst.DropFirst(n) + fd.buf = fd.buf[n:] + if len(fd.buf) == 0 { + atomic.StoreUint32(&fd.haveBuf, 0) + fd.buf = nil + } + bufN = int64(n) + if offset >= 0 { + offset += bufN + } + } + fd.bufMu.Unlock() + if err != nil { + return bufN, err + } + } + + // Going through dst.CopyOutFrom() would hold MM locks around file + // operations of unknown duration. For regularFileFD, doing so is necessary + // to support mmap due to lock ordering; MM locks precede dentry.dataMu. + // That doesn't hold here since specialFileFD doesn't client-cache data. + // Just buffer the read instead. buf := make([]byte, dst.NumBytes()) n, err := fd.handle.readToBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) if err == syserror.EAGAIN { err = syserror.ErrWouldBlock } if n == 0 { - return 0, err + return bufN, err } if cp, cperr := dst.CopyOut(ctx, buf[:n]); cperr != nil { - return int64(cp), cperr + return bufN + int64(cp), cperr } - return int64(n), err + return bufN + int64(n), err } // Read implements vfs.FileDescriptionImpl.Read. @@ -204,16 +253,16 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off } d := fd.dentry() - // If the regular file fd was opened with O_APPEND, make sure the file size - // is updated. There is a possible race here if size is modified externally - // after metadata cache is updated. - if fd.isRegularFile && fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 && !d.cachedMetadataAuthoritative() { - if err := d.updateFromGetattr(ctx); err != nil { - return 0, offset, err + if fd.isRegularFile { + // If the regular file fd was opened with O_APPEND, make sure the file + // size is updated. There is a possible race here if size is modified + // externally after metadata cache is updated. + if fd.vfsfd.StatusFlags()&linux.O_APPEND != 0 && !d.cachedMetadataAuthoritative() { + if err := d.updateFromGetattr(ctx); err != nil { + return 0, offset, err + } } - } - if fd.isRegularFile { // We need to hold the metadataMu *while* writing to a regular file. d.metadataMu.Lock() defer d.metadataMu.Unlock() @@ -235,11 +284,12 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off d.touchCMtime() } buf := make([]byte, src.NumBytes()) - // Don't do partial writes if we get a partial read from src. - if _, err := src.CopyIn(ctx, buf); err != nil { - return 0, offset, err + copied, copyErr := src.CopyIn(ctx, buf) + if copied == 0 && copyErr != nil { + // Only return the error if we didn't get any data. + return 0, offset, copyErr } - n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf)), uint64(offset)) + n, err := fd.handle.writeFromBlocksAt(ctx, safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:copied])), uint64(offset)) if err == syserror.EAGAIN { err = syserror.ErrWouldBlock } @@ -256,7 +306,10 @@ func (fd *specialFileFD) pwrite(ctx context.Context, src usermem.IOSequence, off atomic.StoreUint64(&d.size, uint64(offset)) } } - return int64(n), offset, err + if err != nil { + return int64(n), offset, err + } + return int64(n), offset, copyErr } // Write implements vfs.FileDescriptionImpl.Write. @@ -289,13 +342,31 @@ func (fd *specialFileFD) Seek(ctx context.Context, offset int64, whence int32) ( // Sync implements vfs.FileDescriptionImpl.Sync. func (fd *specialFileFD) Sync(ctx context.Context) error { - // If we have a host FD, fsyncing it is likely to be faster than an fsync - // RPC. - if fd.handle.fd >= 0 { - ctx.UninterruptibleSleepStart(false) - err := syscall.Fsync(int(fd.handle.fd)) - ctx.UninterruptibleSleepFinish(false) - return err + return fd.sync(ctx, false /* forFilesystemSync */) +} + +func (fd *specialFileFD) sync(ctx context.Context, forFilesystemSync bool) error { + err := func() error { + // If we have a host FD, fsyncing it is likely to be faster than an fsync + // RPC. + if fd.handle.fd >= 0 { + ctx.UninterruptibleSleepStart(false) + err := syscall.Fsync(int(fd.handle.fd)) + ctx.UninterruptibleSleepFinish(false) + return err + } + return fd.handle.file.fsync(ctx) + }() + if err != nil { + if !forFilesystemSync { + return err + } + // Only return err if we can reasonably have expected sync to succeed + // (fd represents a regular file that was opened for writing). + if fd.isRegularFile && fd.vfsfd.IsWritable() { + return err + } + ctx.Debugf("gofer.specialFileFD.sync: syncing non-writable or non-regular-file FD failed: %v", err) } - return fd.handle.file.fsync(ctx) + return nil } diff --git a/pkg/sentry/fsimpl/gofer/time.go b/pkg/sentry/fsimpl/gofer/time.go index e59d07e90..9cbe805b9 100644 --- a/pkg/sentry/fsimpl/gofer/time.go +++ b/pkg/sentry/fsimpl/gofer/time.go @@ -17,7 +17,6 @@ package gofer import ( "sync/atomic" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -25,17 +24,6 @@ func dentryTimestampFromP9(s, ns uint64) int64 { return int64(s*1e9 + ns) } -func dentryTimestampFromStatx(ts linux.StatxTimestamp) int64 { - return ts.Sec*1e9 + int64(ts.Nsec) -} - -func statxTimestampFromDentry(ns int64) linux.StatxTimestamp { - return linux.StatxTimestamp{ - Sec: ns / 1e9, - Nsec: uint32(ns % 1e9), - } -} - // Preconditions: d.cachedMetadataAuthoritative() == true. func (d *dentry) touchAtime(mnt *vfs.Mount) { if mnt.Flags.NoATime || mnt.ReadOnly() { @@ -52,8 +40,23 @@ func (d *dentry) touchAtime(mnt *vfs.Mount) { mnt.EndWrite() } -// Preconditions: d.cachedMetadataAuthoritative() == true. The caller has -// successfully called vfs.Mount.CheckBeginWrite(). +// Preconditions: d.metadataMu is locked. d.cachedMetadataAuthoritative() == true. +func (d *dentry) touchAtimeLocked(mnt *vfs.Mount) { + if mnt.Flags.NoATime || mnt.ReadOnly() { + return + } + if err := mnt.CheckBeginWrite(); err != nil { + return + } + now := d.fs.clock.Now().Nanoseconds() + atomic.StoreInt64(&d.atime, now) + atomic.StoreUint32(&d.atimeDirty, 1) + mnt.EndWrite() +} + +// Preconditions: +// * d.cachedMetadataAuthoritative() == true. +// * The caller has successfully called vfs.Mount.CheckBeginWrite(). func (d *dentry) touchCtime() { now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() @@ -61,8 +64,9 @@ func (d *dentry) touchCtime() { d.metadataMu.Unlock() } -// Preconditions: d.cachedMetadataAuthoritative() == true. The caller has -// successfully called vfs.Mount.CheckBeginWrite(). +// Preconditions: +// * d.cachedMetadataAuthoritative() == true. +// * The caller has successfully called vfs.Mount.CheckBeginWrite(). func (d *dentry) touchCMtime() { now := d.fs.clock.Now().Nanoseconds() d.metadataMu.Lock() @@ -72,8 +76,9 @@ func (d *dentry) touchCMtime() { d.metadataMu.Unlock() } -// Preconditions: d.cachedMetadataAuthoritative() == true. The caller has -// locked d.metadataMu. +// Preconditions: +// * d.cachedMetadataAuthoritative() == true. +// * The caller has locked d.metadataMu. func (d *dentry) touchCMtimeLocked() { now := d.fs.clock.Now().Nanoseconds() atomic.StoreInt64(&d.mtime, now) diff --git a/pkg/sentry/fsimpl/host/BUILD b/pkg/sentry/fsimpl/host/BUILD index bd701bbc7..dc0f86061 100644 --- a/pkg/sentry/fsimpl/host/BUILD +++ b/pkg/sentry/fsimpl/host/BUILD @@ -1,14 +1,40 @@ load("//tools:defs.bzl", "go_library") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "inode_refs", + out = "inode_refs.go", + package = "host", + prefix = "inode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "inode", + }, +) + +go_template_instance( + name = "connected_endpoint_refs", + out = "connected_endpoint_refs.go", + package = "host", + prefix = "ConnectedEndpoint", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "ConnectedEndpoint", + }, +) + go_library( name = "host", srcs = [ + "connected_endpoint_refs.go", "control.go", "host.go", + "inode_refs.go", "ioctl_unsafe.go", "mmap.go", + "save_restore.go", "socket.go", "socket_iovec.go", "socket_unsafe.go", @@ -24,7 +50,9 @@ go_library( "//pkg/fspath", "//pkg/iovec", "//pkg/log", + "//pkg/marshal/primitive", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs/fsutil", diff --git a/pkg/sentry/fsimpl/host/control.go b/pkg/sentry/fsimpl/host/control.go index 0135e4428..13ef48cb5 100644 --- a/pkg/sentry/fsimpl/host/control.go +++ b/pkg/sentry/fsimpl/host/control.go @@ -79,7 +79,7 @@ func fdsToFiles(ctx context.Context, fds []int) []*vfs.FileDescription { } // Create the file backed by hostFD. - file, err := ImportFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, false /* isTTY */) + file, err := NewFD(ctx, kernel.KernelFromContext(ctx).HostMount(), fd, &NewFDOptions{}) if err != nil { ctx.Warningf("Error creating file from host FD: %v", err) break diff --git a/pkg/sentry/fsimpl/host/host.go b/pkg/sentry/fsimpl/host/host.go index 56869f59a..eeed0f97d 100644 --- a/pkg/sentry/fsimpl/host/host.go +++ b/pkg/sentry/fsimpl/host/host.go @@ -19,6 +19,7 @@ package host import ( "fmt" "math" + "sync/atomic" "syscall" "golang.org/x/sys/unix" @@ -27,7 +28,6 @@ import ( "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/fspath" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/refs" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/hostfd" @@ -41,8 +41,123 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// inode implements kernfs.Inode. +// +// +stateify savable +type inode struct { + kernfs.InodeNoStatFS + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid. + + locks vfs.FileLocks + + // When the reference count reaches zero, the host fd is closed. + inodeRefs + + // hostFD contains the host fd that this file was originally created from, + // which must be available at time of restore. + // + // This field is initialized at creation time and is immutable. + hostFD int + + // ino is an inode number unique within this filesystem. + // + // This field is initialized at creation time and is immutable. + ino uint64 + + // ftype is the file's type (a linux.S_IFMT mask). + // + // This field is initialized at creation time and is immutable. + ftype uint16 + + // mayBlock is true if hostFD is non-blocking, and operations on it may + // return EAGAIN or EWOULDBLOCK instead of blocking. + // + // This field is initialized at creation time and is immutable. + mayBlock bool + + // seekable is false if lseek(hostFD) returns ESPIPE. We assume that file + // offsets are meaningful iff seekable is true. + // + // This field is initialized at creation time and is immutable. + seekable bool + + // isTTY is true if this file represents a TTY. + // + // This field is initialized at creation time and is immutable. + isTTY bool + + // savable is true if hostFD may be saved/restored by its numeric value. + // + // This field is initialized at creation time and is immutable. + savable bool + + // Event queue for blocking operations. + queue waiter.Queue + + // mapsMu protects mappings. + mapsMu sync.Mutex `state:"nosave"` + + // If this file is mmappable, mappings tracks mappings of hostFD into + // memmap.MappingSpaces. + mappings memmap.MappingSet + + // pf implements platform.File for mappings of hostFD. + pf inodePlatformFile + + // If haveBuf is non-zero, hostFD represents a pipe, and buf contains data + // read from the pipe from previous calls to inode.beforeSave(). haveBuf + // and buf are protected by bufMu. haveBuf is accessed using atomic memory + // operations. + bufMu sync.Mutex `state:"nosave"` + haveBuf uint32 + buf []byte +} + +func newInode(ctx context.Context, fs *filesystem, hostFD int, savable bool, fileType linux.FileMode, isTTY bool) (*inode, error) { + // Determine if hostFD is seekable. + _, err := unix.Seek(hostFD, 0, linux.SEEK_CUR) + seekable := err != syserror.ESPIPE + // We expect regular files to be seekable, as this is required for them to + // be memory-mappable. + if !seekable && fileType == syscall.S_IFREG { + ctx.Infof("host.newInode: host FD %d is a non-seekable regular file", hostFD) + return nil, syserror.ESPIPE + } + + i := &inode{ + hostFD: hostFD, + ino: fs.NextIno(), + ftype: uint16(fileType), + mayBlock: fileType != syscall.S_IFREG && fileType != syscall.S_IFDIR, + seekable: seekable, + isTTY: isTTY, + savable: savable, + } + i.pf.inode = i + i.EnableLeakCheck() + + // If the hostFD can return EWOULDBLOCK when set to non-blocking, do so and + // handle blocking behavior in the sentry. + if i.mayBlock { + if err := syscall.SetNonblock(i.hostFD, true); err != nil { + return nil, err + } + if err := fdnotifier.AddFD(int32(i.hostFD), &i.queue); err != nil { + return nil, err + } + } + return i, nil +} + // NewFDOptions contains options to NewFD. type NewFDOptions struct { + // If Savable is true, the host file descriptor may be saved/restored by + // numeric value; the sandbox API requires a corresponding host FD with the + // same numeric value to be provieded at time of restore. + Savable bool + // If IsTTY is true, the file descriptor is a TTY. IsTTY bool @@ -76,45 +191,12 @@ func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions) flags = uint32(flagsInt) } - fileMode := linux.FileMode(s.Mode) - fileType := fileMode.FileType() - - // Determine if hostFD is seekable. If not, this syscall will return ESPIPE - // (see fs/read_write.c:llseek), e.g. for pipes, sockets, and some character - // devices. - _, err := unix.Seek(hostFD, 0, linux.SEEK_CUR) - seekable := err != syserror.ESPIPE - - i := &inode{ - hostFD: hostFD, - ino: fs.NextIno(), - isTTY: opts.IsTTY, - wouldBlock: wouldBlock(uint32(fileType)), - seekable: seekable, - // NOTE(b/38213152): Technically, some obscure char devices can be memory - // mapped, but we only allow regular files. - canMap: fileType == linux.S_IFREG, - } - i.pf.inode = i - - // Non-seekable files can't be memory mapped, assert this. - if !i.seekable && i.canMap { - panic("files that can return EWOULDBLOCK (sockets, pipes, etc.) cannot be memory mapped") - } - - // If the hostFD would block, we must set it to non-blocking and handle - // blocking behavior in the sentry. - if i.wouldBlock { - if err := syscall.SetNonblock(i.hostFD, true); err != nil { - return nil, err - } - if err := fdnotifier.AddFD(int32(i.hostFD), &i.queue); err != nil { - return nil, err - } - } - d := &kernfs.Dentry{} - d.Init(i) + i, err := newInode(ctx, fs, hostFD, opts.Savable, linux.FileMode(s.Mode).FileType(), opts.IsTTY) + if err != nil { + return nil, err + } + d.Init(&fs.Filesystem, i) // i.open will take a reference on d. defer d.DecRef(ctx) @@ -122,29 +204,35 @@ func NewFD(ctx context.Context, mnt *vfs.Mount, hostFD int, opts *NewFDOptions) // For simplicity, fileDescription.offset is set to 0. Technically, we // should only set to 0 on files that are not seekable (sockets, pipes, // etc.), and use the offset from the host fd otherwise when importing. - return i.open(ctx, d.VFSDentry(), mnt, flags) + return i.open(ctx, d, mnt, flags) } // ImportFD sets up and returns a vfs.FileDescription from a donated fd. func ImportFD(ctx context.Context, mnt *vfs.Mount, hostFD int, isTTY bool) (*vfs.FileDescription, error) { return NewFD(ctx, mnt, hostFD, &NewFDOptions{ - IsTTY: isTTY, + Savable: true, + IsTTY: isTTY, }) } // filesystemType implements vfs.FilesystemType. +// +// +stateify savable type filesystemType struct{} -// GetFilesystem implements FilesystemType.GetFilesystem. +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (filesystemType) GetFilesystem(context.Context, *vfs.VirtualFilesystem, *auth.Credentials, string, vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { panic("host.filesystemType.GetFilesystem should never be called") } -// Name implements FilesystemType.Name. +// Name implements vfs.FilesystemType.Name. func (filesystemType) Name() string { return "none" } +// Release implements vfs.FilesystemType.Release. +func (filesystemType) Release(ctx context.Context) {} + // NewFilesystem sets up and returns a new hostfs filesystem. // // Note that there should only ever be one instance of host.filesystem, @@ -162,6 +250,8 @@ func NewFilesystem(vfsObj *vfs.VirtualFilesystem) (*vfs.Filesystem, error) { } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -180,65 +270,7 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe return vfs.PrependPathSyntheticError{} } -// inode implements kernfs.Inode. -type inode struct { - kernfs.InodeNotDirectory - kernfs.InodeNotSymlink - - locks vfs.FileLocks - - // When the reference count reaches zero, the host fd is closed. - refs.AtomicRefCount - - // hostFD contains the host fd that this file was originally created from, - // which must be available at time of restore. - // - // This field is initialized at creation time and is immutable. - hostFD int - - // ino is an inode number unique within this filesystem. - // - // This field is initialized at creation time and is immutable. - ino uint64 - - // isTTY is true if this file represents a TTY. - // - // This field is initialized at creation time and is immutable. - isTTY bool - - // seekable is false if the host fd points to a file representing a stream, - // e.g. a socket or a pipe. Such files are not seekable and can return - // EWOULDBLOCK for I/O operations. - // - // This field is initialized at creation time and is immutable. - seekable bool - - // wouldBlock is true if the host FD would return EWOULDBLOCK for - // operations that would block. - // - // This field is initialized at creation time and is immutable. - wouldBlock bool - - // Event queue for blocking operations. - queue waiter.Queue - - // canMap specifies whether we allow the file to be memory mapped. - // - // This field is initialized at creation time and is immutable. - canMap bool - - // mapsMu protects mappings. - mapsMu sync.Mutex - - // If canMap is true, mappings tracks mappings of hostFD into - // memmap.MappingSpaces. - mappings memmap.MappingSet - - // pf implements platform.File for mappings of hostFD. - pf inodePlatformFile -} - -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { var s syscall.Stat_t if err := syscall.Fstat(i.hostFD, &s); err != nil { @@ -247,7 +279,7 @@ func (i *inode) CheckPermissions(ctx context.Context, creds *auth.Credentials, a return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(s.Mode), auth.KUID(s.Uid), auth.KGID(s.Gid)) } -// Mode implements kernfs.Inode. +// Mode implements kernfs.Inode.Mode. func (i *inode) Mode() linux.FileMode { var s syscall.Stat_t if err := syscall.Fstat(i.hostFD, &s); err != nil { @@ -258,7 +290,7 @@ func (i *inode) Mode() linux.FileMode { return linux.FileMode(s.Mode) } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *inode) Stat(ctx context.Context, vfsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { if opts.Mask&linux.STATX__RESERVED != 0 { return linux.Statx{}, syserror.EINVAL @@ -371,7 +403,7 @@ func (i *inode) fstat(fs *filesystem) (linux.Statx, error) { }, nil } -// SetStat implements kernfs.Inode. +// SetStat implements kernfs.Inode.SetStat. func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { s := &opts.Stat @@ -430,10 +462,10 @@ func (i *inode) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Cre return nil } -// DecRef implements kernfs.Inode. +// DecRef implements kernfs.Inode.DecRef. func (i *inode) DecRef(ctx context.Context) { - i.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { - if i.wouldBlock { + i.inodeRefs.DecRef(func() { + if i.mayBlock { fdnotifier.RemoveFD(int32(i.hostFD)) } if err := unix.Close(i.hostFD); err != nil { @@ -442,16 +474,16 @@ func (i *inode) DecRef(ctx context.Context) { }) } -// Open implements kernfs.Inode. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +// Open implements kernfs.Inode.Open. +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { // Once created, we cannot re-open a socket fd through /proc/[pid]/fd/. if i.Mode().FileType() == linux.S_IFSOCK { return nil, syserror.ENXIO } - return i.open(ctx, vfsd, rp.Mount(), opts.Flags) + return i.open(ctx, d, rp.Mount(), opts.Flags) } -func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, error) { +func (i *inode) open(ctx context.Context, d *kernfs.Dentry, mnt *vfs.Mount, flags uint32) (*vfs.FileDescription, error) { var s syscall.Stat_t if err := syscall.Fstat(i.hostFD, &s); err != nil { return nil, err @@ -475,17 +507,17 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u return nil, err } // Currently, we only allow Unix sockets to be imported. - return unixsocket.NewFileDescription(ep, ep.Type(), flags, mnt, d, &i.locks) + return unixsocket.NewFileDescription(ep, ep.Type(), flags, mnt, d.VFSDentry(), &i.locks) case syscall.S_IFREG, syscall.S_IFIFO, syscall.S_IFCHR: if i.isTTY { fd := &TTYFileDescription{ fileDescription: fileDescription{inode: i}, - termios: linux.DefaultSlaveTermios, + termios: linux.DefaultReplicaTermios, } fd.LockFD.Init(&i.locks) vfsfd := &fd.vfsfd - if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := vfsfd.Init(fd, flags, mnt, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return vfsfd, nil @@ -494,7 +526,7 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u fd := &fileDescription{inode: i} fd.LockFD.Init(&i.locks) vfsfd := &fd.vfsfd - if err := vfsfd.Init(fd, flags, mnt, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := vfsfd.Init(fd, flags, mnt, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return vfsfd, nil @@ -506,6 +538,8 @@ func (i *inode) open(ctx context.Context, d *vfs.Dentry, mnt *vfs.Mount, flags u } // fileDescription is embedded by host fd implementations of FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -520,41 +554,43 @@ type fileDescription struct { inode *inode // offsetMu protects offset. - offsetMu sync.Mutex + offsetMu sync.Mutex `state:"nosave"` // offset specifies the current file offset. It is only meaningful when // inode.seekable is true. offset int64 } -// SetStat implements vfs.FileDescriptionImpl. +// SetStat implements vfs.FileDescriptionImpl.SetStat. func (f *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { creds := auth.CredentialsFromContext(ctx) return f.inode.SetStat(ctx, f.vfsfd.Mount().Filesystem(), creds, opts) } -// Stat implements vfs.FileDescriptionImpl. +// Stat implements vfs.FileDescriptionImpl.Stat. func (f *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { return f.inode.Stat(ctx, f.vfsfd.Mount().Filesystem(), opts) } -// Release implements vfs.FileDescriptionImpl. +// Release implements vfs.FileDescriptionImpl.Release. func (f *fileDescription) Release(context.Context) { // noop } -// Allocate implements vfs.FileDescriptionImpl. +// Allocate implements vfs.FileDescriptionImpl.Allocate. func (f *fileDescription) Allocate(ctx context.Context, mode, offset, length uint64) error { - if !f.inode.seekable { - return syserror.ESPIPE - } - - // TODO(gvisor.dev/issue/3589): Implement Allocate for non-pipe hostfds. - return syserror.EOPNOTSUPP + return unix.Fallocate(f.inode.hostFD, uint32(mode), int64(offset), int64(length)) } -// PRead implements FileDescriptionImpl. +// PRead implements vfs.FileDescriptionImpl.PRead. func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, syserror.EOPNOTSUPP + } + i := f.inode if !i.seekable { return 0, syserror.ESPIPE @@ -563,21 +599,33 @@ func (f *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, off return readFromHostFD(ctx, i.hostFD, dst, offset, opts.Flags) } -// Read implements FileDescriptionImpl. +// Read implements vfs.FileDescriptionImpl.Read. func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // Check that flags are supported. + // + // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. + if opts.Flags&^linux.RWF_HIPRI != 0 { + return 0, syserror.EOPNOTSUPP + } + i := f.inode if !i.seekable { + bufN, err := i.readFromBuf(ctx, &dst) + if err != nil { + return bufN, err + } n, err := readFromHostFD(ctx, i.hostFD, dst, -1, opts.Flags) + total := bufN + n 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 { + if total != 0 { err = nil } else { err = syserror.ErrWouldBlock } } - return n, err + return total, err } f.offsetMu.Lock() @@ -587,20 +635,33 @@ func (f *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts return n, err } -func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) { - // Check that flags are supported. - // - // TODO(gvisor.dev/issue/2601): Support select preadv2 flags. - if flags&^linux.RWF_HIPRI != 0 { - return 0, syserror.EOPNOTSUPP +func (i *inode) readFromBuf(ctx context.Context, dst *usermem.IOSequence) (int64, error) { + if atomic.LoadUint32(&i.haveBuf) == 0 { + return 0, nil + } + i.bufMu.Lock() + defer i.bufMu.Unlock() + if len(i.buf) == 0 { + return 0, nil } + n, err := dst.CopyOut(ctx, i.buf) + *dst = dst.DropFirst(n) + i.buf = i.buf[n:] + if len(i.buf) == 0 { + atomic.StoreUint32(&i.haveBuf, 0) + i.buf = nil + } + return int64(n), err +} + +func readFromHostFD(ctx context.Context, hostFD int, dst usermem.IOSequence, offset int64, flags uint32) (int64, error) { reader := hostfd.GetReadWriterAt(int32(hostFD), offset, flags) n, err := dst.CopyOutFrom(ctx, reader) hostfd.PutReadWriterAt(reader) return int64(n), err } -// PWrite implements FileDescriptionImpl. +// PWrite implements vfs.FileDescriptionImpl.PWrite. func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { if !f.inode.seekable { return 0, syserror.ESPIPE @@ -609,7 +670,7 @@ func (f *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, of return f.writeToHostFD(ctx, src, offset, opts.Flags) } -// Write implements FileDescriptionImpl. +// Write implements vfs.FileDescriptionImpl.Write. func (f *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { i := f.inode if !i.seekable { @@ -657,7 +718,7 @@ func (f *fileDescription) writeToHostFD(ctx context.Context, src usermem.IOSeque return int64(n), err } -// Seek implements FileDescriptionImpl. +// Seek implements vfs.FileDescriptionImpl.Seek. // // Note that we do not support seeking on directories, since we do not even // allow directory fds to be imported at all. @@ -722,15 +783,17 @@ func (f *fileDescription) Seek(_ context.Context, offset int64, whence int32) (i return f.offset, nil } -// Sync implements FileDescriptionImpl. -func (f *fileDescription) Sync(context.Context) error { +// Sync implements vfs.FileDescriptionImpl.Sync. +func (f *fileDescription) Sync(ctx context.Context) error { // TODO(gvisor.dev/issue/1897): Currently, we always sync everything. return unix.Fsync(f.inode.hostFD) } -// ConfigureMMap implements FileDescriptionImpl. +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts) error { - if !f.inode.canMap { + // NOTE(b/38213152): Technically, some obscure char devices can be memory + // mapped, but we only allow regular files. + if f.inode.ftype != syscall.S_IFREG { return syserror.ENODEV } i := f.inode @@ -741,13 +804,17 @@ func (f *fileDescription) ConfigureMMap(_ context.Context, opts *memmap.MMapOpts // EventRegister implements waiter.Waitable.EventRegister. func (f *fileDescription) EventRegister(e *waiter.Entry, mask waiter.EventMask) { f.inode.queue.EventRegister(e, mask) - fdnotifier.UpdateFD(int32(f.inode.hostFD)) + if f.inode.mayBlock { + fdnotifier.UpdateFD(int32(f.inode.hostFD)) + } } // EventUnregister implements waiter.Waitable.EventUnregister. func (f *fileDescription) EventUnregister(e *waiter.Entry) { f.inode.queue.EventUnregister(e) - fdnotifier.UpdateFD(int32(f.inode.hostFD)) + if f.inode.mayBlock { + fdnotifier.UpdateFD(int32(f.inode.hostFD)) + } } // Readiness uses the poll() syscall to check the status of the underlying FD. diff --git a/pkg/sentry/fsimpl/host/mmap.go b/pkg/sentry/fsimpl/host/mmap.go index 65d3af38c..3d7eb2f96 100644 --- a/pkg/sentry/fsimpl/host/mmap.go +++ b/pkg/sentry/fsimpl/host/mmap.go @@ -27,11 +27,13 @@ import ( // cannot implement both kernfs.Inode.IncRef and memmap.File.IncRef. // // inodePlatformFile should only be used if inode.canMap is true. +// +// +stateify savable type inodePlatformFile struct { *inode // fdRefsMu protects fdRefs. - fdRefsMu sync.Mutex + fdRefsMu sync.Mutex `state:"nosave"` // fdRefs counts references on memmap.File offsets. It is used solely for // memory accounting. @@ -41,7 +43,7 @@ type inodePlatformFile struct { fileMapper fsutil.HostFileMapper // fileMapperInitOnce is used to lazily initialize fileMapper. - fileMapperInitOnce sync.Once + fileMapperInitOnce sync.Once `state:"nosave"` } // IncRef implements memmap.File.IncRef. diff --git a/pkg/sentry/fsimpl/host/save_restore.go b/pkg/sentry/fsimpl/host/save_restore.go new file mode 100644 index 000000000..7e32a8863 --- /dev/null +++ b/pkg/sentry/fsimpl/host/save_restore.go @@ -0,0 +1,78 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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" + "sync/atomic" + "syscall" + + "gvisor.dev/gvisor/pkg/fdnotifier" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/hostfd" + "gvisor.dev/gvisor/pkg/usermem" +) + +// beforeSave is invoked by stateify. +func (i *inode) beforeSave() { + if !i.savable { + panic("host.inode is not savable") + } + if i.ftype == syscall.S_IFIFO { + // If this pipe FD is readable, drain it so that bytes in the pipe can + // be read after restore. (This is a legacy VFS1 feature.) We don't + // know if the pipe FD is readable, so just try reading and tolerate + // EBADF from the read. + i.bufMu.Lock() + defer i.bufMu.Unlock() + var buf [usermem.PageSize]byte + for { + n, err := hostfd.Preadv2(int32(i.hostFD), safemem.BlockSeqOf(safemem.BlockFromSafeSlice(buf[:])), -1 /* offset */, 0 /* flags */) + if n != 0 { + i.buf = append(i.buf, buf[:n]...) + } + if err != nil { + if err == io.EOF || err == syscall.EAGAIN || err == syscall.EBADF { + break + } + panic(fmt.Errorf("host.inode.beforeSave: buffering from pipe failed: %v", err)) + } + } + if len(i.buf) != 0 { + atomic.StoreUint32(&i.haveBuf, 1) + } + } +} + +// afterLoad is invoked by stateify. +func (i *inode) afterLoad() { + if i.mayBlock { + if err := syscall.SetNonblock(i.hostFD, true); err != nil { + panic(fmt.Sprintf("host.inode.afterLoad: failed to set host FD %d non-blocking: %v", i.hostFD, err)) + } + if err := fdnotifier.AddFD(int32(i.hostFD), &i.queue); err != nil { + panic(fmt.Sprintf("host.inode.afterLoad: fdnotifier.AddFD(%d) failed: %v", i.hostFD, err)) + } + } +} + +// afterLoad is invoked by stateify. +func (i *inodePlatformFile) afterLoad() { + if i.fileMapper.IsInited() { + // Ensure that we don't call i.fileMapper.Init() again. + i.fileMapperInitOnce.Do(func() {}) + } +} diff --git a/pkg/sentry/fsimpl/host/socket.go b/pkg/sentry/fsimpl/host/socket.go index 4979dd0a9..8a447e29f 100644 --- a/pkg/sentry/fsimpl/host/socket.go +++ b/pkg/sentry/fsimpl/host/socket.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/socket/control" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/uniqueid" @@ -59,8 +58,7 @@ func newEndpoint(ctx context.Context, hostFD int, queue *waiter.Queue) (transpor // // +stateify savable type ConnectedEndpoint struct { - // ref keeps track of references to a ConnectedEndpoint. - ref refs.AtomicRefCount + ConnectedEndpointRefs // mu protects fd below. mu sync.RWMutex `state:"nosave"` @@ -132,9 +130,9 @@ func NewConnectedEndpoint(ctx context.Context, hostFD int, addr string, saveable return nil, err } - // AtomicRefCounters start off with a single reference. We need two. - e.ref.IncRef() - e.ref.EnableLeakCheck("host.ConnectedEndpoint") + // ConnectedEndpointRefs start off with a single reference. We need two. + e.IncRef() + e.EnableLeakCheck() return &e, nil } @@ -318,7 +316,7 @@ func (c *ConnectedEndpoint) destroyLocked() { // Release implements transport.ConnectedEndpoint.Release and // transport.Receiver.Release. func (c *ConnectedEndpoint) Release(ctx context.Context) { - c.ref.DecRefWithDestructor(ctx, func(context.Context) { + c.DecRef(func() { c.mu.Lock() c.destroyLocked() c.mu.Unlock() @@ -348,12 +346,12 @@ func (e *SCMConnectedEndpoint) Init() error { // Release implements transport.ConnectedEndpoint.Release and // transport.Receiver.Release. func (e *SCMConnectedEndpoint) Release(ctx context.Context) { - e.ref.DecRefWithDestructor(ctx, func(context.Context) { + e.DecRef(func() { e.mu.Lock() + fdnotifier.RemoveFD(int32(e.fd)) if err := syscall.Close(e.fd); err != nil { log.Warningf("Failed to close host fd %d: %v", err) } - fdnotifier.RemoveFD(int32(e.fd)) e.destroyLocked() e.mu.Unlock() }) @@ -378,8 +376,8 @@ func NewSCMEndpoint(ctx context.Context, hostFD int, queue *waiter.Queue, addr s return nil, err } - // AtomicRefCounters start off with a single reference. We need two. - e.ref.IncRef() - e.ref.EnableLeakCheck("host.SCMConnectedEndpoint") + // ConnectedEndpointRefs start off with a single reference. We need two. + e.IncRef() + e.EnableLeakCheck() return &e, nil } diff --git a/pkg/sentry/fsimpl/host/socket_unsafe.go b/pkg/sentry/fsimpl/host/socket_unsafe.go index 35ded24bc..c0bf45f08 100644 --- a/pkg/sentry/fsimpl/host/socket_unsafe.go +++ b/pkg/sentry/fsimpl/host/socket_unsafe.go @@ -63,10 +63,10 @@ func fdReadVec(fd int, bufs [][]byte, control []byte, peek bool, maxlen int64) ( controlTrunc = msg.Flags&syscall.MSG_CTRUNC == syscall.MSG_CTRUNC if n > length { - return length, n, msg.Controllen, controlTrunc, err + return length, n, msg.Controllen, controlTrunc, nil } - return n, n, msg.Controllen, controlTrunc, err + return n, n, msg.Controllen, controlTrunc, nil } // fdWriteVec sends from bufs to fd. diff --git a/pkg/sentry/fsimpl/host/tty.go b/pkg/sentry/fsimpl/host/tty.go index 27cbd3059..f5c596fec 100644 --- a/pkg/sentry/fsimpl/host/tty.go +++ b/pkg/sentry/fsimpl/host/tty.go @@ -17,6 +17,7 @@ package host import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -29,6 +30,8 @@ import ( // TTYFileDescription implements vfs.FileDescriptionImpl for a host file // descriptor that wraps a TTY FD. +// +// +stateify savable type TTYFileDescription struct { fileDescription @@ -75,7 +78,7 @@ func (t *TTYFileDescription) Release(ctx context.Context) { t.fileDescription.Release(ctx) } -// PRead implements vfs.FileDescriptionImpl. +// PRead implements vfs.FileDescriptionImpl.PRead. // // Reading from a TTY is only allowed for foreground process groups. Background // process groups will either get EIO or a SIGTTIN. @@ -93,7 +96,7 @@ func (t *TTYFileDescription) PRead(ctx context.Context, dst usermem.IOSequence, return t.fileDescription.PRead(ctx, dst, offset, opts) } -// Read implements vfs.FileDescriptionImpl. +// Read implements vfs.FileDescriptionImpl.Read. // // Reading from a TTY is only allowed for foreground process groups. Background // process groups will either get EIO or a SIGTTIN. @@ -111,7 +114,7 @@ func (t *TTYFileDescription) Read(ctx context.Context, dst usermem.IOSequence, o return t.fileDescription.Read(ctx, dst, opts) } -// PWrite implements vfs.FileDescriptionImpl. +// PWrite implements vfs.FileDescriptionImpl.PWrite. func (t *TTYFileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { t.mu.Lock() defer t.mu.Unlock() @@ -126,7 +129,7 @@ func (t *TTYFileDescription) PWrite(ctx context.Context, src usermem.IOSequence, return t.fileDescription.PWrite(ctx, src, offset, opts) } -// Write implements vfs.FileDescriptionImpl. +// Write implements vfs.FileDescriptionImpl.Write. func (t *TTYFileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { t.mu.Lock() defer t.mu.Unlock() @@ -141,8 +144,13 @@ func (t *TTYFileDescription) Write(ctx context.Context, src usermem.IOSequence, return t.fileDescription.Write(ctx, src, opts) } -// Ioctl implements vfs.FileDescriptionImpl. +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArguments) (uintptr, error) { + task := kernel.TaskFromContext(ctx) + if task == nil { + return 0, syserror.ENOTTY + } + // Ignore arg[0]. This is the real FD: fd := t.inode.hostFD ioctl := args[1].Uint64() @@ -152,9 +160,7 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch if err != nil { return 0, err } - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), termios, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err = termios.CopyOut(task, args[2].Pointer()) return 0, err case linux.TCSETS, linux.TCSETSW, linux.TCSETSF: @@ -166,9 +172,7 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch } var termios linux.Termios - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &termios, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := termios.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } err := ioctlSetTermios(fd, ioctl, &termios) @@ -192,10 +196,8 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch defer t.mu.Unlock() // Map the ProcessGroup into a ProcessGroupID in the task's PID namespace. - pgID := pidns.IDOfProcessGroup(t.fgProcessGroup) - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ - AddressSpaceActive: true, - }) + pgID := primitive.Int32(pidns.IDOfProcessGroup(t.fgProcessGroup)) + _, err := pgID.CopyOut(task, args[2].Pointer()) return 0, err case linux.TIOCSPGRP: @@ -203,11 +205,6 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch // Equivalent to tcsetpgrp(fd, *argp). // Set the foreground process group ID of this terminal. - task := kernel.TaskFromContext(ctx) - if task == nil { - return 0, syserror.ENOTTY - } - t.mu.Lock() defer t.mu.Unlock() @@ -226,12 +223,11 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch return 0, syserror.ENOTTY } - var pgID kernel.ProcessGroupID - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &pgID, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + var pgIDP primitive.Int32 + if _, err := pgIDP.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } + pgID := kernel.ProcessGroupID(pgIDP) // pgID must be non-negative. if pgID < 0 { @@ -260,9 +256,7 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch if err != nil { return 0, err } - _, err = usermem.CopyObjectOut(ctx, io, args[2].Pointer(), winsize, usermem.IOOpts{ - AddressSpaceActive: true, - }) + _, err = winsize.CopyOut(task, args[2].Pointer()) return 0, err case linux.TIOCSWINSZ: @@ -273,9 +267,7 @@ func (t *TTYFileDescription) Ioctl(ctx context.Context, io usermem.IO, args arch // set the winsize. var winsize linux.Winsize - if _, err := usermem.CopyObjectIn(ctx, io, args[2].Pointer(), &winsize, usermem.IOOpts{ - AddressSpaceActive: true, - }); err != nil { + if _, err := winsize.CopyIn(task, args[2].Pointer()); err != nil { return 0, err } err := ioctlSetWinsize(fd, &winsize) diff --git a/pkg/sentry/fsimpl/host/util.go b/pkg/sentry/fsimpl/host/util.go index 412bdb2eb..b2f43a119 100644 --- a/pkg/sentry/fsimpl/host/util.go +++ b/pkg/sentry/fsimpl/host/util.go @@ -43,12 +43,6 @@ func timespecToStatxTimestamp(ts unix.Timespec) linux.StatxTimestamp { return linux.StatxTimestamp{Sec: int64(ts.Sec), Nsec: uint32(ts.Nsec)} } -// wouldBlock returns true for file types that can return EWOULDBLOCK -// for blocking operations, e.g. pipes, character devices, and sockets. -func wouldBlock(fileType uint32) bool { - return fileType == syscall.S_IFIFO || fileType == syscall.S_IFCHR || fileType == syscall.S_IFSOCK -} - // isBlockError checks if an error is EAGAIN or EWOULDBLOCK. // If so, they can be transformed into syserror.ErrWouldBlock. func isBlockError(err error) bool { diff --git a/pkg/sentry/fsimpl/kernfs/BUILD b/pkg/sentry/fsimpl/kernfs/BUILD index 3835557fe..aaad67ab8 100644 --- a/pkg/sentry/fsimpl/kernfs/BUILD +++ b/pkg/sentry/fsimpl/kernfs/BUILD @@ -4,6 +4,18 @@ load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) go_template_instance( + name = "dentry_list", + out = "dentry_list.go", + package = "kernfs", + prefix = "dentry", + template = "//pkg/ilist:generic_list", + types = { + "Element": "*Dentry", + "Linker": "*Dentry", + }, +) + +go_template_instance( name = "fstree", out = "fstree.go", package = "kernfs", @@ -26,9 +38,54 @@ go_template_instance( }, ) +go_template_instance( + name = "static_directory_refs", + out = "static_directory_refs.go", + package = "kernfs", + prefix = "StaticDirectory", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "StaticDirectory", + }, +) + +go_template_instance( + name = "dir_refs", + out = "dir_refs.go", + package = "kernfs_test", + prefix = "dir", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "dir", + }, +) + +go_template_instance( + name = "readonly_dir_refs", + out = "readonly_dir_refs.go", + package = "kernfs_test", + prefix = "readonlyDir", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "readonlyDir", + }, +) + +go_template_instance( + name = "synthetic_directory_refs", + out = "synthetic_directory_refs.go", + package = "kernfs", + prefix = "syntheticDirectory", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "syntheticDirectory", + }, +) + go_library( name = "kernfs", srcs = [ + "dentry_list.go", "dynamic_bytes_file.go", "fd_impl_util.go", "filesystem.go", @@ -36,7 +93,10 @@ go_library( "inode_impl_util.go", "kernfs.go", "slot_list.go", + "static_directory_refs.go", "symlink.go", + "synthetic_directory.go", + "synthetic_directory_refs.go", ], visibility = ["//pkg/sentry:internal"], deps = [ @@ -45,8 +105,11 @@ go_library( "//pkg/fspath", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", + "//pkg/sentry/fs/fsutil", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", + "//pkg/sentry/kernel/time", "//pkg/sentry/memmap", "//pkg/sentry/socket/unix/transport", "//pkg/sentry/vfs", @@ -59,11 +122,18 @@ go_library( go_test( name = "kernfs_test", size = "small", - srcs = ["kernfs_test.go"], + srcs = [ + "dir_refs.go", + "kernfs_test.go", + "readonly_dir_refs.go", + ], deps = [ ":kernfs", "//pkg/abi/linux", "//pkg/context", + "//pkg/log", + "//pkg/refs", + "//pkg/refsvfs2", "//pkg/sentry/contexttest", "//pkg/sentry/fsimpl/testutil", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go index 12adf727a..485504995 100644 --- a/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go +++ b/pkg/sentry/fsimpl/kernfs/dynamic_bytes_file.go @@ -35,6 +35,7 @@ import ( // +stateify savable type DynamicBytesFile struct { InodeAttrs + InodeNoStatFS InodeNoopRefCount InodeNotDirectory InodeNotSymlink @@ -46,18 +47,18 @@ type DynamicBytesFile struct { var _ Inode = (*DynamicBytesFile)(nil) // Init initializes a dynamic bytes file. -func (f *DynamicBytesFile) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) { +func (f *DynamicBytesFile) Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) { if perm&^linux.PermissionsMask != 0 { panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) } - f.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeRegular|perm) + f.InodeAttrs.Init(ctx, creds, devMajor, devMinor, ino, linux.ModeRegular|perm) f.data = data } // Open implements Inode.Open. -func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (f *DynamicBytesFile) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &DynamicBytesFD{} - if err := fd.Init(rp.Mount(), vfsd, f.data, &f.locks, opts.Flags); err != nil { + if err := fd.Init(rp.Mount(), d, f.data, &f.locks, opts.Flags); err != nil { return nil, err } return &fd.vfsfd, nil @@ -86,12 +87,12 @@ type DynamicBytesFD struct { } // Init initializes a DynamicBytesFD. -func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *vfs.Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { +func (fd *DynamicBytesFD) Init(m *vfs.Mount, d *Dentry, data vfs.DynamicBytesSource, locks *vfs.FileLocks, flags uint32) error { fd.LockFD.Init(locks) - if err := fd.vfsfd.Init(fd, flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return err } - fd.inode = d.Impl().(*Dentry).inode + fd.inode = d.inode fd.SetDataSource(data) return nil } diff --git a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go index fcee6200a..f8dae22f8 100644 --- a/pkg/sentry/fsimpl/kernfs/fd_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/fd_impl_util.go @@ -15,7 +15,7 @@ package kernfs import ( - "math" + "fmt" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -28,9 +28,29 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// SeekEndConfig describes the SEEK_END behaviour for FDs. +// +// +stateify savable +type SeekEndConfig int + +// Constants related to SEEK_END behaviour for FDs. +const ( + // Consider the end of the file to be after the final static entry. This is + // the default option. + SeekEndStaticEntries = iota + // Consider the end of the file to be at offset 0. + SeekEndZero +) + +// GenericDirectoryFDOptions contains configuration for a GenericDirectoryFD. +// +// +stateify savable +type GenericDirectoryFDOptions struct { + SeekEnd SeekEndConfig +} + // GenericDirectoryFD implements vfs.FileDescriptionImpl for a generic directory -// inode that uses OrderChildren to track child nodes. GenericDirectoryFD is not -// compatible with dynamic directories. +// inode that uses OrderChildren to track child nodes. // // Note that GenericDirectoryFD holds a lock over OrderedChildren while calling // IterDirents callback. The IterDirents callback therefore cannot hash or @@ -40,16 +60,21 @@ import ( // Must be initialize with Init before first use. // // Lock ordering: mu => children.mu. +// +// +stateify savable type GenericDirectoryFD struct { vfs.FileDescriptionDefaultImpl vfs.DirectoryFileDescriptionDefaultImpl vfs.LockFD + // Immutable. + seekEnd SeekEndConfig + vfsfd vfs.FileDescription children *OrderedChildren // mu protects the fields below. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // off is the current directory offset. Protected by "mu". off int64 @@ -57,12 +82,12 @@ type GenericDirectoryFD struct { // NewGenericDirectoryFD creates a new GenericDirectoryFD and returns its // dentry. -func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions) (*GenericDirectoryFD, error) { +func NewGenericDirectoryFD(m *vfs.Mount, d *Dentry, children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) (*GenericDirectoryFD, error) { fd := &GenericDirectoryFD{} - if err := fd.Init(children, locks, opts); err != nil { + if err := fd.Init(children, locks, opts, fdOpts); err != nil { return nil, err } - if err := fd.vfsfd.Init(fd, opts.Flags, m, d, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, m, d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return fd, nil @@ -71,12 +96,13 @@ func NewGenericDirectoryFD(m *vfs.Mount, d *vfs.Dentry, children *OrderedChildre // Init initializes a GenericDirectoryFD. Use it when overriding // GenericDirectoryFD. Caller must call fd.VFSFileDescription.Init() with the // correct implementation. -func (fd *GenericDirectoryFD) Init(children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions) error { +func (fd *GenericDirectoryFD) Init(children *OrderedChildren, locks *vfs.FileLocks, opts *vfs.OpenOptions, fdOpts GenericDirectoryFDOptions) error { if vfs.AccessTypesForOpenFlags(opts)&vfs.MayWrite != 0 { // Can't open directories for writing. return syserror.EISDIR } fd.LockFD.Init(locks) + fd.seekEnd = fdOpts.SeekEnd fd.children = children return nil } @@ -119,8 +145,12 @@ func (fd *GenericDirectoryFD) filesystem() *vfs.Filesystem { return fd.vfsfd.VirtualDentry().Mount().Filesystem() } +func (fd *GenericDirectoryFD) dentry() *Dentry { + return fd.vfsfd.Dentry().Impl().(*Dentry) +} + func (fd *GenericDirectoryFD) inode() Inode { - return fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode + return fd.dentry().inode } // IterDirents implements vfs.FileDescriptionImpl.IterDirents. IterDirents holds @@ -150,8 +180,7 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // Handle "..". if fd.off == 1 { - vfsd := fd.vfsfd.VirtualDentry().Dentry() - parentInode := genericParentOrSelf(vfsd.Impl().(*Dentry)).inode + parentInode := genericParentOrSelf(fd.dentry()).inode stat, err := parentInode.Stat(ctx, fd.filesystem(), opts) if err != nil { return err @@ -175,13 +204,12 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent // these. childIdx := fd.off - 2 for it := fd.children.nthLocked(childIdx); it != nil; it = it.Next() { - inode := it.Dentry.Impl().(*Dentry).inode - stat, err := inode.Stat(ctx, fd.filesystem(), opts) + stat, err := it.inode.Stat(ctx, fd.filesystem(), opts) if err != nil { return err } dirent := vfs.Dirent{ - Name: it.Name, + Name: it.name, Type: linux.FileMode(stat.Mode).DirentType(), Ino: stat.Ino, NextOff: fd.off + 1, @@ -194,7 +222,7 @@ func (fd *GenericDirectoryFD) IterDirents(ctx context.Context, cb vfs.IterDirent var err error relOffset := fd.off - int64(len(fd.children.set)) - 2 - fd.off, err = fd.inode().IterDirents(ctx, cb, fd.off, relOffset) + fd.off, err = fd.inode().IterDirents(ctx, fd.vfsfd.Mount(), cb, fd.off, relOffset) return err } @@ -209,9 +237,17 @@ func (fd *GenericDirectoryFD) Seek(ctx context.Context, offset int64, whence int case linux.SEEK_CUR: offset += fd.off case linux.SEEK_END: - // TODO(gvisor.dev/issue/1193): This can prevent new files from showing up - // if they are added after SEEK_END. - offset = math.MaxInt64 + switch fd.seekEnd { + case SeekEndStaticEntries: + fd.children.mu.RLock() + offset += int64(len(fd.children.set)) + offset += 2 // '.' and '..' aren't tracked in children. + fd.children.mu.RUnlock() + case SeekEndZero: + // No-op: offset += 0. + default: + panic(fmt.Sprintf("Invalid GenericDirectoryFD.seekEnd = %v", fd.seekEnd)) + } default: return 0, syserror.EINVAL } @@ -232,8 +268,7 @@ func (fd *GenericDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (l // SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *GenericDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { creds := auth.CredentialsFromContext(ctx) - inode := fd.vfsfd.VirtualDentry().Dentry().Impl().(*Dentry).inode - return inode.SetStat(ctx, fd.filesystem(), creds, opts) + return fd.inode().SetStat(ctx, fd.filesystem(), creds, opts) } // Allocate implements vfs.FileDescriptionImpl.Allocate. diff --git a/pkg/sentry/fsimpl/kernfs/filesystem.go b/pkg/sentry/fsimpl/kernfs/filesystem.go index 3e5192edd..399895f3e 100644 --- a/pkg/sentry/fsimpl/kernfs/filesystem.go +++ b/pkg/sentry/fsimpl/kernfs/filesystem.go @@ -32,11 +32,12 @@ import ( // // stepExistingLocked is loosely analogous to fs/namei.c:walk_component(). // -// Preconditions: Filesystem.mu must be locked for at least reading. !rp.Done(). +// Preconditions: +// * Filesystem.mu must be locked for at least reading. +// * !rp.Done(). // // Postcondition: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, mayFollowSymlinks bool) (*vfs.Dentry, error) { - d := vfsd.Impl().(*Dentry) +func (fs *Filesystem) stepExistingLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, mayFollowSymlinks bool) (*Dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR } @@ -53,20 +54,20 @@ afterSymlink: // calls d_revalidate(), but walk_component() => handle_dots() does not. if name == "." { rp.Advance() - return vfsd, nil + return d, nil } if name == ".." { - if isRoot, err := rp.CheckRoot(ctx, vfsd); err != nil { + if isRoot, err := rp.CheckRoot(ctx, d.VFSDentry()); err != nil { return nil, err } else if isRoot || d.parent == nil { rp.Advance() - return vfsd, nil + return d, nil } - if err := rp.CheckMount(ctx, &d.parent.vfsd); err != nil { + if err := rp.CheckMount(ctx, d.parent.VFSDentry()); err != nil { return nil, err } rp.Advance() - return &d.parent.vfsd, nil + return d.parent, nil } if len(name) > linux.NAME_MAX { return nil, syserror.ENAMETOOLONG @@ -77,7 +78,7 @@ afterSymlink: if err != nil { return nil, err } - if err := rp.CheckMount(ctx, &next.vfsd); err != nil { + if err := rp.CheckMount(ctx, next.VFSDentry()); err != nil { return nil, err } // Resolve any symlink at current path component. @@ -88,7 +89,7 @@ afterSymlink: } if targetVD.Ok() { err := rp.HandleJump(targetVD) - targetVD.DecRef(ctx) + fs.deferDecRefVD(ctx, targetVD) if err != nil { return nil, err } @@ -100,15 +101,18 @@ afterSymlink: goto afterSymlink } rp.Advance() - return &next.vfsd, nil + return next, nil } // revalidateChildLocked must be called after a call to parent.vfsd.Child(name) // or vfs.ResolvingPath.ResolveChild(name) returns childVFSD (which may be // nil) to verify that the returned child (or lack thereof) is correct. // -// Preconditions: Filesystem.mu must be locked for at least reading. -// parent.dirMu must be locked. parent.isDir(). name is not "." or "..". +// Preconditions: +// * Filesystem.mu must be locked for at least reading. +// * parent.dirMu must be locked. +// * parent.isDir(). +// * name is not "." or "..". // // Postconditions: Caller must call fs.processDeferredDecRefs*. func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.VirtualFilesystem, parent *Dentry, name string, child *Dentry) (*Dentry, error) { @@ -116,26 +120,33 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // Cached dentry exists, revalidate. if !child.inode.Valid(ctx) { delete(parent.children, name) - vfsObj.InvalidateDentry(ctx, &child.vfsd) - fs.deferDecRef(&child.vfsd) // Reference from Lookup. + if child.inode.Keep() { + // Drop the ref owned by kernfs. + fs.deferDecRef(child) + } + vfsObj.InvalidateDentry(ctx, child.VFSDentry()) child = nil } } if child == nil { - // Dentry isn't cached; it either doesn't exist or failed - // revalidation. Attempt to resolve it via Lookup. - // - // FIXME(gvisor.dev/issue/1193): Inode.Lookup() should return - // *(kernfs.)Dentry, not *vfs.Dentry, since (kernfs.)Filesystem assumes - // that all dentries in the filesystem are (kernfs.)Dentry and performs - // vfs.DentryImpl casts accordingly. - childVFSD, err := parent.inode.Lookup(ctx, name) + // Dentry isn't cached; it either doesn't exist or failed revalidation. + // Attempt to resolve it via Lookup. + childInode, err := parent.inode.Lookup(ctx, name) if err != nil { return nil, err } - // Reference on childVFSD dropped by a corresponding Valid. - child = childVFSD.Impl().(*Dentry) - parent.insertChildLocked(name, child) + var newChild Dentry + newChild.Init(fs, childInode) // childInode's ref is transferred to newChild. + parent.insertChildLocked(name, &newChild) + child = &newChild + + // Drop the ref on newChild. This will cause the dentry to get pruned + // from the dentry tree by the end of current filesystem operation + // (before returning to the VFS layer) if another ref is not picked on + // this dentry. + if !childInode.Keep() { + fs.deferDecRef(&newChild) + } } return child, nil } @@ -148,20 +159,19 @@ func (fs *Filesystem) revalidateChildLocked(ctx context.Context, vfsObj *vfs.Vir // Preconditions: Filesystem.mu must be locked for at least reading. // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Done() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if rp.MustBeDir() && !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -171,32 +181,34 @@ func (fs *Filesystem) walkExistingLocked(ctx context.Context, rp *vfs.ResolvingP // walkParentDirLocked is loosely analogous to Linux's // fs/namei.c:path_parentat(). // -// Preconditions: Filesystem.mu must be locked for at least reading. !rp.Done(). +// Preconditions: +// * Filesystem.mu must be locked for at least reading. +// * !rp.Done(). // // Postconditions: Caller must call fs.processDeferredDecRefs*. -func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*vfs.Dentry, Inode, error) { - vfsd := rp.Start() +func (fs *Filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath) (*Dentry, error) { + d := rp.Start().Impl().(*Dentry) for !rp.Final() { var err error - vfsd, err = fs.stepExistingLocked(ctx, rp, vfsd, true /* mayFollowSymlinks */) + d, err = fs.stepExistingLocked(ctx, rp, d, true /* mayFollowSymlinks */) if err != nil { - return nil, nil, err + return nil, err } } - d := vfsd.Impl().(*Dentry) if !d.isDir() { - return nil, nil, syserror.ENOTDIR + return nil, syserror.ENOTDIR } - return vfsd, d.inode, nil + return d, nil } // checkCreateLocked checks that a file named rp.Component() may be created in -// directory parentVFSD, then returns rp.Component(). +// directory parent, then returns rp.Component(). // -// Preconditions: Filesystem.mu must be locked for at least reading. parentInode -// == parentVFSD.Impl().(*Dentry).Inode. isDir(parentInode) == true. -func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *vfs.Dentry, parentInode Inode) (string, error) { - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { +// Preconditions: +// * Filesystem.mu must be locked for at least reading. +// * isDir(parentInode) == true. +func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parent *Dentry) (string, error) { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite|vfs.MayExec); err != nil { return "", err } pc := rp.Component() @@ -206,11 +218,10 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v if len(pc) > linux.NAME_MAX { return "", syserror.ENAMETOOLONG } - // FIXME(gvisor.dev/issue/1193): Data race due to not holding dirMu. - if _, ok := parentVFSD.Impl().(*Dentry).children[pc]; ok { + if _, ok := parent.children[pc]; ok { return "", syserror.EEXIST } - if parentVFSD.IsDead() { + if parent.VFSDentry().IsDead() { return "", syserror.ENOENT } return pc, nil @@ -219,8 +230,8 @@ func checkCreateLocked(ctx context.Context, rp *vfs.ResolvingPath, parentVFSD *v // checkDeleteLocked checks that the file represented by vfsd may be deleted. // // Preconditions: Filesystem.mu must be locked for at least reading. -func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry) error { - parent := vfsd.Impl().(*Dentry).parent +func checkDeleteLocked(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry) error { + parent := d.parent if parent == nil { return syserror.EBUSY } @@ -249,11 +260,11 @@ func (fs *Filesystem) AccessAt(ctx context.Context, rp *vfs.ResolvingPath, creds defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - _, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } - return inode.CheckPermissions(ctx, creds, ats) + return d.inode.CheckPermissions(ctx, creds, ats) } // GetDentryAt implements vfs.FilesystemImpl.GetDentryAt. @@ -261,20 +272,20 @@ func (fs *Filesystem) GetDentryAt(ctx context.Context, rp *vfs.ResolvingPath, op fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return nil, err } if opts.CheckSearchable { - d := vfsd.Impl().(*Dentry) if !d.isDir() { return nil, syserror.ENOTDIR } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } } + vfsd := d.VFSDentry() vfsd.IncRef() // Ownership transferred to caller. return vfsd, nil } @@ -284,12 +295,12 @@ func (fs *Filesystem) GetParentDentryAt(ctx context.Context, rp *vfs.ResolvingPa fs.mu.RLock() defer fs.processDeferredDecRefs(ctx) defer fs.mu.RUnlock() - vfsd, _, err := fs.walkParentDirLocked(ctx, rp) + d, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return nil, err } - vfsd.IncRef() // Ownership transferred to caller. - return vfsd, nil + d.IncRef() // Ownership transferred to caller. + return d.VFSDentry(), nil } // LinkAt implements vfs.FilesystemImpl.LinkAt. @@ -298,13 +309,16 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EEXIST } fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + parent, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -321,11 +335,13 @@ func (fs *Filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. return syserror.EPERM } - childVFSD, err := parentInode.NewLink(ctx, pc, d.inode) + childI, err := parent.inode.NewLink(ctx, pc, d.inode) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + var child Dentry + child.Init(fs, childI) + parent.insertChildLocked(pc, &child) return nil } @@ -335,13 +351,16 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return syserror.EEXIST } fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + parent, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -349,11 +368,16 @@ func (fs *Filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewDir(ctx, pc, opts) + childI, err := parent.inode.NewDir(ctx, pc, opts) if err != nil { - return err + if !opts.ForSyntheticMountpoint || err == syserror.EEXIST { + return err + } + childI = newSyntheticDirectory(ctx, rp.Credentials(), opts.Mode) } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + var child Dentry + child.Init(fs, childI) + parent.insertChildLocked(pc, &child) return nil } @@ -363,13 +387,16 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return syserror.EEXIST } fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + parent, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -377,11 +404,13 @@ func (fs *Filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v return err } defer rp.Mount().EndWrite() - newVFSD, err := parentInode.NewNode(ctx, pc, opts) + newI, err := parent.inode.NewNode(ctx, pc, opts) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, newVFSD.Impl().(*Dentry)) + var newD Dentry + newD.Init(fs, newI) + parent.insertChildLocked(pc, &newD) return nil } @@ -397,28 +426,28 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf // Do not create new file. if opts.Flags&linux.O_CREAT == 0 { fs.mu.RLock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) + defer fs.processDeferredDecRefs(ctx) + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + // Open may block so we need to unlock fs.mu. IncRef d to prevent + // its destruction while fs.mu is unlocked. + d.IncRef() fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) - return inode.Open(ctx, rp, vfsd, opts) + fd, err := d.inode.Open(ctx, rp, d, opts) + d.DecRef(ctx) + return fd, err } // May create new file. mustCreate := opts.Flags&linux.O_EXCL != 0 - vfsd := rp.Start() - inode := vfsd.Impl().(*Dentry).inode + d := rp.Start().Impl().(*Dentry) fs.mu.Lock() unlocked := false unlock := func() { @@ -427,6 +456,10 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf unlocked = true } } + // Process all to-be-decref'd dentries at the end at once. + // Since we defer unlock() AFTER this, fs.mu is guaranteed to be unlocked + // when this is executed. + defer fs.processDeferredDecRefs(ctx) defer unlock() if rp.Done() { if rp.MustBeDir() { @@ -435,22 +468,24 @@ func (fs *Filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if mustCreate { return nil, syserror.EEXIST } - if err := inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - inode.IncRef() - defer inode.DecRef(ctx) + // Open may block so we need to unlock fs.mu. IncRef d to prevent + // its destruction while fs.mu is unlocked. + d.IncRef() unlock() - return inode.Open(ctx, rp, vfsd, opts) + fd, err := d.inode.Open(ctx, rp, d, opts) + d.DecRef(ctx) + return fd, err } afterTrailingSymlink: - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + parent, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return nil, err } // Check for search permission in the parent directory. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayExec); err != nil { return nil, err } // Reject attempts to open directories with O_CREAT. @@ -465,10 +500,10 @@ afterTrailingSymlink: return nil, syserror.ENAMETOOLONG } // Determine whether or not we need to create a file. - childVFSD, err := fs.stepExistingLocked(ctx, rp, parentVFSD, false /* mayFollowSymlinks */) + child, err := fs.stepExistingLocked(ctx, rp, parent, false /* mayFollowSymlinks */) if err == syserror.ENOENT { // Already checked for searchability above; now check for writability. - if err := parentInode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := parent.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } if err := rp.Mount().CheckBeginWrite(); err != nil { @@ -476,16 +511,20 @@ afterTrailingSymlink: } defer rp.Mount().EndWrite() // Create and open the child. - childVFSD, err = parentInode.NewFile(ctx, pc, opts) + childI, err := parent.inode.NewFile(ctx, pc, opts) if err != nil { return nil, err } - child := childVFSD.Impl().(*Dentry) - parentVFSD.Impl().(*Dentry).InsertChild(pc, child) - child.inode.IncRef() - defer child.inode.DecRef(ctx) + var child Dentry + child.Init(fs, childI) + parent.insertChild(pc, &child) + // Open may block so we need to unlock fs.mu. IncRef child to prevent + // its destruction while fs.mu is unlocked. + child.IncRef() unlock() - return child.inode.Open(ctx, rp, childVFSD, opts) + fd, err := child.inode.Open(ctx, rp, &child, opts) + child.DecRef(ctx) + return fd, err } if err != nil { return nil, err @@ -494,7 +533,6 @@ afterTrailingSymlink: if mustCreate { return nil, syserror.EEXIST } - child := childVFSD.Impl().(*Dentry) if rp.ShouldFollowSymlink() && child.isSymlink() { targetVD, targetPathname, err := child.inode.Getlink(ctx, rp.Mount()) if err != nil { @@ -502,7 +540,7 @@ afterTrailingSymlink: } if targetVD.Ok() { err := rp.HandleJump(targetVD) - targetVD.DecRef(ctx) + fs.deferDecRefVD(ctx, targetVD) if err != nil { return nil, err } @@ -518,25 +556,28 @@ afterTrailingSymlink: if err := child.inode.CheckPermissions(ctx, rp.Credentials(), ats); err != nil { return nil, err } - child.inode.IncRef() - defer child.inode.DecRef(ctx) + // Open may block so we need to unlock fs.mu. IncRef child to prevent + // its destruction while fs.mu is unlocked. + child.IncRef() unlock() - return child.inode.Open(ctx, rp, &child.vfsd, opts) + fd, err := child.inode.Open(ctx, rp, child, opts) + child.DecRef(ctx) + return fd, err } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. func (fs *Filesystem) ReadlinkAt(ctx context.Context, rp *vfs.ResolvingPath) (string, error) { fs.mu.RLock() - d, inode, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return "", err } - if !d.Impl().(*Dentry).isSymlink() { + if !d.isSymlink() { return "", syserror.EINVAL } - return inode.Readlink(ctx) + return d.inode.Readlink(ctx, rp.Mount()) } // RenameAt implements vfs.FilesystemImpl.RenameAt. @@ -548,16 +589,15 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa noReplace := opts.Flags&linux.RENAME_NOREPLACE != 0 fs.mu.Lock() - defer fs.processDeferredDecRefsLocked(ctx) + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() // Resolve the destination directory first to verify that it's on this // Mount. - dstDirVFSD, dstDirInode, err := fs.walkParentDirLocked(ctx, rp) + dstDir, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - dstDir := dstDirVFSD.Impl().(*Dentry) mnt := rp.Mount() if mnt != oldParentVD.Mount() { return syserror.EXDEV @@ -575,16 +615,15 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa if err != nil { return err } - srcVFSD := &src.vfsd // Can we remove the src dentry? - if err := checkDeleteLocked(ctx, rp, srcVFSD); err != nil { + if err := checkDeleteLocked(ctx, rp, src); err != nil { return err } // Can we create the dst dentry? var dst *Dentry - pc, err := checkCreateLocked(ctx, rp, dstDirVFSD, dstDirInode) + pc, err := checkCreateLocked(ctx, rp, dstDir) switch err { case nil: // Ok, continue with rename as replacement. @@ -595,14 +634,14 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } dst = dstDir.children[pc] if dst == nil { - panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDirVFSD)) + panic(fmt.Sprintf("Child %q for parent Dentry %+v disappeared inside atomic section?", pc, dstDir)) } default: return err } var dstVFSD *vfs.Dentry if dst != nil { - dstVFSD = &dst.vfsd + dstVFSD = dst.VFSDentry() } mntns := vfs.MountNamespaceFromContext(ctx) @@ -618,35 +657,44 @@ func (fs *Filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa defer dstDir.dirMu.Unlock() } + srcVFSD := src.VFSDentry() if err := virtfs.PrepareRenameDentry(mntns, srcVFSD, dstVFSD); err != nil { return err } - replaced, err := srcDir.inode.Rename(ctx, src.name, pc, srcVFSD, dstDirVFSD) + err = srcDir.inode.Rename(ctx, src.name, pc, src.inode, dstDir.inode) if err != nil { virtfs.AbortRenameDentry(srcVFSD, dstVFSD) return err } delete(srcDir.children, src.name) if srcDir != dstDir { - fs.deferDecRef(srcDirVFSD) - dstDir.IncRef() + fs.deferDecRef(srcDir) // child (src) drops ref on old parent. + dstDir.IncRef() // child (src) takes a ref on the new parent. } src.parent = dstDir src.name = pc if dstDir.children == nil { dstDir.children = make(map[string]*Dentry) } + replaced := dstDir.children[pc] dstDir.children[pc] = src - virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaced) + var replaceVFSD *vfs.Dentry + if replaced != nil { + // deferDecRef so that fs.mu and dstDir.mu are unlocked by then. + fs.deferDecRef(replaced) + replaceVFSD = replaced.VFSDentry() + } + virtfs.CommitRenameReplaceDentry(ctx, srcVFSD, replaceVFSD) return nil } // RmdirAt implements vfs.FilesystemImpl.RmdirAt. func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - vfsd, inode, err := fs.walkExistingLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } @@ -654,14 +702,13 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if !d.isDir() { return syserror.ENOTDIR } - if inode.HasChildren() { + if d.inode.HasChildren() { return syserror.ENOTEMPTY } virtfs := rp.VirtualFilesystem() @@ -671,13 +718,18 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.RmDir(ctx, rp.Component(), vfsd); err != nil { + + if err := parentDentry.inode.RmDir(ctx, d.name, d.inode); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } + delete(parentDentry.children, d.name) + // Defer decref so that fs.mu and parentDentry.dirMu are unlocked by then. + fs.deferDecRef(d) virtfs.CommitDeleteDentry(ctx, vfsd) return nil } @@ -685,41 +737,40 @@ func (fs *Filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error // SetStatAt implements vfs.FilesystemImpl.SetStatAt. func (fs *Filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } if opts.Stat.Mask == 0 { return nil } - return inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) + return d.inode.SetStat(ctx, fs.VFSFilesystem(), rp.Credentials(), opts) } // StatAt implements vfs.FilesystemImpl.StatAt. func (fs *Filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return linux.Statx{}, err } - return inode.Stat(ctx, fs.VFSFilesystem(), opts) + return d.inode.Stat(ctx, fs.VFSFilesystem(), opts) } // StatFSAt implements vfs.FilesystemImpl.StatFSAt. func (fs *Filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linux.Statfs, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return linux.Statfs{}, err } - // TODO(gvisor.dev/issue/1193): actually implement statfs. - return linux.Statfs{}, syserror.ENOSYS + return d.inode.StatFS(ctx, fs.VFSFilesystem()) } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. @@ -728,13 +779,16 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return syserror.EEXIST } fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - parentVFSD, parentInode, err := fs.walkParentDirLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + parent, err := fs.walkParentDirLocked(ctx, rp) if err != nil { return err } - pc, err := checkCreateLocked(ctx, rp, parentVFSD, parentInode) + parent.dirMu.Lock() + defer parent.dirMu.Unlock() + + pc, err := checkCreateLocked(ctx, rp, parent) if err != nil { return err } @@ -742,20 +796,23 @@ func (fs *Filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ return err } defer rp.Mount().EndWrite() - childVFSD, err := parentInode.NewSymlink(ctx, pc, target) + childI, err := parent.inode.NewSymlink(ctx, pc, target) if err != nil { return err } - parentVFSD.Impl().(*Dentry).InsertChild(pc, childVFSD.Impl().(*Dentry)) + var child Dentry + child.Init(fs, childI) + parent.insertChildLocked(pc, &child) return nil } // UnlinkAt implements vfs.FilesystemImpl.UnlinkAt. func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error { fs.mu.Lock() + defer fs.processDeferredDecRefs(ctx) defer fs.mu.Unlock() - vfsd, _, err := fs.walkExistingLocked(ctx, rp) - fs.processDeferredDecRefsLocked(ctx) + + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } @@ -763,10 +820,9 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } defer rp.Mount().EndWrite() - if err := checkDeleteLocked(ctx, rp, vfsd); err != nil { + if err := checkDeleteLocked(ctx, rp, d); err != nil { return err } - d := vfsd.Impl().(*Dentry) if d.isDir() { return syserror.EISDIR } @@ -776,38 +832,42 @@ func (fs *Filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error defer parentDentry.dirMu.Unlock() mntns := vfs.MountNamespaceFromContext(ctx) defer mntns.DecRef(ctx) + vfsd := d.VFSDentry() if err := virtfs.PrepareDeleteDentry(mntns, vfsd); err != nil { return err } - if err := parentDentry.inode.Unlink(ctx, rp.Component(), vfsd); err != nil { + if err := parentDentry.inode.Unlink(ctx, d.name, d.inode); err != nil { virtfs.AbortDeleteDentry(vfsd) return err } + delete(parentDentry.children, d.name) + // Defer decref so that fs.mu and parentDentry.dirMu are unlocked by then. + fs.deferDecRef(d) virtfs.CommitDeleteDentry(ctx, vfsd) return nil } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *Filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { fs.mu.RLock() - _, inode, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + d, err := fs.walkExistingLocked(ctx, rp) if err != nil { return nil, err } - if err := inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { + if err := d.inode.CheckPermissions(ctx, rp.Credentials(), vfs.MayWrite); err != nil { return nil, err } return nil, syserror.ECONNREFUSED } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *Filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + _, err := fs.walkExistingLocked(ctx, rp) if err != nil { return nil, err } @@ -815,12 +875,12 @@ func (fs *Filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, si return nil, syserror.ENOTSUP } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *Filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + _, err := fs.walkExistingLocked(ctx, rp) if err != nil { return "", err } @@ -828,12 +888,12 @@ func (fs *Filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt return "", syserror.ENOTSUP } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *Filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *Filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + _, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } @@ -841,12 +901,12 @@ func (fs *Filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt return syserror.ENOTSUP } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *Filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *Filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() - _, _, err := fs.walkExistingLocked(ctx, rp) - fs.mu.RUnlock() - fs.processDeferredDecRefs(ctx) + defer fs.processDeferredDecRefs(ctx) + defer fs.mu.RUnlock() + _, err := fs.walkExistingLocked(ctx, rp) if err != nil { return err } @@ -860,3 +920,16 @@ func (fs *Filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe defer fs.mu.RUnlock() return genericPrependPath(vfsroot, vd.Mount(), vd.Dentry().Impl().(*Dentry), b) } + +func (fs *Filesystem) deferDecRefVD(ctx context.Context, vd vfs.VirtualDentry) { + if d, ok := vd.Dentry().Impl().(*Dentry); ok && d.fs == fs { + // The following is equivalent to vd.DecRef(ctx). This is needed + // because if d belongs to this filesystem, we can not DecRef it right + // away as we may be holding fs.mu. d.DecRef may acquire fs.mu. So we + // defer the DecRef to when locks are dropped. + vd.Mount().DecRef(ctx) + fs.deferDecRef(d) + } else { + vd.DecRef(ctx) + } +} diff --git a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go index fe8a1e710..d9d76758a 100644 --- a/pkg/sentry/fsimpl/kernfs/inode_impl_util.go +++ b/pkg/sentry/fsimpl/kernfs/inode_impl_util.go @@ -20,11 +20,12 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // InodeNoopRefCount partially implements the Inode interface, specifically the @@ -32,7 +33,10 @@ import ( // count for inodes, performing no extra actions when references are obtained or // released. This is suitable for simple file inodes that don't reference any // resources. +// +// +stateify savable type InodeNoopRefCount struct { + InodeTemporary } // IncRef implements Inode.IncRef. @@ -51,30 +55,32 @@ func (InodeNoopRefCount) TryIncRef() bool { // InodeDirectoryNoNewChildren partially implements the Inode interface. // InodeDirectoryNoNewChildren represents a directory inode which does not // support creation of new children. +// +// +stateify savable type InodeDirectoryNoNewChildren struct{} // NewFile implements Inode.NewFile. -func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewFile(context.Context, string, vfs.OpenOptions) (Inode, error) { return nil, syserror.EPERM } // NewDir implements Inode.NewDir. -func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewDir(context.Context, string, vfs.MkdirOptions) (Inode, error) { return nil, syserror.EPERM } // NewLink implements Inode.NewLink. -func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewLink(context.Context, string, Inode) (Inode, error) { return nil, syserror.EPERM } // NewSymlink implements Inode.NewSymlink. -func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewSymlink(context.Context, string, string) (Inode, error) { return nil, syserror.EPERM } // NewNode implements Inode.NewNode. -func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOptions) (Inode, error) { return nil, syserror.EPERM } @@ -82,7 +88,10 @@ func (InodeDirectoryNoNewChildren) NewNode(context.Context, string, vfs.MknodOpt // inodeDirectory and inodeDynamicDirectory sub interfaces. Inodes that do not // represent directories can embed this to provide no-op implementations for // directory-related functions. +// +// +stateify savable type InodeNotDirectory struct { + InodeAlwaysValid } // HasChildren implements Inode.HasChildren. @@ -91,89 +100,64 @@ func (InodeNotDirectory) HasChildren() bool { } // NewFile implements Inode.NewFile. -func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewFile(context.Context, string, vfs.OpenOptions) (Inode, error) { panic("NewFile called on non-directory inode") } // NewDir implements Inode.NewDir. -func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewDir(context.Context, string, vfs.MkdirOptions) (Inode, error) { panic("NewDir called on non-directory inode") } // NewLink implements Inode.NewLinkink. -func (InodeNotDirectory) NewLink(context.Context, string, Inode) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewLink(context.Context, string, Inode) (Inode, error) { panic("NewLink called on non-directory inode") } // NewSymlink implements Inode.NewSymlink. -func (InodeNotDirectory) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewSymlink(context.Context, string, string) (Inode, error) { panic("NewSymlink called on non-directory inode") } // NewNode implements Inode.NewNode. -func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (InodeNotDirectory) NewNode(context.Context, string, vfs.MknodOptions) (Inode, error) { panic("NewNode called on non-directory inode") } // Unlink implements Inode.Unlink. -func (InodeNotDirectory) Unlink(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) Unlink(context.Context, string, Inode) error { panic("Unlink called on non-directory inode") } // RmDir implements Inode.RmDir. -func (InodeNotDirectory) RmDir(context.Context, string, *vfs.Dentry) error { +func (InodeNotDirectory) RmDir(context.Context, string, Inode) error { panic("RmDir called on non-directory inode") } // Rename implements Inode.Rename. -func (InodeNotDirectory) Rename(context.Context, string, string, *vfs.Dentry, *vfs.Dentry) (*vfs.Dentry, error) { +func (InodeNotDirectory) Rename(context.Context, string, string, Inode, Inode) error { panic("Rename called on non-directory inode") } // Lookup implements Inode.Lookup. -func (InodeNotDirectory) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +func (InodeNotDirectory) Lookup(ctx context.Context, name string) (Inode, error) { panic("Lookup called on non-directory inode") } // IterDirents implements Inode.IterDirents. -func (InodeNotDirectory) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { +func (InodeNotDirectory) IterDirents(ctx context.Context, mnt *vfs.Mount, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { panic("IterDirents called on non-directory inode") } -// Valid implements Inode.Valid. -func (InodeNotDirectory) Valid(context.Context) bool { - return true -} - -// InodeNoDynamicLookup partially implements the Inode interface, specifically -// the inodeDynamicLookup sub interface. Directory inodes that do not support -// dymanic entries (i.e. entries that are not "hashed" into the -// vfs.Dentry.children) can embed this to provide no-op implementations for -// functions related to dynamic entries. -type InodeNoDynamicLookup struct{} - -// Lookup implements Inode.Lookup. -func (InodeNoDynamicLookup) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { - return nil, syserror.ENOENT -} - -// IterDirents implements Inode.IterDirents. -func (InodeNoDynamicLookup) IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { - return offset, nil -} - -// Valid implements Inode.Valid. -func (InodeNoDynamicLookup) Valid(ctx context.Context) bool { - return true -} - // InodeNotSymlink partially implements the Inode interface, specifically the // inodeSymlink sub interface. All inodes that are not symlinks may embed this // to return the appropriate errors from symlink-related functions. +// +// +stateify savable type InodeNotSymlink struct{} // Readlink implements Inode.Readlink. -func (InodeNotSymlink) Readlink(context.Context) (string, error) { +func (InodeNotSymlink) Readlink(context.Context, *vfs.Mount) (string, error) { return "", syserror.EINVAL } @@ -187,18 +171,26 @@ func (InodeNotSymlink) Getlink(context.Context, *vfs.Mount) (vfs.VirtualDentry, // inode attributes. // // Must be initialized by Init prior to first use. +// +// +stateify savable type InodeAttrs struct { - devMajor uint32 - devMinor uint32 - ino uint64 - mode uint32 - uid uint32 - gid uint32 - nlink uint32 + devMajor uint32 + devMinor uint32 + ino uint64 + mode uint32 + uid uint32 + gid uint32 + nlink uint32 + blockSize uint32 + + // Timestamps, all nsecs from the Unix epoch. + atime int64 + mtime int64 + ctime int64 } // Init initializes this InodeAttrs. -func (a *InodeAttrs) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, mode linux.FileMode) { +func (a *InodeAttrs) Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, mode linux.FileMode) { if mode.FileType() == 0 { panic(fmt.Sprintf("No file type specified in 'mode' for InodeAttrs.Init(): mode=0%o", mode)) } @@ -214,6 +206,11 @@ func (a *InodeAttrs) Init(creds *auth.Credentials, devMajor, devMinor uint32, in atomic.StoreUint32(&a.uid, uint32(creds.EffectiveKUID)) atomic.StoreUint32(&a.gid, uint32(creds.EffectiveKGID)) atomic.StoreUint32(&a.nlink, nlink) + atomic.StoreUint32(&a.blockSize, usermem.PageSize) + now := ktime.NowFromContext(ctx).Nanoseconds() + atomic.StoreInt64(&a.atime, now) + atomic.StoreInt64(&a.mtime, now) + atomic.StoreInt64(&a.ctime, now) } // DevMajor returns the device major number. @@ -236,12 +233,33 @@ func (a *InodeAttrs) Mode() linux.FileMode { return linux.FileMode(atomic.LoadUint32(&a.mode)) } +// TouchAtime updates a.atime to the current time. +func (a *InodeAttrs) TouchAtime(ctx context.Context, mnt *vfs.Mount) { + if mnt.Flags.NoATime || mnt.ReadOnly() { + return + } + if err := mnt.CheckBeginWrite(); err != nil { + return + } + atomic.StoreInt64(&a.atime, ktime.NowFromContext(ctx).Nanoseconds()) + mnt.EndWrite() +} + +// TouchCMtime updates a.{c/m}time to the current time. The caller should +// synchronize calls to this so that ctime and mtime are updated to the same +// value. +func (a *InodeAttrs) TouchCMtime(ctx context.Context) { + now := ktime.NowFromContext(ctx).Nanoseconds() + atomic.StoreInt64(&a.mtime, now) + atomic.StoreInt64(&a.ctime, now) +} + // Stat partially implements Inode.Stat. Note that this function doesn't provide // all the stat fields, and the embedder should consider extending the result // with filesystem-specific fields. func (a *InodeAttrs) Stat(context.Context, *vfs.Filesystem, vfs.StatOptions) (linux.Statx, error) { var stat linux.Statx - stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK + stat.Mask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO | linux.STATX_NLINK | linux.STATX_ATIME | linux.STATX_MTIME | linux.STATX_CTIME stat.DevMajor = a.devMajor stat.DevMinor = a.devMinor stat.Ino = atomic.LoadUint64(&a.ino) @@ -249,9 +267,10 @@ func (a *InodeAttrs) Stat(context.Context, *vfs.Filesystem, vfs.StatOptions) (li stat.UID = atomic.LoadUint32(&a.uid) stat.GID = atomic.LoadUint32(&a.gid) stat.Nlink = atomic.LoadUint32(&a.nlink) - - // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. - + stat.Blksize = atomic.LoadUint32(&a.blockSize) + stat.Atime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&a.atime)) + stat.Mtime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&a.mtime)) + stat.Ctime = linux.NsecToStatxTimestamp(atomic.LoadInt64(&a.ctime)) return stat, nil } @@ -260,9 +279,17 @@ func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *aut if opts.Stat.Mask == 0 { return nil } - if opts.Stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID) != 0 { + + // Note that not all fields are modifiable. For example, the file type and + // inode numbers are immutable after node creation. Setting the size is often + // allowed by kernfs files but does not do anything. If some other behavior is + // needed, the embedder should consider extending SetStat. + if opts.Stat.Mask&^(linux.STATX_MODE|linux.STATX_UID|linux.STATX_GID|linux.STATX_ATIME|linux.STATX_MTIME|linux.STATX_SIZE) != 0 { return syserror.EPERM } + if opts.Stat.Mask&linux.STATX_SIZE != 0 && a.Mode().IsDir() { + return syserror.EISDIR + } if err := vfs.CheckSetStat(ctx, creds, &opts, a.Mode(), auth.KUID(atomic.LoadUint32(&a.uid)), auth.KGID(atomic.LoadUint32(&a.gid))); err != nil { return err } @@ -285,12 +312,19 @@ func (a *InodeAttrs) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *aut atomic.StoreUint32(&a.gid, stat.GID) } - // Note that not all fields are modifiable. For example, the file type and - // inode numbers are immutable after node creation. - - // TODO(gvisor.dev/issue/1193): Implement other stat fields like timestamps. - // Also, STATX_SIZE will need some special handling, because read-only static - // files should return EIO for truncate operations. + now := ktime.NowFromContext(ctx).Nanoseconds() + if stat.Mask&linux.STATX_ATIME != 0 { + if stat.Atime.Nsec == linux.UTIME_NOW { + stat.Atime = linux.NsecToStatxTimestamp(now) + } + atomic.StoreInt64(&a.atime, stat.Atime.ToNsec()) + } + if stat.Mask&linux.STATX_MTIME != 0 { + if stat.Mtime.Nsec == linux.UTIME_NOW { + stat.Mtime = linux.NsecToStatxTimestamp(now) + } + atomic.StoreInt64(&a.mtime, stat.Mtime.ToNsec()) + } return nil } @@ -321,13 +355,17 @@ func (a *InodeAttrs) DecLinks() { } } +// +stateify savable type slot struct { - Name string - Dentry *vfs.Dentry + name string + inode Inode + static bool slotEntry } // OrderedChildrenOptions contains initialization options for OrderedChildren. +// +// +stateify savable type OrderedChildrenOptions struct { // Writable indicates whether vfs.FilesystemImpl methods implemented by // OrderedChildren may modify the tracked children. This applies to @@ -337,20 +375,28 @@ type OrderedChildrenOptions struct { } // OrderedChildren partially implements the Inode interface. OrderedChildren can -// be embedded in directory inodes to keep track of the children in the +// be embedded in directory inodes to keep track of children in the // directory, and can then be used to implement a generic directory FD -- see -// GenericDirectoryFD. OrderedChildren is not compatible with dynamic -// directories. +// GenericDirectoryFD. +// +// OrderedChildren can represent a node in an Inode tree. The children inodes +// might be directories themselves using OrderedChildren; hence extending the +// tree. The parent inode (OrderedChildren user) holds a ref on all its static +// children. This lets the static inodes outlive their associated dentry. +// While the dentry might have to be regenerated via a Lookup() call, we can +// keep reusing the same static inode. These static children inodes are finally +// DecRef'd when this directory inode is being destroyed. This makes +// OrderedChildren suitable for static directory entries as well. // // Must be initialize with Init before first use. +// +// +stateify savable type OrderedChildren struct { - refs.AtomicRefCount - // Can children be modified by user syscalls? It set to false, interface // methods that would modify the children return EPERM. Immutable. writable bool - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` order slotList set map[string]*slot } @@ -361,36 +407,66 @@ func (o *OrderedChildren) Init(opts OrderedChildrenOptions) { o.set = make(map[string]*slot) } -// DecRef implements Inode.DecRef. -func (o *OrderedChildren) DecRef(ctx context.Context) { - o.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { - o.mu.Lock() - defer o.mu.Unlock() - o.order.Reset() - o.set = nil - }) +// Destroy clears the children stored in o. It should be called by structs +// embedding OrderedChildren upon destruction, i.e. when their reference count +// reaches zero. +func (o *OrderedChildren) Destroy(ctx context.Context) { + o.mu.Lock() + defer o.mu.Unlock() + // Drop the ref that o owns on the static inodes it holds. + for _, s := range o.set { + if s.static { + s.inode.DecRef(ctx) + } + } + o.order.Reset() + o.set = nil } -// Populate inserts children into this OrderedChildren, and d's dentry -// cache. Populate returns the number of directories inserted, which the caller +// Populate inserts static children into this OrderedChildren. +// Populate returns the number of directories inserted, which the caller // may use to update the link count for the parent directory. // -// Precondition: d must represent a directory inode. children must not contain -// any conflicting entries already in o. -func (o *OrderedChildren) Populate(d *Dentry, children map[string]*Dentry) uint32 { +// Precondition: +// * d must represent a directory inode. +// * children must not contain any conflicting entries already in o. +// * Caller must hold a reference on all inodes passed. +// +// Postcondition: Caller's references on inodes are transferred to o. +func (o *OrderedChildren) Populate(children map[string]Inode) uint32 { var links uint32 for name, child := range children { - if child.isDir() { + if child.Mode().IsDir() { links++ } - if err := o.Insert(name, child.VFSDentry()); err != nil { - panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v) into %+v", name, child, d)) + if err := o.insert(name, child, true); err != nil { + panic(fmt.Sprintf("Collision when attempting to insert child %q (%+v)", name, child)) } - d.InsertChild(name, child) } return links } +// Lookup implements Inode.Lookup. +func (o *OrderedChildren) Lookup(ctx context.Context, name string) (Inode, error) { + o.mu.RLock() + defer o.mu.RUnlock() + + s, ok := o.set[name] + if !ok { + return nil, syserror.ENOENT + } + + s.inode.IncRef() // This ref is passed to the dentry upon creation via Init. + return s.inode, nil +} + +// IterDirents implements Inode.IterDirents. +func (o *OrderedChildren) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { + // All entries from OrderedChildren have already been handled in + // GenericDirectoryFD.IterDirents. + return offset, nil +} + // HasChildren implements Inode.HasChildren. func (o *OrderedChildren) HasChildren() bool { o.mu.RLock() @@ -398,17 +474,27 @@ func (o *OrderedChildren) HasChildren() bool { return len(o.set) > 0 } -// Insert inserts child into o. This ignores the writability of o, as this is -// not part of the vfs.FilesystemImpl interface, and is a lower-level operation. -func (o *OrderedChildren) Insert(name string, child *vfs.Dentry) error { +// Insert inserts a dynamic child into o. This ignores the writability of o, as +// this is not part of the vfs.FilesystemImpl interface, and is a lower-level operation. +func (o *OrderedChildren) Insert(name string, child Inode) error { + return o.insert(name, child, false) +} + +// insert inserts child into o. +// +// Precondition: Caller must be holding a ref on child if static is true. +// +// Postcondition: Caller's ref on child is transferred to o if static is true. +func (o *OrderedChildren) insert(name string, child Inode, static bool) error { o.mu.Lock() defer o.mu.Unlock() if _, ok := o.set[name]; ok { return syserror.EEXIST } s := &slot{ - Name: name, - Dentry: child, + name: name, + inode: child, + static: static, } o.order.PushBack(s) o.set[name] = s @@ -418,44 +504,49 @@ func (o *OrderedChildren) Insert(name string, child *vfs.Dentry) error { // Precondition: caller must hold o.mu for writing. func (o *OrderedChildren) removeLocked(name string) { if s, ok := o.set[name]; ok { + if s.static { + panic(fmt.Sprintf("removeLocked called on a static inode: %v", s.inode)) + } delete(o.set, name) o.order.Remove(s) } } // Precondition: caller must hold o.mu for writing. -func (o *OrderedChildren) replaceChildLocked(name string, new *vfs.Dentry) *vfs.Dentry { +func (o *OrderedChildren) replaceChildLocked(ctx context.Context, name string, newI Inode) { if s, ok := o.set[name]; ok { + if s.static { + panic(fmt.Sprintf("replacing a static inode: %v", s.inode)) + } + // Existing slot with given name, simply replace the dentry. - var old *vfs.Dentry - old, s.Dentry = s.Dentry, new - return old + s.inode = newI } // No existing slot with given name, create and hash new slot. s := &slot{ - Name: name, - Dentry: new, + name: name, + inode: newI, + static: false, } o.order.PushBack(s) o.set[name] = s - return nil } // Precondition: caller must hold o.mu for reading or writing. -func (o *OrderedChildren) checkExistingLocked(name string, child *vfs.Dentry) error { +func (o *OrderedChildren) checkExistingLocked(name string, child Inode) error { s, ok := o.set[name] if !ok { return syserror.ENOENT } - if s.Dentry != child { - panic(fmt.Sprintf("Dentry hashed into inode doesn't match what vfs thinks! OrderedChild: %+v, vfs: %+v", s.Dentry, child)) + if s.inode != child { + panic(fmt.Sprintf("Inode doesn't match what kernfs thinks! OrderedChild: %+v, kernfs: %+v", s.inode, child)) } return nil } // Unlink implements Inode.Unlink. -func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.Dentry) error { +func (o *OrderedChildren) Unlink(ctx context.Context, name string, child Inode) error { if !o.writable { return syserror.EPERM } @@ -470,13 +561,14 @@ func (o *OrderedChildren) Unlink(ctx context.Context, name string, child *vfs.De return nil } -// Rmdir implements Inode.Rmdir. -func (o *OrderedChildren) RmDir(ctx context.Context, name string, child *vfs.Dentry) error { +// RmDir implements Inode.RmDir. +func (o *OrderedChildren) RmDir(ctx context.Context, name string, child Inode) error { // We're not responsible for checking that child is a directory, that it's // empty, or updating any link counts; so this is the same as unlink. return o.Unlink(ctx, name, child) } +// +stateify savable type renameAcrossDifferentImplementationsError struct{} func (renameAcrossDifferentImplementationsError) Error() string { @@ -492,13 +584,13 @@ func (renameAcrossDifferentImplementationsError) Error() string { // that will support Rename. // // Postcondition: reference on any replaced dentry transferred to caller. -func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (*vfs.Dentry, error) { - dst, ok := dstDir.Impl().(*Dentry).inode.(interface{}).(*OrderedChildren) +func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, child, dstDir Inode) error { + dst, ok := dstDir.(interface{}).(*OrderedChildren) if !ok { - return nil, renameAcrossDifferentImplementationsError{} + return renameAcrossDifferentImplementationsError{} } if !o.writable || !dst.writable { - return nil, syserror.EPERM + return syserror.EPERM } // Note: There's a potential deadlock below if concurrent calls to Rename // refer to the same src and dst directories in reverse. We avoid any @@ -511,12 +603,12 @@ func (o *OrderedChildren) Rename(ctx context.Context, oldname, newname string, c defer dst.mu.Unlock() } if err := o.checkExistingLocked(oldname, child); err != nil { - return nil, err + return err } // TODO(gvisor.dev/issue/3027): Check sticky bit before removing. - replaced := dst.replaceChildLocked(newname, child) - return replaced, nil + dst.replaceChildLocked(ctx, newname, child) + return nil } // nthLocked returns an iterator to the nth child tracked by this object. The @@ -535,12 +627,14 @@ func (o *OrderedChildren) nthLocked(i int64) *slot { } // InodeSymlink partially implements Inode interface for symlinks. +// +// +stateify savable type InodeSymlink struct { InodeNotDirectory } // Open implements Inode.Open. -func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ELOOP } @@ -549,43 +643,46 @@ func (InodeSymlink) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.D // // +stateify savable type StaticDirectory struct { - InodeNotSymlink - InodeDirectoryNoNewChildren + InodeAlwaysValid InodeAttrs - InodeNoDynamicLookup + InodeDirectoryNoNewChildren + InodeNoStatFS + InodeNotSymlink + InodeTemporary OrderedChildren + StaticDirectoryRefs - locks vfs.FileLocks + locks vfs.FileLocks + fdOpts GenericDirectoryFDOptions } var _ Inode = (*StaticDirectory)(nil) // NewStaticDir creates a new static directory and returns its dentry. -func NewStaticDir(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, children map[string]*Dentry) *Dentry { +func NewStaticDir(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, children map[string]Inode, fdOpts GenericDirectoryFDOptions) Inode { inode := &StaticDirectory{} - inode.Init(creds, devMajor, devMinor, ino, perm) - - dentry := &Dentry{} - dentry.Init(inode) + inode.Init(ctx, creds, devMajor, devMinor, ino, perm, fdOpts) + inode.EnableLeakCheck() inode.OrderedChildren.Init(OrderedChildrenOptions{}) - links := inode.OrderedChildren.Populate(dentry, children) + links := inode.OrderedChildren.Populate(children) inode.IncLinks(links) - return dentry + return inode } // Init initializes StaticDirectory. -func (s *StaticDirectory) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { +func (s *StaticDirectory) Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode, fdOpts GenericDirectoryFDOptions) { if perm&^linux.PermissionsMask != 0 { panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) } - s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeDirectory|perm) + s.fdOpts = fdOpts + s.InodeAttrs.Init(ctx, creds, devMajor, devMinor, ino, linux.ModeDirectory|perm) } -// Open implements kernfs.Inode. -func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := NewGenericDirectoryFD(rp.Mount(), vfsd, &s.OrderedChildren, &s.locks, &opts) +// Open implements Inode.Open. +func (s *StaticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &s.OrderedChildren, &s.locks, &opts, s.fdOpts) if err != nil { return nil, err } @@ -597,10 +694,38 @@ func (*StaticDirectory) SetStat(context.Context, *vfs.Filesystem, *auth.Credenti return syserror.EPERM } -// AlwaysValid partially implements kernfs.inodeDynamicLookup. -type AlwaysValid struct{} +// DecRef implements Inode.DecRef. +func (s *StaticDirectory) DecRef(ctx context.Context) { + s.StaticDirectoryRefs.DecRef(func() { s.Destroy(ctx) }) +} -// Valid implements kernfs.inodeDynamicLookup. -func (*AlwaysValid) Valid(context.Context) bool { +// InodeAlwaysValid partially implements Inode. +// +// +stateify savable +type InodeAlwaysValid struct{} + +// Valid implements Inode.Valid. +func (*InodeAlwaysValid) Valid(context.Context) bool { return true } + +// InodeTemporary partially implements Inode. +// +// +stateify savable +type InodeTemporary struct{} + +// Keep implements Inode.Keep. +func (*InodeTemporary) Keep() bool { + return false +} + +// InodeNoStatFS partially implements the Inode interface, where the client +// filesystem doesn't support statfs(2). +// +// +stateify savable +type InodeNoStatFS struct{} + +// StatFS implements Inode.StatFS. +func (*InodeNoStatFS) StatFS(context.Context, *vfs.Filesystem) (linux.Statfs, error) { + return linux.Statfs{}, syserror.ENOSYS +} diff --git a/pkg/sentry/fsimpl/kernfs/kernfs.go b/pkg/sentry/fsimpl/kernfs/kernfs.go index 51dbc050c..5c5e09ac5 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs.go @@ -29,12 +29,16 @@ // // Reference Model: // -// Kernfs dentries represents named pointers to inodes. Dentries and inode have +// Kernfs dentries represents named pointers to inodes. Kernfs is solely +// reponsible for maintaining and modifying its dentry tree; inode +// implementations can not access the tree. Dentries and inodes have // independent lifetimes and reference counts. A child dentry unconditionally // holds a reference on its parent directory's dentry. A dentry also holds a -// reference on the inode it points to. Multiple dentries can point to the same -// inode (for example, in the case of hardlinks). File descriptors hold a -// reference to the dentry they're opened on. +// reference on the inode it points to (although that might not be the only +// reference on the inode). Due to this inodes can outlive the dentries that +// point to them. Multiple dentries can point to the same inode (for example, +// in the case of hardlinks). File descriptors hold a reference to the dentry +// they're opened on. // // Dentries are guaranteed to exist while holding Filesystem.mu for // reading. Dropping dentries require holding Filesystem.mu for writing. To @@ -47,8 +51,8 @@ // kernfs.Dentry.dirMu // vfs.VirtualFilesystem.mountMu // vfs.Dentry.mu -// kernfs.Filesystem.droppedDentriesMu // (inode implementation locks, if any) +// kernfs.Filesystem.droppedDentriesMu package kernfs import ( @@ -57,7 +61,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -66,15 +69,17 @@ import ( // Filesystem mostly implements vfs.FilesystemImpl for a generic in-memory // filesystem. Concrete implementations are expected to embed this in their own // Filesystem type. +// +// +stateify savable type Filesystem struct { vfsfs vfs.Filesystem - droppedDentriesMu sync.Mutex + droppedDentriesMu sync.Mutex `state:"nosave"` // droppedDentries is a list of dentries waiting to be DecRef()ed. This is // used to defer dentry destruction until mu can be acquired for // writing. Protected by droppedDentriesMu. - droppedDentries []*vfs.Dentry + droppedDentries []*Dentry // mu synchronizes the lifetime of Dentries on this filesystem. Holding it // for reading guarantees continued existence of any resolved dentries, but @@ -93,22 +98,32 @@ type Filesystem struct { // example: // // fs.mu.RLock() - // fs.mu.processDeferredDecRefs() + // defer fs.processDeferredDecRefs() // defer fs.mu.RUnlock() // ... // fs.deferDecRef(dentry) - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` // nextInoMinusOne is used to to allocate inode numbers on this // filesystem. Must be accessed by atomic operations. nextInoMinusOne uint64 + + // cachedDentries contains all dentries with 0 references. (Due to race + // conditions, it may also contain dentries with non-zero references.) + // cachedDentriesLen is the number of dentries in cachedDentries. These + // fields are protected by mu. + cachedDentries dentryList + cachedDentriesLen uint64 + + // MaxCachedDentries is the maximum size of cachedDentries. If not set, + // defaults to 0 and kernfs does not cache any dentries. This is immutable. + MaxCachedDentries uint64 } // deferDecRef defers dropping a dentry ref until the next call to // processDeferredDecRefs{,Locked}. See comment on Filesystem.mu. -// -// Precondition: d must not already be pending destruction. -func (fs *Filesystem) deferDecRef(d *vfs.Dentry) { +// This may be called while Filesystem.mu or Dentry.dirMu is locked. +func (fs *Filesystem) deferDecRef(d *Dentry) { fs.droppedDentriesMu.Lock() fs.droppedDentries = append(fs.droppedDentries, d) fs.droppedDentriesMu.Unlock() @@ -116,17 +131,14 @@ func (fs *Filesystem) deferDecRef(d *vfs.Dentry) { // processDeferredDecRefs calls vfs.Dentry.DecRef on all dentries in the // droppedDentries list. See comment on Filesystem.mu. +// +// Precondition: Filesystem.mu or Dentry.dirMu must NOT be locked. func (fs *Filesystem) processDeferredDecRefs(ctx context.Context) { - fs.mu.Lock() - fs.processDeferredDecRefsLocked(ctx) - fs.mu.Unlock() -} - -// Precondition: fs.mu must be held for writing. -func (fs *Filesystem) processDeferredDecRefsLocked(ctx context.Context) { fs.droppedDentriesMu.Lock() for _, d := range fs.droppedDentries { - d.DecRef(ctx) + // Defer the DecRef call so that we are not holding droppedDentriesMu + // when DecRef is called. + defer d.DecRef(ctx) } fs.droppedDentries = fs.droppedDentries[:0] // Keep slice memory for reuse. fs.droppedDentriesMu.Unlock() @@ -155,15 +167,24 @@ const ( // // A kernfs dentry is similar to a dentry in a traditional filesystem: it's a // named reference to an inode. A dentry generally lives as long as it's part of -// a mounted filesystem tree. Kernfs doesn't cache dentries once all references -// to them are removed. Dentries hold a single reference to the inode they point +// a mounted filesystem tree. Kernfs drops dentries once all references to them +// are dropped. Dentries hold a single reference to the inode they point // to, and child dentries hold a reference on their parent. // // Must be initialized by Init prior to first use. +// +// +stateify savable type Dentry struct { vfsd vfs.Dentry - refs.AtomicRefCount + // refs is the reference count. When refs reaches 0, the dentry may be + // added to the cache or destroyed. If refs == -1, the dentry has already + // been destroyed. refs are allowed to go to 0 and increase again. refs is + // accessed using atomic memory operations. + refs int64 + + // fs is the owning filesystem. fs is immutable. + fs *Filesystem // flags caches useful information about the dentry from the inode. See the // dflags* consts above. Must be accessed by atomic ops. @@ -172,21 +193,177 @@ type Dentry struct { parent *Dentry name string + // If cached is true, dentryEntry links dentry into + // Filesystem.cachedDentries. cached and dentryEntry are protected by + // Filesystem.mu. + cached bool + dentryEntry + // dirMu protects children and the names of child Dentries. - dirMu sync.Mutex + // + // Note that holding fs.mu for writing is not sufficient; + // revalidateChildLocked(), which is a very hot path, may modify children with + // fs.mu acquired for reading only. + dirMu sync.Mutex `state:"nosave"` children map[string]*Dentry inode Inode } +// IncRef implements vfs.DentryImpl.IncRef. +func (d *Dentry) IncRef() { + // d.refs may be 0 if d.fs.mu is locked, which serializes against + // d.cacheLocked(). + atomic.AddInt64(&d.refs, 1) +} + +// TryIncRef implements vfs.DentryImpl.TryIncRef. +func (d *Dentry) TryIncRef() bool { + for { + refs := atomic.LoadInt64(&d.refs) + if refs <= 0 { + return false + } + if atomic.CompareAndSwapInt64(&d.refs, refs, refs+1) { + return true + } + } +} + +// DecRef implements vfs.DentryImpl.DecRef. +func (d *Dentry) DecRef(ctx context.Context) { + if refs := atomic.AddInt64(&d.refs, -1); refs == 0 { + d.fs.mu.Lock() + d.cacheLocked(ctx) + d.fs.mu.Unlock() + } else if refs < 0 { + panic("kernfs.Dentry.DecRef() called without holding a reference") + } +} + +// cacheLocked should be called after d's reference count becomes 0. The ref +// count check may happen before acquiring d.fs.mu so there might be a race +// condition where the ref count is increased again by the time the caller +// acquires d.fs.mu. This race is handled. +// Only reachable dentries are added to the cache. However, a dentry might +// become unreachable *while* it is in the cache due to invalidation. +// +// Preconditions: d.fs.mu must be locked for writing. +func (d *Dentry) cacheLocked(ctx context.Context) { + // Dentries with a non-zero reference count must be retained. (The only way + // to obtain a reference on a dentry with zero references is via path + // resolution, which requires d.fs.mu, so if d.refs is zero then it will + // remain zero while we hold d.fs.mu for writing.) + refs := atomic.LoadInt64(&d.refs) + if refs == -1 { + // Dentry has already been destroyed. + panic(fmt.Sprintf("cacheLocked called on a dentry which has already been destroyed: %v", d)) + } + if refs > 0 { + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + return + } + // If the dentry is deleted and invalidated or has no parent, then it is no + // longer reachable by path resolution and should be dropped immediately + // because it has zero references. + // Note that a dentry may not always have a parent; for example magic links + // as described in Inode.Getlink. + if isDead := d.VFSDentry().IsDead(); isDead || d.parent == nil { + if !isDead { + d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, d.VFSDentry()) + } + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentriesLen-- + d.cached = false + } + d.destroyLocked(ctx) + return + } + // If d is already cached, just move it to the front of the LRU. + if d.cached { + d.fs.cachedDentries.Remove(d) + d.fs.cachedDentries.PushFront(d) + return + } + // Cache the dentry, then evict the least recently used cached dentry if + // the cache becomes over-full. + d.fs.cachedDentries.PushFront(d) + d.fs.cachedDentriesLen++ + d.cached = true + if d.fs.cachedDentriesLen <= d.fs.MaxCachedDentries { + return + } + // Evict the least recently used dentry because cache size is greater than + // max cache size (configured on mount). + victim := d.fs.cachedDentries.Back() + d.fs.cachedDentries.Remove(victim) + d.fs.cachedDentriesLen-- + victim.cached = false + // victim.refs may have become non-zero from an earlier path resolution + // after it was inserted into fs.cachedDentries. + if atomic.LoadInt64(&victim.refs) == 0 { + if !victim.vfsd.IsDead() { + victim.parent.dirMu.Lock() + // Note that victim can't be a mount point (in any mount + // namespace), since VFS holds references on mount points. + d.fs.vfsfs.VirtualFilesystem().InvalidateDentry(ctx, victim.VFSDentry()) + delete(victim.parent.children, victim.name) + victim.parent.dirMu.Unlock() + } + victim.destroyLocked(ctx) + } + // Whether or not victim was destroyed, we brought fs.cachedDentriesLen + // back down to fs.MaxCachedDentries, so we don't loop. +} + +// destroyLocked destroys the dentry. +// +// Preconditions: +// * d.fs.mu must be locked for writing. +// * d.refs == 0. +// * d should have been removed from d.parent.children, i.e. d is not reachable +// by path traversal. +// * d.vfsd.IsDead() is true. +func (d *Dentry) destroyLocked(ctx context.Context) { + switch atomic.LoadInt64(&d.refs) { + case 0: + // Mark the dentry destroyed. + atomic.StoreInt64(&d.refs, -1) + case -1: + panic("dentry.destroyLocked() called on already destroyed dentry") + default: + panic("dentry.destroyLocked() called with references on the dentry") + } + + d.inode.DecRef(ctx) // IncRef from Init. + d.inode = nil + + // Drop the reference held by d on its parent without recursively locking + // d.fs.mu. + if d.parent != nil { + if refs := atomic.AddInt64(&d.parent.refs, -1); refs == 0 { + d.parent.cacheLocked(ctx) + } else if refs < 0 { + panic("kernfs.Dentry.DecRef() called without holding a reference") + } + } +} + // Init initializes this dentry. // // Precondition: Caller must hold a reference on inode. // // Postcondition: Caller's reference on inode is transferred to the dentry. -func (d *Dentry) Init(inode Inode) { +func (d *Dentry) Init(fs *Filesystem, inode Inode) { d.vfsd.Init(d) + d.fs = fs d.inode = inode + atomic.StoreInt64(&d.refs, 1) ftype := inode.Mode().FileType() if ftype == linux.ModeDirectory { d.flags |= dflagsIsDir @@ -211,20 +388,6 @@ func (d *Dentry) isSymlink() bool { return atomic.LoadUint32(&d.flags)&dflagsIsSymlink != 0 } -// DecRef implements vfs.DentryImpl.DecRef. -func (d *Dentry) DecRef(ctx context.Context) { - d.AtomicRefCount.DecRefWithDestructor(ctx, d.destroy) -} - -// Precondition: Dentry must be removed from VFS' dentry cache. -func (d *Dentry) destroy(ctx context.Context) { - d.inode.DecRef(ctx) // IncRef from Init. - d.inode = nil - if d.parent != nil { - d.parent.DecRef(ctx) // IncRef from Dentry.InsertChild. - } -} - // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. // // Although Linux technically supports inotify on pseudo filesystems (inotify @@ -240,25 +403,29 @@ func (d *Dentry) Watches() *vfs.Watches { // OnZeroWatches implements vfs.Dentry.OnZeroWatches. func (d *Dentry) OnZeroWatches(context.Context) {} -// InsertChild inserts child into the vfs dentry cache with the given name under -// this dentry. This does not update the directory inode, so calling this on -// its own isn't sufficient to insert a child into a directory. InsertChild -// updates the link count on d if required. +// insertChild inserts child into the vfs dentry cache with the given name under +// this dentry. This does not update the directory inode, so calling this on its +// own isn't sufficient to insert a child into a directory. // -// Precondition: d must represent a directory inode. -func (d *Dentry) InsertChild(name string, child *Dentry) { +// Preconditions: +// * d must represent a directory inode. +// * d.fs.mu must be locked for at least reading. +func (d *Dentry) insertChild(name string, child *Dentry) { d.dirMu.Lock() d.insertChildLocked(name, child) d.dirMu.Unlock() } -// insertChildLocked is equivalent to InsertChild, with additional +// insertChildLocked is equivalent to insertChild, with additional // preconditions. // -// Precondition: d.dirMu must be locked. +// Preconditions: +// * d must represent a directory inode. +// * d.dirMu must be locked. +// * d.fs.mu must be locked for at least reading. func (d *Dentry) insertChildLocked(name string, child *Dentry) { if !d.isDir() { - panic(fmt.Sprintf("InsertChild called on non-directory Dentry: %+v.", d)) + panic(fmt.Sprintf("insertChildLocked called on non-directory Dentry: %+v.", d)) } d.IncRef() // DecRef in child's Dentry.destroy. child.parent = d @@ -289,7 +456,6 @@ func (d *Dentry) Inode() Inode { // // - Checking that dentries passed to methods are of the appropriate file type. // - Checking permissions. -// - Updating link and reference counts. // // Specific responsibilities of implementations are documented below. type Inode interface { @@ -299,7 +465,8 @@ type Inode interface { inodeRefs // Methods related to node metadata. A generic implementation is provided by - // InodeAttrs. + // InodeAttrs. Note that a concrete filesystem using kernfs is responsible for + // managing link counts. inodeMetadata // Method for inodes that represent symlink. InodeNotSymlink provides a @@ -310,18 +477,26 @@ type Inode interface { // a blanket implementation for all non-directory inodes. inodeDirectory - // Method for inodes that represent dynamic directories and their - // children. InodeNoDynamicLookup provides a blanket implementation for all - // non-dynamic-directory inodes. - inodeDynamicLookup - // Open creates a file description for the filesystem object represented by // this inode. The returned file description should hold a reference on the - // inode for its lifetime. + // dentry for its lifetime. // // Precondition: rp.Done(). vfsd.Impl() must be the kernfs Dentry containing // the inode on which Open() is being called. - Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) + Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) + + // StatFS returns filesystem statistics for the client filesystem. This + // corresponds to vfs.FilesystemImpl.StatFSAt. If the client filesystem + // doesn't support statfs(2), this should return ENOSYS. + StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, error) + + // Keep indicates whether the dentry created after Inode.Lookup should be + // kept in the kernfs dentry tree. + Keep() bool + + // Valid should return true if this inode is still valid, or needs to + // be resolved again by a call to Lookup. + Valid(ctx context.Context) bool } type inodeRefs interface { @@ -354,8 +529,8 @@ type inodeMetadata interface { // Precondition: All methods in this interface may only be called on directory // inodes. type inodeDirectory interface { - // The New{File,Dir,Node,Symlink} methods below should return a new inode - // hashed into this inode. + // The New{File,Dir,Node,Link,Symlink} methods below should return a new inode + // that will be hashed into the dentry tree. // // These inode constructors are inode-level operations rather than // filesystem-level operations to allow client filesystems to mix different @@ -366,75 +541,69 @@ type inodeDirectory interface { HasChildren() bool // NewFile creates a new regular file inode. - NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) + NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (Inode, error) // NewDir creates a new directory inode. - NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) + NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (Inode, error) // NewLink creates a new hardlink to a specified inode in this // directory. Implementations should create a new kernfs Dentry pointing to // target, and update target's link count. - NewLink(ctx context.Context, name string, target Inode) (*vfs.Dentry, error) + NewLink(ctx context.Context, name string, target Inode) (Inode, error) // NewSymlink creates a new symbolic link inode. - NewSymlink(ctx context.Context, name, target string) (*vfs.Dentry, error) + NewSymlink(ctx context.Context, name, target string) (Inode, error) // NewNode creates a new filesystem node for a mknod syscall. - NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (*vfs.Dentry, error) + NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (Inode, error) // Unlink removes a child dentry from this directory inode. - Unlink(ctx context.Context, name string, child *vfs.Dentry) error + Unlink(ctx context.Context, name string, child Inode) error // RmDir removes an empty child directory from this directory // inode. Implementations must update the parent directory's link count, // if required. Implementations are not responsible for checking that child // is a directory, checking for an empty directory. - RmDir(ctx context.Context, name string, child *vfs.Dentry) error + RmDir(ctx context.Context, name string, child Inode) error // Rename is called on the source directory containing an inode being // renamed. child should point to the resolved child in the source - // directory. If Rename replaces a dentry in the destination directory, it - // should return the replaced dentry or nil otherwise. + // directory. // // Precondition: Caller must serialize concurrent calls to Rename. - Rename(ctx context.Context, oldname, newname string, child, dstDir *vfs.Dentry) (replaced *vfs.Dentry, err error) -} + Rename(ctx context.Context, oldname, newname string, child, dstDir Inode) error -type inodeDynamicLookup interface { - // Lookup should return an appropriate dentry if name should resolve to a - // child of this dynamic directory inode. This gives the directory an - // opportunity on every lookup to resolve additional entries that aren't - // hashed into the directory. This is only called when the inode is a - // directory. If the inode is not a directory, or if the directory only - // contains a static set of children, the implementer can unconditionally - // return an appropriate error (ENOTDIR and ENOENT respectively). + // Lookup should return an appropriate inode if name should resolve to a + // child of this directory inode. This gives the directory an opportunity + // on every lookup to resolve additional entries. This is only called when + // the inode is a directory. // - // The child returned by Lookup will be hashed into the VFS dentry tree. Its - // lifetime can be controlled by the filesystem implementation with an - // appropriate implementation of Valid. + // The child returned by Lookup will be hashed into the VFS dentry tree, + // at least for the duration of the current FS operation. // - // Lookup returns the child with an extra reference and the caller owns this - // reference. - Lookup(ctx context.Context, name string) (*vfs.Dentry, error) - - // Valid should return true if this inode is still valid, or needs to - // be resolved again by a call to Lookup. - Valid(ctx context.Context) bool + // Lookup must return the child with an extra reference whose ownership is + // transferred to the dentry that is created to point to that inode. If + // Inode.Keep returns false, that new dentry will be dropped at the end of + // the current filesystem operation (before returning back to the VFS + // layer) if no other ref is picked on that dentry. If Inode.Keep returns + // true, then the dentry will be cached into the dentry tree until it is + // Unlink'd or RmDir'd. + Lookup(ctx context.Context, name string) (Inode, error) // IterDirents is used to iterate over dynamically created entries. It invokes - // cb on each entry in the directory represented by the FileDescription. + // cb on each entry in the directory represented by the Inode. // 'offset' is the offset for the entire IterDirents call, which may include // results from the caller (e.g. "." and ".."). 'relOffset' is the offset // inside the entries returned by this IterDirents invocation. In other words, // 'offset' should be used to calculate each vfs.Dirent.NextOff as well as // the return value, while 'relOffset' is the place to start iteration. - IterDirents(ctx context.Context, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) + IterDirents(ctx context.Context, mnt *vfs.Mount, callback vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) } type inodeSymlink interface { // Readlink returns the target of a symbolic link. If an inode is not a // symlink, the implementation should return EINVAL. - Readlink(ctx context.Context) (string, error) + Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) // Getlink returns the target of a symbolic link, as used by path // resolution: diff --git a/pkg/sentry/fsimpl/kernfs/kernfs_test.go b/pkg/sentry/fsimpl/kernfs/kernfs_test.go index c5d5afedf..2418eec44 100644 --- a/pkg/sentry/fsimpl/kernfs/kernfs_test.go +++ b/pkg/sentry/fsimpl/kernfs/kernfs_test.go @@ -36,7 +36,7 @@ const staticFileContent = "This is sample content for a static test file." // RootDentryFn is a generator function for creating the root dentry of a test // filesystem. See newTestSystem. -type RootDentryFn func(*auth.Credentials, *filesystem) *kernfs.Dentry +type RootDentryFn func(context.Context, *auth.Credentials, *filesystem) kernfs.Inode // newTestSystem sets up a minimal environment for running a test, including an // instance of a test filesystem. Tests can control the contents of the @@ -52,7 +52,7 @@ func newTestSystem(t *testing.T, rootFn RootDentryFn) *testutil.System { v.MustRegisterFilesystemType("testfs", &fsType{rootFn: rootFn}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mns, err := v.NewMountNamespace(ctx, creds, "", "testfs", &vfs.GetFilesystemOptions{}) + mns, err := v.NewMountNamespace(ctx, creds, "", "testfs", &vfs.MountOptions{}) if err != nil { t.Fatalf("Failed to create testfs root mount: %v", err) } @@ -72,14 +72,11 @@ type file struct { content string } -func (fs *filesystem) newFile(creds *auth.Credentials, content string) *kernfs.Dentry { +func (fs *filesystem) newFile(ctx context.Context, creds *auth.Credentials, content string) kernfs.Inode { f := &file{} f.content = content - f.DynamicBytesFile.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), f, 0777) - - d := &kernfs.Dentry{} - d.Init(f) - return d + f.DynamicBytesFile.Init(ctx, creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), f, 0777) + return f } func (f *file) Generate(ctx context.Context, buf *bytes.Buffer) error { @@ -96,100 +93,112 @@ func (*attrs) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.S } type readonlyDir struct { + readonlyDirRefs attrs - kernfs.InodeNotSymlink - kernfs.InodeNoDynamicLookup + kernfs.InodeAlwaysValid kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNoStatFS + kernfs.InodeNotSymlink + kernfs.InodeTemporary kernfs.OrderedChildren locks vfs.FileLocks - - dentry kernfs.Dentry } -func (fs *filesystem) newReadonlyDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { +func (fs *filesystem) newReadonlyDir(ctx context.Context, creds *auth.Credentials, mode linux.FileMode, contents map[string]kernfs.Inode) kernfs.Inode { dir := &readonlyDir{} - dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) + dir.attrs.Init(ctx, creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - dir.dentry.Init(dir) - - dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) - - return &dir.dentry + dir.EnableLeakCheck() + dir.IncLinks(dir.OrderedChildren.Populate(contents)) + return dir } -func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &d.locks, &opts) +func (d *readonlyDir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndStaticEntries, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } +func (d *readonlyDir) DecRef(ctx context.Context) { + d.readonlyDirRefs.DecRef(func() { d.Destroy(ctx) }) +} + type dir struct { + dirRefs attrs + kernfs.InodeAlwaysValid kernfs.InodeNotSymlink - kernfs.InodeNoDynamicLookup + kernfs.InodeNoStatFS + kernfs.InodeTemporary kernfs.OrderedChildren locks vfs.FileLocks - fs *filesystem - dentry kernfs.Dentry + fs *filesystem } -func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { +func (fs *filesystem) newDir(ctx context.Context, creds *auth.Credentials, mode linux.FileMode, contents map[string]kernfs.Inode) kernfs.Inode { dir := &dir{} dir.fs = fs - dir.attrs.Init(creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) + dir.attrs.Init(ctx, creds, 0 /* devMajor */, 0 /* devMinor */, fs.NextIno(), linux.ModeDirectory|mode) dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{Writable: true}) - dir.dentry.Init(dir) + dir.EnableLeakCheck() - dir.IncLinks(dir.OrderedChildren.Populate(&dir.dentry, contents)) - - return &dir.dentry + dir.IncLinks(dir.OrderedChildren.Populate(contents)) + return dir } -func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &d.locks, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndStaticEntries, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } -func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (*vfs.Dentry, error) { +func (d *dir) DecRef(ctx context.Context) { + d.dirRefs.DecRef(func() { d.Destroy(ctx) }) +} + +func (d *dir) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (kernfs.Inode, error) { creds := auth.CredentialsFromContext(ctx) - dir := d.fs.newDir(creds, opts.Mode, nil) - dirVFSD := dir.VFSDentry() - if err := d.OrderedChildren.Insert(name, dirVFSD); err != nil { + dir := d.fs.newDir(ctx, creds, opts.Mode, nil) + if err := d.OrderedChildren.Insert(name, dir); err != nil { dir.DecRef(ctx) return nil, err } + d.TouchCMtime(ctx) d.IncLinks(1) - return dirVFSD, nil + return dir, nil } -func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (*vfs.Dentry, error) { +func (d *dir) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (kernfs.Inode, error) { creds := auth.CredentialsFromContext(ctx) - f := d.fs.newFile(creds, "") - fVFSD := f.VFSDentry() - if err := d.OrderedChildren.Insert(name, fVFSD); err != nil { + f := d.fs.newFile(ctx, creds, "") + if err := d.OrderedChildren.Insert(name, f); err != nil { f.DecRef(ctx) return nil, err } - return fVFSD, nil + d.TouchCMtime(ctx) + return f, nil } -func (*dir) NewLink(context.Context, string, kernfs.Inode) (*vfs.Dentry, error) { +func (*dir) NewLink(context.Context, string, kernfs.Inode) (kernfs.Inode, error) { return nil, syserror.EPERM } -func (*dir) NewSymlink(context.Context, string, string) (*vfs.Dentry, error) { +func (*dir) NewSymlink(context.Context, string, string) (kernfs.Inode, error) { return nil, syserror.EPERM } -func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (*vfs.Dentry, error) { +func (*dir) NewNode(context.Context, string, vfs.MknodOptions) (kernfs.Inode, error) { return nil, syserror.EPERM } @@ -197,19 +206,23 @@ func (fsType) Name() string { return "kernfs" } +func (fsType) Release(ctx context.Context) {} + func (fst fsType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opt vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { fs := &filesystem{} fs.VFSFilesystem().Init(vfsObj, &fst, fs) - root := fst.rootFn(creds, fs) - return fs.VFSFilesystem(), root.VFSDentry(), nil + root := fst.rootFn(ctx, creds, fs) + var d kernfs.Dentry + d.Init(&fs.Filesystem, root) + return fs.VFSFilesystem(), d.VFSDentry(), nil } // -------------------- Remainder of the file are test cases -------------------- func TestBasic(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ - "file1": fs.newFile(creds, staticFileContent), + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, map[string]kernfs.Inode{ + "file1": fs.newFile(ctx, creds, staticFileContent), }) }) defer sys.Destroy() @@ -217,9 +230,9 @@ func TestBasic(t *testing.T) { } func TestMkdirGetDentry(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ - "dir1": fs.newDir(creds, 0755, nil), + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, map[string]kernfs.Inode{ + "dir1": fs.newDir(ctx, creds, 0755, nil), }) }) defer sys.Destroy() @@ -232,9 +245,9 @@ func TestMkdirGetDentry(t *testing.T) { } func TestReadStaticFile(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ - "file1": fs.newFile(creds, staticFileContent), + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, map[string]kernfs.Inode{ + "file1": fs.newFile(ctx, creds, staticFileContent), }) }) defer sys.Destroy() @@ -258,9 +271,9 @@ func TestReadStaticFile(t *testing.T) { } func TestCreateNewFileInStaticDir(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ - "dir1": fs.newDir(creds, 0755, nil), + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, map[string]kernfs.Inode{ + "dir1": fs.newDir(ctx, creds, 0755, nil), }) }) defer sys.Destroy() @@ -285,8 +298,8 @@ func TestCreateNewFileInStaticDir(t *testing.T) { } func TestDirFDReadWrite(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, nil) + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, nil) }) defer sys.Destroy() @@ -309,14 +322,14 @@ func TestDirFDReadWrite(t *testing.T) { } func TestDirFDIterDirents(t *testing.T) { - sys := newTestSystem(t, func(creds *auth.Credentials, fs *filesystem) *kernfs.Dentry { - return fs.newReadonlyDir(creds, 0755, map[string]*kernfs.Dentry{ + sys := newTestSystem(t, func(ctx context.Context, creds *auth.Credentials, fs *filesystem) kernfs.Inode { + return fs.newReadonlyDir(ctx, creds, 0755, map[string]kernfs.Inode{ // Fill root with nodes backed by various inode implementations. - "dir1": fs.newReadonlyDir(creds, 0755, nil), - "dir2": fs.newDir(creds, 0755, map[string]*kernfs.Dentry{ - "dir3": fs.newDir(creds, 0755, nil), + "dir1": fs.newReadonlyDir(ctx, creds, 0755, nil), + "dir2": fs.newDir(ctx, creds, 0755, map[string]kernfs.Inode{ + "dir3": fs.newDir(ctx, creds, 0755, nil), }), - "file1": fs.newFile(creds, staticFileContent), + "file1": fs.newFile(ctx, creds, staticFileContent), }) }) defer sys.Destroy() diff --git a/pkg/sentry/fsimpl/kernfs/symlink.go b/pkg/sentry/fsimpl/kernfs/symlink.go index 2ab3f53fd..a0736c0d6 100644 --- a/pkg/sentry/fsimpl/kernfs/symlink.go +++ b/pkg/sentry/fsimpl/kernfs/symlink.go @@ -24,10 +24,13 @@ import ( // StaticSymlink provides an Inode implementation for symlinks that point to // a immutable target. +// +// +stateify savable type StaticSymlink struct { InodeAttrs InodeNoopRefCount InodeSymlink + InodeNoStatFS target string } @@ -35,23 +38,20 @@ type StaticSymlink struct { var _ Inode = (*StaticSymlink)(nil) // NewStaticSymlink creates a new symlink file pointing to 'target'. -func NewStaticSymlink(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, target string) *Dentry { +func NewStaticSymlink(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, target string) Inode { inode := &StaticSymlink{} - inode.Init(creds, devMajor, devMinor, ino, target) - - d := &Dentry{} - d.Init(inode) - return d + inode.Init(ctx, creds, devMajor, devMinor, ino, target) + return inode } // Init initializes the instance. -func (s *StaticSymlink) Init(creds *auth.Credentials, devMajor uint32, devMinor uint32, ino uint64, target string) { +func (s *StaticSymlink) Init(ctx context.Context, creds *auth.Credentials, devMajor uint32, devMinor uint32, ino uint64, target string) { s.target = target - s.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeSymlink|0777) + s.InodeAttrs.Init(ctx, creds, devMajor, devMinor, ino, linux.ModeSymlink|0777) } -// Readlink implements Inode. -func (s *StaticSymlink) Readlink(_ context.Context) (string, error) { +// Readlink implements Inode.Readlink. +func (s *StaticSymlink) Readlink(_ context.Context, _ *vfs.Mount) (string, error) { return s.target, nil } diff --git a/pkg/sentry/fsimpl/kernfs/synthetic_directory.go b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go new file mode 100644 index 000000000..463d77d79 --- /dev/null +++ b/pkg/sentry/fsimpl/kernfs/synthetic_directory.go @@ -0,0 +1,113 @@ +// Copyright 2019 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package kernfs + +import ( + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" +) + +// syntheticDirectory implements kernfs.Inode for a directory created by +// MkdirAt(ForSyntheticMountpoint=true). +// +// +stateify savable +type syntheticDirectory struct { + InodeAlwaysValid + InodeAttrs + InodeNoStatFS + InodeNotSymlink + OrderedChildren + syntheticDirectoryRefs + + locks vfs.FileLocks +} + +var _ Inode = (*syntheticDirectory)(nil) + +func newSyntheticDirectory(ctx context.Context, creds *auth.Credentials, perm linux.FileMode) Inode { + inode := &syntheticDirectory{} + inode.Init(ctx, creds, 0 /* devMajor */, 0 /* devMinor */, 0 /* ino */, perm) + return inode +} + +func (dir *syntheticDirectory) Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { + if perm&^linux.PermissionsMask != 0 { + panic(fmt.Sprintf("perm contains non-permission bits: %#o", perm)) + } + dir.InodeAttrs.Init(ctx, creds, devMajor, devMinor, ino, linux.S_IFDIR|perm) + dir.OrderedChildren.Init(OrderedChildrenOptions{ + Writable: true, + }) +} + +// Open implements Inode.Open. +func (dir *syntheticDirectory) Open(ctx context.Context, rp *vfs.ResolvingPath, d *Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := NewGenericDirectoryFD(rp.Mount(), d, &dir.OrderedChildren, &dir.locks, &opts, GenericDirectoryFDOptions{}) + if err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// NewFile implements Inode.NewFile. +func (dir *syntheticDirectory) NewFile(ctx context.Context, name string, opts vfs.OpenOptions) (Inode, error) { + return nil, syserror.EPERM +} + +// NewDir implements Inode.NewDir. +func (dir *syntheticDirectory) NewDir(ctx context.Context, name string, opts vfs.MkdirOptions) (Inode, error) { + if !opts.ForSyntheticMountpoint { + return nil, syserror.EPERM + } + subdirI := newSyntheticDirectory(ctx, auth.CredentialsFromContext(ctx), opts.Mode&linux.PermissionsMask) + if err := dir.OrderedChildren.Insert(name, subdirI); err != nil { + subdirI.DecRef(ctx) + return nil, err + } + dir.TouchCMtime(ctx) + return subdirI, nil +} + +// NewLink implements Inode.NewLink. +func (dir *syntheticDirectory) NewLink(ctx context.Context, name string, target Inode) (Inode, error) { + return nil, syserror.EPERM +} + +// NewSymlink implements Inode.NewSymlink. +func (dir *syntheticDirectory) NewSymlink(ctx context.Context, name, target string) (Inode, error) { + return nil, syserror.EPERM +} + +// NewNode implements Inode.NewNode. +func (dir *syntheticDirectory) NewNode(ctx context.Context, name string, opts vfs.MknodOptions) (Inode, error) { + return nil, syserror.EPERM +} + +// DecRef implements Inode.DecRef. +func (dir *syntheticDirectory) DecRef(ctx context.Context) { + dir.syntheticDirectoryRefs.DecRef(func() { dir.Destroy(ctx) }) +} + +// Keep implements Inode.Keep. This is redundant because inodes will never be +// created via Lookup and inodes are always valid. Makes sense to return true +// because these inodes are not temporary and should only be removed on RmDir. +func (dir *syntheticDirectory) Keep() bool { + return true +} diff --git a/pkg/sentry/fsimpl/overlay/BUILD b/pkg/sentry/fsimpl/overlay/BUILD index 8cf5b35d3..fd6c55921 100644 --- a/pkg/sentry/fsimpl/overlay/BUILD +++ b/pkg/sentry/fsimpl/overlay/BUILD @@ -21,14 +21,18 @@ go_library( "directory.go", "filesystem.go", "fstree.go", - "non_directory.go", "overlay.go", + "regular_file.go", + "save_restore.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", "//pkg/fspath", + "//pkg/log", + "//pkg/refsvfs2", + "//pkg/sentry/arch", "//pkg/sentry/fs/lock", "//pkg/sentry/kernel/auth", "//pkg/sentry/memmap", @@ -37,5 +41,6 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/usermem", + "//pkg/waiter", ], ) diff --git a/pkg/sentry/fsimpl/overlay/copy_up.go b/pkg/sentry/fsimpl/overlay/copy_up.go index b3d19ff82..4506642ca 100644 --- a/pkg/sentry/fsimpl/overlay/copy_up.go +++ b/pkg/sentry/fsimpl/overlay/copy_up.go @@ -22,6 +22,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -40,6 +42,10 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { return nil } + // Attach our credentials to the context, as some VFS operations use + // credentials from context rather an take an explicit creds parameter. + ctx = auth.ContextWithCredentials(ctx, d.fs.creds) + ftype := atomic.LoadUint32(&d.mode) & linux.S_IFMT switch ftype { case linux.S_IFREG, linux.S_IFDIR, linux.S_IFLNK, linux.S_IFBLK, linux.S_IFCHR: @@ -69,13 +75,28 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { return syserror.ENOENT } - // Perform copy-up. + // Obtain settable timestamps from the lower layer. vfsObj := d.fs.vfsfs.VirtualFilesystem() + oldpop := vfs.PathOperation{ + Root: d.lowerVDs[0], + Start: d.lowerVDs[0], + } + const timestampsMask = linux.STATX_ATIME | linux.STATX_MTIME + oldStat, err := vfsObj.StatAt(ctx, d.fs.creds, &oldpop, &vfs.StatOptions{ + Mask: timestampsMask, + }) + if err != nil { + return err + } + + // Perform copy-up. newpop := vfs.PathOperation{ Root: d.parent.upperVD, Start: d.parent.upperVD, Path: fspath.Parse(d.name), } + // Used during copy-up of memory-mapped regular files. + var mmapOpts *memmap.MMapOpts cleanupUndoCopyUp := func() { var err error if ftype == linux.S_IFDIR { @@ -84,15 +105,16 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { err = vfsObj.UnlinkAt(ctx, d.fs.creds, &newpop) } if err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after copy-up error: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after copy-up error: %v", err)) + } + if d.upperVD.Ok() { + d.upperVD.DecRef(ctx) + d.upperVD = vfs.VirtualDentry{} } } switch ftype { case linux.S_IFREG: - oldFD, err := vfsObj.OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ - Root: d.lowerVDs[0], - Start: d.lowerVDs[0], - }, &vfs.OpenOptions{ + oldFD, err := vfsObj.OpenAt(ctx, d.fs.creds, &oldpop, &vfs.OpenOptions{ Flags: linux.O_RDONLY, }) if err != nil { @@ -127,11 +149,32 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { break } } + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + if d.wrappedMappable != nil { + // We may have memory mappings of the file on the lower layer. + // Switch to mapping the file on the upper layer instead. + mmapOpts = &memmap.MMapOpts{ + Perms: usermem.ReadWrite, + MaxPerms: usermem.ReadWrite, + } + if err := newFD.ConfigureMMap(ctx, mmapOpts); err != nil { + cleanupUndoCopyUp() + return err + } + if mmapOpts.MappingIdentity != nil { + mmapOpts.MappingIdentity.DecRef(ctx) + } + // Don't actually switch Mappables until the end of copy-up; see + // below for why. + } if err := newFD.SetStat(ctx, vfs.SetStatOptions{ Stat: linux.Statx{ - Mask: linux.STATX_UID | linux.STATX_GID, - UID: d.uid, - GID: d.gid, + Mask: linux.STATX_UID | linux.STATX_GID | oldStat.Mask×tampsMask, + UID: d.uid, + GID: d.gid, + Atime: oldStat.Atime, + Mtime: oldStat.Mtime, }, }); err != nil { cleanupUndoCopyUp() @@ -148,9 +191,11 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { } if err := vfsObj.SetStatAt(ctx, d.fs.creds, &newpop, &vfs.SetStatOptions{ Stat: linux.Statx{ - Mask: linux.STATX_UID | linux.STATX_GID, - UID: d.uid, - GID: d.gid, + Mask: linux.STATX_UID | linux.STATX_GID | oldStat.Mask×tampsMask, + UID: d.uid, + GID: d.gid, + Atime: oldStat.Atime, + Mtime: oldStat.Mtime, }, }); err != nil { cleanupUndoCopyUp() @@ -164,10 +209,7 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { d.upperVD = upperVD case linux.S_IFLNK: - target, err := vfsObj.ReadlinkAt(ctx, d.fs.creds, &vfs.PathOperation{ - Root: d.lowerVDs[0], - Start: d.lowerVDs[0], - }) + target, err := vfsObj.ReadlinkAt(ctx, d.fs.creds, &oldpop) if err != nil { return err } @@ -176,10 +218,12 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { } if err := vfsObj.SetStatAt(ctx, d.fs.creds, &newpop, &vfs.SetStatOptions{ Stat: linux.Statx{ - Mask: linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID, - Mode: uint16(d.mode), - UID: d.uid, - GID: d.gid, + Mask: linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | oldStat.Mask×tampsMask, + Mode: uint16(d.mode), + UID: d.uid, + GID: d.gid, + Atime: oldStat.Atime, + Mtime: oldStat.Mtime, }, }); err != nil { cleanupUndoCopyUp() @@ -193,25 +237,20 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { d.upperVD = upperVD case linux.S_IFBLK, linux.S_IFCHR: - lowerStat, err := vfsObj.StatAt(ctx, d.fs.creds, &vfs.PathOperation{ - Root: d.lowerVDs[0], - Start: d.lowerVDs[0], - }, &vfs.StatOptions{}) - if err != nil { - return err - } if err := vfsObj.MknodAt(ctx, d.fs.creds, &newpop, &vfs.MknodOptions{ Mode: linux.FileMode(d.mode), - DevMajor: lowerStat.RdevMajor, - DevMinor: lowerStat.RdevMinor, + DevMajor: oldStat.RdevMajor, + DevMinor: oldStat.RdevMinor, }); err != nil { return err } if err := vfsObj.SetStatAt(ctx, d.fs.creds, &newpop, &vfs.SetStatOptions{ Stat: linux.Statx{ - Mask: linux.STATX_UID | linux.STATX_GID, - UID: d.uid, - GID: d.gid, + Mask: linux.STATX_UID | linux.STATX_GID | oldStat.Mask×tampsMask, + UID: d.uid, + GID: d.gid, + Atime: oldStat.Atime, + Mtime: oldStat.Mtime, }, }); err != nil { cleanupUndoCopyUp() @@ -229,7 +268,10 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { panic(fmt.Sprintf("unexpected file type %o", ftype)) } - // TODO(gvisor.dev/issue/1199): copy up xattrs + if err := d.copyXattrsLocked(ctx); err != nil { + cleanupUndoCopyUp() + return err + } // Update the dentry's device and inode numbers (except for directories, // for which these remain overlay-assigned). @@ -241,14 +283,10 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { Mask: linux.STATX_INO, }) if err != nil { - d.upperVD.DecRef(ctx) - d.upperVD = vfs.VirtualDentry{} cleanupUndoCopyUp() return err } if upperStat.Mask&linux.STATX_INO == 0 { - d.upperVD.DecRef(ctx) - d.upperVD = vfs.VirtualDentry{} cleanupUndoCopyUp() return syserror.EREMOTE } @@ -257,6 +295,135 @@ func (d *dentry) copyUpLocked(ctx context.Context) error { atomic.StoreUint64(&d.ino, upperStat.Ino) } + if mmapOpts != nil && mmapOpts.Mappable != nil { + // Note that if mmapOpts != nil, then d.mapsMu is locked for writing + // (from the S_IFREG path above). + + // Propagate mappings of d to the new Mappable. Remember which mappings + // we added so we can remove them on failure. + upperMappable := mmapOpts.Mappable + allAdded := make(map[memmap.MappableRange]memmap.MappingsOfRange) + for seg := d.lowerMappings.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(), m.Writable); err != nil { + for m := range added { + upperMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, seg.Start(), m.Writable) + } + for mr, mappings := range allAdded { + for m := range mappings { + upperMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, mr.Start, m.Writable) + } + } + return err + } + added[m] = struct{}{} + } + allAdded[seg.Range()] = added + } + + // Switch to the new Mappable. We do this at the end of copy-up + // because: + // + // - We need to switch Mappables (by changing d.wrappedMappable) before + // invalidating Translations from the old Mappable (to pick up + // Translations from the new one). + // + // - We need to lock d.dataMu while changing d.wrappedMappable, but + // must invalidate Translations with d.dataMu unlocked (due to lock + // ordering). + // + // - Consequently, once we unlock d.dataMu, other threads may + // immediately observe the new (copied-up) Mappable, which we want to + // delay until copy-up is guaranteed to succeed. + d.dataMu.Lock() + lowerMappable := d.wrappedMappable + d.wrappedMappable = upperMappable + d.dataMu.Unlock() + d.lowerMappings.InvalidateAll(memmap.InvalidateOpts{}) + + // Remove mappings from the old Mappable. + for seg := d.lowerMappings.FirstSegment(); seg.Ok(); seg = seg.NextSegment() { + for m := range seg.Value() { + lowerMappable.RemoveMapping(ctx, m.MappingSpace, m.AddrRange, seg.Start(), m.Writable) + } + } + d.lowerMappings.RemoveAll() + } + atomic.StoreUint32(&d.copiedUp, 1) return nil } + +// copyXattrsLocked copies a subset of lower's extended attributes to upper. +// Attributes that configure an overlay in the lower are not copied up. +// +// Preconditions: d.copyMu must be locked for writing. +func (d *dentry) copyXattrsLocked(ctx context.Context) error { + vfsObj := d.fs.vfsfs.VirtualFilesystem() + lowerPop := &vfs.PathOperation{Root: d.lowerVDs[0], Start: d.lowerVDs[0]} + upperPop := &vfs.PathOperation{Root: d.upperVD, Start: d.upperVD} + + lowerXattrs, err := vfsObj.ListXattrAt(ctx, d.fs.creds, lowerPop, 0) + if err != nil { + if err == syserror.EOPNOTSUPP { + // There are no guarantees as to the contents of lowerXattrs. + return nil + } + ctx.Infof("failed to copy up xattrs because ListXattrAt failed: %v", err) + return err + } + + for _, name := range lowerXattrs { + // Do not copy up overlay attributes. + if isOverlayXattr(name) { + continue + } + + value, err := vfsObj.GetXattrAt(ctx, d.fs.creds, lowerPop, &vfs.GetXattrOptions{Name: name, Size: 0}) + if err != nil { + ctx.Infof("failed to copy up xattrs because GetXattrAt failed: %v", err) + return err + } + + if err := vfsObj.SetXattrAt(ctx, d.fs.creds, upperPop, &vfs.SetXattrOptions{Name: name, Value: value}); err != nil { + ctx.Infof("failed to copy up xattrs because SetXattrAt failed: %v", err) + return err + } + } + return nil +} + +// copyUpDescendantsLocked ensures that all descendants of d are copied up. +// +// Preconditions: +// * filesystem.renameMu must be locked. +// * d.dirMu must be locked. +// * d.isDir(). +func (d *dentry) copyUpDescendantsLocked(ctx context.Context, ds **[]*dentry) error { + dirents, err := d.getDirentsLocked(ctx) + if err != nil { + return err + } + for _, dirent := range dirents { + if dirent.Name == "." || dirent.Name == ".." { + continue + } + child, err := d.fs.getChildLocked(ctx, d, dirent.Name, ds) + if err != nil { + return err + } + if err := child.copyUpLocked(ctx); err != nil { + return err + } + if child.isDir() { + child.dirMu.Lock() + err := child.copyUpDescendantsLocked(ctx, ds) + child.dirMu.Unlock() + if err != nil { + return err + } + } + } + return nil +} diff --git a/pkg/sentry/fsimpl/overlay/directory.go b/pkg/sentry/fsimpl/overlay/directory.go index 6a79f7ffe..df4492346 100644 --- a/pkg/sentry/fsimpl/overlay/directory.go +++ b/pkg/sentry/fsimpl/overlay/directory.go @@ -29,7 +29,9 @@ func (d *dentry) isDir() bool { return atomic.LoadUint32(&d.mode)&linux.S_IFMT == linux.S_IFDIR } -// Preconditions: d.dirMu must be locked. d.isDir(). +// Preconditions: +// * d.dirMu must be locked. +// * d.isDir(). func (d *dentry) collectWhiteoutsForRmdirLocked(ctx context.Context) (map[string]bool, error) { vfsObj := d.fs.vfsfs.VirtualFilesystem() var readdirErr error @@ -98,12 +100,13 @@ func (d *dentry) collectWhiteoutsForRmdirLocked(ctx context.Context) (map[string return whiteouts, readdirErr } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl - mu sync.Mutex + mu sync.Mutex `state:"nosave"` off int64 dirents []vfs.Dirent } @@ -114,10 +117,12 @@ func (fd *directoryFD) Release(ctx context.Context) { // IterDirents implements vfs.FileDescriptionImpl.IterDirents. func (fd *directoryFD) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback) error { + d := fd.dentry() + defer d.InotifyWithParent(ctx, linux.IN_ACCESS, 0, vfs.PathEvent) + fd.mu.Lock() defer fd.mu.Unlock() - d := fd.dentry() if fd.dirents == nil { ds, err := d.getDirents(ctx) if err != nil { @@ -141,7 +146,14 @@ func (d *dentry) getDirents(ctx context.Context) ([]vfs.Dirent, error) { defer d.fs.renameMu.RUnlock() d.dirMu.Lock() defer d.dirMu.Unlock() + return d.getDirentsLocked(ctx) +} +// Preconditions: +// * filesystem.renameMu must be locked. +// * d.dirMu must be locked. +// * d.isDir(). +func (d *dentry) getDirentsLocked(ctx context.Context) ([]vfs.Dirent, error) { if d.dirents != nil { return d.dirents, nil } diff --git a/pkg/sentry/fsimpl/overlay/filesystem.go b/pkg/sentry/fsimpl/overlay/filesystem.go index 86d0164b4..10161a08d 100644 --- a/pkg/sentry/fsimpl/overlay/filesystem.go +++ b/pkg/sentry/fsimpl/overlay/filesystem.go @@ -15,6 +15,8 @@ package overlay import ( + "fmt" + "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" @@ -27,10 +29,15 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// _OVL_XATTR_PREFIX is an extended attribute key prefix to identify overlayfs +// attributes. +// Linux: fs/overlayfs/overlayfs.h:OVL_XATTR_PREFIX +const _OVL_XATTR_PREFIX = linux.XATTR_TRUSTED_PREFIX + "overlay." + // _OVL_XATTR_OPAQUE is an extended attribute key whose value is set to "y" for // opaque directories. // Linux: fs/overlayfs/overlayfs.h:OVL_XATTR_OPAQUE -const _OVL_XATTR_OPAQUE = "trusted.overlay.opaque" +const _OVL_XATTR_OPAQUE = _OVL_XATTR_PREFIX + "opaque" func isWhiteout(stat *linux.Statx) bool { return stat.Mode&linux.S_IFMT == linux.S_IFCHR && stat.RdevMajor == 0 && stat.RdevMinor == 0 @@ -110,8 +117,10 @@ func (fs *filesystem) renameMuUnlockAndCheckDrop(ctx context.Context, ds **[]*de // Dentries which may have a reference count of zero, and which therefore // should be dropped once traversal is complete, are appended to ds. // -// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. -// !rp.Done(). +// Preconditions: +// * fs.renameMu must be locked. +// * d.dirMu must be locked. +// * !rp.Done(). func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, mayFollowSymlinks bool, ds **[]*dentry) (*dentry, error) { if !d.isDir() { return nil, syserror.ENOTDIR @@ -159,7 +168,9 @@ afterSymlink: return child, nil } -// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +// Preconditions: +// * fs.renameMu must be locked. +// * d.dirMu must be locked. func (fs *filesystem) getChildLocked(ctx context.Context, parent *dentry, name string, ds **[]*dentry) (*dentry, error) { if child, ok := parent.children[name]; ok { return child, nil @@ -177,7 +188,9 @@ func (fs *filesystem) getChildLocked(ctx context.Context, parent *dentry, name s return child, nil } -// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. +// Preconditions: +// * fs.renameMu must be locked. +// * parent.dirMu must be locked. func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name string) (*dentry, error) { childPath := fspath.Parse(name) child := fs.newDentry() @@ -199,6 +212,7 @@ func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name str lookupErr = err return false } + defer childVD.DecRef(ctx) mask := uint32(linux.STATX_TYPE) if !existsOnAnyLayer { @@ -237,6 +251,7 @@ func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name str } // Update child to include this layer. + childVD.IncRef() if isUpper { child.upperVD = childVD child.copiedUp = 1 @@ -261,10 +276,10 @@ func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name str // Directories are merged with directories from lower layers if they // are not explicitly opaque. - opaqueVal, err := vfsObj.GetxattrAt(ctx, fs.creds, &vfs.PathOperation{ + opaqueVal, err := vfsObj.GetXattrAt(ctx, fs.creds, &vfs.PathOperation{ Root: childVD, Start: childVD, - }, &vfs.GetxattrOptions{ + }, &vfs.GetXattrOptions{ Name: _OVL_XATTR_OPAQUE, Size: 1, }) @@ -287,8 +302,14 @@ func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name str child.devMinor = fs.dirDevMinor child.ino = fs.newDirIno() } else if !child.upperVD.Ok() { + childDevMinor, err := fs.getLowerDevMinor(child.devMajor, child.devMinor) + if err != nil { + ctx.Infof("overlay.filesystem.lookupLocked: failed to map lower layer device number (%d, %d) to an overlay-specific device number: %v", child.devMajor, child.devMinor, err) + child.destroyLocked(ctx) + return nil, err + } child.devMajor = linux.UNNAMED_MAJOR - child.devMinor = fs.lowerDevMinors[child.lowerVDs[0].Mount().Filesystem()] + child.devMinor = childDevMinor } parent.IncRef() @@ -300,7 +321,9 @@ func (fs *filesystem) lookupLocked(ctx context.Context, parent *dentry, name str // lookupLayerLocked is similar to lookupLocked, but only returns information // about the file rather than a dentry. // -// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. +// Preconditions: +// * fs.renameMu must be locked. +// * parent.dirMu must be locked. func (fs *filesystem) lookupLayerLocked(ctx context.Context, parent *dentry, name string) (lookupLayer, error) { childPath := fspath.Parse(name) lookupLayer := lookupLayerNone @@ -385,7 +408,9 @@ func (ll lookupLayer) existsInOverlay() bool { // rp.Start().Impl().(*dentry)). It does not check that the returned directory // is searchable by the provider of rp. // -// Preconditions: fs.renameMu must be locked. !rp.Done(). +// Preconditions: +// * fs.renameMu must be locked. +// * !rp.Done(). func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { for !rp.Final() { d.dirMu.Lock() @@ -425,8 +450,9 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, // doCreateAt checks that creating a file at rp is permitted, then invokes // create to do so. // -// Preconditions: !rp.Done(). For the final path component in rp, -// !rp.ShouldFollowSymlink(). +// Preconditions: +// * !rp.Done(). +// * For the final path component in rp, !rp.ShouldFollowSymlink(). func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parent *dentry, name string, haveUpperWhiteout bool) error) error { var ds *[]*dentry fs.renameMu.RLock() @@ -479,7 +505,13 @@ func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir if err := create(parent, name, childLayer == lookupLayerUpperWhiteout); err != nil { return err } + parent.dirents = nil + ev := linux.IN_CREATE + if dir { + ev |= linux.IN_ISDIR + } + parent.watches.Notify(ctx, name, uint32(ev), 0 /* cookie */, vfs.InodeEvent, false /* unlinked */) return nil } @@ -493,7 +525,7 @@ func (fs *filesystem) createWhiteout(ctx context.Context, vfsObj *vfs.VirtualFil func (fs *filesystem) cleanupRecreateWhiteout(ctx context.Context, vfsObj *vfs.VirtualFilesystem, pop *vfs.PathOperation) { if err := fs.createWhiteout(ctx, vfsObj, pop); err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to recreate whiteout after failed file creation: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate whiteout after failed file creation: %v", err)) } } @@ -605,12 +637,13 @@ func (fs *filesystem) LinkAt(ctx context.Context, rp *vfs.ResolvingPath, vd vfs. }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &newpop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after LinkAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after LinkAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &newpop) } return err } + old.watches.Notify(ctx, "", linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent, false /* unlinked */) return nil }) } @@ -644,7 +677,7 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v }, }); err != nil { if cleanupErr := vfsObj.RmdirAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -654,12 +687,12 @@ func (fs *filesystem) MkdirAt(ctx context.Context, rp *vfs.ResolvingPath, opts v // There may be directories on lower layers (previously hidden by // the whiteout) that the new directory should not be merged with. // Mark it opaque to prevent merging. - if err := vfsObj.SetxattrAt(ctx, fs.creds, &pop, &vfs.SetxattrOptions{ + if err := vfsObj.SetXattrAt(ctx, fs.creds, &pop, &vfs.SetXattrOptions{ Name: _OVL_XATTR_OPAQUE, Value: "y", }); err != nil { if cleanupErr := vfsObj.RmdirAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt set-opaque failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer directory after MkdirAt set-opaque failure: %v", cleanupErr)) } else { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -703,7 +736,7 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after MknodAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after MknodAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -732,10 +765,13 @@ func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf start := rp.Start().Impl().(*dentry) if rp.Done() { + if mayCreate && rp.MustBeDir() { + return nil, syserror.EISDIR + } if mustCreate { return nil, syserror.EEXIST } - if mayWrite { + if start.isRegularFile() && mayWrite { if err := start.copyUpLocked(ctx); err != nil { return nil, err } @@ -755,6 +791,10 @@ afterTrailingSymlink: if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { return nil, err } + // Reject attempts to open directories with O_CREAT. + if mayCreate && rp.MustBeDir() { + return nil, syserror.EISDIR + } // Determine whether or not we need to create a file. parent.dirMu.Lock() child, err := fs.stepLocked(ctx, rp, parent, false /* mayFollowSymlinks */, &ds) @@ -763,12 +803,11 @@ afterTrailingSymlink: parent.dirMu.Unlock() return fd, err } + parent.dirMu.Unlock() if err != nil { - parent.dirMu.Unlock() return nil, err } // Open existing child or follow symlink. - parent.dirMu.Unlock() if mustCreate { return nil, syserror.EEXIST } @@ -783,7 +822,10 @@ afterTrailingSymlink: start = parent goto afterTrailingSymlink } - if mayWrite { + if rp.MustBeDir() && !child.isDir() { + return nil, syserror.ENOTDIR + } + if child.isRegularFile() && mayWrite { if err := child.copyUpLocked(ctx); err != nil { return nil, err } @@ -836,8 +878,11 @@ func (d *dentry) openCopiedUp(ctx context.Context, rp *vfs.ResolvingPath, opts * if err != nil { return nil, err } + if ftype != linux.S_IFREG { + return layerFD, nil + } layerFlags := layerFD.StatusFlags() - fd := &nonDirectoryFD{ + fd := ®ularFileFD{ copiedUp: isUpper, cachedFD: layerFD, cachedFlags: layerFlags, @@ -851,8 +896,9 @@ func (d *dentry) openCopiedUp(ctx context.Context, rp *vfs.ResolvingPath, opts * return &fd.vfsfd, nil } -// Preconditions: parent.dirMu must be locked. parent does not already contain -// a child named rp.Component(). +// Preconditions: +// * parent.dirMu must be locked. +// * parent does not already contain a child named rp.Component(). func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.ResolvingPath, parent *dentry, opts *vfs.OpenOptions, ds **[]*dentry) (*vfs.FileDescription, error) { creds := rp.Credentials() if err := parent.checkPermissions(creds, vfs.MayWrite); err != nil { @@ -913,7 +959,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -924,7 +970,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving child, err := fs.getChildLocked(ctx, parent, childName, ds) if err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) dentry lookup failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after OpenAt(O_CREAT) dentry lookup failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -932,7 +978,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving } // Finally construct the overlay FD. upperFlags := upperFD.StatusFlags() - fd := &nonDirectoryFD{ + fd := ®ularFileFD{ copiedUp: true, cachedFD: upperFD, cachedFlags: upperFlags, @@ -945,6 +991,7 @@ func (fs *filesystem) createAndOpenLocked(ctx context.Context, rp *vfs.Resolving // just can't open it anymore for some reason. return nil, err } + parent.watches.Notify(ctx, childName, linux.IN_CREATE, 0 /* cookie */, vfs.PathEvent, false /* unlinked */) return &fd.vfsfd, nil } @@ -990,9 +1037,224 @@ func (fs *filesystem) RenameAt(ctx context.Context, rp *vfs.ResolvingPath, oldPa } defer mnt.EndWrite() - // FIXME(gvisor.dev/issue/1199): Actually implement rename. - _ = newParent - return syserror.EXDEV + oldParent := oldParentVD.Dentry().Impl().(*dentry) + creds := rp.Credentials() + if err := oldParent.checkPermissions(creds, vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + // We need a dentry representing the renamed file since, if it's a + // directory, we need to check for write permission on it. + oldParent.dirMu.Lock() + defer oldParent.dirMu.Unlock() + renamed, err := fs.getChildLocked(ctx, oldParent, oldName, &ds) + if err != nil { + return err + } + if err := vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&oldParent.mode)), auth.KUID(atomic.LoadUint32(&renamed.uid))); err != nil { + return err + } + if renamed.isDir() { + if renamed == newParent || genericIsAncestorDentry(renamed, newParent) { + return syserror.EINVAL + } + if oldParent != newParent { + if err := renamed.checkPermissions(creds, vfs.MayWrite); err != nil { + return err + } + } + } else { + if opts.MustBeDir || rp.MustBeDir() { + return syserror.ENOTDIR + } + } + + if oldParent != newParent { + if err := newParent.checkPermissions(creds, vfs.MayWrite|vfs.MayExec); err != nil { + return err + } + newParent.dirMu.Lock() + defer newParent.dirMu.Unlock() + } + if newParent.vfsd.IsDead() { + return syserror.ENOENT + } + replacedLayer, err := fs.lookupLayerLocked(ctx, newParent, newName) + if err != nil { + return err + } + var ( + replaced *dentry + replacedVFSD *vfs.Dentry + whiteouts map[string]bool + ) + if replacedLayer.existsInOverlay() { + replaced, err = fs.getChildLocked(ctx, newParent, newName, &ds) + if err != nil { + return err + } + replacedVFSD = &replaced.vfsd + if replaced.isDir() { + if !renamed.isDir() { + return syserror.EISDIR + } + if genericIsAncestorDentry(replaced, renamed) { + return syserror.ENOTEMPTY + } + replaced.dirMu.Lock() + defer replaced.dirMu.Unlock() + whiteouts, err = replaced.collectWhiteoutsForRmdirLocked(ctx) + if err != nil { + return err + } + } else { + if rp.MustBeDir() || renamed.isDir() { + return syserror.ENOTDIR + } + } + } + + if oldParent == newParent && oldName == newName { + return nil + } + + // renamed and oldParent need to be copied-up before they're renamed on the + // upper layer. + if err := renamed.copyUpLocked(ctx); err != nil { + return err + } + // If renamed is a directory, all of its descendants need to be copied-up + // before they're renamed on the upper layer. + if renamed.isDir() { + if err := renamed.copyUpDescendantsLocked(ctx, &ds); err != nil { + return err + } + } + // newParent must be copied-up before it can contain renamed on the upper + // layer. + if err := newParent.copyUpLocked(ctx); err != nil { + return err + } + // If replaced exists, it doesn't need to be copied-up, but we do need to + // serialize with copy-up. Holding renameMu for writing should be + // sufficient, but out of an abundance of caution... + if replaced != nil { + replaced.copyMu.RLock() + defer replaced.copyMu.RUnlock() + } + + vfsObj := rp.VirtualFilesystem() + mntns := vfs.MountNamespaceFromContext(ctx) + defer mntns.DecRef(ctx) + if err := vfsObj.PrepareRenameDentry(mntns, &renamed.vfsd, replacedVFSD); err != nil { + return err + } + + newpop := vfs.PathOperation{ + Root: newParent.upperVD, + Start: newParent.upperVD, + Path: fspath.Parse(newName), + } + + needRecreateWhiteouts := false + cleanupRecreateWhiteouts := func() { + if !needRecreateWhiteouts { + return + } + for whiteoutName, whiteoutUpper := range whiteouts { + if !whiteoutUpper { + continue + } + if err := fs.createWhiteout(ctx, vfsObj, &vfs.PathOperation{ + Root: replaced.upperVD, + Start: replaced.upperVD, + Path: fspath.Parse(whiteoutName), + }); err != nil && err != syserror.EEXIST { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RenameAt failure: %v", err)) + } + } + } + if renamed.isDir() { + if replacedLayer == lookupLayerUpper { + // Remove whiteouts from the directory being replaced. + needRecreateWhiteouts = true + for whiteoutName, whiteoutUpper := range whiteouts { + if !whiteoutUpper { + continue + } + if err := vfsObj.UnlinkAt(ctx, fs.creds, &vfs.PathOperation{ + Root: replaced.upperVD, + Start: replaced.upperVD, + Path: fspath.Parse(whiteoutName), + }); err != nil { + cleanupRecreateWhiteouts() + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + } else if replacedLayer == lookupLayerUpperWhiteout { + // We need to explicitly remove the whiteout since otherwise rename + // on the upper layer will fail with ENOTDIR. + if err := vfsObj.UnlinkAt(ctx, fs.creds, &newpop); err != nil { + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + } + } + + // Essentially no gVisor filesystem supports RENAME_WHITEOUT, so just do a + // regular rename and create the whiteout at the origin manually. Unlike + // RENAME_WHITEOUT, this isn't atomic with respect to other users of the + // upper filesystem, but this is already the case for virtually all other + // overlay filesystem operations too. + oldpop := vfs.PathOperation{ + Root: oldParent.upperVD, + Start: oldParent.upperVD, + Path: fspath.Parse(oldName), + } + if err := vfsObj.RenameAt(ctx, creds, &oldpop, &newpop, &opts); err != nil { + cleanupRecreateWhiteouts() + vfsObj.AbortRenameDentry(&renamed.vfsd, replacedVFSD) + return err + } + + // Below this point, the renamed dentry is now at newpop, and anything we + // replaced is gone forever. Commit the rename, update the overlay + // filesystem tree, and abandon attempts to recover from errors. + vfsObj.CommitRenameReplaceDentry(ctx, &renamed.vfsd, replacedVFSD) + delete(oldParent.children, oldName) + if replaced != nil { + ds = appendDentry(ds, replaced) + } + if oldParent != newParent { + newParent.dirents = nil + // This can't drop the last reference on oldParent because one is held + // by oldParentVD, so lock recursion is impossible. + oldParent.DecRef(ctx) + ds = appendDentry(ds, oldParent) + newParent.IncRef() + renamed.parent = newParent + } + renamed.name = newName + if newParent.children == nil { + newParent.children = make(map[string]*dentry) + } + newParent.children[newName] = renamed + oldParent.dirents = nil + + if err := fs.createWhiteout(ctx, vfsObj, &oldpop); err != nil { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout at origin after RenameAt: %v", err)) + } + if renamed.isDir() { + if err := vfsObj.SetXattrAt(ctx, fs.creds, &newpop, &vfs.SetXattrOptions{ + Name: _OVL_XATTR_OPAQUE, + Value: "y", + }); err != nil { + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to make renamed directory opaque: %v", err)) + } + } + + vfs.InotifyRename(ctx, &renamed.watches, &oldParent.watches, &newParent.watches, oldName, newName, renamed.isDir()) + return nil } // RmdirAt implements vfs.FilesystemImpl.RmdirAt. @@ -1040,6 +1302,9 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error if !child.isDir() { return syserror.ENOTDIR } + if err := vfs.CheckDeleteSticky(rp.Credentials(), linux.FileMode(atomic.LoadUint32(&parent.mode)), auth.KUID(atomic.LoadUint32(&child.uid))); err != nil { + return err + } child.dirMu.Lock() defer child.dirMu.Unlock() whiteouts, err := child.collectWhiteoutsForRmdirLocked(ctx) @@ -1071,7 +1336,7 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error Start: child.upperVD, Path: fspath.Parse(whiteoutName), }); err != nil && err != syserror.EEXIST { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RmdirAt failure: %v", err) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to recreate deleted whiteout after RmdirAt failure: %v", err)) } } } @@ -1101,15 +1366,14 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error // Don't attempt to recover from this: the original directory is // already gone, so any dentries representing it are invalid, and // creating a new directory won't undo that. - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to create whiteout during RmdirAt: %v", err) - vfsObj.AbortDeleteDentry(&child.vfsd) - return err + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout during RmdirAt: %v", err)) } vfsObj.CommitDeleteDentry(ctx, &child.vfsd) delete(parent.children, name) ds = appendDentry(ds, child) parent.dirents = nil + parent.watches.Notify(ctx, name, linux.IN_DELETE|linux.IN_ISDIR, 0 /* cookie */, vfs.InodeEvent, true /* unlinked */) return nil } @@ -1117,12 +1381,25 @@ func (fs *filesystem) RmdirAt(ctx context.Context, rp *vfs.ResolvingPath) error func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) return err } + err = d.setStatLocked(ctx, rp, opts) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { + return err + } + + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ctx, ev, 0 /* cookie */, vfs.InodeEvent) + } + return nil +} +// Precondition: d.fs.renameMu must be held for reading. +func (d *dentry) setStatLocked(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetStatOptions) error { mode := linux.FileMode(atomic.LoadUint32(&d.mode)) if err := vfs.CheckSetStat(ctx, rp.Credentials(), &opts, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { return err @@ -1217,7 +1494,7 @@ func (fs *filesystem) SymlinkAt(ctx context.Context, rp *vfs.ResolvingPath, targ }, }); err != nil { if cleanupErr := vfsObj.UnlinkAt(ctx, fs.creds, &pop); cleanupErr != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to delete upper layer file after SymlinkAt metadata update failure: %v", cleanupErr) + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to delete upper layer file after SymlinkAt metadata update failure: %v", cleanupErr)) } else if haveUpperWhiteout { fs.cleanupRecreateWhiteout(ctx, vfsObj, &pop) } @@ -1263,12 +1540,38 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return err } + parentMode := atomic.LoadUint32(&parent.mode) child := parent.children[name] var childLayer lookupLayer + if child == nil { + if parentMode&linux.S_ISVTX != 0 { + // If the parent's sticky bit is set, we need a child dentry to get + // its owner. + child, err = fs.getChildLocked(ctx, parent, name, &ds) + if err != nil { + return err + } + } else { + // Determine if the file being unlinked actually exists. Holding + // parent.dirMu prevents a dentry from being instantiated for the file, + // which in turn prevents it from being copied-up, so this result is + // stable. + childLayer, err = fs.lookupLayerLocked(ctx, parent, name) + if err != nil { + return err + } + if !childLayer.existsInOverlay() { + return syserror.ENOENT + } + } + } if child != nil { if child.isDir() { return syserror.EISDIR } + if err := vfs.CheckDeleteSticky(rp.Credentials(), linux.FileMode(parentMode), auth.KUID(atomic.LoadUint32(&child.uid))); err != nil { + return err + } if err := vfsObj.PrepareDeleteDentry(mntns, &child.vfsd); err != nil { return err } @@ -1281,18 +1584,6 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } else { childLayer = lookupLayerLower } - } else { - // Determine if the file being unlinked actually exists. Holding - // parent.dirMu prevents a dentry from being instantiated for the file, - // which in turn prevents it from being copied-up, so this result is - // stable. - childLayer, err = fs.lookupLayerLocked(ctx, parent, name) - if err != nil { - return err - } - if !childLayer.existsInOverlay() { - return syserror.ENOENT - } } pop := vfs.PathOperation{ @@ -1310,70 +1601,175 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error } } if err := fs.createWhiteout(ctx, vfsObj, &pop); err != nil { - ctx.Warningf("Unrecoverable overlayfs inconsistency: failed to create whiteout during UnlinkAt: %v", err) - if child != nil { - vfsObj.AbortDeleteDentry(&child.vfsd) - } - return err + panic(fmt.Sprintf("unrecoverable overlayfs inconsistency: failed to create whiteout during UnlinkAt: %v", err)) } + var cw *vfs.Watches if child != nil { vfsObj.CommitDeleteDentry(ctx, &child.vfsd) delete(parent.children, name) ds = appendDentry(ds, child) + cw = &child.watches } + vfs.InotifyRemoveChild(ctx, cw, &parent.watches, name) parent.dirents = nil return nil } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// isOverlayXattr returns whether the given extended attribute configures the +// overlay. +func isOverlayXattr(name string) bool { + return strings.HasPrefix(name, _OVL_XATTR_PREFIX) +} + +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) - _, err := fs.resolveLocked(ctx, rp, &ds) + d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { return nil, err } - // TODO(gvisor.dev/issue/1199): Linux overlayfs actually allows listxattr, - // but not any other xattr syscalls. For now we just reject all of them. - return nil, syserror.ENOTSUP + + return fs.listXattr(ctx, d, size) } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +func (fs *filesystem) listXattr(ctx context.Context, d *dentry, size uint64) ([]string, error) { + vfsObj := d.fs.vfsfs.VirtualFilesystem() + top := d.topLayer() + names, err := vfsObj.ListXattrAt(ctx, fs.creds, &vfs.PathOperation{Root: top, Start: top}, size) + if err != nil { + return nil, err + } + + // Filter out all overlay attributes. + n := 0 + for _, name := range names { + if !isOverlayXattr(name) { + names[n] = name + n++ + } + } + return names[:n], err +} + +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) - _, err := fs.resolveLocked(ctx, rp, &ds) + d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { return "", err } - return "", syserror.ENOTSUP + + return fs.getXattr(ctx, d, rp.Credentials(), &opts) +} + +func (fs *filesystem) getXattr(ctx context.Context, d *dentry, creds *auth.Credentials, opts *vfs.GetXattrOptions) (string, error) { + if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayRead); err != nil { + return "", err + } + + // Return EOPNOTSUPP when fetching an overlay attribute. + // See fs/overlayfs/super.c:ovl_own_xattr_get(). + if isOverlayXattr(opts.Name) { + return "", syserror.EOPNOTSUPP + } + + // Analogous to fs/overlayfs/super.c:ovl_other_xattr_get(). + vfsObj := d.fs.vfsfs.VirtualFilesystem() + top := d.topLayer() + return vfsObj.GetXattrAt(ctx, fs.creds, &vfs.PathOperation{Root: top, Start: top}, opts) } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) - _, err := fs.resolveLocked(ctx, rp, &ds) + d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + return err + } + + err = fs.setXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), &opts) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent) + return nil +} + +// Precondition: fs.renameMu must be locked. +func (fs *filesystem) setXattrLocked(ctx context.Context, d *dentry, mnt *vfs.Mount, creds *auth.Credentials, opts *vfs.SetXattrOptions) error { + if err := d.checkXattrPermissions(creds, opts.Name, vfs.MayWrite); err != nil { + return err + } + + // Return EOPNOTSUPP when setting an overlay attribute. + // See fs/overlayfs/super.c:ovl_own_xattr_set(). + if isOverlayXattr(opts.Name) { + return syserror.EOPNOTSUPP + } + + // Analogous to fs/overlayfs/super.c:ovl_other_xattr_set(). + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + if err := d.copyUpLocked(ctx); err != nil { return err } - return syserror.ENOTSUP + vfsObj := d.fs.vfsfs.VirtualFilesystem() + return vfsObj.SetXattrAt(ctx, fs.creds, &vfs.PathOperation{Root: d.upperVD, Start: d.upperVD}, opts) } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { var ds *[]*dentry fs.renameMu.RLock() - defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) - _, err := fs.resolveLocked(ctx, rp, &ds) + d, err := fs.resolveLocked(ctx, rp, &ds) if err != nil { + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + return err + } + + err = fs.removeXattrLocked(ctx, d, rp.Mount(), rp.Credentials(), name) + fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0 /* cookie */, vfs.InodeEvent) + return nil +} + +// Precondition: fs.renameMu must be locked. +func (fs *filesystem) removeXattrLocked(ctx context.Context, d *dentry, mnt *vfs.Mount, creds *auth.Credentials, name string) error { + if err := d.checkXattrPermissions(creds, name, vfs.MayWrite); err != nil { + return err + } + + // Like SetXattrAt, return EOPNOTSUPP when removing an overlay attribute. + // Linux passes the remove request to xattr_handler->set. + // See fs/xattr.c:vfs_removexattr(). + if isOverlayXattr(name) { + return syserror.EOPNOTSUPP + } + + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + if err := d.copyUpLocked(ctx); err != nil { return err } - return syserror.ENOTSUP + vfsObj := d.fs.vfsfs.VirtualFilesystem() + return vfsObj.RemoveXattrAt(ctx, fs.creds, &vfs.PathOperation{Root: d.upperVD, Start: d.upperVD}, name) } // PrependPath implements vfs.FilesystemImpl.PrependPath. diff --git a/pkg/sentry/fsimpl/overlay/non_directory.go b/pkg/sentry/fsimpl/overlay/non_directory.go deleted file mode 100644 index d3060a481..000000000 --- a/pkg/sentry/fsimpl/overlay/non_directory.go +++ /dev/null @@ -1,266 +0,0 @@ -// Copyright 2020 The gVisor Authors. -// -// Licensed under the Apache License, Version 2.0 (the "License"); -// you may not use this file except in compliance with the License. -// You may obtain a copy of the License at -// -// http://www.apache.org/licenses/LICENSE-2.0 -// -// Unless required by applicable law or agreed to in writing, software -// distributed under the License is distributed on an "AS IS" BASIS, -// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. -// See the License for the specific language governing permissions and -// limitations under the License. - -package overlay - -import ( - "sync/atomic" - - "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/sentry/kernel/auth" - "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/vfs" - "gvisor.dev/gvisor/pkg/sync" - "gvisor.dev/gvisor/pkg/usermem" -) - -func (d *dentry) isSymlink() bool { - return atomic.LoadUint32(&d.mode)&linux.S_IFMT == linux.S_IFLNK -} - -func (d *dentry) readlink(ctx context.Context) (string, error) { - layerVD := d.topLayer() - return d.fs.vfsfs.VirtualFilesystem().ReadlinkAt(ctx, d.fs.creds, &vfs.PathOperation{ - Root: layerVD, - Start: layerVD, - }) -} - -type nonDirectoryFD struct { - fileDescription - - // If copiedUp is false, cachedFD represents - // fileDescription.dentry().lowerVDs[0]; otherwise, cachedFD represents - // fileDescription.dentry().upperVD. cachedFlags is the last known value of - // cachedFD.StatusFlags(). copiedUp, cachedFD, and cachedFlags are - // protected by mu. - mu sync.Mutex - copiedUp bool - cachedFD *vfs.FileDescription - cachedFlags uint32 -} - -func (fd *nonDirectoryFD) getCurrentFD(ctx context.Context) (*vfs.FileDescription, error) { - fd.mu.Lock() - defer fd.mu.Unlock() - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - return nil, err - } - wrappedFD.IncRef() - return wrappedFD, nil -} - -func (fd *nonDirectoryFD) currentFDLocked(ctx context.Context) (*vfs.FileDescription, error) { - d := fd.dentry() - statusFlags := fd.vfsfd.StatusFlags() - if !fd.copiedUp && d.isCopiedUp() { - // Switch to the copied-up file. - upperVD := d.topLayer() - upperFD, err := fd.filesystem().vfsfs.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ - Root: upperVD, - Start: upperVD, - }, &vfs.OpenOptions{ - Flags: statusFlags, - }) - if err != nil { - return nil, err - } - oldOff, oldOffErr := fd.cachedFD.Seek(ctx, 0, linux.SEEK_CUR) - if oldOffErr == nil { - if _, err := upperFD.Seek(ctx, oldOff, linux.SEEK_SET); err != nil { - upperFD.DecRef(ctx) - return nil, err - } - } - fd.cachedFD.DecRef(ctx) - fd.copiedUp = true - fd.cachedFD = upperFD - fd.cachedFlags = statusFlags - } else if fd.cachedFlags != statusFlags { - if err := fd.cachedFD.SetStatusFlags(ctx, d.fs.creds, statusFlags); err != nil { - return nil, err - } - fd.cachedFlags = statusFlags - } - return fd.cachedFD, nil -} - -// Release implements vfs.FileDescriptionImpl.Release. -func (fd *nonDirectoryFD) Release(ctx context.Context) { - fd.cachedFD.DecRef(ctx) - fd.cachedFD = nil -} - -// OnClose implements vfs.FileDescriptionImpl.OnClose. -func (fd *nonDirectoryFD) OnClose(ctx context.Context) error { - // Linux doesn't define ovl_file_operations.flush at all (i.e. its - // equivalent to OnClose is a no-op). We pass through to - // fd.cachedFD.OnClose() without upgrading if fd.dentry() has been - // copied-up, since OnClose is mostly used to define post-close writeback, - // and if fd.cachedFD hasn't been updated then it can't have been used to - // mutate fd.dentry() anyway. - fd.mu.Lock() - if statusFlags := fd.vfsfd.StatusFlags(); fd.cachedFlags != statusFlags { - if err := fd.cachedFD.SetStatusFlags(ctx, fd.filesystem().creds, statusFlags); err != nil { - fd.mu.Unlock() - return err - } - fd.cachedFlags = statusFlags - } - wrappedFD := fd.cachedFD - defer wrappedFD.IncRef() - fd.mu.Unlock() - return wrappedFD.OnClose(ctx) -} - -// Stat implements vfs.FileDescriptionImpl.Stat. -func (fd *nonDirectoryFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { - var stat linux.Statx - if layerMask := opts.Mask &^ statInternalMask; layerMask != 0 { - wrappedFD, err := fd.getCurrentFD(ctx) - if err != nil { - return linux.Statx{}, err - } - stat, err = wrappedFD.Stat(ctx, vfs.StatOptions{ - Mask: layerMask, - Sync: opts.Sync, - }) - wrappedFD.DecRef(ctx) - if err != nil { - return linux.Statx{}, err - } - } - fd.dentry().statInternalTo(ctx, &opts, &stat) - return stat, nil -} - -// SetStat implements vfs.FileDescriptionImpl.SetStat. -func (fd *nonDirectoryFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { - d := fd.dentry() - mode := linux.FileMode(atomic.LoadUint32(&d.mode)) - if err := vfs.CheckSetStat(ctx, auth.CredentialsFromContext(ctx), &opts, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { - return err - } - mnt := fd.vfsfd.Mount() - if err := mnt.CheckBeginWrite(); err != nil { - return err - } - defer mnt.EndWrite() - if err := d.copyUpLocked(ctx); err != nil { - return err - } - // Changes to d's attributes are serialized by d.copyMu. - d.copyMu.Lock() - defer d.copyMu.Unlock() - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - return err - } - if err := wrappedFD.SetStat(ctx, opts); err != nil { - return err - } - d.updateAfterSetStatLocked(&opts) - return nil -} - -// StatFS implements vfs.FileDescriptionImpl.StatFS. -func (fd *nonDirectoryFD) StatFS(ctx context.Context) (linux.Statfs, error) { - return fd.filesystem().statFS(ctx) -} - -// PRead implements vfs.FileDescriptionImpl.PRead. -func (fd *nonDirectoryFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { - wrappedFD, err := fd.getCurrentFD(ctx) - if err != nil { - return 0, err - } - defer wrappedFD.DecRef(ctx) - return wrappedFD.PRead(ctx, dst, offset, opts) -} - -// Read implements vfs.FileDescriptionImpl.Read. -func (fd *nonDirectoryFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { - // Hold fd.mu during the read to serialize the file offset. - fd.mu.Lock() - defer fd.mu.Unlock() - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - return 0, err - } - return wrappedFD.Read(ctx, dst, opts) -} - -// PWrite implements vfs.FileDescriptionImpl.PWrite. -func (fd *nonDirectoryFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { - wrappedFD, err := fd.getCurrentFD(ctx) - if err != nil { - return 0, err - } - defer wrappedFD.DecRef(ctx) - return wrappedFD.PWrite(ctx, src, offset, opts) -} - -// Write implements vfs.FileDescriptionImpl.Write. -func (fd *nonDirectoryFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { - // Hold fd.mu during the write to serialize the file offset. - fd.mu.Lock() - defer fd.mu.Unlock() - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - return 0, err - } - return wrappedFD.Write(ctx, src, opts) -} - -// Seek implements vfs.FileDescriptionImpl.Seek. -func (fd *nonDirectoryFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { - // Hold fd.mu during the seek to serialize the file offset. - fd.mu.Lock() - defer fd.mu.Unlock() - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - return 0, err - } - return wrappedFD.Seek(ctx, offset, whence) -} - -// Sync implements vfs.FileDescriptionImpl.Sync. -func (fd *nonDirectoryFD) Sync(ctx context.Context) error { - fd.mu.Lock() - if !fd.dentry().isCopiedUp() { - fd.mu.Unlock() - return nil - } - wrappedFD, err := fd.currentFDLocked(ctx) - if err != nil { - fd.mu.Unlock() - return err - } - wrappedFD.IncRef() - defer wrappedFD.DecRef(ctx) - fd.mu.Unlock() - return wrappedFD.Sync(ctx) -} - -// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. -func (fd *nonDirectoryFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { - wrappedFD, err := fd.getCurrentFD(ctx) - if err != nil { - return err - } - defer wrappedFD.DecRef(ctx) - return wrappedFD.ConfigureMMap(ctx, opts) -} diff --git a/pkg/sentry/fsimpl/overlay/overlay.go b/pkg/sentry/fsimpl/overlay/overlay.go index 75cc006bf..c812f0a70 100644 --- a/pkg/sentry/fsimpl/overlay/overlay.go +++ b/pkg/sentry/fsimpl/overlay/overlay.go @@ -18,10 +18,15 @@ // // Lock order: // -// directoryFD.mu / nonDirectoryFD.mu +// directoryFD.mu / regularFileFD.mu // filesystem.renameMu // dentry.dirMu // dentry.copyMu +// filesystem.devMu +// *** "memmap.Mappable locks" below this point +// dentry.mapsMu +// *** "memmap.Mappable locks taken by Translate" below this point +// dentry.dataMu // // Locking dentry.dirMu in multiple dentries requires that parent dentries are // locked before child dentries, and that filesystem.renameMu is locked to @@ -29,14 +34,17 @@ package overlay import ( + "fmt" "strings" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/refsvfs2" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" @@ -46,6 +54,8 @@ import ( const Name = "overlay" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // Name implements vfs.FilesystemType.Name. @@ -53,8 +63,13 @@ func (FilesystemType) Name() string { return Name } +// Release implements FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // FilesystemOptions may be passed as vfs.GetFilesystemOptions.InternalData to // FilesystemType.GetFilesystem. +// +// +stateify savable type FilesystemOptions struct { // Callers passing FilesystemOptions to // overlay.FilesystemType.GetFilesystem() are responsible for ensuring that @@ -71,6 +86,8 @@ type FilesystemOptions struct { } // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -85,97 +102,93 @@ type filesystem struct { // is immutable. dirDevMinor uint32 - // lowerDevMinors maps lower layer filesystems to device minor numbers - // assigned to non-directory files originating from that filesystem. - // lowerDevMinors is immutable. - lowerDevMinors map[*vfs.Filesystem]uint32 + // lowerDevMinors maps device numbers from lower layer filesystems to + // device minor numbers assigned to non-directory files originating from + // that filesystem. (This remapping is necessary for lower layers because a + // file on a lower layer, and that same file on an overlay, are + // distinguishable because they will diverge after copy-up; this isn't true + // for non-directory files already on the upper layer.) lowerDevMinors is + // protected by devMu. + devMu sync.Mutex `state:"nosave"` + lowerDevMinors map[layerDevNumber]uint32 // renameMu synchronizes renaming with non-renaming operations in order to // ensure consistent lock ordering between dentry.dirMu in different // dentries. - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` // lastDirIno is the last inode number assigned to a directory. lastDirIno // is accessed using atomic memory operations. lastDirIno uint64 } +// +stateify savable +type layerDevNumber struct { + major uint32 + minor uint32 +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { mopts := vfs.GenericParseMountOptions(opts.Data) fsoptsRaw := opts.InternalData - fsopts, haveFSOpts := fsoptsRaw.(FilesystemOptions) - if fsoptsRaw != nil && !haveFSOpts { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted overlay.FilesystemOptions or nil", fsoptsRaw) + fsopts, ok := fsoptsRaw.(FilesystemOptions) + if fsoptsRaw != nil && !ok { + ctx.Infof("overlay.FilesystemType.GetFilesystem: GetFilesystemOptions.InternalData has type %T, wanted overlay.FilesystemOptions or nil", fsoptsRaw) return nil, nil, syserror.EINVAL } - if haveFSOpts { - if len(fsopts.LowerRoots) == 0 { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: LowerRoots must be non-empty") + vfsroot := vfs.RootFromContext(ctx) + if vfsroot.Ok() { + defer vfsroot.DecRef(ctx) + } + + if upperPathname, ok := mopts["upperdir"]; ok { + if fsopts.UpperRoot.Ok() { + ctx.Infof("overlay.FilesystemType.GetFilesystem: both upperdir and FilesystemOptions.UpperRoot are specified") return nil, nil, syserror.EINVAL } - if len(fsopts.LowerRoots) < 2 && !fsopts.UpperRoot.Ok() { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: at least two LowerRoots are required when UpperRoot is unspecified") + delete(mopts, "upperdir") + // Linux overlayfs also requires a workdir when upperdir is + // specified; we don't, so silently ignore this option. + delete(mopts, "workdir") + upperPath := fspath.Parse(upperPathname) + if !upperPath.Absolute { + ctx.Infof("overlay.FilesystemType.GetFilesystem: upperdir %q must be absolute", upperPathname) return nil, nil, syserror.EINVAL } - // We don't enforce a maximum number of lower layers when not - // configured by applications; the sandbox owner can have an overlay - // filesystem with any number of lower layers. - } else { - vfsroot := vfs.RootFromContext(ctx) - defer vfsroot.DecRef(ctx) - upperPathname, ok := mopts["upperdir"] - if ok { - delete(mopts, "upperdir") - // Linux overlayfs also requires a workdir when upperdir is - // specified; we don't, so silently ignore this option. - delete(mopts, "workdir") - upperPath := fspath.Parse(upperPathname) - if !upperPath.Absolute { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: upperdir %q must be absolute", upperPathname) - return nil, nil, syserror.EINVAL - } - upperRoot, err := vfsObj.GetDentryAt(ctx, creds, &vfs.PathOperation{ - Root: vfsroot, - Start: vfsroot, - Path: upperPath, - FollowFinalSymlink: true, - }, &vfs.GetDentryOptions{ - CheckSearchable: true, - }) - if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to resolve upperdir %q: %v", upperPathname, err) - return nil, nil, err - } - defer upperRoot.DecRef(ctx) - privateUpperRoot, err := clonePrivateMount(vfsObj, upperRoot, false /* forceReadOnly */) - if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of upperdir %q: %v", upperPathname, err) - return nil, nil, err - } - defer privateUpperRoot.DecRef(ctx) - fsopts.UpperRoot = privateUpperRoot + upperRoot, err := vfsObj.GetDentryAt(ctx, creds, &vfs.PathOperation{ + Root: vfsroot, + Start: vfsroot, + Path: upperPath, + FollowFinalSymlink: true, + }, &vfs.GetDentryOptions{ + CheckSearchable: true, + }) + if err != nil { + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to resolve upperdir %q: %v", upperPathname, err) + return nil, nil, err + } + privateUpperRoot, err := clonePrivateMount(vfsObj, upperRoot, false /* forceReadOnly */) + upperRoot.DecRef(ctx) + if err != nil { + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of upperdir %q: %v", upperPathname, err) + return nil, nil, err } - lowerPathnamesStr, ok := mopts["lowerdir"] - if !ok { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: missing required option lowerdir") + defer privateUpperRoot.DecRef(ctx) + fsopts.UpperRoot = privateUpperRoot + } + + if lowerPathnamesStr, ok := mopts["lowerdir"]; ok { + if len(fsopts.LowerRoots) != 0 { + ctx.Infof("overlay.FilesystemType.GetFilesystem: both lowerdir and FilesystemOptions.LowerRoots are specified") return nil, nil, syserror.EINVAL } delete(mopts, "lowerdir") lowerPathnames := strings.Split(lowerPathnamesStr, ":") - const maxLowerLayers = 500 // Linux: fs/overlay/super.c:OVL_MAX_STACK - if len(lowerPathnames) < 2 && !fsopts.UpperRoot.Ok() { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: at least two lowerdirs are required when upperdir is unspecified") - return nil, nil, syserror.EINVAL - } - if len(lowerPathnames) > maxLowerLayers { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: %d lowerdirs specified, maximum %d", len(lowerPathnames), maxLowerLayers) - return nil, nil, syserror.EINVAL - } for _, lowerPathname := range lowerPathnames { lowerPath := fspath.Parse(lowerPathname) if !lowerPath.Absolute { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: lowerdir %q must be absolute", lowerPathname) + ctx.Infof("overlay.FilesystemType.GetFilesystem: lowerdir %q must be absolute", lowerPathname) return nil, nil, syserror.EINVAL } lowerRoot, err := vfsObj.GetDentryAt(ctx, creds, &vfs.PathOperation{ @@ -187,44 +200,44 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt CheckSearchable: true, }) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to resolve lowerdir %q: %v", lowerPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to resolve lowerdir %q: %v", lowerPathname, err) return nil, nil, err } - defer lowerRoot.DecRef(ctx) privateLowerRoot, err := clonePrivateMount(vfsObj, lowerRoot, true /* forceReadOnly */) + lowerRoot.DecRef(ctx) if err != nil { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of lowerdir %q: %v", lowerPathname, err) + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to make private bind mount of lowerdir %q: %v", lowerPathname, err) return nil, nil, err } defer privateLowerRoot.DecRef(ctx) fsopts.LowerRoots = append(fsopts.LowerRoots, privateLowerRoot) } } + if len(mopts) != 0 { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: unused options: %v", mopts) + ctx.Infof("overlay.FilesystemType.GetFilesystem: unused options: %v", mopts) + return nil, nil, syserror.EINVAL + } + + if len(fsopts.LowerRoots) == 0 { + ctx.Infof("overlay.FilesystemType.GetFilesystem: at least one lower layer is required") + return nil, nil, syserror.EINVAL + } + if len(fsopts.LowerRoots) < 2 && !fsopts.UpperRoot.Ok() { + ctx.Infof("overlay.FilesystemType.GetFilesystem: at least two lower layers are required when no upper layer is present") + return nil, nil, syserror.EINVAL + } + const maxLowerLayers = 500 // Linux: fs/overlay/super.c:OVL_MAX_STACK + if len(fsopts.LowerRoots) > maxLowerLayers { + ctx.Infof("overlay.FilesystemType.GetFilesystem: %d lower layers specified, maximum %d", len(fsopts.LowerRoots), maxLowerLayers) return nil, nil, syserror.EINVAL } - // Allocate device numbers. + // Allocate dirDevMinor. lowerDevMinors are allocated dynamically. dirDevMinor, err := vfsObj.GetAnonBlockDevMinor() if err != nil { return nil, nil, err } - lowerDevMinors := make(map[*vfs.Filesystem]uint32) - for _, lowerRoot := range fsopts.LowerRoots { - lowerFS := lowerRoot.Mount().Filesystem() - if _, ok := lowerDevMinors[lowerFS]; !ok { - devMinor, err := vfsObj.GetAnonBlockDevMinor() - if err != nil { - vfsObj.PutAnonBlockDevMinor(dirDevMinor) - for _, lowerDevMinor := range lowerDevMinors { - vfsObj.PutAnonBlockDevMinor(lowerDevMinor) - } - return nil, nil, err - } - lowerDevMinors[lowerFS] = devMinor - } - } // Take extra references held by the filesystem. if fsopts.UpperRoot.Ok() { @@ -238,7 +251,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt opts: fsopts, creds: creds.Fork(), dirDevMinor: dirDevMinor, - lowerDevMinors: lowerDevMinors, + lowerDevMinors: make(map[layerDevNumber]uint32), } fs.vfsfs.Init(vfsObj, &fstype, fs) @@ -274,7 +287,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, syserror.EREMOTE } if isWhiteout(&rootStat) { - ctx.Warningf("overlay.FilesystemType.GetFilesystem: filesystem root is a whiteout") + ctx.Infof("overlay.FilesystemType.GetFilesystem: filesystem root is a whiteout") root.destroyLocked(ctx) fs.vfsfs.DecRef(ctx) return nil, nil, syserror.EINVAL @@ -288,7 +301,14 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt root.ino = fs.newDirIno() } else if !root.upperVD.Ok() { root.devMajor = linux.UNNAMED_MAJOR - root.devMinor = fs.lowerDevMinors[root.lowerVDs[0].Mount().Filesystem()] + rootDevMinor, err := fs.getLowerDevMinor(rootStat.DevMajor, rootStat.DevMinor) + if err != nil { + ctx.Infof("overlay.FilesystemType.GetFilesystem: failed to get device number for root: %v", err) + root.destroyLocked(ctx) + fs.vfsfs.DecRef(ctx) + return nil, nil, err + } + root.devMinor = rootDevMinor root.ino = rootStat.Ino } else { root.devMajor = rootStat.DevMajor @@ -315,7 +335,11 @@ func clonePrivateMount(vfsObj *vfs.VirtualFilesystem, vd vfs.VirtualDentry, forc if err != nil { return vfs.VirtualDentry{}, err } - return vfs.MakeVirtualDentry(newmnt, vd.Dentry()), nil + // Take a reference on the dentry which will be owned by the returned + // VirtualDentry. + d := vd.Dentry() + d.IncRef() + return vfs.MakeVirtualDentry(newmnt, d), nil } // Release implements vfs.FilesystemImpl.Release. @@ -357,7 +381,24 @@ func (fs *filesystem) newDirIno() uint64 { return atomic.AddUint64(&fs.lastDirIno, 1) } +func (fs *filesystem) getLowerDevMinor(layerMajor, layerMinor uint32) (uint32, error) { + fs.devMu.Lock() + defer fs.devMu.Unlock() + orig := layerDevNumber{layerMajor, layerMinor} + if minor, ok := fs.lowerDevMinors[orig]; ok { + return minor, nil + } + minor, err := fs.vfsfs.VirtualFilesystem().GetAnonBlockDevMinor() + if err != nil { + return 0, err + } + fs.lowerDevMinors[orig] = minor + return minor, nil +} + // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -390,7 +431,7 @@ type dentry struct { // and dirents (if not nil) is a cache of dirents as returned by // directoryFDs representing this directory. children is protected by // dirMu. - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` children map[string]*dentry dirents []vfs.Dirent @@ -400,7 +441,7 @@ type dentry struct { // If !upperVD.Ok(), it can transition to a valid vfs.VirtualDentry (i.e. // be copied up) with copyMu locked for writing; otherwise, it is // immutable. lowerVDs is always immutable. - copyMu sync.RWMutex + copyMu sync.RWMutex `state:"nosave"` upperVD vfs.VirtualDentry lowerVDs []vfs.VirtualDentry @@ -415,7 +456,43 @@ type dentry struct { devMinor uint32 ino uint64 + // If this dentry represents a regular file, then: + // + // - mapsMu is used to synchronize between copy-up and memmap.Mappable + // methods on dentry preceding mm.MemoryManager.activeMu in the lock order. + // + // - dataMu is used to synchronize between copy-up and + // dentry.(memmap.Mappable).Translate. + // + // - lowerMappings tracks memory mappings of the file. lowerMappings is + // used to invalidate mappings of the lower layer when the file is copied + // up to ensure that they remain coherent with subsequent writes to the + // file. (Note that, as of this writing, Linux overlayfs does not do this; + // this feature is a gVisor extension.) lowerMappings is protected by + // mapsMu. + // + // - If this dentry is copied-up, then wrappedMappable is the Mappable + // obtained from a call to the current top layer's + // FileDescription.ConfigureMMap(). Once wrappedMappable becomes non-nil + // (from a call to regularFileFD.ensureMappable()), it cannot become nil. + // wrappedMappable is protected by mapsMu and dataMu. + // + // - isMappable is non-zero iff wrappedMappable is non-nil. isMappable is + // accessed using atomic memory operations. + mapsMu sync.Mutex `state:"nosave"` + lowerMappings memmap.MappingSet + dataMu sync.RWMutex `state:"nosave"` + wrappedMappable memmap.Mappable + isMappable uint32 + locks vfs.FileLocks + + // watches is the set of inotify watches on the file repesented by this dentry. + // + // Note that hard links to the same file will not share the same set of + // watches, due to the fact that we do not have inode structures in this + // overlay implementation. + watches vfs.Watches } // newDentry creates a new dentry. The dentry initially has no references; it @@ -428,6 +505,9 @@ func (fs *filesystem) newDentry() *dentry { } d.lowerVDs = d.inlineLowerVDs[:0] d.vfsd.Init(d) + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Register(d, "overlay.dentry") + } return d } @@ -475,6 +555,14 @@ func (d *dentry) checkDropLocked(ctx context.Context) { if atomic.LoadInt64(&d.refs) != 0 { return } + + // Make sure that we do not lose watches on dentries that have not been + // deleted. Note that overlayfs never calls VFS.InvalidateDentry(), so + // d.vfsd.IsDead() indicates that d was deleted. + if !d.vfsd.IsDead() && d.watches.Size() > 0 { + return + } + // Refs is still zero; destroy it. d.destroyLocked(ctx) return @@ -482,7 +570,9 @@ func (d *dentry) checkDropLocked(ctx context.Context) { // destroyLocked destroys the dentry. // -// Preconditions: d.fs.renameMu must be locked for writing. d.refs == 0. +// Preconditions: +// * d.fs.renameMu must be locked for writing. +// * d.refs == 0. func (d *dentry) destroyLocked(ctx context.Context) { switch atomic.LoadInt64(&d.refs) { case 0: @@ -501,6 +591,8 @@ func (d *dentry) destroyLocked(ctx context.Context) { lowerVD.DecRef(ctx) } + d.watches.HandleDeletion(ctx) + if d.parent != nil { d.parent.dirMu.Lock() if !d.vfsd.IsDead() { @@ -515,23 +607,48 @@ func (d *dentry) destroyLocked(ctx context.Context) { panic("overlay.dentry.DecRef() called without holding a reference") } } + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Unregister(d, "overlay.dentry") + } +} + +// LeakMessage implements refsvfs2.CheckedObject.LeakMessage. +func (d *dentry) LeakMessage() string { + return fmt.Sprintf("[overlay.dentry %p] reference count of %d instead of -1", d, atomic.LoadInt64(&d.refs)) } // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. func (d *dentry) InotifyWithParent(ctx context.Context, events uint32, cookie uint32, et vfs.EventType) { - // TODO(gvisor.dev/issue/1479): Implement inotify. + if d.isDir() { + events |= linux.IN_ISDIR + } + + // overlayfs never calls VFS.InvalidateDentry(), so d.vfsd.IsDead() indicates + // that d was deleted. + deleted := d.vfsd.IsDead() + + d.fs.renameMu.RLock() + // The ordering below is important, Linux always notifies the parent first. + if d.parent != nil { + d.parent.watches.Notify(ctx, d.name, events, cookie, et, deleted) + } + d.watches.Notify(ctx, "", events, cookie, et, deleted) + d.fs.renameMu.RUnlock() } // Watches implements vfs.DentryImpl.Watches. func (d *dentry) Watches() *vfs.Watches { - // TODO(gvisor.dev/issue/1479): Implement inotify. - return nil + return &d.watches } // OnZeroWatches implements vfs.DentryImpl.OnZeroWatches. -// -// TODO(gvisor.dev/issue/1479): Implement inotify. -func (d *dentry) OnZeroWatches(context.Context) {} +func (d *dentry) OnZeroWatches(ctx context.Context) { + if atomic.LoadInt64(&d.refs) == 0 { + d.fs.renameMu.Lock() + d.checkDropLocked(ctx) + d.fs.renameMu.Unlock() + } +} // iterLayers invokes yield on each layer comprising d, from top to bottom. If // any call to yield returns false, iterLayer stops iteration. @@ -564,6 +681,16 @@ func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) } +func (d *dentry) checkXattrPermissions(creds *auth.Credentials, name string, ats vfs.AccessTypes) error { + mode := linux.FileMode(atomic.LoadUint32(&d.mode)) + kuid := auth.KUID(atomic.LoadUint32(&d.uid)) + kgid := auth.KGID(atomic.LoadUint32(&d.gid)) + if err := vfs.GenericCheckPermissions(creds, ats, mode, kuid, kgid); err != nil { + return err + } + return vfs.CheckXattrPermissions(creds, ats, mode, kuid, name) +} + // statInternalMask is the set of stat fields that is set by // dentry.statInternalTo(). const statInternalMask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID | linux.STATX_INO @@ -602,6 +729,8 @@ func (d *dentry) updateAfterSetStatLocked(opts *vfs.SetStatOptions) { // fileDescription is embedded by overlay implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -616,6 +745,48 @@ func (fd *fileDescription) dentry() *dentry { return fd.vfsfd.Dentry().Impl().(*dentry) } +// ListXattr implements vfs.FileDescriptionImpl.ListXattr. +func (fd *fileDescription) ListXattr(ctx context.Context, size uint64) ([]string, error) { + return fd.filesystem().listXattr(ctx, fd.dentry(), size) +} + +// GetXattr implements vfs.FileDescriptionImpl.GetXattr. +func (fd *fileDescription) GetXattr(ctx context.Context, opts vfs.GetXattrOptions) (string, error) { + return fd.filesystem().getXattr(ctx, fd.dentry(), auth.CredentialsFromContext(ctx), &opts) +} + +// SetXattr implements vfs.FileDescriptionImpl.SetXattr. +func (fd *fileDescription) SetXattr(ctx context.Context, opts vfs.SetXattrOptions) error { + fs := fd.filesystem() + d := fd.dentry() + + fs.renameMu.RLock() + err := fs.setXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), &opts) + fs.renameMu.RUnlock() + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil +} + +// RemoveXattr implements vfs.FileDescriptionImpl.RemoveXattr. +func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error { + fs := fd.filesystem() + d := fd.dentry() + + fs.renameMu.RLock() + err := fs.removeXattrLocked(ctx, d, fd.vfsfd.Mount(), auth.CredentialsFromContext(ctx), name) + fs.renameMu.RUnlock() + if err != nil { + return err + } + + d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) + return nil +} + // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) diff --git a/pkg/sentry/fsimpl/overlay/regular_file.go b/pkg/sentry/fsimpl/overlay/regular_file.go new file mode 100644 index 000000000..2b89a7a6d --- /dev/null +++ b/pkg/sentry/fsimpl/overlay/regular_file.go @@ -0,0 +1,456 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package overlay + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/sync" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" + "gvisor.dev/gvisor/pkg/waiter" +) + +func (d *dentry) isRegularFile() bool { + return atomic.LoadUint32(&d.mode)&linux.S_IFMT == linux.S_IFREG +} + +func (d *dentry) isSymlink() bool { + return atomic.LoadUint32(&d.mode)&linux.S_IFMT == linux.S_IFLNK +} + +func (d *dentry) readlink(ctx context.Context) (string, error) { + layerVD := d.topLayer() + return d.fs.vfsfs.VirtualFilesystem().ReadlinkAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: layerVD, + Start: layerVD, + }) +} + +// +stateify savable +type regularFileFD struct { + fileDescription + + // If copiedUp is false, cachedFD represents + // fileDescription.dentry().lowerVDs[0]; otherwise, cachedFD represents + // fileDescription.dentry().upperVD. cachedFlags is the last known value of + // cachedFD.StatusFlags(). copiedUp, cachedFD, and cachedFlags are + // protected by mu. + mu sync.Mutex `state:"nosave"` + copiedUp bool + cachedFD *vfs.FileDescription + cachedFlags uint32 + + // If copiedUp is false, lowerWaiters contains all waiter.Entries + // registered with cachedFD. lowerWaiters is protected by mu. + lowerWaiters map[*waiter.Entry]waiter.EventMask +} + +func (fd *regularFileFD) getCurrentFD(ctx context.Context) (*vfs.FileDescription, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return nil, err + } + wrappedFD.IncRef() + return wrappedFD, nil +} + +func (fd *regularFileFD) currentFDLocked(ctx context.Context) (*vfs.FileDescription, error) { + d := fd.dentry() + statusFlags := fd.vfsfd.StatusFlags() + if !fd.copiedUp && d.isCopiedUp() { + // Switch to the copied-up file. + upperVD := d.topLayer() + upperFD, err := fd.filesystem().vfsfs.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: upperVD, + Start: upperVD, + }, &vfs.OpenOptions{ + Flags: statusFlags, + }) + if err != nil { + return nil, err + } + oldOff, oldOffErr := fd.cachedFD.Seek(ctx, 0, linux.SEEK_CUR) + if oldOffErr == nil { + if _, err := upperFD.Seek(ctx, oldOff, linux.SEEK_SET); err != nil { + upperFD.DecRef(ctx) + return nil, err + } + } + if len(fd.lowerWaiters) != 0 { + ready := upperFD.Readiness(^waiter.EventMask(0)) + for e, mask := range fd.lowerWaiters { + fd.cachedFD.EventUnregister(e) + upperFD.EventRegister(e, mask) + if ready&mask != 0 { + e.Callback.Callback(e) + } + } + } + fd.cachedFD.DecRef(ctx) + fd.copiedUp = true + fd.cachedFD = upperFD + fd.cachedFlags = statusFlags + fd.lowerWaiters = nil + } else if fd.cachedFlags != statusFlags { + if err := fd.cachedFD.SetStatusFlags(ctx, d.fs.creds, statusFlags); err != nil { + return nil, err + } + fd.cachedFlags = statusFlags + } + return fd.cachedFD, nil +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *regularFileFD) Release(ctx context.Context) { + fd.cachedFD.DecRef(ctx) + fd.cachedFD = nil +} + +// OnClose implements vfs.FileDescriptionImpl.OnClose. +func (fd *regularFileFD) OnClose(ctx context.Context) error { + // Linux doesn't define ovl_file_operations.flush at all (i.e. its + // equivalent to OnClose is a no-op). We pass through to + // fd.cachedFD.OnClose() without upgrading if fd.dentry() has been + // copied-up, since OnClose is mostly used to define post-close writeback, + // and if fd.cachedFD hasn't been updated then it can't have been used to + // mutate fd.dentry() anyway. + fd.mu.Lock() + if statusFlags := fd.vfsfd.StatusFlags(); fd.cachedFlags != statusFlags { + if err := fd.cachedFD.SetStatusFlags(ctx, fd.filesystem().creds, statusFlags); err != nil { + fd.mu.Unlock() + return err + } + fd.cachedFlags = statusFlags + } + wrappedFD := fd.cachedFD + fd.mu.Unlock() + return wrappedFD.OnClose(ctx) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *regularFileFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + var stat linux.Statx + if layerMask := opts.Mask &^ statInternalMask; layerMask != 0 { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return linux.Statx{}, err + } + stat, err = wrappedFD.Stat(ctx, vfs.StatOptions{ + Mask: layerMask, + Sync: opts.Sync, + }) + wrappedFD.DecRef(ctx) + if err != nil { + return linux.Statx{}, err + } + } + fd.dentry().statInternalTo(ctx, &opts, &stat) + return stat, nil +} + +// Allocate implements vfs.FileDescriptionImpl.Allocate. +func (fd *regularFileFD) Allocate(ctx context.Context, mode, offset, length uint64) error { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return err + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.Allocate(ctx, mode, offset, length) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *regularFileFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + d := fd.dentry() + mode := linux.FileMode(atomic.LoadUint32(&d.mode)) + if err := vfs.CheckSetStat(ctx, auth.CredentialsFromContext(ctx), &opts, mode, auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))); err != nil { + return err + } + mnt := fd.vfsfd.Mount() + if err := mnt.CheckBeginWrite(); err != nil { + return err + } + defer mnt.EndWrite() + if err := d.copyUpLocked(ctx); err != nil { + return err + } + // Changes to d's attributes are serialized by d.copyMu. + d.copyMu.Lock() + defer d.copyMu.Unlock() + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return err + } + if err := wrappedFD.SetStat(ctx, opts); err != nil { + return err + } + d.updateAfterSetStatLocked(&opts) + if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { + d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) + } + return nil +} + +// StatFS implements vfs.FileDescriptionImpl.StatFS. +func (fd *regularFileFD) StatFS(ctx context.Context) (linux.Statfs, error) { + return fd.filesystem().statFS(ctx) +} + +// Readiness implements waiter.Waitable.Readiness. +func (fd *regularFileFD) Readiness(mask waiter.EventMask) waiter.EventMask { + ctx := context.Background() + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + // TODO(b/171089913): Just use fd.cachedFD since Readiness can't return + // an error. This is obviously wrong, but at least consistent with + // VFS1. + log.Warningf("overlay.regularFileFD.Readiness: currentFDLocked failed: %v", err) + fd.mu.Lock() + wrappedFD = fd.cachedFD + wrappedFD.IncRef() + fd.mu.Unlock() + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.Readiness(mask) +} + +// EventRegister implements waiter.Waitable.EventRegister. +func (fd *regularFileFD) EventRegister(e *waiter.Entry, mask waiter.EventMask) { + fd.mu.Lock() + defer fd.mu.Unlock() + wrappedFD, err := fd.currentFDLocked(context.Background()) + if err != nil { + // TODO(b/171089913): Just use fd.cachedFD since EventRegister can't + // return an error. This is obviously wrong, but at least consistent + // with VFS1. + log.Warningf("overlay.regularFileFD.EventRegister: currentFDLocked failed: %v", err) + wrappedFD = fd.cachedFD + } + wrappedFD.EventRegister(e, mask) + if !fd.copiedUp { + if fd.lowerWaiters == nil { + fd.lowerWaiters = make(map[*waiter.Entry]waiter.EventMask) + } + fd.lowerWaiters[e] = mask + } +} + +// EventUnregister implements waiter.Waitable.EventUnregister. +func (fd *regularFileFD) EventUnregister(e *waiter.Entry) { + fd.mu.Lock() + defer fd.mu.Unlock() + fd.cachedFD.EventUnregister(e) + if !fd.copiedUp { + delete(fd.lowerWaiters, e) + } +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *regularFileFD) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return 0, err + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.PRead(ctx, dst, offset, opts) +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *regularFileFD) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // Hold fd.mu during the read to serialize the file offset. + fd.mu.Lock() + defer fd.mu.Unlock() + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return 0, err + } + return wrappedFD.Read(ctx, dst, opts) +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *regularFileFD) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return 0, err + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.PWrite(ctx, src, offset, opts) +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *regularFileFD) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + // Hold fd.mu during the write to serialize the file offset. + fd.mu.Lock() + defer fd.mu.Unlock() + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return 0, err + } + return wrappedFD.Write(ctx, src, opts) +} + +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *regularFileFD) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + // Hold fd.mu during the seek to serialize the file offset. + fd.mu.Lock() + defer fd.mu.Unlock() + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return 0, err + } + return wrappedFD.Seek(ctx, offset, whence) +} + +// Sync implements vfs.FileDescriptionImpl.Sync. +func (fd *regularFileFD) Sync(ctx context.Context) error { + fd.mu.Lock() + if !fd.dentry().isCopiedUp() { + fd.mu.Unlock() + return nil + } + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + fd.mu.Unlock() + return err + } + wrappedFD.IncRef() + defer wrappedFD.DecRef(ctx) + fd.mu.Unlock() + return wrappedFD.Sync(ctx) +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (fd *regularFileFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + wrappedFD, err := fd.getCurrentFD(ctx) + if err != nil { + return 0, err + } + defer wrappedFD.DecRef(ctx) + return wrappedFD.Ioctl(ctx, uio, args) +} + +// ConfigureMMap implements vfs.FileDescriptionImpl.ConfigureMMap. +func (fd *regularFileFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + if err := fd.ensureMappable(ctx, opts); err != nil { + return err + } + return vfs.GenericConfigureMMap(&fd.vfsfd, fd.dentry(), opts) +} + +// ensureMappable ensures that fd.dentry().wrappedMappable is not nil. +func (fd *regularFileFD) ensureMappable(ctx context.Context, opts *memmap.MMapOpts) error { + d := fd.dentry() + + // Fast path if we already have a Mappable for the current top layer. + if atomic.LoadUint32(&d.isMappable) != 0 { + return nil + } + + // Only permit mmap of regular files, since other file types may have + // unpredictable behavior when mmapped (e.g. /dev/zero). + if atomic.LoadUint32(&d.mode)&linux.S_IFMT != linux.S_IFREG { + return syserror.ENODEV + } + + // Get a Mappable for the current top layer. + fd.mu.Lock() + defer fd.mu.Unlock() + d.copyMu.RLock() + defer d.copyMu.RUnlock() + if atomic.LoadUint32(&d.isMappable) != 0 { + return nil + } + wrappedFD, err := fd.currentFDLocked(ctx) + if err != nil { + return err + } + if err := wrappedFD.ConfigureMMap(ctx, opts); err != nil { + return err + } + if opts.MappingIdentity != nil { + opts.MappingIdentity.DecRef(ctx) + opts.MappingIdentity = nil + } + // Use this Mappable for all mappings of this layer (unless we raced with + // another call to ensureMappable). + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.dataMu.Lock() + defer d.dataMu.Unlock() + if d.wrappedMappable == nil { + d.wrappedMappable = opts.Mappable + atomic.StoreUint32(&d.isMappable, 1) + } + return nil +} + +// AddMapping implements memmap.Mappable.AddMapping. +func (d *dentry) AddMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) error { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + if err := d.wrappedMappable.AddMapping(ctx, ms, ar, offset, writable); err != nil { + return err + } + if !d.isCopiedUp() { + d.lowerMappings.AddMapping(ms, ar, offset, writable) + } + return nil +} + +// RemoveMapping implements memmap.Mappable.RemoveMapping. +func (d *dentry) RemoveMapping(ctx context.Context, ms memmap.MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + d.wrappedMappable.RemoveMapping(ctx, ms, ar, offset, writable) + if !d.isCopiedUp() { + d.lowerMappings.RemoveMapping(ms, ar, offset, writable) + } +} + +// CopyMapping implements memmap.Mappable.CopyMapping. +func (d *dentry) CopyMapping(ctx context.Context, ms memmap.MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + if err := d.wrappedMappable.CopyMapping(ctx, ms, srcAR, dstAR, offset, writable); err != nil { + return err + } + if !d.isCopiedUp() { + d.lowerMappings.AddMapping(ms, dstAR, offset, writable) + } + return nil +} + +// Translate implements memmap.Mappable.Translate. +func (d *dentry) Translate(ctx context.Context, required, optional memmap.MappableRange, at usermem.AccessType) ([]memmap.Translation, error) { + d.dataMu.RLock() + defer d.dataMu.RUnlock() + return d.wrappedMappable.Translate(ctx, required, optional, at) +} + +// InvalidateUnsavable implements memmap.Mappable.InvalidateUnsavable. +func (d *dentry) InvalidateUnsavable(ctx context.Context) error { + d.mapsMu.Lock() + defer d.mapsMu.Unlock() + return d.wrappedMappable.InvalidateUnsavable(ctx) +} diff --git a/pkg/sentry/fsimpl/overlay/save_restore.go b/pkg/sentry/fsimpl/overlay/save_restore.go new file mode 100644 index 000000000..054e17b17 --- /dev/null +++ b/pkg/sentry/fsimpl/overlay/save_restore.go @@ -0,0 +1,27 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package overlay + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/refsvfs2" +) + +func (d *dentry) afterLoad() { + if refsvfs2.LeakCheckEnabled() && atomic.LoadInt64(&d.refs) != -1 { + refsvfs2.Register(d, "overlay.dentry") + } +} diff --git a/pkg/sentry/fsimpl/pipefs/pipefs.go b/pkg/sentry/fsimpl/pipefs/pipefs.go index 2ca793db9..e44b79b68 100644 --- a/pkg/sentry/fsimpl/pipefs/pipefs.go +++ b/pkg/sentry/fsimpl/pipefs/pipefs.go @@ -31,6 +31,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type filesystemType struct{} // Name implements vfs.FilesystemType.Name. @@ -38,11 +39,15 @@ func (filesystemType) Name() string { return "pipefs" } +// Release implements vfs.FilesystemType.Release. +func (filesystemType) Release(ctx context.Context) {} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (filesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { panic("pipefs.filesystemType.GetFilesystem should never be called") } +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -76,6 +81,8 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { kernfs.InodeNotDirectory kernfs.InodeNotSymlink @@ -143,12 +150,14 @@ func (i *inode) SetStat(ctx context.Context, vfsfs *vfs.Filesystem, creds *auth. return syserror.EPERM } -// TODO(gvisor.dev/issue/1193): kernfs does not provide a way to implement -// statfs, from which we should indicate PIPEFS_MAGIC. - // Open implements kernfs.Inode.Open. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - return i.pipe.Open(ctx, rp.Mount(), vfsd, opts.Flags, &i.locks) +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + return i.pipe.Open(ctx, rp.Mount(), d.VFSDentry(), opts.Flags, &i.locks) +} + +// StatFS implements kernfs.Inode.StatFS. +func (i *inode) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.PIPEFS_MAGIC), nil } // NewConnectedPipeFDs returns a pair of FileDescriptions representing the read @@ -159,7 +168,7 @@ func NewConnectedPipeFDs(ctx context.Context, mnt *vfs.Mount, flags uint32) (*vf fs := mnt.Filesystem().Impl().(*filesystem) inode := newInode(ctx, fs) var d kernfs.Dentry - d.Init(inode) + d.Init(&fs.Filesystem, inode) defer d.DecRef(ctx) return inode.pipe.ReaderWriterPair(mnt, d.VFSDentry(), flags) } diff --git a/pkg/sentry/fsimpl/proc/BUILD b/pkg/sentry/fsimpl/proc/BUILD index 14ecfd300..5196a2a80 100644 --- a/pkg/sentry/fsimpl/proc/BUILD +++ b/pkg/sentry/fsimpl/proc/BUILD @@ -1,18 +1,79 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "fd_dir_inode_refs", + out = "fd_dir_inode_refs.go", + package = "proc", + prefix = "fdDirInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "fdDirInode", + }, +) + +go_template_instance( + name = "fd_info_dir_inode_refs", + out = "fd_info_dir_inode_refs.go", + package = "proc", + prefix = "fdInfoDirInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "fdInfoDirInode", + }, +) + +go_template_instance( + name = "subtasks_inode_refs", + out = "subtasks_inode_refs.go", + package = "proc", + prefix = "subtasksInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "subtasksInode", + }, +) + +go_template_instance( + name = "task_inode_refs", + out = "task_inode_refs.go", + package = "proc", + prefix = "taskInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "taskInode", + }, +) + +go_template_instance( + name = "tasks_inode_refs", + out = "tasks_inode_refs.go", + package = "proc", + prefix = "tasksInode", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "tasksInode", + }, +) + go_library( name = "proc", srcs = [ + "fd_dir_inode_refs.go", + "fd_info_dir_inode_refs.go", "filesystem.go", "subtasks.go", + "subtasks_inode_refs.go", "task.go", "task_fds.go", "task_files.go", + "task_inode_refs.go", "task_net.go", "tasks.go", "tasks_files.go", + "tasks_inode_refs.go", "tasks_sys.go", ], visibility = ["//pkg/sentry:internal"], @@ -21,6 +82,7 @@ go_library( "//pkg/context", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/fs/lock", "//pkg/sentry/fsbridge", @@ -39,6 +101,7 @@ go_library( "//pkg/sync", "//pkg/syserror", "//pkg/tcpip/header", + "//pkg/tcpip/network/ipv4", "//pkg/usermem", ], ) diff --git a/pkg/sentry/fsimpl/proc/filesystem.go b/pkg/sentry/fsimpl/proc/filesystem.go index 2463d51cd..99abcab66 100644 --- a/pkg/sentry/fsimpl/proc/filesystem.go +++ b/pkg/sentry/fsimpl/proc/filesystem.go @@ -17,6 +17,7 @@ package proc import ( "fmt" + "strconv" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -24,23 +25,29 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" ) -// Name is the default filesystem name. -const Name = "proc" +const ( + // Name is the default filesystem name. + Name = "proc" + defaultMaxCachedDentries = uint64(1000) +) // FilesystemType is the factory class for procfs. // // +stateify savable type FilesystemType struct{} -var _ vfs.FilesystemType = (*FilesystemType)(nil) - // Name implements vfs.FilesystemType.Name. func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -61,9 +68,22 @@ func (ft FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualF if err != nil { return nil, nil, err } + + mopts := vfs.GenericParseMountOptions(opts.Data) + maxCachedDentries := defaultMaxCachedDentries + if str, ok := mopts["dentry_cache_limit"]; ok { + delete(mopts, "dentry_cache_limit") + maxCachedDentries, err = strconv.ParseUint(str, 10, 64) + if err != nil { + ctx.Warningf("proc.FilesystemType.GetFilesystem: invalid dentry cache limit: dentry_cache_limit=%s", str) + return nil, nil, syserror.EINVAL + } + } + procfs := &filesystem{ devMinor: devMinor, } + procfs.MaxCachedDentries = maxCachedDentries procfs.VFSFilesystem().Init(vfsObj, &ft, procfs) var cgroups map[string]string @@ -72,7 +92,9 @@ func (ft FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualF cgroups = data.Cgroups } - _, dentry := procfs.newTasksInode(k, pidns, cgroups) + inode := procfs.newTasksInode(ctx, k, pidns, cgroups) + var dentry kernfs.Dentry + dentry.Init(&procfs.Filesystem, inode) return procfs.VFSFilesystem(), dentry.VFSDentry(), nil } @@ -84,21 +106,21 @@ func (fs *filesystem) Release(ctx context.Context) { // dynamicInode is an overfitted interface for common Inodes with // dynamicByteSource types used in procfs. +// +// +stateify savable type dynamicInode interface { kernfs.Inode vfs.DynamicBytesSource - Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) + Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, data vfs.DynamicBytesSource, perm linux.FileMode) } -func (fs *filesystem) newDentry(creds *auth.Credentials, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { - inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, inode, perm) - - d := &kernfs.Dentry{} - d.Init(inode) - return d +func (fs *filesystem) newInode(ctx context.Context, creds *auth.Credentials, perm linux.FileMode, inode dynamicInode) dynamicInode { + inode.Init(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), inode, perm) + return inode } +// +stateify savable type staticFile struct { kernfs.DynamicBytesFile vfs.StaticData @@ -110,8 +132,24 @@ func newStaticFile(data string) *staticFile { return &staticFile{StaticData: vfs.StaticData{Data: data}} } +func (fs *filesystem) newStaticDir(ctx context.Context, creds *auth.Credentials, children map[string]kernfs.Inode) kernfs.Inode { + return kernfs.NewStaticDir(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, children, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }) +} + // InternalData contains internal data passed in to the procfs mount via // vfs.GetFilesystemOptions.InternalData. +// +// +stateify savable type InternalData struct { Cgroups map[string]string } + +// +stateify savable +type implStatFS struct{} + +// StatFS implements kernfs.Inode.StatFS. +func (*implStatFS) StatFS(context.Context, *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.PROC_SUPER_MAGIC), nil +} diff --git a/pkg/sentry/fsimpl/proc/subtasks.go b/pkg/sentry/fsimpl/proc/subtasks.go index 79c2725f3..cb3c5e0fd 100644 --- a/pkg/sentry/fsimpl/proc/subtasks.go +++ b/pkg/sentry/fsimpl/proc/subtasks.go @@ -31,11 +31,14 @@ import ( // // +stateify savable type subtasksInode struct { - kernfs.InodeNotSymlink - kernfs.InodeDirectoryNoNewChildren + implStatFS + kernfs.InodeAlwaysValid kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.InodeTemporary kernfs.OrderedChildren - kernfs.AlwaysValid + subtasksInodeRefs locks vfs.FileLocks @@ -47,7 +50,7 @@ type subtasksInode struct { var _ kernfs.Inode = (*subtasksInode)(nil) -func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) *kernfs.Dentry { +func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) kernfs.Inode { subInode := &subtasksInode{ fs: fs, task: task, @@ -55,18 +58,16 @@ func (fs *filesystem) newSubtasks(task *kernel.Task, pidns *kernel.PIDNamespace, cgroupControllers: cgroupControllers, } // Note: credentials are overridden by taskOwnedInode. - subInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + subInode.InodeAttrs.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) subInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + subInode.EnableLeakCheck() inode := &taskOwnedInode{Inode: subInode, owner: task} - dentry := &kernfs.Dentry{} - dentry.Init(inode) - - return dentry + return inode } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDirectory.Lookup. +func (i *subtasksInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { tid, err := strconv.ParseUint(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -79,13 +80,11 @@ func (i *subtasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, e if subTask.ThreadGroup() != i.task.ThreadGroup() { return nil, syserror.ENOENT } - - subTaskDentry := i.fs.newTaskInode(subTask, i.pidns, false, i.cgroupControllers) - return subTaskDentry.VFSDentry(), nil + return i.fs.newTaskInode(subTask, i.pidns, false, i.cgroupControllers) } -// IterDirents implements kernfs.inodeDynamicLookup. -func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +// IterDirents implements kernfs.inodeDirectory.IterDirents. +func (i *subtasksInode) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { tasks := i.task.ThreadGroup().MemberIDs(i.pidns) if len(tasks) == 0 { return offset, syserror.ENOENT @@ -115,6 +114,7 @@ func (i *subtasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallb return offset, nil } +// +stateify savable type subtasksFD struct { kernfs.GenericDirectoryFD @@ -152,19 +152,21 @@ func (fd *subtasksFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) erro return fd.GenericDirectoryFD.SetStat(ctx, opts) } -// Open implements kernfs.Inode. -func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +// Open implements kernfs.Inode.Open. +func (i *subtasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &subtasksFD{task: i.task} - if err := fd.Init(&i.OrderedChildren, &i.locks, &opts); err != nil { + if err := fd.Init(&i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }); err != nil { return nil, err } - if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.VFSFileDescription().Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return fd.VFSFileDescription(), nil } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { stat, err := i.InodeAttrs.Stat(ctx, vsfs, opts) if err != nil { @@ -176,7 +178,12 @@ func (i *subtasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs return stat, nil } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*subtasksInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } + +// DecRef implements kernfs.Inode.DecRef. +func (i *subtasksInode) DecRef(ctx context.Context) { + i.subtasksInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} diff --git a/pkg/sentry/fsimpl/proc/task.go b/pkg/sentry/fsimpl/proc/task.go index a5c7aa470..57cf8ce26 100644 --- a/pkg/sentry/fsimpl/proc/task.go +++ b/pkg/sentry/fsimpl/proc/task.go @@ -32,11 +32,13 @@ import ( // // +stateify savable type taskInode struct { - kernfs.InodeNotSymlink - kernfs.InodeDirectoryNoNewChildren - kernfs.InodeNoDynamicLookup + implStatFS kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.InodeTemporary kernfs.OrderedChildren + taskInodeRefs locks vfs.FileLocks @@ -45,80 +47,92 @@ type taskInode struct { var _ kernfs.Inode = (*taskInode)(nil) -func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) *kernfs.Dentry { - // TODO(gvisor.dev/issue/164): Fail with ESRCH if task exited. - contents := map[string]*kernfs.Dentry{ - "auxv": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &auxvData{task: task}), - "cmdline": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), +func (fs *filesystem) newTaskInode(task *kernel.Task, pidns *kernel.PIDNamespace, isThreadGroup bool, cgroupControllers map[string]string) (kernfs.Inode, error) { + if task.ExitState() == kernel.TaskExitDead { + return nil, syserror.ESRCH + } + + contents := map[string]kernfs.Inode{ + "auxv": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &auxvData{task: task}), + "cmdline": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: cmdlineDataArg}), "comm": fs.newComm(task, fs.NextIno(), 0444), - "environ": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), + "cwd": fs.newCwdSymlink(task, fs.NextIno()), + "environ": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &cmdlineData{task: task, arg: environDataArg}), "exe": fs.newExeSymlink(task, fs.NextIno()), "fd": fs.newFDDirInode(task), "fdinfo": fs.newFDInfoDirInode(task), - "gid_map": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &idMapData{task: task, gids: true}), - "io": fs.newTaskOwnedFile(task, fs.NextIno(), 0400, newIO(task, isThreadGroup)), - "maps": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mapsData{task: task}), - "mountinfo": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mountInfoData{task: task}), - "mounts": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &mountsData{task: task}), + "gid_map": fs.newTaskOwnedInode(task, fs.NextIno(), 0644, &idMapData{task: task, gids: true}), + "io": fs.newTaskOwnedInode(task, fs.NextIno(), 0400, newIO(task, isThreadGroup)), + "maps": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &mapsData{task: task}), + "mountinfo": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &mountInfoData{task: task}), + "mounts": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &mountsData{task: task}), "net": fs.newTaskNetDir(task), - "ns": fs.newTaskOwnedDir(task, fs.NextIno(), 0511, map[string]*kernfs.Dentry{ + "ns": fs.newTaskOwnedDir(task, fs.NextIno(), 0511, map[string]kernfs.Inode{ "net": fs.newNamespaceSymlink(task, fs.NextIno(), "net"), "pid": fs.newNamespaceSymlink(task, fs.NextIno(), "pid"), "user": fs.newNamespaceSymlink(task, fs.NextIno(), "user"), }), - "oom_score": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, newStaticFile("0\n")), - "oom_score_adj": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &oomScoreAdj{task: task}), - "smaps": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &smapsData{task: task}), - "stat": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), - "statm": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &statmData{task: task}), - "status": fs.newTaskOwnedFile(task, fs.NextIno(), 0444, &statusData{task: task, pidns: pidns}), - "uid_map": fs.newTaskOwnedFile(task, fs.NextIno(), 0644, &idMapData{task: task, gids: false}), + "oom_score": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, newStaticFile("0\n")), + "oom_score_adj": fs.newTaskOwnedInode(task, fs.NextIno(), 0644, &oomScoreAdj{task: task}), + "smaps": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &smapsData{task: task}), + "stat": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &taskStatData{task: task, pidns: pidns, tgstats: isThreadGroup}), + "statm": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &statmData{task: task}), + "status": fs.newTaskOwnedInode(task, fs.NextIno(), 0444, &statusData{task: task, pidns: pidns}), + "uid_map": fs.newTaskOwnedInode(task, fs.NextIno(), 0644, &idMapData{task: task, gids: false}), } if isThreadGroup { contents["task"] = fs.newSubtasks(task, pidns, cgroupControllers) } if len(cgroupControllers) > 0 { - contents["cgroup"] = fs.newTaskOwnedFile(task, fs.NextIno(), 0444, newCgroupData(cgroupControllers)) + contents["cgroup"] = fs.newTaskOwnedInode(task, fs.NextIno(), 0444, newCgroupData(cgroupControllers)) } taskInode := &taskInode{task: task} // Note: credentials are overridden by taskOwnedInode. - taskInode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + taskInode.InodeAttrs.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + taskInode.EnableLeakCheck() inode := &taskOwnedInode{Inode: taskInode, owner: task} - dentry := &kernfs.Dentry{} - dentry.Init(inode) taskInode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - links := taskInode.OrderedChildren.Populate(dentry, contents) + links := taskInode.OrderedChildren.Populate(contents) taskInode.IncLinks(links) - return dentry + return inode, nil } -// Valid implements kernfs.inodeDynamicLookup. This inode remains valid as long +// Valid implements kernfs.Inode.Valid. This inode remains valid as long // as the task is still running. When it's dead, another tasks with the same // PID could replace it. func (i *taskInode) Valid(ctx context.Context) bool { return i.task.ExitState() != kernel.TaskExitDead } -// Open implements kernfs.Inode. -func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +// Open implements kernfs.Inode.Open. +func (i *taskInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*taskInode) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } +// DecRef implements kernfs.Inode.DecRef. +func (i *taskInode) DecRef(ctx context.Context) { + i.taskInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + // taskOwnedInode implements kernfs.Inode and overrides inode owner with task // effective user and group. +// +// +stateify savable type taskOwnedInode struct { kernfs.Inode @@ -128,34 +142,26 @@ type taskOwnedInode struct { var _ kernfs.Inode = (*taskOwnedInode)(nil) -func (fs *filesystem) newTaskOwnedFile(task *kernel.Task, ino uint64, perm linux.FileMode, inode dynamicInode) *kernfs.Dentry { +func (fs *filesystem) newTaskOwnedInode(task *kernel.Task, ino uint64, perm linux.FileMode, inode dynamicInode) kernfs.Inode { // Note: credentials are overridden by taskOwnedInode. - inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, inode, perm) + inode.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, inode, perm) - taskInode := &taskOwnedInode{Inode: inode, owner: task} - d := &kernfs.Dentry{} - d.Init(taskInode) - return d + return &taskOwnedInode{Inode: inode, owner: task} } -func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, children map[string]*kernfs.Dentry) *kernfs.Dentry { - dir := &kernfs.StaticDirectory{} - +func (fs *filesystem) newTaskOwnedDir(task *kernel.Task, ino uint64, perm linux.FileMode, children map[string]kernfs.Inode) kernfs.Inode { // Note: credentials are overridden by taskOwnedInode. - dir.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, perm) - - inode := &taskOwnedInode{Inode: dir, owner: task} - d := &kernfs.Dentry{} - d.Init(inode) + fdOpts := kernfs.GenericDirectoryFDOptions{SeekEnd: kernfs.SeekEndZero} + dir := kernfs.NewStaticDir(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, perm, children, fdOpts) - dir.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - links := dir.OrderedChildren.Populate(d, children) - dir.IncLinks(links) + return &taskOwnedInode{Inode: dir, owner: task} +} - return d +func (i *taskOwnedInode) Valid(ctx context.Context) bool { + return i.owner.ExitState() != kernel.TaskExitDead && i.Inode.Valid(ctx) } -// Stat implements kernfs.Inode. +// Stat implements kernfs.Inode.Stat. func (i *taskOwnedInode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs.StatOptions) (linux.Statx, error) { stat, err := i.Inode.Stat(ctx, fs, opts) if err != nil { @@ -173,7 +179,7 @@ func (i *taskOwnedInode) Stat(ctx context.Context, fs *vfs.Filesystem, opts vfs. return stat, nil } -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. func (i *taskOwnedInode) CheckPermissions(_ context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { mode := i.Mode() uid, gid := i.getOwner(mode) diff --git a/pkg/sentry/fsimpl/proc/task_fds.go b/pkg/sentry/fsimpl/proc/task_fds.go index f0d3f7f5e..d268b44be 100644 --- a/pkg/sentry/fsimpl/proc/task_fds.go +++ b/pkg/sentry/fsimpl/proc/task_fds.go @@ -22,7 +22,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -52,6 +51,7 @@ func taskFDExists(ctx context.Context, t *kernel.Task, fd int32) bool { return true } +// +stateify savable type fdDir struct { locks vfs.FileLocks @@ -63,8 +63,8 @@ type fdDir struct { produceSymlink bool } -// IterDirents implements kernfs.inodeDynamicLookup. -func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { +// IterDirents implements kernfs.inodeDirectory.IterDirents. +func (i *fdDir) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { var fds []int32 i.task.WithMuLocked(func(t *kernel.Task) { if fdTable := t.FDTable(); fdTable != nil { @@ -87,31 +87,39 @@ func (i *fdDir) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, off Name: strconv.FormatUint(uint64(fd), 10), Type: typ, Ino: i.fs.NextIno(), - NextOff: offset + 1, + NextOff: int64(fd) + 3, } if err := cb.Handle(dirent); err != nil { - return offset, err + // Getdents should iterate correctly despite mutation + // of fds, so we return the next fd to serialize plus + // 2 (which accounts for the "." and ".." tracked by + // kernfs) as the offset. + return int64(fd) + 2, err } - offset++ } - return offset, nil + // We serialized them all. Next offset should be higher than last + // serialized fd. + return int64(fds[len(fds)-1]) + 3, nil } // fdDirInode represents the inode for /proc/[pid]/fd directory. // // +stateify savable type fdDirInode struct { - kernfs.InodeNotSymlink - kernfs.InodeDirectoryNoNewChildren + fdDir + fdDirInodeRefs + implStatFS + kernfs.InodeAlwaysValid kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.InodeTemporary kernfs.OrderedChildren - kernfs.AlwaysValid - fdDir } var _ kernfs.Inode = (*fdDirInode)(nil) -func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry { +func (fs *filesystem) newFDDirInode(task *kernel.Task) kernfs.Inode { inode := &fdDirInode{ fdDir: fdDir{ fs: fs, @@ -119,17 +127,19 @@ func (fs *filesystem) newFDDirInode(task *kernel.Task) *kernfs.Dentry { produceSymlink: true, }, } - inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) - - dentry := &kernfs.Dentry{} - dentry.Init(inode) + inode.InodeAttrs.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) + return inode +} - return dentry +// IterDirents implements kernfs.inodeDirectory.IterDirents. +func (i *fdDirInode) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (int64, error) { + return i.fdDir.IterDirents(ctx, mnt, cb, offset, relOffset) } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDirectory.Lookup. +func (i *fdDirInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { fdInt, err := strconv.ParseInt(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -138,20 +148,21 @@ func (i *fdDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro if !taskFDExists(ctx, i.task, fd) { return nil, syserror.ENOENT } - taskDentry := i.fs.newFDSymlink(i.task, fd, i.fs.NextIno()) - return taskDentry.VFSDentry(), nil + return i.fs.newFDSymlink(i.task, fd, i.fs.NextIno()), nil } -// Open implements kernfs.Inode. -func (i *fdDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +// Open implements kernfs.Inode.Open. +func (i *fdDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } -// CheckPermissions implements kernfs.Inode. +// CheckPermissions implements kernfs.Inode.CheckPermissions. // // 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 @@ -173,10 +184,16 @@ func (i *fdDirInode) CheckPermissions(ctx context.Context, creds *auth.Credentia return err } +// DecRef implements kernfs.Inode.DecRef. +func (i *fdDirInode) DecRef(ctx context.Context) { + i.fdDirInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + // fdSymlink is an symlink for the /proc/[pid]/fd/[fd] file. // // +stateify savable type fdSymlink struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeSymlink @@ -187,19 +204,16 @@ type fdSymlink struct { var _ kernfs.Inode = (*fdSymlink)(nil) -func (fs *filesystem) newFDSymlink(task *kernel.Task, fd int32, ino uint64) *kernfs.Dentry { +func (fs *filesystem) newFDSymlink(task *kernel.Task, fd int32, ino uint64) kernfs.Inode { inode := &fdSymlink{ task: task, fd: fd, } - inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) - - d := &kernfs.Dentry{} - d.Init(inode) - return d + inode.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) + return inode } -func (s *fdSymlink) Readlink(ctx context.Context) (string, error) { +func (s *fdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { file, _ := getTaskFD(s.task, s.fd) if file == nil { return "", syserror.ENOENT @@ -221,38 +235,43 @@ func (s *fdSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDen return vd, "", nil } +// Valid implements kernfs.Inode.Valid. +func (s *fdSymlink) Valid(ctx context.Context) bool { + return taskFDExists(ctx, s.task, s.fd) +} + // fdInfoDirInode represents the inode for /proc/[pid]/fdinfo directory. // // +stateify savable type fdInfoDirInode struct { - kernfs.InodeNotSymlink - kernfs.InodeDirectoryNoNewChildren + fdDir + fdInfoDirInodeRefs + implStatFS + kernfs.InodeAlwaysValid kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.InodeTemporary kernfs.OrderedChildren - kernfs.AlwaysValid - fdDir } var _ kernfs.Inode = (*fdInfoDirInode)(nil) -func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) *kernfs.Dentry { +func (fs *filesystem) newFDInfoDirInode(task *kernel.Task) kernfs.Inode { inode := &fdInfoDirInode{ fdDir: fdDir{ fs: fs, task: task, }, } - inode.InodeAttrs.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) - - dentry := &kernfs.Dentry{} - dentry.Init(inode) + inode.InodeAttrs.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - - return dentry + return inode } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { +// Lookup implements kernfs.inodeDirectory.Lookup. +func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { fdInt, err := strconv.ParseInt(name, 10, 32) if err != nil { return nil, syserror.ENOENT @@ -265,25 +284,35 @@ func (i *fdInfoDirInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, task: i.task, fd: fd, } - dentry := i.fs.newTaskOwnedFile(i.task, i.fs.NextIno(), 0444, data) - return dentry.VFSDentry(), nil + return i.fs.newTaskOwnedInode(i.task, i.fs.NextIno(), 0444, data), nil } -// Open implements kernfs.Inode. -func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +// IterDirents implements Inode.IterDirents. +func (i *fdInfoDirInode) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, relOffset int64) (newOffset int64, err error) { + return i.fdDir.IterDirents(ctx, mnt, cb, offset, relOffset) +} + +// Open implements kernfs.Inode.Open. +func (i *fdInfoDirInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } +// DecRef implements kernfs.Inode.DecRef. +func (i *fdInfoDirInode) DecRef(ctx context.Context) { + i.fdInfoDirInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + // fdInfoData implements vfs.DynamicBytesSource for /proc/[pid]/fdinfo/[fd]. // // +stateify savable type fdInfoData struct { kernfs.DynamicBytesFile - refs.AtomicRefCount task *kernel.Task fd int32 @@ -305,3 +334,8 @@ func (d *fdInfoData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "flags:\t0%o\n", flags) return nil } + +// Valid implements kernfs.Inode.Valid. +func (d *fdInfoData) Valid(ctx context.Context) bool { + return taskFDExists(ctx, d.task, d.fd) +} diff --git a/pkg/sentry/fsimpl/proc/task_files.go b/pkg/sentry/fsimpl/proc/task_files.go index 830b78949..d3f4e259b 100644 --- a/pkg/sentry/fsimpl/proc/task_files.go +++ b/pkg/sentry/fsimpl/proc/task_files.go @@ -247,13 +247,10 @@ type commInode struct { task *kernel.Task } -func (fs *filesystem) newComm(task *kernel.Task, ino uint64, perm linux.FileMode) *kernfs.Dentry { +func (fs *filesystem) newComm(task *kernel.Task, ino uint64, perm linux.FileMode) kernfs.Inode { inode := &commInode{task: task} - inode.DynamicBytesFile.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, &commData{task: task}, perm) - - d := &kernfs.Dentry{} - d.Init(inode) - return d + inode.DynamicBytesFile.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, &commData{task: task}, perm) + return inode } func (i *commInode) CheckPermissions(ctx context.Context, creds *auth.Credentials, ats vfs.AccessTypes) error { @@ -543,7 +540,7 @@ func (s *statusData) Generate(ctx context.Context, buf *bytes.Buffer) error { var vss, rss, data uint64 s.task.WithMuLocked(func(t *kernel.Task) { if fdTable := t.FDTable(); fdTable != nil { - fds = fdTable.Size() + fds = fdTable.CurrentMaxFDs() } if mm := t.MemoryManager(); mm != nil { vss = mm.VirtualMemorySize() @@ -648,6 +645,7 @@ func (o *oomScoreAdj) Write(ctx context.Context, src usermem.IOSequence, offset // // +stateify savable type exeSymlink struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeSymlink @@ -657,29 +655,30 @@ type exeSymlink struct { var _ kernfs.Inode = (*exeSymlink)(nil) -func (fs *filesystem) newExeSymlink(task *kernel.Task, ino uint64) *kernfs.Dentry { +func (fs *filesystem) newExeSymlink(task *kernel.Task, ino uint64) kernfs.Inode { inode := &exeSymlink{task: task} - inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) - - d := &kernfs.Dentry{} - d.Init(inode) - return d + inode.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) + return inode } -// Readlink implements kernfs.Inode. -func (s *exeSymlink) Readlink(ctx context.Context) (string, error) { - if !kernel.ContextCanTrace(ctx, s.task, false) { - return "", syserror.EACCES - } - - // Pull out the executable for /proc/[pid]/exe. - exec, err := s.executable() +// Readlink implements kernfs.Inode.Readlink. +func (s *exeSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { + exec, _, err := s.Getlink(ctx, nil) if err != nil { return "", err } defer exec.DecRef(ctx) - return exec.PathnameWithDeleted(ctx), nil + root := vfs.RootFromContext(ctx) + if !root.Ok() { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + vfsObj := exec.Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, exec) + return name, nil } // Getlink implements kernfs.Inode.Getlink. @@ -687,23 +686,12 @@ func (s *exeSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDent if !kernel.ContextCanTrace(ctx, s.task, false) { return vfs.VirtualDentry{}, "", syserror.EACCES } - - exec, err := s.executable() - if err != nil { - return vfs.VirtualDentry{}, "", err - } - defer exec.DecRef(ctx) - - vd := exec.(*fsbridge.VFSFile).FileDescription().VirtualDentry() - vd.IncRef() - return vd, "", nil -} - -func (s *exeSymlink) executable() (file fsbridge.File, err error) { if err := checkTaskState(s.task); err != nil { - return nil, err + return vfs.VirtualDentry{}, "", err } + var err error + var exec fsbridge.File s.task.WithMuLocked(func(t *kernel.Task) { mm := t.MemoryManager() if mm == nil { @@ -714,12 +702,75 @@ func (s *exeSymlink) executable() (file fsbridge.File, err error) { // The MemoryManager may be destroyed, in which case // MemoryManager.destroy will simply set the executable to nil // (with locks held). - file = mm.Executable() - if file == nil { + exec = mm.Executable() + if exec == nil { err = syserror.ESRCH } }) - return + if err != nil { + return vfs.VirtualDentry{}, "", err + } + defer exec.DecRef(ctx) + + vd := exec.(*fsbridge.VFSFile).FileDescription().VirtualDentry() + vd.IncRef() + return vd, "", nil +} + +// cwdSymlink is an symlink for the /proc/[pid]/cwd file. +// +// +stateify savable +type cwdSymlink struct { + implStatFS + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeSymlink + + task *kernel.Task +} + +var _ kernfs.Inode = (*cwdSymlink)(nil) + +func (fs *filesystem) newCwdSymlink(task *kernel.Task, ino uint64) kernfs.Inode { + inode := &cwdSymlink{task: task} + inode.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) + return inode +} + +// Readlink implements kernfs.Inode.Readlink. +func (s *cwdSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { + cwd, _, err := s.Getlink(ctx, nil) + if err != nil { + return "", err + } + defer cwd.DecRef(ctx) + + root := vfs.RootFromContext(ctx) + if !root.Ok() { + // It could have raced with process deletion. + return "", syserror.ESRCH + } + defer root.DecRef(ctx) + + vfsObj := cwd.Mount().Filesystem().VirtualFilesystem() + name, _ := vfsObj.PathnameWithDeleted(ctx, root, cwd) + return name, nil +} + +// Getlink implements kernfs.Inode.Getlink. +func (s *cwdSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { + if !kernel.ContextCanTrace(ctx, s.task, false) { + return vfs.VirtualDentry{}, "", syserror.EACCES + } + if err := checkTaskState(s.task); err != nil { + return vfs.VirtualDentry{}, "", err + } + cwd := s.task.FSContext().WorkingDirectoryVFS2() + if !cwd.Ok() { + // It could have raced with process deletion. + return vfs.VirtualDentry{}, "", syserror.ESRCH + } + return cwd, "", nil } // mountInfoData is used to implement /proc/[pid]/mountinfo. @@ -784,13 +835,14 @@ func (i *mountsData) Generate(ctx context.Context, buf *bytes.Buffer) error { return nil } +// +stateify savable type namespaceSymlink struct { kernfs.StaticSymlink task *kernel.Task } -func (fs *filesystem) newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) *kernfs.Dentry { +func (fs *filesystem) newNamespaceSymlink(task *kernel.Task, ino uint64, ns string) kernfs.Inode { // 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. @@ -798,40 +850,44 @@ func (fs *filesystem) newNamespaceSymlink(task *kernel.Task, ino uint64, ns stri inode := &namespaceSymlink{task: task} // Note: credentials are overridden by taskOwnedInode. - inode.Init(task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, target) + inode.Init(task, task.Credentials(), linux.UNNAMED_MAJOR, fs.devMinor, ino, target) taskInode := &taskOwnedInode{Inode: inode, owner: task} - d := &kernfs.Dentry{} - d.Init(taskInode) - return d + return taskInode } -// Readlink implements Inode. -func (s *namespaceSymlink) Readlink(ctx context.Context) (string, error) { +// Readlink implements kernfs.Inode.Readlink. +func (s *namespaceSymlink) Readlink(ctx context.Context, mnt *vfs.Mount) (string, error) { if err := checkTaskState(s.task); err != nil { return "", err } - return s.StaticSymlink.Readlink(ctx) + return s.StaticSymlink.Readlink(ctx, mnt) } -// Getlink implements Inode.Getlink. +// Getlink implements kernfs.Inode.Getlink. func (s *namespaceSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { if err := checkTaskState(s.task); err != nil { return vfs.VirtualDentry{}, "", err } // Create a synthetic inode to represent the namespace. + fs := mnt.Filesystem().Impl().(*filesystem) + nsInode := &namespaceInode{} + nsInode.Init(ctx, auth.CredentialsFromContext(ctx), linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0444) dentry := &kernfs.Dentry{} - dentry.Init(&namespaceInode{}) + dentry.Init(&fs.Filesystem, nsInode) vd := vfs.MakeVirtualDentry(mnt, dentry.VFSDentry()) - vd.IncRef() - dentry.DecRef(ctx) + // Only IncRef vd.Mount() because vd.Dentry() already holds a ref of 1. + mnt.IncRef() return vd, "", nil } // namespaceInode is a synthetic inode created to represent a namespace in // /proc/[pid]/ns/*. +// +// +stateify savable type namespaceInode struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeNotDirectory @@ -843,19 +899,19 @@ type namespaceInode struct { var _ kernfs.Inode = (*namespaceInode)(nil) // Init initializes a namespace inode. -func (i *namespaceInode) Init(creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { +func (i *namespaceInode) Init(ctx context.Context, creds *auth.Credentials, devMajor, devMinor uint32, ino uint64, perm linux.FileMode) { if perm&^linux.PermissionsMask != 0 { panic(fmt.Sprintf("Only permission mask must be set: %x", perm&linux.PermissionsMask)) } - i.InodeAttrs.Init(creds, devMajor, devMinor, ino, linux.ModeRegular|perm) + i.InodeAttrs.Init(ctx, creds, devMajor, devMinor, ino, linux.ModeRegular|perm) } -// Open implements Inode.Open. -func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +// Open implements kernfs.Inode.Open. +func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { fd := &namespaceFD{inode: i} i.IncRef() fd.LockFD.Init(&i.locks) - if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), vfsd, &vfs.FileDescriptionOptions{}); err != nil { + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{}); err != nil { return nil, err } return &fd.vfsfd, nil @@ -863,6 +919,8 @@ func (i *namespaceInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd * // namespace FD is a synthetic file that represents a namespace in // /proc/[pid]/ns/*. +// +// +stateify savable type namespaceFD struct { vfs.FileDescriptionDefaultImpl vfs.LockFD @@ -873,20 +931,20 @@ type namespaceFD struct { var _ vfs.FileDescriptionImpl = (*namespaceFD)(nil) -// Stat implements FileDescriptionImpl. +// Stat implements vfs.FileDescriptionImpl.Stat. func (fd *namespaceFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() return fd.inode.Stat(ctx, vfs, opts) } -// SetStat implements FileDescriptionImpl. +// SetStat implements vfs.FileDescriptionImpl.SetStat. func (fd *namespaceFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { vfs := fd.vfsfd.VirtualDentry().Mount().Filesystem() creds := auth.CredentialsFromContext(ctx) return fd.inode.SetStat(ctx, vfs, creds, opts) } -// Release implements FileDescriptionImpl. +// Release implements vfs.FileDescriptionImpl.Release. func (fd *namespaceFD) Release(ctx context.Context) { fd.inode.DecRef(ctx) } diff --git a/pkg/sentry/fsimpl/proc/task_net.go b/pkg/sentry/fsimpl/proc/task_net.go index a4c884bf9..5a9ee111f 100644 --- a/pkg/sentry/fsimpl/proc/task_net.go +++ b/pkg/sentry/fsimpl/proc/task_net.go @@ -37,12 +37,12 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -func (fs *filesystem) newTaskNetDir(task *kernel.Task) *kernfs.Dentry { +func (fs *filesystem) newTaskNetDir(task *kernel.Task) kernfs.Inode { k := task.Kernel() pidns := task.PIDNamespace() root := auth.NewRootCredentials(pidns.UserNamespace()) - var contents map[string]*kernfs.Dentry + var contents map[string]kernfs.Inode if stack := task.NetworkNamespace().Stack(); stack != nil { const ( arp = "IP address HW type Flags HW address Mask Device\n" @@ -56,34 +56,34 @@ func (fs *filesystem) newTaskNetDir(task *kernel.Task) *kernfs.Dentry { // TODO(gvisor.dev/issue/1833): Make sure file contents reflect the task // network namespace. - contents = map[string]*kernfs.Dentry{ - "dev": fs.newDentry(root, fs.NextIno(), 0444, &netDevData{stack: stack}), - "snmp": fs.newDentry(root, fs.NextIno(), 0444, &netSnmpData{stack: stack}), + contents = map[string]kernfs.Inode{ + "dev": fs.newInode(task, root, 0444, &netDevData{stack: stack}), + "snmp": fs.newInode(task, root, 0444, &netSnmpData{stack: stack}), // The following files are simple stubs until they are implemented in // netstack, if the file contains a header the stub is just the header // otherwise it is an empty file. - "arp": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(arp)), - "netlink": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(netlink)), - "netstat": fs.newDentry(root, fs.NextIno(), 0444, &netStatData{}), - "packet": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(packet)), - "protocols": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(protocols)), + "arp": fs.newInode(task, root, 0444, newStaticFile(arp)), + "netlink": fs.newInode(task, root, 0444, newStaticFile(netlink)), + "netstat": fs.newInode(task, root, 0444, &netStatData{}), + "packet": fs.newInode(task, root, 0444, newStaticFile(packet)), + "protocols": fs.newInode(task, root, 0444, newStaticFile(protocols)), // Linux sets psched values to: nsec per usec, psched tick in ns, 1000000, // high res timer ticks per sec (ClockGetres returns 1ns resolution). - "psched": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(psched)), - "ptype": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(ptype)), - "route": fs.newDentry(root, fs.NextIno(), 0444, &netRouteData{stack: stack}), - "tcp": fs.newDentry(root, fs.NextIno(), 0444, &netTCPData{kernel: k}), - "udp": fs.newDentry(root, fs.NextIno(), 0444, &netUDPData{kernel: k}), - "unix": fs.newDentry(root, fs.NextIno(), 0444, &netUnixData{kernel: k}), + "psched": fs.newInode(task, root, 0444, newStaticFile(psched)), + "ptype": fs.newInode(task, root, 0444, newStaticFile(ptype)), + "route": fs.newInode(task, root, 0444, &netRouteData{stack: stack}), + "tcp": fs.newInode(task, root, 0444, &netTCPData{kernel: k}), + "udp": fs.newInode(task, root, 0444, &netUDPData{kernel: k}), + "unix": fs.newInode(task, root, 0444, &netUnixData{kernel: k}), } if stack.SupportsIPv6() { - contents["if_inet6"] = fs.newDentry(root, fs.NextIno(), 0444, &ifinet6{stack: stack}) - contents["ipv6_route"] = fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")) - contents["tcp6"] = fs.newDentry(root, fs.NextIno(), 0444, &netTCP6Data{kernel: k}) - contents["udp6"] = fs.newDentry(root, fs.NextIno(), 0444, newStaticFile(upd6)) + contents["if_inet6"] = fs.newInode(task, root, 0444, &ifinet6{stack: stack}) + contents["ipv6_route"] = fs.newInode(task, root, 0444, newStaticFile("")) + contents["tcp6"] = fs.newInode(task, root, 0444, &netTCP6Data{kernel: k}) + contents["udp6"] = fs.newInode(task, root, 0444, newStaticFile(upd6)) } } @@ -262,7 +262,7 @@ func (n *netUnixData) Generate(ctx context.Context, buf *bytes.Buffer) error { // For now, we always redact this pointer. fmt.Fprintf(buf, "%#016p: %08X %08X %08X %04X %02X %8d", (*unix.SocketOperations)(nil), // Num, pointer to kernel socket struct. - s.Refs()-1, // RefCount, don't count our own ref. + s.ReadRefs()-1, // RefCount, don't count our own ref. 0, // Protocol, always 0 for UDS. sockFlags, // Flags. sops.Endpoint().Type(), // Type. @@ -430,7 +430,7 @@ func commonGenerateTCP(ctx context.Context, buf *bytes.Buffer, k *kernel.Kernel, // Field: refcount. Don't count the ref we obtain while deferencing // the weakref to this socket. - fmt.Fprintf(buf, "%d ", s.Refs()-1) + fmt.Fprintf(buf, "%d ", s.ReadRefs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -589,7 +589,7 @@ func (d *netUDPData) Generate(ctx context.Context, buf *bytes.Buffer) error { // Field: ref; reference count on the socket inode. Don't count the ref // we obtain while deferencing the weakref to this socket. - fmt.Fprintf(buf, "%d ", s.Refs()-1) + fmt.Fprintf(buf, "%d ", s.ReadRefs()-1) // Field: Socket struct address. Redacted due to the same reason as // the 'Num' field in /proc/net/unix, see netUnix.ReadSeqFileData. @@ -616,6 +616,7 @@ type netSnmpData struct { var _ dynamicInode = (*netSnmpData)(nil) +// +stateify savable type snmpLine struct { prefix string header string @@ -660,7 +661,7 @@ func sprintSlice(s []uint64) string { return r[1 : len(r)-1] // Remove "[]" introduced by fmt of slice. } -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *netSnmpData) Generate(ctx context.Context, buf *bytes.Buffer) error { types := []interface{}{ &inet.StatSNMPIP{}, @@ -709,7 +710,7 @@ type netRouteData struct { var _ dynamicInode = (*netRouteData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. // See Linux's net/ipv4/fib_trie.c:fib_route_seq_show. func (d *netRouteData) Generate(ctx context.Context, buf *bytes.Buffer) error { fmt.Fprintf(buf, "%-127s\n", "Iface\tDestination\tGateway\tFlags\tRefCnt\tUse\tMetric\tMask\tMTU\tWindow\tIRTT") @@ -773,7 +774,7 @@ type netStatData struct { var _ dynamicInode = (*netStatData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. // See Linux's net/ipv4/fib_trie.c:fib_route_seq_show. func (d *netStatData) Generate(ctx context.Context, buf *bytes.Buffer) error { buf.WriteString("TcpExt: SyncookiesSent SyncookiesRecv SyncookiesFailed " + diff --git a/pkg/sentry/fsimpl/proc/tasks.go b/pkg/sentry/fsimpl/proc/tasks.go index 6d2b90a8b..b81ea14bf 100644 --- a/pkg/sentry/fsimpl/proc/tasks.go +++ b/pkg/sentry/fsimpl/proc/tasks.go @@ -37,11 +37,14 @@ const ( // // +stateify savable type tasksInode struct { - kernfs.InodeNotSymlink - kernfs.InodeDirectoryNoNewChildren + implStatFS + kernfs.InodeAlwaysValid kernfs.InodeAttrs + kernfs.InodeDirectoryNoNewChildren + kernfs.InodeNotSymlink + kernfs.InodeTemporary // This holds no meaning as this inode can't be Looked up and is always valid. kernfs.OrderedChildren - kernfs.AlwaysValid + tasksInodeRefs locks vfs.FileLocks @@ -50,8 +53,6 @@ type tasksInode struct { // '/proc/self' and '/proc/thread-self' have custom directory offsets in // Linux. So handle them outside of OrderedChildren. - selfSymlink *vfs.Dentry - threadSelfSymlink *vfs.Dentry // cgroupControllers is a map of controller name to directory in the // cgroup hierarchy. These controllers are immutable and will be listed @@ -61,51 +62,53 @@ type tasksInode struct { var _ kernfs.Inode = (*tasksInode)(nil) -func (fs *filesystem) newTasksInode(k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) (*tasksInode, *kernfs.Dentry) { +func (fs *filesystem) newTasksInode(ctx context.Context, k *kernel.Kernel, pidns *kernel.PIDNamespace, cgroupControllers map[string]string) *tasksInode { root := auth.NewRootCredentials(pidns.UserNamespace()) - contents := map[string]*kernfs.Dentry{ - "cpuinfo": fs.newDentry(root, fs.NextIno(), 0444, newStaticFileSetStat(cpuInfoData(k))), - "filesystems": fs.newDentry(root, fs.NextIno(), 0444, &filesystemsData{}), - "loadavg": fs.newDentry(root, fs.NextIno(), 0444, &loadavgData{}), - "sys": fs.newSysDir(root, k), - "meminfo": fs.newDentry(root, fs.NextIno(), 0444, &meminfoData{}), - "mounts": kernfs.NewStaticSymlink(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/mounts"), - "net": kernfs.NewStaticSymlink(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/net"), - "stat": fs.newDentry(root, fs.NextIno(), 0444, &statData{}), - "uptime": fs.newDentry(root, fs.NextIno(), 0444, &uptimeData{}), - "version": fs.newDentry(root, fs.NextIno(), 0444, &versionData{}), + contents := map[string]kernfs.Inode{ + "cpuinfo": fs.newInode(ctx, root, 0444, newStaticFileSetStat(cpuInfoData(k))), + "filesystems": fs.newInode(ctx, root, 0444, &filesystemsData{}), + "loadavg": fs.newInode(ctx, root, 0444, &loadavgData{}), + "sys": fs.newSysDir(ctx, root, k), + "meminfo": fs.newInode(ctx, root, 0444, &meminfoData{}), + "mounts": kernfs.NewStaticSymlink(ctx, root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/mounts"), + "net": kernfs.NewStaticSymlink(ctx, root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), "self/net"), + "stat": fs.newInode(ctx, root, 0444, &statData{}), + "uptime": fs.newInode(ctx, root, 0444, &uptimeData{}), + "version": fs.newInode(ctx, root, 0444, &versionData{}), } inode := &tasksInode{ pidns: pidns, fs: fs, - selfSymlink: fs.newSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), - threadSelfSymlink: fs.newThreadSelfSymlink(root, fs.NextIno(), pidns).VFSDentry(), cgroupControllers: cgroupControllers, } - inode.InodeAttrs.Init(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) - - dentry := &kernfs.Dentry{} - dentry.Init(inode) + inode.InodeAttrs.Init(ctx, root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0555) + inode.EnableLeakCheck() inode.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - links := inode.OrderedChildren.Populate(dentry, contents) + links := inode.OrderedChildren.Populate(contents) inode.IncLinks(links) - return inode, dentry + return inode } -// Lookup implements kernfs.inodeDynamicLookup. -func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, error) { - // Try to lookup a corresponding task. +// Lookup implements kernfs.inodeDirectory.Lookup. +func (i *tasksInode) Lookup(ctx context.Context, name string) (kernfs.Inode, error) { + // Check if a static entry was looked up. + if d, err := i.OrderedChildren.Lookup(ctx, name); err == nil { + return d, nil + } + + // Not a static entry. Try to lookup a corresponding task. tid, err := strconv.ParseUint(name, 10, 64) if err != nil { + root := auth.NewRootCredentials(i.pidns.UserNamespace()) // If it failed to parse, check if it's one of the special handled files. switch name { case selfName: - return i.selfSymlink, nil + return i.newSelfSymlink(ctx, root), nil case threadSelfName: - return i.threadSelfSymlink, nil + return i.newThreadSelfSymlink(ctx, root), nil } return nil, syserror.ENOENT } @@ -115,12 +118,11 @@ func (i *tasksInode) Lookup(ctx context.Context, name string) (*vfs.Dentry, erro return nil, syserror.ENOENT } - taskDentry := i.fs.newTaskInode(task, i.pidns, true, i.cgroupControllers) - return taskDentry.VFSDentry(), nil + return i.fs.newTaskInode(task, i.pidns, true, i.cgroupControllers) } -// IterDirents implements kernfs.inodeDynamicLookup. -func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { +// IterDirents implements kernfs.inodeDirectory.IterDirents. +func (i *tasksInode) IterDirents(ctx context.Context, mnt *vfs.Mount, cb vfs.IterDirentsCallback, offset, _ int64) (int64, error) { // fs/proc/internal.h: #define FIRST_PROCESS_ENTRY 256 const FIRST_PROCESS_ENTRY = 256 @@ -197,9 +199,11 @@ func (i *tasksInode) IterDirents(ctx context.Context, cb vfs.IterDirentsCallback return maxTaskID, nil } -// Open implements kernfs.Inode. -func (i *tasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &i.OrderedChildren, &i.locks, &opts) +// Open implements kernfs.Inode.Open. +func (i *tasksInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), d, &i.OrderedChildren, &i.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndZero, + }) if err != nil { return nil, err } @@ -224,9 +228,16 @@ func (i *tasksInode) Stat(ctx context.Context, vsfs *vfs.Filesystem, opts vfs.St return stat, nil } +// DecRef implements kernfs.Inode.DecRef. +func (i *tasksInode) DecRef(ctx context.Context) { + i.tasksInodeRefs.DecRef(func() { i.Destroy(ctx) }) +} + // staticFileSetStat implements a special static file that allows inode // attributes to be set. This is to support /proc files that are readonly, but // allow attributes to be set. +// +// +stateify savable type staticFileSetStat struct { dynamicBytesFileSetAttr vfs.StaticData diff --git a/pkg/sentry/fsimpl/proc/tasks_files.go b/pkg/sentry/fsimpl/proc/tasks_files.go index 7d8983aa5..01b7a6678 100644 --- a/pkg/sentry/fsimpl/proc/tasks_files.go +++ b/pkg/sentry/fsimpl/proc/tasks_files.go @@ -31,7 +31,9 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type selfSymlink struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeSymlink @@ -41,16 +43,13 @@ type selfSymlink struct { var _ kernfs.Inode = (*selfSymlink)(nil) -func (fs *filesystem) newSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { - inode := &selfSymlink{pidns: pidns} - inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) - - d := &kernfs.Dentry{} - d.Init(inode) - return d +func (i *tasksInode) newSelfSymlink(ctx context.Context, creds *auth.Credentials) kernfs.Inode { + inode := &selfSymlink{pidns: i.pidns} + inode.Init(ctx, creds, linux.UNNAMED_MAJOR, i.fs.devMinor, i.fs.NextIno(), linux.ModeSymlink|0777) + return inode } -func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { +func (s *selfSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { t := kernel.TaskFromContext(ctx) if t == nil { // Who is reading this link? @@ -63,17 +62,19 @@ func (s *selfSymlink) Readlink(ctx context.Context) (string, error) { return strconv.FormatUint(uint64(tgid), 10), nil } -func (s *selfSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { - target, err := s.Readlink(ctx) +func (s *selfSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx, mnt) return vfs.VirtualDentry{}, target, err } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*selfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } +// +stateify savable type threadSelfSymlink struct { + implStatFS kernfs.InodeAttrs kernfs.InodeNoopRefCount kernfs.InodeSymlink @@ -83,16 +84,13 @@ type threadSelfSymlink struct { var _ kernfs.Inode = (*threadSelfSymlink)(nil) -func (fs *filesystem) newThreadSelfSymlink(creds *auth.Credentials, ino uint64, pidns *kernel.PIDNamespace) *kernfs.Dentry { - inode := &threadSelfSymlink{pidns: pidns} - inode.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, ino, linux.ModeSymlink|0777) - - d := &kernfs.Dentry{} - d.Init(inode) - return d +func (i *tasksInode) newThreadSelfSymlink(ctx context.Context, creds *auth.Credentials) kernfs.Inode { + inode := &threadSelfSymlink{pidns: i.pidns} + inode.Init(ctx, creds, linux.UNNAMED_MAJOR, i.fs.devMinor, i.fs.NextIno(), linux.ModeSymlink|0777) + return inode } -func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { +func (s *threadSelfSymlink) Readlink(ctx context.Context, _ *vfs.Mount) (string, error) { t := kernel.TaskFromContext(ctx) if t == nil { // Who is reading this link? @@ -106,12 +104,12 @@ func (s *threadSelfSymlink) Readlink(ctx context.Context) (string, error) { return fmt.Sprintf("%d/task/%d", tgid, tid), nil } -func (s *threadSelfSymlink) Getlink(ctx context.Context, _ *vfs.Mount) (vfs.VirtualDentry, string, error) { - target, err := s.Readlink(ctx) +func (s *threadSelfSymlink) Getlink(ctx context.Context, mnt *vfs.Mount) (vfs.VirtualDentry, string, error) { + target, err := s.Readlink(ctx, mnt) return vfs.VirtualDentry{}, target, err } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } @@ -119,16 +117,20 @@ func (*threadSelfSymlink) SetStat(context.Context, *vfs.Filesystem, *auth.Creden // dynamicBytesFileSetAttr implements a special file that allows inode // attributes to be set. This is to support /proc files that are readonly, but // allow attributes to be set. +// +// +stateify savable type dynamicBytesFileSetAttr struct { kernfs.DynamicBytesFile } -// SetStat implements Inode.SetStat. +// SetStat implements kernfs.Inode.SetStat. func (d *dynamicBytesFileSetAttr) SetStat(ctx context.Context, fs *vfs.Filesystem, creds *auth.Credentials, opts vfs.SetStatOptions) error { return d.DynamicBytesFile.InodeAttrs.SetStat(ctx, fs, creds, opts) } // cpuStats contains the breakdown of CPU time for /proc/stat. +// +// +stateify savable type cpuStats struct { // user is time spent in userspace tasks with non-positive niceness. user uint64 diff --git a/pkg/sentry/fsimpl/proc/tasks_sys.go b/pkg/sentry/fsimpl/proc/tasks_sys.go index 6435385ef..7c7afdcfa 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys.go @@ -27,9 +27,11 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type tcpMemDir int const ( @@ -38,92 +40,93 @@ const ( ) // newSysDir returns the dentry corresponding to /proc/sys directory. -func (fs *filesystem) newSysDir(root *auth.Credentials, k *kernel.Kernel) *kernfs.Dentry { - return kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ - "kernel": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ - "hostname": fs.newDentry(root, fs.NextIno(), 0444, &hostnameData{}), - "shmall": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMALL)), - "shmmax": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMMAX)), - "shmmni": fs.newDentry(root, fs.NextIno(), 0444, shmData(linux.SHMMNI)), +func (fs *filesystem) newSysDir(ctx context.Context, root *auth.Credentials, k *kernel.Kernel) kernfs.Inode { + return fs.newStaticDir(ctx, root, map[string]kernfs.Inode{ + "kernel": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{ + "hostname": fs.newInode(ctx, root, 0444, &hostnameData{}), + "shmall": fs.newInode(ctx, root, 0444, shmData(linux.SHMALL)), + "shmmax": fs.newInode(ctx, root, 0444, shmData(linux.SHMMAX)), + "shmmni": fs.newInode(ctx, root, 0444, shmData(linux.SHMMNI)), }), - "vm": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ - "mmap_min_addr": fs.newDentry(root, fs.NextIno(), 0444, &mmapMinAddrData{k: k}), - "overcommit_memory": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0\n")), + "vm": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{ + "mmap_min_addr": fs.newInode(ctx, root, 0444, &mmapMinAddrData{k: k}), + "overcommit_memory": fs.newInode(ctx, root, 0444, newStaticFile("0\n")), }), - "net": fs.newSysNetDir(root, k), + "net": fs.newSysNetDir(ctx, root, k), }) } // newSysNetDir returns the dentry corresponding to /proc/sys/net directory. -func (fs *filesystem) newSysNetDir(root *auth.Credentials, k *kernel.Kernel) *kernfs.Dentry { - var contents map[string]*kernfs.Dentry +func (fs *filesystem) newSysNetDir(ctx context.Context, root *auth.Credentials, k *kernel.Kernel) kernfs.Inode { + var contents map[string]kernfs.Inode // TODO(gvisor.dev/issue/1833): Support for using the network stack in the // network namespace of the calling process. if stack := k.RootNetworkNamespace().Stack(); stack != nil { - contents = map[string]*kernfs.Dentry{ - "ipv4": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ - "tcp_recovery": fs.newDentry(root, fs.NextIno(), 0644, &tcpRecoveryData{stack: stack}), - "tcp_rmem": fs.newDentry(root, fs.NextIno(), 0644, &tcpMemData{stack: stack, dir: tcpRMem}), - "tcp_sack": fs.newDentry(root, fs.NextIno(), 0644, &tcpSackData{stack: stack}), - "tcp_wmem": fs.newDentry(root, fs.NextIno(), 0644, &tcpMemData{stack: stack, dir: tcpWMem}), + contents = map[string]kernfs.Inode{ + "ipv4": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{ + "tcp_recovery": fs.newInode(ctx, root, 0644, &tcpRecoveryData{stack: stack}), + "tcp_rmem": fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpRMem}), + "tcp_sack": fs.newInode(ctx, root, 0644, &tcpSackData{stack: stack}), + "tcp_wmem": fs.newInode(ctx, root, 0644, &tcpMemData{stack: stack, dir: tcpWMem}), + "ip_forward": fs.newInode(ctx, root, 0444, &ipForwarding{stack: stack}), // The following files are simple stubs until they are implemented in // netstack, most of these files are configuration related. We use the // value closest to the actual netstack behavior or any empty file, all // of these files will have mode 0444 (read-only for all users). - "ip_local_port_range": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("16000 65535")), - "ip_local_reserved_ports": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), - "ipfrag_time": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("30")), - "ip_nonlocal_bind": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "ip_no_pmtu_disc": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), + "ip_local_port_range": fs.newInode(ctx, root, 0444, newStaticFile("16000 65535")), + "ip_local_reserved_ports": fs.newInode(ctx, root, 0444, newStaticFile("")), + "ipfrag_time": fs.newInode(ctx, root, 0444, newStaticFile("30")), + "ip_nonlocal_bind": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "ip_no_pmtu_disc": fs.newInode(ctx, root, 0444, newStaticFile("1")), // tcp_allowed_congestion_control tell the user what they are able to // do as an unprivledged process so we leave it empty. - "tcp_allowed_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), - "tcp_available_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("reno")), - "tcp_congestion_control": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("reno")), + "tcp_allowed_congestion_control": fs.newInode(ctx, root, 0444, newStaticFile("")), + "tcp_available_congestion_control": fs.newInode(ctx, root, 0444, newStaticFile("reno")), + "tcp_congestion_control": fs.newInode(ctx, root, 0444, newStaticFile("reno")), // Many of the following stub files are features netstack doesn't // support. The unsupported features return "0" to indicate they are // disabled. - "tcp_base_mss": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1280")), - "tcp_dsack": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_early_retrans": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_fack": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_fastopen": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_fastopen_key": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("")), - "tcp_invalid_ratelimit": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_intvl": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_probes": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_keepalive_time": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("7200")), - "tcp_mtu_probing": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_no_metrics_save": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), - "tcp_probe_interval": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_probe_threshold": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "tcp_retries1": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("3")), - "tcp_retries2": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("15")), - "tcp_rfc1337": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), - "tcp_slow_start_after_idle": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), - "tcp_synack_retries": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("5")), - "tcp_syn_retries": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("3")), - "tcp_timestamps": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("1")), + "tcp_base_mss": fs.newInode(ctx, root, 0444, newStaticFile("1280")), + "tcp_dsack": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_early_retrans": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_fack": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_fastopen": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_fastopen_key": fs.newInode(ctx, root, 0444, newStaticFile("")), + "tcp_invalid_ratelimit": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_keepalive_intvl": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_keepalive_probes": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_keepalive_time": fs.newInode(ctx, root, 0444, newStaticFile("7200")), + "tcp_mtu_probing": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_no_metrics_save": fs.newInode(ctx, root, 0444, newStaticFile("1")), + "tcp_probe_interval": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_probe_threshold": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "tcp_retries1": fs.newInode(ctx, root, 0444, newStaticFile("3")), + "tcp_retries2": fs.newInode(ctx, root, 0444, newStaticFile("15")), + "tcp_rfc1337": fs.newInode(ctx, root, 0444, newStaticFile("1")), + "tcp_slow_start_after_idle": fs.newInode(ctx, root, 0444, newStaticFile("1")), + "tcp_synack_retries": fs.newInode(ctx, root, 0444, newStaticFile("5")), + "tcp_syn_retries": fs.newInode(ctx, root, 0444, newStaticFile("3")), + "tcp_timestamps": fs.newInode(ctx, root, 0444, newStaticFile("1")), }), - "core": kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, map[string]*kernfs.Dentry{ - "default_qdisc": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("pfifo_fast")), - "message_burst": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("10")), - "message_cost": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("5")), - "optmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("0")), - "rmem_default": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), - "rmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), - "somaxconn": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("128")), - "wmem_default": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), - "wmem_max": fs.newDentry(root, fs.NextIno(), 0444, newStaticFile("212992")), + "core": fs.newStaticDir(ctx, root, map[string]kernfs.Inode{ + "default_qdisc": fs.newInode(ctx, root, 0444, newStaticFile("pfifo_fast")), + "message_burst": fs.newInode(ctx, root, 0444, newStaticFile("10")), + "message_cost": fs.newInode(ctx, root, 0444, newStaticFile("5")), + "optmem_max": fs.newInode(ctx, root, 0444, newStaticFile("0")), + "rmem_default": fs.newInode(ctx, root, 0444, newStaticFile("212992")), + "rmem_max": fs.newInode(ctx, root, 0444, newStaticFile("212992")), + "somaxconn": fs.newInode(ctx, root, 0444, newStaticFile("128")), + "wmem_default": fs.newInode(ctx, root, 0444, newStaticFile("212992")), + "wmem_max": fs.newInode(ctx, root, 0444, newStaticFile("212992")), }), } } - return kernfs.NewStaticDir(root, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), 0555, contents) + return fs.newStaticDir(ctx, root, contents) } // mmapMinAddrData implements vfs.DynamicBytesSource for @@ -174,7 +177,7 @@ type tcpSackData struct { var _ vfs.WritableDynamicBytesSource = (*tcpSackData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpSackData) Generate(ctx context.Context, buf *bytes.Buffer) error { if d.enabled == nil { sack, err := d.stack.TCPSACKEnabled() @@ -232,7 +235,7 @@ type tcpRecoveryData struct { var _ vfs.WritableDynamicBytesSource = (*tcpRecoveryData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpRecoveryData) Generate(ctx context.Context, buf *bytes.Buffer) error { recovery, err := d.stack.TCPRecovery() if err != nil { @@ -284,7 +287,7 @@ type tcpMemData struct { var _ vfs.WritableDynamicBytesSource = (*tcpMemData)(nil) -// Generate implements vfs.DynamicBytesSource. +// Generate implements vfs.DynamicBytesSource.Generate. func (d *tcpMemData) Generate(ctx context.Context, buf *bytes.Buffer) error { d.mu.Lock() defer d.mu.Unlock() @@ -354,3 +357,63 @@ func (d *tcpMemData) writeSizeLocked(size inet.TCPBufferSize) error { panic(fmt.Sprintf("unknown tcpMemFile type: %v", d.dir)) } } + +// ipForwarding implements vfs.WritableDynamicBytesSource for +// /proc/sys/net/ipv4/ip_forwarding. +// +// +stateify savable +type ipForwarding struct { + kernfs.DynamicBytesFile + + stack inet.Stack `state:"wait"` + enabled *bool +} + +var _ vfs.WritableDynamicBytesSource = (*ipForwarding)(nil) + +// Generate implements vfs.DynamicBytesSource.Generate. +func (ipf *ipForwarding) Generate(ctx context.Context, buf *bytes.Buffer) error { + if ipf.enabled == nil { + enabled := ipf.stack.Forwarding(ipv4.ProtocolNumber) + ipf.enabled = &enabled + } + + val := "0\n" + if *ipf.enabled { + // 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" + } + buf.WriteString(val) + + return nil +} + +// Write implements vfs.WritableDynamicBytesSource.Write. +func (ipf *ipForwarding) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) { + if offset != 0 { + // No need to handle partial writes thus far. + return 0, syserror.EINVAL + } + if src.NumBytes() == 0 { + return 0, nil + } + + // Limit input size so as not to impact performance if input size is large. + 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 0, err + } + if ipf.enabled == nil { + ipf.enabled = new(bool) + } + *ipf.enabled = v != 0 + if err := ipf.stack.SetForwarding(ipv4.ProtocolNumber, *ipf.enabled); err != nil { + return 0, err + } + return n, nil +} diff --git a/pkg/sentry/fsimpl/proc/tasks_sys_test.go b/pkg/sentry/fsimpl/proc/tasks_sys_test.go index be54897bb..6cee22823 100644 --- a/pkg/sentry/fsimpl/proc/tasks_sys_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_sys_test.go @@ -20,8 +20,10 @@ import ( "testing" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/contexttest" "gvisor.dev/gvisor/pkg/sentry/inet" + "gvisor.dev/gvisor/pkg/usermem" ) func newIPv6TestStack() *inet.TestStack { @@ -76,3 +78,72 @@ func TestIfinet6(t *testing.T) { t.Errorf("Got n.contents() = %v, want = %v", got, want) } } + +// TestIPForwarding tests the implementation of +// /proc/sys/net/ipv4/ip_forwarding +func TestConfigureIPForwarding(t *testing.T) { + ctx := context.Background() + s := inet.NewTestStack() + + var cases = []struct { + comment string + initial bool + str string + final bool + }{ + { + comment: `Forwarding is disabled; write 1 and enable forwarding`, + initial: false, + str: "1", + final: true, + }, + { + comment: `Forwarding is disabled; write 0 and disable forwarding`, + initial: false, + str: "0", + final: false, + }, + { + comment: `Forwarding is enabled; write 1 and enable forwarding`, + initial: true, + str: "1", + final: true, + }, + { + comment: `Forwarding is enabled; write 0 and disable forwarding`, + initial: true, + str: "0", + final: false, + }, + { + comment: `Forwarding is disabled; write 2404 and enable forwarding`, + initial: false, + str: "2404", + final: true, + }, + { + comment: `Forwarding is enabled; write 2404 and enable forwarding`, + initial: true, + str: "2404", + final: true, + }, + } + for _, c := range cases { + t.Run(c.comment, func(t *testing.T) { + s.IPForwarding = c.initial + + file := &ipForwarding{stack: s, enabled: &c.initial} + + // Write the values. + src := usermem.BytesIOSequence([]byte(c.str)) + if n, err := file.Write(ctx, src, 0); n != int64(len(c.str)) || err != nil { + t.Errorf("file.Write(ctx, nil, %q, 0) = (%d, %v); want (%d, nil)", c.str, n, err, len(c.str)) + } + + // Read the values from the stack and check them. + if got, want := s.IPForwarding, c.final; got != want { + t.Errorf("s.IPForwarding incorrect; got: %v, want: %v", got, want) + } + }) + } +} diff --git a/pkg/sentry/fsimpl/proc/tasks_test.go b/pkg/sentry/fsimpl/proc/tasks_test.go index 3c9297dee..2582ababd 100644 --- a/pkg/sentry/fsimpl/proc/tasks_test.go +++ b/pkg/sentry/fsimpl/proc/tasks_test.go @@ -67,6 +67,7 @@ var ( taskStaticFiles = map[string]testutil.DirentType{ "auxv": linux.DT_REG, "cgroup": linux.DT_REG, + "cwd": linux.DT_LNK, "cmdline": linux.DT_REG, "comm": linux.DT_REG, "environ": linux.DT_REG, @@ -104,13 +105,16 @@ func setup(t *testing.T) *testutil.System { AllowUserMount: true, }) - mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", tmpfs.Name, &vfs.GetFilesystemOptions{}) + mntns, err := k.VFS().NewMountNamespace(ctx, creds, "", tmpfs.Name, &vfs.MountOptions{}) if err != nil { t.Fatalf("NewMountNamespace(): %v", err) } + root := mntns.Root() + root.IncRef() + defer root.DecRef(ctx) pop := &vfs.PathOperation{ - Root: mntns.Root(), - Start: mntns.Root(), + Root: root, + Start: root, Path: fspath.Parse("/proc"), } if err := k.VFS().MkdirAt(ctx, creds, pop, &vfs.MkdirOptions{Mode: 0777}); err != nil { @@ -118,8 +122,8 @@ func setup(t *testing.T) *testutil.System { } pop = &vfs.PathOperation{ - Root: mntns.Root(), - Start: mntns.Root(), + Root: root, + Start: root, Path: fspath.Parse("/proc"), } mntOpts := &vfs.MountOptions{ @@ -132,7 +136,7 @@ func setup(t *testing.T) *testutil.System { }, }, } - if err := k.VFS().MountAt(ctx, creds, "", pop, Name, mntOpts); err != nil { + if _, err := k.VFS().MountAt(ctx, creds, "", pop, Name, mntOpts); err != nil { t.Fatalf("MountAt(/proc): %v", err) } return testutil.NewSystem(ctx, t, k.VFS(), mntns) diff --git a/pkg/sentry/fsimpl/signalfd/BUILD b/pkg/sentry/fsimpl/signalfd/BUILD index 067c1657f..adb610213 100644 --- a/pkg/sentry/fsimpl/signalfd/BUILD +++ b/pkg/sentry/fsimpl/signalfd/BUILD @@ -8,7 +8,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/sentry/kernel", "//pkg/sentry/vfs", diff --git a/pkg/sentry/fsimpl/signalfd/signalfd.go b/pkg/sentry/fsimpl/signalfd/signalfd.go index 6297e1df4..10f1452ef 100644 --- a/pkg/sentry/fsimpl/signalfd/signalfd.go +++ b/pkg/sentry/fsimpl/signalfd/signalfd.go @@ -16,7 +16,6 @@ package signalfd import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -26,7 +25,9 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// SignalFileDescription implements FileDescriptionImpl for signal fds. +// SignalFileDescription implements vfs.FileDescriptionImpl for signal fds. +// +// +stateify savable type SignalFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -43,7 +44,7 @@ type SignalFileDescription struct { target *kernel.Task // mu protects mask. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // mask is the signal mask. Protected by mu. mask linux.SignalSet @@ -83,7 +84,7 @@ func (sfd *SignalFileDescription) SetMask(mask linux.SignalSet) { sfd.mask = mask } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequence, _ vfs.ReadOptions) (int64, error) { // Attempt to dequeue relevant signals. info, err := sfd.target.Sigtimedwait(sfd.Mask(), 0) @@ -93,8 +94,7 @@ func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequen } // Copy out the signal info using the specified format. - var buf [128]byte - binary.Marshal(buf[:0], usermem.ByteOrder, &linux.SignalfdSiginfo{ + infoNative := linux.SignalfdSiginfo{ Signo: uint32(info.Signo), Errno: info.Errno, Code: info.Code, @@ -103,9 +103,13 @@ func (sfd *SignalFileDescription) Read(ctx context.Context, dst usermem.IOSequen Status: info.Status(), Overrun: uint32(info.Overrun()), Addr: info.Addr(), - }) - n, err := dst.CopyOut(ctx, buf[:]) - return int64(n), err + } + n, err := infoNative.WriteTo(dst.Writer(ctx)) + if err == usermem.ErrEndOfIOSequence { + // Partial copy-out ok. + err = nil + } + return n, err } // Readiness implements waiter.Waitable.Readiness. @@ -132,5 +136,5 @@ func (sfd *SignalFileDescription) EventUnregister(entry *waiter.Entry) { sfd.target.SignalUnregister(entry) } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (sfd *SignalFileDescription) Release(context.Context) {} diff --git a/pkg/sentry/fsimpl/sockfs/sockfs.go b/pkg/sentry/fsimpl/sockfs/sockfs.go index c61818ff6..fda1fa942 100644 --- a/pkg/sentry/fsimpl/sockfs/sockfs.go +++ b/pkg/sentry/fsimpl/sockfs/sockfs.go @@ -28,14 +28,16 @@ import ( ) // filesystemType implements vfs.FilesystemType. +// +// +stateify savable type filesystemType struct{} -// GetFilesystem implements FilesystemType.GetFilesystem. +// GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fsType filesystemType) GetFilesystem(_ context.Context, vfsObj *vfs.VirtualFilesystem, _ *auth.Credentials, _ string, _ vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { panic("sockfs.filesystemType.GetFilesystem should never be called") } -// Name implements FilesystemType.Name. +// Name implements vfs.FilesystemType.Name. // // Note that registering sockfs is unnecessary, except for the fact that it // will not show up under /proc/filesystems as a result. This is a very minor @@ -44,6 +46,10 @@ func (filesystemType) Name() string { return "sockfs" } +// Release implements vfs.FilesystemType.Release. +func (filesystemType) Release(ctx context.Context) {} + +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -80,30 +86,37 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } // inode implements kernfs.Inode. +// +// +stateify savable type inode struct { - kernfs.InodeNotDirectory - kernfs.InodeNotSymlink kernfs.InodeAttrs kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink } // Open implements kernfs.Inode.Open. -func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { +func (i *inode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { return nil, syserror.ENXIO } +// StatFS implements kernfs.Inode.StatFS. +func (i *inode) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.SOCKFS_MAGIC), nil +} + // NewDentry constructs and returns a sockfs dentry. // // Preconditions: mnt.Filesystem() must have been returned by NewFilesystem(). -func NewDentry(creds *auth.Credentials, mnt *vfs.Mount) *vfs.Dentry { +func NewDentry(ctx context.Context, mnt *vfs.Mount) *vfs.Dentry { fs := mnt.Filesystem().Impl().(*filesystem) // File mode matches net/socket.c:sock_alloc. filemode := linux.FileMode(linux.S_IFSOCK | 0600) i := &inode{} - i.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.Filesystem.NextIno(), filemode) + i.InodeAttrs.Init(ctx, auth.CredentialsFromContext(ctx), linux.UNNAMED_MAJOR, fs.devMinor, fs.Filesystem.NextIno(), filemode) d := &kernfs.Dentry{} - d.Init(i) + d.Init(&fs.Filesystem, i) return d.VFSDentry() } diff --git a/pkg/sentry/fsimpl/sys/BUILD b/pkg/sentry/fsimpl/sys/BUILD index 1b548ccd4..09043b572 100644 --- a/pkg/sentry/fsimpl/sys/BUILD +++ b/pkg/sentry/fsimpl/sys/BUILD @@ -1,21 +1,42 @@ load("//tools:defs.bzl", "go_library", "go_test") +load("//tools/go_generics:defs.bzl", "go_template_instance") licenses(["notice"]) +go_template_instance( + name = "dir_refs", + out = "dir_refs.go", + package = "sys", + prefix = "dir", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "dir", + }, +) + go_library( name = "sys", srcs = [ + "dir_refs.go", + "kcov.go", "sys.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", "//pkg/context", + "//pkg/coverage", + "//pkg/log", + "//pkg/refs", + "//pkg/refsvfs2", + "//pkg/sentry/arch", "//pkg/sentry/fsimpl/kernfs", "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", + "//pkg/sentry/memmap", "//pkg/sentry/vfs", "//pkg/syserror", + "//pkg/usermem", ], ) diff --git a/pkg/sentry/fsimpl/sys/kcov.go b/pkg/sentry/fsimpl/sys/kcov.go new file mode 100644 index 000000000..b13f141a8 --- /dev/null +++ b/pkg/sentry/fsimpl/sys/kcov.go @@ -0,0 +1,118 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +func (fs *filesystem) newKcovFile(ctx context.Context, creds *auth.Credentials) kernfs.Inode { + k := &kcovInode{} + k.InodeAttrs.Init(ctx, creds, 0, 0, fs.NextIno(), linux.S_IFREG|0600) + return k +} + +// kcovInode implements kernfs.Inode. +// +// +stateify savable +type kcovInode struct { + kernfs.InodeAttrs + kernfs.InodeNoopRefCount + kernfs.InodeNotDirectory + kernfs.InodeNotSymlink + implStatFS +} + +func (i *kcovInode) Open(ctx context.Context, rp *vfs.ResolvingPath, d *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + k := kernel.KernelFromContext(ctx) + if k == nil { + panic("KernelFromContext returned nil") + } + fd := &kcovFD{ + inode: i, + kcov: k.NewKcov(), + } + + if err := fd.vfsfd.Init(fd, opts.Flags, rp.Mount(), d.VFSDentry(), &vfs.FileDescriptionOptions{ + DenyPRead: true, + DenyPWrite: true, + }); err != nil { + return nil, err + } + return &fd.vfsfd, nil +} + +// +stateify savable +type kcovFD struct { + vfs.FileDescriptionDefaultImpl + vfs.NoLockFD + + vfsfd vfs.FileDescription + inode *kcovInode + kcov *kernel.Kcov +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (fd *kcovFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + cmd := uint32(args[1].Int()) + arg := args[2].Uint64() + switch uint32(cmd) { + case linux.KCOV_INIT_TRACE: + return 0, fd.kcov.InitTrace(arg) + case linux.KCOV_ENABLE: + return 0, fd.kcov.EnableTrace(ctx, uint8(arg)) + case linux.KCOV_DISABLE: + if arg != 0 { + // This arg is unused; it should be 0. + return 0, syserror.EINVAL + } + return 0, fd.kcov.DisableTrace(ctx) + default: + return 0, syserror.ENOTTY + } +} + +// ConfigureMmap implements vfs.FileDescriptionImpl.ConfigureMmap. +func (fd *kcovFD) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + return fd.kcov.ConfigureMMap(ctx, opts) +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (fd *kcovFD) Release(ctx context.Context) { + // kcov instances have reference counts in Linux, but this seems sufficient + // for our purposes. + fd.kcov.Clear(ctx) +} + +// SetStat implements vfs.FileDescriptionImpl.SetStat. +func (fd *kcovFD) SetStat(ctx context.Context, opts vfs.SetStatOptions) error { + creds := auth.CredentialsFromContext(ctx) + fs := fd.vfsfd.VirtualDentry().Mount().Filesystem() + return fd.inode.SetStat(ctx, fs, creds, opts) +} + +// Stat implements vfs.FileDescriptionImpl.Stat. +func (fd *kcovFD) Stat(ctx context.Context, opts vfs.StatOptions) (linux.Statx, error) { + return fd.inode.Stat(ctx, fd.vfsfd.Mount().Filesystem(), opts) +} diff --git a/pkg/sentry/fsimpl/sys/sys.go b/pkg/sentry/fsimpl/sys/sys.go index 0401726b6..7d2147141 100644 --- a/pkg/sentry/fsimpl/sys/sys.go +++ b/pkg/sentry/fsimpl/sys/sys.go @@ -18,9 +18,11 @@ package sys import ( "bytes" "fmt" + "strconv" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/coverage" "gvisor.dev/gvisor/pkg/sentry/fsimpl/kernfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -28,14 +30,21 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) -// Name is the default filesystem name. -const Name = "sysfs" -const defaultSysDirMode = linux.FileMode(0755) +const ( + // Name is the default filesystem name. + Name = "sysfs" + defaultSysDirMode = linux.FileMode(0755) + defaultMaxCachedDentries = uint64(1000) +) // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { kernfs.Filesystem @@ -47,6 +56,9 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { devMinor, err := vfsObj.GetAnonBlockDevMinor() @@ -54,44 +66,73 @@ func (fsType FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt return nil, nil, err } + mopts := vfs.GenericParseMountOptions(opts.Data) + maxCachedDentries := defaultMaxCachedDentries + if str, ok := mopts["dentry_cache_limit"]; ok { + delete(mopts, "dentry_cache_limit") + maxCachedDentries, err = strconv.ParseUint(str, 10, 64) + if err != nil { + ctx.Warningf("sys.FilesystemType.GetFilesystem: invalid dentry cache limit: dentry_cache_limit=%s", str) + return nil, nil, syserror.EINVAL + } + } + fs := &filesystem{ devMinor: devMinor, } + fs.MaxCachedDentries = maxCachedDentries fs.VFSFilesystem().Init(vfsObj, &fsType, fs) - root := fs.newDir(creds, defaultSysDirMode, map[string]*kernfs.Dentry{ - "block": fs.newDir(creds, defaultSysDirMode, nil), - "bus": fs.newDir(creds, defaultSysDirMode, nil), - "class": fs.newDir(creds, defaultSysDirMode, map[string]*kernfs.Dentry{ - "power_supply": fs.newDir(creds, defaultSysDirMode, nil), + root := fs.newDir(ctx, creds, defaultSysDirMode, map[string]kernfs.Inode{ + "block": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "bus": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "class": fs.newDir(ctx, creds, defaultSysDirMode, map[string]kernfs.Inode{ + "power_supply": fs.newDir(ctx, creds, defaultSysDirMode, nil), }), - "dev": fs.newDir(creds, defaultSysDirMode, nil), - "devices": fs.newDir(creds, defaultSysDirMode, map[string]*kernfs.Dentry{ - "system": fs.newDir(creds, defaultSysDirMode, map[string]*kernfs.Dentry{ + "dev": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "devices": fs.newDir(ctx, creds, defaultSysDirMode, map[string]kernfs.Inode{ + "system": fs.newDir(ctx, creds, defaultSysDirMode, map[string]kernfs.Inode{ "cpu": cpuDir(ctx, fs, creds), }), }), - "firmware": fs.newDir(creds, defaultSysDirMode, nil), - "fs": fs.newDir(creds, defaultSysDirMode, nil), - "kernel": fs.newDir(creds, defaultSysDirMode, nil), - "module": fs.newDir(creds, defaultSysDirMode, nil), - "power": fs.newDir(creds, defaultSysDirMode, nil), + "firmware": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "fs": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "kernel": kernelDir(ctx, fs, creds), + "module": fs.newDir(ctx, creds, defaultSysDirMode, nil), + "power": fs.newDir(ctx, creds, defaultSysDirMode, nil), }) - return fs.VFSFilesystem(), root.VFSDentry(), nil + var rootD kernfs.Dentry + rootD.Init(&fs.Filesystem, root) + return fs.VFSFilesystem(), rootD.VFSDentry(), nil } -func cpuDir(ctx context.Context, fs *filesystem, creds *auth.Credentials) *kernfs.Dentry { +func cpuDir(ctx context.Context, fs *filesystem, creds *auth.Credentials) kernfs.Inode { k := kernel.KernelFromContext(ctx) maxCPUCores := k.ApplicationCores() - children := map[string]*kernfs.Dentry{ - "online": fs.newCPUFile(creds, maxCPUCores, linux.FileMode(0444)), - "possible": fs.newCPUFile(creds, maxCPUCores, linux.FileMode(0444)), - "present": fs.newCPUFile(creds, maxCPUCores, linux.FileMode(0444)), + children := map[string]kernfs.Inode{ + "online": fs.newCPUFile(ctx, creds, maxCPUCores, linux.FileMode(0444)), + "possible": fs.newCPUFile(ctx, creds, maxCPUCores, linux.FileMode(0444)), + "present": fs.newCPUFile(ctx, creds, maxCPUCores, linux.FileMode(0444)), } for i := uint(0); i < maxCPUCores; i++ { - children[fmt.Sprintf("cpu%d", i)] = fs.newDir(creds, linux.FileMode(0555), nil) + children[fmt.Sprintf("cpu%d", i)] = fs.newDir(ctx, creds, linux.FileMode(0555), nil) } - return fs.newDir(creds, defaultSysDirMode, children) + return fs.newDir(ctx, creds, defaultSysDirMode, children) +} + +func kernelDir(ctx context.Context, fs *filesystem, creds *auth.Credentials) kernfs.Inode { + // If kcov is available, set up /sys/kernel/debug/kcov. Technically, debugfs + // should be mounted at debug/, but for our purposes, it is sufficient to + // keep it in sys. + var children map[string]kernfs.Inode + if coverage.KcovAvailable() { + children = map[string]kernfs.Inode{ + "debug": fs.newDir(ctx, creds, linux.FileMode(0700), map[string]kernfs.Inode{ + "kcov": fs.newKcovFile(ctx, creds), + }), + } + } + return fs.newDir(ctx, creds, defaultSysDirMode, children) } // Release implements vfs.FilesystemImpl.Release. @@ -101,46 +142,62 @@ func (fs *filesystem) Release(ctx context.Context) { } // dir implements kernfs.Inode. +// +// +stateify savable type dir struct { + dirRefs + kernfs.InodeAlwaysValid kernfs.InodeAttrs - kernfs.InodeNoDynamicLookup kernfs.InodeNotSymlink kernfs.InodeDirectoryNoNewChildren + kernfs.InodeTemporary kernfs.OrderedChildren locks vfs.FileLocks - - dentry kernfs.Dentry } -func (fs *filesystem) newDir(creds *auth.Credentials, mode linux.FileMode, contents map[string]*kernfs.Dentry) *kernfs.Dentry { +func (fs *filesystem) newDir(ctx context.Context, creds *auth.Credentials, mode linux.FileMode, contents map[string]kernfs.Inode) kernfs.Inode { d := &dir{} - d.InodeAttrs.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) + d.InodeAttrs.Init(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), linux.ModeDirectory|0755) d.OrderedChildren.Init(kernfs.OrderedChildrenOptions{}) - d.dentry.Init(d) - - d.IncLinks(d.OrderedChildren.Populate(&d.dentry, contents)) - - return &d.dentry + d.EnableLeakCheck() + d.IncLinks(d.OrderedChildren.Populate(contents)) + return d } -// SetStat implements Inode.SetStat not allowing inode attributes to be changed. +// SetStat implements kernfs.Inode.SetStat not allowing inode attributes to be changed. func (*dir) SetStat(context.Context, *vfs.Filesystem, *auth.Credentials, vfs.SetStatOptions) error { return syserror.EPERM } // Open implements kernfs.Inode.Open. -func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, vfsd *vfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), vfsd, &d.OrderedChildren, &d.locks, &opts) +func (d *dir) Open(ctx context.Context, rp *vfs.ResolvingPath, kd *kernfs.Dentry, opts vfs.OpenOptions) (*vfs.FileDescription, error) { + fd, err := kernfs.NewGenericDirectoryFD(rp.Mount(), kd, &d.OrderedChildren, &d.locks, &opts, kernfs.GenericDirectoryFDOptions{ + SeekEnd: kernfs.SeekEndStaticEntries, + }) if err != nil { return nil, err } return fd.VFSFileDescription(), nil } +// DecRef implements kernfs.Inode.DecRef. +func (d *dir) DecRef(ctx context.Context) { + d.dirRefs.DecRef(func() { d.Destroy(ctx) }) +} + +// StatFS implements kernfs.Inode.StatFS. +func (d *dir) StatFS(ctx context.Context, fs *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.SYSFS_MAGIC), nil +} + // cpuFile implements kernfs.Inode. +// +// +stateify savable type cpuFile struct { + implStatFS kernfs.DynamicBytesFile + maxCores uint } @@ -150,10 +207,16 @@ func (c *cpuFile) Generate(ctx context.Context, buf *bytes.Buffer) error { return nil } -func (fs *filesystem) newCPUFile(creds *auth.Credentials, maxCores uint, mode linux.FileMode) *kernfs.Dentry { +func (fs *filesystem) newCPUFile(ctx context.Context, creds *auth.Credentials, maxCores uint, mode linux.FileMode) kernfs.Inode { c := &cpuFile{maxCores: maxCores} - c.DynamicBytesFile.Init(creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), c, mode) - d := &kernfs.Dentry{} - d.Init(c) - return d + c.DynamicBytesFile.Init(ctx, creds, linux.UNNAMED_MAJOR, fs.devMinor, fs.NextIno(), c, mode) + return c +} + +// +stateify savable +type implStatFS struct{} + +// StatFS implements kernfs.Inode.StatFS. +func (*implStatFS) StatFS(context.Context, *vfs.Filesystem) (linux.Statfs, error) { + return vfs.GenericStatFS(linux.SYSFS_MAGIC), nil } diff --git a/pkg/sentry/fsimpl/sys/sys_test.go b/pkg/sentry/fsimpl/sys/sys_test.go index 9fd38b295..0a0d914cc 100644 --- a/pkg/sentry/fsimpl/sys/sys_test.go +++ b/pkg/sentry/fsimpl/sys/sys_test.go @@ -38,7 +38,7 @@ func newTestSystem(t *testing.T) *testutil.System { AllowUserMount: true, }) - mns, err := k.VFS().NewMountNamespace(ctx, creds, "", sys.Name, &vfs.GetFilesystemOptions{}) + mns, err := k.VFS().NewMountNamespace(ctx, creds, "", sys.Name, &vfs.MountOptions{}) if err != nil { t.Fatalf("Failed to create new mount namespace: %v", err) } diff --git a/pkg/sentry/fsimpl/testutil/kernel.go b/pkg/sentry/fsimpl/testutil/kernel.go index 1813269e0..738c0c9cc 100644 --- a/pkg/sentry/fsimpl/testutil/kernel.go +++ b/pkg/sentry/fsimpl/testutil/kernel.go @@ -147,7 +147,12 @@ func CreateTask(ctx context.Context, name string, tc *kernel.ThreadGroup, mntns FSContext: kernel.NewFSContextVFS2(root, cwd, 0022), FDTable: k.NewFDTable(), } - return k.TaskSet().NewTask(config) + t, err := k.TaskSet().NewTask(ctx, config) + if err != nil { + config.ThreadGroup.Release(ctx) + return nil, err + } + return t, nil } func newFakeExecutable(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, root vfs.VirtualDentry) (*vfs.FileDescription, error) { diff --git a/pkg/sentry/fsimpl/testutil/testutil.go b/pkg/sentry/fsimpl/testutil/testutil.go index 568132121..1a8525b06 100644 --- a/pkg/sentry/fsimpl/testutil/testutil.go +++ b/pkg/sentry/fsimpl/testutil/testutil.go @@ -46,16 +46,18 @@ type System struct { // NewSystem constructs a System. // -// Precondition: Caller must hold a reference on MntNs, whose ownership +// Precondition: Caller must hold a reference on mns, whose ownership // is transferred to the new System. func NewSystem(ctx context.Context, t *testing.T, v *vfs.VirtualFilesystem, mns *vfs.MountNamespace) *System { + root := mns.Root() + root.IncRef() s := &System{ t: t, Ctx: ctx, Creds: auth.CredentialsFromContext(ctx), VFS: v, MntNs: mns, - Root: mns.Root(), + Root: root, } return s } @@ -254,10 +256,10 @@ func (d *DirentCollector) Contains(name string, typ uint8) error { defer d.mu.Unlock() dirent, ok := d.dirents[name] if !ok { - return fmt.Errorf("No dirent named %q found", name) + return fmt.Errorf("no dirent named %q found", name) } if dirent.Type != typ { - return fmt.Errorf("Dirent named %q found, but was expecting type %s, got: %+v", name, linux.DirentType.Parse(uint64(typ)), dirent) + return fmt.Errorf("dirent named %q found, but was expecting type %s, got: %+v", name, linux.DirentType.Parse(uint64(typ)), dirent) } return nil } diff --git a/pkg/sentry/fsimpl/timerfd/timerfd.go b/pkg/sentry/fsimpl/timerfd/timerfd.go index 86beaa0a8..8853c8ad2 100644 --- a/pkg/sentry/fsimpl/timerfd/timerfd.go +++ b/pkg/sentry/fsimpl/timerfd/timerfd.go @@ -26,8 +26,10 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) -// TimerFileDescription implements FileDescriptionImpl for timer fds. It also +// TimerFileDescription implements vfs.FileDescriptionImpl for timer fds. It also // implements ktime.TimerListener. +// +// +stateify savable type TimerFileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -62,7 +64,7 @@ func New(ctx context.Context, vfsObj *vfs.VirtualFilesystem, clock ktime.Clock, return &tfd.vfsfd, nil } -// Read implements FileDescriptionImpl.Read. +// Read implements vfs.FileDescriptionImpl.Read. func (tfd *TimerFileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { const sizeofUint64 = 8 if dst.NumBytes() < sizeofUint64 { @@ -128,7 +130,7 @@ func (tfd *TimerFileDescription) ResumeTimer() { tfd.timer.Resume() } -// Release implements FileDescriptionImpl.Release() +// Release implements vfs.FileDescriptionImpl.Release. func (tfd *TimerFileDescription) Release(context.Context) { tfd.timer.Destroy() } diff --git a/pkg/sentry/fsimpl/tmpfs/BUILD b/pkg/sentry/fsimpl/tmpfs/BUILD index 5cd428d64..fe520b6fd 100644 --- a/pkg/sentry/fsimpl/tmpfs/BUILD +++ b/pkg/sentry/fsimpl/tmpfs/BUILD @@ -31,7 +31,7 @@ go_template_instance( out = "inode_refs.go", package = "tmpfs", prefix = "inode", - template = "//pkg/refs_vfs2:refs_template", + template = "//pkg/refsvfs2:refs_template", types = { "T": "inode", }, @@ -48,6 +48,7 @@ go_library( "inode_refs.go", "named_pipe.go", "regular_file.go", + "save_restore.go", "socket_file.go", "symlink.go", "tmpfs.go", @@ -60,6 +61,7 @@ go_library( "//pkg/fspath", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs", diff --git a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go index d263147c2..3cc63e732 100644 --- a/pkg/sentry/fsimpl/tmpfs/benchmark_test.go +++ b/pkg/sentry/fsimpl/tmpfs/benchmark_test.go @@ -182,7 +182,7 @@ func BenchmarkVFS2TmpfsStat(b *testing.B) { vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) if err != nil { b.Fatalf("failed to create tmpfs root mount: %v", err) } @@ -193,6 +193,7 @@ func BenchmarkVFS2TmpfsStat(b *testing.B) { // Create nested directories with given depth. root := mntns.Root() + root.IncRef() defer root.DecRef(ctx) vd := root vd.IncRef() @@ -376,7 +377,7 @@ func BenchmarkVFS2TmpfsMountStat(b *testing.B) { vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) if err != nil { b.Fatalf("failed to create tmpfs root mount: %v", err) } @@ -387,6 +388,7 @@ func BenchmarkVFS2TmpfsMountStat(b *testing.B) { // Create the mount point. root := mntns.Root() + root.IncRef() defer root.DecRef(ctx) pop := vfs.PathOperation{ Root: root, @@ -405,7 +407,7 @@ func BenchmarkVFS2TmpfsMountStat(b *testing.B) { } defer mountPoint.DecRef(ctx) // Create and mount the submount. - if err := vfsObj.MountAt(ctx, creds, "", &pop, "tmpfs", &vfs.MountOptions{}); err != nil { + if _, err := vfsObj.MountAt(ctx, creds, "", &pop, "tmpfs", &vfs.MountOptions{}); err != nil { b.Fatalf("failed to mount tmpfs submount: %v", err) } filePathBuilder.WriteString(mountPointName) diff --git a/pkg/sentry/fsimpl/tmpfs/device_file.go b/pkg/sentry/fsimpl/tmpfs/device_file.go index ac54d420d..9129d35b7 100644 --- a/pkg/sentry/fsimpl/tmpfs/device_file.go +++ b/pkg/sentry/fsimpl/tmpfs/device_file.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/vfs" ) +// +stateify savable type deviceFile struct { inode inode kind vfs.DeviceKind diff --git a/pkg/sentry/fsimpl/tmpfs/directory.go b/pkg/sentry/fsimpl/tmpfs/directory.go index 78b4fc5be..e90669cf0 100644 --- a/pkg/sentry/fsimpl/tmpfs/directory.go +++ b/pkg/sentry/fsimpl/tmpfs/directory.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/syserror" ) +// +stateify savable type directory struct { // Since directories can't be hard-linked, each directory can only be // associated with a single dentry, which we can store in the directory @@ -44,7 +45,7 @@ type directory struct { // (with inode == nil) that represent the iteration position of // directoryFDs. childList is used to support directoryFD.IterDirents() // efficiently. childList is protected by iterMu. - iterMu sync.Mutex + iterMu sync.Mutex `state:"nosave"` childList dentryList } @@ -57,8 +58,9 @@ func (fs *filesystem) newDirectory(kuid auth.KUID, kgid auth.KGID, mode linux.Fi return dir } -// Preconditions: filesystem.mu must be locked for writing. dir must not -// already contain a child with the given name. +// Preconditions: +// * filesystem.mu must be locked for writing. +// * dir must not already contain a child with the given name. func (dir *directory) insertChildLocked(child *dentry, name string) { child.parent = &dir.dentry child.name = name @@ -85,6 +87,7 @@ func (dir *directory) mayDelete(creds *auth.Credentials, child *dentry) error { return vfs.CheckDeleteSticky(creds, linux.FileMode(atomic.LoadUint32(&dir.inode.mode)), auth.KUID(atomic.LoadUint32(&child.inode.uid))) } +// +stateify savable type directoryFD struct { fileDescription vfs.DirectoryFileDescriptionDefaultImpl diff --git a/pkg/sentry/fsimpl/tmpfs/filesystem.go b/pkg/sentry/fsimpl/tmpfs/filesystem.go index b0ec177e6..e39cd305b 100644 --- a/pkg/sentry/fsimpl/tmpfs/filesystem.go +++ b/pkg/sentry/fsimpl/tmpfs/filesystem.go @@ -25,7 +25,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" ) // Sync implements vfs.FilesystemImpl.Sync. @@ -39,7 +38,9 @@ func (fs *filesystem) Sync(ctx context.Context) error { // // stepLocked is loosely analogous to fs/namei.c:walk_component(). // -// Preconditions: filesystem.mu must be locked. !rp.Done(). +// Preconditions: +// * filesystem.mu must be locked. +// * !rp.Done(). func stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry) (*dentry, error) { dir, ok := d.inode.impl.(*directory) if !ok { @@ -97,7 +98,9 @@ afterSymlink: // walkParentDirLocked is loosely analogous to Linux's // fs/namei.c:path_parentat(). // -// Preconditions: filesystem.mu must be locked. !rp.Done(). +// Preconditions: +// * filesystem.mu must be locked. +// * !rp.Done(). func walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry) (*directory, error) { for !rp.Final() { next, err := stepLocked(ctx, rp, d) @@ -139,8 +142,9 @@ func resolveLocked(ctx context.Context, rp *vfs.ResolvingPath) (*dentry, error) // doCreateAt is loosely analogous to a conjunction of Linux's // fs/namei.c:filename_create() and done_path_create(). // -// Preconditions: !rp.Done(). For the final path component in rp, -// !rp.ShouldFollowSymlink(). +// Preconditions: +// * !rp.Done(). +// * For the final path component in rp, !rp.ShouldFollowSymlink(). func (fs *filesystem) doCreateAt(ctx context.Context, rp *vfs.ResolvingPath, dir bool, create func(parentDir *directory, name string) error) error { fs.mu.Lock() defer fs.mu.Unlock() @@ -669,11 +673,11 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts fs.mu.RUnlock() return err } - if err := d.inode.setStat(ctx, rp.Credentials(), &opts); err != nil { - fs.mu.RUnlock() + err = d.inode.setStat(ctx, rp.Credentials(), &opts) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() if ev := vfs.InotifyEventFromStatMask(opts.Stat.Mask); ev != 0 { d.InotifyWithParent(ctx, ev, 0, vfs.InodeEvent) @@ -701,16 +705,7 @@ func (fs *filesystem) StatFSAt(ctx context.Context, rp *vfs.ResolvingPath) (linu if _, err := resolveLocked(ctx, rp); err != nil { return linux.Statfs{}, err } - statfs := linux.Statfs{ - Type: linux.TMPFS_MAGIC, - BlockSize: usermem.PageSize, - FragmentSize: usermem.PageSize, - NameLength: linux.NAME_MAX, - // TODO(b/29637826): Allow configuring a tmpfs size and enforce it. - Blocks: 0, - BlocksFree: 0, - } - return statfs, nil + return globalStatfs, nil } // SymlinkAt implements vfs.FilesystemImpl.SymlinkAt. @@ -775,7 +770,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return nil } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { fs.mu.RLock() defer fs.mu.RUnlock() @@ -788,65 +783,68 @@ func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath } switch impl := d.inode.impl.(type) { case *socketFile: + if impl.ep == nil { + return nil, syserror.ECONNREFUSED + } return impl.ep, nil default: return nil, syserror.ECONNREFUSED } } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { fs.mu.RLock() defer fs.mu.RUnlock() d, err := resolveLocked(ctx, rp) if err != nil { return nil, err } - return d.inode.listxattr(size) + return d.inode.listXattr(size) } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { fs.mu.RLock() defer fs.mu.RUnlock() d, err := resolveLocked(ctx, rp) if err != nil { return "", err } - return d.inode.getxattr(rp.Credentials(), &opts) + return d.inode.getXattr(rp.Credentials(), &opts) } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { fs.mu.RLock() d, err := resolveLocked(ctx, rp) if err != nil { fs.mu.RUnlock() return err } - if err := d.inode.setxattr(rp.Credentials(), &opts); err != nil { - fs.mu.RUnlock() + err = d.inode.setXattr(rp.Credentials(), &opts) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { fs.mu.RLock() d, err := resolveLocked(ctx, rp) if err != nil { fs.mu.RUnlock() return err } - if err := d.inode.removexattr(rp.Credentials(), name); err != nil { - fs.mu.RUnlock() + err = d.inode.removeXattr(rp.Credentials(), name) + fs.mu.RUnlock() + if err != nil { return err } - fs.mu.RUnlock() d.InotifyWithParent(ctx, linux.IN_ATTRIB, 0, vfs.InodeEvent) return nil @@ -867,8 +865,16 @@ func (fs *filesystem) PrependPath(ctx context.Context, vfsroot, vd vfs.VirtualDe } if d.parent == nil { if d.name != "" { - // This must be an anonymous memfd file. + // This file must have been created by + // newUnlinkedRegularFileDescription(). In Linux, + // mm/shmem.c:__shmem_file_setup() => + // fs/file_table.c:alloc_file_pseudo() sets the created + // dentry's dentry_operations to anon_ops, for which d_dname == + // simple_dname. fs/d_path.c:simple_dname() defines the + // dentry's pathname to be its name, prefixed with "/" and + // suffixed with " (deleted)". b.PrependComponent("/" + d.name) + b.AppendString(" (deleted)") return vfs.PrependPathSyntheticError{} } return vfs.PrependPathAtNonMountRootError{} diff --git a/pkg/sentry/fsimpl/tmpfs/named_pipe.go b/pkg/sentry/fsimpl/tmpfs/named_pipe.go index 739350cf0..d772db9e9 100644 --- a/pkg/sentry/fsimpl/tmpfs/named_pipe.go +++ b/pkg/sentry/fsimpl/tmpfs/named_pipe.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) +// +stateify savable type namedPipe struct { inode inode @@ -28,8 +29,8 @@ type namedPipe struct { } // Preconditions: -// * fs.mu must be locked. -// * rp.Mount().CheckBeginWrite() has been called successfully. +// * fs.mu must be locked. +// * rp.Mount().CheckBeginWrite() has been called successfully. func (fs *filesystem) newNamedPipe(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode { file := &namedPipe{pipe: pipe.NewVFSPipe(true /* isNamed */, pipe.DefaultPipeSize, usermem.PageSize)} file.inode.init(file, fs, kuid, kgid, linux.S_IFIFO|mode) diff --git a/pkg/sentry/fsimpl/tmpfs/pipe_test.go b/pkg/sentry/fsimpl/tmpfs/pipe_test.go index ec2701d8b..2f856ce36 100644 --- a/pkg/sentry/fsimpl/tmpfs/pipe_test.go +++ b/pkg/sentry/fsimpl/tmpfs/pipe_test.go @@ -158,13 +158,14 @@ func setup(t *testing.T) (context.Context, *auth.Credentials, *vfs.VirtualFilesy vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) if err != nil { t.Fatalf("failed to create tmpfs root mount: %v", err) } // Create the pipe. root := mntns.Root() + root.IncRef() pop := vfs.PathOperation{ Root: root, Start: root, diff --git a/pkg/sentry/fsimpl/tmpfs/regular_file.go b/pkg/sentry/fsimpl/tmpfs/regular_file.go index 0710b65db..98680fde9 100644 --- a/pkg/sentry/fsimpl/tmpfs/regular_file.go +++ b/pkg/sentry/fsimpl/tmpfs/regular_file.go @@ -36,11 +36,17 @@ import ( ) // regularFile is a regular (=S_IFREG) tmpfs file. +// +// +stateify savable type regularFile struct { inode inode // memFile is a platform.File used to allocate pages to this regularFile. - memFile *pgalloc.MemoryFile + memFile *pgalloc.MemoryFile `state:"nosave"` + + // memoryUsageKind is the memory accounting category under which pages backing + // this regularFile's contents are accounted. + memoryUsageKind usage.MemoryKind // mapsMu protects mappings. mapsMu sync.Mutex `state:"nosave"` @@ -62,7 +68,7 @@ type regularFile struct { writableMappingPages uint64 // dataMu protects the fields below. - dataMu sync.RWMutex + dataMu sync.RWMutex `state:"nosave"` // data maps offsets into the file to offsets into memFile that store // the file's data. @@ -86,14 +92,75 @@ type regularFile struct { func (fs *filesystem) newRegularFile(kuid auth.KUID, kgid auth.KGID, mode linux.FileMode) *inode { file := ®ularFile{ - memFile: fs.memFile, - seals: linux.F_SEAL_SEAL, + memFile: fs.mfp.MemoryFile(), + memoryUsageKind: usage.Tmpfs, + seals: linux.F_SEAL_SEAL, } file.inode.init(file, fs, kuid, kgid, linux.S_IFREG|mode) file.inode.nlink = 1 // from parent directory return &file.inode } +// newUnlinkedRegularFileDescription creates a regular file on the tmpfs +// filesystem represented by mount and returns an FD representing that file. +// The new file is not reachable by path traversal from any other file. +// +// newUnlinkedRegularFileDescription is analogous to Linux's +// mm/shmem.c:__shmem_file_setup(). +// +// Preconditions: mount must be a tmpfs mount. +func newUnlinkedRegularFileDescription(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, name string) (*regularFileFD, error) { + fs, ok := mount.Filesystem().Impl().(*filesystem) + if !ok { + panic("tmpfs.newUnlinkedRegularFileDescription() called with non-tmpfs mount") + } + + inode := fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, 0777) + d := fs.newDentry(inode) + defer d.DecRef(ctx) + d.name = name + + fd := ®ularFileFD{} + fd.Init(&inode.locks) + flags := uint32(linux.O_RDWR) + if err := fd.vfsfd.Init(fd, flags, mount, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { + return nil, err + } + return fd, nil +} + +// NewZeroFile creates a new regular file and file description as for +// mmap(MAP_SHARED | MAP_ANONYMOUS). The file has the given size and is +// initially (implicitly) filled with zeroes. +// +// Preconditions: mount must be a tmpfs mount. +func NewZeroFile(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, size uint64) (*vfs.FileDescription, error) { + // Compare mm/shmem.c:shmem_zero_setup(). + fd, err := newUnlinkedRegularFileDescription(ctx, creds, mount, "dev/zero") + if err != nil { + return nil, err + } + rf := fd.inode().impl.(*regularFile) + rf.memoryUsageKind = usage.Anonymous + rf.size = size + return &fd.vfsfd, err +} + +// NewMemfd creates a new regular file and file description as for +// memfd_create. +// +// Preconditions: mount must be a tmpfs mount. +func NewMemfd(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, allowSeals bool, name string) (*vfs.FileDescription, error) { + fd, err := newUnlinkedRegularFileDescription(ctx, creds, mount, name) + if err != nil { + return nil, err + } + if allowSeals { + fd.inode().impl.(*regularFile).seals = 0 + } + return &fd.vfsfd, nil +} + // truncate grows or shrinks the file to the given size. It returns true if the // file size was updated. func (rf *regularFile) truncate(newSize uint64) (bool, error) { @@ -226,7 +293,7 @@ func (rf *regularFile) Translate(ctx context.Context, required, optional memmap. optional.End = pgend } - cerr := rf.data.Fill(ctx, required, optional, rf.memFile, usage.Tmpfs, func(_ context.Context, dsts safemem.BlockSeq, _ uint64) (uint64, error) { + cerr := rf.data.Fill(ctx, required, optional, rf.size, rf.memFile, rf.memoryUsageKind, 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 }) @@ -260,13 +327,14 @@ func (*regularFile) InvalidateUnsavable(context.Context) error { return nil } +// +stateify savable type regularFileFD struct { fileDescription // off is the file offset. off is accessed using atomic memory operations. // offMu serializes operations that may mutate off. off int64 - offMu sync.Mutex + offMu sync.Mutex `state:"nosave"` } // Release implements vfs.FileDescriptionImpl.Release. @@ -575,7 +643,7 @@ func (rw *regularFileReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, case gap.Ok(): // Allocate memory for the write. gapMR := gap.Range().Intersect(pgMR) - fr, err := rw.file.memFile.Allocate(gapMR.Length(), usage.Tmpfs) + fr, err := rw.file.memFile.Allocate(gapMR.Length(), rw.file.memoryUsageKind) if err != nil { retErr = err goto exitLoop diff --git a/pkg/sentry/fsimpl/tmpfs/save_restore.go b/pkg/sentry/fsimpl/tmpfs/save_restore.go new file mode 100644 index 000000000..b27f75cc2 --- /dev/null +++ b/pkg/sentry/fsimpl/tmpfs/save_restore.go @@ -0,0 +1,20 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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 + +// afterLoad is called by stateify. +func (rf *regularFile) afterLoad() { + rf.memFile = rf.inode.fs.mfp.MemoryFile() +} diff --git a/pkg/sentry/fsimpl/tmpfs/socket_file.go b/pkg/sentry/fsimpl/tmpfs/socket_file.go index 3ed650474..5699d5975 100644 --- a/pkg/sentry/fsimpl/tmpfs/socket_file.go +++ b/pkg/sentry/fsimpl/tmpfs/socket_file.go @@ -21,6 +21,8 @@ import ( ) // socketFile is a socket (=S_IFSOCK) tmpfs file. +// +// +stateify savable type socketFile struct { inode inode ep transport.BoundEndpoint diff --git a/pkg/sentry/fsimpl/tmpfs/symlink.go b/pkg/sentry/fsimpl/tmpfs/symlink.go index b0de5fabe..a102a2ee2 100644 --- a/pkg/sentry/fsimpl/tmpfs/symlink.go +++ b/pkg/sentry/fsimpl/tmpfs/symlink.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ) +// +stateify savable type symlink struct { inode inode target string // immutable diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs.go b/pkg/sentry/fsimpl/tmpfs/tmpfs.go index de2af6d01..4ce859d57 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs.go @@ -51,14 +51,19 @@ import ( const Name = "tmpfs" // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem - // memFile is used to allocate pages to for regular files. - memFile *pgalloc.MemoryFile + // mfp is used to allocate memory that stores regular file contents. mfp is + // immutable. + mfp pgalloc.MemoryFileProvider // clock is a realtime clock used to set timestamps in file operations. clock time.Clock @@ -67,9 +72,11 @@ type filesystem struct { devMinor uint32 // mu serializes changes to the Dentry tree. - mu sync.RWMutex + mu sync.RWMutex `state:"nosave"` nextInoMinusOne uint64 // accessed using atomic memory operations + + root *dentry } // Name implements vfs.FilesystemType.Name. @@ -77,7 +84,12 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + // FilesystemOpts is used to pass configuration data to tmpfs. +// +// +stateify savable type FilesystemOpts struct { // RootFileType is the FileType of the filesystem root. Valid values // are: S_IFDIR, S_IFREG, and S_IFLNK. Defaults to S_IFDIR. @@ -95,8 +107,8 @@ type FilesystemOpts struct { // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, _ string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { - memFileProvider := pgalloc.MemoryFileProviderFromContext(ctx) - if memFileProvider == nil { + mfp := pgalloc.MemoryFileProviderFromContext(ctx) + if mfp == nil { panic("MemoryFileProviderFromContext returned nil") } @@ -170,7 +182,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt } clock := time.RealtimeClockFromContext(ctx) fs := filesystem{ - memFile: memFileProvider.MemoryFile(), + mfp: mfp, clock: clock, devMinor: devMinor, } @@ -188,6 +200,7 @@ func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.Virt fs.vfsfs.DecRef(ctx) return nil, nil, fmt.Errorf("invalid tmpfs root file type: %#o", rootFileType) } + fs.root = root return &fs.vfsfs, &root.vfsd, nil } @@ -199,9 +212,61 @@ func NewFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *au // Release implements vfs.FilesystemImpl.Release. func (fs *filesystem) Release(ctx context.Context) { fs.vfsfs.VirtualFilesystem().PutAnonBlockDevMinor(fs.devMinor) + fs.mu.Lock() + if fs.root.inode.isDir() { + fs.root.releaseChildrenLocked(ctx) + } + fs.mu.Unlock() +} + +// releaseChildrenLocked is called on the mount point by filesystem.Release() to +// destroy all objects in the mount. It performs a depth-first walk of the +// filesystem and "unlinks" everything by decrementing link counts +// appropriately. There should be no open file descriptors when this is called, +// so each inode should only have one outstanding reference that is removed once +// its link count hits zero. +// +// Note that we do not update filesystem state precisely while tearing down (for +// instance, the child maps are ignored)--we only care to remove all remaining +// references so that every filesystem object gets destroyed. Also note that we +// do not need to trigger DecRef on the mount point itself or any child mount; +// these are taken care of by the destructor of the enclosing MountNamespace. +// +// Precondition: filesystem.mu is held. +func (d *dentry) releaseChildrenLocked(ctx context.Context) { + dir := d.inode.impl.(*directory) + for _, child := range dir.childMap { + if child.inode.isDir() { + child.releaseChildrenLocked(ctx) + child.inode.decLinksLocked(ctx) // link for child/. + dir.inode.decLinksLocked(ctx) // link for child/.. + } + child.inode.decLinksLocked(ctx) // link for child + } +} + +// immutable +var globalStatfs = linux.Statfs{ + Type: linux.TMPFS_MAGIC, + BlockSize: usermem.PageSize, + FragmentSize: usermem.PageSize, + NameLength: linux.NAME_MAX, + + // tmpfs currently does not support configurable size limits. In Linux, + // such a tmpfs mount will return f_blocks == f_bfree == f_bavail == 0 from + // statfs(2). However, many applications treat this as having a size limit + // of 0. To work around this, claim to have a very large but non-zero size, + // chosen to ensure that BlockSize * Blocks does not overflow int64 (which + // applications may also handle incorrectly). + // TODO(b/29637826): allow configuring a tmpfs size and enforce it. + Blocks: math.MaxInt64 / usermem.PageSize, + BlocksFree: math.MaxInt64 / usermem.PageSize, + BlocksAvailable: math.MaxInt64 / usermem.PageSize, } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -281,6 +346,8 @@ func (d *dentry) Watches() *vfs.Watches { func (d *dentry) OnZeroWatches(context.Context) {} // inode represents a filesystem object. +// +// +stateify savable type inode struct { // fs is the owning filesystem. fs is immutable. fs *filesystem @@ -297,12 +364,12 @@ type inode struct { // Inode metadata. Writing multiple fields atomically requires holding // mu, othewise atomic operations can be used. - mu sync.Mutex - mode uint32 // file type and mode - nlink uint32 // protected by filesystem.mu instead of inode.mu - uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic - gid uint32 // auth.KGID, but ... - ino uint64 // immutable + mu sync.Mutex `state:"nosave"` + mode uint32 // file type and mode + nlink uint32 // protected by filesystem.mu instead of inode.mu + uid uint32 // auth.KUID, but stored as raw uint32 for sync/atomic + gid uint32 // auth.KGID, but ... + ino uint64 // immutable // Linux's tmpfs has no concept of btime. atime int64 // nanoseconds @@ -340,8 +407,10 @@ func (i *inode) init(impl interface{}, fs *filesystem, kuid auth.KUID, kgid auth // incLinksLocked increments i's link count. // -// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. -// i.nlink < maxLinks. +// Preconditions: +// * filesystem.mu must be locked for writing. +// * i.nlink != 0. +// * i.nlink < maxLinks. func (i *inode) incLinksLocked() { if i.nlink == 0 { panic("tmpfs.inode.incLinksLocked() called with no existing links") @@ -355,7 +424,9 @@ func (i *inode) incLinksLocked() { // decLinksLocked decrements i's link count. If the link count reaches 0, we // remove a reference on i as well. // -// Preconditions: filesystem.mu must be locked for writing. i.nlink != 0. +// Preconditions: +// * filesystem.mu must be locked for writing. +// * i.nlink != 0. func (i *inode) decLinksLocked(ctx context.Context) { if i.nlink == 0 { panic("tmpfs.inode.decLinksLocked() called with no existing links") @@ -594,66 +665,59 @@ func (i *inode) touchCMtime() { i.mu.Unlock() } -// Preconditions: The caller has called vfs.Mount.CheckBeginWrite() and holds -// inode.mu. +// Preconditions: +// * The caller has called vfs.Mount.CheckBeginWrite(). +// * inode.mu must be locked. func (i *inode) touchCMtimeLocked() { now := i.fs.clock.Now().Nanoseconds() atomic.StoreInt64(&i.mtime, now) atomic.StoreInt64(&i.ctime, now) } -func (i *inode) listxattr(size uint64) ([]string, error) { - return i.xattrs.Listxattr(size) +func (i *inode) listXattr(size uint64) ([]string, error) { + return i.xattrs.ListXattr(size) } -func (i *inode) getxattr(creds *auth.Credentials, opts *vfs.GetxattrOptions) (string, error) { - if err := i.checkPermissions(creds, vfs.MayRead); err != nil { +func (i *inode) getXattr(creds *auth.Credentials, opts *vfs.GetXattrOptions) (string, error) { + if err := i.checkXattrPermissions(creds, opts.Name, vfs.MayRead); err != nil { return "", err } - if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { - return "", syserror.EOPNOTSUPP - } - if !i.userXattrSupported() { - return "", syserror.ENODATA - } - return i.xattrs.Getxattr(opts) + return i.xattrs.GetXattr(opts) } -func (i *inode) setxattr(creds *auth.Credentials, opts *vfs.SetxattrOptions) error { - if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { +func (i *inode) setXattr(creds *auth.Credentials, opts *vfs.SetXattrOptions) error { + if err := i.checkXattrPermissions(creds, opts.Name, vfs.MayWrite); err != nil { return err } - if !strings.HasPrefix(opts.Name, linux.XATTR_USER_PREFIX) { - return syserror.EOPNOTSUPP - } - if !i.userXattrSupported() { - return syserror.EPERM - } - return i.xattrs.Setxattr(opts) + return i.xattrs.SetXattr(opts) } -func (i *inode) removexattr(creds *auth.Credentials, name string) error { - if err := i.checkPermissions(creds, vfs.MayWrite); err != nil { +func (i *inode) removeXattr(creds *auth.Credentials, name string) error { + if err := i.checkXattrPermissions(creds, name, vfs.MayWrite); err != nil { return err } - if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) { + return i.xattrs.RemoveXattr(name) +} + +func (i *inode) checkXattrPermissions(creds *auth.Credentials, name string, ats vfs.AccessTypes) error { + // We currently only support extended attributes in the user.* and + // trusted.* namespaces. See b/148380782. + if !strings.HasPrefix(name, linux.XATTR_USER_PREFIX) && !strings.HasPrefix(name, linux.XATTR_TRUSTED_PREFIX) { return syserror.EOPNOTSUPP } - if !i.userXattrSupported() { - return syserror.EPERM + mode := linux.FileMode(atomic.LoadUint32(&i.mode)) + kuid := auth.KUID(atomic.LoadUint32(&i.uid)) + kgid := auth.KGID(atomic.LoadUint32(&i.gid)) + if err := vfs.GenericCheckPermissions(creds, ats, mode, kuid, kgid); err != nil { + return err } - return i.xattrs.Removexattr(name) -} - -// Extended attributes in the user.* namespace are only supported for regular -// files and directories. -func (i *inode) userXattrSupported() bool { - filetype := linux.S_IFMT & atomic.LoadUint32(&i.mode) - return filetype == linux.S_IFREG || filetype == linux.S_IFDIR + return vfs.CheckXattrPermissions(creds, ats, mode, kuid, name) } // fileDescription is embedded by tmpfs implementations of // vfs.FileDescriptionImpl. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -693,20 +757,25 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) return nil } -// Listxattr implements vfs.FileDescriptionImpl.Listxattr. -func (fd *fileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { - return fd.inode().listxattr(size) +// StatFS implements vfs.FileDescriptionImpl.StatFS. +func (fd *fileDescription) StatFS(ctx context.Context) (linux.Statfs, error) { + return globalStatfs, nil } -// Getxattr implements vfs.FileDescriptionImpl.Getxattr. -func (fd *fileDescription) Getxattr(ctx context.Context, opts vfs.GetxattrOptions) (string, error) { - return fd.inode().getxattr(auth.CredentialsFromContext(ctx), &opts) +// ListXattr implements vfs.FileDescriptionImpl.ListXattr. +func (fd *fileDescription) ListXattr(ctx context.Context, size uint64) ([]string, error) { + return fd.inode().listXattr(size) } -// Setxattr implements vfs.FileDescriptionImpl.Setxattr. -func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOptions) error { +// GetXattr implements vfs.FileDescriptionImpl.GetXattr. +func (fd *fileDescription) GetXattr(ctx context.Context, opts vfs.GetXattrOptions) (string, error) { + return fd.inode().getXattr(auth.CredentialsFromContext(ctx), &opts) +} + +// SetXattr implements vfs.FileDescriptionImpl.SetXattr. +func (fd *fileDescription) SetXattr(ctx context.Context, opts vfs.SetXattrOptions) error { d := fd.dentry() - if err := d.inode.setxattr(auth.CredentialsFromContext(ctx), &opts); err != nil { + if err := d.inode.setXattr(auth.CredentialsFromContext(ctx), &opts); err != nil { return err } @@ -715,10 +784,10 @@ func (fd *fileDescription) Setxattr(ctx context.Context, opts vfs.SetxattrOption return nil } -// Removexattr implements vfs.FileDescriptionImpl.Removexattr. -func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { +// RemoveXattr implements vfs.FileDescriptionImpl.RemoveXattr. +func (fd *fileDescription) RemoveXattr(ctx context.Context, name string) error { d := fd.dentry() - if err := d.inode.removexattr(auth.CredentialsFromContext(ctx), name); err != nil { + if err := d.inode.removeXattr(auth.CredentialsFromContext(ctx), name); err != nil { return err } @@ -727,37 +796,6 @@ func (fd *fileDescription) Removexattr(ctx context.Context, name string) error { return nil } -// NewMemfd creates a new tmpfs regular file and file description that can back -// an anonymous fd created by memfd_create. -func NewMemfd(ctx context.Context, creds *auth.Credentials, mount *vfs.Mount, allowSeals bool, name string) (*vfs.FileDescription, error) { - fs, ok := mount.Filesystem().Impl().(*filesystem) - if !ok { - panic("NewMemfd() called with non-tmpfs mount") - } - - // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd inodes are set up with - // S_IRWXUGO. - inode := fs.newRegularFile(creds.EffectiveKUID, creds.EffectiveKGID, 0777) - rf := inode.impl.(*regularFile) - if allowSeals { - rf.seals = 0 - } - - d := fs.newDentry(inode) - defer d.DecRef(ctx) - d.name = name - - // Per Linux, mm/shmem.c:__shmem_file_setup(), memfd files are set up with - // FMODE_READ | FMODE_WRITE. - var fd regularFileFD - fd.Init(&inode.locks) - flags := uint32(linux.O_RDWR) - if err := fd.vfsfd.Init(&fd, flags, mount, &d.vfsd, &vfs.FileDescriptionOptions{}); err != nil { - return nil, err - } - return &fd.vfsfd, nil -} - // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) diff --git a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go index 6f3e3ae6f..fc5323abc 100644 --- a/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go +++ b/pkg/sentry/fsimpl/tmpfs/tmpfs_test.go @@ -41,11 +41,12 @@ func newTmpfsRoot(ctx context.Context) (*vfs.VirtualFilesystem, vfs.VirtualDentr vfsObj.MustRegisterFilesystemType("tmpfs", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ AllowUserMount: true, }) - mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.GetFilesystemOptions{}) + mntns, err := vfsObj.NewMountNamespace(ctx, creds, "", "tmpfs", &vfs.MountOptions{}) if err != nil { return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("failed to create tmpfs root mount: %v", err) } root := mntns.Root() + root.IncRef() return vfsObj, root, func() { root.DecRef(ctx) mntns.DecRef(ctx) diff --git a/pkg/sentry/fsimpl/verity/BUILD b/pkg/sentry/fsimpl/verity/BUILD index 28d2a4bcb..e265be0ee 100644 --- a/pkg/sentry/fsimpl/verity/BUILD +++ b/pkg/sentry/fsimpl/verity/BUILD @@ -1,4 +1,4 @@ -load("//tools:defs.bzl", "go_library") +load("//tools:defs.bzl", "go_library", "go_test") licenses(["notice"]) @@ -6,6 +6,7 @@ go_library( name = "verity", srcs = [ "filesystem.go", + "save_restore.go", "verity.go", ], visibility = ["//pkg/sentry:internal"], @@ -13,11 +14,38 @@ go_library( "//pkg/abi/linux", "//pkg/context", "//pkg/fspath", + "//pkg/marshal/primitive", + "//pkg/merkletree", + "//pkg/refsvfs2", + "//pkg/sentry/arch", "//pkg/sentry/fs/lock", + "//pkg/sentry/kernel", "//pkg/sentry/kernel/auth", "//pkg/sentry/socket/unix/transport", "//pkg/sentry/vfs", "//pkg/sync", "//pkg/syserror", + "//pkg/usermem", + ], +) + +go_test( + name = "verity_test", + srcs = [ + "verity_test.go", + ], + library = ":verity", + deps = [ + "//pkg/abi/linux", + "//pkg/context", + "//pkg/fspath", + "//pkg/sentry/arch", + "//pkg/sentry/fsimpl/testutil", + "//pkg/sentry/fsimpl/tmpfs", + "//pkg/sentry/kernel", + "//pkg/sentry/kernel/auth", + "//pkg/sentry/vfs", + "//pkg/syserror", + "//pkg/usermem", ], ) diff --git a/pkg/sentry/fsimpl/verity/filesystem.go b/pkg/sentry/fsimpl/verity/filesystem.go index 78c6074bd..81dfed266 100644 --- a/pkg/sentry/fsimpl/verity/filesystem.go +++ b/pkg/sentry/fsimpl/verity/filesystem.go @@ -15,9 +15,17 @@ package verity import ( + "bytes" + "fmt" + "io" + "strconv" + "strings" + "sync/atomic" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/merkletree" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sentry/vfs" @@ -91,10 +99,498 @@ func (fs *filesystem) renameMuUnlockAndCheckDrop(ctx context.Context, ds **[]*de putDentrySlice(*ds) } -// resolveLocked resolves rp to an existing file. -func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { - // TODO(b/159261227): Implement resolveLocked. - return nil, nil +// stepLocked resolves rp.Component() to an existing file, starting from the +// given directory. +// +// Dentries which may have a reference count of zero, and which therefore +// should be dropped once traversal is complete, are appended to ds. +// +// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +// !rp.Done(). +func (fs *filesystem) stepLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, mayFollowSymlinks bool, ds **[]*dentry) (*dentry, error) { + if !d.isDir() { + return nil, syserror.ENOTDIR + } + + if err := d.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } + +afterSymlink: + name := rp.Component() + if name == "." { + rp.Advance() + return d, nil + } + if name == ".." { + if isRoot, err := rp.CheckRoot(ctx, &d.vfsd); err != nil { + return nil, err + } else if isRoot || d.parent == nil { + rp.Advance() + return d, nil + } + if err := rp.CheckMount(ctx, &d.parent.vfsd); err != nil { + return nil, err + } + rp.Advance() + return d.parent, nil + } + child, err := fs.getChildLocked(ctx, d, name, ds) + if err != nil { + return nil, err + } + if err := rp.CheckMount(ctx, &child.vfsd); err != nil { + return nil, err + } + if child.isSymlink() && mayFollowSymlinks && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + goto afterSymlink // don't check the current directory again + } + rp.Advance() + return child, nil +} + +// verifyChild verifies the hash of child against the already verified hash of +// the parent to ensure the child is expected. verifyChild triggers a sentry +// panic if unexpected modifications to the file system are detected. In +// noCrashOnVerificationFailure mode it returns a syserror instead. +// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +// TODO(b/166474175): Investigate all possible errors returned in this +// function, and make sure we differentiate all errors that indicate unexpected +// modifications to the file system from the ones that are not harmful. +func (fs *filesystem) verifyChild(ctx context.Context, parent *dentry, child *dentry) (*dentry, error) { + vfsObj := fs.vfsfs.VirtualFilesystem() + + // Get the path to the child dentry. This is only used to provide path + // information in failure case. + childPath, err := vfsObj.PathnameWithDeleted(ctx, child.fs.rootDentry.lowerVD, child.lowerVD) + if err != nil { + return nil, err + } + + fs.verityMu.RLock() + defer fs.verityMu.RUnlock() + // Read the offset of the child from the extended attributes of the + // corresponding Merkle tree file. + // This is the offset of the hash for child in its parent's Merkle tree + // file. + off, err := vfsObj.GetXattrAt(ctx, fs.creds, &vfs.PathOperation{ + Root: child.lowerMerkleVD, + Start: child.lowerMerkleVD, + }, &vfs.GetXattrOptions{ + Name: merkleOffsetInParentXattr, + Size: sizeOfStringInt32, + }) + + // The Merkle tree file for the child should have been created and + // contains the expected xattrs. If the file or the xattr does not + // exist, it indicates unexpected modifications to the file system. + if err == syserror.ENOENT || err == syserror.ENODATA { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleOffsetInParentXattr, childPath, err)) + } + if err != nil { + return nil, err + } + // The offset xattr should be an integer. If it's not, it indicates + // unexpected modifications to the file system. + offset, err := strconv.Atoi(off) + if err != nil { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleOffsetInParentXattr, childPath, err)) + } + + // Open parent Merkle tree file to read and verify child's hash. + parentMerkleFD, err := vfsObj.OpenAt(ctx, fs.creds, &vfs.PathOperation{ + Root: parent.lowerMerkleVD, + Start: parent.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + + // The parent Merkle tree file should have been created. If it's + // missing, it indicates an unexpected modification to the file system. + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to open parent Merkle file for %s: %v", childPath, err)) + } + if err != nil { + return nil, err + } + + // dataSize is the size of raw data for the Merkle tree. For a file, + // dataSize is the size of the whole file. For a directory, dataSize is + // the size of all its children's hashes. + dataSize, err := parentMerkleFD.GetXattr(ctx, &vfs.GetXattrOptions{ + Name: merkleSizeXattr, + Size: sizeOfStringInt32, + }) + + // The Merkle tree file for the child should have been created and + // contains the expected xattrs. If the file or the xattr does not + // exist, it indicates unexpected modifications to the file system. + if err == syserror.ENOENT || err == syserror.ENODATA { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to get xattr %s for %s: %v", merkleSizeXattr, childPath, err)) + } + if err != nil { + return nil, err + } + + // The dataSize xattr should be an integer. If it's not, it indicates + // unexpected modifications to the file system. + parentSize, err := strconv.Atoi(dataSize) + if err != nil { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err)) + } + + fdReader := vfs.FileReadWriteSeeker{ + FD: parentMerkleFD, + Ctx: ctx, + } + + parentStat, err := vfsObj.StatAt(ctx, fs.creds, &vfs.PathOperation{ + Root: parent.lowerVD, + Start: parent.lowerVD, + }, &vfs.StatOptions{}) + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to get parent stat for %s: %v", childPath, err)) + } + if err != nil { + return nil, err + } + + // Since we are verifying against a directory Merkle tree, buf should + // contain the hash of the children in the parent Merkle tree when + // Verify returns with success. + var buf bytes.Buffer + if _, err := merkletree.Verify(&merkletree.VerifyParams{ + Out: &buf, + File: &fdReader, + Tree: &fdReader, + Size: int64(parentSize), + Name: parent.name, + Mode: uint32(parentStat.Mode), + UID: parentStat.UID, + GID: parentStat.GID, + ReadOffset: int64(offset), + ReadSize: int64(merkletree.DigestSize()), + Expected: parent.hash, + DataAndTreeInSameFile: true, + }); err != nil && err != io.EOF { + return nil, alertIntegrityViolation(fmt.Sprintf("Verification for %s failed: %v", childPath, err)) + } + + // Cache child hash when it's verified the first time. + if len(child.hash) == 0 { + child.hash = buf.Bytes() + } + return child, nil +} + +// verifyStat verifies the stat against the verified hash. The mode/uid/gid of +// the file is cached after verified. +func (fs *filesystem) verifyStat(ctx context.Context, d *dentry, stat linux.Statx) error { + vfsObj := fs.vfsfs.VirtualFilesystem() + + // Get the path to the child dentry. This is only used to provide path + // information in failure case. + childPath, err := vfsObj.PathnameWithDeleted(ctx, d.fs.rootDentry.lowerVD, d.lowerVD) + if err != nil { + return err + } + + fs.verityMu.RLock() + defer fs.verityMu.RUnlock() + + fd, err := vfsObj.OpenAt(ctx, fs.creds, &vfs.PathOperation{ + Root: d.lowerMerkleVD, + Start: d.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + if err == syserror.ENOENT { + return alertIntegrityViolation(fmt.Sprintf("Failed to open merkle file for %s: %v", childPath, err)) + } + if err != nil { + return err + } + + merkleSize, err := fd.GetXattr(ctx, &vfs.GetXattrOptions{ + Name: merkleSizeXattr, + Size: sizeOfStringInt32, + }) + + if err == syserror.ENODATA { + return alertIntegrityViolation(fmt.Sprintf("Failed to get xattr %s for merkle file of %s: %v", merkleSizeXattr, childPath, err)) + } + if err != nil { + return err + } + + size, err := strconv.Atoi(merkleSize) + if err != nil { + return alertIntegrityViolation(fmt.Sprintf("Failed to convert xattr %s for %s to int: %v", merkleSizeXattr, childPath, err)) + } + + fdReader := vfs.FileReadWriteSeeker{ + FD: fd, + Ctx: ctx, + } + + var buf bytes.Buffer + params := &merkletree.VerifyParams{ + Out: &buf, + Tree: &fdReader, + Size: int64(size), + Name: d.name, + Mode: uint32(stat.Mode), + UID: stat.UID, + GID: stat.GID, + ReadOffset: 0, + // Set read size to 0 so only the metadata is verified. + ReadSize: 0, + Expected: d.hash, + DataAndTreeInSameFile: false, + } + if atomic.LoadUint32(&d.mode)&linux.S_IFMT == linux.S_IFDIR { + params.DataAndTreeInSameFile = true + } + + if _, err := merkletree.Verify(params); err != nil && err != io.EOF { + return alertIntegrityViolation(fmt.Sprintf("Verification stat for %s failed: %v", childPath, err)) + } + d.mode = uint32(stat.Mode) + d.uid = stat.UID + d.gid = stat.GID + d.size = uint32(size) + return nil +} + +// Preconditions: fs.renameMu must be locked. d.dirMu must be locked. +func (fs *filesystem) getChildLocked(ctx context.Context, parent *dentry, name string, ds **[]*dentry) (*dentry, error) { + if child, ok := parent.children[name]; ok { + // If verity is enabled on child, we should check again whether + // the file and the corresponding Merkle tree are as expected, + // in order to catch deletion/renaming after the last time it's + // accessed. + if child.verityEnabled() { + vfsObj := fs.vfsfs.VirtualFilesystem() + // Get the path to the child dentry. This is only used + // to provide path information in failure case. + path, err := vfsObj.PathnameWithDeleted(ctx, child.fs.rootDentry.lowerVD, child.lowerVD) + if err != nil { + return nil, err + } + + childVD, err := parent.getLowerAt(ctx, vfsObj, name) + if err == syserror.ENOENT { + // The file was previously accessed. If the + // file does not exist now, it indicates an + // unexpected modification to the file system. + return nil, alertIntegrityViolation(fmt.Sprintf("Target file %s is expected but missing", path)) + } + if err != nil { + return nil, err + } + defer childVD.DecRef(ctx) + + childMerkleVD, err := parent.getLowerAt(ctx, vfsObj, merklePrefix+name) + // The Merkle tree file was previous accessed. If it + // does not exist now, it indicates an unexpected + // modification to the file system. + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("Expected Merkle file for target %s but none found", path)) + } + if err != nil { + return nil, err + } + + defer childMerkleVD.DecRef(ctx) + } + + // If enabling verification on files/directories is not allowed + // during runtime, all cached children are already verified. If + // runtime enable is allowed and the parent directory is + // enabled, we should verify the child hash here because it may + // be cached before enabled. + if fs.allowRuntimeEnable { + if parent.verityEnabled() { + if _, err := fs.verifyChild(ctx, parent, child); err != nil { + return nil, err + } + } + if child.verityEnabled() { + vfsObj := fs.vfsfs.VirtualFilesystem() + mask := uint32(linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID) + stat, err := vfsObj.StatAt(ctx, fs.creds, &vfs.PathOperation{ + Root: child.lowerVD, + Start: child.lowerVD, + }, &vfs.StatOptions{ + Mask: mask, + }) + if err != nil { + return nil, err + } + if err := fs.verifyStat(ctx, child, stat); err != nil { + return nil, err + } + } + } + return child, nil + } + child, err := fs.lookupAndVerifyLocked(ctx, parent, name) + if err != nil { + return nil, err + } + if parent.children == nil { + parent.children = make(map[string]*dentry) + } + parent.children[name] = child + // child's refcount is initially 0, so it may be dropped after traversal. + *ds = appendDentry(*ds, child) + return child, nil +} + +// Preconditions: fs.renameMu must be locked. parent.dirMu must be locked. +func (fs *filesystem) lookupAndVerifyLocked(ctx context.Context, parent *dentry, name string) (*dentry, error) { + vfsObj := fs.vfsfs.VirtualFilesystem() + + childVD, childErr := parent.getLowerAt(ctx, vfsObj, name) + // We will handle ENOENT separately, as it may indicate unexpected + // modifications to the file system, and may cause a sentry panic. + if childErr != nil && childErr != syserror.ENOENT { + return nil, childErr + } + + // The dentry needs to be cleaned up if any error occurs. IncRef will be + // called if a verity child dentry is successfully created. + if childErr == nil { + defer childVD.DecRef(ctx) + } + + childMerkleVD, childMerkleErr := parent.getLowerAt(ctx, vfsObj, merklePrefix+name) + // We will handle ENOENT separately, as it may indicate unexpected + // modifications to the file system, and may cause a sentry panic. + if childMerkleErr != nil && childMerkleErr != syserror.ENOENT { + return nil, childMerkleErr + } + + // The dentry needs to be cleaned up if any error occurs. IncRef will be + // called if a verity child dentry is successfully created. + if childMerkleErr == nil { + defer childMerkleVD.DecRef(ctx) + } + + // Get the path to the parent dentry. This is only used to provide path + // information in failure case. + parentPath, err := vfsObj.PathnameWithDeleted(ctx, parent.fs.rootDentry.lowerVD, parent.lowerVD) + if err != nil { + return nil, err + } + + // TODO(b/166474175): Investigate all possible errors of childErr and + // childMerkleErr, and make sure we differentiate all errors that + // indicate unexpected modifications to the file system from the ones + // that are not harmful. + if childErr == syserror.ENOENT && childMerkleErr == nil { + // Failed to get child file/directory dentry. However the + // corresponding Merkle tree is found. This indicates an + // unexpected modification to the file system that + // removed/renamed the child. + return nil, alertIntegrityViolation(fmt.Sprintf("Target file %s is expected but missing", parentPath+"/"+name)) + } else if childErr == nil && childMerkleErr == syserror.ENOENT { + // If in allowRuntimeEnable mode, and the Merkle tree file is + // not created yet, we create an empty Merkle tree file, so that + // if the file is enabled through ioctl, we have the Merkle tree + // file open and ready to use. + // This may cause empty and unused Merkle tree files in + // allowRuntimeEnable mode, if they are never enabled. This + // does not affect verification, as we rely on cached hash to + // decide whether to perform verification, not the existence of + // the Merkle tree file. Also, those Merkle tree files are + // always hidden and cannot be accessed by verity fs users. + if fs.allowRuntimeEnable { + childMerkleFD, err := vfsObj.OpenAt(ctx, fs.creds, &vfs.PathOperation{ + Root: parent.lowerVD, + Start: parent.lowerVD, + Path: fspath.Parse(merklePrefix + name), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT, + Mode: 0644, + }) + if err != nil { + return nil, err + } + childMerkleFD.DecRef(ctx) + childMerkleVD, err = parent.getLowerAt(ctx, vfsObj, merklePrefix+name) + if err != nil { + return nil, err + } + } else { + // If runtime enable is not allowed. This indicates an + // unexpected modification to the file system that + // removed/renamed the Merkle tree file. + return nil, alertIntegrityViolation(fmt.Sprintf("Expected Merkle file for target %s but none found", parentPath+"/"+name)) + } + } else if childErr == syserror.ENOENT && childMerkleErr == syserror.ENOENT { + // Both the child and the corresponding Merkle tree are missing. + // This could be an unexpected modification or due to incorrect + // parameter. + // TODO(b/167752508): Investigate possible ways to differentiate + // cases that both files are deleted from cases that they never + // exist in the file system. + return nil, alertIntegrityViolation(fmt.Sprintf("Failed to find file %s", parentPath+"/"+name)) + } + + mask := uint32(linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID) + stat, err := vfsObj.StatAt(ctx, fs.creds, &vfs.PathOperation{ + Root: childVD, + Start: childVD, + }, &vfs.StatOptions{ + Mask: mask, + }) + if err != nil { + return nil, err + } + + child := fs.newDentry() + child.lowerVD = childVD + child.lowerMerkleVD = childMerkleVD + + // Increase the reference for both childVD and childMerkleVD as they are + // held by child. If this function fails and the child is destroyed, the + // references will be decreased in destroyLocked. + childVD.IncRef() + childMerkleVD.IncRef() + + parent.IncRef() + child.parent = parent + child.name = name + + child.mode = uint32(stat.Mode) + child.uid = stat.UID + child.gid = stat.GID + + // Verify child hash. This should always be performed unless in + // allowRuntimeEnable mode and the parent directory hasn't been enabled + // yet. + if parent.verityEnabled() { + if _, err := fs.verifyChild(ctx, parent, child); err != nil { + child.destroyLocked(ctx) + return nil, err + } + } + if child.verityEnabled() { + if err := fs.verifyStat(ctx, child, stat); err != nil { + child.destroyLocked(ctx) + return nil, err + } + } + + return child, nil } // walkParentDirLocked resolves all but the last path component of rp to an @@ -104,8 +600,39 @@ func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, // // Preconditions: fs.renameMu must be locked. !rp.Done(). func (fs *filesystem) walkParentDirLocked(ctx context.Context, rp *vfs.ResolvingPath, d *dentry, ds **[]*dentry) (*dentry, error) { - // TODO(b/159261227): Implement walkParentDirLocked. - return nil, nil + for !rp.Final() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, true /* mayFollowSymlinks */, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil +} + +// resolveLocked resolves rp to an existing file. +// +// Preconditions: fs.renameMu must be locked. +func (fs *filesystem) resolveLocked(ctx context.Context, rp *vfs.ResolvingPath, ds **[]*dentry) (*dentry, error) { + d := rp.Start().Impl().(*dentry) + for !rp.Done() { + d.dirMu.Lock() + next, err := fs.stepLocked(ctx, rp, d, true /* mayFollowSymlinks */, ds) + d.dirMu.Unlock() + if err != nil { + return nil, err + } + d = next + } + if rp.MustBeDir() && !d.isDir() { + return nil, syserror.ENOTDIR + } + return d, nil } // AccessAt implements vfs.Filesystem.Impl.AccessAt. @@ -179,8 +706,183 @@ func (fs *filesystem) MknodAt(ctx context.Context, rp *vfs.ResolvingPath, opts v // OpenAt implements vfs.FilesystemImpl.OpenAt. func (fs *filesystem) OpenAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.OpenOptions) (*vfs.FileDescription, error) { - //TODO(b/159261227): Implement OpenAt. - return nil, nil + // Verity fs is read-only. + if opts.Flags&(linux.O_WRONLY|linux.O_CREAT) != 0 { + return nil, syserror.EROFS + } + + var ds *[]*dentry + fs.renameMu.RLock() + defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) + + start := rp.Start().Impl().(*dentry) + if rp.Done() { + return start.openLocked(ctx, rp, &opts) + } + +afterTrailingSymlink: + parent, err := fs.walkParentDirLocked(ctx, rp, start, &ds) + if err != nil { + return nil, err + } + + // Check for search permission in the parent directory. + if err := parent.checkPermissions(rp.Credentials(), vfs.MayExec); err != nil { + return nil, err + } + + // Open existing child or follow symlink. + parent.dirMu.Lock() + child, err := fs.stepLocked(ctx, rp, parent, false /*mayFollowSymlinks*/, &ds) + parent.dirMu.Unlock() + if err != nil { + return nil, err + } + if child.isSymlink() && rp.ShouldFollowSymlink() { + target, err := child.readlink(ctx) + if err != nil { + return nil, err + } + if err := rp.HandleSymlink(target); err != nil { + return nil, err + } + start = parent + goto afterTrailingSymlink + } + return child.openLocked(ctx, rp, &opts) +} + +// Preconditions: fs.renameMu must be locked. +func (d *dentry) openLocked(ctx context.Context, rp *vfs.ResolvingPath, opts *vfs.OpenOptions) (*vfs.FileDescription, error) { + // Users should not open the Merkle tree files. Those are for verity fs + // use only. + if strings.Contains(d.name, merklePrefix) { + return nil, syserror.EPERM + } + ats := vfs.AccessTypesForOpenFlags(opts) + if err := d.checkPermissions(rp.Credentials(), ats); err != nil { + return nil, err + } + + // Verity fs is read-only. + if ats&vfs.MayWrite != 0 { + return nil, syserror.EROFS + } + + // Get the path to the target file. This is only used to provide path + // information in failure case. + path, err := d.fs.vfsfs.VirtualFilesystem().PathnameWithDeleted(ctx, d.fs.rootDentry.lowerVD, d.lowerVD) + if err != nil { + return nil, err + } + + // Open the file in the underlying file system. + lowerFD, err := rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerVD, + Start: d.lowerVD, + }, opts) + + // The file should exist, as we succeeded in finding its dentry. If it's + // missing, it indicates an unexpected modification to the file system. + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("File %s expected but not found", path)) + } + return nil, err + } + + // lowerFD needs to be cleaned up if any error occurs. IncRef will be + // called if a verity FD is successfully created. + defer lowerFD.DecRef(ctx) + + // Open the Merkle tree file corresponding to the current file/directory + // to be used later for verifying Read/Walk. + merkleReader, err := rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerMerkleVD, + Start: d.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + + // The Merkle tree file should exist, as we succeeded in finding its + // dentry. If it's missing, it indicates an unexpected modification to + // the file system. + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("Merkle file for %s expected but not found", path)) + } + return nil, err + } + + // merkleReader needs to be cleaned up if any error occurs. IncRef will + // be called if a verity FD is successfully created. + defer merkleReader.DecRef(ctx) + + lowerFlags := lowerFD.StatusFlags() + lowerFDOpts := lowerFD.Options() + var merkleWriter *vfs.FileDescription + var parentMerkleWriter *vfs.FileDescription + + // Only open the Merkle tree files for write if in allowRuntimeEnable + // mode. + if d.fs.allowRuntimeEnable { + merkleWriter, err = rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerMerkleVD, + Start: d.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_WRONLY | linux.O_APPEND, + }) + if err != nil { + if err == syserror.ENOENT { + return nil, alertIntegrityViolation(fmt.Sprintf("Merkle file for %s expected but not found", path)) + } + return nil, err + } + // merkleWriter is cleaned up if any error occurs. IncRef will + // be called if a verity FD is created successfully. + defer merkleWriter.DecRef(ctx) + + if d.parent != nil { + parentMerkleWriter, err = rp.VirtualFilesystem().OpenAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.parent.lowerMerkleVD, + Start: d.parent.lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_WRONLY | linux.O_APPEND, + }) + if err != nil { + if err == syserror.ENOENT { + parentPath, _ := d.fs.vfsfs.VirtualFilesystem().PathnameWithDeleted(ctx, d.fs.rootDentry.lowerVD, d.parent.lowerVD) + return nil, alertIntegrityViolation(fmt.Sprintf("Merkle file for %s expected but not found", parentPath)) + } + return nil, err + } + // parentMerkleWriter is cleaned up if any error occurs. IncRef + // will be called if a verity FD is created successfully. + defer parentMerkleWriter.DecRef(ctx) + } + } + + fd := &fileDescription{ + d: d, + lowerFD: lowerFD, + merkleReader: merkleReader, + merkleWriter: merkleWriter, + parentMerkleWriter: parentMerkleWriter, + isDir: d.isDir(), + } + + if err := fd.vfsfd.Init(fd, lowerFlags, rp.Mount(), &d.vfsd, &lowerFDOpts); err != nil { + return nil, err + } + lowerFD.IncRef() + merkleReader.IncRef() + if merkleWriter != nil { + merkleWriter.IncRef() + } + if parentMerkleWriter != nil { + parentMerkleWriter.IncRef() + } + return &fd.vfsfd, err } // ReadlinkAt implements vfs.FilesystemImpl.ReadlinkAt. @@ -218,6 +920,8 @@ func (fs *filesystem) SetStatAt(ctx context.Context, rp *vfs.ResolvingPath, opts } // StatAt implements vfs.FilesystemImpl.StatAt. +// TODO(b/170157489): Investigate whether stats other than Mode/UID/GID should +// be verified. func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.StatOptions) (linux.Statx, error) { var ds *[]*dentry fs.renameMu.RLock() @@ -235,6 +939,11 @@ func (fs *filesystem) StatAt(ctx context.Context, rp *vfs.ResolvingPath, opts vf if err != nil { return linux.Statx{}, err } + if d.verityEnabled() { + if err := fs.verifyStat(ctx, d, stat); err != nil { + return linux.Statx{}, err + } + } return stat, nil } @@ -256,7 +965,7 @@ func (fs *filesystem) UnlinkAt(ctx context.Context, rp *vfs.ResolvingPath) error return syserror.EROFS } -// BoundEndpointAt implements FilesystemImpl.BoundEndpointAt. +// BoundEndpointAt implements vfs.FilesystemImpl.BoundEndpointAt. func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.BoundEndpointOptions) (transport.BoundEndpoint, error) { var ds *[]*dentry fs.renameMu.RLock() @@ -267,8 +976,8 @@ func (fs *filesystem) BoundEndpointAt(ctx context.Context, rp *vfs.ResolvingPath return nil, syserror.ECONNREFUSED } -// ListxattrAt implements vfs.FilesystemImpl.ListxattrAt. -func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements vfs.FilesystemImpl.ListXattrAt. +func (fs *filesystem) ListXattrAt(ctx context.Context, rp *vfs.ResolvingPath, size uint64) ([]string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) @@ -277,14 +986,14 @@ func (fs *filesystem) ListxattrAt(ctx context.Context, rp *vfs.ResolvingPath, si return nil, err } lowerVD := d.lowerVD - return fs.vfsfs.VirtualFilesystem().ListxattrAt(ctx, d.fs.creds, &vfs.PathOperation{ + return fs.vfsfs.VirtualFilesystem().ListXattrAt(ctx, d.fs.creds, &vfs.PathOperation{ Root: lowerVD, Start: lowerVD, }, size) } -// GetxattrAt implements vfs.FilesystemImpl.GetxattrAt. -func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetxattrOptions) (string, error) { +// GetXattrAt implements vfs.FilesystemImpl.GetXattrAt. +func (fs *filesystem) GetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.GetXattrOptions) (string, error) { var ds *[]*dentry fs.renameMu.RLock() defer fs.renameMuRUnlockAndCheckDrop(ctx, &ds) @@ -293,20 +1002,20 @@ func (fs *filesystem) GetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opt return "", err } lowerVD := d.lowerVD - return fs.vfsfs.VirtualFilesystem().GetxattrAt(ctx, d.fs.creds, &vfs.PathOperation{ + return fs.vfsfs.VirtualFilesystem().GetXattrAt(ctx, d.fs.creds, &vfs.PathOperation{ Root: lowerVD, Start: lowerVD, }, &opts) } -// SetxattrAt implements vfs.FilesystemImpl.SetxattrAt. -func (fs *filesystem) SetxattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetxattrOptions) error { +// SetXattrAt implements vfs.FilesystemImpl.SetXattrAt. +func (fs *filesystem) SetXattrAt(ctx context.Context, rp *vfs.ResolvingPath, opts vfs.SetXattrOptions) error { // Verity file system is read-only. return syserror.EROFS } -// RemovexattrAt implements vfs.FilesystemImpl.RemovexattrAt. -func (fs *filesystem) RemovexattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { +// RemoveXattrAt implements vfs.FilesystemImpl.RemoveXattrAt. +func (fs *filesystem) RemoveXattrAt(ctx context.Context, rp *vfs.ResolvingPath, name string) error { // Verity file system is read-only. return syserror.EROFS } diff --git a/pkg/sentry/fsimpl/verity/save_restore.go b/pkg/sentry/fsimpl/verity/save_restore.go new file mode 100644 index 000000000..4a161163c --- /dev/null +++ b/pkg/sentry/fsimpl/verity/save_restore.go @@ -0,0 +1,27 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package verity + +import ( + "sync/atomic" + + "gvisor.dev/gvisor/pkg/refsvfs2" +) + +func (d *dentry) afterLoad() { + if refsvfs2.LeakCheckEnabled() && atomic.LoadInt64(&d.refs) != -1 { + refsvfs2.Register(d, "verity.dentry") + } +} diff --git a/pkg/sentry/fsimpl/verity/verity.go b/pkg/sentry/fsimpl/verity/verity.go index cb29d33a5..e2cbb206f 100644 --- a/pkg/sentry/fsimpl/verity/verity.go +++ b/pkg/sentry/fsimpl/verity/verity.go @@ -22,29 +22,62 @@ package verity import ( + "fmt" + "math" + "strconv" "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/marshal/primitive" + "gvisor.dev/gvisor/pkg/merkletree" + "gvisor.dev/gvisor/pkg/refsvfs2" + "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" + "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" ) // Name is the default filesystem name. const Name = "verity" -// testOnlyDebugging allows verity file system to return error instead of -// crashing the application when a malicious action is detected. This should -// only be set for tests. -var testOnlyDebugging bool +// merklePrefix is the prefix of the Merkle tree files. For example, the Merkle +// tree file for "/foo" is "/.merkle.verity.foo". +const merklePrefix = ".merkle.verity." + +// merkleoffsetInParentXattr is the extended attribute name specifying the +// offset of child hash in its parent's Merkle tree. +const merkleOffsetInParentXattr = "user.merkle.offset" + +// merkleSizeXattr is the extended attribute name specifying the size of data +// hashed by the corresponding Merkle tree. For a file, it's the size of the +// whole file. For a directory, it's the size of all its children's hashes. +const merkleSizeXattr = "user.merkle.size" + +// sizeOfStringInt32 is the size for a 32 bit integer stored as string in +// extended attributes. The maximum value of a 32 bit integer is 10 digits. +const sizeOfStringInt32 = 10 + +// noCrashOnVerificationFailure indicates whether the sandbox should panic +// whenever verification fails. If true, an error is returned instead of +// panicking. This should only be set for tests. +// TOOD(b/165661693): Decide whether to panic or return error based on this +// flag. +var noCrashOnVerificationFailure bool // FilesystemType implements vfs.FilesystemType. +// +// +stateify savable type FilesystemType struct{} // filesystem implements vfs.FilesystemImpl. +// +// +stateify savable type filesystem struct { vfsfs vfs.Filesystem @@ -69,11 +102,24 @@ type filesystem struct { // renameMu synchronizes renaming with non-renaming operations in order // to ensure consistent lock ordering between dentry.dirMu in different // dentries. - renameMu sync.RWMutex + renameMu sync.RWMutex `state:"nosave"` + + // verityMu synchronizes enabling verity files, protects files or + // directories from being enabled by different threads simultaneously. + // It also ensures that verity does not access files that are being + // enabled. + // + // Also, the directory Merkle trees depends on the generated trees of + // its children. So they shouldn't be enabled the same time. This lock + // is for the whole file system to ensure that no more than one file is + // enabled the same time. + verityMu sync.RWMutex } // InternalFilesystemOptions may be passed as // vfs.GetFilesystemOptions.InternalData to FilesystemType.GetFilesystem. +// +// +stateify savable type InternalFilesystemOptions struct { // RootMerkleFileName is the name of the verity root Merkle tree file. RootMerkleFileName string @@ -93,10 +139,10 @@ type InternalFilesystemOptions struct { // system wrapped by verity file system. LowerGetFSOptions vfs.GetFilesystemOptions - // TestOnlyDebugging allows verity file system to return error instead - // of crashing the application when a malicious action is detected. This - // should only be set for tests. - TestOnlyDebugging bool + // NoCrashOnVerificationFailure indicates whether the sandbox should + // panic whenever verification fails. If true, an error is returned + // instead of panicking. This should only be set for tests. + NoCrashOnVerificationFailure bool } // Name implements vfs.FilesystemType.Name. @@ -104,10 +150,129 @@ func (FilesystemType) Name() string { return Name } +// Release implements vfs.FilesystemType.Release. +func (FilesystemType) Release(ctx context.Context) {} + +// alertIntegrityViolation alerts a violation of integrity, which usually means +// unexpected modification to the file system is detected. In +// noCrashOnVerificationFailure mode, it returns EIO, otherwise it panic. +func alertIntegrityViolation(msg string) error { + if noCrashOnVerificationFailure { + return syserror.EIO + } + panic(msg) +} + // GetFilesystem implements vfs.FilesystemType.GetFilesystem. func (fstype FilesystemType) GetFilesystem(ctx context.Context, vfsObj *vfs.VirtualFilesystem, creds *auth.Credentials, source string, opts vfs.GetFilesystemOptions) (*vfs.Filesystem, *vfs.Dentry, error) { - //TODO(b/159261227): Implement GetFilesystem. - return nil, nil, nil + iopts, ok := opts.InternalData.(InternalFilesystemOptions) + if !ok { + ctx.Warningf("verity.FilesystemType.GetFilesystem: missing verity configs") + return nil, nil, syserror.EINVAL + } + noCrashOnVerificationFailure = iopts.NoCrashOnVerificationFailure + + // Mount the lower file system. The lower file system is wrapped inside + // verity, and should not be exposed or connected. + mopts := &vfs.MountOptions{ + GetFilesystemOptions: iopts.LowerGetFSOptions, + InternalMount: true, + } + mnt, err := vfsObj.MountDisconnected(ctx, creds, "", iopts.LowerName, mopts) + if err != nil { + return nil, nil, err + } + + fs := &filesystem{ + creds: creds.Fork(), + lowerMount: mnt, + allowRuntimeEnable: iopts.AllowRuntimeEnable, + } + fs.vfsfs.Init(vfsObj, &fstype, fs) + + // Construct the root dentry. + d := fs.newDentry() + d.refs = 1 + lowerVD := vfs.MakeVirtualDentry(mnt, mnt.Root()) + lowerVD.IncRef() + d.lowerVD = lowerVD + + rootMerkleName := merklePrefix + iopts.RootMerkleFileName + + lowerMerkleVD, err := vfsObj.GetDentryAt(ctx, fs.creds, &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + Path: fspath.Parse(rootMerkleName), + }, &vfs.GetDentryOptions{}) + + // If runtime enable is allowed, the root merkle tree may be absent. We + // should create the tree file. + if err == syserror.ENOENT && fs.allowRuntimeEnable { + lowerMerkleFD, err := vfsObj.OpenAt(ctx, fs.creds, &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + Path: fspath.Parse(rootMerkleName), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT, + Mode: 0644, + }) + if err != nil { + fs.vfsfs.DecRef(ctx) + d.DecRef(ctx) + return nil, nil, err + } + lowerMerkleFD.DecRef(ctx) + lowerMerkleVD, err = vfsObj.GetDentryAt(ctx, fs.creds, &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + Path: fspath.Parse(rootMerkleName), + }, &vfs.GetDentryOptions{}) + if err != nil { + fs.vfsfs.DecRef(ctx) + d.DecRef(ctx) + return nil, nil, err + } + } else if err != nil { + // Failed to get dentry for the root Merkle file. This + // indicates an unexpected modification that removed/renamed + // the root Merkle file, or it's never generated. + fs.vfsfs.DecRef(ctx) + d.DecRef(ctx) + return nil, nil, alertIntegrityViolation("Failed to find root Merkle file") + } + d.lowerMerkleVD = lowerMerkleVD + + // Get metadata from the underlying file system. + const statMask = linux.STATX_TYPE | linux.STATX_MODE | linux.STATX_UID | linux.STATX_GID + stat, err := vfsObj.StatAt(ctx, creds, &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + }, &vfs.StatOptions{ + Mask: statMask, + }) + if err != nil { + fs.vfsfs.DecRef(ctx) + d.DecRef(ctx) + return nil, nil, err + } + + d.mode = uint32(stat.Mode) + d.uid = stat.UID + d.gid = stat.GID + d.hash = make([]byte, len(iopts.RootHash)) + + if !fs.allowRuntimeEnable { + if err := fs.verifyStat(ctx, d, stat); err != nil { + return nil, nil, err + } + } + + copy(d.hash, iopts.RootHash) + d.vfsd.Init(d) + + fs.rootDentry = d + + return &fs.vfsfs, &d.vfsd, nil } // Release implements vfs.FilesystemImpl.Release. @@ -116,6 +281,8 @@ func (fs *filesystem) Release(ctx context.Context) { } // dentry implements vfs.DentryImpl. +// +// +stateify savable type dentry struct { vfsd vfs.Dentry @@ -124,11 +291,12 @@ type dentry struct { // fs is the owning filesystem. fs is immutable. fs *filesystem - // mode, uid and gid are the file mode, owner, and group of the file in - // the underlying file system. + // mode, uid, gid and size are the file mode, owner, group, and size of + // the file in the underlying file system. mode uint32 uid uint32 gid uint32 + size uint32 // parent is the dentry corresponding to this dentry's parent directory. // name is this dentry's name in parent. If this dentry is a filesystem @@ -142,7 +310,7 @@ type dentry struct { // and dirents (if not nil) is a cache of dirents as returned by // directoryFDs representing this directory. children is protected by // dirMu. - dirMu sync.Mutex + dirMu sync.Mutex `state:"nosave"` children map[string]*dentry // lowerVD is the VirtualDentry in the underlying file system. @@ -152,8 +320,8 @@ type dentry struct { // in the underlying file system. lowerMerkleVD vfs.VirtualDentry - // rootHash is the rootHash for the current file or directory. - rootHash []byte + // hash is the calculated hash for the current file or directory. + hash []byte } // newDentry creates a new dentry representing the given verity file. The @@ -166,6 +334,9 @@ func (fs *filesystem) newDentry() *dentry { fs: fs, } d.vfsd.Init(d) + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Register(d, "verity.dentry") + } return d } @@ -228,6 +399,9 @@ func (d *dentry) destroyLocked(ctx context.Context) { if d.lowerVD.Ok() { d.lowerVD.DecRef(ctx) } + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Unregister(d, "verity.dentry") + } if d.lowerMerkleVD.Ok() { d.lowerMerkleVD.DecRef(ctx) @@ -247,6 +421,11 @@ func (d *dentry) destroyLocked(ctx context.Context) { } } +// LeakMessage implements refsvfs2.CheckedObject.LeakMessage. +func (d *dentry) LeakMessage() string { + return fmt.Sprintf("[verity.dentry %p] reference count of %d instead of -1", d, atomic.LoadInt64(&d.refs)) +} + // InotifyWithParent implements vfs.DentryImpl.InotifyWithParent. func (d *dentry) InotifyWithParent(ctx context.Context, events, cookie uint32, et vfs.EventType) { //TODO(b/159261227): Implement InotifyWithParent. @@ -275,6 +454,24 @@ func (d *dentry) checkPermissions(creds *auth.Credentials, ats vfs.AccessTypes) return vfs.GenericCheckPermissions(creds, ats, linux.FileMode(atomic.LoadUint32(&d.mode)), auth.KUID(atomic.LoadUint32(&d.uid)), auth.KGID(atomic.LoadUint32(&d.gid))) } +// verityEnabled checks whether the file is enabled with verity features. It +// should always be true if runtime enable is not allowed. In runtime enable +// mode, it returns true if the target has been enabled with +// ioctl(FS_IOC_ENABLE_VERITY). +func (d *dentry) verityEnabled() bool { + return !d.fs.allowRuntimeEnable || len(d.hash) != 0 +} + +// getLowerAt returns the dentry in the underlying file system, which is +// represented by filename relative to d. +func (d *dentry) getLowerAt(ctx context.Context, vfsObj *vfs.VirtualFilesystem, filename string) (vfs.VirtualDentry, error) { + return vfsObj.GetDentryAt(ctx, d.fs.creds, &vfs.PathOperation{ + Root: d.lowerVD, + Start: d.lowerVD, + Path: fspath.Parse(filename), + }, &vfs.GetDentryOptions{}) +} + func (d *dentry) readlink(ctx context.Context) (string, error) { return d.fs.vfsfs.VirtualFilesystem().ReadlinkAt(ctx, d.fs.creds, &vfs.PathOperation{ Root: d.lowerVD, @@ -286,6 +483,8 @@ func (d *dentry) readlink(ctx context.Context) (string, error) { // FileDescription is a wrapper of the underlying lowerFD, with support to build // Merkle trees through the Linux fs-verity API to verify contents read from // lowerFD. +// +// +stateify savable type fileDescription struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -314,6 +513,10 @@ type fileDescription struct { // directory that contains the current file/directory. This is only used // if allowRuntimeEnable is set to true. parentMerkleWriter *vfs.FileDescription + + // off is the file offset. off is protected by mu. + mu sync.Mutex `state:"nosave"` + off int64 } // Release implements vfs.FileDescriptionImpl.Release. @@ -335,6 +538,11 @@ func (fd *fileDescription) Stat(ctx context.Context, opts vfs.StatOptions) (linu if err != nil { return linux.Statx{}, err } + if fd.d.verityEnabled() { + if err := fd.d.fs.verifyStat(ctx, fd.d, stat); err != nil { + return linux.Statx{}, err + } + } return stat, nil } @@ -344,12 +552,325 @@ func (fd *fileDescription) SetStat(ctx context.Context, opts vfs.SetStatOptions) return syserror.EPERM } +// Seek implements vfs.FileDescriptionImpl.Seek. +func (fd *fileDescription) Seek(ctx context.Context, offset int64, whence int32) (int64, error) { + fd.mu.Lock() + defer fd.mu.Unlock() + n := int64(0) + switch whence { + case linux.SEEK_SET: + // use offset as specified + case linux.SEEK_CUR: + n = fd.off + case linux.SEEK_END: + n = int64(fd.d.size) + default: + return 0, syserror.EINVAL + } + if offset > math.MaxInt64-n { + return 0, syserror.EINVAL + } + offset += n + if offset < 0 { + return 0, syserror.EINVAL + } + fd.off = offset + return offset, nil +} + +// generateMerkle generates a Merkle tree file for fd. If fd points to a file +// /foo/bar, a Merkle tree file /foo/.merkle.verity.bar is generated. The hash +// of the generated Merkle tree and the data size is returned. If fd points to +// a regular file, the data is the content of the file. If fd points to a +// directory, the data is all hahes of its children, written to the Merkle tree +// file. +func (fd *fileDescription) generateMerkle(ctx context.Context) ([]byte, uint64, error) { + fdReader := vfs.FileReadWriteSeeker{ + FD: fd.lowerFD, + Ctx: ctx, + } + merkleReader := vfs.FileReadWriteSeeker{ + FD: fd.merkleReader, + Ctx: ctx, + } + merkleWriter := vfs.FileReadWriteSeeker{ + FD: fd.merkleWriter, + Ctx: ctx, + } + params := &merkletree.GenerateParams{ + TreeReader: &merkleReader, + TreeWriter: &merkleWriter, + } + + switch atomic.LoadUint32(&fd.d.mode) & linux.S_IFMT { + case linux.S_IFREG: + // For a regular file, generate a Merkle tree based on its + // content. + var err error + stat, err := fd.lowerFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return nil, 0, err + } + + params.File = &fdReader + params.Size = int64(stat.Size) + params.Name = fd.d.name + params.Mode = uint32(stat.Mode) + params.UID = stat.UID + params.GID = stat.GID + params.DataAndTreeInSameFile = false + case linux.S_IFDIR: + // For a directory, generate a Merkle tree based on the hashes + // of its children that has already been written to the Merkle + // tree file. + merkleStat, err := fd.merkleReader.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return nil, 0, err + } + + params.Size = int64(merkleStat.Size) + + stat, err := fd.lowerFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return nil, 0, err + } + + params.File = &merkleReader + params.Name = fd.d.name + params.Mode = uint32(stat.Mode) + params.UID = stat.UID + params.GID = stat.GID + params.DataAndTreeInSameFile = true + default: + // TODO(b/167728857): Investigate whether and how we should + // enable other types of file. + return nil, 0, syserror.EINVAL + } + hash, err := merkletree.Generate(params) + return hash, uint64(params.Size), err +} + +// enableVerity enables verity features on fd by generating a Merkle tree file +// and stores its hash in its parent directory's Merkle tree. +func (fd *fileDescription) enableVerity(ctx context.Context, uio usermem.IO) (uintptr, error) { + if !fd.d.fs.allowRuntimeEnable { + return 0, syserror.EPERM + } + + fd.d.fs.verityMu.Lock() + defer fd.d.fs.verityMu.Unlock() + + // In allowRuntimeEnable mode, the underlying fd and read/write fd for + // the Merkle tree file should have all been initialized. For any file + // or directory other than the root, the parent Merkle tree file should + // have also been initialized. + if fd.lowerFD == nil || fd.merkleReader == nil || fd.merkleWriter == nil || (fd.parentMerkleWriter == nil && fd.d != fd.d.fs.rootDentry) { + return 0, alertIntegrityViolation("Unexpected verity fd: missing expected underlying fds") + } + + hash, dataSize, err := fd.generateMerkle(ctx) + if err != nil { + return 0, err + } + + if fd.parentMerkleWriter != nil { + stat, err := fd.parentMerkleWriter.Stat(ctx, vfs.StatOptions{}) + if err != nil { + return 0, err + } + + // Write the hash of fd to the parent directory's Merkle tree + // file, as it should be part of the parent Merkle tree data. + // parentMerkleWriter is open with O_APPEND, so it should write + // directly to the end of the file. + if _, err = fd.parentMerkleWriter.Write(ctx, usermem.BytesIOSequence(hash), vfs.WriteOptions{}); err != nil { + return 0, err + } + + // Record the offset of the hash of fd in parent directory's + // Merkle tree file. + if err := fd.merkleWriter.SetXattr(ctx, &vfs.SetXattrOptions{ + Name: merkleOffsetInParentXattr, + Value: strconv.Itoa(int(stat.Size)), + }); err != nil { + return 0, err + } + } + + // Record the size of the data being hashed for fd. + if err := fd.merkleWriter.SetXattr(ctx, &vfs.SetXattrOptions{ + Name: merkleSizeXattr, + Value: strconv.Itoa(int(dataSize)), + }); err != nil { + return 0, err + } + fd.d.hash = append(fd.d.hash, hash...) + return 0, nil +} + +// measureVerity returns the hash of fd, saved in verityDigest. +func (fd *fileDescription) measureVerity(ctx context.Context, uio usermem.IO, verityDigest usermem.Addr) (uintptr, error) { + t := kernel.TaskFromContext(ctx) + if t == nil { + return 0, syserror.EINVAL + } + var metadata linux.DigestMetadata + + // If allowRuntimeEnable is true, an empty fd.d.hash indicates that + // verity is not enabled for the file. If allowRuntimeEnable is false, + // this is an integrity violation because all files should have verity + // enabled, in which case fd.d.hash should be set. + if len(fd.d.hash) == 0 { + if fd.d.fs.allowRuntimeEnable { + return 0, syserror.ENODATA + } + return 0, alertIntegrityViolation("Ioctl measureVerity: no hash found") + } + + // The first part of VerityDigest is the metadata. + if _, err := metadata.CopyIn(t, verityDigest); err != nil { + return 0, err + } + if metadata.DigestSize < uint16(len(fd.d.hash)) { + return 0, syserror.EOVERFLOW + } + + // Populate the output digest size, since DigestSize is both input and + // output. + metadata.DigestSize = uint16(len(fd.d.hash)) + + // First copy the metadata. + if _, err := metadata.CopyOut(t, verityDigest); err != nil { + return 0, err + } + + // Now copy the root hash bytes to the memory after metadata. + _, err := t.CopyOutBytes(usermem.Addr(uintptr(verityDigest)+linux.SizeOfDigestMetadata), fd.d.hash) + return 0, err +} + +func (fd *fileDescription) verityFlags(ctx context.Context, uio usermem.IO, flags usermem.Addr) (uintptr, error) { + f := int32(0) + + // All enabled files should store a hash. This flag is not settable via + // FS_IOC_SETFLAGS. + if len(fd.d.hash) != 0 { + f |= linux.FS_VERITY_FL + } + + t := kernel.TaskFromContext(ctx) + if t == nil { + return 0, syserror.EINVAL + } + _, err := primitive.CopyInt32Out(t, flags, f) + return 0, err +} + +// Ioctl implements vfs.FileDescriptionImpl.Ioctl. +func (fd *fileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) { + switch cmd := args[1].Uint(); cmd { + case linux.FS_IOC_ENABLE_VERITY: + return fd.enableVerity(ctx, uio) + case linux.FS_IOC_MEASURE_VERITY: + return fd.measureVerity(ctx, uio, args[2].Pointer()) + case linux.FS_IOC_GETFLAGS: + return fd.verityFlags(ctx, uio, args[2].Pointer()) + default: + // TODO(b/169682228): Investigate which ioctl commands should + // be allowed. + return 0, syserror.ENOSYS + } +} + +// Read implements vfs.FileDescriptionImpl.Read. +func (fd *fileDescription) Read(ctx context.Context, dst usermem.IOSequence, opts vfs.ReadOptions) (int64, error) { + // Implement Read with PRead by setting offset. + fd.mu.Lock() + n, err := fd.PRead(ctx, dst, fd.off, opts) + fd.off += n + fd.mu.Unlock() + return n, err +} + +// PRead implements vfs.FileDescriptionImpl.PRead. +func (fd *fileDescription) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { + // No need to verify if the file is not enabled yet in + // allowRuntimeEnable mode. + if !fd.d.verityEnabled() { + return fd.lowerFD.PRead(ctx, dst, offset, opts) + } + + fd.d.fs.verityMu.RLock() + defer fd.d.fs.verityMu.RUnlock() + // dataSize is the size of the whole file. + dataSize, err := fd.merkleReader.GetXattr(ctx, &vfs.GetXattrOptions{ + Name: merkleSizeXattr, + Size: sizeOfStringInt32, + }) + + // The Merkle tree file for the child should have been created and + // contains the expected xattrs. If the xattr does not exist, it + // indicates unexpected modifications to the file system. + if err == syserror.ENODATA { + return 0, alertIntegrityViolation(fmt.Sprintf("Failed to get xattr %s: %v", merkleSizeXattr, err)) + } + if err != nil { + return 0, err + } + + // The dataSize xattr should be an integer. If it's not, it indicates + // unexpected modifications to the file system. + size, err := strconv.Atoi(dataSize) + if err != nil { + return 0, alertIntegrityViolation(fmt.Sprintf("Failed to convert xattr %s to int: %v", merkleSizeXattr, err)) + } + + dataReader := vfs.FileReadWriteSeeker{ + FD: fd.lowerFD, + Ctx: ctx, + } + + merkleReader := vfs.FileReadWriteSeeker{ + FD: fd.merkleReader, + Ctx: ctx, + } + + n, err := merkletree.Verify(&merkletree.VerifyParams{ + Out: dst.Writer(ctx), + File: &dataReader, + Tree: &merkleReader, + Size: int64(size), + Name: fd.d.name, + Mode: fd.d.mode, + UID: fd.d.uid, + GID: fd.d.gid, + ReadOffset: offset, + ReadSize: dst.NumBytes(), + Expected: fd.d.hash, + DataAndTreeInSameFile: false, + }) + if err != nil { + return 0, alertIntegrityViolation(fmt.Sprintf("Verification failed: %v", err)) + } + return n, err +} + +// PWrite implements vfs.FileDescriptionImpl.PWrite. +func (fd *fileDescription) PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EROFS +} + +// Write implements vfs.FileDescriptionImpl.Write. +func (fd *fileDescription) Write(ctx context.Context, src usermem.IOSequence, opts vfs.WriteOptions) (int64, error) { + return 0, syserror.EROFS +} + // LockPOSIX implements vfs.FileDescriptionImpl.LockPOSIX. func (fd *fileDescription) LockPOSIX(ctx context.Context, uid fslock.UniqueID, t fslock.LockType, start, length uint64, whence int16, block fslock.Blocker) error { - return fd.Locks().LockPOSIX(ctx, &fd.vfsfd, uid, t, start, length, whence, block) + return fd.lowerFD.LockPOSIX(ctx, uid, t, start, length, whence, block) } // UnlockPOSIX implements vfs.FileDescriptionImpl.UnlockPOSIX. func (fd *fileDescription) UnlockPOSIX(ctx context.Context, uid fslock.UniqueID, start, length uint64, whence int16) error { - return fd.Locks().UnlockPOSIX(ctx, &fd.vfsfd, uid, start, length, whence) + return fd.lowerFD.UnlockPOSIX(ctx, uid, start, length, whence) } diff --git a/pkg/sentry/fsimpl/verity/verity_test.go b/pkg/sentry/fsimpl/verity/verity_test.go new file mode 100644 index 000000000..c647cbfd3 --- /dev/null +++ b/pkg/sentry/fsimpl/verity/verity_test.go @@ -0,0 +1,700 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package verity + +import ( + "fmt" + "io" + "math/rand" + "testing" + "time" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/testutil" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/sentry/kernel/auth" + "gvisor.dev/gvisor/pkg/sentry/vfs" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// rootMerkleFilename is the name of the root Merkle tree file. +const rootMerkleFilename = "root.verity" + +// maxDataSize is the maximum data size written to the file for test. +const maxDataSize = 100000 + +// newVerityRoot creates a new verity mount, and returns the root. The +// underlying file system is tmpfs. If the error is not nil, then cleanup +// should be called when the root is no longer needed. +func newVerityRoot(t *testing.T) (*vfs.VirtualFilesystem, vfs.VirtualDentry, *kernel.Task, error) { + k, err := testutil.Boot() + if err != nil { + t.Fatalf("testutil.Boot: %v", err) + } + + ctx := k.SupervisorContext() + + rand.Seed(time.Now().UnixNano()) + vfsObj := &vfs.VirtualFilesystem{} + if err := vfsObj.Init(ctx); err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("VFS init: %v", err) + } + + vfsObj.MustRegisterFilesystemType("verity", FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + + vfsObj.MustRegisterFilesystemType("tmpfs", tmpfs.FilesystemType{}, &vfs.RegisterFilesystemTypeOptions{ + AllowUserMount: true, + }) + + mntns, err := vfsObj.NewMountNamespace(ctx, auth.CredentialsFromContext(ctx), "", "verity", &vfs.MountOptions{ + GetFilesystemOptions: vfs.GetFilesystemOptions{ + InternalData: InternalFilesystemOptions{ + RootMerkleFileName: rootMerkleFilename, + LowerName: "tmpfs", + AllowRuntimeEnable: true, + NoCrashOnVerificationFailure: true, + }, + }, + }) + if err != nil { + return nil, vfs.VirtualDentry{}, nil, fmt.Errorf("NewMountNamespace: %v", err) + } + root := mntns.Root() + root.IncRef() + + // Use lowerRoot in the task as we modify the lower file system + // directly in many tests. + lowerRoot := root.Dentry().Impl().(*dentry).lowerVD + tc := k.NewThreadGroup(nil, k.RootPIDNamespace(), kernel.NewSignalHandlers(), linux.SIGCHLD, k.GlobalInit().Limits()) + task, err := testutil.CreateTask(ctx, "name", tc, mntns, lowerRoot, lowerRoot) + if err != nil { + t.Fatalf("testutil.CreateTask: %v", err) + } + + t.Helper() + t.Cleanup(func() { + root.DecRef(ctx) + mntns.DecRef(ctx) + }) + return vfsObj, root, task, nil +} + +// newFileFD creates a new file in the verity mount, and returns the FD. The FD +// points to a file that has random data generated. +func newFileFD(ctx context.Context, vfsObj *vfs.VirtualFilesystem, root vfs.VirtualDentry, filePath string, mode linux.FileMode) (*vfs.FileDescription, int, error) { + creds := auth.CredentialsFromContext(ctx) + lowerRoot := root.Dentry().Impl().(*dentry).lowerVD + + // Create the file in the underlying file system. + lowerFD, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(filePath), + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR | linux.O_CREAT | linux.O_EXCL, + Mode: linux.ModeRegular | mode, + }) + if err != nil { + return nil, 0, err + } + + // Generate random data to be written to the file. + dataSize := rand.Intn(maxDataSize) + 1 + data := make([]byte, dataSize) + rand.Read(data) + + // Write directly to the underlying FD, since verity FD is read-only. + n, err := lowerFD.Write(ctx, usermem.BytesIOSequence(data), vfs.WriteOptions{}) + if err != nil { + return nil, 0, err + } + + if n != int64(len(data)) { + return nil, 0, fmt.Errorf("lowerFD.Write got write length %d, want %d", n, len(data)) + } + + lowerFD.DecRef(ctx) + + // Now open the verity file descriptor. + fd, err := vfsObj.OpenAt(ctx, creds, &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filePath), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular | mode, + }) + return fd, dataSize, err +} + +// corruptRandomBit randomly flips a bit in the file represented by fd. +func corruptRandomBit(ctx context.Context, fd *vfs.FileDescription, size int) error { + // Flip a random bit in the underlying file. + randomPos := int64(rand.Intn(size)) + byteToModify := make([]byte, 1) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(byteToModify), randomPos, vfs.ReadOptions{}); err != nil { + return fmt.Errorf("lowerFD.PRead: %v", err) + } + byteToModify[0] ^= 1 + if _, err := fd.PWrite(ctx, usermem.BytesIOSequence(byteToModify), randomPos, vfs.WriteOptions{}); err != nil { + return fmt.Errorf("lowerFD.PWrite: %v", err) + } + return nil +} + +// TestOpen ensures that when a file is created, the corresponding Merkle tree +// file and the root Merkle tree file exist. +func TestOpen(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + if _, _, err := newFileFD(ctx, vfsObj, root, filename, 0644); err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Ensure that the corresponding Merkle tree file is created. + lowerRoot := root.Dentry().Impl().(*dentry).lowerVD + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(merklePrefix + filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }); err != nil { + t.Errorf("OpenAt Merkle tree file %s: %v", merklePrefix+filename, err) + } + + // Ensure the root merkle tree file is created. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerRoot, + Start: lowerRoot, + Path: fspath.Parse(merklePrefix + rootMerkleFilename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }); err != nil { + t.Errorf("OpenAt root Merkle tree file %s: %v", merklePrefix+rootMerkleFilename, err) + } +} + +// TestPReadUnmodifiedFileSucceeds ensures that pread from an untouched verity +// file succeeds after enabling verity for it. +func TestPReadUnmodifiedFileSucceeds(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirm a normal read succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + buf := make([]byte, size) + n, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}) + if err != nil && err != io.EOF { + t.Fatalf("fd.PRead: %v", err) + } + + if n != int64(size) { + t.Errorf("fd.PRead got read length %d, want %d", n, size) + } +} + +// TestReadUnmodifiedFileSucceeds ensures that read from an untouched verity +// file succeeds after enabling verity for it. +func TestReadUnmodifiedFileSucceeds(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirm a normal read succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + buf := make([]byte, size) + n, err := fd.Read(ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) + if err != nil && err != io.EOF { + t.Fatalf("fd.Read: %v", err) + } + + if n != int64(size) { + t.Errorf("fd.PRead got read length %d, want %d", n, size) + } +} + +// TestReopenUnmodifiedFileSucceeds ensures that reopen an untouched verity file +// succeeds after enabling verity for it. +func TestReopenUnmodifiedFileSucceeds(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirms a normal read succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Ensure reopening the verity enabled file succeeds. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular, + }); err != nil { + t.Errorf("reopen enabled file failed: %v", err) + } +} + +// TestPReadModifiedFileFails ensures that read from a modified verity file +// fails. +func TestPReadModifiedFileFails(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerFD that's read/writable. + lowerVD := fd.Impl().(*fileDescription).d.lowerVD + + lowerFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + if err := corruptRandomBit(ctx, lowerFD, size); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + // Confirm that read from the modified file fails. + buf := make([]byte, size) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}); err == nil { + t.Fatalf("fd.PRead succeeded, expected failure") + } +} + +// TestReadModifiedFileFails ensures that read from a modified verity file +// fails. +func TestReadModifiedFileFails(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerFD that's read/writable. + lowerVD := fd.Impl().(*fileDescription).d.lowerVD + + lowerFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerVD, + Start: lowerVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + if err := corruptRandomBit(ctx, lowerFD, size); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + // Confirm that read from the modified file fails. + buf := make([]byte, size) + if _, err := fd.Read(ctx, usermem.BytesIOSequence(buf), vfs.ReadOptions{}); err == nil { + t.Fatalf("fd.Read succeeded, expected failure") + } +} + +// TestModifiedMerkleFails ensures that read from a verity file fails if the +// corresponding Merkle tree file is modified. +func TestModifiedMerkleFails(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, size, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerMerkleFD that's read/writable. + lowerMerkleVD := fd.Impl().(*fileDescription).d.lowerMerkleVD + + lowerMerkleFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: lowerMerkleVD, + Start: lowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + // Flip a random bit in the Merkle tree file. + stat, err := lowerMerkleFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + t.Fatalf("stat: %v", err) + } + merkleSize := int(stat.Size) + if err := corruptRandomBit(ctx, lowerMerkleFD, merkleSize); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + // Confirm that read from a file with modified Merkle tree fails. + buf := make([]byte, size) + if _, err := fd.PRead(ctx, usermem.BytesIOSequence(buf), 0 /* offset */, vfs.ReadOptions{}); err == nil { + fmt.Println(buf) + t.Fatalf("fd.PRead succeeded with modified Merkle file") + } +} + +// TestModifiedParentMerkleFails ensures that open a verity enabled file in a +// verity enabled directory fails if the hashes related to the target file in +// the parent Merkle tree file is modified. +func TestModifiedParentMerkleFails(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Enable verity on the parent directory. + parentFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + if _, err := parentFD.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + // Open a new lowerMerkleFD that's read/writable. + parentLowerMerkleVD := fd.Impl().(*fileDescription).d.parent.lowerMerkleVD + + parentLowerMerkleFD, err := vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: parentLowerMerkleVD, + Start: parentLowerMerkleVD, + }, &vfs.OpenOptions{ + Flags: linux.O_RDWR, + }) + if err != nil { + t.Fatalf("OpenAt: %v", err) + } + + // Flip a random bit in the parent Merkle tree file. + // This parent directory contains only one child, so any random + // modification in the parent Merkle tree should cause verification + // failure when opening the child file. + stat, err := parentLowerMerkleFD.Stat(ctx, vfs.StatOptions{}) + if err != nil { + t.Fatalf("stat: %v", err) + } + parentMerkleSize := int(stat.Size) + if err := corruptRandomBit(ctx, parentLowerMerkleFD, parentMerkleSize); err != nil { + t.Fatalf("corruptRandomBit: %v", err) + } + + parentLowerMerkleFD.DecRef(ctx) + + // Ensure reopening the verity enabled file fails. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular, + }); err == nil { + t.Errorf("OpenAt file with modified parent Merkle succeeded") + } +} + +// TestUnmodifiedStatSucceeds ensures that stat of an untouched verity file +// succeeds after enabling verity for it. +func TestUnmodifiedStatSucceeds(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file and confirms stat succeeds. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("fd.Ioctl: %v", err) + } + + if _, err := fd.Stat(ctx, vfs.StatOptions{}); err != nil { + t.Errorf("fd.Stat: %v", err) + } +} + +// TestModifiedStatFails checks that getting stat for a file with modified stat +// should fail. +func TestModifiedStatFails(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("fd.Ioctl: %v", err) + } + + lowerFD := fd.Impl().(*fileDescription).lowerFD + // Change the stat of the underlying file, and check that stat fails. + if err := lowerFD.SetStat(ctx, vfs.SetStatOptions{ + Stat: linux.Statx{ + Mask: uint32(linux.STATX_MODE), + Mode: 0777, + }, + }); err != nil { + t.Fatalf("lowerFD.SetStat: %v", err) + } + + if _, err := fd.Stat(ctx, vfs.StatOptions{}); err == nil { + t.Errorf("fd.Stat succeeded when it should fail") + } +} + +// TestOpenDeletedOrRenamedFileFails ensures that opening a deleted/renamed +// verity enabled file or the corresponding Merkle tree file fails with the +// verify error. +func TestOpenDeletedFileFails(t *testing.T) { + testCases := []struct { + // Tests removing files is remove is true. Otherwise tests + // renaming files. + remove bool + // The original file is removed/renamed if changeFile is true. + changeFile bool + // The Merkle tree file is removed/renamed if changeMerkleFile + // is true. + changeMerkleFile bool + }{ + { + remove: true, + changeFile: true, + changeMerkleFile: false, + }, + { + remove: true, + changeFile: false, + changeMerkleFile: true, + }, + { + remove: false, + changeFile: true, + changeMerkleFile: false, + }, + { + remove: false, + changeFile: true, + changeMerkleFile: false, + }, + } + for _, tc := range testCases { + t.Run(fmt.Sprintf("remove:%t", tc.remove), func(t *testing.T) { + vfsObj, root, ctx, err := newVerityRoot(t) + if err != nil { + t.Fatalf("newVerityRoot: %v", err) + } + + filename := "verity-test-file" + fd, _, err := newFileFD(ctx, vfsObj, root, filename, 0644) + if err != nil { + t.Fatalf("newFileFD: %v", err) + } + + // Enable verity on the file. + var args arch.SyscallArguments + args[1] = arch.SyscallArgument{Value: linux.FS_IOC_ENABLE_VERITY} + if _, err := fd.Ioctl(ctx, nil /* uio */, args); err != nil { + t.Fatalf("Ioctl: %v", err) + } + + rootLowerVD := root.Dentry().Impl().(*dentry).lowerVD + if tc.remove { + if tc.changeFile { + if err := vfsObj.UnlinkAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(filename), + }); err != nil { + t.Fatalf("UnlinkAt: %v", err) + } + } + if tc.changeMerkleFile { + if err := vfsObj.UnlinkAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(merklePrefix + filename), + }); err != nil { + t.Fatalf("UnlinkAt: %v", err) + } + } + } else { + newFilename := "renamed-test-file" + if tc.changeFile { + if err := vfsObj.RenameAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(filename), + }, &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(newFilename), + }, &vfs.RenameOptions{}); err != nil { + t.Fatalf("RenameAt: %v", err) + } + } + if tc.changeMerkleFile { + if err := vfsObj.RenameAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(merklePrefix + filename), + }, &vfs.PathOperation{ + Root: rootLowerVD, + Start: rootLowerVD, + Path: fspath.Parse(merklePrefix + newFilename), + }, &vfs.RenameOptions{}); err != nil { + t.Fatalf("UnlinkAt: %v", err) + } + } + } + + // Ensure reopening the verity enabled file fails. + if _, err = vfsObj.OpenAt(ctx, auth.CredentialsFromContext(ctx), &vfs.PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(filename), + }, &vfs.OpenOptions{ + Flags: linux.O_RDONLY, + Mode: linux.ModeRegular, + }); err != syserror.EIO { + t.Errorf("got OpenAt error: %v, expected EIO", err) + } + }) + } +} diff --git a/pkg/sentry/hostmm/BUILD b/pkg/sentry/hostmm/BUILD index 61c78569d..300b7ccce 100644 --- a/pkg/sentry/hostmm/BUILD +++ b/pkg/sentry/hostmm/BUILD @@ -7,11 +7,14 @@ go_library( srcs = [ "cgroup.go", "hostmm.go", + "membarrier.go", ], visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", "//pkg/fd", "//pkg/log", "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/hostmm/membarrier.go b/pkg/sentry/hostmm/membarrier.go new file mode 100644 index 000000000..4468d75f1 --- /dev/null +++ b/pkg/sentry/hostmm/membarrier.go @@ -0,0 +1,90 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package hostmm + +import ( + "syscall" + + "golang.org/x/sys/unix" + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/log" +) + +var ( + haveMembarrierGlobal = false + haveMembarrierPrivateExpedited = false +) + +func init() { + supported, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_QUERY, 0 /* flags */, 0 /* unused */) + if e != 0 { + if e != syscall.ENOSYS { + log.Warningf("membarrier(MEMBARRIER_CMD_QUERY) failed: %s", e.Error()) + } + return + } + // We don't use MEMBARRIER_CMD_GLOBAL_EXPEDITED because this sends IPIs to + // all CPUs running tasks that have previously invoked + // MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED, which presents a DOS risk. + // (MEMBARRIER_CMD_GLOBAL is synchronize_rcu(), i.e. it waits for an RCU + // grace period to elapse without bothering other CPUs. + // MEMBARRIER_CMD_PRIVATE_EXPEDITED sends IPIs only to CPUs running tasks + // sharing the caller's MM.) + if supported&linux.MEMBARRIER_CMD_GLOBAL != 0 { + haveMembarrierGlobal = true + } + if req := uintptr(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED | linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED); supported&req == req { + if _, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED, 0 /* flags */, 0 /* unused */); e != 0 { + log.Warningf("membarrier(MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED) failed: %s", e.Error()) + } else { + haveMembarrierPrivateExpedited = true + } + } +} + +// HaveGlobalMemoryBarrier returns true if GlobalMemoryBarrier is supported. +func HaveGlobalMemoryBarrier() bool { + return haveMembarrierGlobal +} + +// GlobalMemoryBarrier blocks until "all running threads [in the host OS] have +// passed through a state where all memory accesses to user-space addresses +// match program order between entry to and return from [GlobalMemoryBarrier]", +// as for membarrier(2). +// +// Preconditions: HaveGlobalMemoryBarrier() == true. +func GlobalMemoryBarrier() error { + if _, _, e := syscall.Syscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_GLOBAL, 0 /* flags */, 0 /* unused */); e != 0 { + return e + } + return nil +} + +// HaveProcessMemoryBarrier returns true if ProcessMemoryBarrier is supported. +func HaveProcessMemoryBarrier() bool { + return haveMembarrierPrivateExpedited +} + +// ProcessMemoryBarrier is equivalent to GlobalMemoryBarrier, but only +// synchronizes with threads sharing a virtual address space (from the host OS' +// perspective) with the calling thread. +// +// Preconditions: HaveProcessMemoryBarrier() == true. +func ProcessMemoryBarrier() error { + if _, _, e := syscall.RawSyscall(unix.SYS_MEMBARRIER, linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED, 0 /* flags */, 0 /* unused */); e != 0 { + return e + } + return nil +} diff --git a/pkg/sentry/inet/BUILD b/pkg/sentry/inet/BUILD index 07bf39fed..5bba9de0b 100644 --- a/pkg/sentry/inet/BUILD +++ b/pkg/sentry/inet/BUILD @@ -15,6 +15,7 @@ go_library( ], deps = [ "//pkg/context", + "//pkg/tcpip", "//pkg/tcpip/stack", ], ) diff --git a/pkg/sentry/inet/inet.go b/pkg/sentry/inet/inet.go index c0b4831d1..f31277d30 100644 --- a/pkg/sentry/inet/inet.go +++ b/pkg/sentry/inet/inet.go @@ -15,7 +15,10 @@ // Package inet defines semantics for IP stacks. package inet -import "gvisor.dev/gvisor/pkg/tcpip/stack" +import ( + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/stack" +) // Stack represents a TCP/IP stack. type Stack interface { @@ -29,9 +32,13 @@ type Stack interface { InterfaceAddrs() map[int32][]InterfaceAddr // AddInterfaceAddr adds an address to the network interface identified by - // index. + // idx. AddInterfaceAddr(idx int32, addr InterfaceAddr) error + // RemoveInterfaceAddr removes an address from the network interface + // identified by idx. + RemoveInterfaceAddr(idx int32, addr InterfaceAddr) error + // SupportsIPv6 returns true if the stack supports IPv6 connectivity. SupportsIPv6() bool @@ -80,6 +87,12 @@ type Stack interface { // RestoreCleanupEndpoints adds endpoints to cleanup tracking. This is useful // for restoring a stack after a save. RestoreCleanupEndpoints([]stack.TransportEndpoint) + + // Forwarding returns if packet forwarding between NICs is enabled. + Forwarding(protocol tcpip.NetworkProtocolNumber) bool + + // SetForwarding enables or disables packet forwarding between NICs. + SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error } // Interface contains information about a network interface. diff --git a/pkg/sentry/inet/test_stack.go b/pkg/sentry/inet/test_stack.go index 9771f01fc..9ebeba8a3 100644 --- a/pkg/sentry/inet/test_stack.go +++ b/pkg/sentry/inet/test_stack.go @@ -14,7 +14,13 @@ package inet -import "gvisor.dev/gvisor/pkg/tcpip/stack" +import ( + "bytes" + "fmt" + + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/stack" +) // TestStack is a dummy implementation of Stack for tests. type TestStack struct { @@ -26,6 +32,7 @@ type TestStack struct { TCPSendBufSize TCPBufferSize TCPSACKFlag bool Recovery TCPLossRecovery + IPForwarding bool } // NewTestStack returns a TestStack with no network interfaces. The value of @@ -54,6 +61,24 @@ func (s *TestStack) AddInterfaceAddr(idx int32, addr InterfaceAddr) error { return nil } +// RemoveInterfaceAddr implements Stack.RemoveInterfaceAddr. +func (s *TestStack) RemoveInterfaceAddr(idx int32, addr InterfaceAddr) error { + interfaceAddrs, ok := s.InterfaceAddrsMap[idx] + if !ok { + return fmt.Errorf("unknown idx: %d", idx) + } + + var filteredAddrs []InterfaceAddr + for _, interfaceAddr := range interfaceAddrs { + if !bytes.Equal(interfaceAddr.Addr, addr.Addr) { + filteredAddrs = append(filteredAddrs, addr) + } + } + s.InterfaceAddrsMap[idx] = filteredAddrs + + return nil +} + // SupportsIPv6 implements Stack.SupportsIPv6. func (s *TestStack) SupportsIPv6() bool { return s.SupportsIPv6Flag @@ -128,3 +153,14 @@ func (s *TestStack) CleanupEndpoints() []stack.TransportEndpoint { // RestoreCleanupEndpoints implements inet.Stack.RestoreCleanupEndpoints. func (s *TestStack) RestoreCleanupEndpoints([]stack.TransportEndpoint) {} + +// Forwarding implements inet.Stack.Forwarding. +func (s *TestStack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + return s.IPForwarding +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *TestStack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error { + s.IPForwarding = enable + return nil +} diff --git a/pkg/sentry/kernel/BUILD b/pkg/sentry/kernel/BUILD index 5416a310d..90dd4a047 100644 --- a/pkg/sentry/kernel/BUILD +++ b/pkg/sentry/kernel/BUILD @@ -69,8 +69,63 @@ go_template_instance( prefix = "socket", template = "//pkg/ilist:generic_list", types = { - "Element": "*SocketEntry", - "Linker": "*SocketEntry", + "Element": "*SocketRecordVFS1", + "Linker": "*SocketRecordVFS1", + }, +) + +go_template_instance( + name = "fd_table_refs", + out = "fd_table_refs.go", + package = "kernel", + prefix = "FDTable", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "FDTable", + }, +) + +go_template_instance( + name = "fs_context_refs", + out = "fs_context_refs.go", + package = "kernel", + prefix = "FSContext", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "FSContext", + }, +) + +go_template_instance( + name = "ipc_namespace_refs", + out = "ipc_namespace_refs.go", + package = "kernel", + prefix = "IPCNamespace", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "IPCNamespace", + }, +) + +go_template_instance( + name = "process_group_refs", + out = "process_group_refs.go", + package = "kernel", + prefix = "ProcessGroup", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "ProcessGroup", + }, +) + +go_template_instance( + name = "session_refs", + out = "session_refs.go", + package = "kernel", + prefix = "Session", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "Session", }, ) @@ -88,9 +143,14 @@ go_library( "aio.go", "context.go", "fd_table.go", + "fd_table_refs.go", "fd_table_unsafe.go", "fs_context.go", + "fs_context_refs.go", "ipc_namespace.go", + "ipc_namespace_refs.go", + "kcov.go", + "kcov_unsafe.go", "kernel.go", "kernel_opts.go", "kernel_state.go", @@ -99,6 +159,7 @@ go_library( "pending_signals_state.go", "posixtimer.go", "process_group_list.go", + "process_group_refs.go", "ptrace.go", "ptrace_amd64.go", "ptrace_arm64.go", @@ -106,6 +167,7 @@ go_library( "seccomp.go", "seqatomic_taskgoroutineschedinfo_unsafe.go", "session_list.go", + "session_refs.go", "sessions.go", "signal.go", "signal_handlers.go", @@ -147,23 +209,27 @@ go_library( "gvisor.dev/gvisor/pkg/sentry/device", "gvisor.dev/gvisor/pkg/tcpip", ], + marshal = True, visibility = ["//:sandbox"], deps = [ ":uncaught_signal_go_proto", "//pkg/abi", "//pkg/abi/linux", "//pkg/amutex", - "//pkg/binary", "//pkg/bits", "//pkg/bpf", + "//pkg/cleanup", "//pkg/context", + "//pkg/coverage", "//pkg/cpuid", "//pkg/eventchannel", "//pkg/fspath", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/metric", "//pkg/refs", - "//pkg/refs_vfs2", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/secio", "//pkg/sentry/arch", @@ -210,7 +276,6 @@ go_library( "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/kernel/abstract_socket_namespace.go b/pkg/sentry/kernel/abstract_socket_namespace.go index 1b9721534..0ddbe5ff6 100644 --- a/pkg/sentry/kernel/abstract_socket_namespace.go +++ b/pkg/sentry/kernel/abstract_socket_namespace.go @@ -19,7 +19,7 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs_vfs2" + "gvisor.dev/gvisor/pkg/refsvfs2" "gvisor.dev/gvisor/pkg/sentry/socket/unix/transport" "gvisor.dev/gvisor/pkg/sync" ) @@ -27,7 +27,7 @@ import ( // +stateify savable type abstractEndpoint struct { ep transport.BoundEndpoint - socket refs_vfs2.RefCounter + socket refsvfs2.RefCounter name string ns *AbstractSocketNamespace } @@ -57,7 +57,7 @@ func NewAbstractSocketNamespace() *AbstractSocketNamespace { // its backing socket. type boundEndpoint struct { transport.BoundEndpoint - socket refs_vfs2.RefCounter + socket refsvfs2.RefCounter } // Release implements transport.BoundEndpoint.Release. @@ -89,7 +89,7 @@ func (a *AbstractSocketNamespace) BoundEndpoint(name string) transport.BoundEndp // // When the last reference managed by socket is dropped, ep may be removed from the // namespace. -func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep transport.BoundEndpoint, socket refs_vfs2.RefCounter) error { +func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep transport.BoundEndpoint, socket refsvfs2.RefCounter) error { a.mu.Lock() defer a.mu.Unlock() @@ -109,7 +109,7 @@ func (a *AbstractSocketNamespace) Bind(ctx context.Context, name string, ep tran // Remove removes the specified socket at name from the abstract socket // namespace, if it has not yet been replaced. -func (a *AbstractSocketNamespace) Remove(name string, socket refs_vfs2.RefCounter) { +func (a *AbstractSocketNamespace) Remove(name string, socket refsvfs2.RefCounter) { a.mu.Lock() defer a.mu.Unlock() diff --git a/pkg/sentry/kernel/auth/BUILD b/pkg/sentry/kernel/auth/BUILD index 2bc49483a..869e49ebc 100644 --- a/pkg/sentry/kernel/auth/BUILD +++ b/pkg/sentry/kernel/auth/BUILD @@ -57,6 +57,7 @@ go_library( "id_map_set.go", "user_namespace.go", ], + marshal = True, visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", diff --git a/pkg/sentry/kernel/auth/context.go b/pkg/sentry/kernel/auth/context.go index ef5723127..c08d47787 100644 --- a/pkg/sentry/kernel/auth/context.go +++ b/pkg/sentry/kernel/auth/context.go @@ -34,3 +34,23 @@ func CredentialsFromContext(ctx context.Context) *Credentials { } return NewAnonymousCredentials() } + +// ContextWithCredentials returns a copy of ctx carrying creds. +func ContextWithCredentials(ctx context.Context, creds *Credentials) context.Context { + return &authContext{ctx, creds} +} + +type authContext struct { + context.Context + creds *Credentials +} + +// Value implements context.Context. +func (ac *authContext) Value(key interface{}) interface{} { + switch key { + case CtxCredentials: + return ac.creds + default: + return ac.Context.Value(key) + } +} diff --git a/pkg/sentry/kernel/auth/id.go b/pkg/sentry/kernel/auth/id.go index 0a58ba17c..4c32ee703 100644 --- a/pkg/sentry/kernel/auth/id.go +++ b/pkg/sentry/kernel/auth/id.go @@ -19,9 +19,13 @@ import ( ) // UID is a user ID in an unspecified user namespace. +// +// +marshal type UID uint32 // GID is a group ID in an unspecified user namespace. +// +// +marshal slice:GIDSlice type GID uint32 // In the root user namespace, user/group IDs have a 1-to-1 relationship with diff --git a/pkg/sentry/kernel/context.go b/pkg/sentry/kernel/context.go index dd5f0f5fa..bb94769c4 100644 --- a/pkg/sentry/kernel/context.go +++ b/pkg/sentry/kernel/context.go @@ -81,7 +81,8 @@ func UTSNamespaceFromContext(ctx context.Context) *UTSNamespace { } // IPCNamespaceFromContext returns the IPC namespace in which ctx is executing, -// or nil if there is no such IPC namespace. +// or nil if there is no such IPC namespace. It takes a reference on the +// namespace. func IPCNamespaceFromContext(ctx context.Context) *IPCNamespace { if v := ctx.Value(CtxIPCNamespace); v != nil { return v.(*IPCNamespace) diff --git a/pkg/sentry/kernel/fd_table.go b/pkg/sentry/kernel/fd_table.go index ce53af69b..7aba31587 100644 --- a/pkg/sentry/kernel/fd_table.go +++ b/pkg/sentry/kernel/fd_table.go @@ -23,7 +23,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -78,7 +77,8 @@ type descriptor struct { // // +stateify savable type FDTable struct { - refs.AtomicRefCount + FDTableRefs + k *Kernel // mu protects below. @@ -110,9 +110,12 @@ func (f *FDTable) saveDescriptorTable() map[int32]descriptor { func (f *FDTable) loadDescriptorTable(m map[int32]descriptor) { ctx := context.Background() - f.init() // Initialize table. + f.initNoLeakCheck() // Initialize table. + f.used = 0 for fd, d := range m { - f.setAll(fd, d.file, d.fileVFS2, d.flags) + if file, fileVFS2 := f.setAll(ctx, fd, d.file, d.fileVFS2, d.flags); file != nil || fileVFS2 != nil { + panic("VFS1 or VFS2 files set") + } // Note that we do _not_ need to acquire a extra table reference here. The // table reference will already be accounted for in the file, so we drop the @@ -127,7 +130,7 @@ func (f *FDTable) loadDescriptorTable(m map[int32]descriptor) { } // drop drops the table reference. -func (f *FDTable) drop(file *fs.File) { +func (f *FDTable) drop(ctx context.Context, file *fs.File) { // Release locks. file.Dirent.Inode.LockCtx.Posix.UnlockRegion(f, lock.LockRange{0, lock.LockEOF}) @@ -145,14 +148,13 @@ func (f *FDTable) drop(file *fs.File) { d.InotifyEvent(ev, 0) // Drop the table reference. - file.DecRef(context.Background()) + file.DecRef(ctx) } // dropVFS2 drops the table reference. -func (f *FDTable) dropVFS2(file *vfs.FileDescription) { +func (f *FDTable) dropVFS2(ctx context.Context, file *vfs.FileDescription) { // Release any POSIX lock possibly held by the FDTable. Range {0, 0} means the // entire file. - ctx := context.Background() err := file.UnlockPOSIX(ctx, f, 0, 0, linux.SEEK_SET) if err != nil && err != syserror.ENOLCK { panic(fmt.Sprintf("UnlockPOSIX failed: %v", err)) @@ -176,22 +178,15 @@ func (k *Kernel) NewFDTable() *FDTable { return f } -// destroy removes all of the file descriptors from the map. -func (f *FDTable) destroy(ctx context.Context) { - f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool { - return true - }) -} - -// DecRef implements RefCounter.DecRef with destructor f.destroy. +// DecRef implements RefCounter.DecRef. +// +// If f reaches zero references, all of its file descriptors are removed. func (f *FDTable) DecRef(ctx context.Context) { - f.DecRefWithDestructor(ctx, f.destroy) -} - -// Size returns the number of file descriptor slots currently allocated. -func (f *FDTable) Size() int { - size := atomic.LoadInt32(&f.used) - return int(size) + f.FDTableRefs.DecRef(func() { + f.RemoveIf(ctx, func(*fs.File, *vfs.FileDescription, FDFlags) bool { + return true + }) + }) } // forEach iterates over all non-nil files in sorted order. @@ -245,6 +240,10 @@ func (f *FDTable) String() string { case fileVFS2 != nil: vfsObj := fileVFS2.Mount().Filesystem().VirtualFilesystem() + vd := fileVFS2.VirtualDentry() + if vd.Dentry() == nil { + panic(fmt.Sprintf("fd %d (type %T) has nil dentry: %#v", fd, fileVFS2.Impl(), fileVFS2)) + } name, err := vfsObj.PathnameWithDeleted(ctx, vfs.VirtualDentry{}, fileVFS2.VirtualDentry()) if err != nil { fmt.Fprintf(&buf, "<err: %v>\n", err) @@ -280,7 +279,6 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags } f.mu.Lock() - defer f.mu.Unlock() // From f.next to find available fd. if fd < f.next { @@ -290,15 +288,25 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags // Install all entries. for i := fd; i < end && len(fds) < len(files); i++ { if d, _, _ := f.get(i); d == nil { - f.set(i, files[len(fds)], flags) // Set the descriptor. - fds = append(fds, i) // Record the file descriptor. + // Set the descriptor. + f.set(ctx, i, files[len(fds)], flags) + fds = append(fds, i) // Record the file descriptor. } } // Failure? Unwind existing FDs. if len(fds) < len(files) { for _, i := range fds { - f.set(i, nil, FDFlags{}) // Zap entry. + f.set(ctx, i, nil, FDFlags{}) + } + f.mu.Unlock() + + // Drop the reference taken by the call to f.set() that + // originally installed the file. Don't call f.drop() + // (generating inotify events, etc.) since the file should + // appear to have never been inserted into f. + for _, file := range files[:len(fds)] { + file.DecRef(ctx) } return nil, syscall.EMFILE } @@ -308,6 +316,7 @@ func (f *FDTable) NewFDs(ctx context.Context, fd int32, files []*fs.File, flags f.next = fds[len(fds)-1] + 1 } + f.mu.Unlock() return fds, nil } @@ -335,7 +344,6 @@ func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDes } f.mu.Lock() - defer f.mu.Unlock() // From f.next to find available fd. if fd < f.next { @@ -345,15 +353,25 @@ func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDes // Install all entries. for i := fd; i < end && len(fds) < len(files); i++ { if d, _, _ := f.getVFS2(i); d == nil { - f.setVFS2(i, files[len(fds)], flags) // Set the descriptor. - fds = append(fds, i) // Record the file descriptor. + // Set the descriptor. + f.setVFS2(ctx, i, files[len(fds)], flags) + fds = append(fds, i) // Record the file descriptor. } } // Failure? Unwind existing FDs. if len(fds) < len(files) { for _, i := range fds { - f.setVFS2(i, nil, FDFlags{}) // Zap entry. + f.setVFS2(ctx, i, nil, FDFlags{}) + } + f.mu.Unlock() + + // Drop the reference taken by the call to f.setVFS2() that + // originally installed the file. Don't call f.dropVFS2() + // (generating inotify events, etc.) since the file should + // appear to have never been inserted into f. + for _, file := range files[:len(fds)] { + file.DecRef(ctx) } return nil, syscall.EMFILE } @@ -363,6 +381,7 @@ func (f *FDTable) NewFDsVFS2(ctx context.Context, fd int32, files []*vfs.FileDes f.next = fds[len(fds)-1] + 1 } + f.mu.Unlock() return fds, nil } @@ -398,7 +417,7 @@ func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDesc } for fd < end { if d, _, _ := f.getVFS2(fd); d == nil { - f.setVFS2(fd, file, flags) + f.setVFS2(ctx, fd, file, flags) if fd == f.next { // Update next search start position. f.next = fd + 1 @@ -414,40 +433,55 @@ func (f *FDTable) NewFDVFS2(ctx context.Context, minfd int32, file *vfs.FileDesc // reference for that FD, the ref count for that existing reference is // decremented. func (f *FDTable) NewFDAt(ctx context.Context, fd int32, file *fs.File, flags FDFlags) error { - return f.newFDAt(ctx, fd, file, nil, flags) + df, _, err := f.newFDAt(ctx, fd, file, nil, flags) + if err != nil { + return err + } + if df != nil { + f.drop(ctx, df) + } + return nil } // NewFDAtVFS2 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 *FDTable) NewFDAtVFS2(ctx context.Context, fd int32, file *vfs.FileDescription, flags FDFlags) error { - return f.newFDAt(ctx, fd, nil, file, flags) + _, dfVFS2, err := f.newFDAt(ctx, fd, nil, file, flags) + if err != nil { + return err + } + if dfVFS2 != nil { + f.dropVFS2(ctx, dfVFS2) + } + return nil } -func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) error { +func (f *FDTable) newFDAt(ctx context.Context, fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) (*fs.File, *vfs.FileDescription, error) { if fd < 0 { // Don't accept negative FDs. - return syscall.EBADF + return nil, nil, syscall.EBADF } // Check the limit for the provided file. if limitSet := limits.FromContext(ctx); limitSet != nil { if lim := limitSet.Get(limits.NumberOfFiles); lim.Cur != limits.Infinity && uint64(fd) >= lim.Cur { - return syscall.EMFILE + return nil, nil, syscall.EMFILE } } // Install the entry. f.mu.Lock() defer f.mu.Unlock() - f.setAll(fd, file, fileVFS2, flags) - return nil + + df, dfVFS2 := f.setAll(ctx, fd, file, fileVFS2, flags) + return df, dfVFS2, nil } // SetFlags sets the flags for the given file descriptor. // // True is returned iff flags were changed. -func (f *FDTable) SetFlags(fd int32, flags FDFlags) error { +func (f *FDTable) SetFlags(ctx context.Context, fd int32, flags FDFlags) error { if fd < 0 { // Don't accept negative FDs. return syscall.EBADF @@ -463,14 +497,14 @@ func (f *FDTable) SetFlags(fd int32, flags FDFlags) error { } // Update the flags. - f.set(fd, file, flags) + f.set(ctx, fd, file, flags) return nil } // SetFlagsVFS2 sets the flags for the given file descriptor. // // True is returned iff flags were changed. -func (f *FDTable) SetFlagsVFS2(fd int32, flags FDFlags) error { +func (f *FDTable) SetFlagsVFS2(ctx context.Context, fd int32, flags FDFlags) error { if fd < 0 { // Don't accept negative FDs. return syscall.EBADF @@ -486,7 +520,7 @@ func (f *FDTable) SetFlagsVFS2(fd int32, flags FDFlags) error { } // Update the flags. - f.setVFS2(fd, file, flags) + f.setVFS2(ctx, fd, file, flags) return nil } @@ -552,30 +586,6 @@ func (f *FDTable) GetFDs(ctx context.Context) []int32 { return 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 *FDTable) GetRefs(ctx context.Context) []*fs.File { - files := make([]*fs.File, 0, f.Size()) - f.forEach(ctx, func(_ int32, file *fs.File, _ *vfs.FileDescription, _ FDFlags) { - file.IncRef() // Acquire a reference for caller. - files = append(files, file) - }) - return files -} - -// GetRefsVFS2 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 *FDTable) GetRefsVFS2(ctx context.Context) []*vfs.FileDescription { - files := make([]*vfs.FileDescription, 0, f.Size()) - f.forEach(ctx, func(_ int32, _ *fs.File, file *vfs.FileDescription, _ FDFlags) { - file.IncRef() // Acquire a reference for caller. - files = append(files, file) - }) - return files -} - // Fork returns an independent FDTable. func (f *FDTable) Fork(ctx context.Context) *FDTable { clone := f.k.NewFDTable() @@ -583,11 +593,8 @@ func (f *FDTable) Fork(ctx context.Context) *FDTable { f.forEach(ctx, func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { // The set function here will acquire an appropriate table // reference for the clone. We don't need anything else. - switch { - case file != nil: - clone.set(fd, file, flags) - case fileVFS2 != nil: - clone.setVFS2(fd, fileVFS2, flags) + if df, dfVFS2 := clone.setAll(ctx, fd, file, fileVFS2, flags); df != nil || dfVFS2 != nil { + panic("VFS1 or VFS2 files set") } }) return clone @@ -596,13 +603,12 @@ func (f *FDTable) Fork(ctx context.Context) *FDTable { // Remove removes an FD from and returns a non-file iff successful. // // N.B. Callers are required to use DecRef when they are done. -func (f *FDTable) Remove(fd int32) (*fs.File, *vfs.FileDescription) { +func (f *FDTable) Remove(ctx context.Context, fd int32) (*fs.File, *vfs.FileDescription) { if fd < 0 { return nil, nil } f.mu.Lock() - defer f.mu.Unlock() // Update current available position. if fd < f.next { @@ -618,24 +624,51 @@ func (f *FDTable) Remove(fd int32) (*fs.File, *vfs.FileDescription) { case orig2 != nil: orig2.IncRef() } + if orig != nil || orig2 != nil { - f.setAll(fd, nil, nil, FDFlags{}) // Zap entry. + orig, orig2 = f.setAll(ctx, fd, nil, nil, FDFlags{}) // Zap entry. } + f.mu.Unlock() + + if orig != nil { + f.drop(ctx, orig) + } + if orig2 != nil { + f.dropVFS2(ctx, orig2) + } + return orig, orig2 } // RemoveIf removes all FDs where cond is true. func (f *FDTable) RemoveIf(ctx context.Context, cond func(*fs.File, *vfs.FileDescription, FDFlags) bool) { - f.mu.Lock() - defer f.mu.Unlock() + // TODO(gvisor.dev/issue/1624): Remove fs.File slice. + var files []*fs.File + var filesVFS2 []*vfs.FileDescription + f.mu.Lock() f.forEach(ctx, func(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { if cond(file, fileVFS2, flags) { - f.set(fd, nil, FDFlags{}) // Clear from table. + df, dfVFS2 := f.setAll(ctx, fd, nil, nil, FDFlags{}) // Clear from table. + if df != nil { + files = append(files, df) + } + if dfVFS2 != nil { + filesVFS2 = append(filesVFS2, dfVFS2) + } // Update current available position. if fd < f.next { f.next = fd } } }) + f.mu.Unlock() + + for _, file := range files { + f.drop(ctx, file) + } + + for _, file := range filesVFS2 { + f.dropVFS2(ctx, file) + } } diff --git a/pkg/sentry/kernel/fd_table_test.go b/pkg/sentry/kernel/fd_table_test.go index e3f30ba2a..bf5460083 100644 --- a/pkg/sentry/kernel/fd_table_test.go +++ b/pkg/sentry/kernel/fd_table_test.go @@ -72,7 +72,7 @@ func TestFDTableMany(t *testing.T) { } i := int32(2) - fdTable.Remove(i) + fdTable.Remove(ctx, i) if fds, err := fdTable.NewFDs(ctx, 0, []*fs.File{file}, FDFlags{}); err != nil || fds[0] != i { t.Fatalf("Allocated %v FDs but wanted to allocate %v: %v", i, maxFD, err) } @@ -93,7 +93,7 @@ func TestFDTableOverLimit(t *testing.T) { t.Fatalf("fdTable.NewFDs(maxFD-3, {f,f,f}): got %v, wanted nil", err) } else { for _, fd := range fds { - fdTable.Remove(fd) + fdTable.Remove(ctx, fd) } } @@ -150,13 +150,13 @@ func TestFDTable(t *testing.T) { t.Fatalf("fdTable.Get(2): got a %v, wanted nil", ref) } - ref, _ := fdTable.Remove(1) + ref, _ := fdTable.Remove(ctx, 1) if ref == nil { t.Fatalf("fdTable.Remove(1) for an existing FD: failed, want success") } ref.DecRef(ctx) - if ref, _ := fdTable.Remove(1); ref != nil { + if ref, _ := fdTable.Remove(ctx, 1); ref != nil { t.Fatalf("r.Remove(1) for a removed FD: got success, want failure") } }) diff --git a/pkg/sentry/kernel/fd_table_unsafe.go b/pkg/sentry/kernel/fd_table_unsafe.go index 7fd97dc53..3476551f3 100644 --- a/pkg/sentry/kernel/fd_table_unsafe.go +++ b/pkg/sentry/kernel/fd_table_unsafe.go @@ -18,6 +18,7 @@ import ( "sync/atomic" "unsafe" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/vfs" ) @@ -30,12 +31,21 @@ type descriptorTable struct { slice unsafe.Pointer `state:".(map[int32]*descriptor)"` } -// init initializes the table. -func (f *FDTable) init() { +// initNoLeakCheck initializes the table without enabling leak checking. +// +// This is used when loading an FDTable after S/R, during which the ref count +// object itself will enable leak checking if necessary. +func (f *FDTable) initNoLeakCheck() { var slice []unsafe.Pointer // Empty slice. atomic.StorePointer(&f.slice, unsafe.Pointer(&slice)) } +// init initializes the table with leak checking. +func (f *FDTable) init() { + f.initNoLeakCheck() + f.EnableLeakCheck() +} + // get gets a file entry. // // The boolean indicates whether this was in range. @@ -76,33 +86,37 @@ func (f *FDTable) getAll(fd int32) (*fs.File, *vfs.FileDescription, FDFlags, boo return d.file, d.fileVFS2, d.flags, true } -// set sets an entry. -// -// This handles accounting changes, as well as acquiring and releasing the -// reference needed by the table iff the file is different. +// CurrentMaxFDs returns the number of file descriptors that may be stored in f +// without reallocation. +func (f *FDTable) CurrentMaxFDs() int { + slice := *(*[]unsafe.Pointer)(atomic.LoadPointer(&f.slice)) + return len(slice) +} + +// set sets an entry for VFS1, refer to setAll(). // // Precondition: mu must be held. -func (f *FDTable) set(fd int32, file *fs.File, flags FDFlags) { - f.setAll(fd, file, nil, flags) +func (f *FDTable) set(ctx context.Context, fd int32, file *fs.File, flags FDFlags) *fs.File { + dropFile, _ := f.setAll(ctx, fd, file, nil, flags) + return dropFile } -// setVFS2 sets an entry. -// -// This handles accounting changes, as well as acquiring and releasing the -// reference needed by the table iff the file is different. +// setVFS2 sets an entry for VFS2, refer to setAll(). // // Precondition: mu must be held. -func (f *FDTable) setVFS2(fd int32, file *vfs.FileDescription, flags FDFlags) { - f.setAll(fd, nil, file, flags) +func (f *FDTable) setVFS2(ctx context.Context, fd int32, file *vfs.FileDescription, flags FDFlags) *vfs.FileDescription { + _, dropFile := f.setAll(ctx, fd, nil, file, flags) + return dropFile } -// setAll sets an entry. -// -// This handles accounting changes, as well as acquiring and releasing the -// reference needed by the table iff the file is different. +// setAll sets the file description referred to by fd to file/fileVFS2. If +// file/fileVFS2 are non-nil, it takes a reference on them. If setAll replaces +// an existing file description, it returns it with the FDTable's reference +// transferred to the caller, which must call f.drop/dropVFS2() on the returned +// file after unlocking f.mu. // // Precondition: mu must be held. -func (f *FDTable) setAll(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) { +func (f *FDTable) setAll(ctx context.Context, fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, flags FDFlags) (*fs.File, *vfs.FileDescription) { if file != nil && fileVFS2 != nil { panic("VFS1 and VFS2 files set") } @@ -145,25 +159,25 @@ func (f *FDTable) setAll(fd int32, file *fs.File, fileVFS2 *vfs.FileDescription, } } - // Drop the table reference. + // Adjust used. + switch { + case orig == nil && desc != nil: + atomic.AddInt32(&f.used, 1) + case orig != nil && desc == nil: + atomic.AddInt32(&f.used, -1) + } + if orig != nil { switch { case orig.file != nil: if desc == nil || desc.file != orig.file { - f.drop(orig.file) + return orig.file, nil } case orig.fileVFS2 != nil: if desc == nil || desc.fileVFS2 != orig.fileVFS2 { - f.dropVFS2(orig.fileVFS2) + return nil, orig.fileVFS2 } } } - - // Adjust used. - switch { - case orig == nil && desc != nil: - atomic.AddInt32(&f.used, 1) - case orig != nil && desc == nil: - atomic.AddInt32(&f.used, -1) - } + return nil, nil } diff --git a/pkg/sentry/kernel/fs_context.go b/pkg/sentry/kernel/fs_context.go index 8f2d36d5a..41fb2a784 100644 --- a/pkg/sentry/kernel/fs_context.go +++ b/pkg/sentry/kernel/fs_context.go @@ -18,7 +18,6 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sync" @@ -30,7 +29,7 @@ import ( // // +stateify savable type FSContext struct { - refs.AtomicRefCount + FSContextRefs // mu protects below. mu sync.Mutex `state:"nosave"` @@ -64,7 +63,7 @@ func newFSContext(root, cwd *fs.Dirent, umask uint) *FSContext { cwd: cwd, umask: umask, } - f.EnableLeakCheck("kernel.FSContext") + f.EnableLeakCheck() return &f } @@ -77,95 +76,103 @@ func NewFSContextVFS2(root, cwd vfs.VirtualDentry, umask uint) *FSContext { cwdVFS2: cwd, umask: umask, } - f.EnableLeakCheck("kernel.FSContext") + f.EnableLeakCheck() return &f } -// destroy is the destructor for an FSContext. +// DecRef implements RefCounter.DecRef. // -// This will call DecRef on both root and cwd Dirents. If either call to -// DecRef returns an error, then it will be propagated. If both calls to -// DecRef return an error, then the one from root.DecRef will be propagated. +// When f reaches zero references, DecRef will be called on both root and cwd +// Dirents. // // 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(ctx context.Context) { - // Hold f.mu so that we don't race with RootDirectory() and - // WorkingDirectory(). - f.mu.Lock() - defer f.mu.Unlock() - - if VFS2Enabled { - f.rootVFS2.DecRef(ctx) - f.rootVFS2 = vfs.VirtualDentry{} - f.cwdVFS2.DecRef(ctx) - f.cwdVFS2 = vfs.VirtualDentry{} - } else { - f.root.DecRef(ctx) - f.root = nil - f.cwd.DecRef(ctx) - f.cwd = nil - } -} - -// DecRef implements RefCounter.DecRef with destructor f.destroy. func (f *FSContext) DecRef(ctx context.Context) { - f.DecRefWithDestructor(ctx, f.destroy) + f.FSContextRefs.DecRef(func() { + // Hold f.mu so that we don't race with RootDirectory() and + // WorkingDirectory(). + f.mu.Lock() + defer f.mu.Unlock() + + if VFS2Enabled { + f.rootVFS2.DecRef(ctx) + f.rootVFS2 = vfs.VirtualDentry{} + f.cwdVFS2.DecRef(ctx) + f.cwdVFS2 = vfs.VirtualDentry{} + } else { + f.root.DecRef(ctx) + f.root = nil + f.cwd.DecRef(ctx) + f.cwd = nil + } + }) } // Fork forks this FSContext. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) Fork() *FSContext { f.mu.Lock() defer f.mu.Unlock() if VFS2Enabled { + if !f.cwdVFS2.Ok() { + panic("FSContext.Fork() called after destroy") + } f.cwdVFS2.IncRef() f.rootVFS2.IncRef() } else { + if f.cwd == nil { + panic("FSContext.Fork() called after destroy") + } f.cwd.IncRef() f.root.IncRef() } - return &FSContext{ + ctx := &FSContext{ cwd: f.cwd, root: f.root, cwdVFS2: f.cwdVFS2, rootVFS2: f.rootVFS2, umask: f.umask, } + ctx.EnableLeakCheck() + return ctx } // WorkingDirectory returns the current working directory. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) WorkingDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() - f.cwd.IncRef() + if f.cwd != nil { + f.cwd.IncRef() + } return f.cwd } // WorkingDirectoryVFS2 returns the current working directory. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return an empty vfs.VirtualDentry if called after f is +// destroyed, otherwise it will return a Dirent with a reference taken. func (f *FSContext) WorkingDirectoryVFS2() vfs.VirtualDentry { f.mu.Lock() defer f.mu.Unlock() - f.cwdVFS2.IncRef() + if f.cwdVFS2.Ok() { + f.cwdVFS2.IncRef() + } return f.cwdVFS2 } // SetWorkingDirectory sets the current working directory. // This will take an extra reference on the Dirent. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetWorkingDirectory called with nil dirent") @@ -187,11 +194,15 @@ func (f *FSContext) SetWorkingDirectory(ctx context.Context, d *fs.Dirent) { // SetWorkingDirectoryVFS2 sets the current working directory. // This will take an extra reference on the VirtualDentry. // -// This is not a valid call after destroy. +// This is not a valid call after f is destroyed. func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDentry) { f.mu.Lock() defer f.mu.Unlock() + if !f.cwdVFS2.Ok() { + panic(fmt.Sprintf("FSContext.SetWorkingDirectoryVFS2(%v)) called after destroy", d)) + } + old := f.cwdVFS2 f.cwdVFS2 = d d.IncRef() @@ -200,8 +211,8 @@ func (f *FSContext) SetWorkingDirectoryVFS2(ctx context.Context, d vfs.VirtualDe // RootDirectory returns the current filesystem root. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return nil if called after f is destroyed, otherwise it will return +// a Dirent with a reference taken. func (f *FSContext) RootDirectory() *fs.Dirent { f.mu.Lock() defer f.mu.Unlock() @@ -213,20 +224,22 @@ func (f *FSContext) RootDirectory() *fs.Dirent { // RootDirectoryVFS2 returns the current filesystem root. // -// This will return nil if called after destroy(), otherwise it will return a -// Dirent with a reference taken. +// This will return an empty vfs.VirtualDentry if called after f is +// destroyed, otherwise it will return a Dirent with a reference taken. func (f *FSContext) RootDirectoryVFS2() vfs.VirtualDentry { f.mu.Lock() defer f.mu.Unlock() - f.rootVFS2.IncRef() + if f.rootVFS2.Ok() { + f.rootVFS2.IncRef() + } return f.rootVFS2 } // SetRootDirectory sets the root directory. // This will take an extra reference on the Dirent. // -// This is not a valid call after free. +// This is not a valid call after f is destroyed. func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) { if d == nil { panic("FSContext.SetRootDirectory called with nil dirent") @@ -247,7 +260,7 @@ func (f *FSContext) SetRootDirectory(ctx context.Context, d *fs.Dirent) { // SetRootDirectoryVFS2 sets the root directory. It takes a reference on vd. // -// This is not a valid call after free. +// This is not a valid call after f is destroyed. func (f *FSContext) SetRootDirectoryVFS2(ctx context.Context, vd vfs.VirtualDentry) { if !vd.Ok() { panic("FSContext.SetRootDirectoryVFS2 called with zero-value VirtualDentry") diff --git a/pkg/sentry/kernel/ipc_namespace.go b/pkg/sentry/kernel/ipc_namespace.go index 80a070d7e..b87e40dd1 100644 --- a/pkg/sentry/kernel/ipc_namespace.go +++ b/pkg/sentry/kernel/ipc_namespace.go @@ -15,6 +15,7 @@ package kernel import ( + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/sentry/kernel/semaphore" "gvisor.dev/gvisor/pkg/sentry/kernel/shm" @@ -24,6 +25,8 @@ import ( // // +stateify savable type IPCNamespace struct { + IPCNamespaceRefs + // User namespace which owns this IPC namespace. Immutable. userNS *auth.UserNamespace @@ -33,11 +36,13 @@ type IPCNamespace struct { // NewIPCNamespace creates a new IPC namespace. func NewIPCNamespace(userNS *auth.UserNamespace) *IPCNamespace { - return &IPCNamespace{ + ns := &IPCNamespace{ userNS: userNS, semaphores: semaphore.NewRegistry(userNS), shms: shm.NewRegistry(userNS), } + ns.EnableLeakCheck() + return ns } // SemaphoreRegistry returns the semaphore set registry for this namespace. @@ -50,6 +55,13 @@ func (i *IPCNamespace) ShmRegistry() *shm.Registry { return i.shms } +// DecRef implements refsvfs2.RefCounter.DecRef. +func (i *IPCNamespace) DecRef(ctx context.Context) { + i.IPCNamespaceRefs.DecRef(func() { + i.shms.Release(ctx) + }) +} + // IPCNamespace returns the task's IPC namespace. func (t *Task) IPCNamespace() *IPCNamespace { t.mu.Lock() diff --git a/pkg/sentry/kernel/kcov.go b/pkg/sentry/kernel/kcov.go new file mode 100644 index 000000000..4fcdfc541 --- /dev/null +++ b/pkg/sentry/kernel/kcov.go @@ -0,0 +1,338 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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" + "io" + "sync" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/coverage" + "gvisor.dev/gvisor/pkg/safemem" + "gvisor.dev/gvisor/pkg/sentry/memmap" + "gvisor.dev/gvisor/pkg/sentry/mm" + "gvisor.dev/gvisor/pkg/sentry/pgalloc" + "gvisor.dev/gvisor/pkg/sentry/usage" + "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/usermem" +) + +// kcovAreaSizeMax is the maximum number of uint64 entries allowed in the kcov +// area. On Linux, the maximum is INT_MAX / 8. +const kcovAreaSizeMax = 10 * 1024 * 1024 + +// Kcov provides kernel coverage data to userspace through a memory-mapped +// region, as kcov does in Linux. +// +// To give the illusion that the data is always up to date, we update the shared +// memory every time before we return to userspace. +type Kcov struct { + // mfp provides application memory. It is immutable after creation. + mfp pgalloc.MemoryFileProvider + + // mu protects all of the fields below. + mu sync.RWMutex + + // mode is the current kcov mode. + mode uint8 + + // size is the size of the mapping through which the kernel conveys coverage + // information to userspace. + size uint64 + + // owningTask is the task that currently owns coverage data on the system. The + // interface for kcov essentially requires that coverage is only going to a + // single task. Note that kcov should only generate coverage data for the + // owning task, but we currently generate global coverage. + owningTask *Task + + // count is a locally cached version of the first uint64 in the kcov data, + // which is the number of subsequent entries representing PCs. + // + // It is used with kcovInode.countBlock(), to copy in/out the first element of + // the actual data in an efficient manner, avoid boilerplate, and prevent + // accidental garbage escapes by the temporary counts. + count uint64 + + mappable *mm.SpecialMappable +} + +// NewKcov creates and returns a Kcov instance. +func (k *Kernel) NewKcov() *Kcov { + return &Kcov{ + mfp: k, + } +} + +var coveragePool = sync.Pool{ + New: func() interface{} { + return make([]byte, 0) + }, +} + +// TaskWork implements TaskWorker.TaskWork. +func (kcov *Kcov) TaskWork(t *Task) { + kcov.mu.Lock() + defer kcov.mu.Unlock() + + if kcov.mode != linux.KCOV_MODE_TRACE_PC { + return + } + + rw := &kcovReadWriter{ + mf: kcov.mfp.MemoryFile(), + fr: kcov.mappable.FileRange(), + } + + // Read in the PC count. + if _, err := safemem.ReadFullToBlocks(rw, kcov.countBlock()); err != nil { + panic(fmt.Sprintf("Internal error reading count from kcov area: %v", err)) + } + + rw.off = 8 * (1 + kcov.count) + n := coverage.ConsumeCoverageData(&kcovIOWriter{rw}) + + // Update the pc count, based on the number of entries written. Note that if + // we reached the end of the kcov area, we may not have written everything in + // output. + kcov.count += uint64(n / 8) + rw.off = 0 + if _, err := safemem.WriteFullFromBlocks(rw, kcov.countBlock()); err != nil { + panic(fmt.Sprintf("Internal error writing count to kcov area: %v", err)) + } + + // Re-register for future work. + t.RegisterWork(kcov) +} + +// InitTrace performs the KCOV_INIT_TRACE ioctl. +func (kcov *Kcov) InitTrace(size uint64) error { + kcov.mu.Lock() + defer kcov.mu.Unlock() + + if kcov.mode != linux.KCOV_MODE_DISABLED { + return syserror.EBUSY + } + + // To simplify all the logic around mapping, we require that the length of the + // shared region is a multiple of the system page size. + if (8*size)&(usermem.PageSize-1) != 0 { + return syserror.EINVAL + } + + // We need space for at least two uint64s to hold current position and a + // single PC. + if size < 2 || size > kcovAreaSizeMax { + return syserror.EINVAL + } + + kcov.size = size + kcov.mode = linux.KCOV_MODE_INIT + return nil +} + +// EnableTrace performs the KCOV_ENABLE_TRACE ioctl. +func (kcov *Kcov) EnableTrace(ctx context.Context, traceKind uint8) error { + t := TaskFromContext(ctx) + if t == nil { + panic("kcovInode.EnableTrace() cannot be used outside of a task goroutine") + } + + kcov.mu.Lock() + defer kcov.mu.Unlock() + + // KCOV_ENABLE must be preceded by KCOV_INIT_TRACE and an mmap call. + if kcov.mode != linux.KCOV_MODE_INIT || kcov.mappable == nil { + return syserror.EINVAL + } + + switch traceKind { + case linux.KCOV_TRACE_PC: + kcov.mode = linux.KCOV_MODE_TRACE_PC + case linux.KCOV_TRACE_CMP: + // We do not support KCOV_MODE_TRACE_CMP. + return syserror.ENOTSUP + default: + return syserror.EINVAL + } + + if kcov.owningTask != nil && kcov.owningTask != t { + return syserror.EBUSY + } + + kcov.owningTask = t + t.SetKcov(kcov) + t.RegisterWork(kcov) + + // Clear existing coverage data; the task expects to read only coverage data + // from the time it is activated. + coverage.ClearCoverageData() + return nil +} + +// DisableTrace performs the KCOV_DISABLE_TRACE ioctl. +func (kcov *Kcov) DisableTrace(ctx context.Context) error { + kcov.mu.Lock() + defer kcov.mu.Unlock() + + t := TaskFromContext(ctx) + if t == nil { + panic("kcovInode.EnableTrace() cannot be used outside of a task goroutine") + } + + if t != kcov.owningTask { + return syserror.EINVAL + } + kcov.mode = linux.KCOV_MODE_INIT + kcov.owningTask = nil + if kcov.mappable != nil { + kcov.mappable.DecRef(ctx) + kcov.mappable = nil + } + return nil +} + +// Clear resets the mode and clears the owning task and memory mapping for kcov. +// It is called when the fd corresponding to kcov is closed. Note that the mode +// needs to be set so that the next call to kcov.TaskWork() will exit early. +func (kcov *Kcov) Clear(ctx context.Context) { + kcov.mu.Lock() + kcov.mode = linux.KCOV_MODE_INIT + kcov.owningTask = nil + if kcov.mappable != nil { + kcov.mappable.DecRef(ctx) + kcov.mappable = nil + } + kcov.mu.Unlock() +} + +// OnTaskExit is called when the owning task exits. It is similar to +// kcov.Clear(), except the memory mapping is not cleared, so that the same +// mapping can be used in the future if kcov is enabled again by another task. +func (kcov *Kcov) OnTaskExit() { + kcov.mu.Lock() + kcov.mode = linux.KCOV_MODE_INIT + kcov.owningTask = nil + kcov.mu.Unlock() +} + +// ConfigureMMap is called by the vfs.FileDescription for this kcov instance to +// implement vfs.FileDescription.ConfigureMMap. +func (kcov *Kcov) ConfigureMMap(ctx context.Context, opts *memmap.MMapOpts) error { + kcov.mu.Lock() + defer kcov.mu.Unlock() + + if kcov.mode != linux.KCOV_MODE_INIT { + return syserror.EINVAL + } + + if kcov.mappable == nil { + // Set up the kcov area. + fr, err := kcov.mfp.MemoryFile().Allocate(kcov.size*8, usage.Anonymous) + if err != nil { + return err + } + + // Get the thread id for the mmap name. + t := TaskFromContext(ctx) + if t == nil { + panic("ThreadFromContext returned nil") + } + // For convenience, a special mappable is used here. Note that these mappings + // will look different under /proc/[pid]/maps than they do on Linux. + kcov.mappable = mm.NewSpecialMappable(fmt.Sprintf("[kcov:%d]", t.ThreadID()), kcov.mfp, fr) + } + kcov.mappable.IncRef() + opts.Mappable = kcov.mappable + opts.MappingIdentity = kcov.mappable + return nil +} + +// kcovReadWriter implements safemem.Reader and safemem.Writer. +type kcovReadWriter struct { + off uint64 + mf *pgalloc.MemoryFile + fr memmap.FileRange +} + +// ReadToBlocks implements safemem.Reader.ReadToBlocks. +func (rw *kcovReadWriter) ReadToBlocks(dsts safemem.BlockSeq) (uint64, error) { + if dsts.IsEmpty() { + return 0, nil + } + + // Limit the read to the kcov range and check for overflow. + if rw.fr.Length() <= rw.off { + return 0, io.EOF + } + start := rw.fr.Start + rw.off + end := rw.fr.Start + rw.fr.Length() + if rend := start + dsts.NumBytes(); rend < end { + end = rend + } + + // Get internal mappings. + bs, err := rw.mf.MapInternal(memmap.FileRange{start, end}, usermem.Read) + if err != nil { + return 0, err + } + + // Copy from internal mappings. + n, err := safemem.CopySeq(dsts, bs) + rw.off += n + return n, err +} + +// WriteFromBlocks implements safemem.Writer.WriteFromBlocks. +func (rw *kcovReadWriter) WriteFromBlocks(srcs safemem.BlockSeq) (uint64, error) { + if srcs.IsEmpty() { + return 0, nil + } + + // Limit the write to the kcov area and check for overflow. + if rw.fr.Length() <= rw.off { + return 0, io.EOF + } + start := rw.fr.Start + rw.off + end := rw.fr.Start + rw.fr.Length() + if wend := start + srcs.NumBytes(); wend < end { + end = wend + } + + // Get internal mapping. + bs, err := rw.mf.MapInternal(memmap.FileRange{start, end}, usermem.Write) + if err != nil { + return 0, err + } + + // Copy to internal mapping. + n, err := safemem.CopySeq(bs, srcs) + rw.off += n + return n, err +} + +// kcovIOWriter implements io.Writer as a basic wrapper over kcovReadWriter. +type kcovIOWriter struct { + rw *kcovReadWriter +} + +// Write implements io.Writer.Write. +func (w *kcovIOWriter) Write(p []byte) (int, error) { + bs := safemem.BlockSeqOf(safemem.BlockFromSafeSlice(p)) + n, err := safemem.WriteFullFromBlocks(w.rw, bs) + return int(n), err +} diff --git a/pkg/sentry/kernel/kcov_unsafe.go b/pkg/sentry/kernel/kcov_unsafe.go new file mode 100644 index 000000000..6f8a0266b --- /dev/null +++ b/pkg/sentry/kernel/kcov_unsafe.go @@ -0,0 +1,28 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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 ( + "unsafe" + + "gvisor.dev/gvisor/pkg/safemem" +) + +// countBlock provides a safemem.BlockSeq for kcov.count. +// +// Like k.count, the block returned is protected by k.mu. +func (kcov *Kcov) countBlock() safemem.BlockSeq { + return safemem.BlockSeqOf(safemem.BlockFromSafePointer(unsafe.Pointer(&kcov.count), int(unsafe.Sizeof(kcov.count)))) +} diff --git a/pkg/sentry/kernel/kernel.go b/pkg/sentry/kernel/kernel.go index 1028d13c6..9b2be44d4 100644 --- a/pkg/sentry/kernel/kernel.go +++ b/pkg/sentry/kernel/kernel.go @@ -39,6 +39,7 @@ import ( "time" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/cleanup" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/eventchannel" @@ -220,13 +221,18 @@ type Kernel struct { // danglingEndpoints is used to save / restore tcpip.DanglingEndpoints. danglingEndpoints struct{} `state:".([]tcpip.Endpoint)"` - // sockets is the list of all network sockets the system. Protected by - // extMu. + // sockets is the list of all network sockets in the system. + // Protected by extMu. + // TODO(gvisor.dev/issue/1624): Only used by VFS1. sockets socketList - // nextSocketEntry is the next entry number to use in sockets. Protected + // socketsVFS2 records all network sockets in the system. Protected by + // extMu. + socketsVFS2 map[*vfs.FileDescription]*SocketRecord + + // nextSocketRecord is the next entry number to use in sockets. Protected // by extMu. - nextSocketEntry uint64 + nextSocketRecord uint64 // deviceRegistry is used to save/restore device.SimpleDevices. deviceRegistry struct{} `state:".(*device.Registry)"` @@ -248,7 +254,7 @@ type Kernel struct { // SpecialOpts contains special kernel options. SpecialOpts - // VFS keeps the filesystem state used across the kernel. + // vfs keeps the filesystem state used across the kernel. vfs vfs.VirtualFilesystem // hostMount is the Mount used for file descriptors that were imported @@ -335,7 +341,7 @@ func (k *Kernel) Init(args InitKernelArgs) error { return fmt.Errorf("Timekeeper is nil") } if args.Timekeeper.clocks == nil { - return fmt.Errorf("Must call Timekeeper.SetClocks() before Kernel.Init()") + return fmt.Errorf("must call Timekeeper.SetClocks() before Kernel.Init()") } if args.RootUserNamespace == nil { return fmt.Errorf("RootUserNamespace is nil") @@ -360,7 +366,7 @@ func (k *Kernel) Init(args InitKernelArgs) error { k.useHostCores = true maxCPU, err := hostcpu.MaxPossibleCPU() if err != nil { - return fmt.Errorf("Failed to get maximum CPU number: %v", err) + return fmt.Errorf("failed to get maximum CPU number: %v", err) } minAppCores := uint(maxCPU) + 1 if k.applicationCores < minAppCores { @@ -414,6 +420,8 @@ func (k *Kernel) Init(args InitKernelArgs) error { return fmt.Errorf("failed to create sockfs mount: %v", err) } k.socketMount = socketMount + + k.socketsVFS2 = make(map[*vfs.FileDescription]*SocketRecord) } return nil @@ -422,9 +430,8 @@ func (k *Kernel) Init(args InitKernelArgs) error { // SaveTo saves the state of k to w. // // Preconditions: The kernel must be paused throughout the call to SaveTo. -func (k *Kernel) SaveTo(w wire.Writer) error { +func (k *Kernel) SaveTo(ctx context.Context, w wire.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() @@ -438,38 +445,55 @@ func (k *Kernel) SaveTo(w wire.Writer) error { k.mf.StartEvictions() k.mf.WaitForEvictions() - // Flush write operations on open files so data reaches backing storage. - // This must come after MemoryFile eviction since eviction may cause file - // writes. - if err := k.tasks.flushWritesToFiles(ctx); err != nil { - return err - } + if VFS2Enabled { + // Discard unsavable mappings, such as those for host file descriptors. + if err := k.invalidateUnsavableMappings(ctx); err != nil { + return fmt.Errorf("failed to invalidate unsavable mappings: %v", 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(ctx) + // Prepare filesystems for saving. This must be done after + // invalidateUnsavableMappings(), since dropping memory mappings may + // affect filesystem state (e.g. page cache reference counts). + if err := k.vfs.PrepareSave(ctx); err != nil { + return err + } + } else { + // Flush cached file writes to backing storage. This must come after + // MemoryFile eviction since eviction may cause file writes. + if err := k.flushWritesToFiles(ctx); err != nil { + return err + } - // 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. - if err := k.flushMountSourceRefs(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(ctx) - // Ensure that all inode and mount release operations have completed. - fs.AsyncBarrier() + // 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. + if err := k.flushMountSourceRefs(ctx); err != nil { + return err + } + + // Ensure that all inode and mount release operations have completed. + 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() + // 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) + // 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. + // + // TODO(gvisor.dev/issue/1624): This rationale is believed to be + // obsolete since AIO callbacks are now waited-for by Kernel.Pause(), + // but this order is conservatively retained for VFS1. + if err := k.invalidateUnsavableMappings(ctx); err != nil { + return fmt.Errorf("failed to invalidate unsavable mappings: %v", err) + } } // Save the CPUID FeatureSet before the rest of the kernel so we can @@ -478,14 +502,14 @@ func (k *Kernel) SaveTo(w wire.Writer) error { // // N.B. This will also be saved along with the full kernel save below. cpuidStart := time.Now() - if _, err := state.Save(k.SupervisorContext(), w, k.FeatureSet()); err != nil { + if _, err := state.Save(ctx, w, k.FeatureSet()); err != nil { return err } log.Infof("CPUID save took [%s].", time.Since(cpuidStart)) // Save the kernel state. kernelStart := time.Now() - stats, err := state.Save(k.SupervisorContext(), w, k) + stats, err := state.Save(ctx, w, k) if err != nil { return err } @@ -494,7 +518,7 @@ func (k *Kernel) SaveTo(w wire.Writer) error { // Save the memory file's state. memoryStart := time.Now() - if err := k.mf.SaveTo(k.SupervisorContext(), w); err != nil { + if err := k.mf.SaveTo(ctx, w); err != nil { return err } log.Infof("Memory save took [%s].", time.Since(memoryStart)) @@ -506,6 +530,8 @@ func (k *Kernel) SaveTo(w wire.Writer) error { // flushMountSourceRefs flushes the MountSources for all mounted filesystems // and open FDs. +// +// Preconditions: !VFS2Enabled. func (k *Kernel) flushMountSourceRefs(ctx context.Context) error { // Flush all mount sources for currently mounted filesystems in each task. flushed := make(map[*fs.MountNamespace]struct{}) @@ -533,11 +559,6 @@ func (k *Kernel) flushMountSourceRefs(ctx context.Context) error { // // Precondition: Must be called with the kernel paused. func (ts *TaskSet) forEachFDPaused(ctx context.Context, f func(*fs.File, *vfs.FileDescription) error) (err error) { - // TODO(gvisor.dev/issue/1663): Add save support for VFS2. - if VFS2Enabled { - return nil - } - ts.mu.RLock() defer ts.mu.RUnlock() for t := range ts.Root.tids { @@ -554,9 +575,9 @@ func (ts *TaskSet) forEachFDPaused(ctx context.Context, f func(*fs.File, *vfs.Fi return err } -func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { - // TODO(gvisor.dev/issue/1663): Add save support for VFS2. - return ts.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error { +// Preconditions: !VFS2Enabled. +func (k *Kernel) flushWritesToFiles(ctx context.Context) error { + return k.tasks.forEachFDPaused(ctx, func(file *fs.File, _ *vfs.FileDescription) error { if flags := file.Flags(); !flags.Write { return nil } @@ -578,37 +599,8 @@ func (ts *TaskSet) flushWritesToFiles(ctx context.Context) error { }) } -// 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 -} - +// Preconditions: !VFS2Enabled. func (ts *TaskSet) unregisterEpollWaiters(ctx context.Context) { - // TODO(gvisor.dev/issue/1663): Add save support for VFS2. - if VFS2Enabled { - return - } - ts.mu.RLock() defer ts.mu.RUnlock() @@ -633,8 +625,33 @@ func (ts *TaskSet) unregisterEpollWaiters(ctx context.Context) { } } +// 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 +} + // LoadFrom returns a new Kernel loaded from args. -func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clocks) error { +func (k *Kernel) LoadFrom(ctx context.Context, r wire.Reader, net inet.Stack, clocks sentrytime.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error { loadStart := time.Now() initAppCores := k.applicationCores @@ -645,7 +662,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock // don't need to explicitly install it in the Kernel. cpuidStart := time.Now() var features cpuid.FeatureSet - if _, err := state.Load(k.SupervisorContext(), r, &features); err != nil { + if _, err := state.Load(ctx, r, &features); err != nil { return err } log.Infof("CPUID load took [%s].", time.Since(cpuidStart)) @@ -660,7 +677,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock // Load the kernel state. kernelStart := time.Now() - stats, err := state.Load(k.SupervisorContext(), r, k) + stats, err := state.Load(ctx, r, k) if err != nil { return err } @@ -673,7 +690,7 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock // Load the memory file's state. memoryStart := time.Now() - if err := k.mf.LoadFrom(k.SupervisorContext(), r); err != nil { + if err := k.mf.LoadFrom(ctx, r); err != nil { return err } log.Infof("Memory load took [%s].", time.Since(memoryStart)) @@ -685,11 +702,17 @@ func (k *Kernel) LoadFrom(r wire.Reader, net inet.Stack, clocks sentrytime.Clock net.Resume() } - // Ensure that all pending asynchronous work is complete: - // - namedpipe opening - // - inode file opening - if err := fs.AsyncErrorBarrier(); err != nil { - return err + if VFS2Enabled { + if err := k.vfs.CompleteRestore(ctx, vfsOpts); err != nil { + return err + } + } else { + // Ensure that all pending asynchronous work is complete: + // - namedpipe opening + // - inode file opening + if err := fs.AsyncErrorBarrier(); err != nil { + return err + } } tcpip.AsyncLoading.Wait() @@ -818,7 +841,9 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { case CtxUTSNamespace: return ctx.args.UTSNamespace case CtxIPCNamespace: - return ctx.args.IPCNamespace + ipcns := ctx.args.IPCNamespace + ipcns.IncRef() + return ipcns case auth.CtxCredentials: return ctx.args.Credentials case fs.CtxRoot: @@ -831,14 +856,16 @@ func (ctx *createProcessContext) Value(key interface{}) interface{} { if ctx.args.MountNamespaceVFS2 == nil { return nil } - // MountNamespaceVFS2.Root() takes a reference on the root dirent for us. - return ctx.args.MountNamespaceVFS2.Root() + root := ctx.args.MountNamespaceVFS2.Root() + root.IncRef() + return root case vfs.CtxMountNamespace: if ctx.k.globalInit == nil { return nil } - // MountNamespaceVFS2 takes a reference for us. - return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + mntns.IncRef() + return mntns case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter case inet.CtxStack: @@ -888,19 +915,19 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, opener fsbridge.Lookup fsContext *FSContext mntns *fs.MountNamespace + mntnsVFS2 *vfs.MountNamespace ) if VFS2Enabled { - mntnsVFS2 := args.MountNamespaceVFS2 + mntnsVFS2 = args.MountNamespaceVFS2 if mntnsVFS2 == nil { - // MountNamespaceVFS2 adds a reference to the namespace, which is - // transferred to the new process. + // Add a reference to the namespace, which is transferred to the new process. mntnsVFS2 = k.globalInit.Leader().MountNamespaceVFS2() + mntnsVFS2.IncRef() } // Get the root directory from the MountNamespace. - root := args.MountNamespaceVFS2.Root() - // The call to newFSContext below will take a reference on root, so we - // don't need to hold this one. + root := mntnsVFS2.Root() + root.IncRef() defer root.DecRef(ctx) // Grab the working directory. @@ -952,6 +979,10 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, } tg := k.NewThreadGroup(mntns, args.PIDNamespace, NewSignalHandlers(), linux.SIGCHLD, args.Limits) + cu := cleanup.Make(func() { + tg.Release(ctx) + }) + defer cu.Clean() // Check which file to start from. switch { @@ -1008,16 +1039,17 @@ func (k *Kernel) CreateProcess(args CreateProcessArgs) (*ThreadGroup, ThreadID, UTSNamespace: args.UTSNamespace, IPCNamespace: args.IPCNamespace, AbstractSocketNamespace: args.AbstractSocketNamespace, - MountNamespaceVFS2: args.MountNamespaceVFS2, + MountNamespaceVFS2: mntnsVFS2, ContainerID: args.ContainerID, } - t, err := k.tasks.NewTask(config) + t, err := k.tasks.NewTask(ctx, config) if err != nil { return nil, 0, err } t.traceExecEvent(tc) // Simulate exec for tracing. // Success. + cu.Release() tgid := k.tasks.Root.IDOfThreadGroup(tg) if k.globalInit == nil { k.globalInit = tg @@ -1067,8 +1099,9 @@ func (k *Kernel) Start() error { // pauseTimeLocked pauses all Timers and Timekeeper updates. // -// Preconditions: Any task goroutines running in k must be stopped. k.extMu -// must be locked. +// Preconditions: +// * Any task goroutines running in k must be stopped. +// * k.extMu must be locked. func (k *Kernel) pauseTimeLocked(ctx context.Context) { // k.cpuClockTicker may be nil since Kernel.SaveTo() may be called before // Kernel.Start(). @@ -1111,8 +1144,9 @@ func (k *Kernel) pauseTimeLocked(ctx context.Context) { // 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. +// Preconditions: +// * Any task goroutines running in k must be stopped. +// * k.extMu must be locked. func (k *Kernel) resumeTimeLocked(ctx context.Context) { if k.cpuClockTicker != nil { k.cpuClockTicker.Resume() @@ -1360,8 +1394,9 @@ func (k *Kernel) RootUTSNamespace() *UTSNamespace { return k.rootUTSNamespace } -// RootIPCNamespace returns the root IPCNamespace. +// RootIPCNamespace takes a reference and returns the root IPCNamespace. func (k *Kernel) RootIPCNamespace() *IPCNamespace { + k.rootIPCNamespace.IncRef() return k.rootIPCNamespace } @@ -1506,20 +1541,27 @@ func (k *Kernel) SupervisorContext() context.Context { } } -// SocketEntry represents a socket recorded in Kernel.sockets. It implements +// SocketRecord represents a socket recorded in Kernel.socketsVFS2. +// +// +stateify savable +type SocketRecord struct { + k *Kernel + Sock *refs.WeakRef // TODO(gvisor.dev/issue/1624): Only used by VFS1. + SockVFS2 *vfs.FileDescription // Only used by VFS2. + ID uint64 // Socket table entry number. +} + +// SocketRecordVFS1 represents a socket recorded in Kernel.sockets. It implements // refs.WeakRefUser for sockets stored in the socket table. // // +stateify savable -type SocketEntry struct { +type SocketRecordVFS1 struct { socketEntry - k *Kernel - Sock *refs.WeakRef - SockVFS2 *vfs.FileDescription - ID uint64 // Socket table entry number. + SocketRecord } // WeakRefGone implements refs.WeakRefUser.WeakRefGone. -func (s *SocketEntry) WeakRefGone(context.Context) { +func (s *SocketRecordVFS1) WeakRefGone(context.Context) { s.k.extMu.Lock() s.k.sockets.Remove(s) s.k.extMu.Unlock() @@ -1530,9 +1572,14 @@ func (s *SocketEntry) WeakRefGone(context.Context) { // Precondition: Caller must hold a reference to sock. func (k *Kernel) RecordSocket(sock *fs.File) { k.extMu.Lock() - id := k.nextSocketEntry - k.nextSocketEntry++ - s := &SocketEntry{k: k, ID: id} + id := k.nextSocketRecord + k.nextSocketRecord++ + s := &SocketRecordVFS1{ + SocketRecord: SocketRecord{ + k: k, + ID: id, + }, + } s.Sock = refs.NewWeakRef(sock, s) k.sockets.PushBack(s) k.extMu.Unlock() @@ -1544,29 +1591,45 @@ func (k *Kernel) RecordSocket(sock *fs.File) { // Precondition: Caller must hold a reference to sock. // // Note that the socket table will not hold a reference on the -// vfs.FileDescription, because we do not support weak refs on VFS2 files. +// vfs.FileDescription. func (k *Kernel) RecordSocketVFS2(sock *vfs.FileDescription) { k.extMu.Lock() - id := k.nextSocketEntry - k.nextSocketEntry++ - s := &SocketEntry{ + if _, ok := k.socketsVFS2[sock]; ok { + panic(fmt.Sprintf("Socket %p added twice", sock)) + } + id := k.nextSocketRecord + k.nextSocketRecord++ + s := &SocketRecord{ k: k, ID: id, SockVFS2: sock, } - k.sockets.PushBack(s) + k.socketsVFS2[sock] = s + k.extMu.Unlock() +} + +// DeleteSocketVFS2 removes a VFS2 socket from the system-wide socket table. +func (k *Kernel) DeleteSocketVFS2(sock *vfs.FileDescription) { + k.extMu.Lock() + delete(k.socketsVFS2, sock) k.extMu.Unlock() } // ListSockets returns a snapshot of all sockets. // -// Callers of ListSockets() in VFS2 should use SocketEntry.SockVFS2.TryIncRef() +// Callers of ListSockets() in VFS2 should use SocketRecord.SockVFS2.TryIncRef() // to get a reference on a socket in the table. -func (k *Kernel) ListSockets() []*SocketEntry { +func (k *Kernel) ListSockets() []*SocketRecord { k.extMu.Lock() - var socks []*SocketEntry - for s := k.sockets.Front(); s != nil; s = s.Next() { - socks = append(socks, s) + var socks []*SocketRecord + if VFS2Enabled { + for _, s := range k.socketsVFS2 { + socks = append(socks, s) + } + } else { + for s := k.sockets.Front(); s != nil; s = s.Next() { + socks = append(socks, &s.SocketRecord) + } } k.extMu.Unlock() return socks @@ -1594,7 +1657,9 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { case CtxUTSNamespace: return ctx.k.rootUTSNamespace case CtxIPCNamespace: - return ctx.k.rootIPCNamespace + ipcns := ctx.k.rootIPCNamespace + ipcns.IncRef() + return ipcns case auth.CtxCredentials: // The supervisor context is global root. return auth.NewRootCredentials(ctx.k.rootUserNamespace) @@ -1607,16 +1672,16 @@ func (ctx supervisorContext) Value(key interface{}) interface{} { if ctx.k.globalInit == nil { return vfs.VirtualDentry{} } - mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() - defer mntns.DecRef(ctx) - // Root() takes a reference on the root dirent for us. - return mntns.Root() + root := ctx.k.GlobalInit().Leader().MountNamespaceVFS2().Root() + root.IncRef() + return root case vfs.CtxMountNamespace: if ctx.k.globalInit == nil { return nil } - // MountNamespaceVFS2() takes a reference for us. - return ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + mntns := ctx.k.GlobalInit().Leader().MountNamespaceVFS2() + mntns.IncRef() + return mntns case fs.CtxDirentCacheLimiter: return ctx.k.DirentCacheLimiter case inet.CtxStack: @@ -1697,3 +1762,20 @@ func (k *Kernel) ShmMount() *vfs.Mount { func (k *Kernel) SocketMount() *vfs.Mount { return k.socketMount } + +// Release releases resources owned by k. +// +// Precondition: This should only be called after the kernel is fully +// initialized, e.g. after k.Start() has been called. +func (k *Kernel) Release() { + ctx := k.SupervisorContext() + if VFS2Enabled { + k.hostMount.DecRef(ctx) + k.pipeMount.DecRef(ctx) + k.shmMount.DecRef(ctx) + k.socketMount.DecRef(ctx) + k.vfs.Release(ctx) + } + k.timekeeper.Destroy() + k.vdso.Release(ctx) +} diff --git a/pkg/sentry/kernel/pipe/BUILD b/pkg/sentry/kernel/pipe/BUILD index 449643118..99134e634 100644 --- a/pkg/sentry/kernel/pipe/BUILD +++ b/pkg/sentry/kernel/pipe/BUILD @@ -21,6 +21,7 @@ go_library( "//pkg/amutex", "//pkg/buffer", "//pkg/context", + "//pkg/marshal/primitive", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", diff --git a/pkg/sentry/kernel/pipe/pipe.go b/pkg/sentry/kernel/pipe/pipe.go index 297e8f28f..67beb0ad6 100644 --- a/pkg/sentry/kernel/pipe/pipe.go +++ b/pkg/sentry/kernel/pipe/pipe.go @@ -17,6 +17,7 @@ package pipe import ( "fmt" + "io" "sync/atomic" "syscall" @@ -200,22 +201,22 @@ type readOps struct { // // Precondition: this pipe must have readers. func (p *Pipe) read(ctx context.Context, ops readOps) (int64, error) { - // Don't block for a zero-length read even if the pipe is empty. - if ops.left() == 0 { - return 0, nil - } - p.mu.Lock() defer p.mu.Unlock() return p.readLocked(ctx, ops) } func (p *Pipe) readLocked(ctx context.Context, ops readOps) (int64, error) { + // Don't block for a zero-length read even if the pipe is empty. + if ops.left() == 0 { + return 0, nil + } + // Is the pipe empty? if p.view.Size() == 0 { if !p.HasWriters() { // There are no writers, return EOF. - return 0, nil + return 0, io.EOF } return 0, syserror.ErrWouldBlock } @@ -388,6 +389,10 @@ func (p *Pipe) rwReadiness() waiter.EventMask { func (p *Pipe) queued() int64 { p.mu.Lock() defer p.mu.Unlock() + return p.queuedLocked() +} + +func (p *Pipe) queuedLocked() int64 { return p.view.Size() } diff --git a/pkg/sentry/kernel/pipe/pipe_util.go b/pkg/sentry/kernel/pipe/pipe_util.go index 6d58b682f..f665920cb 100644 --- a/pkg/sentry/kernel/pipe/pipe_util.go +++ b/pkg/sentry/kernel/pipe/pipe_util.go @@ -23,6 +23,7 @@ import ( "gvisor.dev/gvisor/pkg/amutex" "gvisor.dev/gvisor/pkg/buffer" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/usermem" @@ -145,9 +146,14 @@ func (p *Pipe) Ioctl(ctx context.Context, io usermem.IO, args arch.SyscallArgume v = math.MaxInt32 // Silently truncate. } // Copy result to userspace. - _, err := usermem.CopyObjectOut(ctx, io, args[2].Pointer(), int32(v), usermem.IOOpts{ - AddressSpaceActive: true, - }) + iocc := primitive.IOCopyContext{ + IO: io, + Ctx: ctx, + Opts: usermem.IOOpts{ + AddressSpaceActive: true, + }, + } + _, err := primitive.CopyInt32Out(&iocc, args[2].Pointer(), int32(v)) return 0, err default: return 0, syscall.ENOTTY diff --git a/pkg/sentry/kernel/pipe/vfs.go b/pkg/sentry/kernel/pipe/vfs.go index 28f998e45..d96bf253b 100644 --- a/pkg/sentry/kernel/pipe/vfs.go +++ b/pkg/sentry/kernel/pipe/vfs.go @@ -33,6 +33,8 @@ import ( // VFSPipe represents the actual pipe, analagous to an inode. VFSPipes should // not be copied. +// +// +stateify savable type VFSPipe struct { // mu protects the fields below. mu sync.Mutex `state:"nosave"` @@ -67,6 +69,11 @@ func (vp *VFSPipe) ReaderWriterPair(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlag return vp.newFD(mnt, vfsd, linux.O_RDONLY|statusFlags, locks), vp.newFD(mnt, vfsd, linux.O_WRONLY|statusFlags, locks) } +// Allocate implements vfs.FileDescriptionImpl.Allocate. +func (*VFSPipe) Allocate(context.Context, uint64, uint64, uint64) error { + return syserror.ESPIPE +} + // Open opens the pipe represented by vp. func (vp *VFSPipe) Open(ctx context.Context, mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, locks *vfs.FileLocks) (*vfs.FileDescription, error) { vp.mu.Lock() @@ -159,6 +166,8 @@ func (vp *VFSPipe) newFD(mnt *vfs.Mount, vfsd *vfs.Dentry, statusFlags uint32, l // VFSPipeFD implements vfs.FileDescriptionImpl for pipes. It also implements // non-atomic usermem.IO methods, allowing it to be passed as usermem.IO to // other FileDescriptions for splice(2) and tee(2). +// +// +stateify savable type VFSPipeFD struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -232,8 +241,7 @@ func (fd *VFSPipeFD) Ioctl(ctx context.Context, uio usermem.IO, args arch.Syscal // PipeSize implements fcntl(F_GETPIPE_SZ). func (fd *VFSPipeFD) PipeSize() int64 { - // Inline Pipe.FifoSize() rather than calling it with nil Context and - // fs.File and ignoring the returned error (which is always nil). + // Inline Pipe.FifoSize() since we don't have a fs.File. fd.pipe.mu.Lock() defer fd.pipe.mu.Unlock() return fd.pipe.max @@ -244,19 +252,57 @@ func (fd *VFSPipeFD) SetPipeSize(size int64) (int64, error) { return fd.pipe.SetFifoSize(size) } -// IOSequence returns a useremm.IOSequence that reads up to count bytes from, -// or writes up to count bytes to, fd. -func (fd *VFSPipeFD) IOSequence(count int64) usermem.IOSequence { - return usermem.IOSequence{ +// SpliceToNonPipe performs a splice operation from fd to a non-pipe file. +func (fd *VFSPipeFD) SpliceToNonPipe(ctx context.Context, out *vfs.FileDescription, off, count int64) (int64, error) { + fd.pipe.mu.Lock() + defer fd.pipe.mu.Unlock() + + // Cap the sequence at number of bytes actually available. + v := fd.pipe.queuedLocked() + if v < count { + count = v + } + src := usermem.IOSequence{ IO: fd, Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}), } + + var ( + n int64 + err error + ) + if off == -1 { + n, err = out.Write(ctx, src, vfs.WriteOptions{}) + } else { + n, err = out.PWrite(ctx, src, off, vfs.WriteOptions{}) + } + if n > 0 { + fd.pipe.view.TrimFront(n) + } + return n, err +} + +// SpliceFromNonPipe performs a splice operation from a non-pipe file to fd. +func (fd *VFSPipeFD) SpliceFromNonPipe(ctx context.Context, in *vfs.FileDescription, off, count int64) (int64, error) { + fd.pipe.mu.Lock() + defer fd.pipe.mu.Unlock() + + dst := usermem.IOSequence{ + IO: fd, + Addrs: usermem.AddrRangeSeqOf(usermem.AddrRange{0, usermem.Addr(count)}), + } + + if off == -1 { + return in.Read(ctx, dst, vfs.ReadOptions{}) + } + return in.PRead(ctx, dst, off, vfs.ReadOptions{}) } -// CopyIn implements usermem.IO.CopyIn. +// CopyIn implements usermem.IO.CopyIn. Note that it is the caller's +// responsibility to trim fd.pipe.view after the read is completed. func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, opts usermem.IOOpts) (int, error) { origCount := int64(len(dst)) - n, err := fd.pipe.read(ctx, readOps{ + n, err := fd.pipe.readLocked(ctx, readOps{ left: func() int64 { return int64(len(dst)) }, @@ -265,7 +311,6 @@ func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, }, read: func(view *buffer.View) (int64, error) { n, err := view.ReadAt(dst, 0) - view.TrimFront(int64(n)) return int64(n), err }, }) @@ -281,7 +326,7 @@ func (fd *VFSPipeFD) CopyIn(ctx context.Context, addr usermem.Addr, dst []byte, // CopyOut implements usermem.IO.CopyOut. func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, opts usermem.IOOpts) (int, error) { origCount := int64(len(src)) - n, err := fd.pipe.write(ctx, writeOps{ + n, err := fd.pipe.writeLocked(ctx, writeOps{ left: func() int64 { return int64(len(src)) }, @@ -305,7 +350,7 @@ func (fd *VFSPipeFD) CopyOut(ctx context.Context, addr usermem.Addr, src []byte, // ZeroOut implements usermem.IO.ZeroOut. func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int64, opts usermem.IOOpts) (int64, error) { origCount := toZero - n, err := fd.pipe.write(ctx, writeOps{ + n, err := fd.pipe.writeLocked(ctx, writeOps{ left: func() int64 { return toZero }, @@ -326,14 +371,15 @@ func (fd *VFSPipeFD) ZeroOut(ctx context.Context, addr usermem.Addr, toZero int6 return n, err } -// CopyInTo implements usermem.IO.CopyInTo. +// CopyInTo implements usermem.IO.CopyInTo. Note that it is the caller's +// responsibility to trim fd.pipe.view after the read is completed. func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst safemem.Writer, opts usermem.IOOpts) (int64, error) { count := ars.NumBytes() if count == 0 { return 0, nil } origCount := count - n, err := fd.pipe.read(ctx, readOps{ + n, err := fd.pipe.readLocked(ctx, readOps{ left: func() int64 { return count }, @@ -342,7 +388,6 @@ func (fd *VFSPipeFD) CopyInTo(ctx context.Context, ars usermem.AddrRangeSeq, dst }, read: func(view *buffer.View) (int64, error) { n, err := view.ReadToSafememWriter(dst, uint64(count)) - view.TrimFront(int64(n)) return int64(n), err }, }) @@ -362,7 +407,7 @@ func (fd *VFSPipeFD) CopyOutFrom(ctx context.Context, ars usermem.AddrRangeSeq, return 0, nil } origCount := count - n, err := fd.pipe.write(ctx, writeOps{ + n, err := fd.pipe.writeLocked(ctx, writeOps{ left: func() int64 { return count }, diff --git a/pkg/sentry/kernel/ptrace.go b/pkg/sentry/kernel/ptrace.go index 619b0cb7c..1145faf13 100644 --- a/pkg/sentry/kernel/ptrace.go +++ b/pkg/sentry/kernel/ptrace.go @@ -18,6 +18,7 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/mm" "gvisor.dev/gvisor/pkg/syserror" @@ -224,8 +225,9 @@ func (s *ptraceStop) Killable() bool { // 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. +// 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() @@ -270,8 +272,9 @@ func (t *Task) ptraceTrapLocked(code int32) { // 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. +// 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() @@ -301,8 +304,9 @@ func (t *Task) ptraceUnfreeze() { t.ptraceUnfreezeLocked() } -// Preconditions: t must be in a frozen ptraceStop. t's signal mutex must be -// locked. +// Preconditions: +// * t must be in a frozen ptraceStop. +// * t's signal mutex must be locked. func (t *Task) ptraceUnfreezeLocked() { // Do this even if the task has been killed to ensure a panic if t.stop is // nil or not a ptraceStop. @@ -497,8 +501,9 @@ func (t *Task) forgetTracerLocked() { // 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. +// 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 @@ -828,8 +833,9 @@ func (t *Task) ptraceInterrupt(target *Task) error { return nil } -// Preconditions: The TaskSet mutex must be locked for writing. t must have a -// tracer. +// Preconditions: +// * The TaskSet mutex must be locked for writing. +// * t must have a tracer. func (t *Task) ptraceSetOptionsLocked(opts uintptr) error { const valid = uintptr(linux.PTRACE_O_EXITKILL | linux.PTRACE_O_TRACESYSGOOD | @@ -994,18 +1000,15 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { // 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 { + if _, err := word.CopyIn(target.AsCopyContext(usermem.IOOpts{IgnorePermissions: true}), addr); err != nil { return err } - _, err := t.CopyOut(data, word) + _, err := word.CopyOut(t, data) return err case linux.PTRACE_POKETEXT, linux.PTRACE_POKEDATA: - _, err := usermem.CopyObjectOut(t, target.MemoryManager(), addr, t.Arch().Native(uintptr(data)), usermem.IOOpts{ - IgnorePermissions: true, - }) + word := t.Arch().Native(uintptr(data)) + _, err := word.CopyOut(target.AsCopyContext(usermem.IOOpts{IgnorePermissions: true}), addr) return err case linux.PTRACE_GETREGSET: @@ -1073,12 +1076,12 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { if target.ptraceSiginfo == nil { return syserror.EINVAL } - _, err := t.CopyOut(data, target.ptraceSiginfo) + _, err := target.ptraceSiginfo.CopyOut(t, data) return err case linux.PTRACE_SETSIGINFO: var info arch.SignalInfo - if _, err := t.CopyIn(data, &info); err != nil { + if _, err := info.CopyIn(t, data); err != nil { return err } t.tg.pidns.owner.mu.RLock() @@ -1093,7 +1096,8 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { if addr != linux.SignalSetSize { return syserror.EINVAL } - _, err := t.CopyOut(data, target.SignalMask()) + mask := target.SignalMask() + _, err := mask.CopyOut(t, data) return err case linux.PTRACE_SETSIGMASK: @@ -1101,7 +1105,7 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { return syserror.EINVAL } var mask linux.SignalSet - if _, err := t.CopyIn(data, &mask); err != nil { + if _, err := mask.CopyIn(t, data); err != nil { return err } // The target's task goroutine is stopped, so this is safe: @@ -1116,7 +1120,7 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error { case linux.PTRACE_GETEVENTMSG: t.tg.pidns.owner.mu.RLock() defer t.tg.pidns.owner.mu.RUnlock() - _, err := t.CopyOut(usermem.Addr(data), target.ptraceEventMsg) + _, err := primitive.CopyUint64Out(t, usermem.Addr(data), target.ptraceEventMsg) return err // PEEKSIGINFO is unimplemented but seems to have no users anywhere. diff --git a/pkg/sentry/kernel/ptrace_amd64.go b/pkg/sentry/kernel/ptrace_amd64.go index cef1276ec..609ad3941 100644 --- a/pkg/sentry/kernel/ptrace_amd64.go +++ b/pkg/sentry/kernel/ptrace_amd64.go @@ -30,7 +30,7 @@ func (t *Task) ptraceArch(target *Task, req int64, addr, data usermem.Addr) erro if err != nil { return err } - _, err = t.CopyOut(data, n) + _, err = n.CopyOut(t, data) return err case linux.PTRACE_POKEUSR: // aka PTRACE_POKEUSER diff --git a/pkg/sentry/kernel/rseq.go b/pkg/sentry/kernel/rseq.go index 18416643b..2a9023fdf 100644 --- a/pkg/sentry/kernel/rseq.go +++ b/pkg/sentry/kernel/rseq.go @@ -173,8 +173,10 @@ func (t *Task) OldRSeqCPUAddr() usermem.Addr { // SetOldRSeqCPUAddr replaces the address that old 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. +// Preconditions: +// * t.RSeqAvailable() == true. +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) SetOldRSeqCPUAddr(addr usermem.Addr) error { t.oldRSeqCPUAddr = addr @@ -189,8 +191,9 @@ func (t *Task) SetOldRSeqCPUAddr(addr usermem.Addr) error { return nil } -// Preconditions: The caller must be running on the task goroutine. t's -// AddressSpace must be active. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) rseqUpdateCPU() error { if t.rseqAddr == 0 && t.oldRSeqCPUAddr == 0 { t.rseqCPU = -1 @@ -209,8 +212,9 @@ func (t *Task) rseqUpdateCPU() error { return oerr } -// Preconditions: The caller must be running on the task goroutine. t's -// AddressSpace must be active. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) oldRSeqCopyOutCPU() error { if t.oldRSeqCPUAddr == 0 { return nil @@ -222,8 +226,9 @@ func (t *Task) oldRSeqCopyOutCPU() error { return err } -// Preconditions: The caller must be running on the task goroutine. t's -// AddressSpace must be active. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) rseqCopyOutCPU() error { if t.rseqAddr == 0 { return nil @@ -240,8 +245,9 @@ func (t *Task) rseqCopyOutCPU() error { return err } -// Preconditions: The caller must be running on the task goroutine. t's -// AddressSpace must be active. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) rseqClearCPU() error { buf := t.CopyScratchBuffer(8) // CPUIDStart and CPUID are the first two fields in linux.RSeq. @@ -269,8 +275,9 @@ func (t *Task) rseqClearCPU() error { // // See kernel/rseq.c:rseq_ip_fixup for reference. // -// Preconditions: The caller must be running on the task goroutine. t's -// AddressSpace must be active. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t's AddressSpace must be active. func (t *Task) rseqAddrInterrupt() { if t.rseqAddr == 0 { return diff --git a/pkg/sentry/kernel/seccomp.go b/pkg/sentry/kernel/seccomp.go index c38c5a40c..387edfa91 100644 --- a/pkg/sentry/kernel/seccomp.go +++ b/pkg/sentry/kernel/seccomp.go @@ -18,7 +18,6 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/bpf" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/syserror" @@ -27,25 +26,18 @@ import ( const maxSyscallFilterInstructions = 1 << 15 -// 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} +// dataAsBPFInput returns a serialized BPF program, only valid on the current task +// goroutine. +// +// Note: this is called for every syscall, which is a very hot path. +func dataAsBPFInput(t *Task, d *linux.SeccompData) bpf.Input { + buf := t.CopyScratchBuffer(d.SizeBytes()) + d.MarshalUnsafe(buf) + return bpf.InputBytes{ + Data: buf, + // Go-marshal always uses the native byte order. + Order: usermem.ByteOrder, + } } func seccompSiginfo(t *Task, errno, sysno int32, ip usermem.Addr) *arch.SignalInfo { @@ -112,20 +104,20 @@ func (t *Task) checkSeccompSyscall(sysno int32, args arch.SyscallArguments, ip u } 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 := linux.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) { + if i >= len(data.Args) { break } - data.args[i] = arg.Uint64() + data.Args[i] = arg.Uint64() } - input := data.asBPFInput() + input := dataAsBPFInput(t, &data) ret := uint32(linux.SECCOMP_RET_ALLOW) f := t.syscallFilters.Load() diff --git a/pkg/sentry/kernel/semaphore/semaphore.go b/pkg/sentry/kernel/semaphore/semaphore.go index c00fa1138..c39ecfb8f 100644 --- a/pkg/sentry/kernel/semaphore/semaphore.go +++ b/pkg/sentry/kernel/semaphore/semaphore.go @@ -283,6 +283,33 @@ func (s *Set) Change(ctx context.Context, creds *auth.Credentials, owner fs.File return nil } +// GetStat extracts semid_ds information from the set. +func (s *Set) GetStat(creds *auth.Credentials) (*linux.SemidDS, 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 nil, syserror.EACCES + } + + ds := &linux.SemidDS{ + SemPerm: linux.IPCPerm{ + Key: uint32(s.key), + UID: uint32(creds.UserNamespace.MapFromKUID(s.owner.UID)), + GID: uint32(creds.UserNamespace.MapFromKGID(s.owner.GID)), + CUID: uint32(creds.UserNamespace.MapFromKUID(s.creator.UID)), + CGID: uint32(creds.UserNamespace.MapFromKGID(s.creator.GID)), + Mode: uint16(s.perms.LinuxMode()), + Seq: 0, // IPC sequence not supported. + }, + SemOTime: s.opTime.TimeT(), + SemCTime: s.changeTime.TimeT(), + SemNSems: uint64(s.Size()), + } + return ds, 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, pid int32) error { if val < 0 || val > valueMax { @@ -320,7 +347,7 @@ func (s *Set) SetValAll(ctx context.Context, vals []uint16, creds *auth.Credenti } for _, val := range vals { - if val < 0 || val > valueMax { + if val > valueMax { return syserror.ERANGE } } diff --git a/pkg/sentry/kernel/sessions.go b/pkg/sentry/kernel/sessions.go index 5c4c622c2..df5c8421b 100644 --- a/pkg/sentry/kernel/sessions.go +++ b/pkg/sentry/kernel/sessions.go @@ -16,8 +16,6 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/syserror" ) @@ -32,7 +30,7 @@ type ProcessGroupID ThreadID // // +stateify savable type Session struct { - refs refs.AtomicRefCount + SessionRefs // leader is the originator of the Session. // @@ -62,16 +60,11 @@ type Session struct { sessionEntry } -// incRef grabs a reference. -func (s *Session) incRef() { - s.refs.IncRef() -} - -// decRef drops a reference. +// DecRef drops a reference. // // Precondition: callers must hold TaskSet.mu for writing. -func (s *Session) decRef() { - s.refs.DecRefWithDestructor(nil, func(context.Context) { +func (s *Session) DecRef() { + s.SessionRefs.DecRef(func() { // Remove translations from the leader. for ns := s.leader.pidns; ns != nil; ns = ns.parent { id := ns.sids[s] @@ -88,7 +81,7 @@ func (s *Session) decRef() { // // +stateify savable type ProcessGroup struct { - refs refs.AtomicRefCount // not exported. + refs ProcessGroupRefs // originator is the originator of the group. // @@ -163,7 +156,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) { } alive := true - pg.refs.DecRefWithDestructor(nil, func(context.Context) { + pg.refs.DecRef(func() { alive = false // don't bother with handleOrphan. // Remove translations from the originator. @@ -175,7 +168,7 @@ func (pg *ProcessGroup) decRefWithParent(parentPG *ProcessGroup) { // Remove the list of process groups. pg.session.processGroups.Remove(pg) - pg.session.decRef() + pg.session.DecRef() }) if alive { pg.handleOrphan() @@ -302,7 +295,7 @@ func (tg *ThreadGroup) createSession() error { id: SessionID(id), leader: tg, } - s.refs.EnableLeakCheck("kernel.Session") + s.EnableLeakCheck() // Create a new ProcessGroup, belonging to that Session. // This also has a single reference (assigned below). @@ -316,7 +309,7 @@ func (tg *ThreadGroup) createSession() error { session: s, ancestors: 0, } - pg.refs.EnableLeakCheck("kernel.ProcessGroup") + pg.refs.EnableLeakCheck() // Tie them and return the result. s.processGroups.PushBack(pg) @@ -396,13 +389,13 @@ func (tg *ThreadGroup) CreateProcessGroup() error { // // We manually adjust the ancestors if the parent is in the same // session. - tg.processGroup.session.incRef() + tg.processGroup.session.IncRef() pg := ProcessGroup{ id: ProcessGroupID(id), originator: tg, session: tg.processGroup.session, } - pg.refs.EnableLeakCheck("kernel.ProcessGroup") + pg.refs.EnableLeakCheck() if tg.leader.parent != nil && tg.leader.parent.tg.processGroup.session == pg.session { pg.ancestors++ diff --git a/pkg/sentry/kernel/shm/BUILD b/pkg/sentry/kernel/shm/BUILD index c211fc8d0..80a592c8f 100644 --- a/pkg/sentry/kernel/shm/BUILD +++ b/pkg/sentry/kernel/shm/BUILD @@ -1,12 +1,25 @@ load("//tools:defs.bzl", "go_library") +load("//tools/go_generics:defs.bzl", "go_template_instance") package(licenses = ["notice"]) +go_template_instance( + name = "shm_refs", + out = "shm_refs.go", + package = "shm", + prefix = "Shm", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "Shm", + }, +) + go_library( name = "shm", srcs = [ "device.go", "shm.go", + "shm_refs.go", ], visibility = ["//pkg/sentry:internal"], deps = [ @@ -14,6 +27,7 @@ go_library( "//pkg/context", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/kernel/auth", diff --git a/pkg/sentry/kernel/shm/shm.go b/pkg/sentry/kernel/shm/shm.go index 13ec7afe0..ebbebf46b 100644 --- a/pkg/sentry/kernel/shm/shm.go +++ b/pkg/sentry/kernel/shm/shm.go @@ -39,7 +39,6 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -252,7 +251,7 @@ func (r *Registry) newShm(ctx context.Context, pid int32, key Key, creator fs.Fi creatorPID: pid, changeTime: ktime.NowFromContext(ctx), } - shm.EnableLeakCheck("kernel.Shm") + shm.EnableLeakCheck() // Find the next available ID. for id := r.lastIDUsed + 1; id != r.lastIDUsed; id++ { @@ -322,9 +321,32 @@ func (r *Registry) remove(s *Shm) { r.totalPages -= s.effectiveSize / usermem.PageSize } +// Release drops the self-reference of each active shm segment in the registry. +// It is called when the kernel.IPCNamespace containing r is being destroyed. +func (r *Registry) Release(ctx context.Context) { + // Because Shm.DecRef() may acquire the same locks, collect the segments to + // release first. Note that this should not race with any updates to r, since + // the IPC namespace containing it has no more references. + toRelease := make([]*Shm, 0) + r.mu.Lock() + for _, s := range r.keysToShms { + s.mu.Lock() + if !s.pendingDestruction { + toRelease = append(toRelease, s) + } + s.mu.Unlock() + } + r.mu.Unlock() + + for _, s := range toRelease { + r.dissociateKey(s) + s.DecRef(ctx) + } +} + // Shm represents a single shared memory segment. // -// Shm segment are backed directly by an allocation from platform memory. +// Shm segments are backed directly by an allocation from platform memory. // Segments are always mapped as a whole, greatly simplifying how mappings are // tracked. However note that mremap and munmap calls may cause the vma for a // segment to become fragmented; which requires special care when unmapping a @@ -337,14 +359,14 @@ func (r *Registry) remove(s *Shm) { // // +stateify savable type Shm struct { - // AtomicRefCount tracks the number of references to this segment. + // ShmRefs tracks the number of references to this segment. // // A segment holds a reference to itself until it is marked for // destruction. // // In addition to direct users, the MemoryManager will hold references // via MappingIdentity. - refs.AtomicRefCount + ShmRefs mfp pgalloc.MemoryFileProvider @@ -428,11 +450,14 @@ func (s *Shm) InodeID() uint64 { return uint64(s.ID) } -// DecRef overrides refs.RefCount.DecRef with a destructor. +// DecRef drops a reference on s. // // Precondition: Caller must not hold s.mu. func (s *Shm) DecRef(ctx context.Context) { - s.DecRefWithDestructor(ctx, s.destroy) + s.ShmRefs.DecRef(func() { + s.mfp.MemoryFile().DecRef(s.fr) + s.registry.remove(s) + }) } // Msync implements memmap.MappingIdentity.Msync. Msync is a no-op for shm @@ -642,11 +667,6 @@ func (s *Shm) Set(ctx context.Context, ds *linux.ShmidDS) error { return nil } -func (s *Shm) destroy(context.Context) { - s.mfp.MemoryFile().DecRef(s.fr) - s.registry.remove(s) -} - // MarkDestroyed marks a segment for destruction. The segment is actually // destroyed once it has no references. MarkDestroyed may be called multiple // times, and is safe to call after a segment has already been destroyed. See @@ -655,17 +675,20 @@ func (s *Shm) MarkDestroyed(ctx context.Context) { s.registry.dissociateKey(s) s.mu.Lock() - defer s.mu.Unlock() - if !s.pendingDestruction { - s.pendingDestruction = true - // Drop the self-reference so destruction occurs when all - // external references are gone. - // - // N.B. This cannot be the final DecRef, as the caller also - // holds a reference. - s.DecRef(ctx) + if s.pendingDestruction { + s.mu.Unlock() return } + s.pendingDestruction = true + s.mu.Unlock() + + // Drop the self-reference so destruction occurs when all + // external references are gone. + // + // N.B. This cannot be the final DecRef, as the caller also + // holds a reference. + s.DecRef(ctx) + return } // checkOwnership verifies whether a segment may be accessed by ctx as an diff --git a/pkg/sentry/kernel/signalfd/BUILD b/pkg/sentry/kernel/signalfd/BUILD index 3eb78e91b..76d472292 100644 --- a/pkg/sentry/kernel/signalfd/BUILD +++ b/pkg/sentry/kernel/signalfd/BUILD @@ -8,7 +8,6 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/sentry/fs", "//pkg/sentry/fs/anon", diff --git a/pkg/sentry/kernel/signalfd/signalfd.go b/pkg/sentry/kernel/signalfd/signalfd.go index b07e1c1bd..78f718cfe 100644 --- a/pkg/sentry/kernel/signalfd/signalfd.go +++ b/pkg/sentry/kernel/signalfd/signalfd.go @@ -17,7 +17,6 @@ package signalfd import ( "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/anon" @@ -103,8 +102,7 @@ func (s *SignalOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS } // Copy out the signal info using the specified format. - var buf [128]byte - binary.Marshal(buf[:0], usermem.ByteOrder, &linux.SignalfdSiginfo{ + infoNative := linux.SignalfdSiginfo{ Signo: uint32(info.Signo), Errno: info.Errno, Code: info.Code, @@ -113,9 +111,13 @@ func (s *SignalOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS Status: info.Status(), Overrun: uint32(info.Overrun()), Addr: info.Addr(), - }) - n, err := dst.CopyOut(ctx, buf[:]) - return int64(n), err + } + n, err := infoNative.WriteTo(dst.Writer(ctx)) + if err == usermem.ErrEndOfIOSequence { + // Partial copy-out ok. + err = nil + } + return n, err } // Readiness implements waiter.Waitable.Readiness. diff --git a/pkg/sentry/kernel/syscalls.go b/pkg/sentry/kernel/syscalls.go index 413111faf..332bdb8e8 100644 --- a/pkg/sentry/kernel/syscalls.go +++ b/pkg/sentry/kernel/syscalls.go @@ -348,6 +348,16 @@ func (s *SyscallTable) LookupName(sysno uintptr) string { return fmt.Sprintf("sys_%d", sysno) // Unlikely. } +// LookupNo looks up a syscall number by name. +func (s *SyscallTable) LookupNo(name string) (uintptr, error) { + for i, syscall := range s.Table { + if syscall.Name == name { + return uintptr(i), nil + } + } + return 0, fmt.Errorf("syscall %q not found", name) +} + // LookupEmulate looks up an emulation syscall number. func (s *SyscallTable) LookupEmulate(addr usermem.Addr) (uintptr, bool) { sysno, ok := s.Emulate[addr] diff --git a/pkg/sentry/kernel/task.go b/pkg/sentry/kernel/task.go index 5aee699e7..037971393 100644 --- a/pkg/sentry/kernel/task.go +++ b/pkg/sentry/kernel/task.go @@ -574,6 +574,11 @@ type Task struct { // // startTime is protected by mu. startTime ktime.Time + + // kcov is the kcov instance providing code coverage owned by this task. + // + // kcov is exclusive to the task goroutine. + kcov *Kcov } func (t *Task) savePtraceTracer() *Task { @@ -651,7 +656,9 @@ func (t *Task) Value(key interface{}) interface{} { case CtxUTSNamespace: return t.utsns case CtxIPCNamespace: - return t.ipcns + ipcns := t.IPCNamespace() + ipcns.IncRef() + return ipcns case CtxTask: return t case auth.CtxCredentials: @@ -730,7 +737,6 @@ func (t *Task) SyscallRestartBlock() SyscallRestartBlock { func (t *Task) IsChrooted() bool { if VFS2Enabled { realRoot := t.mountNamespaceVFS2.Root() - defer realRoot.DecRef(t) root := t.fsContext.RootDirectoryVFS2() defer root.DecRef(t) return root != realRoot @@ -863,7 +869,6 @@ func (t *Task) MountNamespace() *fs.MountNamespace { func (t *Task) MountNamespaceVFS2() *vfs.MountNamespace { t.mu.Lock() defer t.mu.Unlock() - t.mountNamespaceVFS2.IncRef() return t.mountNamespaceVFS2 } @@ -903,3 +908,16 @@ func (t *Task) UID() uint32 { func (t *Task) GID() uint32 { return uint32(t.Credentials().EffectiveKGID) } + +// SetKcov sets the kcov instance associated with t. +func (t *Task) SetKcov(k *Kcov) { + t.kcov = k +} + +// ResetKcov clears the kcov instance associated with t. +func (t *Task) ResetKcov() { + if t.kcov != nil { + t.kcov.OnTaskExit() + t.kcov = nil + } +} diff --git a/pkg/sentry/kernel/task_clone.go b/pkg/sentry/kernel/task_clone.go index 9d7a9128f..682080c14 100644 --- a/pkg/sentry/kernel/task_clone.go +++ b/pkg/sentry/kernel/task_clone.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bpf" + "gvisor.dev/gvisor/pkg/cleanup" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -203,7 +204,13 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { // Note that "If CLONE_NEWIPC is set, then create the process in a new IPC // namespace" ipcns = NewIPCNamespace(userns) + } else { + ipcns.IncRef() } + cu := cleanup.Make(func() { + ipcns.DecRef(t) + }) + defer cu.Clean() netns := t.NetworkNamespace() if opts.NewNetworkNamespace { @@ -214,12 +221,18 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { mntnsVFS2 := t.mountNamespaceVFS2 if mntnsVFS2 != nil { mntnsVFS2.IncRef() + cu.Add(func() { + mntnsVFS2.DecRef(t) + }) } tc, err := t.tc.Fork(t, t.k, !opts.NewAddressSpace) if err != nil { return 0, nil, err } + cu.Add(func() { + tc.release() + }) // clone() returns 0 in the child. tc.Arch.SetReturn(0) if opts.Stack != 0 { @@ -295,11 +308,11 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { } else { cfg.InheritParent = t } - nt, err := t.tg.pidns.owner.NewTask(cfg) + nt, err := t.tg.pidns.owner.NewTask(t, cfg) + // If NewTask succeeds, we transfer references to nt. If NewTask fails, it does + // the cleanup for us. + cu.Release() if err != nil { - if opts.NewThreadGroup { - tg.release(t) - } return 0, nil, err } @@ -341,12 +354,12 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) { 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{}) + ctid := nt.ThreadID() + ctid.CopyOut(nt.AsCopyContext(usermem.IOOpts{AddressSpaceActive: false}), opts.ChildTID) } ntid := t.tg.pidns.IDOfTask(nt) if opts.ParentSetTID { - t.CopyOut(opts.ParentTID, ntid) + ntid.CopyOut(t, opts.ParentTID) } kind := ptraceCloneKindClone @@ -509,6 +522,7 @@ func (t *Task) Unshare(opts *SharingOptions) error { } // Note that "If CLONE_NEWIPC is set, then create the process in a new IPC // namespace" + t.ipcns.DecRef(t) t.ipcns = NewIPCNamespace(creds.UserNamespace) } var oldFDTable *FDTable diff --git a/pkg/sentry/kernel/task_context.go b/pkg/sentry/kernel/task_context.go index 9fa528384..d1136461a 100644 --- a/pkg/sentry/kernel/task_context.go +++ b/pkg/sentry/kernel/task_context.go @@ -126,7 +126,11 @@ func (t *Task) SyscallTable() *SyscallTable { // 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())} + return &arch.Stack{ + Arch: t.Arch(), + IO: t.MemoryManager(), + Bottom: usermem.Addr(t.Arch().Stack()), + } } // LoadTaskImage loads a specified file into a new TaskContext. diff --git a/pkg/sentry/kernel/task_exec.go b/pkg/sentry/kernel/task_exec.go index 5e4fb3e3a..412d471d3 100644 --- a/pkg/sentry/kernel/task_exec.go +++ b/pkg/sentry/kernel/task_exec.go @@ -237,9 +237,10 @@ func (r *runSyscallAfterExecStop) execute(t *Task) taskRunState { // 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. +// 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 { diff --git a/pkg/sentry/kernel/task_exit.go b/pkg/sentry/kernel/task_exit.go index c165d6cb1..ce7b9641d 100644 --- a/pkg/sentry/kernel/task_exit.go +++ b/pkg/sentry/kernel/task_exit.go @@ -239,6 +239,8 @@ func (*runExitMain) execute(t *Task) taskRunState { t.traceExitEvent() lastExiter := t.exitThreadGroup() + t.ResetKcov() + // 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 { @@ -246,7 +248,8 @@ func (*runExitMain) execute(t *Task) taskRunState { 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 { + zero := ThreadID(0) + if _, err := zero.CopyOut(t, t.cleartid); err == nil { t.Futex().Wake(t, t.cleartid, false, ^uint32(0), 1) } // If the CopyOut fails, there's nothing we can do. @@ -277,12 +280,13 @@ func (*runExitMain) execute(t *Task) taskRunState { t.mountNamespaceVFS2.DecRef(t) t.mountNamespaceVFS2 = nil } + t.ipcns.DecRef(t) t.mu.Unlock() // If this is the last task to exit from the thread group, release the // thread group's resources. if lastExiter { - t.tg.release(t) + t.tg.Release(t) } // Detach tracees. diff --git a/pkg/sentry/kernel/task_futex.go b/pkg/sentry/kernel/task_futex.go index 4b535c949..c80391475 100644 --- a/pkg/sentry/kernel/task_futex.go +++ b/pkg/sentry/kernel/task_futex.go @@ -16,6 +16,7 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/usermem" ) @@ -87,7 +88,7 @@ func (t *Task) exitRobustList() { return } - next := rl.List + next := primitive.Uint64(rl.List) done := 0 var pendingLockAddr usermem.Addr if rl.ListOpPending != 0 { @@ -99,12 +100,12 @@ func (t *Task) exitRobustList() { // We traverse to the next element of the list before we // actually wake anything. This prevents the race where waking // this futex causes a modification of the list. - thisLockAddr := usermem.Addr(next + rl.FutexOffset) + thisLockAddr := usermem.Addr(uint64(next) + rl.FutexOffset) // Try to decode the next element in the list before waking the // current futex. But don't check the error until after we've // woken the current futex. Linux does it in this order too - _, nextErr := t.CopyIn(usermem.Addr(next), &next) + _, nextErr := next.CopyIn(t, usermem.Addr(next)) // Wakeup the current futex if it's not pending. if thisLockAddr != pendingLockAddr { diff --git a/pkg/sentry/kernel/task_run.go b/pkg/sentry/kernel/task_run.go index aa3a573c0..8dc3fec90 100644 --- a/pkg/sentry/kernel/task_run.go +++ b/pkg/sentry/kernel/task_run.go @@ -141,7 +141,7 @@ func (*runApp) handleCPUIDInstruction(t *Task) error { region := trace.StartRegion(t.traceContext, cpuidRegion) expected := arch.CPUIDInstruction[:] found := make([]byte, len(expected)) - _, err := t.CopyIn(usermem.Addr(t.Arch().IP()), &found) + _, err := t.CopyInBytes(usermem.Addr(t.Arch().IP()), found) if err == nil && bytes.Equal(expected, found) { // Skip the cpuid instruction. t.Arch().CPUIDEmulate(t) diff --git a/pkg/sentry/kernel/task_sched.go b/pkg/sentry/kernel/task_sched.go index 09366b60c..52c55d13d 100644 --- a/pkg/sentry/kernel/task_sched.go +++ b/pkg/sentry/kernel/task_sched.go @@ -133,9 +133,10 @@ func (t *Task) accountTaskGoroutineEnter(state TaskGoroutineState) { } } -// Preconditions: The caller must be running on the task goroutine, and leaving -// a state indicated by a previous call to -// t.accountTaskGoroutineEnter(state). +// Preconditions: +// * The caller must be running on the task goroutine +// * The caller must be leaving a state indicated by a previous call to +// t.accountTaskGoroutineEnter(state). func (t *Task) accountTaskGoroutineLeave(state TaskGoroutineState) { if state != TaskGoroutineRunningApp { // Task is unblocking/continuing. @@ -191,8 +192,8 @@ func (tg *ThreadGroup) CPUStats() usage.CPUStats { return tg.cpuStatsAtLocked(tg.leader.k.CPUClockNow()) } -// Preconditions: As for TaskGoroutineSchedInfo.userTicksAt. The TaskSet mutex -// must be locked. +// Preconditions: Same as TaskGoroutineSchedInfo.userTicksAt, plus: +// * The TaskSet mutex must be locked. func (tg *ThreadGroup) cpuStatsAtLocked(now uint64) usage.CPUStats { stats := tg.exitedCPUStats // Account for live tasks. diff --git a/pkg/sentry/kernel/task_signals.go b/pkg/sentry/kernel/task_signals.go index d6a2040bc..ebdb83061 100644 --- a/pkg/sentry/kernel/task_signals.go +++ b/pkg/sentry/kernel/task_signals.go @@ -259,7 +259,11 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct) // 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(), mm, sp} + st := &arch.Stack{ + Arch: t.Arch(), + IO: mm, + Bottom: sp, + } mask := t.signalMask if t.haveSavedSignalMask { mask = t.savedSignalMask @@ -319,8 +323,9 @@ func (t *Task) SignalReturn(rt bool) (*SyscallControl, error) { // Sigtimedwait implements the semantics of sigtimedwait(2). // -// Preconditions: The caller must be running on the task goroutine. t.exitState -// < TaskExitZombie. +// Preconditions: +// * The caller must be running on the task goroutine. +// * t.exitState < TaskExitZombie. func (t *Task) Sigtimedwait(set linux.SignalSet, timeout time.Duration) (*arch.SignalInfo, error) { // set is the set of signals we're interested in; invert it to get the set // of signals to block. @@ -584,8 +589,9 @@ func (t *Task) SignalMask() linux.SignalSet { // SetSignalMask sets t's signal mask. // -// Preconditions: SetSignalMask can only be called by the task goroutine. -// t.exitState < TaskExitZombie. +// Preconditions: +// * The caller must be running on 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. @@ -631,7 +637,7 @@ func (t *Task) setSignalMaskLocked(mask linux.SignalSet) { // SetSavedSignalMask sets the saved signal mask (see Task.savedSignalMask's // comment). // -// Preconditions: SetSavedSignalMask can only be called by the task goroutine. +// Preconditions: The caller must be running on the task goroutine. func (t *Task) SetSavedSignalMask(mask linux.SignalSet) { t.savedSignalMask = mask t.haveSavedSignalMask = true diff --git a/pkg/sentry/kernel/task_start.go b/pkg/sentry/kernel/task_start.go index 64c1e120a..8e28230cc 100644 --- a/pkg/sentry/kernel/task_start.go +++ b/pkg/sentry/kernel/task_start.go @@ -16,6 +16,7 @@ package kernel import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -98,14 +99,18 @@ type TaskConfig struct { // NewTask creates a new task defined by cfg. // // NewTask does not start the returned task; the caller must call Task.Start. -func (ts *TaskSet) NewTask(cfg *TaskConfig) (*Task, error) { +// +// If successful, NewTask transfers references held by cfg to the new task. +// Otherwise, NewTask releases them. +func (ts *TaskSet) NewTask(ctx context.Context, cfg *TaskConfig) (*Task, error) { t, err := ts.newTask(cfg) if err != nil { cfg.TaskContext.release() - cfg.FSContext.DecRef(t) - cfg.FDTable.DecRef(t) + cfg.FSContext.DecRef(ctx) + cfg.FDTable.DecRef(ctx) + cfg.IPCNamespace.DecRef(ctx) if cfg.MountNamespaceVFS2 != nil { - cfg.MountNamespaceVFS2.DecRef(t) + cfg.MountNamespaceVFS2.DecRef(ctx) } return nil, err } diff --git a/pkg/sentry/kernel/task_stop.go b/pkg/sentry/kernel/task_stop.go index 296735d32..a35948a5f 100644 --- a/pkg/sentry/kernel/task_stop.go +++ b/pkg/sentry/kernel/task_stop.go @@ -99,8 +99,9 @@ type TaskStop interface { // 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. +// Preconditions: +// * The caller must be running on the task goroutine. +// * The task must not already be in an internal stop (i.e. t.stop == nil). func (t *Task) beginInternalStop(s TaskStop) { t.tg.pidns.owner.mu.RLock() defer t.tg.pidns.owner.mu.RUnlock() @@ -109,8 +110,8 @@ func (t *Task) beginInternalStop(s TaskStop) { t.beginInternalStopLocked(s) } -// Preconditions: The signal mutex must be locked. All preconditions for -// Task.beginInternalStop also apply. +// Preconditions: Same as beginInternalStop, plus: +// * The signal mutex must be locked. 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)) @@ -128,8 +129,9 @@ func (t *Task) beginInternalStopLocked(s TaskStop) { // 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). +// 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") diff --git a/pkg/sentry/kernel/task_syscall.go b/pkg/sentry/kernel/task_syscall.go index 2dbf86547..0141459e7 100644 --- a/pkg/sentry/kernel/task_syscall.go +++ b/pkg/sentry/kernel/task_syscall.go @@ -22,6 +22,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bits" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/metric" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/memmap" @@ -287,7 +288,7 @@ 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 { + if _, err := caller.CopyIn(t, usermem.Addr(t.Arch().Stack())); err != nil { t.Debugf("vsyscall %d: error reading return address from stack: %v", sysno, err) t.forceSignal(linux.SIGSEGV, false /* unconditional */) t.SendSignal(SignalInfoPriv(linux.SIGSEGV)) @@ -323,7 +324,7 @@ func (t *Task) doVsyscall(addr usermem.Addr, sysno uintptr) taskRunState { type runVsyscallAfterPtraceEventSeccomp struct { addr usermem.Addr sysno uintptr - caller interface{} + caller marshal.Marshallable } func (r *runVsyscallAfterPtraceEventSeccomp) execute(t *Task) taskRunState { @@ -346,7 +347,7 @@ func (r *runVsyscallAfterPtraceEventSeccomp) execute(t *Task) taskRunState { return t.doVsyscallInvoke(sysno, t.Arch().SyscallArgs(), r.caller) } -func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, caller interface{}) taskRunState { +func (t *Task) doVsyscallInvoke(sysno uintptr, args arch.SyscallArguments, caller marshal.Marshallable) 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) diff --git a/pkg/sentry/kernel/task_usermem.go b/pkg/sentry/kernel/task_usermem.go index b02044ad2..ce134bf54 100644 --- a/pkg/sentry/kernel/task_usermem.go +++ b/pkg/sentry/kernel/task_usermem.go @@ -18,6 +18,7 @@ import ( "math" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -43,17 +44,6 @@ func (t *Task) Deactivate() { } } -// 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. // @@ -64,17 +54,6 @@ func (t *Task) CopyInBytes(addr usermem.Addr, dst []byte) (int, error) { }) } -// 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. // @@ -114,7 +93,7 @@ func (t *Task) CopyInVector(addr usermem.Addr, maxElemSize, maxTotalSize int) ([ var v []string for { argAddr := t.Arch().Native(0) - if _, err := t.CopyIn(addr, argAddr); err != nil { + if _, err := argAddr.CopyIn(t, addr); err != nil { return v, err } if t.Arch().Value(argAddr) == 0 { @@ -143,8 +122,9 @@ func (t *Task) CopyInVector(addr usermem.Addr, maxElemSize, maxTotalSize int) ([ // 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. +// Preconditions: Same as usermem.IO.CopyOut, plus: +// * 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: @@ -191,8 +171,9 @@ func (t *Task) CopyOutIovecs(addr usermem.Addr, src usermem.AddrRangeSeq) error // 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. +// Preconditions: Same as usermem.IO.CopyIn, plus: +// * 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 @@ -284,7 +265,7 @@ func (t *Task) SingleIOSequence(addr usermem.Addr, length int, opts usermem.IOOp // // IovecsIOSequence is analogous to Linux's lib/iov_iter.c:import_iovec(). // -// Preconditions: As for Task.CopyInIovecs. +// Preconditions: Same as 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 @@ -299,3 +280,30 @@ func (t *Task) IovecsIOSequence(addr usermem.Addr, iovcnt int, opts usermem.IOOp Opts: opts, }, nil } + +// copyContext implements marshal.CopyContext. It wraps a task to allow copying +// memory to and from the task memory with custom usermem.IOOpts. +type copyContext struct { + *Task + opts usermem.IOOpts +} + +// AsCopyContext wraps the task and returns it as CopyContext. +func (t *Task) AsCopyContext(opts usermem.IOOpts) marshal.CopyContext { + return ©Context{t, opts} +} + +// CopyInString copies a string in from the task's memory. +func (t *copyContext) CopyInString(addr usermem.Addr, maxLen int) (string, error) { + return usermem.CopyStringIn(t, t.MemoryManager(), addr, maxLen, t.opts) +} + +// CopyInBytes copies task memory into dst from an IO context. +func (t *copyContext) CopyInBytes(addr usermem.Addr, dst []byte) (int, error) { + return t.MemoryManager().CopyIn(t, addr, dst, t.opts) +} + +// CopyOutBytes copies src into task memoryfrom an IO context. +func (t *copyContext) CopyOutBytes(addr usermem.Addr, src []byte) (int, error) { + return t.MemoryManager().CopyOut(t, addr, src, t.opts) +} diff --git a/pkg/sentry/kernel/thread_group.go b/pkg/sentry/kernel/thread_group.go index 0b34c0099..a183b28c1 100644 --- a/pkg/sentry/kernel/thread_group.go +++ b/pkg/sentry/kernel/thread_group.go @@ -18,6 +18,7 @@ import ( "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" @@ -307,8 +308,8 @@ func (tg *ThreadGroup) Limits() *limits.LimitSet { return tg.limits } -// release releases the thread group's resources. -func (tg *ThreadGroup) release(t *Task) { +// Release releases the thread group's resources. +func (tg *ThreadGroup) Release(ctx context.Context) { // Timers must be destroyed without holding the TaskSet or signal mutexes // since timers send signals with Timer.mu locked. tg.itimerRealTimer.Destroy() @@ -325,7 +326,7 @@ func (tg *ThreadGroup) release(t *Task) { it.DestroyTimer() } if tg.mounts != nil { - tg.mounts.DecRef(t) + tg.mounts.DecRef(ctx) } } diff --git a/pkg/sentry/kernel/threads.go b/pkg/sentry/kernel/threads.go index 872e1a82d..fdadb52c0 100644 --- a/pkg/sentry/kernel/threads.go +++ b/pkg/sentry/kernel/threads.go @@ -36,6 +36,8 @@ import ( const TasksLimit = (1 << 16) // ThreadID is a generic thread identifier. +// +// +marshal type ThreadID int32 // String returns a decimal representation of the ThreadID. @@ -263,6 +265,13 @@ func (ns *PIDNamespace) Tasks() []*Task { return tasks } +// NumTasks returns the number of tasks in ns. +func (ns *PIDNamespace) NumTasks() int { + ns.owner.mu.RLock() + defer ns.owner.mu.RUnlock() + return len(ns.tids) +} + // ThreadGroups returns a snapshot of the thread groups in ns. func (ns *PIDNamespace) ThreadGroups() []*ThreadGroup { return ns.ThreadGroupsAppend(nil) diff --git a/pkg/sentry/kernel/time/time.go b/pkg/sentry/kernel/time/time.go index e959700f2..f61a8e164 100644 --- a/pkg/sentry/kernel/time/time.go +++ b/pkg/sentry/kernel/time/time.go @@ -616,8 +616,10 @@ func (t *Timer) Swap(s Setting) (Time, Setting) { // 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. +// 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() diff --git a/pkg/sentry/kernel/vdso.go b/pkg/sentry/kernel/vdso.go index 290c32466..9bc452e67 100644 --- a/pkg/sentry/kernel/vdso.go +++ b/pkg/sentry/kernel/vdso.go @@ -17,7 +17,6 @@ package kernel import ( "fmt" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" @@ -28,6 +27,8 @@ import ( // // They are exposed to the VDSO via a parameter page managed by VDSOParamPage, // which also includes a sequence counter. +// +// +marshal type vdsoParams struct { monotonicReady uint64 monotonicBaseCycles int64 @@ -68,21 +69,29 @@ type VDSOParamPage struct { // checked in state_test_util tests, causing this field to change across // save / restore. seq uint64 + + // copyScratchBuffer is a temporary buffer used to marshal the params before + // copying it to the real parameter page. The parameter page is typically + // updated at a moderate frequency of ~O(seconds) throughout the lifetime of + // the sentry, so reusing this buffer is a good tradeoff between memory + // usage and the cost of allocation. + copyScratchBuffer []byte } // NewVDSOParamPage returns a VDSOParamPage. // // Preconditions: -// // * fr is a single page allocated from mfp.MemoryFile(). 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. -// // * mfp.MemoryFile().MapInternal(fr) must return a single safemem.Block. func NewVDSOParamPage(mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *VDSOParamPage { - return &VDSOParamPage{mfp: mfp, fr: fr} + return &VDSOParamPage{ + mfp: mfp, + fr: fr, + copyScratchBuffer: make([]byte, (*vdsoParams)(nil).SizeBytes()), + } } // access returns a mapping of the param page. @@ -136,7 +145,8 @@ func (v *VDSOParamPage) Write(f func() vdsoParams) error { // Get the new params. p := f() - buf := binary.Marshal(nil, usermem.ByteOrder, p) + buf := v.copyScratchBuffer[:p.SizeBytes()] + p.MarshalUnsafe(buf) // Skip the sequence counter. if _, err := safemem.Copy(paramPage.DropFirst(8), safemem.BlockFromSafeSlice(buf)); err != nil { diff --git a/pkg/sentry/limits/context.go b/pkg/sentry/limits/context.go index 77e1fe217..0bade6e57 100644 --- a/pkg/sentry/limits/context.go +++ b/pkg/sentry/limits/context.go @@ -33,3 +33,12 @@ func FromContext(ctx context.Context) *LimitSet { } return nil } + +// FromContextOrDie returns FromContext(ctx) if the latter is not nil. +// Otherwise, panic is triggered. +func FromContextOrDie(ctx context.Context) *LimitSet { + if v := ctx.Value(CtxLimits); v != nil { + return v.(*LimitSet) + } + panic("failed to create limit set from context") +} diff --git a/pkg/sentry/loader/elf.go b/pkg/sentry/loader/elf.go index 20dd1cc21..98af2cc38 100644 --- a/pkg/sentry/loader/elf.go +++ b/pkg/sentry/loader/elf.go @@ -194,6 +194,10 @@ func parseHeader(ctx context.Context, f fullReader) (elfInfo, error) { log.Infof("Too many phdrs (%d): total size %d > %d", hdr.Phnum, totalPhdrSize, maxTotalPhdrSize) return elfInfo{}, syserror.ENOEXEC } + if int64(hdr.Phoff) < 0 || int64(hdr.Phoff+uint64(totalPhdrSize)) < 0 { + ctx.Infof("Unsupported phdr offset %d", hdr.Phoff) + return elfInfo{}, syserror.ENOEXEC + } phdrBuf := make([]byte, totalPhdrSize) _, err = f.ReadFull(ctx, usermem.BytesIOSequence(phdrBuf), int64(hdr.Phoff)) @@ -402,8 +406,7 @@ type loadedELF struct { // // It does not load the ELF interpreter, or return any auxv entries. // -// Preconditions: -// * f is an ELF file +// Preconditions: f is an ELF file. func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, info elfInfo, sharedLoadOffset usermem.Addr) (loadedELF, error) { first := true var start, end usermem.Addr @@ -438,6 +441,10 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, in ctx.Infof("PT_INTERP path too big: %v", phdr.Filesz) return loadedELF{}, syserror.ENOEXEC } + if int64(phdr.Off) < 0 || int64(phdr.Off+phdr.Filesz) < 0 { + ctx.Infof("Unsupported PT_INTERP offset %d", phdr.Off) + return loadedELF{}, syserror.ENOEXEC + } path := make([]byte, phdr.Filesz) _, err := f.ReadFull(ctx, usermem.BytesIOSequence(path), int64(phdr.Off)) @@ -571,8 +578,8 @@ func loadParsedELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, in // 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 +// * 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 fsbridge.File) (loadedELF, arch.Context, error) { info, err := parseHeader(ctx, f) if err != nil { @@ -609,8 +616,7 @@ func loadInitialELF(ctx context.Context, m *mm.MemoryManager, fs *cpuid.FeatureS // // It does not return any auxv entries. // -// Preconditions: -// * f is an ELF file +// Preconditions: f is an ELF file. func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.File, initial loadedELF) (loadedELF, error) { info, err := parseHeader(ctx, f) if err != nil { @@ -640,8 +646,7 @@ func loadInterpreterELF(ctx context.Context, m *mm.MemoryManager, f fsbridge.Fil // If loadELF returns ErrSwitchFile it should be called again with the returned // path and argv. // -// Preconditions: -// * args.File is an ELF file +// Preconditions: args.File is an ELF file. func loadELF(ctx context.Context, args LoadArgs) (loadedELF, arch.Context, error) { bin, ac, err := loadInitialELF(ctx, args.MemoryManager, args.Features, args.File) if err != nil { diff --git a/pkg/sentry/loader/loader.go b/pkg/sentry/loader/loader.go index 8d6802ea3..c69b62db9 100644 --- a/pkg/sentry/loader/loader.go +++ b/pkg/sentry/loader/loader.go @@ -122,7 +122,7 @@ func allocStack(ctx context.Context, m *mm.MemoryManager, a arch.Context) (*arch if err != nil { return nil, err } - return &arch.Stack{a, m, ar.End}, nil + return &arch.Stack{Arch: a, IO: m, Bottom: ar.End}, nil } const ( @@ -215,8 +215,8 @@ func loadExecutable(ctx context.Context, args LoadArgs) (loadedELF, arch.Context // path and argv. // // Preconditions: -// * The Task MemoryManager is empty. -// * Load is called on the Task goroutine. +// * The Task MemoryManager is empty. +// * Load is called on the Task goroutine. func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *VDSO) (abi.OS, arch.Context, string, *syserr.Error) { // Load the executable itself. loaded, ac, file, newArgv, err := loadExecutable(ctx, args) @@ -247,20 +247,20 @@ func Load(ctx context.Context, args LoadArgs, extraAuxv []arch.AuxEntry, vdso *V } // Push the original filename to the stack, for AT_EXECFN. - execfn, err := stack.Push(args.Filename) - if err != nil { + if _, err := stack.PushNullTerminatedByteSlice([]byte(args.Filename)); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to push exec filename: %v", err), syserr.FromError(err).ToLinux()) } + execfn := stack.Bottom // Push 16 random bytes on the stack which AT_RANDOM will point to. var b [16]byte if _, err := rand.Read(b[:]); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to read random bytes: %v", err), syserr.FromError(err).ToLinux()) } - random, err := stack.Push(b) - if err != nil { + if _, err = stack.PushNullTerminatedByteSlice(b[:]); err != nil { return 0, nil, "", syserr.NewDynamic(fmt.Sprintf("Failed to push random bytes: %v", err), syserr.FromError(err).ToLinux()) } + random := stack.Bottom c := auth.CredentialsFromContext(ctx) diff --git a/pkg/sentry/loader/vdso.go b/pkg/sentry/loader/vdso.go index 05a294fe6..241d87835 100644 --- a/pkg/sentry/loader/vdso.go +++ b/pkg/sentry/loader/vdso.go @@ -380,3 +380,9 @@ func loadVDSO(ctx context.Context, m *mm.MemoryManager, v *VDSO, bin loadedELF) return vdsoAddr, nil } + +// Release drops references on mappings held by v. +func (v *VDSO) Release(ctx context.Context) { + v.ParamPage.DecRef(ctx) + v.vdso.DecRef(ctx) +} diff --git a/pkg/sentry/memmap/mapping_set.go b/pkg/sentry/memmap/mapping_set.go index d609c1ae0..457ed87f8 100644 --- a/pkg/sentry/memmap/mapping_set.go +++ b/pkg/sentry/memmap/mapping_set.go @@ -177,7 +177,7 @@ func subsetMapping(wholeRange, subsetRange MappableRange, ms MappingSpace, addr // AddMapping adds the given mapping and returns the set of MappableRanges that // previously had no mappings. // -// Preconditions: As for Mappable.AddMapping. +// Preconditions: Same as Mappable.AddMapping. func (s *MappingSet) AddMapping(ms MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) []MappableRange { mr := MappableRange{offset, offset + uint64(ar.Length())} var mapped []MappableRange @@ -204,7 +204,7 @@ func (s *MappingSet) AddMapping(ms MappingSpace, ar usermem.AddrRange, offset ui // RemoveMapping removes the given mapping and returns the set of // MappableRanges that now have no mappings. // -// Preconditions: As for Mappable.RemoveMapping. +// Preconditions: Same as Mappable.RemoveMapping. func (s *MappingSet) RemoveMapping(ms MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) []MappableRange { mr := MappableRange{offset, offset + uint64(ar.Length())} var unmapped []MappableRange diff --git a/pkg/sentry/memmap/memmap.go b/pkg/sentry/memmap/memmap.go index 65d83096f..7fd77925f 100644 --- a/pkg/sentry/memmap/memmap.go +++ b/pkg/sentry/memmap/memmap.go @@ -28,9 +28,9 @@ import ( // // 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. +// All Mappable methods have the following preconditions: +// * usermem.AddrRanges and MappableRanges must be non-empty (Length() != 0). +// * 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. @@ -48,8 +48,10 @@ type Mappable interface { // 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. writable must match the corresponding call to AddMapping. + // Preconditions: + // * offset+ar.Length() does not overflow. + // * The removed mapping must exist. writable must match the + // corresponding call to AddMapping. RemoveMapping(ctx context.Context, ms MappingSpace, ar usermem.AddrRange, offset uint64, writable bool) // CopyMapping notifies the Mappable of an attempt to copy a mapping in ms @@ -60,9 +62,10 @@ type Mappable interface { // 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+srcAR.Length() and offset+dstAR.Length() do not - // overflow. The mapping at srcAR must exist. writable must match the - // corresponding call to AddMapping. + // Preconditions: + // * offset+srcAR.Length() and offset+dstAR.Length() do not overflow. + // * The mapping at srcAR must exist. writable must match the + // corresponding call to AddMapping. CopyMapping(ctx context.Context, ms MappingSpace, srcAR, dstAR usermem.AddrRange, offset uint64, writable bool) error // Translate returns the Mappable's current mappings for at least the range @@ -77,11 +80,14 @@ type Mappable interface { // reference is held on all pages in a File that may be the result // of a valid Translation. // - // 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. + // 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) @@ -118,10 +124,10 @@ func (t Translation) FileRange() FileRange { // CheckTranslateResult returns an error if (ts, terr) does not satisfy all // postconditions for Mappable.Translate(required, optional, at). // -// Preconditions: As for Mappable.Translate. +// Preconditions: Same as Mappable.Translate. func CheckTranslateResult(required, optional MappableRange, at usermem.AccessType, ts []Translation, terr error) error { // Verify that the inputs to Mappable.Translate were valid. - if !required.WellFormed() || required.Length() <= 0 { + 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() { @@ -139,7 +145,7 @@ func CheckTranslateResult(required, optional MappableRange, at usermem.AccessTyp 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 { + 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() { @@ -214,7 +220,9 @@ type MappingSpace interface { // Invalidate must not take any locks preceding mm.MemoryManager.activeMu // in the lock order. // - // Preconditions: ar.Length() != 0. ar must be page-aligned. + // Preconditions: + // * ar.Length() != 0. + // * ar must be page-aligned. Invalidate(ar usermem.AddrRange, opts InvalidateOpts) } @@ -375,16 +383,20 @@ type File interface { // IncRef increments the reference count on all pages in fr. // - // Preconditions: fr.Start and fr.End must be page-aligned. fr.Length() > - // 0. At least one reference must be held on all pages in fr. (The File - // interface does not provide a way to acquire an initial reference; - // implementors may define mechanisms for doing so.) + // Preconditions: + // * fr.Start and fr.End must be page-aligned. + // * fr.Length() > 0. + // * At least one reference must be held on all pages in fr. (The File + // interface does not provide a way to acquire an initial reference; + // implementors may define mechanisms for doing so.) IncRef(fr FileRange) // DecRef decrements the reference count on all pages in fr. // - // Preconditions: fr.Start and fr.End must be page-aligned. fr.Length() > - // 0. At least one reference must be held on all pages in fr. + // Preconditions: + // * fr.Start and fr.End must be page-aligned. + // * fr.Length() > 0. + // * At least one reference must be held on all pages in fr. DecRef(fr FileRange) // MapInternal returns a mapping of the given file offsets in the invoking @@ -392,8 +404,9 @@ type File interface { // // Note that fr.Start and fr.End need not be page-aligned. // - // Preconditions: fr.Length() > 0. At least one reference must be held on - // all pages in fr. + // Preconditions: + // * fr.Length() > 0. + // * At least one reference must be held on all pages in fr. // // Postconditions: The returned mapping is valid as long as at least one // reference is held on the mapped pages. diff --git a/pkg/sentry/mm/BUILD b/pkg/sentry/mm/BUILD index f9d0837a1..6dbeccfe2 100644 --- a/pkg/sentry/mm/BUILD +++ b/pkg/sentry/mm/BUILD @@ -73,12 +73,35 @@ go_template_instance( }, ) +go_template_instance( + name = "aio_mappable_refs", + out = "aio_mappable_refs.go", + package = "mm", + prefix = "aioMappable", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "aioMappable", + }, +) + +go_template_instance( + name = "special_mappable_refs", + out = "special_mappable_refs.go", + package = "mm", + prefix = "SpecialMappable", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "SpecialMappable", + }, +) + go_library( name = "mm", srcs = [ "address_space.go", "aio_context.go", "aio_context_state.go", + "aio_mappable_refs.go", "debug.go", "file_refcount_set.go", "io.go", @@ -92,6 +115,7 @@ go_library( "save_restore.go", "shm.go", "special_mappable.go", + "special_mappable_refs.go", "syscalls.go", "vma.go", "vma_set.go", @@ -103,6 +127,7 @@ go_library( "//pkg/context", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safecopy", "//pkg/safemem", "//pkg/sentry/arch", diff --git a/pkg/sentry/mm/address_space.go b/pkg/sentry/mm/address_space.go index 5c667117c..a93e76c75 100644 --- a/pkg/sentry/mm/address_space.go +++ b/pkg/sentry/mm/address_space.go @@ -166,8 +166,12 @@ func (mm *MemoryManager) Deactivate() { // 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 == mm.pmas.LowerBoundSegment(ar.Start). +// Preconditions: +// * mm.activeMu must be locked. +// * mm.as != nil. +// * ar.Length() != 0. +// * ar must be page-aligned. +// * pseg == mm.pmas.LowerBoundSegment(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. diff --git a/pkg/sentry/mm/aio_context.go b/pkg/sentry/mm/aio_context.go index 16fea53c4..7bf48cb2c 100644 --- a/pkg/sentry/mm/aio_context.go +++ b/pkg/sentry/mm/aio_context.go @@ -17,7 +17,6 @@ package mm import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/usage" @@ -239,7 +238,7 @@ func (ctx *AIOContext) Drain() { // // +stateify savable type aioMappable struct { - refs.AtomicRefCount + aioMappableRefs mfp pgalloc.MemoryFileProvider fr memmap.FileRange @@ -253,13 +252,13 @@ func newAIOMappable(mfp pgalloc.MemoryFileProvider) (*aioMappable, error) { return nil, err } m := aioMappable{mfp: mfp, fr: fr} - m.EnableLeakCheck("mm.aioMappable") + m.EnableLeakCheck() return &m, nil } // DecRef implements refs.RefCounter.DecRef. func (m *aioMappable) DecRef(ctx context.Context) { - m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { + m.aioMappableRefs.DecRef(func() { m.mfp.MemoryFile().DecRef(m.fr) }) } diff --git a/pkg/sentry/mm/io.go b/pkg/sentry/mm/io.go index fa776f9c6..a8ac48080 100644 --- a/pkg/sentry/mm/io.go +++ b/pkg/sentry/mm/io.go @@ -441,7 +441,10 @@ func (mm *MemoryManager) LoadUint32(ctx context.Context, addr usermem.Addr, opts // handleASIOFault handles a page fault at address addr for an AddressSpaceIO // operation spanning ioar. // -// Preconditions: mm.as != nil. ioar.Length() != 0. ioar.Contains(addr). +// 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. @@ -629,7 +632,9 @@ func (mm *MemoryManager) withVecInternalMappings(ctx context.Context, ars userme // 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. +// 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 { diff --git a/pkg/sentry/mm/mm.go b/pkg/sentry/mm/mm.go index 3e85964e4..92cc87d84 100644 --- a/pkg/sentry/mm/mm.go +++ b/pkg/sentry/mm/mm.go @@ -235,6 +235,20 @@ type MemoryManager struct { // vdsoSigReturnAddr is the address of 'vdso_sigreturn'. vdsoSigReturnAddr uint64 + + // membarrierPrivateEnabled is non-zero if EnableMembarrierPrivate has + // previously been called. Since, as of this writing, + // MEMBARRIER_CMD_PRIVATE_EXPEDITED is implemented as a global memory + // barrier, membarrierPrivateEnabled has no other effect. + // + // membarrierPrivateEnabled is accessed using atomic memory operations. + membarrierPrivateEnabled uint32 + + // membarrierRSeqEnabled is non-zero if EnableMembarrierRSeq has previously + // been called. + // + // membarrierRSeqEnabled is accessed using atomic memory operations. + membarrierRSeqEnabled uint32 } // vma represents a virtual memory area. @@ -242,7 +256,7 @@ type MemoryManager struct { // +stateify savable type vma struct { // mappable is the virtual memory object mapped by this vma. If mappable is - // nil, the vma represents a private anonymous mapping. + // nil, the vma represents an anonymous mapping. mappable memmap.Mappable // off is the offset into mappable at which this vma begins. If mappable is diff --git a/pkg/sentry/mm/mm_test.go b/pkg/sentry/mm/mm_test.go index fdc308542..acac3d357 100644 --- a/pkg/sentry/mm/mm_test.go +++ b/pkg/sentry/mm/mm_test.go @@ -51,7 +51,8 @@ func TestUsageASUpdates(t *testing.T) { defer mm.DecUsers(ctx) addr, err := mm.MMap(ctx, memmap.MMapOpts{ - Length: 2 * usermem.PageSize, + Length: 2 * usermem.PageSize, + Private: true, }) if err != nil { t.Fatalf("MMap got err %v want nil", err) diff --git a/pkg/sentry/mm/pma.go b/pkg/sentry/mm/pma.go index 930ec895f..7e5f7de64 100644 --- a/pkg/sentry/mm/pma.go +++ b/pkg/sentry/mm/pma.go @@ -31,10 +31,12 @@ import ( // iterator to the pma containing ar.Start. Otherwise it returns a terminal // iterator. // -// Preconditions: mm.activeMu must be locked. ar.Length() != 0. +// Preconditions: +// * mm.activeMu must be locked. +// * ar.Length() != 0. func (mm *MemoryManager) existingPMAsLocked(ar usermem.AddrRange, at usermem.AccessType, ignorePermissions bool, needInternalMappings bool) pmaIterator { if checkInvariants { - if !ar.WellFormed() || ar.Length() <= 0 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -89,13 +91,16 @@ func (mm *MemoryManager) existingVecPMAsLocked(ars usermem.AddrRangeSeq, at user // // - 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, and support accesses of type at (i.e. permission -// checks must have been performed against vmas). +// 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, and support accesses of type at +// (i.e. permission checks must have been performed against vmas). func (mm *MemoryManager) getPMAsLocked(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, at usermem.AccessType) (pmaIterator, pmaGapIterator, error) { if checkInvariants { - if !ar.WellFormed() || ar.Length() <= 0 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !vseg.Ok() { @@ -135,9 +140,11 @@ func (mm *MemoryManager) getPMAsLocked(ctx context.Context, vseg vmaIterator, ar // 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, and support accesses of -// type at (i.e. permission checks must have been performed against vmas). +// Preconditions: +// * mm.mappingMu must be locked. +// * mm.activeMu must be locked for writing. +// * vmas must exist for all addresses in ars, and support accesses of type at +// (i.e. permission checks must have been performed against vmas). func (mm *MemoryManager) getVecPMAsLocked(ctx context.Context, ars usermem.AddrRangeSeq, at usermem.AccessType) (usermem.AddrRangeSeq, error) { for arsit := ars; !arsit.IsEmpty(); arsit = arsit.Tail() { ar := arsit.Head() @@ -186,7 +193,7 @@ func (mm *MemoryManager) getVecPMAsLocked(ctx context.Context, ars usermem.AddrR // getVecPMAsLocked; other clients should call one of those instead. func (mm *MemoryManager) getPMAsInternalLocked(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, at usermem.AccessType) (pmaIterator, pmaGapIterator, error) { if checkInvariants { - if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !vseg.Ok() { @@ -216,7 +223,7 @@ func (mm *MemoryManager) getPMAsInternalLocked(ctx context.Context, vseg vmaIter // Need a pma here. optAR := vseg.Range().Intersect(pgap.Range()) if checkInvariants { - if optAR.Length() <= 0 { + if optAR.Length() == 0 { panic(fmt.Sprintf("vseg %v and pgap %v do not overlap", vseg, pgap)) } } @@ -518,8 +525,10 @@ func privateAligned(ar usermem.AddrRange) usermem.AddrRange { // the memory it maps, isPMACopyOnWriteLocked will take ownership of the memory // and update the pma to indicate that it does not require copy-on-write. // -// Preconditions: vseg.Range().IsSupersetOf(pseg.Range()). mm.mappingMu must be -// locked. mm.activeMu must be locked for writing. +// Preconditions: +// * vseg.Range().IsSupersetOf(pseg.Range()). +// * mm.mappingMu must be locked. +// * mm.activeMu must be locked for writing. func (mm *MemoryManager) isPMACopyOnWriteLocked(vseg vmaIterator, pseg pmaIterator) bool { pma := pseg.ValuePtr() if !pma.needCOW { @@ -551,7 +560,7 @@ func (mm *MemoryManager) isPMACopyOnWriteLocked(vseg vmaIterator, pseg pmaIterat // 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() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -568,11 +577,13 @@ func (mm *MemoryManager) Invalidate(ar usermem.AddrRange, opts memmap.Invalidate // 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. +// 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() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -613,10 +624,12 @@ func (mm *MemoryManager) invalidateLocked(ar usermem.AddrRange, invalidatePrivat // most I/O. It should only be used in contexts that would use get_user_pages() // in the Linux kernel. // -// Preconditions: ar.Length() != 0. ar must be page-aligned. +// Preconditions: +// * ar.Length() != 0. +// * ar must be page-aligned. func (mm *MemoryManager) Pin(ctx context.Context, ar usermem.AddrRange, at usermem.AccessType, ignorePermissions bool) ([]PinnedRange, error) { if checkInvariants { - if !ar.WellFormed() || ar.Length() <= 0 || !ar.IsPageAligned() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -693,15 +706,19 @@ func Unpin(prs []PinnedRange) { // 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. +// 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() { + if !oldAR.WellFormed() || oldAR.Length() == 0 || !oldAR.IsPageAligned() { panic(fmt.Sprintf("invalid oldAR: %v", oldAR)) } - if !newAR.WellFormed() || newAR.Length() <= 0 || !newAR.IsPageAligned() { + if !newAR.WellFormed() || newAR.Length() == 0 || !newAR.IsPageAligned() { panic(fmt.Sprintf("invalid newAR: %v", newAR)) } if oldAR.Length() > newAR.Length() { @@ -751,15 +768,17 @@ func (mm *MemoryManager) movePMAsLocked(oldAR, newAR usermem.AddrRange) { // - 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. +// 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 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !pseg.Range().Contains(ar.Start) { @@ -783,8 +802,9 @@ func (mm *MemoryManager) getPMAInternalMappingsLocked(pseg pmaIterator, ar userm // 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. +// 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. @@ -803,12 +823,15 @@ func (mm *MemoryManager) getVecPMAInternalMappingsLocked(ars usermem.AddrRangeSe // 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). +// 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 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !pseg.Range().Contains(ar.Start) { @@ -839,8 +862,10 @@ func (mm *MemoryManager) internalMappingsLocked(pseg pmaIterator, ar usermem.Add // 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. +// 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() { @@ -969,7 +994,9 @@ func (pmaSetFunctions) Split(ar usermem.AddrRange, p pma, split usermem.Addr) (p // 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(). +// Preconditions: +// * mm.activeMu must be locked. +// * addr <= pgap.Start(). func (mm *MemoryManager) findOrSeekPrevUpperBoundPMA(addr usermem.Addr, pgap pmaGapIterator) pmaIterator { if checkInvariants { if !pgap.Ok() { @@ -1015,13 +1042,15 @@ func (pseg pmaIterator) fileRange() memmap.FileRange { return pseg.fileRangeOf(pseg.Range()) } -// Preconditions: pseg.Range().IsSupersetOf(ar). ar.Length != 0. +// Preconditions: +// * pseg.Range().IsSupersetOf(ar). +// * ar.Length != 0. func (pseg pmaIterator) fileRangeOf(ar usermem.AddrRange) memmap.FileRange { if checkInvariants { if !pseg.Ok() { panic("terminal pma iterator") } - if !ar.WellFormed() || ar.Length() <= 0 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !pseg.Range().IsSupersetOf(ar) { diff --git a/pkg/sentry/mm/special_mappable.go b/pkg/sentry/mm/special_mappable.go index 4cdb52eb6..2dbe5b751 100644 --- a/pkg/sentry/mm/special_mappable.go +++ b/pkg/sentry/mm/special_mappable.go @@ -16,7 +16,6 @@ package mm import ( "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/sentry/usage" @@ -31,7 +30,7 @@ import ( // // +stateify savable type SpecialMappable struct { - refs.AtomicRefCount + SpecialMappableRefs mfp pgalloc.MemoryFileProvider fr memmap.FileRange @@ -45,13 +44,13 @@ type SpecialMappable struct { // Preconditions: fr.Length() != 0. func NewSpecialMappable(name string, mfp pgalloc.MemoryFileProvider, fr memmap.FileRange) *SpecialMappable { m := SpecialMappable{mfp: mfp, fr: fr, name: name} - m.EnableLeakCheck("mm.SpecialMappable") + m.EnableLeakCheck() return &m } // DecRef implements refs.RefCounter.DecRef. func (m *SpecialMappable) DecRef(ctx context.Context) { - m.AtomicRefCount.DecRefWithDestructor(ctx, func(context.Context) { + m.SpecialMappableRefs.DecRef(func() { m.mfp.MemoryFile().DecRef(m.fr) }) } @@ -137,9 +136,12 @@ func (m *SpecialMappable) Length() uint64 { // NewSharedAnonMappable returns a SpecialMappable that implements the // semantics of mmap(MAP_SHARED|MAP_ANONYMOUS) and mappings of /dev/zero. // -// TODO(jamieliu): 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. +// TODO(gvisor.dev/issue/1624): Linux uses an ephemeral file created by +// mm/shmem.c:shmem_zero_setup(), and VFS2 does something analogous. VFS1 uses +// a SpecialMappable instead, incorrectly getting device and inode IDs of zero +// and causing memory for shared anonymous mappings to be allocated up-front +// instead of on first touch; this is to avoid exacerbating the fs.MountSource +// leak (b/143656263). Delete this function along with VFS1. func NewSharedAnonMappable(length uint64, mfp pgalloc.MemoryFileProvider) (*SpecialMappable, error) { if length == 0 { return nil, syserror.EINVAL diff --git a/pkg/sentry/mm/syscalls.go b/pkg/sentry/mm/syscalls.go index e74d4e1c1..675efdc7c 100644 --- a/pkg/sentry/mm/syscalls.go +++ b/pkg/sentry/mm/syscalls.go @@ -17,6 +17,7 @@ package mm import ( "fmt" mrand "math/rand" + "sync/atomic" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -24,7 +25,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel/futex" "gvisor.dev/gvisor/pkg/sentry/limits" "gvisor.dev/gvisor/pkg/sentry/memmap" - "gvisor.dev/gvisor/pkg/sentry/pgalloc" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" ) @@ -93,18 +93,6 @@ func (mm *MemoryManager) MMap(ctx context.Context, opts memmap.MMapOpts) (userme } } else { opts.Offset = 0 - if !opts.Private { - if opts.MappingIdentity != nil { - return 0, syserror.EINVAL - } - m, err := NewSharedAnonMappable(opts.Length, pgalloc.MemoryFileProviderFromContext(ctx)) - if err != nil { - return 0, err - } - defer m.DecRef(ctx) - opts.MappingIdentity = m - opts.Mappable = m - } } if opts.Addr.RoundDown() != opts.Addr { @@ -166,7 +154,9 @@ func (mm *MemoryManager) MMap(ctx context.Context, opts memmap.MMapOpts) (userme // populateVMA obtains pmas for addresses in ar in the given vma, and maps them // into mm.as if it is active. // -// Preconditions: mm.mappingMu must be locked. vseg.Range().IsSupersetOf(ar). +// Preconditions: +// * mm.mappingMu must be locked. +// * vseg.Range().IsSupersetOf(ar). func (mm *MemoryManager) populateVMA(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, precommit bool) { if !vseg.ValuePtr().effectivePerms.Any() { // Linux doesn't populate inaccessible pages. See @@ -208,8 +198,9 @@ func (mm *MemoryManager) populateVMA(ctx context.Context, vseg vmaIterator, ar u // preferable to populateVMA since it unlocks mm.mappingMu before performing // expensive operations that don't require it to be locked. // -// Preconditions: mm.mappingMu must be locked for writing. -// vseg.Range().IsSupersetOf(ar). +// Preconditions: +// * mm.mappingMu must be locked for writing. +// * vseg.Range().IsSupersetOf(ar). // // Postconditions: mm.mappingMu will be unlocked. func (mm *MemoryManager) populateVMAAndUnlock(ctx context.Context, vseg vmaIterator, ar usermem.AddrRange, precommit bool) { @@ -1284,3 +1275,27 @@ func (mm *MemoryManager) VirtualDataSize() uint64 { defer mm.mappingMu.RUnlock() return mm.dataAS } + +// EnableMembarrierPrivate causes future calls to IsMembarrierPrivateEnabled to +// return true. +func (mm *MemoryManager) EnableMembarrierPrivate() { + atomic.StoreUint32(&mm.membarrierPrivateEnabled, 1) +} + +// IsMembarrierPrivateEnabled returns true if mm.EnableMembarrierPrivate() has +// previously been called. +func (mm *MemoryManager) IsMembarrierPrivateEnabled() bool { + return atomic.LoadUint32(&mm.membarrierPrivateEnabled) != 0 +} + +// EnableMembarrierRSeq causes future calls to IsMembarrierRSeqEnabled to +// return true. +func (mm *MemoryManager) EnableMembarrierRSeq() { + atomic.StoreUint32(&mm.membarrierRSeqEnabled, 1) +} + +// IsMembarrierRSeqEnabled returns true if mm.EnableMembarrierRSeq() has +// previously been called. +func (mm *MemoryManager) IsMembarrierRSeqEnabled() bool { + return atomic.LoadUint32(&mm.membarrierRSeqEnabled) != 0 +} diff --git a/pkg/sentry/mm/vma.go b/pkg/sentry/mm/vma.go index c4e1989ed..b8df72813 100644 --- a/pkg/sentry/mm/vma.go +++ b/pkg/sentry/mm/vma.go @@ -27,8 +27,9 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// Preconditions: mm.mappingMu must be locked for writing. opts must be valid -// as defined by the checks in MMap. +// 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)) @@ -260,11 +261,12 @@ func (mm *MemoryManager) mlockedBytesRangeLocked(ar usermem.AddrRange) uint64 { // // - 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. +// 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 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -342,11 +344,13 @@ 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. +// 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() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -361,11 +365,13 @@ func (mm *MemoryManager) unmapLocked(ctx context.Context, ar usermem.AddrRange) // 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. +// 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() { + if !ar.WellFormed() || ar.Length() == 0 || !ar.IsPageAligned() { panic(fmt.Sprintf("invalid ar: %v", ar)) } } @@ -467,7 +473,9 @@ func (vmaSetFunctions) Split(ar usermem.AddrRange, v vma, split usermem.Addr) (v return v, v2 } -// Preconditions: vseg.ValuePtr().mappable != nil. vseg.Range().Contains(addr). +// Preconditions: +// * vseg.ValuePtr().mappable != nil. +// * vseg.Range().Contains(addr). func (vseg vmaIterator) mappableOffsetAt(addr usermem.Addr) uint64 { if checkInvariants { if !vseg.Ok() { @@ -491,8 +499,10 @@ func (vseg vmaIterator) mappableRange() memmap.MappableRange { return vseg.mappableRangeOf(vseg.Range()) } -// Preconditions: vseg.ValuePtr().mappable != nil. -// vseg.Range().IsSupersetOf(ar). ar.Length() != 0. +// 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() { @@ -501,7 +511,7 @@ func (vseg vmaIterator) mappableRangeOf(ar usermem.AddrRange) memmap.MappableRan if vseg.ValuePtr().mappable == nil { panic("MappableRange is meaningless for anonymous vma") } - if !ar.WellFormed() || ar.Length() <= 0 { + if !ar.WellFormed() || ar.Length() == 0 { panic(fmt.Sprintf("invalid ar: %v", ar)) } if !vseg.Range().IsSupersetOf(ar) { @@ -514,8 +524,10 @@ func (vseg vmaIterator) mappableRangeOf(ar usermem.AddrRange) memmap.MappableRan 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. +// 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() { @@ -524,7 +536,7 @@ func (vseg vmaIterator) addrRangeOf(mr memmap.MappableRange) usermem.AddrRange { if vseg.ValuePtr().mappable == nil { panic("MappableRange is meaningless for anonymous vma") } - if !mr.WellFormed() || mr.Length() <= 0 { + if !mr.WellFormed() || mr.Length() == 0 { panic(fmt.Sprintf("invalid mr: %v", mr)) } if !vseg.mappableRange().IsSupersetOf(mr) { @@ -540,7 +552,9 @@ func (vseg vmaIterator) addrRangeOf(mr memmap.MappableRange) usermem.AddrRange { // seekNextLowerBound returns mm.vmas.LowerBoundSegment(addr), but does so by // scanning linearly forward from vseg. // -// Preconditions: mm.mappingMu must be locked. addr >= vseg.Start(). +// Preconditions: +// * mm.mappingMu must be locked. +// * addr >= vseg.Start(). func (vseg vmaIterator) seekNextLowerBound(addr usermem.Addr) vmaIterator { if checkInvariants { if !vseg.Ok() { diff --git a/pkg/sentry/pgalloc/BUILD b/pkg/sentry/pgalloc/BUILD index 7a3311a70..5b09b9feb 100644 --- a/pkg/sentry/pgalloc/BUILD +++ b/pkg/sentry/pgalloc/BUILD @@ -83,6 +83,7 @@ go_library( ], visibility = ["//pkg/sentry:internal"], deps = [ + "//pkg/abi/linux", "//pkg/context", "//pkg/log", "//pkg/memutil", diff --git a/pkg/sentry/pgalloc/pgalloc.go b/pkg/sentry/pgalloc/pgalloc.go index 46d3be58c..7c297fb9e 100644 --- a/pkg/sentry/pgalloc/pgalloc.go +++ b/pkg/sentry/pgalloc/pgalloc.go @@ -29,6 +29,7 @@ import ( "syscall" "time" + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/safemem" @@ -224,6 +225,18 @@ type usageInfo struct { refs uint64 } +// canCommit returns true if the tracked region can be committed. +func (u *usageInfo) canCommit() bool { + // refs must be greater than 0 because we 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. + return !u.knownCommitted && u.refs != 0 +} + // An EvictableMemoryUser represents a user of MemoryFile-allocated memory that // may be asked to deallocate that memory in the presence of memory pressure. type EvictableMemoryUser interface { @@ -507,7 +520,9 @@ func findAvailableRange(usage *usageSet, fileSize int64, length, alignment uint6 // 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. +// Preconditions: +// * length > 0. +// * length must be page-aligned. func (f *MemoryFile) AllocateAndFill(length uint64, kind usage.MemoryKind, r safemem.Reader) (memmap.FileRange, error) { fr, err := f.Allocate(length, kind) if err != nil { @@ -826,6 +841,11 @@ func (f *MemoryFile) UpdateUsage() error { log.Debugf("UpdateUsage: skipped with usageSwapped!=0.") return nil } + // Linux updates usage values at CONFIG_HZ. + if scanningAfter := time.Now().Sub(f.usageLast).Milliseconds(); scanningAfter < time.Second.Milliseconds()/linux.CLOCKS_PER_SEC { + log.Debugf("UpdateUsage: skipped because previous scan happened %d ms back", scanningAfter) + return nil + } f.usageLast = time.Now() err = f.updateUsageLocked(currentUsage, mincore) @@ -839,7 +859,7 @@ func (f *MemoryFile) UpdateUsage() error { // 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. +// Precondition: f.mu must be held; it may be unlocked and reacquired. func (f *MemoryFile) 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. @@ -866,7 +886,7 @@ func (f *MemoryFile) updateUsageLocked(currentUsage uint64, checkCommitted func( } 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 + // being committed while the below loop was running. Just // report the higher number that we found and ignore swap. usage.MemoryAccounting.Dec(f.usageSwapped, usage.System) f.usageSwapped = 0 @@ -878,21 +898,9 @@ func (f *MemoryFile) updateUsageLocked(currentUsage uint64, checkCommitted func( // 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 { + for seg := f.usage.FirstSegment(); seg.Ok(); { + if !seg.ValuePtr().canCommit() { + seg = seg.NextSegment() continue } @@ -915,56 +923,53 @@ func (f *MemoryFile) updateUsageLocked(currentUsage uint64, checkCommitted func( } // Query for new pages in core. - if err := checkCommitted(s, buf); err != nil { + // NOTE(b/165896008): mincore (which is passed as checkCommitted) + // by f.UpdateUsage() might take a really long time. So unlock f.mu + // while checkCommitted runs. + f.mu.Unlock() + err := checkCommitted(s, buf) + f.mu.Lock() + if 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. + seg := f.usage.LowerBoundSegment(r.Start) + for i := 0; i < bufLen; { + if buf[i]&0x1 == 0 { + i++ continue - case !populated && populatedRun: - // Finish the run by changing this segment. - runRange := memmap.FileRange{ - Start: r.Start + uint64(populatedRunStart*usermem.PageSize), - End: r.Start + uint64(i*usermem.PageSize), + } + // Scan to the end of this committed range. + j := i + 1 + for ; j < bufLen; j++ { + if buf[j]&0x1 == 0 { + break } - 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() + } + committedFR := memmap.FileRange{ + Start: r.Start + uint64(i*usermem.PageSize), + End: r.Start + uint64(j*usermem.PageSize), + } + // Advance seg to committedFR.Start. + for seg.Ok() && seg.End() < committedFR.Start { + seg = seg.NextSegment() + } + // Mark pages overlapping committedFR as committed. + for seg.Ok() && seg.Start() < committedFR.End { + if seg.ValuePtr().canCommit() { + seg = f.usage.Isolate(seg, committedFR) + seg.ValuePtr().knownCommitted = true + amount := seg.Range().Length() + usage.MemoryAccounting.Inc(amount, seg.ValuePtr().kind) + f.usageExpected += amount + changedAny = true } - amount := runRange.Length() - usage.MemoryAccounting.Inc(amount, val.kind) - f.usageExpected += amount - changedAny = true - populatedRun = false + seg = seg.NextSegment() } + // Continue scanning for committed pages. + i = j + 1 } // Advance r.Start. @@ -976,6 +981,9 @@ func (f *MemoryFile) updateUsageLocked(currentUsage uint64, checkCommitted func( if err != nil { return err } + + // Continue with the first segment after r.End. + seg = f.usage.LowerBoundSegment(r.End) } return nil @@ -1167,8 +1175,10 @@ func (f *MemoryFile) startEvictionsLocked() bool { return startedAny } -// Preconditions: info == f.evictable[user]. !info.evicting. f.mu must be -// locked. +// Preconditions: +// * info == f.evictable[user]. +// * !info.evicting. +// * f.mu must be locked. func (f *MemoryFile) startEvictionGoroutineLocked(user EvictableMemoryUser, info *evictableMemoryUserInfo) { info.evicting = true f.evictionWG.Add(1) diff --git a/pkg/sentry/platform/BUILD b/pkg/sentry/platform/BUILD index 209b28053..db7d55ef2 100644 --- a/pkg/sentry/platform/BUILD +++ b/pkg/sentry/platform/BUILD @@ -15,6 +15,7 @@ go_library( "//pkg/context", "//pkg/seccomp", "//pkg/sentry/arch", + "//pkg/sentry/hostmm", "//pkg/sentry/memmap", "//pkg/usermem", ], diff --git a/pkg/sentry/platform/interrupt/interrupt.go b/pkg/sentry/platform/interrupt/interrupt.go index 57be41647..9dfac3eae 100644 --- a/pkg/sentry/platform/interrupt/interrupt.go +++ b/pkg/sentry/platform/interrupt/interrupt.go @@ -54,8 +54,9 @@ type Forwarder struct { // } // defer f.Disable() // -// Preconditions: r must not be nil. f must not already be forwarding -// interrupts to a Receiver. +// 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") diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD index 3970dd81d..8ce411102 100644 --- a/pkg/sentry/platform/kvm/BUILD +++ b/pkg/sentry/platform/kvm/BUILD @@ -9,12 +9,12 @@ go_library( "bluepill.go", "bluepill_allocator.go", "bluepill_amd64.go", - "bluepill_amd64.s", "bluepill_amd64_unsafe.go", "bluepill_arm64.go", "bluepill_arm64.s", "bluepill_arm64_unsafe.go", "bluepill_fault.go", + "bluepill_impl_amd64.s", "bluepill_unsafe.go", "context.go", "filters_amd64.go", @@ -56,6 +56,7 @@ go_library( "//pkg/sentry/time", "//pkg/sync", "//pkg/usermem", + "@org_golang_x_sys//unix:go_default_library", ], ) @@ -63,6 +64,7 @@ go_test( name = "kvm_test", srcs = [ "kvm_amd64_test.go", + "kvm_arm64_test.go", "kvm_test.go", "virtual_map_test.go", ], @@ -78,6 +80,15 @@ go_test( "//pkg/sentry/platform/kvm/testutil", "//pkg/sentry/platform/ring0", "//pkg/sentry/platform/ring0/pagetables", + "//pkg/sentry/time", "//pkg/usermem", ], ) + +genrule( + name = "bluepill_impl_amd64", + srcs = ["bluepill_amd64.s"], + outs = ["bluepill_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"], +) diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.s b/pkg/sentry/platform/kvm/bluepill_amd64.s index 2bc34a435..025ea93b5 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64.s +++ b/pkg/sentry/platform/kvm/bluepill_amd64.s @@ -19,11 +19,6 @@ // 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 @@ -44,7 +39,7 @@ begin: LEAQ VCPU_CPU(AX), BX BYTE CLI; check_vcpu: - MOVQ CPU_SELF(GS), CX + MOVQ ENTRY_CPU_SELF(GS), CX CMPQ BX, CX JE right_vCPU wrong_vcpu: diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go index 03a98512e..0a54dd30d 100644 --- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go @@ -83,5 +83,34 @@ func bluepillStopGuest(c *vCPU) { // //go:nosplit func bluepillReadyStopGuest(c *vCPU) bool { - return c.runData.readyForInterruptInjection != 0 + if c.runData.readyForInterruptInjection == 0 { + return false + } + + if c.runData.ifFlag == 0 { + // This is impossible if readyForInterruptInjection is 1. + throw("interrupts are disabled") + } + + // Disable interrupts if we are in the kernel space. + // + // When the Sentry switches into the kernel mode, it disables + // interrupts. But when goruntime switches on a goroutine which has + // been saved in the host mode, it restores flags and this enables + // interrupts. See the comment of UserFlagsSet for more details. + uregs := userRegs{} + err := c.getUserRegisters(&uregs) + if err != 0 { + throw("failed to get user registers") + } + + if ring0.IsKernelFlags(uregs.RFLAGS) { + uregs.RFLAGS &^= ring0.KernelFlagsClear + err = c.setUserRegisters(&uregs) + if err != 0 { + throw("failed to set user registers") + } + return false + } + return true } diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.go b/pkg/sentry/platform/kvm/bluepill_arm64.go index ed5ae03d3..58f3d6fdd 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.go +++ b/pkg/sentry/platform/kvm/bluepill_arm64.go @@ -39,6 +39,16 @@ var ( } ) +// getTLS returns the value of TPIDR_EL0 register. +// +//go:nosplit +func getTLS() (value uint64) + +// setTLS writes the TPIDR_EL0 value. +// +//go:nosplit +func setTLS(value uint64) + // bluepillArchEnter is called during bluepillEnter. // //go:nosplit @@ -51,6 +61,8 @@ func bluepillArchEnter(context *arch.SignalContext64) (c *vCPU) { regs.Pstate = context.Pstate regs.Pstate &^= uint64(ring0.PsrFlagsClear) regs.Pstate |= ring0.KernelFlagsSet + regs.TPIDR_EL0 = getTLS() + return } @@ -65,6 +77,7 @@ func bluepillArchExit(c *vCPU, context *arch.SignalContext64) { context.Pstate = regs.Pstate context.Pstate &^= uint64(ring0.PsrFlagsClear) context.Pstate |= ring0.UserFlagsSet + setTLS(regs.TPIDR_EL0) lazyVfp := c.GetLazyVFP() if lazyVfp != 0 { diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.s b/pkg/sentry/platform/kvm/bluepill_arm64.s index 04efa0147..09c7e88e5 100644 --- a/pkg/sentry/platform/kvm/bluepill_arm64.s +++ b/pkg/sentry/platform/kvm/bluepill_arm64.s @@ -32,6 +32,18 @@ #define CONTEXT_PC 0x1B8 #define CONTEXT_R0 0xB8 +// getTLS returns the value of TPIDR_EL0 register. +TEXT ·getTLS(SB),NOSPLIT,$0-8 + MRS TPIDR_EL0, R1 + MOVD R1, ret+0(FP) + RET + +// setTLS writes the TPIDR_EL0 value. +TEXT ·setTLS(SB),NOSPLIT,$0-8 + MOVD addr+0(FP), R1 + MSR R1, TPIDR_EL0 + RET + // See bluepill.go. TEXT ·bluepill(SB),NOSPLIT,$0 begin: diff --git a/pkg/sentry/platform/kvm/bluepill_fault.go b/pkg/sentry/platform/kvm/bluepill_fault.go index e34f46aeb..a182e4f22 100644 --- a/pkg/sentry/platform/kvm/bluepill_fault.go +++ b/pkg/sentry/platform/kvm/bluepill_fault.go @@ -98,6 +98,10 @@ func handleBluepillFault(m *machine, physical uintptr, phyRegions []physicalRegi } errno := m.setMemoryRegion(int(slot), physicalStart, length, virtualStart, flags) if errno == 0 { + // Store the physical address in the slot. This is used to + // avoid calls to handleBluepillFault in the future (see + // machine.mapPhysical). + atomic.StoreUintptr(&m.usedSlots[slot], physical) // Successfully added region; we can increment nextSlot and // allow another set to proceed here. atomic.StoreUint32(&m.nextSlot, slot+1) diff --git a/pkg/sentry/platform/kvm/bluepill_unsafe.go b/pkg/sentry/platform/kvm/bluepill_unsafe.go index bf357de1a..eb05950cd 100644 --- a/pkg/sentry/platform/kvm/bluepill_unsafe.go +++ b/pkg/sentry/platform/kvm/bluepill_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.16 +// +build !go1.17 // Check go:linkname function signatures when updating Go version. @@ -62,6 +62,9 @@ func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 { // //go:nosplit func bluepillGuestExit(c *vCPU, context unsafe.Pointer) { + // Increment our counter. + atomic.AddUint64(&c.guestExits, 1) + // Copy out registers. bluepillArchExit(c, bluepillArchContext(context)) @@ -89,9 +92,6 @@ 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) - // Mark this as guest mode. switch atomic.SwapUint32(&c.state, vCPUGuest|vCPUUser) { case vCPUUser: // Expected case. diff --git a/pkg/sentry/platform/kvm/context.go b/pkg/sentry/platform/kvm/context.go index 6e6b76416..17268d127 100644 --- a/pkg/sentry/platform/kvm/context.go +++ b/pkg/sentry/platform/kvm/context.go @@ -15,6 +15,8 @@ package kvm import ( + "sync/atomic" + pkgcontext "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/platform" @@ -75,6 +77,9 @@ func (c *context) Switch(ctx pkgcontext.Context, mm platform.MemoryManager, ac a // Clear the address space. cpu.active.set(nil) + // Increment the number of user exits. + atomic.AddUint64(&cpu.userExits, 1) + // Release resources. c.machine.Put(cpu) diff --git a/pkg/sentry/platform/kvm/filters_amd64.go b/pkg/sentry/platform/kvm/filters_amd64.go index 7d949f1dd..d3d216aa5 100644 --- a/pkg/sentry/platform/kvm/filters_amd64.go +++ b/pkg/sentry/platform/kvm/filters_amd64.go @@ -17,14 +17,23 @@ package kvm import ( "syscall" + "golang.org/x/sys/unix" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" ) // SyscallFilters returns syscalls made exclusively by the KVM platform. func (*KVM) SyscallFilters() seccomp.SyscallRules { return seccomp.SyscallRules{ - syscall.SYS_ARCH_PRCTL: {}, - syscall.SYS_IOCTL: {}, + syscall.SYS_ARCH_PRCTL: {}, + syscall.SYS_IOCTL: {}, + unix.SYS_MEMBARRIER: []seccomp.Rule{ + { + seccomp.EqualTo(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED), + seccomp.EqualTo(0), + }, + }, syscall.SYS_MMAP: {}, syscall.SYS_RT_SIGSUSPEND: {}, syscall.SYS_RT_SIGTIMEDWAIT: {}, diff --git a/pkg/sentry/platform/kvm/filters_arm64.go b/pkg/sentry/platform/kvm/filters_arm64.go index 9245d07c2..21abc2a3d 100644 --- a/pkg/sentry/platform/kvm/filters_arm64.go +++ b/pkg/sentry/platform/kvm/filters_arm64.go @@ -17,13 +17,22 @@ package kvm import ( "syscall" + "golang.org/x/sys/unix" + + "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/seccomp" ) // SyscallFilters returns syscalls made exclusively by the KVM platform. func (*KVM) SyscallFilters() seccomp.SyscallRules { return seccomp.SyscallRules{ - syscall.SYS_IOCTL: {}, + syscall.SYS_IOCTL: {}, + unix.SYS_MEMBARRIER: []seccomp.Rule{ + { + seccomp.EqualTo(linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED), + seccomp.EqualTo(0), + }, + }, syscall.SYS_MMAP: {}, syscall.SYS_RT_SIGSUSPEND: {}, syscall.SYS_RT_SIGTIMEDWAIT: {}, diff --git a/pkg/sentry/platform/kvm/kvm.go b/pkg/sentry/platform/kvm/kvm.go index ae813e24e..dd45ad10b 100644 --- a/pkg/sentry/platform/kvm/kvm.go +++ b/pkg/sentry/platform/kvm/kvm.go @@ -63,6 +63,9 @@ type runData struct { type KVM struct { platform.NoCPUPreemptionDetection + // KVM never changes mm_structs. + platform.UseHostProcessMemoryBarrier + // machine is the backing VM. machine *machine } @@ -156,15 +159,7 @@ func (*KVM) MaxUserAddress() usermem.Addr { func (k *KVM) NewAddressSpace(_ interface{}) (platform.AddressSpace, <-chan struct{}, error) { // Allocate page tables and install system mappings. pageTables := pagetables.New(newAllocator()) - applyPhysicalRegions(func(pr physicalRegion) bool { - // Map the kernel in the upper half. - pageTables.Map( - usermem.Addr(ring0.KernelStartAddress|pr.virtual), - pr.length, - pagetables.MapOpts{AccessType: usermem.AnyAccess}, - pr.physical) - return true // Keep iterating. - }) + k.machine.mapUpperHalf(pageTables) // Return the new address space. return &addressSpace{ diff --git a/pkg/sentry/platform/kvm/kvm_arm64_test.go b/pkg/sentry/platform/kvm/kvm_arm64_test.go new file mode 100644 index 000000000..0e3d84d95 --- /dev/null +++ b/pkg/sentry/platform/kvm/kvm_arm64_test.go @@ -0,0 +1,31 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT 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 arm64 + +package kvm + +import ( + "testing" + + "gvisor.dev/gvisor/pkg/sentry/platform/kvm/testutil" +) + +func TestKernelTLS(t *testing.T) { + bluepillTest(t, func(c *vCPU) { + if !testutil.TLSWorks() { + t.Errorf("tls does not work, and it should!") + } + }) +} diff --git a/pkg/sentry/platform/kvm/kvm_const.go b/pkg/sentry/platform/kvm/kvm_const.go index 3bf918446..6abaa21c4 100644 --- a/pkg/sentry/platform/kvm/kvm_const.go +++ b/pkg/sentry/platform/kvm/kvm_const.go @@ -26,12 +26,16 @@ const ( _KVM_RUN = 0xae80 _KVM_NMI = 0xae9a _KVM_CHECK_EXTENSION = 0xae03 + _KVM_GET_TSC_KHZ = 0xaea3 + _KVM_SET_TSC_KHZ = 0xaea2 _KVM_INTERRUPT = 0x4004ae86 _KVM_SET_MSRS = 0x4008ae89 _KVM_SET_USER_MEMORY_REGION = 0x4020ae46 _KVM_SET_REGS = 0x4090ae82 _KVM_SET_SREGS = 0x4138ae84 + _KVM_GET_MSRS = 0xc008ae88 _KVM_GET_REGS = 0x8090ae81 + _KVM_GET_SREGS = 0x8138ae83 _KVM_GET_SUPPORTED_CPUID = 0xc008ae05 _KVM_SET_CPUID2 = 0x4008ae90 _KVM_SET_SIGNAL_MASK = 0x4004ae8b @@ -56,6 +60,7 @@ const ( // KVM capability options. const ( + _KVM_CAP_MAX_MEMSLOTS = 0x0a _KVM_CAP_MAX_VCPUS = 0x42 _KVM_CAP_ARM_VM_IPA_SIZE = 0xa5 _KVM_CAP_VCPU_EVENTS = 0x29 @@ -64,6 +69,7 @@ const ( // KVM limits. const ( + _KVM_NR_MEMSLOTS = 0x100 _KVM_NR_VCPUS = 0xff _KVM_NR_INTERRUPTS = 0x100 _KVM_NR_CPUID_ENTRIES = 0x100 @@ -77,11 +83,14 @@ const ( ) // KVM hypercall list. +// // Canonical list of hypercalls supported. const ( // On amd64, it uses 'HLT' to leave the guest. + // // Unlike amd64, arm64 can only uses mmio_exit/psci to leave the guest. - // _KVM_HYPERCALL_VMEXIT is only used on Arm64 for now. + // + // _KVM_HYPERCALL_VMEXIT is only used on arm64 for now. _KVM_HYPERCALL_VMEXIT int = iota _KVM_HYPERCALL_MAX ) diff --git a/pkg/sentry/platform/kvm/kvm_const_arm64.go b/pkg/sentry/platform/kvm/kvm_const_arm64.go index 9a7be3655..84df0f878 100644 --- a/pkg/sentry/platform/kvm/kvm_const_arm64.go +++ b/pkg/sentry/platform/kvm/kvm_const_arm64.go @@ -101,13 +101,20 @@ const ( // Arm64: Memory Attribute Indirection Register EL1. const ( - _MT_DEVICE_nGnRnE = 0 - _MT_DEVICE_nGnRE = 1 - _MT_DEVICE_GRE = 2 - _MT_NORMAL_NC = 3 - _MT_NORMAL = 4 - _MT_NORMAL_WT = 5 - _MT_EL1_INIT = (0 << _MT_DEVICE_nGnRnE) | (0x4 << _MT_DEVICE_nGnRE * 8) | (0xc << _MT_DEVICE_GRE * 8) | (0x44 << _MT_NORMAL_NC * 8) | (0xff << _MT_NORMAL * 8) | (0xbb << _MT_NORMAL_WT * 8) + _MT_DEVICE_nGnRnE = 0 + _MT_DEVICE_nGnRE = 1 + _MT_DEVICE_GRE = 2 + _MT_NORMAL_NC = 3 + _MT_NORMAL = 4 + _MT_NORMAL_WT = 5 + _MT_ATTR_DEVICE_nGnRnE = 0x00 + _MT_ATTR_DEVICE_nGnRE = 0x04 + _MT_ATTR_DEVICE_GRE = 0x0c + _MT_ATTR_NORMAL_NC = 0x44 + _MT_ATTR_NORMAL_WT = 0xbb + _MT_ATTR_NORMAL = 0xff + _MT_ATTR_MASK = 0xff + _MT_EL1_INIT = (_MT_ATTR_DEVICE_nGnRnE << (_MT_DEVICE_nGnRnE * 8)) | (_MT_ATTR_DEVICE_nGnRE << (_MT_DEVICE_nGnRE * 8)) | (_MT_ATTR_DEVICE_GRE << (_MT_DEVICE_GRE * 8)) | (_MT_ATTR_NORMAL_NC << (_MT_NORMAL_NC * 8)) | (_MT_ATTR_NORMAL << (_MT_NORMAL * 8)) | (_MT_ATTR_NORMAL_WT << (_MT_NORMAL_WT * 8)) ) const ( diff --git a/pkg/sentry/platform/kvm/kvm_test.go b/pkg/sentry/platform/kvm/kvm_test.go index 45b3180f1..e58acc071 100644 --- a/pkg/sentry/platform/kvm/kvm_test.go +++ b/pkg/sentry/platform/kvm/kvm_test.go @@ -27,6 +27,7 @@ import ( "gvisor.dev/gvisor/pkg/sentry/platform/kvm/testutil" "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" + ktime "gvisor.dev/gvisor/pkg/sentry/time" "gvisor.dev/gvisor/pkg/usermem" ) @@ -411,9 +412,9 @@ func TestWrongVCPU(t *testing.T) { // Basic test, one then the other. bluepill(c1) bluepill(c2) - if c2.switches == 0 { + if c2.guestExits == 0 { // Don't allow the test to proceed if this fails. - t.Fatalf("wrong vCPU#2 switches: vCPU1=%+v,vCPU2=%+v", c1, c2) + t.Fatalf("wrong vCPU#2 exits: vCPU1=%+v,vCPU2=%+v", c1, c2) } // Alternate vCPUs; we expect to need to trigger the @@ -422,11 +423,11 @@ func TestWrongVCPU(t *testing.T) { bluepill(c1) bluepill(c2) } - if count := c1.switches; count < 90 { - t.Errorf("wrong vCPU#1 switches: vCPU1=%+v,vCPU2=%+v", c1, c2) + if count := c1.guestExits; count < 90 { + t.Errorf("wrong vCPU#1 exits: vCPU1=%+v,vCPU2=%+v", c1, c2) } - if count := c2.switches; count < 90 { - t.Errorf("wrong vCPU#2 switches: vCPU1=%+v,vCPU2=%+v", c1, c2) + if count := c2.guestExits; count < 90 { + t.Errorf("wrong vCPU#2 exits: vCPU1=%+v,vCPU2=%+v", c1, c2) } return false }) @@ -442,6 +443,22 @@ func TestWrongVCPU(t *testing.T) { }) } +func TestRdtsc(t *testing.T) { + var i int // Iteration count. + kvmTest(t, nil, func(c *vCPU) bool { + start := ktime.Rdtsc() + bluepill(c) + guest := ktime.Rdtsc() + redpill() + end := ktime.Rdtsc() + if start > guest || guest > end { + t.Errorf("inconsistent time: start=%d, guest=%d, end=%d", start, guest, end) + } + i++ + return i < 100 + }) +} + func BenchmarkApplicationSyscall(b *testing.B) { var ( i int // Iteration includes machine.Get() / machine.Put(). diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go index 6c54712d1..61ed24d01 100644 --- a/pkg/sentry/platform/kvm/machine.go +++ b/pkg/sentry/platform/kvm/machine.go @@ -43,9 +43,6 @@ type machine struct { // kernel is the set of global structures. kernel ring0.Kernel - // mappingCache is used for mapPhysical. - mappingCache sync.Map - // mu protects vCPUs. mu sync.RWMutex @@ -63,6 +60,12 @@ type machine struct { // maxVCPUs is the maximum number of vCPUs supported by the machine. maxVCPUs int + // maxSlots is the maximum number of memory slots supported by the machine. + maxSlots int + + // usedSlots is the set of used physical addresses (sorted). + usedSlots []uintptr + // nextID is the next vCPU ID. nextID uint32 } @@ -100,8 +103,11 @@ type vCPU struct { // tid is the last set tid. tid uint64 - // switches is a count of world switches (informational only). - switches uint32 + // userExits is the count of user exits. + userExits uint64 + + // guestExits is the count of guest to host world switches. + guestExits uint64 // faults is a count of world faults (informational only). faults uint32 @@ -124,6 +130,7 @@ type vCPU struct { // vCPUArchState is the architecture-specific state. vCPUArchState + // dieState holds state related to vCPU death. dieState dieState } @@ -152,7 +159,7 @@ func (m *machine) newVCPU() *vCPU { fd: int(fd), machine: m, } - c.CPU.Init(&m.kernel, c) + c.CPU.Init(&m.kernel, c.id, c) m.vCPUsByID[c.id] = c // Ensure the signal mask is correct. @@ -180,10 +187,8 @@ func newMachine(vm int) (*machine, error) { // Create the machine. m := &machine{fd: vm} m.available.L = &m.mu - m.kernel.Init(ring0.KernelOpts{ - PageTables: pagetables.New(newAllocator()), - }) + // Pull the maximum vCPUs. maxVCPUs, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CHECK_EXTENSION, _KVM_CAP_MAX_VCPUS) if errno != 0 { m.maxVCPUs = _KVM_NR_VCPUS @@ -191,10 +196,21 @@ func newMachine(vm int) (*machine, error) { m.maxVCPUs = int(maxVCPUs) } log.Debugf("The maximum number of vCPUs is %d.", m.maxVCPUs) - - // Create the vCPUs map/slices. m.vCPUsByTID = make(map[uint64]*vCPU) m.vCPUsByID = make([]*vCPU, m.maxVCPUs) + m.kernel.Init(ring0.KernelOpts{ + PageTables: pagetables.New(newAllocator()), + }, m.maxVCPUs) + + // Pull the maximum slots. + maxSlots, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CHECK_EXTENSION, _KVM_CAP_MAX_MEMSLOTS) + if errno != 0 { + m.maxSlots = _KVM_NR_MEMSLOTS + } else { + m.maxSlots = int(maxSlots) + } + log.Debugf("The maximum number of slots is %d.", m.maxSlots) + m.usedSlots = make([]uintptr, m.maxSlots) // Apply the physical mappings. Note that these mappings may point to // guest physical addresses that are not actually available. These @@ -207,15 +223,9 @@ func newMachine(vm int) (*machine, error) { pagetables.MapOpts{AccessType: usermem.AnyAccess}, pr.physical) - // And keep everything in the upper half. - m.kernel.PageTables.Map( - usermem.Addr(ring0.KernelStartAddress|pr.virtual), - pr.length, - pagetables.MapOpts{AccessType: usermem.AnyAccess}, - pr.physical) - return true // Keep iterating. }) + m.mapUpperHalf(m.kernel.PageTables) var physicalRegionsReadOnly []physicalRegion var physicalRegionsAvailable []physicalRegion @@ -272,6 +282,20 @@ func newMachine(vm int) (*machine, error) { return m, nil } +// hasSlot returns true iff the given address is mapped. +// +// This must be done via a linear scan. +// +//go:nosplit +func (m *machine) hasSlot(physical uintptr) bool { + for i := 0; i < len(m.usedSlots); i++ { + if p := atomic.LoadUintptr(&m.usedSlots[i]); p == physical { + return true + } + } + return false +} + // 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. // @@ -286,8 +310,8 @@ func (m *machine) mapPhysical(physical, length uintptr, phyRegions []physicalReg panic("mapPhysical on unknown physical address") } - if _, ok := m.mappingCache.LoadOrStore(physicalStart, true); !ok { - // Not present in the cache; requires setting the slot. + // Is this already mapped? Check the usedSlots. + if !m.hasSlot(physicalStart) { if _, ok := handleBluepillFault(m, physical, phyRegions, flags); !ok { panic("handleBluepillFault failed") } @@ -339,6 +363,11 @@ func (m *machine) Destroy() { // Get gets an available vCPU. // // This will return with the OS thread locked. +// +// It is guaranteed that if any OS thread TID is in guest, m.vCPUs[TID] points +// to the vCPU in which the OS thread TID is running. So if Get() returns with +// the corrent context in guest, the vCPU of it must be the same as what +// Get() returns. func (m *machine) Get() *vCPU { m.mu.RLock() runtime.LockOSThread() @@ -443,6 +472,19 @@ func (m *machine) newDirtySet() *dirtySet { } } +// dropPageTables drops cached page table entries. +func (m *machine) dropPageTables(pt *pagetables.PageTables) { + m.mu.Lock() + defer m.mu.Unlock() + + // Clear from all PCIDs. + for _, c := range m.vCPUsByID { + if c != nil && c.PCIDs != nil { + c.PCIDs.Drop(pt) + } + } +} + // lock marks the vCPU as in user mode. // // This should only be called directly when known to be safe, i.e. when @@ -502,6 +544,8 @@ var pid = syscall.Getpid() // // This effectively unwinds the state machine. func (c *vCPU) bounce(forceGuestExit bool) { + origGuestExits := atomic.LoadUint64(&c.guestExits) + origUserExits := atomic.LoadUint64(&c.userExits) for { switch state := atomic.LoadUint32(&c.state); state { case vCPUReady, vCPUWaiter: @@ -557,6 +601,14 @@ func (c *vCPU) bounce(forceGuestExit bool) { // Should not happen: the above is exhaustive. panic("invalid state") } + + // Check if we've missed the state transition, but + // we can safely return at this point in time. + newGuestExits := atomic.LoadUint64(&c.guestExits) + newUserExits := atomic.LoadUint64(&c.userExits) + if newUserExits != origUserExits && (!forceGuestExit || newGuestExits != origGuestExits) { + return + } } } diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go index acc823ba6..c67127d95 100644 --- a/pkg/sentry/platform/kvm/machine_amd64.go +++ b/pkg/sentry/platform/kvm/machine_amd64.go @@ -18,14 +18,17 @@ package kvm import ( "fmt" + "math/big" "reflect" "runtime/debug" "syscall" + "gvisor.dev/gvisor/pkg/cpuid" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/platform" "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" + ktime "gvisor.dev/gvisor/pkg/sentry/time" "gvisor.dev/gvisor/pkg/usermem" ) @@ -84,19 +87,6 @@ const ( poolPCIDs = 8 ) -// dropPageTables drops cached page table entries. -func (m *machine) dropPageTables(pt *pagetables.PageTables) { - m.mu.Lock() - defer m.mu.Unlock() - - // Clear from all PCIDs. - for _, c := range m.vCPUsByID { - if c != nil && c.PCIDs != nil { - c.PCIDs.Drop(pt) - } - } -} - // initArchState initializes architecture-specific state. func (c *vCPU) initArchState() error { var ( @@ -144,6 +134,7 @@ func (c *vCPU) initArchState() error { // Set the entrypoint for the kernel. kernelUserRegs.RIP = uint64(reflect.ValueOf(ring0.Start).Pointer()) kernelUserRegs.RAX = uint64(reflect.ValueOf(&c.CPU).Pointer()) + kernelUserRegs.RSP = c.StackTop() kernelUserRegs.RFLAGS = ring0.KernelFlagsSet // Set the system registers. @@ -152,8 +143,8 @@ func (c *vCPU) initArchState() error { } // Set the user registers. - if err := c.setUserRegisters(&kernelUserRegs); err != nil { - return err + if errno := c.setUserRegisters(&kernelUserRegs); errno != 0 { + return fmt.Errorf("error setting user registers: %v", errno) } // Allocate some floating point state save area for the local vCPU. @@ -166,6 +157,133 @@ func (c *vCPU) initArchState() error { return c.setSystemTime() } +// bitsForScaling returns the bits available for storing the fraction component +// of the TSC scaling ratio. This allows us to replicate the (bad) math done by +// the kernel below in scaledTSC, and ensure we can compute an exact zero +// offset in setSystemTime. +// +// These constants correspond to kvm_tsc_scaling_ratio_frac_bits. +var bitsForScaling = func() int64 { + fs := cpuid.HostFeatureSet() + if fs.Intel() { + return 48 // See vmx.c (kvm sources). + } else if fs.AMD() { + return 32 // See svm.c (svm sources). + } else { + return 63 // Unknown: theoretical maximum. + } +}() + +// scaledTSC returns the host TSC scaled by the given frequency. +// +// This assumes a current frequency of 1. We require only the unitless ratio of +// rawFreq to some current frequency. See setSystemTime for context. +// +// The kernel math guarantees that all bits of the multiplication and division +// will be correctly preserved and applied. However, it is not possible to +// actually store the ratio correctly. So we need to use the same schema in +// order to calculate the scaled frequency and get the same result. +// +// We can assume that the current frequency is (1), so we are calculating a +// strict inverse of this value. This simplifies this function considerably. +// +// Roughly, the returned value "scaledTSC" will have: +// scaledTSC/hostTSC == 1/rawFreq +// +//go:nosplit +func scaledTSC(rawFreq uintptr) int64 { + scale := int64(1 << bitsForScaling) + ratio := big.NewInt(scale / int64(rawFreq)) + ratio.Mul(ratio, big.NewInt(int64(ktime.Rdtsc()))) + ratio.Div(ratio, big.NewInt(scale)) + return ratio.Int64() +} + +// setSystemTime sets the vCPU to the system time. +func (c *vCPU) setSystemTime() error { + // First, scale down the clock frequency to the lowest value allowed by + // the API itself. How low we can go depends on the underlying + // hardware, but it is typically ~1/2^48 for Intel, ~1/2^32 for AMD. + // Even the lower bound here will take a 4GHz frequency down to 1Hz, + // meaning that everything should be able to handle a Khz setting of 1 + // with bits to spare. + // + // Note that reducing the clock does not typically require special + // capabilities as it is emulated in KVM. We don't actually use this + // capability, but it means that this method should be robust to + // different hardware configurations. + rawFreq, err := c.getTSCFreq() + if err != nil { + return c.setSystemTimeLegacy() + } + if err := c.setTSCFreq(1); err != nil { + return c.setSystemTimeLegacy() + } + + // Always restore the original frequency. + defer func() { + if err := c.setTSCFreq(rawFreq); err != nil { + panic(err.Error()) + } + }() + + // Attempt to set the system time in this compressed world. The + // calculation for offset normally looks like: + // + // offset = target_tsc - kvm_scale_tsc(vcpu, rdtsc()); + // + // So as long as the kvm_scale_tsc component is constant before and + // after the call to set the TSC value (and it is passes as the + // target_tsc), we will compute an offset value of zero. + // + // This is effectively cheating to make our "setSystemTime" call so + // unbelievably, incredibly fast that we do it "instantly" and all the + // calculations result in an offset of zero. + lastTSC := scaledTSC(rawFreq) + for { + if err := c.setTSC(uint64(lastTSC)); err != nil { + return err + } + nextTSC := scaledTSC(rawFreq) + if lastTSC == nextTSC { + return nil + } + lastTSC = nextTSC // Try again. + } +} + +// setSystemTimeLegacy calibrates and sets an approximate system time. +func (c *vCPU) setSystemTimeLegacy() error { + const minIterations = 10 + minimum := uint64(0) + for iter := 0; ; iter++ { + // Try to set the TSC to an estimate of where it will be + // on the host during a "fast" system call iteration. + start := uint64(ktime.Rdtsc()) + if err := c.setTSC(start + (minimum / 2)); err != nil { + return err + } + // See if this is our new minimum call time. Note that this + // serves two functions: one, we make sure that we are + // accurately predicting the offset we need to set. Second, we + // don't want to do the final set on a slow call, which could + // produce a really bad result. + end := uint64(ktime.Rdtsc()) + if end < start { + continue // Totally bogus: unstable TSC? + } + current := end - start + if current < minimum || iter == 0 { + minimum = current // Set our new minimum. + } + // Is this past minIterations and within ~10% of minimum? + upperThreshold := (((minimum << 3) + minimum) >> 3) + if iter >= minIterations && current <= upperThreshold { + return nil + } + } +} + // nonCanonical generates a canonical address return. // //go:nosplit @@ -345,3 +463,41 @@ func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { func availableRegionsForSetMem() (phyRegions []physicalRegion) { return physicalRegions } + +var execRegions = func() (regions []region) { + applyVirtualRegions(func(vr virtualRegion) { + if excludeVirtualRegion(vr) || vr.filename == "[vsyscall]" { + return + } + if vr.accessType.Execute { + regions = append(regions, vr.region) + } + }) + return +}() + +func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) { + for _, r := range execRegions { + physical, length, ok := translateToPhysical(r.virtual) + if !ok || length < r.length { + panic("impossilbe translation") + } + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|r.virtual), + r.length, + pagetables.MapOpts{AccessType: usermem.Execute}, + physical) + } + for start, end := range m.kernel.EntryRegions() { + regionLen := end - start + physical, length, ok := translateToPhysical(start) + if !ok || length < regionLen { + panic("impossible translation") + } + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|start), + regionLen, + pagetables.MapOpts{AccessType: usermem.ReadWrite}, + physical) + } +} diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go index 290f035dd..b430f92c6 100644 --- a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go @@ -23,7 +23,6 @@ import ( "unsafe" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/time" ) // loadSegments copies the current segments. @@ -61,91 +60,63 @@ func (c *vCPU) setCPUID() error { return nil } -// setSystemTime sets the TSC for the vCPU. +// getTSCFreq gets the TSC frequency. // -// This has to make the call many times in order to minimize the intrinsic -// error in the offset. Unfortunately KVM does not expose a relative offset via -// the API, so this is an approximation. We do this via an iterative algorithm. -// This has the advantage that it can generally deal with highly variable -// system call times and should converge on the correct offset. -func (c *vCPU) setSystemTime() error { - const ( - _MSR_IA32_TSC = 0x00000010 - calibrateTries = 10 - ) - registers := modelControlRegisters{ - nmsrs: 1, - } - registers.entries[0] = modelControlRegister{ - index: _MSR_IA32_TSC, +// If mustSucceed is true, then this function panics on error. +func (c *vCPU) getTSCFreq() (uintptr, error) { + rawFreq, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_GET_TSC_KHZ, + 0 /* ignored */) + if errno != 0 { + return 0, errno } - target := uint64(^uint32(0)) - for done := 0; done < calibrateTries; { - start := uint64(time.Rdtsc()) - registers.entries[0].data = start + target - if _, _, errno := syscall.RawSyscall( - syscall.SYS_IOCTL, - uintptr(c.fd), - _KVM_SET_MSRS, - uintptr(unsafe.Pointer(®isters))); errno != 0 { - return fmt.Errorf("error setting system time: %v", errno) - } - // See if this is our new minimum call time. Note that this - // serves two functions: one, we make sure that we are - // accurately predicting the offset we need to set. Second, we - // don't want to do the final set on a slow call, which could - // produce a really bad result. So we only count attempts - // within +/- 6.25% of our minimum as an attempt. - end := uint64(time.Rdtsc()) - if end < start { - continue // Totally bogus. - } - half := (end - start) / 2 - if half < target { - target = half - } - if (half - target) < target/8 { - done++ - } + return rawFreq, nil +} + +// setTSCFreq sets the TSC frequency. +func (c *vCPU) setTSCFreq(freq uintptr) error { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_SET_TSC_KHZ, + freq /* khz */); errno != 0 { + return fmt.Errorf("error setting TSC frequency: %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 +// setTSC sets the TSC value. +func (c *vCPU) setTSC(value uint64) error { + const _MSR_IA32_TSC = 0x00000010 + registers := modelControlRegisters{ + nmsrs: 1, } - data.length = 8 // Fixed sigset size. - data.mask1 = ^uint32(bounceSignalMask & 0xffffffff) - data.mask2 = ^uint32(bounceSignalMask >> 32) + registers.entries[0].index = _MSR_IA32_TSC + registers.entries[0].data = value 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) + _KVM_SET_MSRS, + uintptr(unsafe.Pointer(®isters))); errno != 0 { + return fmt.Errorf("error setting tsc: %v", errno) } return nil } // setUserRegisters sets user registers in the vCPU. -func (c *vCPU) setUserRegisters(uregs *userRegs) error { +// +//go:nosplit +func (c *vCPU) setUserRegisters(uregs *userRegs) syscall.Errno { 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 errno } - return nil + return 0 } // getUserRegisters reloads user registers in the vCPU. @@ -175,3 +146,17 @@ func (c *vCPU) setSystemRegisters(sregs *systemRegs) error { } return nil } + +// getSystemRegisters sets system registers. +// +//go:nosplit +func (c *vCPU) getSystemRegisters(sregs *systemRegs) syscall.Errno { + if _, _, errno := syscall.RawSyscall( + syscall.SYS_IOCTL, + uintptr(c.fd), + _KVM_GET_SREGS, + uintptr(unsafe.Pointer(sregs))); errno != 0 { + return errno + } + return 0 +} diff --git a/pkg/sentry/platform/kvm/machine_arm64.go b/pkg/sentry/platform/kvm/machine_arm64.go index 9db171af9..54837f20c 100644 --- a/pkg/sentry/platform/kvm/machine_arm64.go +++ b/pkg/sentry/platform/kvm/machine_arm64.go @@ -19,6 +19,7 @@ package kvm import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/platform" + "gvisor.dev/gvisor/pkg/sentry/platform/ring0" "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables" "gvisor.dev/gvisor/pkg/usermem" ) @@ -48,6 +49,18 @@ const ( poolPCIDs = 8 ) +func (m *machine) mapUpperHalf(pageTable *pagetables.PageTables) { + applyPhysicalRegions(func(pr physicalRegion) bool { + pageTable.Map( + usermem.Addr(ring0.KernelStartAddress|pr.virtual), + pr.length, + pagetables.MapOpts{AccessType: usermem.AnyAccess}, + pr.physical) + + return true // Keep iterating. + }) +} + // Get all read-only physicalRegions. func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) { var rdonlyRegions []region @@ -100,19 +113,6 @@ func availableRegionsForSetMem() (phyRegions []physicalRegion) { return phyRegions } -// dropPageTables drops cached page table entries. -func (m *machine) dropPageTables(pt *pagetables.PageTables) { - m.mu.Lock() - defer m.mu.Unlock() - - // Clear from all PCIDs. - for _, c := range m.vCPUsByID { - if c.PCIDs != nil { - c.PCIDs.Drop(pt) - } - } -} - // nonCanonical generates a canonical address return. // //go:nosplit diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go index 905712076..84992c06d 100644 --- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go @@ -79,7 +79,7 @@ func (c *vCPU) initArchState() error { } // tcr_el1 - data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS + data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS | _TCR_A1 reg.id = _KVM_ARM64_REGS_TCR_EL1 if err := c.setOneRegister(®); err != nil { return err @@ -103,7 +103,7 @@ func (c *vCPU) initArchState() error { c.SetTtbr0Kvm(uintptr(data)) // ttbr1_el1 - data = c.machine.kernel.PageTables.TTBR1_EL1(false, 0) + data = c.machine.kernel.PageTables.TTBR1_EL1(false, 1) reg.id = _KVM_ARM64_REGS_TTBR1_EL1 if err := c.setOneRegister(®); err != nil { @@ -191,42 +191,6 @@ func (c *vCPU) getOneRegister(reg *kvmOneReg) error { return nil } -// setCPUID sets the CPUID to be used by the guest. -func (c *vCPU) setCPUID() error { - return nil -} - -// setSystemTime sets the TSC for the vCPU. -func (c *vCPU) setSystemTime() error { - 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 -} - // SwitchToUser unpacks architectural-details. func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo) (usermem.AccessType, error) { // Check for canonical addresses. @@ -271,8 +235,9 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo) return c.fault(int32(syscall.SIGSEGV), info) case ring0.Vector(bounce): // ring0.VirtualizationException return usermem.NoAccess, platform.ErrContextInterrupt - case ring0.El0Sync_undef, - ring0.El1Sync_undef: + case ring0.El0Sync_undef: + return c.fault(int32(syscall.SIGILL), info) + case ring0.El1Sync_undef: *info = arch.SignalInfo{ Signo: int32(syscall.SIGILL), Code: 1, // ILL_ILLOPC (illegal opcode). diff --git a/pkg/sentry/platform/kvm/machine_unsafe.go b/pkg/sentry/platform/kvm/machine_unsafe.go index 9f86f6a7a..1d6ca245a 100644 --- a/pkg/sentry/platform/kvm/machine_unsafe.go +++ b/pkg/sentry/platform/kvm/machine_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.16 +// +build !go1.17 // Check go:linkname function signatures when updating Go version. @@ -143,3 +143,29 @@ func (c *vCPU) waitUntilNot(state uint32) { panic("futex wait error") } } + +// 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/testutil/testutil_arm64.go b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go index 4dad877ba..c5235ca9d 100644 --- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.go +++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go @@ -23,6 +23,11 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" ) +// TLSWorks is a tls test. +// +// It returns true or false. +func TLSWorks() bool + // SetTestTarget sets the rip appropriately. func SetTestTarget(regs *arch.Registers, fn func()) { regs.Pc = uint64(reflect.ValueOf(fn).Pointer()) diff --git a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s index 6caf7282d..7348c29a5 100644 --- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s +++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s @@ -50,6 +50,22 @@ TEXT ·SpinLoop(SB),NOSPLIT,$0 start: B start +TEXT ·TLSWorks(SB),NOSPLIT,$0-8 + NO_LOCAL_POINTERS + MOVD $0x6789, R5 + MSR R5, TPIDR_EL0 + MOVD $SYS_GETPID, R8 // getpid + SVC + MRS TPIDR_EL0, R6 + CMP R5, R6 + BNE isNaN + MOVD $1, R0 + MOVD R0, ret+0(FP) + RET +isNaN: + MOVD $0, ret+0(FP) + RET + TEXT ·FloatingPointWorks(SB),NOSPLIT,$0-8 NO_LOCAL_POINTERS // gc will touch fpsimd, so we should test it. diff --git a/pkg/sentry/platform/kvm/virtual_map.go b/pkg/sentry/platform/kvm/virtual_map.go index c8897d34f..4dcdbf8a7 100644 --- a/pkg/sentry/platform/kvm/virtual_map.go +++ b/pkg/sentry/platform/kvm/virtual_map.go @@ -34,7 +34,7 @@ type virtualRegion struct { } // 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+(.*)") +var mapsLine = regexp.MustCompile("([0-9a-f]+)-([0-9a-f]+) ([r-][w-][x-][sp]) ([0-9a-f]+) [0-9a-f]{2,3}:[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 diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go index ba031516a..dcfe839a7 100644 --- a/pkg/sentry/platform/platform.go +++ b/pkg/sentry/platform/platform.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/hostmm" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/usermem" ) @@ -52,6 +53,10 @@ type Platform interface { // can reliably return ErrContextCPUPreempted. DetectsCPUPreemption() bool + // HaveGlobalMemoryBarrier returns true if the GlobalMemoryBarrier method + // is supported. + HaveGlobalMemoryBarrier() bool + // MapUnit returns the alignment used for optional mappings into this // platform's AddressSpaces. Higher values indicate lower per-page costs // for AddressSpace.MapFile. As a special case, a MapUnit of 0 indicates @@ -97,6 +102,15 @@ type Platform interface { // called. PreemptAllCPUs() error + // GlobalMemoryBarrier blocks until all threads running application code + // (via Context.Switch) and all task goroutines "have passed through a + // state where all memory accesses to user-space addresses match program + // order between entry to and return from [GlobalMemoryBarrier]", as for + // membarrier(2). + // + // Preconditions: HaveGlobalMemoryBarrier() == true. + GlobalMemoryBarrier() error + // SyscallFilters returns syscalls made exclusively by this platform. SyscallFilters() seccomp.SyscallRules } @@ -115,6 +129,43 @@ func (NoCPUPreemptionDetection) PreemptAllCPUs() error { panic("This platform does not support CPU preemption detection") } +// UseHostGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier and +// Platform.GlobalMemoryBarrier by invoking equivalent functionality on the +// host. +type UseHostGlobalMemoryBarrier struct{} + +// HaveGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier. +func (UseHostGlobalMemoryBarrier) HaveGlobalMemoryBarrier() bool { + return hostmm.HaveGlobalMemoryBarrier() +} + +// GlobalMemoryBarrier implements Platform.GlobalMemoryBarrier. +func (UseHostGlobalMemoryBarrier) GlobalMemoryBarrier() error { + return hostmm.GlobalMemoryBarrier() +} + +// UseHostProcessMemoryBarrier implements Platform.HaveGlobalMemoryBarrier and +// Platform.GlobalMemoryBarrier by invoking a process-local memory barrier. +// This is faster than UseHostGlobalMemoryBarrier, but is only appropriate for +// platforms for which application code executes while using the sentry's +// mm_struct. +type UseHostProcessMemoryBarrier struct{} + +// HaveGlobalMemoryBarrier implements Platform.HaveGlobalMemoryBarrier. +func (UseHostProcessMemoryBarrier) HaveGlobalMemoryBarrier() bool { + // Fall back to a global memory barrier if a process-local one isn't + // available. + return hostmm.HaveProcessMemoryBarrier() || hostmm.HaveGlobalMemoryBarrier() +} + +// GlobalMemoryBarrier implements Platform.GlobalMemoryBarrier. +func (UseHostProcessMemoryBarrier) GlobalMemoryBarrier() error { + if hostmm.HaveProcessMemoryBarrier() { + return hostmm.ProcessMemoryBarrier() + } + return hostmm.GlobalMemoryBarrier() +} + // MemoryManager represents an abstraction above the platform address space // which manages memory mappings and their contents. type MemoryManager interface { @@ -245,14 +296,19 @@ type AddressSpace interface { // physical memory) to the mapping. The precommit flag is advisory and // implementations may choose to ignore it. // - // Preconditions: addr and fr must be page-aligned. fr.Length() > 0. - // at.Any() == true. At least one reference must be held on all pages in - // fr, and must continue to be held as long as pages are mapped. + // Preconditions: + // * addr and fr must be page-aligned. + // * fr.Length() > 0. + // * at.Any() == true. + // * At least one reference must be held on all pages in fr, and must + // continue to be held as long as pages are mapped. MapFile(addr usermem.Addr, f memmap.File, fr memmap.FileRange, at usermem.AccessType, precommit bool) error // Unmap unmaps the given range. // - // Preconditions: addr is page-aligned. length > 0. + // Preconditions: + // * addr is page-aligned. + // * length > 0. Unmap(addr usermem.Addr, length uint64) // Release releases this address space. After releasing, a new AddressSpace diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD index e04165fbf..fc43cc3c0 100644 --- a/pkg/sentry/platform/ptrace/BUILD +++ b/pkg/sentry/platform/ptrace/BUILD @@ -30,7 +30,6 @@ go_library( "//pkg/safecopy", "//pkg/seccomp", "//pkg/sentry/arch", - "//pkg/sentry/hostcpu", "//pkg/sentry/memmap", "//pkg/sentry/platform", "//pkg/sentry/platform/interrupt", diff --git a/pkg/sentry/platform/ptrace/filters.go b/pkg/sentry/platform/ptrace/filters.go index 1e07cfd0d..b0970e356 100644 --- a/pkg/sentry/platform/ptrace/filters.go +++ b/pkg/sentry/platform/ptrace/filters.go @@ -24,10 +24,9 @@ import ( // SyscallFilters returns syscalls made exclusively by the ptrace platform. func (*PTrace) SyscallFilters() seccomp.SyscallRules { return seccomp.SyscallRules{ - unix.SYS_GETCPU: {}, - unix.SYS_SCHED_SETAFFINITY: {}, - syscall.SYS_PTRACE: {}, - syscall.SYS_TGKILL: {}, - syscall.SYS_WAIT4: {}, + unix.SYS_GETCPU: {}, + syscall.SYS_PTRACE: {}, + syscall.SYS_TGKILL: {}, + syscall.SYS_WAIT4: {}, } } diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go index b52d0fbd8..f56aa3b79 100644 --- a/pkg/sentry/platform/ptrace/ptrace.go +++ b/pkg/sentry/platform/ptrace/ptrace.go @@ -192,6 +192,7 @@ func (c *context) PullFullState(as platform.AddressSpace, ac arch.Context) {} type PTrace struct { platform.MMapMinAddr platform.NoCPUPreemptionDetection + platform.UseHostGlobalMemoryBarrier } // New returns a new ptrace-based implementation of the platform interface. diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go index e1d54d8a2..812ab80ef 100644 --- a/pkg/sentry/platform/ptrace/subprocess.go +++ b/pkg/sentry/platform/ptrace/subprocess.go @@ -518,11 +518,6 @@ func (s *subprocess) switchToApp(c *context, ac arch.Context) bool { } 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 (%+v) failed: %v", regs, err)) diff --git a/pkg/sentry/platform/ptrace/subprocess_amd64.go b/pkg/sentry/platform/ptrace/subprocess_amd64.go index 84b699f0d..020bbda79 100644 --- a/pkg/sentry/platform/ptrace/subprocess_amd64.go +++ b/pkg/sentry/platform/ptrace/subprocess_amd64.go @@ -201,7 +201,7 @@ func appendArchSeccompRules(rules []seccomp.RuleSet, defaultAction linux.BPFActi seccomp.RuleSet{ Rules: seccomp.SyscallRules{ syscall.SYS_ARCH_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(linux.ARCH_SET_CPUID), seccomp.AllowValue(0)}, + {seccomp.EqualTo(linux.ARCH_SET_CPUID), seccomp.EqualTo(0)}, }, }, Action: linux.SECCOMP_RET_ALLOW, diff --git a/pkg/sentry/platform/ptrace/subprocess_linux.go b/pkg/sentry/platform/ptrace/subprocess_linux.go index 2ce528601..8548853da 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux.go @@ -80,9 +80,9 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro Rules: seccomp.SyscallRules{ syscall.SYS_CLONE: []seccomp.Rule{ // Allow creation of new subprocesses (used by the master). - {seccomp.AllowValue(syscall.CLONE_FILES | syscall.SIGKILL)}, + {seccomp.EqualTo(syscall.CLONE_FILES | syscall.SIGKILL)}, // Allow creation of new threads within a single address space (used by addresss spaces). - {seccomp.AllowValue( + {seccomp.EqualTo( syscall.CLONE_FILES | syscall.CLONE_FS | syscall.CLONE_SIGHAND | @@ -97,14 +97,14 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro // For the stub prctl dance (all). syscall.SYS_PRCTL: []seccomp.Rule{ - {seccomp.AllowValue(syscall.PR_SET_PDEATHSIG), seccomp.AllowValue(syscall.SIGKILL)}, + {seccomp.EqualTo(syscall.PR_SET_PDEATHSIG), seccomp.EqualTo(syscall.SIGKILL)}, }, syscall.SYS_GETPPID: {}, // For the stub to stop itself (all). syscall.SYS_GETPID: {}, syscall.SYS_KILL: []seccomp.Rule{ - {seccomp.AllowAny{}, seccomp.AllowValue(syscall.SIGSTOP)}, + {seccomp.MatchAny{}, seccomp.EqualTo(syscall.SIGSTOP)}, }, // Injected to support the address space operations. @@ -115,7 +115,7 @@ func attachedThread(flags uintptr, defaultAction linux.BPFAction) (*thread, erro }) } rules = appendArchSeccompRules(rules, defaultAction) - instrs, err := seccomp.BuildProgram(rules, defaultAction) + instrs, err := seccomp.BuildProgram(rules, defaultAction, defaultAction) if err != nil { return nil, err } diff --git a/pkg/sentry/platform/ptrace/subprocess_linux_unsafe.go b/pkg/sentry/platform/ptrace/subprocess_linux_unsafe.go index 245b20722..533e45497 100644 --- a/pkg/sentry/platform/ptrace/subprocess_linux_unsafe.go +++ b/pkg/sentry/platform/ptrace/subprocess_linux_unsafe.go @@ -18,29 +18,12 @@ package ptrace import ( - "sync/atomic" "syscall" "unsafe" - "golang.org/x/sys/unix" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/sentry/hostcpu" - "gvisor.dev/gvisor/pkg/sync" ) -// 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 @@ -49,47 +32,3 @@ func unmaskAllSignals() syscall.Errno { _, _, errno := syscall.RawSyscall6(syscall.SYS_RT_SIGPROCMASK, linux.SIG_SETMASK, uintptr(unsafe.Pointer(&set)), 0, linux.SignalSetSize, 0, 0) return errno } - -// 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 := hostcpu.GetCPU() - - 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 index 0bee995e4..7ee20d89a 100644 --- a/pkg/sentry/platform/ptrace/subprocess_unsafe.go +++ b/pkg/sentry/platform/ptrace/subprocess_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.16 +// +build !go1.17 // Check go:linkname function signatures when updating Go version. diff --git a/pkg/sentry/platform/ring0/defs_amd64.go b/pkg/sentry/platform/ring0/defs_amd64.go index 9c6c2cf5c..00899273e 100644 --- a/pkg/sentry/platform/ring0/defs_amd64.go +++ b/pkg/sentry/platform/ring0/defs_amd64.go @@ -76,15 +76,41 @@ type KernelOpts struct { type KernelArchState struct { KernelOpts + // cpuEntries is array of kernelEntry for all cpus + cpuEntries []kernelEntry + // globalIDT is our set of interrupt gates. - globalIDT idt64 + globalIDT *idt64 } -// CPUArchState contains CPU-specific arch state. -type CPUArchState struct { +// kernelEntry contains minimal CPU-specific arch state +// that can be mapped at the upper of the address space. +// Malicious APP might steal info from it via CPU bugs. +type kernelEntry struct { // stack is the stack used for interrupts on this CPU. stack [256]byte + // scratch space for temporary usage. + scratch0 uint64 + + // stackTop is the top of the stack. + stackTop uint64 + + // cpuSelf is back reference to CPU. + cpuSelf *CPU + + // kernelCR3 is the cr3 used for sentry kernel. + kernelCR3 uintptr + + // gdt is the CPU's descriptor table. + gdt descriptorTable + + // tss is the CPU's task state. + tss TaskState64 +} + +// CPUArchState contains CPU-specific arch state. +type CPUArchState struct { // errorCode is the error code from the last exception. errorCode uintptr @@ -97,11 +123,7 @@ type CPUArchState struct { // exception. errorType uintptr - // gdt is the CPU's descriptor table. - gdt descriptorTable - - // tss is the CPU's task state. - tss TaskState64 + *kernelEntry } // ErrorCode returns the last error code. diff --git a/pkg/sentry/platform/ring0/defs_arm64.go b/pkg/sentry/platform/ring0/defs_arm64.go index 0e2ab716c..508236e46 100644 --- a/pkg/sentry/platform/ring0/defs_arm64.go +++ b/pkg/sentry/platform/ring0/defs_arm64.go @@ -77,6 +77,9 @@ type CPUArchState struct { // lazyVFP is the value of cpacr_el1. lazyVFP uintptr + + // appASID is the asid value of guest application. + appASID uintptr } // ErrorCode returns the last error code. diff --git a/pkg/sentry/platform/ring0/entry_amd64.go b/pkg/sentry/platform/ring0/entry_amd64.go index 7fa43c2f5..d87b1fd00 100644 --- a/pkg/sentry/platform/ring0/entry_amd64.go +++ b/pkg/sentry/platform/ring0/entry_amd64.go @@ -36,12 +36,15 @@ func sysenter() // This must be called prior to sysret/iret. func swapgs() +// jumpToKernel jumps to the kernel version of the current RIP. +func jumpToKernel() + // 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, *arch.Registers) Vector +func sysret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector // "iret is the cadillac of CPL switching." // @@ -50,7 +53,7 @@ func sysret(*CPU, *arch.Registers) Vector // 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, *arch.Registers) Vector +func iret(cpu *CPU, regs *arch.Registers, userCR3 uintptr) Vector // exception is the generic exception entry. // diff --git a/pkg/sentry/platform/ring0/entry_amd64.s b/pkg/sentry/platform/ring0/entry_amd64.s index 02df38331..f59747df3 100644 --- a/pkg/sentry/platform/ring0/entry_amd64.s +++ b/pkg/sentry/platform/ring0/entry_amd64.s @@ -63,6 +63,15 @@ MOVQ offset+PTRACE_RSI(reg), SI; \ MOVQ offset+PTRACE_RDI(reg), DI; +// WRITE_CR3() writes the given CR3 value. +// +// The code corresponds to: +// +// mov %rax, %cr3 +// +#define WRITE_CR3() \ + BYTE $0x0f; BYTE $0x22; BYTE $0xd8; + // SWAP_GS swaps the kernel GS (CPU). #define SWAP_GS() \ BYTE $0x0F; BYTE $0x01; BYTE $0xf8; @@ -75,15 +84,9 @@ #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; +#define LOAD_KERNEL_STACK(entry) \ + MOVQ ENTRY_STACK_TOP(entry), SP; // See kernel.go. TEXT ·Halt(SB),NOSPLIT,$0 @@ -95,58 +98,93 @@ TEXT ·swapgs(SB),NOSPLIT,$0 SWAP_GS() 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 + ORQ ·KernelStartAddress(SB), AX // Future return value. + MOVQ AX, 0(SP) + 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) + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX MOVQ SP, CPU_REGISTERS+PTRACE_RSP(BX) MOVQ BP, CPU_REGISTERS+PTRACE_RBP(BX) MOVQ AX, CPU_REGISTERS+PTRACE_RAX(BX) + // save SP AX userCR3 on the kernel stack. + MOVQ CPU_ENTRY(BX), BX + LOAD_KERNEL_STACK(BX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_RAX(AX) + PUSHQ CX + // 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). + + // restore userCR3, AX, SP. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. + POPQ SP // Restore SP. 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) + CALL ·jumpToKernel(SB) + // Save original state and stack. sysenter() or exception() + // from APP(gr3) will switch to this stack, set the return + // value (vector: 32(SP)) and then do RET, which will also + // automatically return to the lower half. + MOVQ cpu+0(FP), BX + MOVQ regs+8(FP), AX + MOVQ userCR3+16(FP), CX 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. + MOVQ CPU_ENTRY(BX), BX 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). + PUSHQ PTRACE_SS(AX) + PUSHQ PTRACE_RSP(AX) + PUSHQ PTRACE_FLAGS(AX) + PUSHQ PTRACE_CS(AX) + PUSHQ PTRACE_RIP(AX) + PUSHQ PTRACE_RAX(AX) // Save AX on kernel stack. + PUSHQ CX // Save userCR3 on kernel stack. + REGISTERS_LOAD(AX, 0) // Restore most registers. + POPQ AX // Get userCR3. + WRITE_CR3() // Switch to userCR3. + POPQ AX // Restore AX. 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 + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + PUSHQ CPU_REGISTERS+PTRACE_SS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RSP(AX) + PUSHQ CPU_REGISTERS+PTRACE_FLAGS(AX) + PUSHQ CPU_REGISTERS+PTRACE_CS(AX) + PUSHQ CPU_REGISTERS+PTRACE_RIP(AX) + REGISTERS_LOAD(AX, CPU_REGISTERS) + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), 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). @@ -155,53 +193,60 @@ TEXT ·Start(SB),NOSPLIT,$0 // 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 + // _RFLAGS_IOPL0 is always set in the user mode and it is never set in + // the kernel mode. See the comment of UserFlagsSet for more details. + TESTL $_RFLAGS_IOPL0, 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). + MOVQ AX, ENTRY_SCRATCH0(GS) // Save user AX on scratch. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user 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. + MOVQ SP, PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), CX // Load saved user AX value. + MOVQ CX, PTRACE_RAX(AX) // Save everything else. + MOVQ CX, PTRACE_ORIGRAX(AX) + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Get stacks. + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. // Return to the kernel, where the frame is: // - // vector (sp+24) + // vector (sp+32) + // userCR3 (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. + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ $Syscall, 32(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. + MOVQ AX, ENTRY_SCRATCH0(GS) + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + MOVQ CX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ R11, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ SP, CPU_REGISTERS+PTRACE_RSP(AX) + MOVQ ENTRY_SCRATCH0(GS), BX + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ $0, CPU_ERROR_CODE(AX) // Clear error code. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. // Call the syscall trampoline. LOAD_KERNEL_STACK(GS) - MOVQ CPU_SELF(GS), AX // Load vCPU. PUSHQ AX // First argument (vCPU). CALL ·kernelSyscall(SB) // Call the trampoline. POPQ AX // Pop vCPU. @@ -230,16 +275,21 @@ TEXT ·exception(SB),NOSPLIT,$0 // ERROR_CODE (sp+8) // VECTOR (sp+0) // - TESTL $_RFLAGS_IF, 32(SP) + TESTL $_RFLAGS_IOPL0, 32(SP) JZ kernel user: SWAP_GS() ADDQ $-8, SP // Adjust for flags. MOVQ $_KERNEL_FLAGS, 0(SP); BYTE $0x9d; // Reset flags (POPFQ). - XCHGQ CPU_REGISTERS+PTRACE_RAX(GS), AX // Swap for user regs. + PUSHQ AX // Save user AX on stack. + MOVQ ENTRY_KERNEL_CR3(GS), AX // Get kernel cr3 on AX. + WRITE_CR3() // Switch to kernel cr3. + + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + MOVQ CPU_REGISTERS+PTRACE_RAX(AX), AX // Get user regs. REGISTERS_SAVE(AX, 0) // Save all except IP, FLAGS, SP, AX. - MOVQ CPU_REGISTERS+PTRACE_RAX(GS), BX // Restore original AX. + POPQ BX // Restore original AX. MOVQ BX, PTRACE_RAX(AX) // Save it. MOVQ BX, PTRACE_ORIGRAX(AX) MOVQ 16(SP), BX; MOVQ BX, PTRACE_RIP(AX) @@ -249,34 +299,36 @@ user: MOVQ 48(SP), SI; MOVQ SI, PTRACE_SS(AX) // Copy out and return. + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. 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. + MOVQ CPU_REGISTERS+PTRACE_RSP(AX), SP // Original stack (kernel version). + MOVQ CPU_REGISTERS+PTRACE_RBP(AX), BP // Original base pointer. + MOVQ CX, CPU_ERROR_CODE(AX) // Set error code. + MOVQ $1, CPU_ERROR_TYPE(AX) // Set error type to user. + MOVQ BX, 32(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) + PUSHQ AX + MOVQ ENTRY_CPU_SELF(GS), AX // Load vCPU. + REGISTERS_SAVE(AX, CPU_REGISTERS) + POPQ BX + MOVQ BX, CPU_REGISTERS+PTRACE_RAX(AX) + MOVQ BX, CPU_REGISTERS+PTRACE_ORIGRAX(AX) + MOVQ 16(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RIP(AX) + MOVQ 32(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_FLAGS(AX) + MOVQ 40(SP), BX; MOVQ BX, CPU_REGISTERS+PTRACE_RSP(AX) // 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 8(SP), BX // Load the error code. + MOVQ BX, CPU_ERROR_CODE(AX) // Copy out to the CPU. + MOVQ $0, CPU_ERROR_TYPE(AX) // Set error type to kernel. MOVQ 0(SP), BX // BX contains the vector. - ADDQ $48, SP // Drop the exception frame. // Call the exception trampoline. LOAD_KERNEL_STACK(GS) - MOVQ CPU_SELF(GS), AX // Load vCPU. PUSHQ BX // Second argument (vector). PUSHQ AX // First argument (vCPU). CALL ·kernelException(SB) // Call the trampoline. diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s index 9d29b7168..1079a024b 100644 --- a/pkg/sentry/platform/ring0/entry_arm64.s +++ b/pkg/sentry/platform/ring0/entry_arm64.s @@ -27,7 +27,9 @@ // ERET returns using the ELR and SPSR for the current exception level. #define ERET() \ - WORD $0xd69f03e0 + WORD $0xd69f03e0; \ + DSB $7; \ + ISB $15; // RSV_REG is a register that holds el1 information temporarily. #define RSV_REG R18_PLATFORM @@ -44,9 +46,11 @@ #define SCTLR_M 1 << 0 #define SCTLR_C 1 << 2 #define SCTLR_I 1 << 12 +#define SCTLR_DZE 1 << 14 #define SCTLR_UCT 1 << 15 +#define SCTLR_UCI 1 << 26 -#define SCTLR_EL1_DEFAULT (SCTLR_M | SCTLR_C | SCTLR_I | SCTLR_UCT) +#define SCTLR_EL1_DEFAULT (SCTLR_M | SCTLR_C | SCTLR_I | SCTLR_UCT | SCTLR_UCI | SCTLR_DZE) // cntkctl_el1: counter-timer kernel control register el1. #define CNTKCTL_EL0PCTEN 1 << 0 @@ -294,23 +298,27 @@ LOAD_KERNEL_ADDRESS(CPU_SELF(from), RSV_REG); \ MOVD $CPU_STACK_TOP(RSV_REG), RSV_REG; \ MOVD RSV_REG, RSP; \ - WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - ISB $15; \ - DSB $15; + WORD $0xd538d092; //MRS TPIDR_EL1, R18 // SWITCH_TO_APP_PAGETABLE sets a new pagetable for a container application. #define SWITCH_TO_APP_PAGETABLE(from) \ - MOVD CPU_TTBR0_APP(from), RSV_REG; \ - WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 + MRS TTBR1_EL1, R0; \ + MOVD CPU_APP_ASID(from), R1; \ + BFI $48, R1, $16, R0; \ + MSR R0, TTBR1_EL1; \ // set the ASID in TTBR1_EL1 (since TCR.A1 is set) ISB $15; \ - DSB $15; + MOVD CPU_TTBR0_APP(from), RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; // SWITCH_TO_KVM_PAGETABLE sets the kvm pagetable. #define SWITCH_TO_KVM_PAGETABLE(from) \ - MOVD CPU_TTBR0_KVM(from), RSV_REG; \ - WORD $0xd5182012; \ // MSR R18, TTBR0_EL1 + MRS TTBR1_EL1, R0; \ + MOVD $1, R1; \ + BFI $48, R1, $16, R0; \ + MSR R0, TTBR1_EL1; \ ISB $15; \ - DSB $15; + MOVD CPU_TTBR0_KVM(from), RSV_REG; \ + MSR RSV_REG, TTBR0_EL1; #define VFP_ENABLE \ MOVD $FPEN_ENABLE, R0; \ @@ -326,29 +334,30 @@ #define KERNEL_ENTRY_FROM_EL0 \ SUB $16, RSP, RSP; \ // step1, save r18, r9 into kernel temporary stack. STP (RSV_REG, RSV_REG_APP), 16*0(RSP); \ - WORD $0xd538d092; \ //MRS TPIDR_EL1, R18, step2, switch user pagetable. - SWITCH_TO_KVM_PAGETABLE(RSV_REG); \ - WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 - MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP; \ // step3, load app context pointer. - REGISTERS_SAVE(RSV_REG_APP, 0); \ // step4, save app context. + WORD $0xd538d092; \ // MRS TPIDR_EL1, R18 + MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP; \ // step2, load app context pointer. + REGISTERS_SAVE(RSV_REG_APP, 0); \ // step3, save app context. MOVD RSV_REG_APP, R20; \ LDP 16*0(RSP), (RSV_REG, RSV_REG_APP); \ ADD $16, RSP, RSP; \ MOVD RSV_REG, PTRACE_R18(R20); \ MOVD RSV_REG_APP, PTRACE_R9(R20); \ - MOVD R20, RSV_REG_APP; \ + MRS TPIDR_EL0, R3; \ + MOVD R3, PTRACE_TLS(R20); \ WORD $0xd5384003; \ // MRS SPSR_EL1, R3 - MOVD R3, PTRACE_PSTATE(RSV_REG_APP); \ + MOVD R3, PTRACE_PSTATE(R20); \ MRS ELR_EL1, R3; \ - MOVD R3, PTRACE_PC(RSV_REG_APP); \ + MOVD R3, PTRACE_PC(R20); \ WORD $0xd5384103; \ // MRS SP_EL0, R3 - MOVD R3, PTRACE_SP(RSV_REG_APP); + MOVD R3, PTRACE_SP(R20); // KERNEL_ENTRY_FROM_EL1 is the entry code of the vcpu from el1 to el1. #define KERNEL_ENTRY_FROM_EL1 \ WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 REGISTERS_SAVE(RSV_REG, CPU_REGISTERS); \ // Save sentry context. MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG); \ + MRS TPIDR_EL0, R4; \ + MOVD R4, CPU_REGISTERS+PTRACE_TLS(RSV_REG); \ WORD $0xd5384004; \ // MRS SPSR_EL1, R4 MOVD R4, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG); \ MRS ELR_EL1, R4; \ @@ -357,6 +366,26 @@ MOVD R4, CPU_REGISTERS+PTRACE_SP(RSV_REG); \ LOAD_KERNEL_STACK(RSV_REG); // Load the temporary stack. +// EXCEPTION_WITH_ERROR is a common exception handler function. +#define EXCEPTION_WITH_ERROR(user, vector) \ + WORD $0xd538d092; \ //MRS TPIDR_EL1, R18 + WORD $0xd538601a; \ //MRS FAR_EL1, R26 + MOVD R26, CPU_FAULT_ADDR(RSV_REG); \ + MOVD $user, R3; \ + MOVD R3, CPU_ERROR_TYPE(RSV_REG); \ // Set error type to user. + MOVD $vector, R3; \ + MOVD R3, CPU_VECTOR_CODE(RSV_REG); \ + MRS ESR_EL1, R3; \ + MOVD R3, CPU_ERROR_CODE(RSV_REG); \ + B ·kernelExitToEl1(SB); + +// storeAppASID writes the application's asid value. +TEXT ·storeAppASID(SB),NOSPLIT,$0-8 + MOVD asid+0(FP), R1 + MRS TPIDR_EL1, RSV_REG + MOVD R1, CPU_APP_ASID(RSV_REG) + RET + // Halt halts execution. TEXT ·Halt(SB),NOSPLIT,$0 // Clear bluepill. @@ -365,8 +394,6 @@ TEXT ·Halt(SB),NOSPLIT,$0 BNE mmio_exit MOVD $0, CPU_REGISTERS+PTRACE_R9(RSV_REG) - // Flush dcache. - WORD $0xd5087e52 // DC CISW mmio_exit: // Disable fpsimd. WORD $0xd5381041 // MRS CPACR_EL1, R1 @@ -384,9 +411,6 @@ mmio_exit: MRS VBAR_EL1, R9 MOVD R0, 0x0(R9) - // Flush dcahce. - WORD $0xd5087e52 // DC CISW - RET // HaltAndResume halts execution and point the pointer to the resume function. @@ -414,7 +438,7 @@ TEXT ·Current(SB),NOSPLIT,$0-8 MOVD R8, ret+0(FP) RET -#define STACK_FRAME_SIZE 16 +#define STACK_FRAME_SIZE 32 // kernelExitToEl0 is the entrypoint for application in guest_el0. // Prepare the vcpu environment for container application. @@ -423,6 +447,8 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 MRS TPIDR_EL1, RSV_REG REGISTERS_SAVE(RSV_REG, CPU_REGISTERS) MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG) + MRS TPIDR_EL0, R3 + MOVD R3, CPU_REGISTERS+PTRACE_TLS(RSV_REG) WORD $0xd5384003 // MRS SPSR_EL1, R3 MOVD R3, CPU_REGISTERS+PTRACE_PSTATE(RSV_REG) @@ -449,8 +475,18 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 MOVD PTRACE_PSTATE(RSV_REG_APP), R1 WORD $0xd5184001 //MSR R1, SPSR_EL1 + // need use kernel space address to excute below code, since + // after SWITCH_TO_APP_PAGETABLE the ASID is changed to app's + // ASID. + WORD $0x10000061 // ADR R1, do_exit_to_el0 + ORR $0xffff000000000000, R1, R1 + JMP (R1) + +do_exit_to_el0: // RSV_REG & RSV_REG_APP will be loaded at the end. REGISTERS_LOAD(RSV_REG_APP, 0) + MOVD PTRACE_TLS(RSV_REG_APP), RSV_REG + MSR RSV_REG, TPIDR_EL0 // switch to user pagetable. MOVD PTRACE_R18(RSV_REG_APP), RSV_REG @@ -458,15 +494,16 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0 SUB $STACK_FRAME_SIZE, RSP, RSP STP (RSV_REG, RSV_REG_APP), 16*0(RSP) + STP (R0, R1), 16*1(RSP) WORD $0xd538d092 //MRS TPIDR_EL1, R18 SWITCH_TO_APP_PAGETABLE(RSV_REG) + LDP 16*1(RSP), (R0, R1) LDP 16*0(RSP), (RSV_REG, RSV_REG_APP) ADD $STACK_FRAME_SIZE, RSP, RSP - ISB $15 ERET() // kernelExitToEl1 is the entrypoint for sentry in guest_el1. @@ -482,6 +519,9 @@ TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R1 MOVD R1, RSP + SWITCH_TO_KVM_PAGETABLE(RSV_REG) + MRS TPIDR_EL1, RSV_REG + REGISTERS_LOAD(RSV_REG, CPU_REGISTERS) MOVD CPU_REGISTERS+PTRACE_R9(RSV_REG), RSV_REG_APP @@ -489,8 +529,6 @@ TEXT ·kernelExitToEl1(SB),NOSPLIT,$0 // Start is the CPU entrypoint. TEXT ·Start(SB),NOSPLIT,$0 - // Flush dcache. - WORD $0xd5087e52 // DC CISW // Init. MOVD $SCTLR_EL1_DEFAULT, R1 MSR R1, SCTLR_EL1 @@ -634,21 +672,7 @@ el0_svc: el0_da: el0_ia: - WORD $0xd538d092 //MRS TPIDR_EL1, R18 - WORD $0xd538601a //MRS FAR_EL1, R26 - - MOVD R26, CPU_FAULT_ADDR(RSV_REG) - - MOVD $1, R3 - MOVD R3, CPU_ERROR_TYPE(RSV_REG) // Set error type to user. - - MOVD $PageFault, R3 - MOVD R3, CPU_VECTOR_CODE(RSV_REG) - - MRS ESR_EL1, R3 - MOVD R3, CPU_ERROR_CODE(RSV_REG) - - B ·kernelExitToEl1(SB) + EXCEPTION_WITH_ERROR(1, PageFault) el0_fpsimd_acc: B ·Shutdown(SB) @@ -663,10 +687,7 @@ el0_sp_pc: B ·Shutdown(SB) el0_undef: - MOVD $El0Sync_undef, R3 - MOVD R3, CPU_VECTOR_CODE(RSV_REG) - - B ·kernelExitToEl1(SB) + EXCEPTION_WITH_ERROR(1, El0Sync_undef) el0_dbg: B ·Shutdown(SB) diff --git a/pkg/sentry/platform/ring0/gen_offsets/BUILD b/pkg/sentry/platform/ring0/gen_offsets/BUILD index 549f3d228..9742308d8 100644 --- a/pkg/sentry/platform/ring0/gen_offsets/BUILD +++ b/pkg/sentry/platform/ring0/gen_offsets/BUILD @@ -24,7 +24,10 @@ go_binary( "defs_impl_arm64.go", "main.go", ], - visibility = ["//pkg/sentry/platform/ring0:__pkg__"], + visibility = [ + "//pkg/sentry/platform/kvm:__pkg__", + "//pkg/sentry/platform/ring0:__pkg__", + ], deps = [ "//pkg/cpuid", "//pkg/sentry/arch", diff --git a/pkg/sentry/platform/ring0/kernel.go b/pkg/sentry/platform/ring0/kernel.go index 021693791..264be23d3 100644 --- a/pkg/sentry/platform/ring0/kernel.go +++ b/pkg/sentry/platform/ring0/kernel.go @@ -19,8 +19,8 @@ package ring0 // N.B. that constraints on KernelOpts must be satisfied. // //go:nosplit -func (k *Kernel) Init(opts KernelOpts) { - k.init(opts) +func (k *Kernel) Init(opts KernelOpts, maxCPUs int) { + k.init(opts, maxCPUs) } // Halt halts execution. @@ -49,6 +49,11 @@ func (defaultHooks) KernelException(Vector) { // kernelSyscall is a trampoline. // +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// // +checkescape:hard,stack // //go:nosplit @@ -58,6 +63,11 @@ func kernelSyscall(c *CPU) { // kernelException is a trampoline. // +// When in amd64, it is called with %rip on the upper half, so it can +// NOT access to any global data which is not mapped on upper and must +// call to function pointers or interfaces to switch to the lower half +// so that callee can access to global data. +// // +checkescape:hard,stack // //go:nosplit @@ -68,10 +78,10 @@ func kernelException(c *CPU, vector Vector) { // Init initializes a new CPU. // // Init allows embedding in other objects. -func (c *CPU) Init(k *Kernel, hooks Hooks) { - c.self = c // Set self reference. - c.kernel = k // Set kernel reference. - c.init() // Perform architectural init. +func (c *CPU) Init(k *Kernel, cpuID int, hooks Hooks) { + c.self = c // Set self reference. + c.kernel = k // Set kernel reference. + c.init(cpuID) // Perform architectural init. // Require hooks. if hooks != nil { diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go index d37981dbf..3a9dff4cc 100644 --- a/pkg/sentry/platform/ring0/kernel_amd64.go +++ b/pkg/sentry/platform/ring0/kernel_amd64.go @@ -18,13 +18,42 @@ package ring0 import ( "encoding/binary" + "reflect" + + "gvisor.dev/gvisor/pkg/usermem" ) // init initializes architecture-specific state. -func (k *Kernel) init(opts KernelOpts) { +func (k *Kernel) init(opts KernelOpts, maxCPUs int) { // Save the root page tables. k.PageTables = opts.PageTables + entrySize := reflect.TypeOf(kernelEntry{}).Size() + var ( + entries []kernelEntry + padding = 1 + ) + for { + entries = make([]kernelEntry, maxCPUs+padding-1) + totalSize := entrySize * uintptr(maxCPUs+padding-1) + addr := reflect.ValueOf(&entries[0]).Pointer() + if addr&(usermem.PageSize-1) == 0 && totalSize >= usermem.PageSize { + // The runtime forces power-of-2 alignment for allocations, and we are therefore + // safe once the first address is aligned and the chunk is at least a full page. + break + } + padding = padding << 1 + } + k.cpuEntries = entries + + k.globalIDT = &idt64{} + if reflect.TypeOf(idt64{}).Size() != usermem.PageSize { + panic("Size of globalIDT should be PageSize") + } + if reflect.ValueOf(k.globalIDT).Pointer()&(usermem.PageSize-1) != 0 { + panic("Allocated globalIDT should be page aligned") + } + // Setup the IDT, which is uniform. for v, handler := range handlers { // Allow Breakpoint and Overflow to be called from all @@ -39,8 +68,26 @@ func (k *Kernel) init(opts KernelOpts) { } } +func (k *Kernel) EntryRegions() map[uintptr]uintptr { + regions := make(map[uintptr]uintptr) + + addr := reflect.ValueOf(&k.cpuEntries[0]).Pointer() + size := reflect.TypeOf(kernelEntry{}).Size() * uintptr(len(k.cpuEntries)) + end, _ := usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + addr = reflect.ValueOf(k.globalIDT).Pointer() + size = reflect.TypeOf(idt64{}).Size() + end, _ = usermem.Addr(addr + size).RoundUp() + regions[uintptr(usermem.Addr(addr).RoundDown())] = uintptr(end) + + return regions +} + // init initializes architecture-specific state. -func (c *CPU) init() { +func (c *CPU) init(cpuID int) { + c.kernelEntry = &c.kernel.cpuEntries[cpuID] + c.cpuSelf = c // Null segment. c.gdt[0].setNull() @@ -65,6 +112,7 @@ func (c *CPU) init() { // Set the kernel stack pointer in the TSS (virtual address). stackAddr := c.StackTop() + c.stackTop = stackAddr c.tss.rsp0Lo = uint32(stackAddr) c.tss.rsp0Hi = uint32(stackAddr >> 32) c.tss.ist1Lo = uint32(stackAddr) @@ -183,7 +231,7 @@ func IsCanonical(addr uint64) bool { //go:nosplit func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { userCR3 := switchOpts.PageTables.CR3(!switchOpts.Flush, switchOpts.UserPCID) - kernelCR3 := c.kernel.PageTables.CR3(true, switchOpts.KernelPCID) + c.kernelCR3 = uintptr(c.kernel.PageTables.CR3(true, switchOpts.KernelPCID)) // Sanitize registers. regs := switchOpts.Registers @@ -197,15 +245,11 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { WriteFS(uintptr(regs.Fs_base)) // escapes: no. Set application FS. WriteGS(uintptr(regs.Gs_base)) // escapes: no. Set application GS. LoadFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy in floating point. - jumpToKernel() // Switch to upper half. - writeCR3(uintptr(userCR3)) // Change to user address space. if switchOpts.FullRestore { - vector = iret(c, regs) + vector = iret(c, regs, uintptr(userCR3)) } else { - vector = sysret(c, regs) + vector = sysret(c, regs, uintptr(userCR3)) } - writeCR3(uintptr(kernelCR3)) // Return to kernel address space. - jumpToUser() // Return to lower half. SaveFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy out floating point. WriteFS(uintptr(c.registers.Fs_base)) // escapes: no. Restore kernel FS. return @@ -219,7 +263,7 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { //go:nosplit func start(c *CPU) { // Save per-cpu & FS segment. - WriteGS(kernelAddr(c)) + WriteGS(kernelAddr(c.kernelEntry)) WriteFS(uintptr(c.registers.Fs_base)) // Initialize floating point. diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go index d0afa1aaa..b294ccc7c 100644 --- a/pkg/sentry/platform/ring0/kernel_arm64.go +++ b/pkg/sentry/platform/ring0/kernel_arm64.go @@ -25,13 +25,13 @@ func HaltAndResume() func HaltEl1SvcAndResume() // init initializes architecture-specific state. -func (k *Kernel) init(opts KernelOpts) { +func (k *Kernel) init(opts KernelOpts, maxCPUs int) { // Save the root page tables. k.PageTables = opts.PageTables } // init initializes architecture-specific state. -func (c *CPU) init() { +func (c *CPU) init(cpuID int) { // Set the kernel stack pointer(virtual address). c.registers.Sp = uint64(c.StackTop()) @@ -53,17 +53,20 @@ func IsCanonical(addr uint64) bool { //go:nosplit func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) { + storeAppASID(uintptr(switchOpts.UserASID)) + if switchOpts.Flush { + FlushTlbAll() + } + regs := switchOpts.Registers regs.Pstate &= ^uint64(PsrFlagsClear) regs.Pstate |= UserFlagsSet LoadFloatingPoint(switchOpts.FloatingPointState) - SetTLS(regs.TPIDR_EL0) kernelExitToEl0() - regs.TPIDR_EL0 = GetTLS() SaveFloatingPoint(switchOpts.FloatingPointState) vector = c.vecCode diff --git a/pkg/sentry/platform/ring0/lib_amd64.go b/pkg/sentry/platform/ring0/lib_amd64.go index ca968a036..0ec5c3bc5 100644 --- a/pkg/sentry/platform/ring0/lib_amd64.go +++ b/pkg/sentry/platform/ring0/lib_amd64.go @@ -61,21 +61,9 @@ func wrgsbase(addr uintptr) // wrgsmsr writes to the GS_BASE MSR. func wrgsmsr(addr uintptr) -// writeCR3 writes the CR3 value. -func writeCR3(phys uintptr) - -// readCR3 reads the current CR3 value. -func readCR3() 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() diff --git a/pkg/sentry/platform/ring0/lib_amd64.s b/pkg/sentry/platform/ring0/lib_amd64.s index 75d742750..2fe83568a 100644 --- a/pkg/sentry/platform/ring0/lib_amd64.s +++ b/pkg/sentry/platform/ring0/lib_amd64.s @@ -127,53 +127,6 @@ TEXT ·wrgsmsr(SB),NOSPLIT,$0-8 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: diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go index 00e52c8af..d91a09de1 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.go +++ b/pkg/sentry/platform/ring0/lib_arm64.go @@ -16,6 +16,15 @@ package ring0 +// storeAppASID writes the application's asid value. +func storeAppASID(asid uintptr) + +// LocalFlushTlbAll same as FlushTlbAll, but only applies to the calling CPU. +func LocalFlushTlbAll() + +// FlushTlbAll flush all tlb. +func FlushTlbAll() + // CPACREL1 returns the value of the CPACR_EL1 register. func CPACREL1() (value uintptr) @@ -44,12 +53,6 @@ func LoadFloatingPoint(*byte) // SaveFloatingPoint saves floating point state. func SaveFloatingPoint(*byte) -// GetTLS returns the value of TPIDR_EL0 register. -func GetTLS() (value uint64) - -// SetTLS writes the TPIDR_EL0 value. -func SetTLS(value uint64) - // Init sets function pointers based on architectural features. // // This must be called prior to using ring0. diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s index 86bfbe46f..da9d3cf55 100644 --- a/pkg/sentry/platform/ring0/lib_arm64.s +++ b/pkg/sentry/platform/ring0/lib_arm64.s @@ -15,14 +15,18 @@ #include "funcdata.h" #include "textflag.h" -TEXT ·GetTLS(SB),NOSPLIT,$0-8 - MRS TPIDR_EL0, R1 - MOVD R1, ret+0(FP) +TEXT ·LocalFlushTlbAll(SB),NOSPLIT,$0 + DSB $6 // dsb(nshst) + WORD $0xd508871f // __tlbi(vmalle1) + DSB $7 // dsb(nsh) + ISB $15 RET -TEXT ·SetTLS(SB),NOSPLIT,$0-8 - MOVD addr+0(FP), R1 - MSR R1, TPIDR_EL0 +TEXT ·FlushTlbAll(SB),NOSPLIT,$0 + DSB $10 // dsb(ishst) + WORD $0xd508831f // __tlbi(vmalle1is) + DSB $11 // dsb(ish) + ISB $15 RET TEXT ·CPACREL1(SB),NOSPLIT,$0-8 diff --git a/pkg/sentry/platform/ring0/offsets_amd64.go b/pkg/sentry/platform/ring0/offsets_amd64.go index b8ab120a0..ca4075b09 100644 --- a/pkg/sentry/platform/ring0/offsets_amd64.go +++ b/pkg/sentry/platform/ring0/offsets_amd64.go @@ -30,14 +30,21 @@ func Emit(w io.Writer) { 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_ENTRY 0x%02x\n", reflect.ValueOf(&c.kernelEntry).Pointer()-reflect.ValueOf(c).Pointer()) + + e := &kernelEntry{} + fmt.Fprintf(w, "\n// CPU entry offsets.\n") + fmt.Fprintf(w, "#define ENTRY_SCRATCH0 0x%02x\n", reflect.ValueOf(&e.scratch0).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_STACK_TOP 0x%02x\n", reflect.ValueOf(&e.stackTop).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_CPU_SELF 0x%02x\n", reflect.ValueOf(&e.cpuSelf).Pointer()-reflect.ValueOf(e).Pointer()) + fmt.Fprintf(w, "#define ENTRY_KERNEL_CR3 0x%02x\n", reflect.ValueOf(&e.kernelCR3).Pointer()-reflect.ValueOf(e).Pointer()) fmt.Fprintf(w, "\n// Bits.\n") fmt.Fprintf(w, "#define _RFLAGS_IF 0x%02x\n", _RFLAGS_IF) + fmt.Fprintf(w, "#define _RFLAGS_IOPL0 0x%02x\n", _RFLAGS_IOPL0) fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) fmt.Fprintf(w, "\n// Vectors.\n") diff --git a/pkg/sentry/platform/ring0/offsets_arm64.go b/pkg/sentry/platform/ring0/offsets_arm64.go index f3de962f0..45eba960d 100644 --- a/pkg/sentry/platform/ring0/offsets_arm64.go +++ b/pkg/sentry/platform/ring0/offsets_arm64.go @@ -41,6 +41,7 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define CPU_VECTOR_CODE 0x%02x\n", reflect.ValueOf(&c.vecCode).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_APP_ADDR 0x%02x\n", reflect.ValueOf(&c.appAddr).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "#define CPU_LAZY_VFP 0x%02x\n", reflect.ValueOf(&c.lazyVFP).Pointer()-reflect.ValueOf(c).Pointer()) + fmt.Fprintf(w, "#define CPU_APP_ASID 0x%02x\n", reflect.ValueOf(&c.appASID).Pointer()-reflect.ValueOf(c).Pointer()) fmt.Fprintf(w, "\n// Bits.\n") fmt.Fprintf(w, "#define _KERNEL_FLAGS 0x%02x\n", KernelFlagsSet) @@ -124,4 +125,5 @@ func Emit(w io.Writer) { fmt.Fprintf(w, "#define PTRACE_SP 0x%02x\n", reflect.ValueOf(&p.Sp).Pointer()-reflect.ValueOf(p).Pointer()) fmt.Fprintf(w, "#define PTRACE_PC 0x%02x\n", reflect.ValueOf(&p.Pc).Pointer()-reflect.ValueOf(p).Pointer()) fmt.Fprintf(w, "#define PTRACE_PSTATE 0x%02x\n", reflect.ValueOf(&p.Pstate).Pointer()-reflect.ValueOf(p).Pointer()) + fmt.Fprintf(w, "#define PTRACE_TLS 0x%02x\n", reflect.ValueOf(&p.TPIDR_EL0).Pointer()-reflect.ValueOf(p).Pointer()) } diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go b/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go index 6409d1d91..520161755 100644 --- a/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go +++ b/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go @@ -78,7 +78,7 @@ const ( const ( executeDisable = xn - optionMask = 0xfff | 0xfff<<48 + optionMask = 0xfff | 0xffff<<48 protDefault = accessed | shared ) @@ -188,7 +188,7 @@ func (p *PTE) Set(addr uintptr, opts MapOpts) { v |= mtNormal } else { v = v &^ user - v |= mtDevicenGnRE // Strong order for the addresses with ring0.KernelStartAddress. + v |= mtNormal } atomic.StoreUintptr((*uintptr)(p), v) } diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_arm64.go b/pkg/sentry/platform/ring0/pagetables/pagetables_arm64.go index 1a49f12a2..5ddd10256 100644 --- a/pkg/sentry/platform/ring0/pagetables/pagetables_arm64.go +++ b/pkg/sentry/platform/ring0/pagetables/pagetables_arm64.go @@ -36,7 +36,7 @@ const ( pudSize = 1 << pudShift pgdSize = 1 << pgdShift - ttbrASIDOffset = 55 + ttbrASIDOffset = 48 ttbrASIDMask = 0xff entriesPerPage = 512 diff --git a/pkg/sentry/platform/ring0/x86.go b/pkg/sentry/platform/ring0/x86.go index 9da0ea685..34fbc1c35 100644 --- a/pkg/sentry/platform/ring0/x86.go +++ b/pkg/sentry/platform/ring0/x86.go @@ -39,7 +39,9 @@ const ( _RFLAGS_AC = 1 << 18 _RFLAGS_NT = 1 << 14 - _RFLAGS_IOPL = 3 << 12 + _RFLAGS_IOPL0 = 1 << 12 + _RFLAGS_IOPL1 = 1 << 13 + _RFLAGS_IOPL = _RFLAGS_IOPL0 | _RFLAGS_IOPL1 _RFLAGS_DF = 1 << 10 _RFLAGS_IF = 1 << 9 _RFLAGS_STEP = 1 << 8 @@ -67,15 +69,45 @@ const ( KernelFlagsSet = _RFLAGS_RESERVED // UserFlagsSet are always set in userspace. - UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF + // + // _RFLAGS_IOPL is a set of two bits and it shows the I/O privilege + // level. The Current Privilege Level (CPL) of the task must be less + // than or equal to the IOPL in order for the task or program to access + // I/O ports. + // + // Here, _RFLAGS_IOPL0 is used only to determine whether the task is + // running in the kernel or userspace mode. In the user mode, the CPL is + // always 3 and it doesn't matter what IOPL is set if it is bellow CPL. + // + // We need to have one bit which will be always different in user and + // kernel modes. And we have to remember that even though we have + // KernelFlagsClear, we still can see some of these flags in the kernel + // mode. This can happen when the goruntime switches on a goroutine + // which has been saved in the host mode. On restore, the popf + // instruction is used to restore flags and this means that all flags + // what the goroutine has in the host mode will be restored in the + // kernel mode. + // + // _RFLAGS_IOPL0 is never set in host and kernel modes and we always set + // it in the user mode. So if this flag is set, the task is running in + // the user mode and if it isn't set, the task is running in the kernel + // mode. + UserFlagsSet = _RFLAGS_RESERVED | _RFLAGS_IF | _RFLAGS_IOPL0 // KernelFlagsClear should always be clear in the kernel. KernelFlagsClear = _RFLAGS_STEP | _RFLAGS_IF | _RFLAGS_IOPL | _RFLAGS_AC | _RFLAGS_NT // UserFlagsClear are always cleared in userspace. - UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL + UserFlagsClear = _RFLAGS_NT | _RFLAGS_IOPL1 ) +// IsKernelFlags returns true if rflags coresponds to the kernel mode. +// +// go:nosplit +func IsKernelFlags(rflags uint64) bool { + return rflags&_RFLAGS_IOPL0 == 0 +} + // Vector is an exception vector. type Vector uintptr @@ -104,7 +136,7 @@ const ( VirtualizationException SecurityException = 0x1e SyscallInt80 = 0x80 - _NR_INTERRUPTS = SyscallInt80 + 1 + _NR_INTERRUPTS = 0x100 ) // System call vectors. diff --git a/pkg/sentry/socket/BUILD b/pkg/sentry/socket/BUILD index c0fd3425b..a3f775d15 100644 --- a/pkg/sentry/socket/BUILD +++ b/pkg/sentry/socket/BUILD @@ -10,6 +10,7 @@ go_library( "//pkg/abi/linux", "//pkg/binary", "//pkg/context", + "//pkg/marshal", "//pkg/sentry/device", "//pkg/sentry/fs", "//pkg/sentry/fs/fsutil", @@ -20,6 +21,5 @@ go_library( "//pkg/syserr", "//pkg/tcpip", "//pkg/usermem", - "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/socket/control/control_vfs2.go b/pkg/sentry/socket/control/control_vfs2.go index d9621968c..37d02948f 100644 --- a/pkg/sentry/socket/control/control_vfs2.go +++ b/pkg/sentry/socket/control/control_vfs2.go @@ -24,6 +24,8 @@ import ( ) // SCMRightsVFS2 represents a SCM_RIGHTS socket control message. +// +// +stateify savable type SCMRightsVFS2 interface { transport.RightsControlMessage @@ -34,9 +36,11 @@ type SCMRightsVFS2 interface { Files(ctx context.Context, max int) (rf RightsFilesVFS2, truncated bool) } -// RightsFiles represents a SCM_RIGHTS socket control message. A reference is -// maintained for each vfs.FileDescription and is release either when an FD is created or -// when the Release method is called. +// RightsFilesVFS2 represents a SCM_RIGHTS socket control message. A reference +// is maintained for each vfs.FileDescription and is release either when an FD +// is created or when the Release method is called. +// +// +stateify savable type RightsFilesVFS2 []*vfs.FileDescription // NewSCMRightsVFS2 creates a new SCM_RIGHTS socket control message diff --git a/pkg/sentry/socket/hostinet/BUILD b/pkg/sentry/socket/hostinet/BUILD index e76e498de..b6ebe29d6 100644 --- a/pkg/sentry/socket/hostinet/BUILD +++ b/pkg/sentry/socket/hostinet/BUILD @@ -21,6 +21,8 @@ go_library( "//pkg/context", "//pkg/fdnotifier", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", @@ -37,11 +39,12 @@ go_library( "//pkg/sentry/vfs", "//pkg/syserr", "//pkg/syserror", + "//pkg/tcpip", + "//pkg/tcpip/network/ipv4", + "//pkg/tcpip/network/ipv6", "//pkg/tcpip/stack", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", - "//tools/go_marshal/primitive", "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/socket/hostinet/socket.go b/pkg/sentry/socket/hostinet/socket.go index 242e6bf76..7d3c4a01c 100644 --- a/pkg/sentry/socket/hostinet/socket.go +++ b/pkg/sentry/socket/hostinet/socket.go @@ -24,6 +24,8 @@ import ( "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fdnotifier" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -36,8 +38,6 @@ import ( "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) const ( diff --git a/pkg/sentry/socket/hostinet/socket_vfs2.go b/pkg/sentry/socket/hostinet/socket_vfs2.go index 8a1d52ebf..9a2cac40b 100644 --- a/pkg/sentry/socket/hostinet/socket_vfs2.go +++ b/pkg/sentry/socket/hostinet/socket_vfs2.go @@ -33,6 +33,7 @@ import ( "gvisor.dev/gvisor/pkg/waiter" ) +// +stateify savable type socketVFS2 struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -51,7 +52,8 @@ var _ = socket.SocketVFS2(&socketVFS2{}) func newVFS2Socket(t *kernel.Task, family int, stype linux.SockType, protocol int, fd int, flags uint32) (*vfs.FileDescription, *syserr.Error) { mnt := t.Kernel().SocketMount() - d := sockfs.NewDentry(t.Credentials(), mnt) + d := sockfs.NewDentry(t, mnt) + defer d.DecRef(t) s := &socketVFS2{ socketOpsCommon: socketOpsCommon{ @@ -77,6 +79,13 @@ func newVFS2Socket(t *kernel.Task, family int, stype linux.SockType, protocol in return vfsfd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *socketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *socketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) @@ -97,11 +106,6 @@ func (s *socketVFS2) Ioctl(ctx context.Context, uio usermem.IO, args arch.Syscal return ioctl(ctx, s.fd, uio, args) } -// Allocate implements vfs.FileDescriptionImpl.Allocate. -func (s *socketVFS2) Allocate(ctx context.Context, mode, offset, length uint64) error { - return syserror.ENODEV -} - // PRead implements vfs.FileDescriptionImpl.PRead. func (s *socketVFS2) PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts vfs.ReadOptions) (int64, error) { return 0, syserror.ESPIPE diff --git a/pkg/sentry/socket/hostinet/stack.go b/pkg/sentry/socket/hostinet/stack.go index fda3dcb35..7e7857ac3 100644 --- a/pkg/sentry/socket/hostinet/stack.go +++ b/pkg/sentry/socket/hostinet/stack.go @@ -30,6 +30,9 @@ import ( "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv4" + "gvisor.dev/gvisor/pkg/tcpip/network/ipv6" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -59,6 +62,8 @@ type Stack struct { tcpSACKEnabled bool netDevFile *os.File netSNMPFile *os.File + ipv4Forwarding bool + ipv6Forwarding bool } // NewStack returns an empty Stack containing no configuration. @@ -118,6 +123,13 @@ func (s *Stack) Configure() error { s.netSNMPFile = f } + s.ipv6Forwarding = false + if ipForwarding, err := ioutil.ReadFile("/proc/sys/net/ipv6/conf/all/forwarding"); err == nil { + s.ipv6Forwarding = strings.TrimSpace(string(ipForwarding)) != "0" + } else { + log.Warningf("Failed to read if ipv6 forwarding is enabled, setting to false") + } + return nil } @@ -312,7 +324,12 @@ func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr { } // AddInterfaceAddr implements inet.Stack.AddInterfaceAddr. -func (s *Stack) AddInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { +func (s *Stack) AddInterfaceAddr(int32, inet.InterfaceAddr) error { + return syserror.EACCES +} + +// RemoveInterfaceAddr implements inet.Stack.RemoveInterfaceAddr. +func (s *Stack) RemoveInterfaceAddr(int32, inet.InterfaceAddr) error { return syserror.EACCES } @@ -347,7 +364,7 @@ func (s *Stack) TCPSACKEnabled() (bool, error) { } // SetTCPSACKEnabled implements inet.Stack.SetTCPSACKEnabled. -func (s *Stack) SetTCPSACKEnabled(enabled bool) error { +func (s *Stack) SetTCPSACKEnabled(bool) error { return syserror.EACCES } @@ -357,7 +374,7 @@ func (s *Stack) TCPRecovery() (inet.TCPLossRecovery, error) { } // SetTCPRecovery implements inet.Stack.SetTCPRecovery. -func (s *Stack) SetTCPRecovery(recovery inet.TCPLossRecovery) error { +func (s *Stack) SetTCPRecovery(inet.TCPLossRecovery) error { return syserror.EACCES } @@ -418,18 +435,18 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { } if rawLine == "" { - return fmt.Errorf("Failed to get raw line") + return fmt.Errorf("failed to get raw line") } parts := strings.SplitN(rawLine, ":", 2) if len(parts) != 2 { - return fmt.Errorf("Failed to get prefix from: %q", rawLine) + return fmt.Errorf("failed to get prefix from: %q", rawLine) } sliceStat = toSlice(stat) fields := strings.Fields(strings.TrimSpace(parts[1])) if len(fields) != len(sliceStat) { - return fmt.Errorf("Failed to parse fields: %q", rawLine) + return fmt.Errorf("failed to parse fields: %q", rawLine) } if _, ok := stat.(*inet.StatSNMPTCP); ok { snmpTCP = true @@ -445,7 +462,7 @@ func (s *Stack) Statistics(stat interface{}, arg string) error { sliceStat[i], err = strconv.ParseUint(fields[i], 10, 64) } if err != nil { - return fmt.Errorf("Failed to parse field %d from: %q, %v", i, rawLine, err) + return fmt.Errorf("failed to parse field %d from: %q, %v", i, rawLine, err) } } @@ -468,3 +485,21 @@ func (s *Stack) CleanupEndpoints() []stack.TransportEndpoint { return nil } // RestoreCleanupEndpoints implements inet.Stack.RestoreCleanupEndpoints. func (s *Stack) RestoreCleanupEndpoints([]stack.TransportEndpoint) {} + +// Forwarding implements inet.Stack.Forwarding. +func (s *Stack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + switch protocol { + case ipv4.ProtocolNumber: + return s.ipv4Forwarding + case ipv6.ProtocolNumber: + return s.ipv6Forwarding + default: + log.Warningf("Forwarding(%v) failed: unsupported protocol", protocol) + return false + } +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *Stack) SetForwarding(tcpip.NetworkProtocolNumber, bool) error { + return syserror.EACCES +} diff --git a/pkg/sentry/socket/netfilter/BUILD b/pkg/sentry/socket/netfilter/BUILD index 721094bbf..8aea0200f 100644 --- a/pkg/sentry/socket/netfilter/BUILD +++ b/pkg/sentry/socket/netfilter/BUILD @@ -6,6 +6,8 @@ go_library( name = "netfilter", srcs = [ "extensions.go", + "ipv4.go", + "ipv6.go", "netfilter.go", "owner_matcher.go", "targets.go", diff --git a/pkg/sentry/socket/netfilter/extensions.go b/pkg/sentry/socket/netfilter/extensions.go index 0336a32d8..549787955 100644 --- a/pkg/sentry/socket/netfilter/extensions.go +++ b/pkg/sentry/socket/netfilter/extensions.go @@ -19,6 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,7 +39,7 @@ type matchMaker interface { // name is the matcher name as stored in the xt_entry_match struct. name() string - // marshal converts from an stack.Matcher to an ABI struct. + // marshal converts from a stack.Matcher to an ABI struct. marshal(matcher stack.Matcher) []byte // unmarshal converts from the ABI matcher struct to an @@ -93,3 +95,71 @@ func unmarshalMatcher(match linux.XTEntryMatch, filter stack.IPHeaderFilter, buf } return matchMaker.unmarshal(buf, filter) } + +// targetMaker knows how to (un)marshal a target. Once registered, +// marshalTarget and unmarshalTarget can be used. +type targetMaker interface { + // id uniquely identifies the target. + id() stack.TargetID + + // marshal converts from a stack.Target to an ABI struct. + marshal(target stack.Target) []byte + + // unmarshal converts from the ABI matcher struct to a stack.Target. + unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) +} + +// targetMakers maps the TargetID of supported targets to the targetMaker that +// marshals and unmarshals it. It is immutable after package initialization. +var targetMakers = map[stack.TargetID]targetMaker{} + +func targetRevision(name string, netProto tcpip.NetworkProtocolNumber, rev uint8) (uint8, bool) { + tid := stack.TargetID{ + Name: name, + NetworkProtocol: netProto, + Revision: rev, + } + if _, ok := targetMakers[tid]; !ok { + return 0, false + } + + // Return the highest supported revision unless rev is higher. + for _, other := range targetMakers { + otherID := other.id() + if name == otherID.Name && netProto == otherID.NetworkProtocol && otherID.Revision > rev { + rev = uint8(otherID.Revision) + } + } + return rev, true +} + +// registerTargetMaker should be called by target extensions to register them +// with the netfilter package. +func registerTargetMaker(tm targetMaker) { + if _, ok := targetMakers[tm.id()]; ok { + panic(fmt.Sprintf("multiple targets registered with name %q.", tm.id())) + } + targetMakers[tm.id()] = tm +} + +func marshalTarget(target stack.Target) []byte { + targetMaker, ok := targetMakers[target.ID()] + if !ok { + panic(fmt.Sprintf("unknown target of type %T with id %+v.", target, target.ID())) + } + return targetMaker.marshal(target) +} + +func unmarshalTarget(target linux.XTEntryTarget, filter stack.IPHeaderFilter, buf []byte) (stack.Target, *syserr.Error) { + tid := stack.TargetID{ + Name: target.Name.String(), + NetworkProtocol: filter.NetworkProtocol(), + Revision: target.Revision, + } + targetMaker, ok := targetMakers[tid] + if !ok { + nflog("unsupported target with name %q", target.Name.String()) + return nil, syserr.ErrInvalidArgument + } + return targetMaker.unmarshal(buf, filter) +} diff --git a/pkg/sentry/socket/netfilter/ipv4.go b/pkg/sentry/socket/netfilter/ipv4.go new file mode 100644 index 000000000..b560fae0d --- /dev/null +++ b/pkg/sentry/socket/netfilter/ipv4.go @@ -0,0 +1,265 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "bytes" + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/usermem" +) + +// emptyIPv4Filter is for comparison with a rule's filters to determine whether +// it is also empty. It is immutable. +var emptyIPv4Filter = stack.IPHeaderFilter{ + Dst: "\x00\x00\x00\x00", + DstMask: "\x00\x00\x00\x00", + Src: "\x00\x00\x00\x00", + SrcMask: "\x00\x00\x00\x00", +} + +// convertNetstackToBinary4 converts the iptables as stored in netstack to the +// format expected by the iptables tool. Linux stores each table as a binary +// blob that can only be traversed by parsing a little data, reading some +// offsets, jumping to those offsets, parsing again, etc. +func convertNetstackToBinary4(stk *stack.Stack, tablename linux.TableName) (linux.KernelIPTGetEntries, linux.IPTGetinfo, error) { + // The table name has to fit in the struct. + if linux.XT_TABLE_MAXNAMELEN < len(tablename) { + return linux.KernelIPTGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("table name %q too long", tablename) + } + + table, ok := stk.IPTables().GetTable(tablename.String(), false) + if !ok { + return linux.KernelIPTGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("couldn't find table %q", tablename) + } + + // Setup the info struct. + entries, info := getEntries4(table, tablename) + return entries, info, nil +} + +func getEntries4(table stack.Table, tablename linux.TableName) (linux.KernelIPTGetEntries, linux.IPTGetinfo) { + var info linux.IPTGetinfo + var entries linux.KernelIPTGetEntries + copy(info.Name[:], tablename[:]) + copy(entries.Name[:], info.Name[:]) + info.ValidHooks = table.ValidHooks() + + for ruleIdx, rule := range table.Rules { + nflog("convert to binary: current offset: %d", entries.Size) + + setHooksAndUnderflow(&info, table, entries.Size, ruleIdx) + // Each rule corresponds to an entry. + entry := linux.KernelIPTEntry{ + Entry: linux.IPTEntry{ + IP: linux.IPTIP{ + Protocol: uint16(rule.Filter.Protocol), + }, + NextOffset: linux.SizeOfIPTEntry, + TargetOffset: linux.SizeOfIPTEntry, + }, + } + copy(entry.Entry.IP.Dst[:], rule.Filter.Dst) + copy(entry.Entry.IP.DstMask[:], rule.Filter.DstMask) + copy(entry.Entry.IP.Src[:], rule.Filter.Src) + copy(entry.Entry.IP.SrcMask[:], rule.Filter.SrcMask) + copy(entry.Entry.IP.OutputInterface[:], rule.Filter.OutputInterface) + copy(entry.Entry.IP.OutputInterfaceMask[:], rule.Filter.OutputInterfaceMask) + if rule.Filter.DstInvert { + entry.Entry.IP.InverseFlags |= linux.IPT_INV_DSTIP + } + if rule.Filter.SrcInvert { + entry.Entry.IP.InverseFlags |= linux.IPT_INV_SRCIP + } + if rule.Filter.OutputInterfaceInvert { + entry.Entry.IP.InverseFlags |= linux.IPT_INV_VIA_OUT + } + + for _, matcher := range rule.Matchers { + // Serialize the matcher and add it to the + // entry. + serialized := marshalMatcher(matcher) + nflog("convert to binary: matcher serialized as: %v", serialized) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("matcher %T is not 64-bit aligned", matcher)) + } + entry.Elems = append(entry.Elems, serialized...) + entry.Entry.NextOffset += uint16(len(serialized)) + entry.Entry.TargetOffset += uint16(len(serialized)) + } + + // Serialize and append the target. + serialized := marshalTarget(rule.Target) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("target %T is not 64-bit aligned", rule.Target)) + } + entry.Elems = append(entry.Elems, serialized...) + entry.Entry.NextOffset += uint16(len(serialized)) + + nflog("convert to binary: adding entry: %+v", entry) + + entries.Size += uint32(entry.Entry.NextOffset) + entries.Entrytable = append(entries.Entrytable, entry) + info.NumEntries++ + } + + info.Size = entries.Size + nflog("convert to binary: finished with an marshalled size of %d", info.Size) + return entries, info +} + +func modifyEntries4(stk *stack.Stack, optVal []byte, replace *linux.IPTReplace, table *stack.Table) (map[uint32]int, *syserr.Error) { + nflog("set entries: setting entries in table %q", replace.Name.String()) + + // Convert input into a list of rules and their offsets. + var offset uint32 + // offsets maps rule byte offsets to their position in table.Rules. + offsets := map[uint32]int{} + for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { + nflog("set entries: processing entry at offset %d", offset) + + // Get the struct ipt_entry. + if len(optVal) < linux.SizeOfIPTEntry { + nflog("optVal has insufficient size for entry %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + var entry linux.IPTEntry + buf := optVal[:linux.SizeOfIPTEntry] + binary.Unmarshal(buf, usermem.ByteOrder, &entry) + initialOptValLen := len(optVal) + optVal = optVal[linux.SizeOfIPTEntry:] + + if entry.TargetOffset < linux.SizeOfIPTEntry { + nflog("entry has too-small target offset %d", entry.TargetOffset) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): We should support more IPTIP + // filtering fields. + filter, err := filterFromIPTIP(entry.IP) + if err != nil { + nflog("bad iptip: %v", err) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): Matchers and targets can specify + // that they only work for certain protocols, hooks, tables. + // Get matchers. + matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry + if len(optVal) < int(matchersSize) { + nflog("entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + matchers, err := parseMatchers(filter, optVal[:matchersSize]) + if err != nil { + nflog("failed to parse matchers: %v", err) + return nil, syserr.ErrInvalidArgument + } + optVal = optVal[matchersSize:] + + // Get the target of the rule. + targetSize := entry.NextOffset - entry.TargetOffset + if len(optVal) < int(targetSize) { + nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + + rule := stack.Rule{ + Filter: filter, + Matchers: matchers, + } + + { + target, err := parseTarget(filter, optVal[:targetSize], false /* ipv6 */) + if err != nil { + nflog("failed to parse target: %v", err) + return nil, err + } + rule.Target = target + } + optVal = optVal[targetSize:] + + table.Rules = append(table.Rules, rule) + offsets[offset] = int(entryIdx) + offset += uint32(entry.NextOffset) + + if initialOptValLen-len(optVal) != int(entry.NextOffset) { + nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + return nil, syserr.ErrInvalidArgument + } + } + return offsets, nil +} + +func filterFromIPTIP(iptip linux.IPTIP) (stack.IPHeaderFilter, error) { + if containsUnsupportedFields4(iptip) { + return stack.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + } + if len(iptip.Dst) != header.IPv4AddressSize || len(iptip.DstMask) != header.IPv4AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of destination (%d) and/or destination mask (%d) fields", len(iptip.Dst), len(iptip.DstMask)) + } + if len(iptip.Src) != header.IPv4AddressSize || len(iptip.SrcMask) != header.IPv4AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of source (%d) and/or source mask (%d) fields", len(iptip.Src), len(iptip.SrcMask)) + } + + n := bytes.IndexByte([]byte(iptip.OutputInterface[:]), 0) + if n == -1 { + n = len(iptip.OutputInterface) + } + ifname := string(iptip.OutputInterface[:n]) + + n = bytes.IndexByte([]byte(iptip.OutputInterfaceMask[:]), 0) + if n == -1 { + n = len(iptip.OutputInterfaceMask) + } + ifnameMask := string(iptip.OutputInterfaceMask[:n]) + + return stack.IPHeaderFilter{ + Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + // A Protocol value of 0 indicates all protocols match. + CheckProtocol: iptip.Protocol != 0, + Dst: tcpip.Address(iptip.Dst[:]), + DstMask: tcpip.Address(iptip.DstMask[:]), + DstInvert: iptip.InverseFlags&linux.IPT_INV_DSTIP != 0, + Src: tcpip.Address(iptip.Src[:]), + SrcMask: tcpip.Address(iptip.SrcMask[:]), + SrcInvert: iptip.InverseFlags&linux.IPT_INV_SRCIP != 0, + OutputInterface: ifname, + OutputInterfaceMask: ifnameMask, + OutputInterfaceInvert: iptip.InverseFlags&linux.IPT_INV_VIA_OUT != 0, + }, nil +} + +func containsUnsupportedFields4(iptip linux.IPTIP) bool { + // The following features are supported: + // - Protocol + // - Dst and DstMask + // - Src and SrcMask + // - The inverse destination IP check flag + // - OutputInterface, OutputInterfaceMask and its inverse. + var emptyInterface = [linux.IFNAMSIZ]byte{} + // Disable any supported inverse flags. + inverseMask := uint8(linux.IPT_INV_DSTIP) | uint8(linux.IPT_INV_SRCIP) | uint8(linux.IPT_INV_VIA_OUT) + return iptip.InputInterface != emptyInterface || + iptip.InputInterfaceMask != emptyInterface || + iptip.Flags != 0 || + iptip.InverseFlags&^inverseMask != 0 +} diff --git a/pkg/sentry/socket/netfilter/ipv6.go b/pkg/sentry/socket/netfilter/ipv6.go new file mode 100644 index 000000000..4253f7bf4 --- /dev/null +++ b/pkg/sentry/socket/netfilter/ipv6.go @@ -0,0 +1,270 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package netfilter + +import ( + "bytes" + "fmt" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" + "gvisor.dev/gvisor/pkg/tcpip" + "gvisor.dev/gvisor/pkg/tcpip/header" + "gvisor.dev/gvisor/pkg/tcpip/stack" + "gvisor.dev/gvisor/pkg/usermem" +) + +// emptyIPv6Filter is for comparison with a rule's filters to determine whether +// it is also empty. It is immutable. +var emptyIPv6Filter = stack.IPHeaderFilter{ + Dst: "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00", + DstMask: "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00", + Src: "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00", + SrcMask: "\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00\x00", +} + +// convertNetstackToBinary6 converts the ip6tables as stored in netstack to the +// format expected by the iptables tool. Linux stores each table as a binary +// blob that can only be traversed by parsing a little data, reading some +// offsets, jumping to those offsets, parsing again, etc. +func convertNetstackToBinary6(stk *stack.Stack, tablename linux.TableName) (linux.KernelIP6TGetEntries, linux.IPTGetinfo, error) { + // The table name has to fit in the struct. + if linux.XT_TABLE_MAXNAMELEN < len(tablename) { + return linux.KernelIP6TGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("table name %q too long", tablename) + } + + table, ok := stk.IPTables().GetTable(tablename.String(), true) + if !ok { + return linux.KernelIP6TGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("couldn't find table %q", tablename) + } + + // Setup the info struct, which is the same in IPv4 and IPv6. + entries, info := getEntries6(table, tablename) + return entries, info, nil +} + +func getEntries6(table stack.Table, tablename linux.TableName) (linux.KernelIP6TGetEntries, linux.IPTGetinfo) { + var info linux.IPTGetinfo + var entries linux.KernelIP6TGetEntries + copy(info.Name[:], tablename[:]) + copy(entries.Name[:], info.Name[:]) + info.ValidHooks = table.ValidHooks() + + for ruleIdx, rule := range table.Rules { + nflog("convert to binary: current offset: %d", entries.Size) + + setHooksAndUnderflow(&info, table, entries.Size, ruleIdx) + // Each rule corresponds to an entry. + entry := linux.KernelIP6TEntry{ + Entry: linux.IP6TEntry{ + IPv6: linux.IP6TIP{ + Protocol: uint16(rule.Filter.Protocol), + }, + NextOffset: linux.SizeOfIP6TEntry, + TargetOffset: linux.SizeOfIP6TEntry, + }, + } + copy(entry.Entry.IPv6.Dst[:], rule.Filter.Dst) + copy(entry.Entry.IPv6.DstMask[:], rule.Filter.DstMask) + copy(entry.Entry.IPv6.Src[:], rule.Filter.Src) + copy(entry.Entry.IPv6.SrcMask[:], rule.Filter.SrcMask) + copy(entry.Entry.IPv6.OutputInterface[:], rule.Filter.OutputInterface) + copy(entry.Entry.IPv6.OutputInterfaceMask[:], rule.Filter.OutputInterfaceMask) + if rule.Filter.DstInvert { + entry.Entry.IPv6.InverseFlags |= linux.IP6T_INV_DSTIP + } + if rule.Filter.SrcInvert { + entry.Entry.IPv6.InverseFlags |= linux.IP6T_INV_SRCIP + } + if rule.Filter.OutputInterfaceInvert { + entry.Entry.IPv6.InverseFlags |= linux.IP6T_INV_VIA_OUT + } + if rule.Filter.CheckProtocol { + entry.Entry.IPv6.Flags |= linux.IP6T_F_PROTO + } + + for _, matcher := range rule.Matchers { + // Serialize the matcher and add it to the + // entry. + serialized := marshalMatcher(matcher) + nflog("convert to binary: matcher serialized as: %v", serialized) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("matcher %T is not 64-bit aligned", matcher)) + } + entry.Elems = append(entry.Elems, serialized...) + entry.Entry.NextOffset += uint16(len(serialized)) + entry.Entry.TargetOffset += uint16(len(serialized)) + } + + // Serialize and append the target. + serialized := marshalTarget(rule.Target) + if len(serialized)%8 != 0 { + panic(fmt.Sprintf("target %T is not 64-bit aligned", rule.Target)) + } + entry.Elems = append(entry.Elems, serialized...) + entry.Entry.NextOffset += uint16(len(serialized)) + + nflog("convert to binary: adding entry: %+v", entry) + + entries.Size += uint32(entry.Entry.NextOffset) + entries.Entrytable = append(entries.Entrytable, entry) + info.NumEntries++ + } + + info.Size = entries.Size + nflog("convert to binary: finished with an marshalled size of %d", info.Size) + return entries, info +} + +func modifyEntries6(stk *stack.Stack, optVal []byte, replace *linux.IPTReplace, table *stack.Table) (map[uint32]int, *syserr.Error) { + nflog("set entries: setting entries in table %q", replace.Name.String()) + + // Convert input into a list of rules and their offsets. + var offset uint32 + // offsets maps rule byte offsets to their position in table.Rules. + offsets := map[uint32]int{} + for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { + nflog("set entries: processing entry at offset %d", offset) + + // Get the struct ipt_entry. + if len(optVal) < linux.SizeOfIP6TEntry { + nflog("optVal has insufficient size for entry %d", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + var entry linux.IP6TEntry + buf := optVal[:linux.SizeOfIP6TEntry] + binary.Unmarshal(buf, usermem.ByteOrder, &entry) + initialOptValLen := len(optVal) + optVal = optVal[linux.SizeOfIP6TEntry:] + + if entry.TargetOffset < linux.SizeOfIP6TEntry { + nflog("entry has too-small target offset %d", entry.TargetOffset) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): We should support more IPTIP + // filtering fields. + filter, err := filterFromIP6TIP(entry.IPv6) + if err != nil { + nflog("bad iptip: %v", err) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): Matchers and targets can specify + // that they only work for certain protocols, hooks, tables. + // Get matchers. + matchersSize := entry.TargetOffset - linux.SizeOfIP6TEntry + if len(optVal) < int(matchersSize) { + nflog("entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + matchers, err := parseMatchers(filter, optVal[:matchersSize]) + if err != nil { + nflog("failed to parse matchers: %v", err) + return nil, syserr.ErrInvalidArgument + } + optVal = optVal[matchersSize:] + + // Get the target of the rule. + targetSize := entry.NextOffset - entry.TargetOffset + if len(optVal) < int(targetSize) { + nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) + return nil, syserr.ErrInvalidArgument + } + + rule := stack.Rule{ + Filter: filter, + Matchers: matchers, + } + + { + target, err := parseTarget(filter, optVal[:targetSize], true /* ipv6 */) + if err != nil { + nflog("failed to parse target: %v", err) + return nil, err + } + rule.Target = target + } + optVal = optVal[targetSize:] + + table.Rules = append(table.Rules, rule) + offsets[offset] = int(entryIdx) + offset += uint32(entry.NextOffset) + + if initialOptValLen-len(optVal) != int(entry.NextOffset) { + nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) + return nil, syserr.ErrInvalidArgument + } + } + return offsets, nil +} + +func filterFromIP6TIP(iptip linux.IP6TIP) (stack.IPHeaderFilter, error) { + if containsUnsupportedFields6(iptip) { + return stack.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) + } + if len(iptip.Dst) != header.IPv6AddressSize || len(iptip.DstMask) != header.IPv6AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of destination (%d) and/or destination mask (%d) fields", len(iptip.Dst), len(iptip.DstMask)) + } + if len(iptip.Src) != header.IPv6AddressSize || len(iptip.SrcMask) != header.IPv6AddressSize { + return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of source (%d) and/or source mask (%d) fields", len(iptip.Src), len(iptip.SrcMask)) + } + + n := bytes.IndexByte([]byte(iptip.OutputInterface[:]), 0) + if n == -1 { + n = len(iptip.OutputInterface) + } + ifname := string(iptip.OutputInterface[:n]) + + n = bytes.IndexByte([]byte(iptip.OutputInterfaceMask[:]), 0) + if n == -1 { + n = len(iptip.OutputInterfaceMask) + } + ifnameMask := string(iptip.OutputInterfaceMask[:n]) + + return stack.IPHeaderFilter{ + Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), + // In ip6tables a flag controls whether to check the protocol. + CheckProtocol: iptip.Flags&linux.IP6T_F_PROTO != 0, + Dst: tcpip.Address(iptip.Dst[:]), + DstMask: tcpip.Address(iptip.DstMask[:]), + DstInvert: iptip.InverseFlags&linux.IP6T_INV_DSTIP != 0, + Src: tcpip.Address(iptip.Src[:]), + SrcMask: tcpip.Address(iptip.SrcMask[:]), + SrcInvert: iptip.InverseFlags&linux.IP6T_INV_SRCIP != 0, + OutputInterface: ifname, + OutputInterfaceMask: ifnameMask, + OutputInterfaceInvert: iptip.InverseFlags&linux.IP6T_INV_VIA_OUT != 0, + }, nil +} + +func containsUnsupportedFields6(iptip linux.IP6TIP) bool { + // The following features are supported: + // - Protocol + // - Dst and DstMask + // - Src and SrcMask + // - The inverse destination IP check flag + // - OutputInterface, OutputInterfaceMask and its inverse. + var emptyInterface = [linux.IFNAMSIZ]byte{} + flagMask := uint8(linux.IP6T_F_PROTO) + // Disable any supported inverse flags. + inverseMask := uint8(linux.IP6T_INV_DSTIP) | uint8(linux.IP6T_INV_SRCIP) | uint8(linux.IP6T_INV_VIA_OUT) + return iptip.InputInterface != emptyInterface || + iptip.InputInterfaceMask != emptyInterface || + iptip.Flags&^flagMask != 0 || + iptip.InverseFlags&^inverseMask != 0 || + iptip.TOS != 0 +} diff --git a/pkg/sentry/socket/netfilter/netfilter.go b/pkg/sentry/socket/netfilter/netfilter.go index e91b0624c..904a12e38 100644 --- a/pkg/sentry/socket/netfilter/netfilter.go +++ b/pkg/sentry/socket/netfilter/netfilter.go @@ -17,7 +17,6 @@ package netfilter import ( - "bytes" "errors" "fmt" @@ -27,7 +26,6 @@ import ( "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" - "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) @@ -37,15 +35,6 @@ import ( // developing iptables, but can pollute sentry logs otherwise. const enableLogging = false -// emptyFilter is for comparison with a rule's filters to determine whether it -// is also empty. It is immutable. -var emptyFilter = stack.IPHeaderFilter{ - Dst: "\x00\x00\x00\x00", - DstMask: "\x00\x00\x00\x00", - Src: "\x00\x00\x00\x00", - SrcMask: "\x00\x00\x00\x00", -} - // nflog logs messages related to the writing and reading of iptables. func nflog(format string, args ...interface{}) { if enableLogging && log.IsLogging(log.Debug) { @@ -54,14 +43,19 @@ func nflog(format string, args ...interface{}) { } // GetInfo returns information about iptables. -func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPTGetinfo, *syserr.Error) { +func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, ipv6 bool) (linux.IPTGetinfo, *syserr.Error) { // Read in the struct and table name. var info linux.IPTGetinfo if _, err := info.CopyIn(t, outPtr); err != nil { return linux.IPTGetinfo{}, syserr.FromError(err) } - _, info, err := convertNetstackToBinary(stack, info.Name) + var err error + if ipv6 { + _, info, err = convertNetstackToBinary6(stack, info.Name) + } else { + _, info, err = convertNetstackToBinary4(stack, info.Name) + } if err != nil { nflog("couldn't convert iptables: %v", err) return linux.IPTGetinfo{}, syserr.ErrInvalidArgument @@ -71,8 +65,8 @@ func GetInfo(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr) (linux.IPT return info, nil } -// GetEntries returns netstack's iptables rules encoded for the iptables tool. -func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen int) (linux.KernelIPTGetEntries, *syserr.Error) { +// GetEntries4 returns netstack's iptables rules. +func GetEntries4(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen int) (linux.KernelIPTGetEntries, *syserr.Error) { // Read in the struct and table name. var userEntries linux.IPTGetEntries if _, err := userEntries.CopyIn(t, outPtr); err != nil { @@ -82,7 +76,7 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen // Convert netstack's iptables rules to something that the iptables // tool can understand. - entries, _, err := convertNetstackToBinary(stack, userEntries.Name) + entries, _, err := convertNetstackToBinary4(stack, userEntries.Name) if err != nil { nflog("couldn't read entries: %v", err) return linux.KernelIPTGetEntries{}, syserr.ErrInvalidArgument @@ -95,112 +89,53 @@ func GetEntries(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen return entries, nil } -// convertNetstackToBinary converts the iptables as stored in netstack to the -// format expected by the iptables tool. Linux stores each table as a binary -// blob that can only be traversed by parsing a bit, reading some offsets, -// jumping to those offsets, parsing again, etc. -func convertNetstackToBinary(stack *stack.Stack, tablename linux.TableName) (linux.KernelIPTGetEntries, linux.IPTGetinfo, error) { - table, ok := stack.IPTables().GetTable(tablename.String()) - if !ok { - return linux.KernelIPTGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("couldn't find table %q", tablename) +// GetEntries6 returns netstack's ip6tables rules. +func GetEntries6(t *kernel.Task, stack *stack.Stack, outPtr usermem.Addr, outLen int) (linux.KernelIP6TGetEntries, *syserr.Error) { + // Read in the struct and table name. IPv4 and IPv6 utilize structs + // with the same layout. + var userEntries linux.IPTGetEntries + if _, err := userEntries.CopyIn(t, outPtr); err != nil { + nflog("couldn't copy in entries %q", userEntries.Name) + return linux.KernelIP6TGetEntries{}, syserr.FromError(err) } - var entries linux.KernelIPTGetEntries - var info linux.IPTGetinfo - info.ValidHooks = table.ValidHooks() - - // The table name has to fit in the struct. - if linux.XT_TABLE_MAXNAMELEN < len(tablename) { - return linux.KernelIPTGetEntries{}, linux.IPTGetinfo{}, fmt.Errorf("table name %q too long", tablename) + // Convert netstack's iptables rules to something that the iptables + // tool can understand. + entries, _, err := convertNetstackToBinary6(stack, userEntries.Name) + if err != nil { + nflog("couldn't read entries: %v", err) + return linux.KernelIP6TGetEntries{}, syserr.ErrInvalidArgument + } + if binary.Size(entries) > uintptr(outLen) { + nflog("insufficient GetEntries output size: %d", uintptr(outLen)) + return linux.KernelIP6TGetEntries{}, syserr.ErrInvalidArgument } - copy(info.Name[:], tablename[:]) - copy(entries.Name[:], tablename[:]) - - for ruleIdx, rule := range table.Rules { - nflog("convert to binary: current offset: %d", entries.Size) - - // Is this a chain entry point? - for hook, hookRuleIdx := range table.BuiltinChains { - if hookRuleIdx == ruleIdx { - nflog("convert to binary: found hook %d at offset %d", hook, entries.Size) - info.HookEntry[hook] = entries.Size - } - } - // Is this a chain underflow point? - for underflow, underflowRuleIdx := range table.Underflows { - if underflowRuleIdx == ruleIdx { - nflog("convert to binary: found underflow %d at offset %d", underflow, entries.Size) - info.Underflow[underflow] = entries.Size - } - } - // Each rule corresponds to an entry. - entry := linux.KernelIPTEntry{ - Entry: linux.IPTEntry{ - IP: linux.IPTIP{ - Protocol: uint16(rule.Filter.Protocol), - }, - NextOffset: linux.SizeOfIPTEntry, - TargetOffset: linux.SizeOfIPTEntry, - }, - } - copy(entry.Entry.IP.Dst[:], rule.Filter.Dst) - copy(entry.Entry.IP.DstMask[:], rule.Filter.DstMask) - copy(entry.Entry.IP.Src[:], rule.Filter.Src) - copy(entry.Entry.IP.SrcMask[:], rule.Filter.SrcMask) - copy(entry.Entry.IP.OutputInterface[:], rule.Filter.OutputInterface) - copy(entry.Entry.IP.OutputInterfaceMask[:], rule.Filter.OutputInterfaceMask) - if rule.Filter.DstInvert { - entry.Entry.IP.InverseFlags |= linux.IPT_INV_DSTIP - } - if rule.Filter.SrcInvert { - entry.Entry.IP.InverseFlags |= linux.IPT_INV_SRCIP - } - if rule.Filter.OutputInterfaceInvert { - entry.Entry.IP.InverseFlags |= linux.IPT_INV_VIA_OUT - } + return entries, nil +} - for _, matcher := range rule.Matchers { - // Serialize the matcher and add it to the - // entry. - serialized := marshalMatcher(matcher) - nflog("convert to binary: matcher serialized as: %v", serialized) - if len(serialized)%8 != 0 { - panic(fmt.Sprintf("matcher %T is not 64-bit aligned", matcher)) - } - entry.Elems = append(entry.Elems, serialized...) - entry.Entry.NextOffset += uint16(len(serialized)) - entry.Entry.TargetOffset += uint16(len(serialized)) +// setHooksAndUnderflow checks whether the rule at ruleIdx is a hook entrypoint +// or underflow, in which case it fills in info.HookEntry and info.Underflows. +func setHooksAndUnderflow(info *linux.IPTGetinfo, table stack.Table, offset uint32, ruleIdx int) { + // Is this a chain entry point? + for hook, hookRuleIdx := range table.BuiltinChains { + if hookRuleIdx == ruleIdx { + nflog("convert to binary: found hook %d at offset %d", hook, offset) + info.HookEntry[hook] = offset } - - // Serialize and append the target. - serialized := marshalTarget(rule.Target) - if len(serialized)%8 != 0 { - panic(fmt.Sprintf("target %T is not 64-bit aligned", rule.Target)) + } + // Is this a chain underflow point? + for underflow, underflowRuleIdx := range table.Underflows { + if underflowRuleIdx == ruleIdx { + nflog("convert to binary: found underflow %d at offset %d", underflow, offset) + info.Underflow[underflow] = offset } - entry.Elems = append(entry.Elems, serialized...) - entry.Entry.NextOffset += uint16(len(serialized)) - - nflog("convert to binary: adding entry: %+v", entry) - - entries.Size += uint32(entry.Entry.NextOffset) - entries.Entrytable = append(entries.Entrytable, entry) - info.NumEntries++ } - - nflog("convert to binary: finished with an marshalled size of %d", info.Size) - info.Size = entries.Size - return entries, info, nil } // SetEntries sets iptables rules for a single table. See // net/ipv4/netfilter/ip_tables.c:translate_table for reference. -func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { - // Get the basic rules data (struct ipt_replace). - if len(optVal) < linux.SizeOfIPTReplace { - nflog("optVal has insufficient size for replace %d", len(optVal)) - return syserr.ErrInvalidArgument - } +func SetEntries(stk *stack.Stack, optVal []byte, ipv6 bool) *syserr.Error { var replace linux.IPTReplace replaceBuf := optVal[:linux.SizeOfIPTReplace] optVal = optVal[linux.SizeOfIPTReplace:] @@ -218,79 +153,15 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { return syserr.ErrInvalidArgument } - nflog("set entries: setting entries in table %q", replace.Name.String()) - - // Convert input into a list of rules and their offsets. - var offset uint32 - // offsets maps rule byte offsets to their position in table.Rules. - offsets := map[uint32]int{} - for entryIdx := uint32(0); entryIdx < replace.NumEntries; entryIdx++ { - nflog("set entries: processing entry at offset %d", offset) - - // Get the struct ipt_entry. - if len(optVal) < linux.SizeOfIPTEntry { - nflog("optVal has insufficient size for entry %d", len(optVal)) - return syserr.ErrInvalidArgument - } - var entry linux.IPTEntry - buf := optVal[:linux.SizeOfIPTEntry] - binary.Unmarshal(buf, usermem.ByteOrder, &entry) - initialOptValLen := len(optVal) - optVal = optVal[linux.SizeOfIPTEntry:] - - if entry.TargetOffset < linux.SizeOfIPTEntry { - nflog("entry has too-small target offset %d", entry.TargetOffset) - return syserr.ErrInvalidArgument - } - - // TODO(gvisor.dev/issue/170): We should support more IPTIP - // filtering fields. - filter, err := filterFromIPTIP(entry.IP) - if err != nil { - nflog("bad iptip: %v", err) - return syserr.ErrInvalidArgument - } - - // TODO(gvisor.dev/issue/170): Matchers and targets can specify - // that they only work for certain protocols, hooks, tables. - // Get matchers. - matchersSize := entry.TargetOffset - linux.SizeOfIPTEntry - if len(optVal) < int(matchersSize) { - nflog("entry doesn't have enough room for its matchers (only %d bytes remain)", len(optVal)) - return syserr.ErrInvalidArgument - } - matchers, err := parseMatchers(filter, optVal[:matchersSize]) - if err != nil { - nflog("failed to parse matchers: %v", err) - return syserr.ErrInvalidArgument - } - optVal = optVal[matchersSize:] - - // Get the target of the rule. - targetSize := entry.NextOffset - entry.TargetOffset - if len(optVal) < int(targetSize) { - nflog("entry doesn't have enough room for its target (only %d bytes remain)", len(optVal)) - return syserr.ErrInvalidArgument - } - target, err := parseTarget(filter, optVal[:targetSize]) - if err != nil { - nflog("failed to parse target: %v", err) - return syserr.ErrInvalidArgument - } - optVal = optVal[targetSize:] - - table.Rules = append(table.Rules, stack.Rule{ - Filter: filter, - Target: target, - Matchers: matchers, - }) - offsets[offset] = int(entryIdx) - offset += uint32(entry.NextOffset) - - if initialOptValLen-len(optVal) != int(entry.NextOffset) { - nflog("entry NextOffset is %d, but entry took up %d bytes", entry.NextOffset, initialOptValLen-len(optVal)) - return syserr.ErrInvalidArgument - } + var err *syserr.Error + var offsets map[uint32]int + if ipv6 { + offsets, err = modifyEntries6(stk, optVal, &replace, &table) + } else { + offsets, err = modifyEntries4(stk, optVal, &replace, &table) + } + if err != nil { + return err } // Go through the list of supported hooks for this table and, for each @@ -305,7 +176,7 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { table.BuiltinChains[hk] = ruleIdx } if offset == replace.Underflow[hook] { - if !validUnderflow(table.Rules[ruleIdx]) { + if !validUnderflow(table.Rules[ruleIdx], ipv6) { nflog("underflow for hook %d isn't an unconditional ACCEPT or DROP", ruleIdx) return syserr.ErrInvalidArgument } @@ -323,9 +194,9 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { } } - // Add the user chains. + // Check the user chains. for ruleIdx, rule := range table.Rules { - if _, ok := rule.Target.(stack.UserChainTarget); !ok { + if _, ok := rule.Target.(*stack.UserChainTarget); !ok { continue } @@ -346,7 +217,7 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { // Set each jump to point to the appropriate rule. Right now they hold byte // offsets. for ruleIdx, rule := range table.Rules { - jump, ok := rule.Target.(JumpTarget) + jump, ok := rule.Target.(*JumpTarget) if !ok { continue } @@ -370,7 +241,7 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { if ruleIdx == stack.HookUnset { continue } - if !isUnconditionalAccept(table.Rules[ruleIdx]) { + if !isUnconditionalAccept(table.Rules[ruleIdx], ipv6) { nflog("hook %d is unsupported.", hook) return syserr.ErrInvalidArgument } @@ -382,7 +253,8 @@ func SetEntries(stk *stack.Stack, optVal []byte) *syserr.Error { // - There are no chains without an unconditional final rule. // - There are no chains without an unconditional underflow rule. - return syserr.TranslateNetstackError(stk.IPTables().ReplaceTable(replace.Name.String(), table)) + return syserr.TranslateNetstackError(stk.IPTables().ReplaceTable(replace.Name.String(), table, ipv6)) + } // parseMatchers parses 0 or more matchers from optVal. optVal should contain @@ -404,7 +276,6 @@ func parseMatchers(filter stack.IPHeaderFilter, optVal []byte) ([]stack.Matcher, // Check some invariants. if match.MatchSize < linux.SizeOfXTEntryMatch { - return nil, fmt.Errorf("match size is too small, must be at least %d", linux.SizeOfXTEntryMatch) } if len(optVal) < int(match.MatchSize) { @@ -429,79 +300,26 @@ func parseMatchers(filter stack.IPHeaderFilter, optVal []byte) ([]stack.Matcher, return matchers, nil } -func filterFromIPTIP(iptip linux.IPTIP) (stack.IPHeaderFilter, error) { - if containsUnsupportedFields(iptip) { - return stack.IPHeaderFilter{}, fmt.Errorf("unsupported fields in struct iptip: %+v", iptip) - } - if len(iptip.Dst) != header.IPv4AddressSize || len(iptip.DstMask) != header.IPv4AddressSize { - return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of destination (%d) and/or destination mask (%d) fields", len(iptip.Dst), len(iptip.DstMask)) - } - if len(iptip.Src) != header.IPv4AddressSize || len(iptip.SrcMask) != header.IPv4AddressSize { - return stack.IPHeaderFilter{}, fmt.Errorf("incorrect length of source (%d) and/or source mask (%d) fields", len(iptip.Src), len(iptip.SrcMask)) - } - - n := bytes.IndexByte([]byte(iptip.OutputInterface[:]), 0) - if n == -1 { - n = len(iptip.OutputInterface) - } - ifname := string(iptip.OutputInterface[:n]) - - n = bytes.IndexByte([]byte(iptip.OutputInterfaceMask[:]), 0) - if n == -1 { - n = len(iptip.OutputInterfaceMask) - } - ifnameMask := string(iptip.OutputInterfaceMask[:n]) - - return stack.IPHeaderFilter{ - Protocol: tcpip.TransportProtocolNumber(iptip.Protocol), - Dst: tcpip.Address(iptip.Dst[:]), - DstMask: tcpip.Address(iptip.DstMask[:]), - DstInvert: iptip.InverseFlags&linux.IPT_INV_DSTIP != 0, - Src: tcpip.Address(iptip.Src[:]), - SrcMask: tcpip.Address(iptip.SrcMask[:]), - SrcInvert: iptip.InverseFlags&linux.IPT_INV_SRCIP != 0, - OutputInterface: ifname, - OutputInterfaceMask: ifnameMask, - OutputInterfaceInvert: iptip.InverseFlags&linux.IPT_INV_VIA_OUT != 0, - }, nil -} - -func containsUnsupportedFields(iptip linux.IPTIP) bool { - // The following features are supported: - // - Protocol - // - Dst and DstMask - // - Src and SrcMask - // - The inverse destination IP check flag - // - OutputInterface, OutputInterfaceMask and its inverse. - var emptyInterface = [linux.IFNAMSIZ]byte{} - // Disable any supported inverse flags. - inverseMask := uint8(linux.IPT_INV_DSTIP) | uint8(linux.IPT_INV_SRCIP) | uint8(linux.IPT_INV_VIA_OUT) - return iptip.InputInterface != emptyInterface || - iptip.InputInterfaceMask != emptyInterface || - iptip.Flags != 0 || - iptip.InverseFlags&^inverseMask != 0 -} - -func validUnderflow(rule stack.Rule) bool { +func validUnderflow(rule stack.Rule, ipv6 bool) bool { if len(rule.Matchers) != 0 { return false } - if rule.Filter != emptyFilter { + if (ipv6 && rule.Filter != emptyIPv6Filter) || (!ipv6 && rule.Filter != emptyIPv4Filter) { return false } switch rule.Target.(type) { - case stack.AcceptTarget, stack.DropTarget: + case *stack.AcceptTarget, *stack.DropTarget: return true default: return false } } -func isUnconditionalAccept(rule stack.Rule) bool { - if !validUnderflow(rule) { +func isUnconditionalAccept(rule stack.Rule, ipv6 bool) bool { + if !validUnderflow(rule, ipv6) { return false } - _, ok := rule.Target.(stack.AcceptTarget) + _, ok := rule.Target.(*stack.AcceptTarget) return ok } @@ -520,3 +338,20 @@ func hookFromLinux(hook int) stack.Hook { } panic(fmt.Sprintf("Unknown hook %d does not correspond to a builtin chain", hook)) } + +// TargetRevision returns a linux.XTGetRevision for a given target. It sets +// Revision to the highest supported value, unless the provided revision number +// is larger. +func TargetRevision(t *kernel.Task, revPtr usermem.Addr, netProto tcpip.NetworkProtocolNumber) (linux.XTGetRevision, *syserr.Error) { + // Read in the target name and version. + var rev linux.XTGetRevision + if _, err := rev.CopyIn(t, revPtr); err != nil { + return linux.XTGetRevision{}, syserr.FromError(err) + } + maxSupported, ok := targetRevision(rev.Name.String(), netProto, rev.Revision) + if !ok { + return linux.XTGetRevision{}, syserr.ErrProtocolNotSupported + } + rev.Revision = maxSupported + return rev, nil +} diff --git a/pkg/sentry/socket/netfilter/targets.go b/pkg/sentry/socket/netfilter/targets.go index 8ebdaff18..0e14447fe 100644 --- a/pkg/sentry/socket/netfilter/targets.go +++ b/pkg/sentry/socket/netfilter/targets.go @@ -15,255 +15,357 @@ package netfilter import ( - "errors" "fmt" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/tcpip/header" "gvisor.dev/gvisor/pkg/tcpip/stack" "gvisor.dev/gvisor/pkg/usermem" ) -// errorTargetName is used to mark targets as error targets. Error targets -// shouldn't be reached - an error has occurred if we fall through to one. -const errorTargetName = "ERROR" +func init() { + // Standard targets include ACCEPT, DROP, RETURN, and JUMP. + registerTargetMaker(&standardTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&standardTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) + + // Both user chains and actual errors are represented in iptables by + // error targets. + registerTargetMaker(&errorTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&errorTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) + + registerTargetMaker(&redirectTargetMaker{ + NetworkProtocol: header.IPv4ProtocolNumber, + }) + registerTargetMaker(&nfNATTargetMaker{ + NetworkProtocol: header.IPv6ProtocolNumber, + }) +} -// redirectTargetName is used to mark targets as redirect targets. Redirect -// targets should be reached for only NAT and Mangle tables. These targets will -// change the destination port/destination IP for packets. -const redirectTargetName = "REDIRECT" +type standardTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} -func marshalTarget(target stack.Target) []byte { +func (sm *standardTargetMaker) id() stack.TargetID { + // Standard targets have the empty string as a name and no revisions. + return stack.TargetID{ + NetworkProtocol: sm.NetworkProtocol, + } +} +func (*standardTargetMaker) marshal(target stack.Target) []byte { + // Translate verdicts the same way as the iptables tool. + var verdict int32 switch tg := target.(type) { - case stack.AcceptTarget: - return marshalStandardTarget(stack.RuleAccept) - case stack.DropTarget: - return marshalStandardTarget(stack.RuleDrop) - case stack.ErrorTarget: - return marshalErrorTarget(errorTargetName) - case stack.UserChainTarget: - return marshalErrorTarget(tg.Name) - case stack.ReturnTarget: - return marshalStandardTarget(stack.RuleReturn) - case stack.RedirectTarget: - return marshalRedirectTarget(tg) - case JumpTarget: - return marshalJumpTarget(tg) + case *stack.AcceptTarget: + verdict = -linux.NF_ACCEPT - 1 + case *stack.DropTarget: + verdict = -linux.NF_DROP - 1 + case *stack.ReturnTarget: + verdict = linux.NF_RETURN + case *JumpTarget: + verdict = int32(tg.Offset) default: panic(fmt.Errorf("unknown target of type %T", target)) } -} - -func marshalStandardTarget(verdict stack.RuleVerdict) []byte { - nflog("convert to binary: marshalling standard target") // The target's name will be the empty string. - target := linux.XTStandardTarget{ + xt := linux.XTStandardTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTStandardTarget, }, - Verdict: translateFromStandardVerdict(verdict), + Verdict: verdict, } ret := make([]byte, 0, linux.SizeOfXTStandardTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + return binary.Marshal(ret, usermem.ByteOrder, xt) +} + +func (*standardTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) != linux.SizeOfXTStandardTarget { + nflog("buf has wrong size for standard target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + var standardTarget linux.XTStandardTarget + buf = buf[:linux.SizeOfXTStandardTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) + + if standardTarget.Verdict < 0 { + // A Verdict < 0 indicates a non-jump verdict. + return translateToStandardTarget(standardTarget.Verdict, filter.NetworkProtocol()) + } + // A verdict >= 0 indicates a jump. + return &JumpTarget{ + Offset: uint32(standardTarget.Verdict), + NetworkProtocol: filter.NetworkProtocol(), + }, nil +} + +type errorTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (em *errorTargetMaker) id() stack.TargetID { + // Error targets have no revision. + return stack.TargetID{ + Name: stack.ErrorTargetName, + NetworkProtocol: em.NetworkProtocol, + } } -func marshalErrorTarget(errorName string) []byte { +func (*errorTargetMaker) marshal(target stack.Target) []byte { + var errorName string + switch tg := target.(type) { + case *stack.ErrorTarget: + errorName = stack.ErrorTargetName + case *stack.UserChainTarget: + errorName = tg.Name + default: + panic(fmt.Sprintf("errorMakerTarget cannot marshal unknown type %T", target)) + } + // This is an error target named error - target := linux.XTErrorTarget{ + xt := linux.XTErrorTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTErrorTarget, }, } - copy(target.Name[:], errorName) - copy(target.Target.Name[:], errorTargetName) + copy(xt.Name[:], errorName) + copy(xt.Target.Name[:], stack.ErrorTargetName) ret := make([]byte, 0, linux.SizeOfXTErrorTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + return binary.Marshal(ret, usermem.ByteOrder, xt) } -func marshalRedirectTarget(rt stack.RedirectTarget) []byte { +func (*errorTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) != linux.SizeOfXTErrorTarget { + nflog("buf has insufficient size for error target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + var errorTarget linux.XTErrorTarget + buf = buf[:linux.SizeOfXTErrorTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &errorTarget) + + // Error targets are used in 2 cases: + // * An actual error case. These rules have an error + // named stack.ErrorTargetName. The last entry of the table + // is usually an error case to catch any packets that + // somehow fall through every rule. + // * To mark the start of a user defined chain. These + // rules have an error with the name of the chain. + switch name := errorTarget.Name.String(); name { + case stack.ErrorTargetName: + return &stack.ErrorTarget{NetworkProtocol: filter.NetworkProtocol()}, nil + default: + // User defined chain. + return &stack.UserChainTarget{ + Name: name, + NetworkProtocol: filter.NetworkProtocol(), + }, nil + } +} + +type redirectTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (rm *redirectTargetMaker) id() stack.TargetID { + return stack.TargetID{ + Name: stack.RedirectTargetName, + NetworkProtocol: rm.NetworkProtocol, + } +} + +func (*redirectTargetMaker) marshal(target stack.Target) []byte { + rt := target.(*stack.RedirectTarget) // This is a redirect target named redirect - target := linux.XTRedirectTarget{ + xt := linux.XTRedirectTarget{ Target: linux.XTEntryTarget{ TargetSize: linux.SizeOfXTRedirectTarget, }, } - copy(target.Target.Name[:], redirectTargetName) + copy(xt.Target.Name[:], stack.RedirectTargetName) ret := make([]byte, 0, linux.SizeOfXTRedirectTarget) - target.NfRange.RangeSize = 1 - if rt.RangeProtoSpecified { - target.NfRange.RangeIPV4.Flags |= linux.NF_NAT_RANGE_PROTO_SPECIFIED + xt.NfRange.RangeSize = 1 + xt.NfRange.RangeIPV4.Flags |= linux.NF_NAT_RANGE_PROTO_SPECIFIED + xt.NfRange.RangeIPV4.MinPort = htons(rt.Port) + xt.NfRange.RangeIPV4.MaxPort = xt.NfRange.RangeIPV4.MinPort + return binary.Marshal(ret, usermem.ByteOrder, xt) +} + +func (*redirectTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if len(buf) < linux.SizeOfXTRedirectTarget { + nflog("redirectTargetMaker: buf has insufficient size for redirect target %d", len(buf)) + return nil, syserr.ErrInvalidArgument + } + + if p := filter.Protocol; p != header.TCPProtocolNumber && p != header.UDPProtocolNumber { + nflog("redirectTargetMaker: bad proto %d", p) + return nil, syserr.ErrInvalidArgument + } + + var redirectTarget linux.XTRedirectTarget + buf = buf[:linux.SizeOfXTRedirectTarget] + binary.Unmarshal(buf, usermem.ByteOrder, &redirectTarget) + + // Copy linux.XTRedirectTarget to stack.RedirectTarget. + target := stack.RedirectTarget{NetworkProtocol: filter.NetworkProtocol()} + + // RangeSize should be 1. + nfRange := redirectTarget.NfRange + if nfRange.RangeSize != 1 { + nflog("redirectTargetMaker: bad rangesize %d", nfRange.RangeSize) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): Check if the flags are valid. + // Also check if we need to map ports or IP. + // For now, redirect target only supports destination port change. + // Port range and IP range are not supported yet. + if nfRange.RangeIPV4.Flags != linux.NF_NAT_RANGE_PROTO_SPECIFIED { + nflog("redirectTargetMaker: invalid range flags %d", nfRange.RangeIPV4.Flags) + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/170): Port range is not supported yet. + if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { + nflog("redirectTargetMaker: MinPort != MaxPort (%d, %d)", nfRange.RangeIPV4.MinPort, nfRange.RangeIPV4.MaxPort) + return nil, syserr.ErrInvalidArgument } - // Convert port from little endian to big endian. - port := make([]byte, 2) - binary.LittleEndian.PutUint16(port, rt.MinPort) - target.NfRange.RangeIPV4.MinPort = binary.BigEndian.Uint16(port) - binary.LittleEndian.PutUint16(port, rt.MaxPort) - target.NfRange.RangeIPV4.MaxPort = binary.BigEndian.Uint16(port) - return binary.Marshal(ret, usermem.ByteOrder, target) + if nfRange.RangeIPV4.MinIP != nfRange.RangeIPV4.MaxIP { + nflog("redirectTargetMaker: MinIP != MaxIP (%d, %d)", nfRange.RangeIPV4.MinPort, nfRange.RangeIPV4.MaxPort) + return nil, syserr.ErrInvalidArgument + } + + target.Addr = tcpip.Address(nfRange.RangeIPV4.MinIP[:]) + target.Port = ntohs(nfRange.RangeIPV4.MinPort) + + return &target, nil } -func marshalJumpTarget(jt JumpTarget) []byte { - nflog("convert to binary: marshalling jump target") +type nfNATTarget struct { + Target linux.XTEntryTarget + Range linux.NFNATRange +} - // The target's name will be the empty string. - target := linux.XTStandardTarget{ +const nfNATMarhsalledSize = linux.SizeOfXTEntryTarget + linux.SizeOfNFNATRange + +type nfNATTargetMaker struct { + NetworkProtocol tcpip.NetworkProtocolNumber +} + +func (rm *nfNATTargetMaker) id() stack.TargetID { + return stack.TargetID{ + Name: stack.RedirectTargetName, + NetworkProtocol: rm.NetworkProtocol, + } +} + +func (*nfNATTargetMaker) marshal(target stack.Target) []byte { + rt := target.(*stack.RedirectTarget) + nt := nfNATTarget{ Target: linux.XTEntryTarget{ - TargetSize: linux.SizeOfXTStandardTarget, + TargetSize: nfNATMarhsalledSize, + }, + Range: linux.NFNATRange{ + Flags: linux.NF_NAT_RANGE_PROTO_SPECIFIED, }, - // Verdict is overloaded by the ABI. When positive, it holds - // the jump offset from the start of the table. - Verdict: int32(jt.Offset), } + copy(nt.Target.Name[:], stack.RedirectTargetName) + copy(nt.Range.MinAddr[:], rt.Addr) + copy(nt.Range.MaxAddr[:], rt.Addr) - ret := make([]byte, 0, linux.SizeOfXTStandardTarget) - return binary.Marshal(ret, usermem.ByteOrder, target) + nt.Range.MinProto = htons(rt.Port) + nt.Range.MaxProto = nt.Range.MinProto + + ret := make([]byte, 0, nfNATMarhsalledSize) + return binary.Marshal(ret, usermem.ByteOrder, nt) } -// translateFromStandardVerdict translates verdicts the same way as the iptables -// tool. -func translateFromStandardVerdict(verdict stack.RuleVerdict) int32 { - switch verdict { - case stack.RuleAccept: - return -linux.NF_ACCEPT - 1 - case stack.RuleDrop: - return -linux.NF_DROP - 1 - case stack.RuleReturn: - return linux.NF_RETURN - default: - // TODO(gvisor.dev/issue/170): Support Jump. - panic(fmt.Sprintf("unknown standard verdict: %d", verdict)) +func (*nfNATTargetMaker) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Target, *syserr.Error) { + if size := nfNATMarhsalledSize; len(buf) < size { + nflog("nfNATTargetMaker: buf has insufficient size (%d) for nfNAT target (%d)", len(buf), size) + return nil, syserr.ErrInvalidArgument } + + if p := filter.Protocol; p != header.TCPProtocolNumber && p != header.UDPProtocolNumber { + nflog("nfNATTargetMaker: bad proto %d", p) + return nil, syserr.ErrInvalidArgument + } + + var natRange linux.NFNATRange + buf = buf[linux.SizeOfXTEntryTarget:nfNATMarhsalledSize] + binary.Unmarshal(buf, usermem.ByteOrder, &natRange) + + // We don't support port or address ranges. + if natRange.MinAddr != natRange.MaxAddr { + nflog("nfNATTargetMaker: MinAddr and MaxAddr are different") + return nil, syserr.ErrInvalidArgument + } + if natRange.MinProto != natRange.MaxProto { + nflog("nfNATTargetMaker: MinProto and MaxProto are different") + return nil, syserr.ErrInvalidArgument + } + + // TODO(gvisor.dev/issue/3549): Check for other flags. + // For now, redirect target only supports destination change. + if natRange.Flags != linux.NF_NAT_RANGE_PROTO_SPECIFIED { + nflog("nfNATTargetMaker: invalid range flags %d", natRange.Flags) + return nil, syserr.ErrInvalidArgument + } + + target := stack.RedirectTarget{ + NetworkProtocol: filter.NetworkProtocol(), + Addr: tcpip.Address(natRange.MinAddr[:]), + Port: ntohs(natRange.MinProto), + } + + return &target, nil } // translateToStandardTarget translates from the value in a // linux.XTStandardTarget to an stack.Verdict. -func translateToStandardTarget(val int32) (stack.Target, error) { +func translateToStandardTarget(val int32, netProto tcpip.NetworkProtocolNumber) (stack.Target, *syserr.Error) { // TODO(gvisor.dev/issue/170): Support other verdicts. switch val { case -linux.NF_ACCEPT - 1: - return stack.AcceptTarget{}, nil + return &stack.AcceptTarget{NetworkProtocol: netProto}, nil case -linux.NF_DROP - 1: - return stack.DropTarget{}, nil + return &stack.DropTarget{NetworkProtocol: netProto}, nil case -linux.NF_QUEUE - 1: - return nil, errors.New("unsupported iptables verdict QUEUE") + nflog("unsupported iptables verdict QUEUE") + return nil, syserr.ErrInvalidArgument case linux.NF_RETURN: - return stack.ReturnTarget{}, nil + return &stack.ReturnTarget{NetworkProtocol: netProto}, nil default: - return nil, fmt.Errorf("unknown iptables verdict %d", val) + nflog("unknown iptables verdict %d", val) + return nil, syserr.ErrInvalidArgument } } // parseTarget parses a target from optVal. optVal should contain only the // target. -func parseTarget(filter stack.IPHeaderFilter, optVal []byte) (stack.Target, error) { +func parseTarget(filter stack.IPHeaderFilter, optVal []byte, ipv6 bool) (stack.Target, *syserr.Error) { nflog("set entries: parsing target of size %d", len(optVal)) if len(optVal) < linux.SizeOfXTEntryTarget { - return nil, fmt.Errorf("optVal has insufficient size for entry target %d", len(optVal)) + nflog("optVal has insufficient size for entry target %d", len(optVal)) + return nil, syserr.ErrInvalidArgument } var target linux.XTEntryTarget buf := optVal[:linux.SizeOfXTEntryTarget] binary.Unmarshal(buf, usermem.ByteOrder, &target) - switch target.Name.String() { - case "": - // Standard target. - if len(optVal) != linux.SizeOfXTStandardTarget { - return nil, fmt.Errorf("optVal has wrong size for standard target %d", len(optVal)) - } - var standardTarget linux.XTStandardTarget - buf = optVal[:linux.SizeOfXTStandardTarget] - binary.Unmarshal(buf, usermem.ByteOrder, &standardTarget) - - if standardTarget.Verdict < 0 { - // A Verdict < 0 indicates a non-jump verdict. - return translateToStandardTarget(standardTarget.Verdict) - } - // A verdict >= 0 indicates a jump. - return JumpTarget{Offset: uint32(standardTarget.Verdict)}, nil - - case errorTargetName: - // Error target. - if len(optVal) != linux.SizeOfXTErrorTarget { - return nil, fmt.Errorf("optVal has insufficient size for error target %d", len(optVal)) - } - var errorTarget linux.XTErrorTarget - buf = optVal[:linux.SizeOfXTErrorTarget] - binary.Unmarshal(buf, usermem.ByteOrder, &errorTarget) - - // Error targets are used in 2 cases: - // * An actual error case. These rules have an error - // named errorTargetName. The last entry of the table - // is usually an error case to catch any packets that - // somehow fall through every rule. - // * To mark the start of a user defined chain. These - // rules have an error with the name of the chain. - switch name := errorTarget.Name.String(); name { - case errorTargetName: - nflog("set entries: error target") - return stack.ErrorTarget{}, nil - default: - // User defined chain. - nflog("set entries: user-defined target %q", name) - return stack.UserChainTarget{Name: name}, nil - } - - case redirectTargetName: - // Redirect target. - if len(optVal) < linux.SizeOfXTRedirectTarget { - return nil, fmt.Errorf("netfilter.SetEntries: optVal has insufficient size for redirect target %d", len(optVal)) - } - - if filter.Protocol != header.TCPProtocolNumber && filter.Protocol != header.UDPProtocolNumber { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") - } - - var redirectTarget linux.XTRedirectTarget - buf = optVal[:linux.SizeOfXTRedirectTarget] - binary.Unmarshal(buf, usermem.ByteOrder, &redirectTarget) - - // Copy linux.XTRedirectTarget to stack.RedirectTarget. - var target stack.RedirectTarget - nfRange := redirectTarget.NfRange - - // RangeSize should be 1. - if nfRange.RangeSize != 1 { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") - } - - // TODO(gvisor.dev/issue/170): Check if the flags are valid. - // Also check if we need to map ports or IP. - // For now, redirect target only supports destination port change. - // Port range and IP range are not supported yet. - if nfRange.RangeIPV4.Flags&linux.NF_NAT_RANGE_PROTO_SPECIFIED == 0 { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") - } - target.RangeProtoSpecified = true - - target.MinIP = tcpip.Address(nfRange.RangeIPV4.MinIP[:]) - target.MaxIP = tcpip.Address(nfRange.RangeIPV4.MaxIP[:]) - - // TODO(gvisor.dev/issue/170): Port range is not supported yet. - if nfRange.RangeIPV4.MinPort != nfRange.RangeIPV4.MaxPort { - return nil, fmt.Errorf("netfilter.SetEntries: invalid argument") - } - - // Convert port from big endian to little endian. - port := make([]byte, 2) - binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MinPort) - target.MinPort = binary.LittleEndian.Uint16(port) - - binary.BigEndian.PutUint16(port, nfRange.RangeIPV4.MaxPort) - target.MaxPort = binary.LittleEndian.Uint16(port) - return target, nil - } - // Unknown target. - return nil, fmt.Errorf("unknown target %q doesn't exist or isn't supported yet", target.Name.String()) + return unmarshalTarget(target, filter, optVal) } // JumpTarget implements stack.Target. @@ -274,9 +376,31 @@ type JumpTarget struct { // RuleNum is the rule to jump to. RuleNum int + + // NetworkProtocol is the network protocol the target is used with. + NetworkProtocol tcpip.NetworkProtocolNumber +} + +// ID implements Target.ID. +func (jt *JumpTarget) ID() stack.TargetID { + return stack.TargetID{ + NetworkProtocol: jt.NetworkProtocol, + } } // Action implements stack.Target.Action. -func (jt JumpTarget) Action(*stack.PacketBuffer, *stack.ConnTrack, stack.Hook, *stack.GSO, *stack.Route, tcpip.Address) (stack.RuleVerdict, int) { +func (jt *JumpTarget) Action(*stack.PacketBuffer, *stack.ConnTrack, stack.Hook, *stack.GSO, *stack.Route, tcpip.Address) (stack.RuleVerdict, int) { return stack.RuleJump, jt.RuleNum } + +func ntohs(port uint16) uint16 { + buf := make([]byte, 2) + binary.BigEndian.PutUint16(buf, port) + return usermem.ByteOrder.Uint16(buf) +} + +func htons(port uint16) uint16 { + buf := make([]byte, 2) + usermem.ByteOrder.PutUint16(buf, port) + return binary.BigEndian.Uint16(buf) +} diff --git a/pkg/sentry/socket/netfilter/tcp_matcher.go b/pkg/sentry/socket/netfilter/tcp_matcher.go index 0bfd6c1f4..352c51390 100644 --- a/pkg/sentry/socket/netfilter/tcp_matcher.go +++ b/pkg/sentry/socket/netfilter/tcp_matcher.go @@ -71,7 +71,7 @@ func (tcpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Ma } if filter.Protocol != header.TCPProtocolNumber { - return nil, fmt.Errorf("TCP matching is only valid for protocol %d.", header.TCPProtocolNumber) + return nil, fmt.Errorf("TCP matching is only valid for protocol %d", header.TCPProtocolNumber) } return &TCPMatcher{ @@ -97,17 +97,33 @@ func (*TCPMatcher) Name() string { // Match implements Matcher.Match. func (tm *TCPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader().View()) + // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved + // into the stack.Check codepath as matchers are added. + switch pkt.NetworkProtocolNumber { + case header.IPv4ProtocolNumber: + netHeader := header.IPv4(pkt.NetworkHeader().View()) + if netHeader.TransportProtocol() != header.TCPProtocolNumber { + return false, false + } - if netHeader.TransportProtocol() != header.TCPProtocolNumber { - return false, false - } + // We don't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false + } - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - return false, true + case header.IPv6ProtocolNumber: + // As in Linux, we do not perform an IPv6 fragment check. See + // xt_action_param.fragoff in + // include/linux/netfilter/x_tables.h. + if header.IPv6(pkt.NetworkHeader().View()).TransportProtocol() != header.TCPProtocolNumber { + return false, false } + + default: + // We don't know the network protocol. return false, false } diff --git a/pkg/sentry/socket/netfilter/udp_matcher.go b/pkg/sentry/socket/netfilter/udp_matcher.go index 7ed05461d..c88d8268d 100644 --- a/pkg/sentry/socket/netfilter/udp_matcher.go +++ b/pkg/sentry/socket/netfilter/udp_matcher.go @@ -68,7 +68,7 @@ func (udpMarshaler) unmarshal(buf []byte, filter stack.IPHeaderFilter) (stack.Ma } if filter.Protocol != header.UDPProtocolNumber { - return nil, fmt.Errorf("UDP matching is only valid for protocol %d.", header.UDPProtocolNumber) + return nil, fmt.Errorf("UDP matching is only valid for protocol %d", header.UDPProtocolNumber) } return &UDPMatcher{ @@ -94,19 +94,33 @@ func (*UDPMatcher) Name() string { // Match implements Matcher.Match. func (um *UDPMatcher) Match(hook stack.Hook, pkt *stack.PacketBuffer, interfaceName string) (bool, bool) { - netHeader := header.IPv4(pkt.NetworkHeader().View()) - // TODO(gvisor.dev/issue/170): Proto checks should ultimately be moved // into the stack.Check codepath as matchers are added. - if netHeader.TransportProtocol() != header.UDPProtocolNumber { - return false, false - } + switch pkt.NetworkProtocolNumber { + case header.IPv4ProtocolNumber: + netHeader := header.IPv4(pkt.NetworkHeader().View()) + if netHeader.TransportProtocol() != header.UDPProtocolNumber { + return false, false + } - // We dont't match fragments. - if frag := netHeader.FragmentOffset(); frag != 0 { - if frag == 1 { - return false, true + // We don't match fragments. + if frag := netHeader.FragmentOffset(); frag != 0 { + if frag == 1 { + return false, true + } + return false, false } + + case header.IPv6ProtocolNumber: + // As in Linux, we do not perform an IPv6 fragment check. See + // xt_action_param.fragoff in + // include/linux/netfilter/x_tables.h. + if header.IPv6(pkt.NetworkHeader().View()).TransportProtocol() != header.UDPProtocolNumber { + return false, false + } + + default: + // We don't know the network protocol. return false, false } diff --git a/pkg/sentry/socket/netlink/BUILD b/pkg/sentry/socket/netlink/BUILD index 0546801bf..1f926aa91 100644 --- a/pkg/sentry/socket/netlink/BUILD +++ b/pkg/sentry/socket/netlink/BUILD @@ -16,6 +16,8 @@ go_library( "//pkg/abi/linux", "//pkg/binary", "//pkg/context", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/sentry/arch", "//pkg/sentry/device", "//pkg/sentry/fs", @@ -36,8 +38,6 @@ go_library( "//pkg/tcpip", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", - "//tools/go_marshal/primitive", ], ) diff --git a/pkg/sentry/socket/netlink/provider_vfs2.go b/pkg/sentry/socket/netlink/provider_vfs2.go index bb205be0d..f061c5d62 100644 --- a/pkg/sentry/socket/netlink/provider_vfs2.go +++ b/pkg/sentry/socket/netlink/provider_vfs2.go @@ -51,7 +51,8 @@ func (*socketProviderVFS2) Socket(t *kernel.Task, stype linux.SockType, protocol vfsfd := &s.vfsfd mnt := t.Kernel().SocketMount() - d := sockfs.NewDentry(t.Credentials(), mnt) + d := sockfs.NewDentry(t, mnt) + defer d.DecRef(t) if err := vfsfd.Init(s, linux.O_RDWR, mnt, d, &vfs.FileDescriptionOptions{ DenyPRead: true, DenyPWrite: true, diff --git a/pkg/sentry/socket/netlink/route/protocol.go b/pkg/sentry/socket/netlink/route/protocol.go index c84d8bd7c..22216158e 100644 --- a/pkg/sentry/socket/netlink/route/protocol.go +++ b/pkg/sentry/socket/netlink/route/protocol.go @@ -36,9 +36,9 @@ type commandKind int const ( kindNew commandKind = 0x0 - kindDel = 0x1 - kindGet = 0x2 - kindSet = 0x3 + kindDel commandKind = 0x1 + kindGet commandKind = 0x2 + kindSet commandKind = 0x3 ) func typeKind(typ uint16) commandKind { @@ -423,6 +423,11 @@ func (p *Protocol) newAddr(ctx context.Context, msg *netlink.Message, ms *netlin } attrs = rest + // NOTE: A netlink message will contain multiple header attributes. + // Both the IFA_ADDRESS and IFA_LOCAL attributes are typically sent + // with IFA_ADDRESS being a prefix address and IFA_LOCAL being the + // local interface address. We add the local interface address here + // and ignore the IFA_ADDRESS. switch ahdr.Type { case linux.IFA_LOCAL: err := stack.AddInterfaceAddr(int32(ifa.Index), inet.InterfaceAddr{ @@ -439,8 +444,57 @@ func (p *Protocol) newAddr(ctx context.Context, msg *netlink.Message, ms *netlin } else if err != nil { return syserr.ErrInvalidArgument } + case linux.IFA_ADDRESS: + default: + return syserr.ErrNotSupported + } + } + return nil +} + +// delAddr handles RTM_DELADDR requests. +func (p *Protocol) delAddr(ctx context.Context, msg *netlink.Message, ms *netlink.MessageSet) *syserr.Error { + stack := inet.StackFromContext(ctx) + if stack == nil { + // No network stack. + return syserr.ErrProtocolNotSupported + } + + var ifa linux.InterfaceAddrMessage + attrs, ok := msg.GetData(&ifa) + if !ok { + return syserr.ErrInvalidArgument + } + + for !attrs.Empty() { + ahdr, value, rest, ok := attrs.ParseFirst() + if !ok { + return syserr.ErrInvalidArgument + } + attrs = rest + + // NOTE: A netlink message will contain multiple header attributes. + // Both the IFA_ADDRESS and IFA_LOCAL attributes are typically sent + // with IFA_ADDRESS being a prefix address and IFA_LOCAL being the + // local interface address. We use the local interface address to + // remove the address and ignore the IFA_ADDRESS. + switch ahdr.Type { + case linux.IFA_LOCAL: + err := stack.RemoveInterfaceAddr(int32(ifa.Index), inet.InterfaceAddr{ + Family: ifa.Family, + PrefixLen: ifa.PrefixLen, + Flags: ifa.Flags, + Addr: value, + }) + if err != nil { + return syserr.ErrInvalidArgument + } + case linux.IFA_ADDRESS: + default: + return syserr.ErrNotSupported } } + return nil } @@ -485,6 +539,8 @@ func (p *Protocol) ProcessMessage(ctx context.Context, msg *netlink.Message, ms return p.dumpRoutes(ctx, msg, ms) case linux.RTM_NEWADDR: return p.newAddr(ctx, msg, ms) + case linux.RTM_DELADDR: + return p.delAddr(ctx, msg, ms) default: return syserr.ErrNotSupported } diff --git a/pkg/sentry/socket/netlink/socket.go b/pkg/sentry/socket/netlink/socket.go index 68a9b9a96..3baad098b 100644 --- a/pkg/sentry/socket/netlink/socket.go +++ b/pkg/sentry/socket/netlink/socket.go @@ -16,11 +16,14 @@ package netlink import ( + "io" "math" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" @@ -38,8 +41,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) const sizeOfInt32 int = 4 @@ -748,6 +749,12 @@ func (s *socketOpsCommon) sendMsg(ctx context.Context, src usermem.IOSequence, t buf := make([]byte, src.NumBytes()) n, err := src.CopyIn(ctx, buf) + // io.EOF can be only returned if src is a file, this means that + // sendMsg is called from splice and the error has to be ignored in + // this case. + if err == io.EOF { + err = nil + } if err != nil { // Don't partially consume messages. return 0, syserr.FromError(err) diff --git a/pkg/sentry/socket/netlink/socket_vfs2.go b/pkg/sentry/socket/netlink/socket_vfs2.go index a38d25da9..461d524e5 100644 --- a/pkg/sentry/socket/netlink/socket_vfs2.go +++ b/pkg/sentry/socket/netlink/socket_vfs2.go @@ -37,6 +37,8 @@ import ( // to/from the kernel. // // SocketVFS2 implements socket.SocketVFS2 and transport.Credentialer. +// +// +stateify savable type SocketVFS2 struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -82,6 +84,13 @@ func NewVFS2(t *kernel.Task, skType linux.SockType, protocol Protocol) (*SocketV return fd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) diff --git a/pkg/sentry/socket/netstack/BUILD b/pkg/sentry/socket/netstack/BUILD index 1fb777a6c..fae3b6783 100644 --- a/pkg/sentry/socket/netstack/BUILD +++ b/pkg/sentry/socket/netstack/BUILD @@ -22,6 +22,8 @@ go_library( "//pkg/binary", "//pkg/context", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/metric", "//pkg/safemem", "//pkg/sentry/arch", @@ -51,8 +53,6 @@ go_library( "//pkg/tcpip/transport/udp", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", - "//tools/go_marshal/primitive", "@org_golang_x_sys//unix:go_default_library", ], ) diff --git a/pkg/sentry/socket/netstack/netstack.go b/pkg/sentry/socket/netstack/netstack.go index e4846bc0b..86c634715 100644 --- a/pkg/sentry/socket/netstack/netstack.go +++ b/pkg/sentry/socket/netstack/netstack.go @@ -40,6 +40,8 @@ import ( "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/metric" "gvisor.dev/gvisor/pkg/safemem" "gvisor.dev/gvisor/pkg/sentry/arch" @@ -62,8 +64,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip/transport/udp" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) func mustCreateMetric(name, description string) *tcpip.StatCounter { @@ -158,6 +158,9 @@ var Metrics = tcpip.Stats{ OutgoingPacketErrors: mustCreateMetric("/netstack/ip/outgoing_packet_errors", "Total number of IP packets which failed to write to a link-layer endpoint."), MalformedPacketsReceived: mustCreateMetric("/netstack/ip/malformed_packets_received", "Total number of IP packets which failed IP header validation checks."), MalformedFragmentsReceived: mustCreateMetric("/netstack/ip/malformed_fragments_received", "Total number of IP fragments which failed IP fragment validation checks."), + IPTablesPreroutingDropped: mustCreateMetric("/netstack/ip/iptables/prerouting_dropped", "Total number of IP packets dropped in the Prerouting chain."), + IPTablesInputDropped: mustCreateMetric("/netstack/ip/iptables/input_dropped", "Total number of IP packets dropped in the Input chain."), + IPTablesOutputDropped: mustCreateMetric("/netstack/ip/iptables/output_dropped", "Total number of IP packets dropped in the Output chain."), }, TCP: tcpip.TCPStats{ ActiveConnectionOpenings: mustCreateMetric("/netstack/tcp/active_connection_openings", "Number of connections opened successfully via Connect."), @@ -195,7 +198,6 @@ var Metrics = tcpip.Stats{ PacketsSent: mustCreateMetric("/netstack/udp/packets_sent", "Number of UDP datagrams sent."), PacketSendErrors: mustCreateMetric("/netstack/udp/packet_send_errors", "Number of UDP datagrams failed to be sent."), ChecksumErrors: mustCreateMetric("/netstack/udp/checksum_errors", "Number of UDP datagrams dropped due to bad checksums."), - InvalidSourceAddress: mustCreateMetric("/netstack/udp/invalid_source", "Number of UDP datagrams dropped due to invalid source address."), }, } @@ -236,7 +238,7 @@ type commonEndpoint interface { // SetSockOpt implements tcpip.Endpoint.SetSockOpt and // transport.Endpoint.SetSockOpt. - SetSockOpt(interface{}) *tcpip.Error + SetSockOpt(tcpip.SettableSocketOption) *tcpip.Error // SetSockOptBool implements tcpip.Endpoint.SetSockOptBool and // transport.Endpoint.SetSockOptBool. @@ -248,7 +250,7 @@ type commonEndpoint interface { // GetSockOpt implements tcpip.Endpoint.GetSockOpt and // transport.Endpoint.GetSockOpt. - GetSockOpt(interface{}) *tcpip.Error + GetSockOpt(tcpip.GettableSocketOption) *tcpip.Error // GetSockOptBool implements tcpip.Endpoint.GetSockOptBool and // transport.Endpoint.GetSockOpt. @@ -257,6 +259,9 @@ type commonEndpoint interface { // GetSockOptInt implements tcpip.Endpoint.GetSockOptInt and // transport.Endpoint.GetSockOpt. GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) + + // LastError implements tcpip.Endpoint.LastError. + LastError() *tcpip.Error } // LINT.IfChange @@ -479,8 +484,35 @@ func (s *socketOpsCommon) fetchReadView() *syserr.Error { } // Release implements fs.FileOperations.Release. -func (s *socketOpsCommon) Release(context.Context) { +func (s *socketOpsCommon) Release(ctx context.Context) { + e, ch := waiter.NewChannelEntry(nil) + s.EventRegister(&e, waiter.EventHUp|waiter.EventErr) + defer s.EventUnregister(&e) + s.Endpoint.Close() + + // SO_LINGER option is valid only for TCP. For other socket types + // return after endpoint close. + if family, skType, _ := s.Type(); skType != linux.SOCK_STREAM || (family != linux.AF_INET && family != linux.AF_INET6) { + return + } + + var v tcpip.LingerOption + if err := s.Endpoint.GetSockOpt(&v); err != nil { + return + } + + // The case for zero timeout is handled in tcp endpoint close function. + // Close is blocked until either: + // 1. The endpoint state is not in any of the states: FIN-WAIT1, + // CLOSING and LAST_ACK. + // 2. Timeout is reached. + if v.Enabled && v.Timeout != 0 { + t := kernel.TaskFromContext(ctx) + start := t.Kernel().MonotonicClock().Now() + deadline := start.Add(v.Timeout) + t.BlockWithDeadline(ch, true, deadline) + } } // Read implements fs.FileOperations.Read. @@ -555,6 +587,11 @@ func (i *ioSequencePayload) Payload(size int) ([]byte, *tcpip.Error) { } v := buffer.NewView(size) if _, err := i.src.CopyIn(i.ctx, v); err != nil { + // EOF can be returned only if src is a file and this means it + // is in a splice syscall and the error has to be ignored. + if err == io.EOF { + return v, nil + } return nil, tcpip.ErrBadAddress } return v, nil @@ -803,7 +840,20 @@ func (s *socketOpsCommon) Bind(t *kernel.Task, sockaddr []byte) *syserr.Error { } // Issue the bind request to the endpoint. - return syserr.TranslateNetstackError(s.Endpoint.Bind(addr)) + err := s.Endpoint.Bind(addr) + if err == tcpip.ErrNoPortAvailable { + // Bind always returns EADDRINUSE irrespective of if the specified port was + // already bound or if an ephemeral port was requested but none were + // available. + // + // tcpip.ErrNoPortAvailable is mapped to EAGAIN in syserr package because + // UDP connect returns EAGAIN on ephemeral port exhaustion. + // + // TCP connect returns EADDRNOTAVAIL on ephemeral port exhaustion. + err = tcpip.ErrPortInUse + } + + return syserr.TranslateNetstackError(err) } // Listen implements the linux syscall listen(2) for sockets backed by @@ -814,7 +864,7 @@ func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { // 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 *socketOpsCommon) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) { +func (s *socketOpsCommon) blockingAccept(t *kernel.Task, peerAddr *tcpip.FullAddress) (tcpip.Endpoint, *waiter.Queue, *syserr.Error) { // Register for notifications. e, ch := waiter.NewChannelEntry(nil) s.EventRegister(&e, waiter.EventIn) @@ -823,7 +873,7 @@ func (s *socketOpsCommon) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waite // 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 { + if ep, wq, err := s.Endpoint.Accept(peerAddr); err != tcpip.ErrWouldBlock { return ep, wq, syserr.TranslateNetstackError(err) } @@ -836,15 +886,18 @@ func (s *socketOpsCommon) blockingAccept(t *kernel.Task) (tcpip.Endpoint, *waite // 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) (int32, linux.SockAddr, uint32, *syserr.Error) { - // Issue the accept request to get the new endpoint. - ep, wq, terr := s.Endpoint.Accept() + var peerAddr *tcpip.FullAddress + if peerRequested { + peerAddr = &tcpip.FullAddress{} + } + ep, wq, terr := s.Endpoint.Accept(peerAddr) if terr != nil { if terr != tcpip.ErrWouldBlock || !blocking { return 0, nil, 0, syserr.TranslateNetstackError(terr) } var err *syserr.Error - ep, wq, err = s.blockingAccept(t) + ep, wq, err = s.blockingAccept(t, peerAddr) if err != nil { return 0, nil, 0, err } @@ -864,13 +917,8 @@ func (s *SocketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, var addr linux.SockAddr 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 - } + if peerAddr != nil { + addr, addrLen = ConvertAddress(s.family, *peerAddr) } fd, e := t.NewFDFrom(0, ns, kernel.FDFlags{ @@ -943,47 +991,12 @@ func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr us return &val, nil } - if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { - switch name { - case linux.IPT_SO_GET_INFO: - if outLen < linux.SizeOfIPTGetinfo { - return nil, syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return nil, syserr.ErrNoDevice - } - info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr) - if err != nil { - return nil, err - } - return &info, nil - - case linux.IPT_SO_GET_ENTRIES: - if outLen < linux.SizeOfIPTGetEntries { - return nil, syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return nil, syserr.ErrNoDevice - } - entries, err := netfilter.GetEntries(t, stack.(*Stack).Stack, outPtr, outLen) - if err != nil { - return nil, err - } - return &entries, nil - - } - } - - return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outLen) + return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outPtr, 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.SocketOps, ep commonEndpoint, family int, skType linux.SockType, level, name, outLen int) (marshal.Marshallable, *syserr.Error) { +func GetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family int, skType linux.SockType, level, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) { switch level { case linux.SOL_SOCKET: return getSockOptSocket(t, s, ep, family, skType, name, outLen) @@ -992,10 +1005,10 @@ func GetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, family in return getSockOptTCP(t, ep, name, outLen) case linux.SOL_IPV6: - return getSockOptIPv6(t, ep, name, outLen) + return getSockOptIPv6(t, s, ep, name, outPtr, outLen) case linux.SOL_IP: - return getSockOptIP(t, ep, name, outLen, family) + return getSockOptIP(t, s, ep, name, outPtr, outLen, family) case linux.SOL_UDP, linux.SOL_ICMPV6, @@ -1025,7 +1038,7 @@ func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, fam } // Get the last error and convert it. - err := ep.GetSockOpt(tcpip.ErrorOption{}) + err := ep.LastError() if err == nil { optP := primitive.Int32(0) return &optP, nil @@ -1176,7 +1189,16 @@ func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, fam return nil, syserr.ErrInvalidArgument } - linger := linux.Linger{} + var v tcpip.LingerOption + var linger linux.Linger + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + if v.Enabled { + linger.OnOff = 1 + } + linger.Linger = int32(v.Timeout.Seconds()) return &linger, nil case linux.SO_SNDTIMEO: @@ -1222,6 +1244,18 @@ func getSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, fam vP := primitive.Int32(boolToInt32(v)) return &vP, nil + case linux.SO_ACCEPTCONN: + if outLen < sizeOfInt32 { + return nil, syserr.ErrInvalidArgument + } + + v, err := ep.GetSockOptBool(tcpip.AcceptConnOption) + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + vP := primitive.Int32(boolToInt32(v)) + return &vP, nil + default: socket.GetSockOptEmitUnimplementedEvent(t, name) } @@ -1390,8 +1424,12 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (marshal if err := ep.GetSockOpt(&v); err != nil { return nil, syserr.TranslateNetstackError(err) } - - lingerTimeout := primitive.Int32(time.Duration(v) / time.Second) + var lingerTimeout primitive.Int32 + if v >= 0 { + lingerTimeout = primitive.Int32(time.Duration(v) / time.Second) + } else { + lingerTimeout = -1 + } return &lingerTimeout, nil case linux.TCP_DEFER_ACCEPT: @@ -1437,7 +1475,7 @@ func getSockOptTCP(t *kernel.Task, ep commonEndpoint, name, outLen int) (marshal } // getSockOptIPv6 implements GetSockOpt when level is SOL_IPV6. -func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (marshal.Marshallable, *syserr.Error) { +func getSockOptIPv6(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) { switch name { case linux.IPV6_V6ONLY: if outLen < sizeOfInt32 { @@ -1490,9 +1528,78 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (marsha vP := primitive.Int32(boolToInt32(v)) return &vP, nil - case linux.SO_ORIGINAL_DST: - // TODO(gvisor.dev/issue/170): ip6tables. - return nil, syserr.ErrInvalidArgument + case linux.IP6T_ORIGINAL_DST: + if outLen < int(binary.Size(linux.SockAddrInet6{})) { + return nil, syserr.ErrInvalidArgument + } + + var v tcpip.OriginalDestinationOption + if err := ep.GetSockOpt(&v); err != nil { + return nil, syserr.TranslateNetstackError(err) + } + + a, _ := ConvertAddress(linux.AF_INET6, tcpip.FullAddress(v)) + return a.(*linux.SockAddrInet6), nil + + case linux.IP6T_SO_GET_INFO: + if outLen < linux.SizeOfIPTGetinfo { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv6 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET6 || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr, true) + if err != nil { + return nil, err + } + return &info, nil + + case linux.IP6T_SO_GET_ENTRIES: + // IPTGetEntries is reused for IPv6. + if outLen < linux.SizeOfIPTGetEntries { + return nil, syserr.ErrInvalidArgument + } + // Only valid for raw IPv6 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET6 || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + entries, err := netfilter.GetEntries6(t, stack.(*Stack).Stack, outPtr, outLen) + if err != nil { + return nil, err + } + return &entries, nil + + case linux.IP6T_SO_GET_REVISION_TARGET: + if outLen < linux.SizeOfXTGetRevision { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv6 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET6 || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + ret, err := netfilter.TargetRevision(t, outPtr, header.IPv6ProtocolNumber) + if err != nil { + return nil, err + } + return &ret, nil default: emitUnimplementedEventIPv6(t, name) @@ -1501,7 +1608,7 @@ func getSockOptIPv6(t *kernel.Task, ep commonEndpoint, name, outLen int) (marsha } // getSockOptIP implements GetSockOpt when level is SOL_IP. -func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family int) (marshal.Marshallable, *syserr.Error) { +func getSockOptIP(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, outPtr usermem.Addr, outLen int, family int) (marshal.Marshallable, *syserr.Error) { switch name { case linux.IP_TTL: if outLen < sizeOfInt32 { @@ -1617,6 +1724,66 @@ func getSockOptIP(t *kernel.Task, ep commonEndpoint, name, outLen int, family in a, _ := ConvertAddress(linux.AF_INET, tcpip.FullAddress(v)) return a.(*linux.SockAddrInet), nil + case linux.IPT_SO_GET_INFO: + if outLen < linux.SizeOfIPTGetinfo { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv4 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr, false) + if err != nil { + return nil, err + } + return &info, nil + + case linux.IPT_SO_GET_ENTRIES: + if outLen < linux.SizeOfIPTGetEntries { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv4 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + entries, err := netfilter.GetEntries4(t, stack.(*Stack).Stack, outPtr, outLen) + if err != nil { + return nil, err + } + return &entries, nil + + case linux.IPT_SO_GET_REVISION_TARGET: + if outLen < linux.SizeOfXTGetRevision { + return nil, syserr.ErrInvalidArgument + } + + // Only valid for raw IPv4 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET || skType != linux.SOCK_RAW { + return nil, syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return nil, syserr.ErrNoDevice + } + ret, err := netfilter.TargetRevision(t, outPtr, header.IPv4ProtocolNumber) + if err != nil { + return nil, err + } + return &ret, nil + default: emitUnimplementedEventIP(t, name) } @@ -1650,26 +1817,6 @@ func (s *SocketOperations) SetSockOpt(t *kernel.Task, level int, name int, optVa return nil } - if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { - switch name { - case linux.IPT_SO_SET_REPLACE: - if len(optVal) < linux.SizeOfIPTReplace { - return syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return syserr.ErrNoDevice - } - // Stack must be a netstack stack. - return netfilter.SetEntries(stack.(*Stack).Stack, optVal) - - case linux.IPT_SO_SET_ADD_COUNTERS: - // TODO(gvisor.dev/issue/170): Counter support. - return nil - } - } - return SetSockOpt(t, s, s.Endpoint, level, name, optVal) } @@ -1684,21 +1831,26 @@ func SetSockOpt(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, level int return setSockOptTCP(t, ep, name, optVal) case linux.SOL_IPV6: - return setSockOptIPv6(t, ep, name, optVal) + return setSockOptIPv6(t, s, ep, name, optVal) case linux.SOL_IP: - return setSockOptIP(t, ep, name, optVal) + return setSockOptIP(t, s, ep, name, optVal) + + case linux.SOL_PACKET: + // gVisor doesn't support any SOL_PACKET options just return not + // supported. Returning nil here will result in tcpdump thinking AF_PACKET + // features are supported and proceed to use them and break. + t.Kernel().EmitUnimplementedEvent(t) + return syserr.ErrProtocolNotAvailable case linux.SOL_UDP, linux.SOL_ICMPV6, - linux.SOL_RAW, - linux.SOL_PACKET: + linux.SOL_RAW: t.Kernel().EmitUnimplementedEvent(t) } - // Default to the old behavior; hand off to network stack. - return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{})) + return nil } // setSockOptSocket implements SetSockOpt when level is SOL_SOCKET. @@ -1743,7 +1895,8 @@ func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, nam } name := string(optVal[:n]) if name == "" { - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.BindToDeviceOption(0))) + v := tcpip.BindToDeviceOption(0) + return syserr.TranslateNetstackError(ep.SetSockOpt(&v)) } s := t.NetworkContext() if s == nil { @@ -1751,7 +1904,8 @@ func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, nam } for nicID, nic := range s.Interfaces() { if nic.Name == name { - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.BindToDeviceOption(nicID))) + v := tcpip.BindToDeviceOption(nicID) + return syserr.TranslateNetstackError(ep.SetSockOpt(&v)) } } return syserr.ErrUnknownDevice @@ -1817,7 +1971,8 @@ func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, nam socket.SetSockOptEmitUnimplementedEvent(t, name) } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.OutOfBandInlineOption(v))) + opt := tcpip.OutOfBandInlineOption(v) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.SO_NO_CHECK: if len(optVal) < sizeOfInt32 { @@ -1839,19 +1994,21 @@ func setSockOptSocket(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, nam socket.SetSockOptEmitUnimplementedEvent(t, name) } - return nil + return syserr.TranslateNetstackError( + ep.SetSockOpt(&tcpip.LingerOption{ + Enabled: v.OnOff != 0, + Timeout: time.Second * time.Duration(v.Linger)})) case linux.SO_DETACH_FILTER: // optval is ignored. var v tcpip.SocketDetachFilterOption - return syserr.TranslateNetstackError(ep.SetSockOpt(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&v)) default: socket.SetSockOptEmitUnimplementedEvent(t, name) } - // Default to the old behavior; hand off to network stack. - return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{})) + return nil } // setSockOptTCP implements SetSockOpt when level is SOL_TCP. @@ -1898,7 +2055,8 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * if v < 1 || v > linux.MAX_TCP_KEEPIDLE { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.KeepaliveIdleOption(time.Second * time.Duration(v)))) + opt := tcpip.KeepaliveIdleOption(time.Second * time.Duration(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.TCP_KEEPINTVL: if len(optVal) < sizeOfInt32 { @@ -1909,7 +2067,8 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * if v < 1 || v > linux.MAX_TCP_KEEPINTVL { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.KeepaliveIntervalOption(time.Second * time.Duration(v)))) + opt := tcpip.KeepaliveIntervalOption(time.Second * time.Duration(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.TCP_KEEPCNT: if len(optVal) < sizeOfInt32 { @@ -1931,11 +2090,12 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * if v < 0 { return syserr.ErrInvalidArgument } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPUserTimeoutOption(time.Millisecond * time.Duration(v)))) + opt := tcpip.TCPUserTimeoutOption(time.Millisecond * time.Duration(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.TCP_CONGESTION: v := tcpip.CongestionControlOption(optVal) - if err := ep.SetSockOpt(v); err != nil { + if err := ep.SetSockOpt(&v); err != nil { return syserr.TranslateNetstackError(err) } return nil @@ -1945,8 +2105,9 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * return syserr.ErrInvalidArgument } - v := usermem.ByteOrder.Uint32(optVal) - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPLingerTimeoutOption(time.Second * time.Duration(v)))) + v := int32(usermem.ByteOrder.Uint32(optVal)) + opt := tcpip.TCPLingerTimeoutOption(time.Second * time.Duration(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.TCP_DEFER_ACCEPT: if len(optVal) < sizeOfInt32 { @@ -1956,7 +2117,8 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * if v < 0 { v = 0 } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.TCPDeferAcceptOption(time.Second * time.Duration(v)))) + opt := tcpip.TCPDeferAcceptOption(time.Second * time.Duration(v)) + return syserr.TranslateNetstackError(ep.SetSockOpt(&opt)) case linux.TCP_SYNCNT: if len(optVal) < sizeOfInt32 { @@ -1981,12 +2143,11 @@ func setSockOptTCP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) * emitUnimplementedEventTCP(t, name) } - // Default to the old behavior; hand off to network stack. - return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{})) + return nil } // setSockOptIPv6 implements SetSockOpt when level is SOL_IPV6. -func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *syserr.Error { +func setSockOptIPv6(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, optVal []byte) *syserr.Error { switch name { case linux.IPV6_V6ONLY: if len(optVal) < sizeOfInt32 { @@ -2035,12 +2196,32 @@ func setSockOptIPv6(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.ReceiveTClassOption, v != 0)) + case linux.IP6T_SO_SET_REPLACE: + if len(optVal) < linux.SizeOfIP6TReplace { + return syserr.ErrInvalidArgument + } + + // Only valid for raw IPv6 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET6 || skType != linux.SOCK_RAW { + return syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return syserr.ErrNoDevice + } + // Stack must be a netstack stack. + return netfilter.SetEntries(stack.(*Stack).Stack, optVal, true) + + case linux.IP6T_SO_SET_ADD_COUNTERS: + // TODO(gvisor.dev/issue/170): Counter support. + return nil + default: emitUnimplementedEventIPv6(t, name) } - // Default to the old behavior; hand off to network stack. - return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{})) + return nil } var ( @@ -2095,7 +2276,7 @@ func parseIntOrChar(buf []byte) (int32, *syserr.Error) { } // setSockOptIP implements SetSockOpt when level is SOL_IP. -func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *syserr.Error { +func setSockOptIP(t *kernel.Task, s socket.SocketOps, ep commonEndpoint, name int, optVal []byte) *syserr.Error { switch name { case linux.IP_MULTICAST_TTL: v, err := parseIntOrChar(optVal) @@ -2118,7 +2299,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s return err } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.AddMembershipOption{ + return syserr.TranslateNetstackError(ep.SetSockOpt(&tcpip.AddMembershipOption{ NIC: tcpip.NICID(req.InterfaceIndex), // TODO(igudger): Change AddMembership to use the standard // any address representation. @@ -2132,7 +2313,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s return err } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.RemoveMembershipOption{ + return syserr.TranslateNetstackError(ep.SetSockOpt(&tcpip.RemoveMembershipOption{ NIC: tcpip.NICID(req.InterfaceIndex), // TODO(igudger): Change DropMembership to use the standard // any address representation. @@ -2146,7 +2327,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s return err } - return syserr.TranslateNetstackError(ep.SetSockOpt(tcpip.MulticastInterfaceOption{ + return syserr.TranslateNetstackError(ep.SetSockOpt(&tcpip.MulticastInterfaceOption{ NIC: tcpip.NICID(req.InterfaceIndex), InterfaceAddr: bytesToIPAddress(req.InterfaceAddr[:]), })) @@ -2215,6 +2396,27 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s } return syserr.TranslateNetstackError(ep.SetSockOptBool(tcpip.IPHdrIncludedOption, v != 0)) + case linux.IPT_SO_SET_REPLACE: + if len(optVal) < linux.SizeOfIPTReplace { + return syserr.ErrInvalidArgument + } + + // Only valid for raw IPv4 sockets. + if family, skType, _ := s.Type(); family != linux.AF_INET || skType != linux.SOCK_RAW { + return syserr.ErrProtocolNotAvailable + } + + stack := inet.StackFromContext(t) + if stack == nil { + return syserr.ErrNoDevice + } + // Stack must be a netstack stack. + return netfilter.SetEntries(stack.(*Stack).Stack, optVal, false) + + case linux.IPT_SO_SET_ADD_COUNTERS: + // TODO(gvisor.dev/issue/170): Counter support. + return nil + case linux.IP_ADD_SOURCE_MEMBERSHIP, linux.IP_BIND_ADDRESS_NO_PORT, linux.IP_BLOCK_SOURCE, @@ -2249,8 +2451,7 @@ func setSockOptIP(t *kernel.Task, ep commonEndpoint, name int, optVal []byte) *s t.Kernel().EmitUnimplementedEvent(t) } - // Default to the old behavior; hand off to network stack. - return syserr.TranslateNetstackError(ep.SetSockOpt(struct{}{})) + return nil } // emitUnimplementedEventTCP emits unimplemented event if name is valid. This diff --git a/pkg/sentry/socket/netstack/netstack_vfs2.go b/pkg/sentry/socket/netstack/netstack_vfs2.go index 3335e7430..b0d9e4d9e 100644 --- a/pkg/sentry/socket/netstack/netstack_vfs2.go +++ b/pkg/sentry/socket/netstack/netstack_vfs2.go @@ -18,25 +18,25 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/amutex" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" - "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/socket" - "gvisor.dev/gvisor/pkg/sentry/socket/netfilter" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) // SocketVFS2 encapsulates all the state needed to represent a network stack // endpoint in the kernel context. +// +// +stateify savable type SocketVFS2 struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl @@ -57,7 +57,8 @@ func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, qu } mnt := t.Kernel().SocketMount() - d := sockfs.NewDentry(t.Credentials(), mnt) + d := sockfs.NewDentry(t, mnt) + defer d.DecRef(t) s := &SocketVFS2{ socketOpsCommon: socketOpsCommon{ @@ -80,6 +81,13 @@ func NewVFS2(t *kernel.Task, family int, skType linux.SockType, protocol int, qu return vfsfd, nil } +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.socketOpsCommon.Release(ctx) +} + // Readiness implements waiter.Waitable.Readiness. func (s *SocketVFS2) Readiness(mask waiter.EventMask) waiter.EventMask { return s.socketOpsCommon.Readiness(mask) @@ -152,14 +160,18 @@ func (s *SocketVFS2) Write(ctx context.Context, src usermem.IOSequence, opts vfs // tcpip.Endpoint. func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { // Issue the accept request to get the new endpoint. - ep, wq, terr := s.Endpoint.Accept() + var peerAddr *tcpip.FullAddress + if peerRequested { + peerAddr = &tcpip.FullAddress{} + } + ep, wq, terr := s.Endpoint.Accept(peerAddr) if terr != nil { if terr != tcpip.ErrWouldBlock || !blocking { return 0, nil, 0, syserr.TranslateNetstackError(terr) } var err *syserr.Error - ep, wq, err = s.blockingAccept(t) + ep, wq, err = s.blockingAccept(t, peerAddr) if err != nil { return 0, nil, 0, err } @@ -177,13 +189,9 @@ func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, block var addr linux.SockAddr var addrLen uint32 - if peerRequested { + if peerAddr != nil { // Get address of the peer and write it to peer slice. - var err *syserr.Error - addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t) - if err != nil { - return 0, nil, 0, err - } + addr, addrLen = ConvertAddress(s.family, *peerAddr) } fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{ @@ -233,42 +241,7 @@ func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem. return &val, nil } - if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { - switch name { - case linux.IPT_SO_GET_INFO: - if outLen < linux.SizeOfIPTGetinfo { - return nil, syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return nil, syserr.ErrNoDevice - } - info, err := netfilter.GetInfo(t, stack.(*Stack).Stack, outPtr) - if err != nil { - return nil, err - } - return &info, nil - - case linux.IPT_SO_GET_ENTRIES: - if outLen < linux.SizeOfIPTGetEntries { - return nil, syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return nil, syserr.ErrNoDevice - } - entries, err := netfilter.GetEntries(t, stack.(*Stack).Stack, outPtr, outLen) - if err != nil { - return nil, err - } - return &entries, nil - - } - } - - return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outLen) + return GetSockOpt(t, s, s.Endpoint, s.family, s.skType, level, name, outPtr, outLen) } // SetSockOpt implements the linux syscall setsockopt(2) for sockets backed by @@ -298,26 +271,6 @@ func (s *SocketVFS2) SetSockOpt(t *kernel.Task, level int, name int, optVal []by return nil } - if s.skType == linux.SOCK_RAW && level == linux.IPPROTO_IP { - switch name { - case linux.IPT_SO_SET_REPLACE: - if len(optVal) < linux.SizeOfIPTReplace { - return syserr.ErrInvalidArgument - } - - stack := inet.StackFromContext(t) - if stack == nil { - return syserr.ErrNoDevice - } - // Stack must be a netstack stack. - return netfilter.SetEntries(stack.(*Stack).Stack, optVal) - - case linux.IPT_SO_SET_ADD_COUNTERS: - // TODO(gvisor.dev/issue/170): Counter support. - return nil - } - } - return SetSockOpt(t, s, s.Endpoint, level, name, optVal) } diff --git a/pkg/sentry/socket/netstack/stack.go b/pkg/sentry/socket/netstack/stack.go index f0fe18684..fa9ac9059 100644 --- a/pkg/sentry/socket/netstack/stack.go +++ b/pkg/sentry/socket/netstack/stack.go @@ -100,62 +100,107 @@ func (s *Stack) InterfaceAddrs() map[int32][]inet.InterfaceAddr { return nicAddrs } -// AddInterfaceAddr implements inet.Stack.AddInterfaceAddr. -func (s *Stack) AddInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { +// convertAddr converts an InterfaceAddr to a ProtocolAddress. +func convertAddr(addr inet.InterfaceAddr) (tcpip.ProtocolAddress, error) { var ( - protocol tcpip.NetworkProtocolNumber - address tcpip.Address + protocol tcpip.NetworkProtocolNumber + address tcpip.Address + protocolAddress tcpip.ProtocolAddress ) switch addr.Family { case linux.AF_INET: - if len(addr.Addr) < header.IPv4AddressSize { - return syserror.EINVAL + if len(addr.Addr) != header.IPv4AddressSize { + return protocolAddress, syserror.EINVAL } if addr.PrefixLen > header.IPv4AddressSize*8 { - return syserror.EINVAL + return protocolAddress, syserror.EINVAL } protocol = ipv4.ProtocolNumber - address = tcpip.Address(addr.Addr[:header.IPv4AddressSize]) - + address = tcpip.Address(addr.Addr) case linux.AF_INET6: - if len(addr.Addr) < header.IPv6AddressSize { - return syserror.EINVAL + if len(addr.Addr) != header.IPv6AddressSize { + return protocolAddress, syserror.EINVAL } if addr.PrefixLen > header.IPv6AddressSize*8 { - return syserror.EINVAL + return protocolAddress, syserror.EINVAL } protocol = ipv6.ProtocolNumber - address = tcpip.Address(addr.Addr[:header.IPv6AddressSize]) - + address = tcpip.Address(addr.Addr) default: - return syserror.ENOTSUP + return protocolAddress, syserror.ENOTSUP } - protocolAddress := tcpip.ProtocolAddress{ + protocolAddress = tcpip.ProtocolAddress{ Protocol: protocol, AddressWithPrefix: tcpip.AddressWithPrefix{ Address: address, PrefixLen: int(addr.PrefixLen), }, } + return protocolAddress, nil +} + +// AddInterfaceAddr implements inet.Stack.AddInterfaceAddr. +func (s *Stack) AddInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { + protocolAddress, err := convertAddr(addr) + if err != nil { + return err + } // Attach address to interface. - if err := s.Stack.AddProtocolAddressWithOptions(tcpip.NICID(idx), protocolAddress, stack.CanBePrimaryEndpoint); err != nil { + nicID := tcpip.NICID(idx) + if err := s.Stack.AddProtocolAddressWithOptions(nicID, protocolAddress, stack.CanBePrimaryEndpoint); err != nil { return syserr.TranslateNetstackError(err).ToError() } - // Add route for local network. - s.Stack.AddRoute(tcpip.Route{ + // Add route for local network if it doesn't exist already. + localRoute := tcpip.Route{ Destination: protocolAddress.AddressWithPrefix.Subnet(), Gateway: "", // No gateway for local network. - NIC: tcpip.NICID(idx), + NIC: nicID, + } + + for _, rt := range s.Stack.GetRouteTable() { + if rt.Equal(localRoute) { + return nil + } + } + + // Local route does not exist yet. Add it. + s.Stack.AddRoute(localRoute) + + return nil +} + +// RemoveInterfaceAddr implements inet.Stack.RemoveInterfaceAddr. +func (s *Stack) RemoveInterfaceAddr(idx int32, addr inet.InterfaceAddr) error { + protocolAddress, err := convertAddr(addr) + if err != nil { + return err + } + + // Remove addresses matching the address and prefix. + nicID := tcpip.NICID(idx) + if err := s.Stack.RemoveAddress(nicID, protocolAddress.AddressWithPrefix.Address); err != nil { + return syserr.TranslateNetstackError(err).ToError() + } + + // Remove the corresponding local network route if it exists. + localRoute := tcpip.Route{ + Destination: protocolAddress.AddressWithPrefix.Subnet(), + Gateway: "", // No gateway for local network. + NIC: nicID, + } + s.Stack.RemoveRoutes(func(rt tcpip.Route) bool { + return rt.Equal(localRoute) }) + return nil } // TCPReceiveBufferSize implements inet.Stack.TCPReceiveBufferSize. func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) { - var rs tcp.ReceiveBufferSizeOption + var rs tcpip.TCPReceiveBufferSizeRangeOption err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &rs) return inet.TCPBufferSize{ Min: rs.Min, @@ -166,17 +211,17 @@ func (s *Stack) TCPReceiveBufferSize() (inet.TCPBufferSize, error) { // SetTCPReceiveBufferSize implements inet.Stack.SetTCPReceiveBufferSize. func (s *Stack) SetTCPReceiveBufferSize(size inet.TCPBufferSize) error { - rs := tcp.ReceiveBufferSizeOption{ + rs := tcpip.TCPReceiveBufferSizeRangeOption{ Min: size.Min, Default: size.Default, Max: size.Max, } - return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, rs)).ToError() + 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 + var ss tcpip.TCPSendBufferSizeRangeOption err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &ss) return inet.TCPBufferSize{ Min: ss.Min, @@ -187,29 +232,30 @@ func (s *Stack) TCPSendBufferSize() (inet.TCPBufferSize, error) { // SetTCPSendBufferSize implements inet.Stack.SetTCPSendBufferSize. func (s *Stack) SetTCPSendBufferSize(size inet.TCPBufferSize) error { - ss := tcp.SendBufferSizeOption{ + ss := tcpip.TCPSendBufferSizeRangeOption{ Min: size.Min, Default: size.Default, Max: size.Max, } - return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, ss)).ToError() + 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 + var sack tcpip.TCPSACKEnabled 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() + opt := tcpip.TCPSACKEnabled(enabled) + return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, &opt)).ToError() } // TCPRecovery implements inet.Stack.TCPRecovery. func (s *Stack) TCPRecovery() (inet.TCPLossRecovery, error) { - var recovery tcp.Recovery + var recovery tcpip.TCPRecovery if err := s.Stack.TransportProtocolOption(tcp.ProtocolNumber, &recovery); err != nil { return 0, syserr.TranslateNetstackError(err).ToError() } @@ -218,7 +264,8 @@ func (s *Stack) TCPRecovery() (inet.TCPLossRecovery, error) { // SetTCPRecovery implements inet.Stack.SetTCPRecovery. func (s *Stack) SetTCPRecovery(recovery inet.TCPLossRecovery) error { - return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, tcp.Recovery(recovery))).ToError() + opt := tcpip.TCPRecovery(recovery) + return syserr.TranslateNetstackError(s.Stack.SetTransportProtocolOption(tcp.ProtocolNumber, &opt)).ToError() } // Statistics implements inet.Stack.Statistics. @@ -410,3 +457,24 @@ func (s *Stack) CleanupEndpoints() []stack.TransportEndpoint { func (s *Stack) RestoreCleanupEndpoints(es []stack.TransportEndpoint) { s.Stack.RestoreCleanupEndpoints(es) } + +// Forwarding implements inet.Stack.Forwarding. +func (s *Stack) Forwarding(protocol tcpip.NetworkProtocolNumber) bool { + switch protocol { + case ipv4.ProtocolNumber, ipv6.ProtocolNumber: + return s.Stack.Forwarding(protocol) + default: + panic(fmt.Sprintf("Forwarding(%v) failed: unsupported protocol", protocol)) + } +} + +// SetForwarding implements inet.Stack.SetForwarding. +func (s *Stack) SetForwarding(protocol tcpip.NetworkProtocolNumber, enable bool) error { + switch protocol { + case ipv4.ProtocolNumber, ipv6.ProtocolNumber: + s.Stack.SetForwarding(protocol, enable) + default: + panic(fmt.Sprintf("SetForwarding(%v) failed: unsupported protocol", protocol)) + } + return nil +} diff --git a/pkg/sentry/socket/socket.go b/pkg/sentry/socket/socket.go index 04b259d27..fd31479e5 100644 --- a/pkg/sentry/socket/socket.go +++ b/pkg/sentry/socket/socket.go @@ -25,6 +25,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/device" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" @@ -35,7 +36,6 @@ import ( "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) // ControlMessages represents the union of unix control messages and tcpip diff --git a/pkg/sentry/socket/unix/BUILD b/pkg/sentry/socket/unix/BUILD index cb953e4dc..cce0acc33 100644 --- a/pkg/sentry/socket/unix/BUILD +++ b/pkg/sentry/socket/unix/BUILD @@ -7,10 +7,21 @@ go_template_instance( name = "socket_refs", out = "socket_refs.go", package = "unix", - prefix = "socketOpsCommon", - template = "//pkg/refs_vfs2:refs_template", + prefix = "socketOperations", + template = "//pkg/refsvfs2:refs_template", types = { - "T": "socketOpsCommon", + "T": "SocketOperations", + }, +) + +go_template_instance( + name = "socket_vfs2_refs", + out = "socket_vfs2_refs.go", + package = "unix", + prefix = "socketVFS2", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "SocketVFS2", }, ) @@ -20,6 +31,7 @@ go_library( "device.go", "io.go", "socket_refs.go", + "socket_vfs2_refs.go", "unix.go", "unix_vfs2.go", ], @@ -29,7 +41,9 @@ go_library( "//pkg/context", "//pkg/fspath", "//pkg/log", + "//pkg/marshal", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/device", @@ -49,6 +63,5 @@ go_library( "//pkg/tcpip", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", ], ) diff --git a/pkg/sentry/socket/unix/transport/BUILD b/pkg/sentry/socket/unix/transport/BUILD index c708b6030..3ebbd28b0 100644 --- a/pkg/sentry/socket/unix/transport/BUILD +++ b/pkg/sentry/socket/unix/transport/BUILD @@ -15,6 +15,17 @@ go_template_instance( }, ) +go_template_instance( + name = "queue_refs", + out = "queue_refs.go", + package = "transport", + prefix = "queue", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "queue", + }, +) + go_library( name = "transport", srcs = [ @@ -22,6 +33,7 @@ go_library( "connectioned_state.go", "connectionless.go", "queue.go", + "queue_refs.go", "transport_message_list.go", "unix.go", ], @@ -32,6 +44,7 @@ go_library( "//pkg/ilist", "//pkg/log", "//pkg/refs", + "//pkg/refsvfs2", "//pkg/sync", "//pkg/syserr", "//pkg/tcpip", diff --git a/pkg/sentry/socket/unix/transport/connectioned.go b/pkg/sentry/socket/unix/transport/connectioned.go index c67b602f0..aa4f3c04d 100644 --- a/pkg/sentry/socket/unix/transport/connectioned.go +++ b/pkg/sentry/socket/unix/transport/connectioned.go @@ -142,9 +142,9 @@ func NewPair(ctx context.Context, stype linux.SockType, uid UniqueIDProvider) (E } q1 := &queue{ReaderQueue: a.Queue, WriterQueue: b.Queue, limit: initialLimit} - q1.EnableLeakCheck("transport.queue") + q1.EnableLeakCheck() q2 := &queue{ReaderQueue: b.Queue, WriterQueue: a.Queue, limit: initialLimit} - q2.EnableLeakCheck("transport.queue") + q2.EnableLeakCheck() if stype == linux.SOCK_STREAM { a.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{q1}} @@ -300,14 +300,14 @@ func (e *connectionedEndpoint) BidirectionalConnect(ctx context.Context, ce Conn } readQueue := &queue{ReaderQueue: ce.WaiterQueue(), WriterQueue: ne.Queue, limit: initialLimit} - readQueue.EnableLeakCheck("transport.queue") + readQueue.EnableLeakCheck() ne.connected = &connectedEndpoint{ endpoint: ce, writeQueue: readQueue, } writeQueue := &queue{ReaderQueue: ne.Queue, WriterQueue: ce.WaiterQueue(), limit: initialLimit} - writeQueue.EnableLeakCheck("transport.queue") + writeQueue.EnableLeakCheck() if e.stype == linux.SOCK_STREAM { ne.receiver = &streamQueueReceiver{queueReceiver: queueReceiver{readQueue: writeQueue}} } else { @@ -391,7 +391,7 @@ func (e *connectionedEndpoint) Listen(backlog int) *syserr.Error { } // Accept accepts a new connection. -func (e *connectionedEndpoint) Accept() (Endpoint, *syserr.Error) { +func (e *connectionedEndpoint) Accept(peerAddr *tcpip.FullAddress) (Endpoint, *syserr.Error) { e.Lock() defer e.Unlock() @@ -401,6 +401,18 @@ func (e *connectionedEndpoint) Accept() (Endpoint, *syserr.Error) { select { case ne := <-e.acceptedChan: + if peerAddr != nil { + ne.Lock() + c := ne.connected + ne.Unlock() + if c != nil { + addr, err := c.GetLocalAddress() + if err != nil { + return nil, syserr.TranslateNetstackError(err) + } + *peerAddr = addr + } + } return ne, nil default: diff --git a/pkg/sentry/socket/unix/transport/connectionless.go b/pkg/sentry/socket/unix/transport/connectionless.go index 70ee8f9b8..f8aacca13 100644 --- a/pkg/sentry/socket/unix/transport/connectionless.go +++ b/pkg/sentry/socket/unix/transport/connectionless.go @@ -42,7 +42,7 @@ var ( func NewConnectionless(ctx context.Context) Endpoint { ep := &connectionlessEndpoint{baseEndpoint{Queue: &waiter.Queue{}}} q := queue{ReaderQueue: ep.Queue, WriterQueue: &waiter.Queue{}, limit: initialLimit} - q.EnableLeakCheck("transport.queue") + q.EnableLeakCheck() ep.receiver = &queueReceiver{readQueue: &q} return ep } @@ -144,12 +144,12 @@ func (e *connectionlessEndpoint) Connect(ctx context.Context, server BoundEndpoi } // Listen starts listening on the connection. -func (e *connectionlessEndpoint) Listen(int) *syserr.Error { +func (*connectionlessEndpoint) Listen(int) *syserr.Error { return syserr.ErrNotSupported } // Accept accepts a new connection. -func (e *connectionlessEndpoint) Accept() (Endpoint, *syserr.Error) { +func (*connectionlessEndpoint) Accept(*tcpip.FullAddress) (Endpoint, *syserr.Error) { return nil, syserr.ErrNotSupported } diff --git a/pkg/sentry/socket/unix/transport/queue.go b/pkg/sentry/socket/unix/transport/queue.go index ef6043e19..342def28f 100644 --- a/pkg/sentry/socket/unix/transport/queue.go +++ b/pkg/sentry/socket/unix/transport/queue.go @@ -16,7 +16,6 @@ package transport import ( "gvisor.dev/gvisor/pkg/context" - "gvisor.dev/gvisor/pkg/refs" "gvisor.dev/gvisor/pkg/sync" "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/tcpip" @@ -28,7 +27,7 @@ import ( // // +stateify savable type queue struct { - refs.AtomicRefCount + queueRefs ReaderQueue *waiter.Queue WriterQueue *waiter.Queue @@ -68,11 +67,13 @@ func (q *queue) Reset(ctx context.Context) { q.mu.Unlock() } -// DecRef implements RefCounter.DecRef with destructor q.Reset. +// DecRef implements RefCounter.DecRef. func (q *queue) DecRef(ctx context.Context) { - q.DecRefWithDestructor(ctx, q.Reset) - // We don't need to notify after resetting because no one cares about - // this queue after all references have been dropped. + q.queueRefs.DecRef(func() { + // We don't need to notify after resetting because no one cares about + // this queue after all references have been dropped. + q.Reset(ctx) + }) } // IsReadable determines if q is currently readable. diff --git a/pkg/sentry/socket/unix/transport/unix.go b/pkg/sentry/socket/unix/transport/unix.go index 475d7177e..b648273a4 100644 --- a/pkg/sentry/socket/unix/transport/unix.go +++ b/pkg/sentry/socket/unix/transport/unix.go @@ -32,6 +32,8 @@ import ( const initialLimit = 16 * 1024 // A RightsControlMessage is a control message containing FDs. +// +// +stateify savable type RightsControlMessage interface { // Clone returns a copy of the RightsControlMessage. Clone() RightsControlMessage @@ -151,7 +153,10 @@ type Endpoint interface { // block if no new connections are available. // // The returned Queue is the wait queue for the newly created endpoint. - Accept() (Endpoint, *syserr.Error) + // + // peerAddr if not nil will be populated with the address of the connected + // peer on a successful accept. + Accept(peerAddr *tcpip.FullAddress) (Endpoint, *syserr.Error) // Bind binds the endpoint to a specific local address and port. // Specifying a NIC is optional. @@ -172,9 +177,8 @@ type Endpoint interface { // connected. GetRemoteAddress() (tcpip.FullAddress, *tcpip.Error) - // SetSockOpt sets a socket option. opt should be one of the tcpip.*Option - // types. - SetSockOpt(opt interface{}) *tcpip.Error + // SetSockOpt sets a socket option. + SetSockOpt(opt tcpip.SettableSocketOption) *tcpip.Error // SetSockOptBool sets a socket option for simple cases when a value has // the int type. @@ -184,9 +188,8 @@ type Endpoint interface { // the int type. SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error - // GetSockOpt gets a socket option. opt should be a pointer to one of the - // tcpip.*Option types. - GetSockOpt(opt interface{}) *tcpip.Error + // GetSockOpt gets a socket option. + GetSockOpt(opt tcpip.GettableSocketOption) *tcpip.Error // GetSockOptBool gets a socket option for simple cases when a return // value has the int type. @@ -199,6 +202,9 @@ type Endpoint interface { // State returns the current state of the socket, as represented by Linux in // procfs. State() uint32 + + // LastError implements tcpip.Endpoint.LastError. + LastError() *tcpip.Error } // A Credentialer is a socket or endpoint that supports the SO_PASSCRED socket @@ -332,7 +338,7 @@ type Receiver interface { RecvMaxQueueSize() int64 // Release releases any resources owned by the Receiver. It should be - // called before droping all references to a Receiver. + // called before dropping all references to a Receiver. Release(ctx context.Context) } @@ -483,7 +489,7 @@ func (q *streamQueueReceiver) Recv(ctx context.Context, data [][]byte, wantCreds c := q.control.Clone() // Don't consume data since we are peeking. - copied, data, _ = vecCopy(data, q.buffer) + copied, _, _ = vecCopy(data, q.buffer) return copied, copied, c, false, q.addr, notify, nil } @@ -568,6 +574,12 @@ func (q *streamQueueReceiver) Recv(ctx context.Context, data [][]byte, wantCreds return copied, copied, c, cmTruncated, q.addr, notify, nil } +// Release implements Receiver.Release. +func (q *streamQueueReceiver) Release(ctx context.Context) { + q.queueReceiver.Release(ctx) + q.control.Release(ctx) +} + // A ConnectedEndpoint is an Endpoint that can be used to send Messages. type ConnectedEndpoint interface { // Passcred implements Endpoint.Passcred. @@ -615,7 +627,7 @@ type ConnectedEndpoint interface { SendMaxQueueSize() int64 // Release releases any resources owned by the ConnectedEndpoint. It should - // be called before droping all references to a ConnectedEndpoint. + // be called before dropping all references to a ConnectedEndpoint. Release(ctx context.Context) // CloseUnread sets the fact that this end is closed with unread data to @@ -742,6 +754,9 @@ type baseEndpoint struct { // path is not empty if the endpoint has been bound, // or may be used if the endpoint is connected. path string + + // linger is used for SO_LINGER socket option. + linger tcpip.LingerOption } // EventRegister implements waiter.Waitable.EventRegister. @@ -837,8 +852,14 @@ func (e *baseEndpoint) SendMsg(ctx context.Context, data [][]byte, c ControlMess return n, err } -// SetSockOpt sets a socket option. Currently not supported. -func (e *baseEndpoint) SetSockOpt(opt interface{}) *tcpip.Error { +// SetSockOpt sets a socket option. +func (e *baseEndpoint) SetSockOpt(opt tcpip.SettableSocketOption) *tcpip.Error { + switch v := opt.(type) { + case *tcpip.LingerOption: + e.Lock() + e.linger = *v + e.Unlock() + } return nil } @@ -866,7 +887,7 @@ func (e *baseEndpoint) SetSockOptInt(opt tcpip.SockOptInt, v int) *tcpip.Error { func (e *baseEndpoint) GetSockOptBool(opt tcpip.SockOptBool) (bool, *tcpip.Error) { switch opt { - case tcpip.KeepaliveEnabledOption: + case tcpip.KeepaliveEnabledOption, tcpip.AcceptConnOption: return false, nil case tcpip.PasscredOption: @@ -940,9 +961,12 @@ func (e *baseEndpoint) GetSockOptInt(opt tcpip.SockOptInt) (int, *tcpip.Error) { } // GetSockOpt implements tcpip.Endpoint.GetSockOpt. -func (e *baseEndpoint) GetSockOpt(opt interface{}) *tcpip.Error { - switch opt.(type) { - case tcpip.ErrorOption: +func (e *baseEndpoint) GetSockOpt(opt tcpip.GettableSocketOption) *tcpip.Error { + switch o := opt.(type) { + case *tcpip.LingerOption: + e.Lock() + *o = e.linger + e.Unlock() return nil default: @@ -951,6 +975,11 @@ func (e *baseEndpoint) GetSockOpt(opt interface{}) *tcpip.Error { } } +// LastError implements Endpoint.LastError. +func (*baseEndpoint) LastError() *tcpip.Error { + return nil +} + // Shutdown closes the read and/or write end of the endpoint connection to its // peer. func (e *baseEndpoint) Shutdown(flags tcpip.ShutdownFlags) *syserr.Error { diff --git a/pkg/sentry/socket/unix/unix.go b/pkg/sentry/socket/unix/unix.go index b7e8e4325..adad485a9 100644 --- a/pkg/sentry/socket/unix/unix.go +++ b/pkg/sentry/socket/unix/unix.go @@ -24,6 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/fsutil" @@ -39,7 +40,6 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) // SocketOperations is a Unix socket. It is similar to a netstack socket, @@ -55,6 +55,7 @@ type SocketOperations struct { fsutil.FileNoopFlush `state:"nosave"` fsutil.FileUseInodeUnstableAttr `state:"nosave"` + socketOperationsRefs socketOpsCommon } @@ -80,15 +81,30 @@ func NewWithDirent(ctx context.Context, d *fs.Dirent, ep transport.Endpoint, sty }, } s.EnableLeakCheck() - return fs.NewFile(ctx, d, flags, &s) } +// DecRef implements RefCounter.DecRef. +func (s *SocketOperations) DecRef(ctx context.Context) { + s.socketOperationsRefs.DecRef(func() { + s.ep.Close(ctx) + if s.abstractNamespace != nil { + s.abstractNamespace.Remove(s.abstractName, s) + } + }) +} + +// Release implemements fs.FileOperations.Release. +func (s *SocketOperations) Release(ctx context.Context) { + // Release only decrements a reference on s because s may be referenced in + // the abstract socket namespace. + s.DecRef(ctx) +} + // socketOpsCommon contains the socket operations common to VFS1 and VFS2. // // +stateify savable type socketOpsCommon struct { - socketOpsCommonRefs socket.SendReceiveTimeout ep transport.Endpoint @@ -101,23 +117,6 @@ type socketOpsCommon struct { abstractNamespace *kernel.AbstractSocketNamespace } -// DecRef implements RefCounter.DecRef. -func (s *socketOpsCommon) DecRef(ctx context.Context) { - s.socketOpsCommonRefs.DecRef(func() { - s.ep.Close(ctx) - if s.abstractNamespace != nil { - s.abstractNamespace.Remove(s.abstractName, s) - } - }) -} - -// Release implemements fs.FileOperations.Release. -func (s *socketOpsCommon) Release(ctx context.Context) { - // Release only decrements a reference on s because s may be referenced in - // the abstract socket namespace. - s.DecRef(ctx) -} - func (s *socketOpsCommon) isPacket() bool { switch s.stype { case linux.SOCK_DGRAM, linux.SOCK_SEQPACKET: @@ -194,7 +193,7 @@ func (s *SocketOperations) Ioctl(ctx context.Context, _ *fs.File, io usermem.IO, // GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by // a transport.Endpoint. func (s *SocketOperations) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) { - return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outLen) + return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outPtr, outLen) } // Listen implements the linux syscall listen(2) for sockets backed by @@ -205,7 +204,7 @@ func (s *socketOpsCommon) Listen(t *kernel.Task, backlog int) *syserr.Error { // 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) (transport.Endpoint, *syserr.Error) { +func (s *SocketOperations) blockingAccept(t *kernel.Task, peerAddr *tcpip.FullAddress) (transport.Endpoint, *syserr.Error) { // Register for notifications. e, ch := waiter.NewChannelEntry(nil) s.EventRegister(&e, waiter.EventIn) @@ -214,7 +213,7 @@ func (s *SocketOperations) blockingAccept(t *kernel.Task) (transport.Endpoint, * // Try to accept the connection; if it fails, then wait until we get a // notification. for { - if ep, err := s.ep.Accept(); err != syserr.ErrWouldBlock { + if ep, err := s.ep.Accept(peerAddr); err != syserr.ErrWouldBlock { return ep, err } @@ -227,15 +226,18 @@ func (s *SocketOperations) blockingAccept(t *kernel.Task) (transport.Endpoint, * // Accept implements the linux syscall accept(2) for sockets backed by // a transport.Endpoint. func (s *SocketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { - // Issue the accept request to get the new endpoint. - ep, err := s.ep.Accept() + var peerAddr *tcpip.FullAddress + if peerRequested { + peerAddr = &tcpip.FullAddress{} + } + ep, err := s.ep.Accept(peerAddr) if err != nil { if err != syserr.ErrWouldBlock || !blocking { return 0, nil, 0, err } var err *syserr.Error - ep, err = s.blockingAccept(t) + ep, err = s.blockingAccept(t, peerAddr) if err != nil { return 0, nil, 0, err } @@ -252,13 +254,8 @@ func (s *SocketOperations) Accept(t *kernel.Task, peerRequested bool, flags int, var addr linux.SockAddr 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 - } + if peerAddr != nil { + addr, addrLen = netstack.ConvertAddress(linux.AF_UNIX, *peerAddr) } fd, e := t.NewFDFrom(0, ns, kernel.FDFlags{ @@ -575,13 +572,17 @@ func (s *SocketOperations) Read(ctx context.Context, _ *fs.File, dst usermem.IOS if dst.NumBytes() == 0 { return 0, nil } - return dst.CopyOutFrom(ctx, &EndpointReader{ + r := &EndpointReader{ Ctx: ctx, Endpoint: s.ep, NumRights: 0, Peek: false, From: nil, - }) + } + n, err := dst.CopyOutFrom(ctx, r) + // Drop control messages. + r.Control.Release(ctx) + return n, err } // RecvMsg implements the linux syscall recvmsg(2) for sockets backed by diff --git a/pkg/sentry/socket/unix/unix_vfs2.go b/pkg/sentry/socket/unix/unix_vfs2.go index d066ef8ab..7a78444dc 100644 --- a/pkg/sentry/socket/unix/unix_vfs2.go +++ b/pkg/sentry/socket/unix/unix_vfs2.go @@ -18,6 +18,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/arch" fslock "gvisor.dev/gvisor/pkg/sentry/fs/lock" "gvisor.dev/gvisor/pkg/sentry/fsimpl/sockfs" @@ -32,17 +33,19 @@ import ( "gvisor.dev/gvisor/pkg/tcpip" "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" - "gvisor.dev/gvisor/tools/go_marshal/marshal" ) // SocketVFS2 implements socket.SocketVFS2 (and by extension, // vfs.FileDescriptionImpl) for Unix sockets. +// +// +stateify savable type SocketVFS2 struct { vfsfd vfs.FileDescription vfs.FileDescriptionDefaultImpl vfs.DentryMetadataFileDescriptionImpl vfs.LockFD + socketVFS2Refs socketOpsCommon } @@ -52,7 +55,8 @@ var _ = socket.SocketVFS2(&SocketVFS2{}) // returns a corresponding file description. func NewSockfsFile(t *kernel.Task, ep transport.Endpoint, stype linux.SockType) (*vfs.FileDescription, *syserr.Error) { mnt := t.Kernel().SocketMount() - d := sockfs.NewDentry(t.Credentials(), mnt) + d := sockfs.NewDentry(t, mnt) + defer d.DecRef(t) fd, err := NewFileDescription(ep, stype, linux.O_RDWR, mnt, d, &vfs.FileLocks{}) if err != nil { @@ -76,6 +80,7 @@ func NewFileDescription(ep transport.Endpoint, stype linux.SockType, flags uint3 stype: stype, }, } + sock.EnableLeakCheck() sock.LockFD.Init(locks) vfsfd := &sock.vfsfd if err := vfsfd.Init(sock, flags, mnt, d, &vfs.FileDescriptionOptions{ @@ -88,15 +93,34 @@ func NewFileDescription(ep transport.Endpoint, stype linux.SockType, flags uint3 return vfsfd, nil } +// DecRef implements RefCounter.DecRef. +func (s *SocketVFS2) DecRef(ctx context.Context) { + s.socketVFS2Refs.DecRef(func() { + t := kernel.TaskFromContext(ctx) + t.Kernel().DeleteSocketVFS2(&s.vfsfd) + s.ep.Close(ctx) + if s.abstractNamespace != nil { + s.abstractNamespace.Remove(s.abstractName, s) + } + }) +} + +// Release implements vfs.FileDescriptionImpl.Release. +func (s *SocketVFS2) Release(ctx context.Context) { + // Release only decrements a reference on s because s may be referenced in + // the abstract socket namespace. + s.DecRef(ctx) +} + // GetSockOpt implements the linux syscall getsockopt(2) for sockets backed by // a transport.Endpoint. func (s *SocketVFS2) GetSockOpt(t *kernel.Task, level, name int, outPtr usermem.Addr, outLen int) (marshal.Marshallable, *syserr.Error) { - return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outLen) + return netstack.GetSockOpt(t, s, s.ep, linux.AF_UNIX, s.ep.Type(), level, name, outPtr, outLen) } // 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 *SocketVFS2) blockingAccept(t *kernel.Task) (transport.Endpoint, *syserr.Error) { +func (s *SocketVFS2) blockingAccept(t *kernel.Task, peerAddr *tcpip.FullAddress) (transport.Endpoint, *syserr.Error) { // Register for notifications. e, ch := waiter.NewChannelEntry(nil) s.socketOpsCommon.EventRegister(&e, waiter.EventIn) @@ -105,7 +129,7 @@ func (s *SocketVFS2) blockingAccept(t *kernel.Task) (transport.Endpoint, *syserr // Try to accept the connection; if it fails, then wait until we get a // notification. for { - if ep, err := s.ep.Accept(); err != syserr.ErrWouldBlock { + if ep, err := s.ep.Accept(peerAddr); err != syserr.ErrWouldBlock { return ep, err } @@ -118,15 +142,18 @@ func (s *SocketVFS2) blockingAccept(t *kernel.Task) (transport.Endpoint, *syserr // Accept implements the linux syscall accept(2) for sockets backed by // a transport.Endpoint. func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, blocking bool) (int32, linux.SockAddr, uint32, *syserr.Error) { - // Issue the accept request to get the new endpoint. - ep, err := s.ep.Accept() + var peerAddr *tcpip.FullAddress + if peerRequested { + peerAddr = &tcpip.FullAddress{} + } + ep, err := s.ep.Accept(peerAddr) if err != nil { if err != syserr.ErrWouldBlock || !blocking { return 0, nil, 0, err } var err *syserr.Error - ep, err = s.blockingAccept(t) + ep, err = s.blockingAccept(t, peerAddr) if err != nil { return 0, nil, 0, err } @@ -144,13 +171,8 @@ func (s *SocketVFS2) Accept(t *kernel.Task, peerRequested bool, flags int, block var addr linux.SockAddr var addrLen uint32 - if peerRequested { - // Get address of the peer. - var err *syserr.Error - addr, addrLen, err = ns.Impl().(*SocketVFS2).GetPeerName(t) - if err != nil { - return 0, nil, 0, err - } + if peerAddr != nil { + addr, addrLen = netstack.ConvertAddress(linux.AF_UNIX, *peerAddr) } fd, e := t.NewFDFromVFS2(0, ns, kernel.FDFlags{ @@ -246,13 +268,17 @@ func (s *SocketVFS2) Read(ctx context.Context, dst usermem.IOSequence, opts vfs. if dst.NumBytes() == 0 { return 0, nil } - return dst.CopyOutFrom(ctx, &EndpointReader{ + r := &EndpointReader{ Ctx: ctx, Endpoint: s.ep, NumRights: 0, Peek: false, From: nil, - }) + } + n, err := dst.CopyOutFrom(ctx, r) + // Drop control messages. + r.Control.Release(ctx) + return n, err } // PWrite implements vfs.FileDescriptionImpl. diff --git a/pkg/sentry/state/BUILD b/pkg/sentry/state/BUILD index 0ea4aab8b..563d60578 100644 --- a/pkg/sentry/state/BUILD +++ b/pkg/sentry/state/BUILD @@ -12,10 +12,12 @@ go_library( visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", + "//pkg/context", "//pkg/log", "//pkg/sentry/inet", "//pkg/sentry/kernel", "//pkg/sentry/time", + "//pkg/sentry/vfs", "//pkg/sentry/watchdog", "//pkg/state/statefile", "//pkg/syserror", diff --git a/pkg/sentry/state/state.go b/pkg/sentry/state/state.go index a06c9b8ab..167754537 100644 --- a/pkg/sentry/state/state.go +++ b/pkg/sentry/state/state.go @@ -19,10 +19,12 @@ import ( "fmt" "io" + "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/log" "gvisor.dev/gvisor/pkg/sentry/inet" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/time" + "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/sentry/watchdog" "gvisor.dev/gvisor/pkg/state/statefile" "gvisor.dev/gvisor/pkg/syserror" @@ -57,12 +59,14 @@ type SaveOpts struct { } // Save saves the system state. -func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error { +func (opts SaveOpts) Save(ctx context.Context, k *kernel.Kernel, w *watchdog.Watchdog) error { log.Infof("Sandbox save started, pausing all tasks.") k.Pause() k.ReceiveTaskStates() - defer k.Unpause() - defer log.Infof("Tasks resumed after save.") + defer func() { + k.Unpause() + log.Infof("Tasks resumed after save.") + }() w.Stop() defer w.Start() @@ -79,7 +83,7 @@ func (opts SaveOpts) Save(k *kernel.Kernel, w *watchdog.Watchdog) error { err = ErrStateFile{err} } else { // Save the kernel. - err = k.SaveTo(wc) + err = k.SaveTo(ctx, wc) // ENOSPC is a state file error. This error can only come from // writing the state file, and not from fs.FileOperations.Fsync @@ -106,7 +110,7 @@ type LoadOpts struct { } // Load loads the given kernel, setting the provided platform and stack. -func (opts LoadOpts) Load(k *kernel.Kernel, n inet.Stack, clocks time.Clocks) error { +func (opts LoadOpts) Load(ctx context.Context, k *kernel.Kernel, n inet.Stack, clocks time.Clocks, vfsOpts *vfs.CompleteRestoreOptions) error { // Open the file. r, m, err := statefile.NewReader(opts.Source, opts.Key) if err != nil { @@ -116,5 +120,5 @@ func (opts LoadOpts) Load(k *kernel.Kernel, n inet.Stack, clocks time.Clocks) er previousMetadata = m // Restore the Kernel object graph. - return k.LoadFrom(r, n, clocks) + return k.LoadFrom(ctx, r, n, clocks, vfsOpts) } diff --git a/pkg/sentry/strace/BUILD b/pkg/sentry/strace/BUILD index 88d5db9fc..a920180d3 100644 --- a/pkg/sentry/strace/BUILD +++ b/pkg/sentry/strace/BUILD @@ -28,6 +28,7 @@ go_library( "//pkg/binary", "//pkg/bits", "//pkg/eventchannel", + "//pkg/marshal/primitive", "//pkg/seccomp", "//pkg/sentry/arch", "//pkg/sentry/kernel", diff --git a/pkg/sentry/strace/epoll.go b/pkg/sentry/strace/epoll.go index 5d51a7792..ae3b998c8 100644 --- a/pkg/sentry/strace/epoll.go +++ b/pkg/sentry/strace/epoll.go @@ -26,7 +26,7 @@ import ( func epollEvent(t *kernel.Task, eventAddr usermem.Addr) string { var e linux.EpollEvent - if _, err := t.CopyIn(eventAddr, &e); err != nil { + if _, err := e.CopyIn(t, eventAddr); err != nil { return fmt.Sprintf("%#x {error reading event: %v}", eventAddr, err) } var sb strings.Builder @@ -41,7 +41,7 @@ func epollEvents(t *kernel.Task, eventsAddr usermem.Addr, numEvents, maxBytes ui addr := eventsAddr for i := uint64(0); i < numEvents; i++ { var e linux.EpollEvent - if _, err := t.CopyIn(addr, &e); err != nil { + if _, err := e.CopyIn(t, addr); err != nil { fmt.Fprintf(&sb, "{error reading event at %#x: %v}", addr, err) continue } diff --git a/pkg/sentry/strace/socket.go b/pkg/sentry/strace/socket.go index b51c4c941..cc5f70cd4 100644 --- a/pkg/sentry/strace/socket.go +++ b/pkg/sentry/strace/socket.go @@ -21,6 +21,7 @@ import ( "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/socket/netlink" "gvisor.dev/gvisor/pkg/sentry/socket/netstack" @@ -166,7 +167,7 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) } buf := make([]byte, length) - if _, err := t.CopyIn(addr, &buf); err != nil { + if _, err := t.CopyInBytes(addr, buf); err != nil { return fmt.Sprintf("%#x (error decoding control: %v)", addr, err) } @@ -302,7 +303,7 @@ func cmsghdr(t *kernel.Task, addr usermem.Addr, length uint64, maxBytes uint64) 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 { + if _, err := msg.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding msghdr: %v)", addr, err) } s := fmt.Sprintf( @@ -380,9 +381,9 @@ func postSockAddr(t *kernel.Task, addr usermem.Addr, lengthPtr usermem.Addr) str 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 + var l primitive.Uint32 + _, err := l.CopyIn(t, addr) + return uint32(l), err } func sockLenPointer(t *kernel.Task, addr usermem.Addr) string { @@ -436,22 +437,22 @@ func getSockOptVal(t *kernel.Task, level, optname uint64, optVal usermem.Addr, o func sockOptVal(t *kernel.Task, level, optname uint64, optVal usermem.Addr, optLen uint64, maximumBlobSize uint) string { switch optLen { case 1: - var v uint8 - _, err := t.CopyIn(optVal, &v) + var v primitive.Uint8 + _, err := v.CopyIn(t, optVal) if err != nil { return fmt.Sprintf("%#x {error reading optval: %v}", optVal, err) } return fmt.Sprintf("%#x {value=%v}", optVal, v) case 2: - var v uint16 - _, err := t.CopyIn(optVal, &v) + var v primitive.Uint16 + _, err := v.CopyIn(t, optVal) if err != nil { return fmt.Sprintf("%#x {error reading optval: %v}", optVal, err) } return fmt.Sprintf("%#x {value=%v}", optVal, v) case 4: - var v uint32 - _, err := t.CopyIn(optVal, &v) + var v primitive.Uint32 + _, err := v.CopyIn(t, optVal) if err != nil { return fmt.Sprintf("%#x {error reading optval: %v}", optVal, err) } @@ -632,6 +633,8 @@ var sockOptNames = map[uint64]abi.ValueSet{ linux.IPV6_UNICAST_IF: "IPV6_UNICAST_IF", linux.MCAST_MSFILTER: "MCAST_MSFILTER", linux.IPV6_ADDRFORM: "IPV6_ADDRFORM", + linux.IP6T_SO_GET_INFO: "IP6T_SO_GET_INFO", + linux.IP6T_SO_GET_ENTRIES: "IP6T_SO_GET_ENTRIES", }, linux.SOL_NETLINK: { linux.NETLINK_BROADCAST_ERROR: "NETLINK_BROADCAST_ERROR", diff --git a/pkg/sentry/strace/strace.go b/pkg/sentry/strace/strace.go index 87b239730..396744597 100644 --- a/pkg/sentry/strace/strace.go +++ b/pkg/sentry/strace/strace.go @@ -17,17 +17,16 @@ package strace import ( - "encoding/binary" "fmt" "strconv" "strings" - "syscall" "time" "gvisor.dev/gvisor/pkg/abi" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/bits" "gvisor.dev/gvisor/pkg/eventchannel" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/seccomp" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -91,7 +90,7 @@ func iovecs(t *kernel.Task, addr usermem.Addr, iovcnt int, printContent bool, ma } b := make([]byte, size) - amt, err := t.CopyIn(ar.Start, b) + amt, err := t.CopyInBytes(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 @@ -118,7 +117,7 @@ func dump(t *kernel.Task, addr usermem.Addr, size uint, maximumBlobSize uint) st } b := make([]byte, size) - amt, err := t.CopyIn(addr, b) + amt, err := t.CopyInBytes(addr, b) if err != nil { return fmt.Sprintf("%#x (error decoding string: %s)", addr, err) } @@ -199,7 +198,7 @@ func fdVFS2(t *kernel.Task, fd int32) string { func fdpair(t *kernel.Task, addr usermem.Addr) string { var fds [2]int32 - _, err := t.CopyIn(addr, &fds) + _, err := primitive.CopyInt32SliceIn(t, addr, fds[:]) if err != nil { return fmt.Sprintf("%#x (error decoding fds: %s)", addr, err) } @@ -209,7 +208,7 @@ func fdpair(t *kernel.Task, addr usermem.Addr) string { func uname(t *kernel.Task, addr usermem.Addr) string { var u linux.UtsName - if _, err := t.CopyIn(addr, &u); err != nil { + if _, err := u.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding utsname: %s)", addr, err) } @@ -222,7 +221,7 @@ func utimensTimespec(t *kernel.Task, addr usermem.Addr) string { } var tim linux.Timespec - if _, err := t.CopyIn(addr, &tim); err != nil { + if _, err := tim.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding timespec: %s)", addr, err) } @@ -244,7 +243,7 @@ func timespec(t *kernel.Task, addr usermem.Addr) string { } var tim linux.Timespec - if _, err := t.CopyIn(addr, &tim); err != nil { + if _, err := tim.CopyIn(t, addr); 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) @@ -256,7 +255,7 @@ func timeval(t *kernel.Task, addr usermem.Addr) string { } var tim linux.Timeval - if _, err := t.CopyIn(addr, &tim); err != nil { + if _, err := tim.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding timeval: %s)", addr, err) } @@ -268,8 +267,8 @@ func utimbuf(t *kernel.Task, addr usermem.Addr) string { return "null" } - var utim syscall.Utimbuf - if _, err := t.CopyIn(addr, &utim); err != nil { + var utim linux.Utime + if _, err := utim.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding utimbuf: %s)", addr, err) } @@ -282,7 +281,7 @@ func stat(t *kernel.Task, addr usermem.Addr) string { } var stat linux.Stat - if _, err := t.CopyIn(addr, &stat); err != nil { + if _, err := stat.CopyIn(t, addr); 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)) @@ -294,7 +293,7 @@ func itimerval(t *kernel.Task, addr usermem.Addr) string { } interval := timeval(t, addr) - value := timeval(t, addr+usermem.Addr(binary.Size(linux.Timeval{}))) + value := timeval(t, addr+usermem.Addr((*linux.Timeval)(nil).SizeBytes())) return fmt.Sprintf("%#x {interval=%s, value=%s}", addr, interval, value) } @@ -304,7 +303,7 @@ func itimerspec(t *kernel.Task, addr usermem.Addr) string { } interval := timespec(t, addr) - value := timespec(t, addr+usermem.Addr(binary.Size(linux.Timespec{}))) + value := timespec(t, addr+usermem.Addr((*linux.Timespec)(nil).SizeBytes())) return fmt.Sprintf("%#x {interval=%s, value=%s}", addr, interval, value) } @@ -330,7 +329,7 @@ func rusage(t *kernel.Task, addr usermem.Addr) string { } var ru linux.Rusage - if _, err := t.CopyIn(addr, &ru); err != nil { + if _, err := ru.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding rusage: %s)", addr, err) } return fmt.Sprintf("%#x %+v", addr, ru) @@ -342,7 +341,7 @@ func capHeader(t *kernel.Task, addr usermem.Addr) string { } var hdr linux.CapUserHeader - if _, err := t.CopyIn(addr, &hdr); err != nil { + if _, err := hdr.CopyIn(t, addr); err != nil { return fmt.Sprintf("%#x (error decoding header: %s)", addr, err) } @@ -367,7 +366,7 @@ func capData(t *kernel.Task, hdrAddr, dataAddr usermem.Addr) string { } var hdr linux.CapUserHeader - if _, err := t.CopyIn(hdrAddr, &hdr); err != nil { + if _, err := hdr.CopyIn(t, hdrAddr); err != nil { return fmt.Sprintf("%#x (error decoding header: %v)", dataAddr, err) } @@ -376,7 +375,7 @@ func capData(t *kernel.Task, hdrAddr, dataAddr usermem.Addr) string { switch hdr.Version { case linux.LINUX_CAPABILITY_VERSION_1: var data linux.CapUserData - if _, err := t.CopyIn(dataAddr, &data); err != nil { + if _, err := data.CopyIn(t, dataAddr); err != nil { return fmt.Sprintf("%#x (error decoding data: %v)", dataAddr, err) } p = uint64(data.Permitted) @@ -384,7 +383,7 @@ func capData(t *kernel.Task, hdrAddr, dataAddr usermem.Addr) string { e = uint64(data.Effective) case linux.LINUX_CAPABILITY_VERSION_2, linux.LINUX_CAPABILITY_VERSION_3: var data [2]linux.CapUserData - if _, err := t.CopyIn(dataAddr, &data); err != nil { + if _, err := linux.CopyCapUserDataSliceIn(t, dataAddr, data[:]); err != nil { return fmt.Sprintf("%#x (error decoding data: %v)", dataAddr, err) } p = uint64(data[0].Permitted) | (uint64(data[1].Permitted) << 32) diff --git a/pkg/sentry/syscalls/linux/BUILD b/pkg/sentry/syscalls/linux/BUILD index 4a9b04fd0..a2e441448 100644 --- a/pkg/sentry/syscalls/linux/BUILD +++ b/pkg/sentry/syscalls/linux/BUILD @@ -21,6 +21,7 @@ go_library( "sys_identity.go", "sys_inotify.go", "sys_lseek.go", + "sys_membarrier.go", "sys_mempolicy.go", "sys_mmap.go", "sys_mount.go", @@ -56,6 +57,7 @@ go_library( "sys_xattr.go", "timespec.go", ], + marshal = True, visibility = ["//:sandbox"], deps = [ "//pkg/abi", @@ -64,6 +66,8 @@ go_library( "//pkg/bpf", "//pkg/context", "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/metric", "//pkg/rand", "//pkg/safemem", @@ -99,7 +103,5 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", - "//tools/go_marshal/primitive", ], ) diff --git a/pkg/sentry/syscalls/linux/linux64.go b/pkg/sentry/syscalls/linux/linux64.go index 80c65164a..36902d177 100644 --- a/pkg/sentry/syscalls/linux/linux64.go +++ b/pkg/sentry/syscalls/linux/linux64.go @@ -118,7 +118,7 @@ var AMD64 = &kernel.SyscallTable{ 63: syscalls.Supported("uname", Uname), 64: syscalls.Supported("semget", Semget), 65: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), - 66: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), + 66: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), 67: syscalls.Supported("shmdt", Shmdt), 68: syscalls.ErrorWithEvent("msgget", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) 69: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) @@ -138,7 +138,7 @@ var AMD64 = &kernel.SyscallTable{ 83: syscalls.Supported("mkdir", Mkdir), 84: syscalls.Supported("rmdir", Rmdir), 85: syscalls.Supported("creat", Creat), - 86: syscalls.Supported("link", Link), + 86: syscalls.PartiallySupported("link", Link, "Limited support with Gofer. Link count and linked files may get out of sync because gVisor is not aware of external hardlinks.", nil), 87: syscalls.Supported("unlink", Unlink), 88: syscalls.Supported("symlink", Symlink), 89: syscalls.Supported("readlink", Readlink), @@ -305,9 +305,9 @@ var AMD64 = &kernel.SyscallTable{ 250: syscalls.Error("keyctl", syserror.EACCES, "Not available to user.", nil), 251: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) 252: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) - 253: syscalls.PartiallySupported("inotify_init", InotifyInit, "inotify events are only available inside the sandbox.", nil), - 254: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), - 255: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), + 253: syscalls.PartiallySupported("inotify_init", InotifyInit, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), + 254: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), + 255: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), 256: syscalls.CapError("migrate_pages", linux.CAP_SYS_NICE, "", nil), 257: syscalls.Supported("openat", Openat), 258: syscalls.Supported("mkdirat", Mkdirat), @@ -317,7 +317,7 @@ var AMD64 = &kernel.SyscallTable{ 262: syscalls.Supported("fstatat", Fstatat), 263: syscalls.Supported("unlinkat", Unlinkat), 264: syscalls.Supported("renameat", Renameat), - 265: syscalls.Supported("linkat", Linkat), + 265: syscalls.PartiallySupported("linkat", Linkat, "See link(2).", nil), 266: syscalls.Supported("symlinkat", Symlinkat), 267: syscalls.Supported("readlinkat", Readlinkat), 268: syscalls.Supported("fchmodat", Fchmodat), @@ -346,7 +346,7 @@ var AMD64 = &kernel.SyscallTable{ 291: syscalls.Supported("epoll_create1", EpollCreate1), 292: syscalls.Supported("dup3", Dup3), 293: syscalls.Supported("pipe2", Pipe2), - 294: syscalls.Supported("inotify_init1", InotifyInit1), + 294: syscalls.PartiallySupported("inotify_init1", InotifyInit1, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), 295: syscalls.Supported("preadv", Preadv), 296: syscalls.Supported("pwritev", Pwritev), 297: syscalls.Supported("rt_tgsigqueueinfo", RtTgsigqueueinfo), @@ -376,7 +376,7 @@ var AMD64 = &kernel.SyscallTable{ 321: syscalls.CapError("bpf", linux.CAP_SYS_ADMIN, "", nil), 322: syscalls.Supported("execveat", Execveat), 323: syscalls.ErrorWithEvent("userfaultfd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/266"}), // TODO(b/118906345) - 324: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) + 324: syscalls.PartiallySupported("membarrier", Membarrier, "Not supported on all platforms.", nil), 325: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), // Syscalls implemented after 325 are "backports" from versions @@ -454,9 +454,9 @@ var ARM64 = &kernel.SyscallTable{ 23: syscalls.Supported("dup", Dup), 24: syscalls.Supported("dup3", Dup3), 25: syscalls.PartiallySupported("fcntl", Fcntl, "Not all options are supported.", nil), - 26: syscalls.Supported("inotify_init1", InotifyInit1), - 27: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "inotify events are only available inside the sandbox.", nil), - 28: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "inotify events are only available inside the sandbox.", nil), + 26: syscalls.PartiallySupported("inotify_init1", InotifyInit1, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), + 27: syscalls.PartiallySupported("inotify_add_watch", InotifyAddWatch, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), + 28: syscalls.PartiallySupported("inotify_rm_watch", InotifyRmWatch, "Inotify events are only available inside the sandbox. Hard links are treated as different watch targets in gofer fs.", nil), 29: syscalls.PartiallySupported("ioctl", Ioctl, "Only a few ioctls are implemented for backing devices and file systems.", nil), 30: syscalls.CapError("ioprio_set", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) 31: syscalls.CapError("ioprio_get", linux.CAP_SYS_ADMIN, "", nil), // requires cap_sys_nice or cap_sys_admin (depending) @@ -527,8 +527,8 @@ var ARM64 = &kernel.SyscallTable{ 96: syscalls.Supported("set_tid_address", SetTidAddress), 97: syscalls.PartiallySupported("unshare", Unshare, "Mount, cgroup namespaces not supported. Network namespaces supported but must be empty.", nil), 98: syscalls.PartiallySupported("futex", Futex, "Robust futexes not supported.", nil), - 99: syscalls.Error("set_robust_list", syserror.ENOSYS, "Obsolete.", nil), - 100: syscalls.Error("get_robust_list", syserror.ENOSYS, "Obsolete.", nil), + 99: syscalls.Supported("set_robust_list", SetRobustList), + 100: syscalls.Supported("get_robust_list", GetRobustList), 101: syscalls.Supported("nanosleep", Nanosleep), 102: syscalls.Supported("getitimer", Getitimer), 103: syscalls.Supported("setitimer", Setitimer), @@ -619,7 +619,7 @@ var ARM64 = &kernel.SyscallTable{ 188: syscalls.ErrorWithEvent("msgrcv", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) 189: syscalls.ErrorWithEvent("msgsnd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/135"}), // TODO(b/29354921) 190: syscalls.Supported("semget", Semget), - 191: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, IPC_STAT, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), + 191: syscalls.PartiallySupported("semctl", Semctl, "Options IPC_INFO, SEM_INFO, SEM_STAT, SEM_STAT_ANY, GETNCNT, GETZCNT not supported.", nil), 192: syscalls.ErrorWithEvent("semtimedop", syserror.ENOSYS, "", []string{"gvisor.dev/issue/137"}), 193: syscalls.PartiallySupported("semop", Semop, "Option SEM_UNDO not supported.", nil), 194: syscalls.PartiallySupported("shmget", Shmget, "Option SHM_HUGETLB is not supported.", nil), @@ -695,7 +695,7 @@ var ARM64 = &kernel.SyscallTable{ 280: syscalls.CapError("bpf", linux.CAP_SYS_ADMIN, "", nil), 281: syscalls.Supported("execveat", Execveat), 282: syscalls.ErrorWithEvent("userfaultfd", syserror.ENOSYS, "", []string{"gvisor.dev/issue/266"}), // TODO(b/118906345) - 283: syscalls.ErrorWithEvent("membarrier", syserror.ENOSYS, "", []string{"gvisor.dev/issue/267"}), // TODO(gvisor.dev/issue/267) + 283: syscalls.PartiallySupported("membarrier", Membarrier, "Not supported on all platforms.", nil), 284: syscalls.PartiallySupported("mlock2", Mlock2, "Stub implementation. The sandbox lacks appropriate permissions.", nil), // Syscalls after 284 are "backports" from versions of Linux after 4.4. diff --git a/pkg/sentry/syscalls/linux/sys_aio.go b/pkg/sentry/syscalls/linux/sys_aio.go index e9d64dec5..0bf313a13 100644 --- a/pkg/sentry/syscalls/linux/sys_aio.go +++ b/pkg/sentry/syscalls/linux/sys_aio.go @@ -17,6 +17,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -36,7 +37,7 @@ func IoSetup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca // // The context pointer _must_ be zero initially. var idIn uint64 - if _, err := t.CopyIn(idAddr, &idIn); err != nil { + if _, err := primitive.CopyUint64In(t, idAddr, &idIn); err != nil { return 0, nil, err } if idIn != 0 { @@ -49,7 +50,7 @@ func IoSetup(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca } // Copy out the new ID. - if _, err := t.CopyOut(idAddr, &id); err != nil { + if _, err := primitive.CopyUint64Out(t, idAddr, id); err != nil { t.MemoryManager().DestroyAIOContext(t, id) return 0, nil, err } @@ -142,7 +143,7 @@ func IoGetevents(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S ev := v.(*linux.IOEvent) // Copy out the result. - if _, err := t.CopyOut(eventsAddr, ev); err != nil { + if _, err := ev.CopyOut(t, eventsAddr); err != nil { if count > 0 { return uintptr(count), nil, nil } @@ -338,21 +339,27 @@ func IoSubmit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } 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 + // Copy in the callback address. + var cbAddr usermem.Addr + switch t.Arch().Width() { + case 8: + var cbAddrP primitive.Uint64 + if _, err := cbAddrP.CopyIn(t, addr); err != nil { + if i > 0 { + // Some successful. + return uintptr(i), nil, nil + } + // Nothing done. + return 0, nil, err } - // Nothing done. - return 0, nil, err + cbAddr = usermem.Addr(cbAddrP) + default: + return 0, nil, syserror.ENOSYS } // Copy in this callback. var cb linux.IOCallback - cbAddr := usermem.Addr(t.Arch().Value(cbAddrNative)) - if _, err := t.CopyIn(cbAddr, &cb); err != nil { + if _, err := cb.CopyIn(t, cbAddr); err != nil { if i > 0 { // Some have been successful. diff --git a/pkg/sentry/syscalls/linux/sys_capability.go b/pkg/sentry/syscalls/linux/sys_capability.go index adf5ea5f2..d3b85e11b 100644 --- a/pkg/sentry/syscalls/linux/sys_capability.go +++ b/pkg/sentry/syscalls/linux/sys_capability.go @@ -45,7 +45,7 @@ func Capget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal dataAddr := args[1].Pointer() var hdr linux.CapUserHeader - if _, err := t.CopyIn(hdrAddr, &hdr); err != nil { + if _, err := hdr.CopyIn(t, hdrAddr); err != nil { return 0, nil, err } // hdr.Pid doesn't need to be valid if this capget() is a "version probe" @@ -65,7 +65,7 @@ func Capget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal Permitted: uint32(p), Inheritable: uint32(i), } - _, err = t.CopyOut(dataAddr, &data) + _, err = data.CopyOut(t, dataAddr) return 0, nil, err case linux.LINUX_CAPABILITY_VERSION_2, linux.LINUX_CAPABILITY_VERSION_3: @@ -88,12 +88,12 @@ func Capget(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal Inheritable: uint32(i >> 32), }, } - _, err = t.CopyOut(dataAddr, &data) + _, err = linux.CopyCapUserDataSliceOut(t, dataAddr, data[:]) return 0, nil, err default: hdr.Version = linux.HighestCapabilityVersion - if _, err := t.CopyOut(hdrAddr, &hdr); err != nil { + if _, err := hdr.CopyOut(t, hdrAddr); err != nil { return 0, nil, err } if dataAddr != 0 { @@ -109,7 +109,7 @@ func Capset(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal dataAddr := args[1].Pointer() var hdr linux.CapUserHeader - if _, err := t.CopyIn(hdrAddr, &hdr); err != nil { + if _, err := hdr.CopyIn(t, hdrAddr); err != nil { return 0, nil, err } switch hdr.Version { @@ -118,7 +118,7 @@ func Capset(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, syserror.EPERM } var data linux.CapUserData - if _, err := t.CopyIn(dataAddr, &data); err != nil { + if _, err := data.CopyIn(t, dataAddr); err != nil { return 0, nil, err } p := auth.CapabilitySet(data.Permitted) & auth.AllCapabilities @@ -131,7 +131,7 @@ func Capset(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return 0, nil, syserror.EPERM } var data [2]linux.CapUserData - if _, err := t.CopyIn(dataAddr, &data); err != nil { + if _, err := linux.CopyCapUserDataSliceIn(t, dataAddr, data[:]); err != nil { return 0, nil, err } p := (auth.CapabilitySet(data[0].Permitted) | (auth.CapabilitySet(data[1].Permitted) << 32)) & auth.AllCapabilities @@ -141,7 +141,7 @@ func Capset(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal default: hdr.Version = linux.HighestCapabilityVersion - if _, err := t.CopyOut(hdrAddr, &hdr); err != nil { + if _, err := hdr.CopyOut(t, hdrAddr); err != nil { return 0, nil, err } return 0, nil, syserror.EINVAL diff --git a/pkg/sentry/syscalls/linux/sys_file.go b/pkg/sentry/syscalls/linux/sys_file.go index 256422689..519066a47 100644 --- a/pkg/sentry/syscalls/linux/sys_file.go +++ b/pkg/sentry/syscalls/linux/sys_file.go @@ -19,6 +19,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fs/lock" @@ -83,6 +84,7 @@ func fileOpOn(t *kernel.Task, dirFD int32, path string, resolve bool, fn func(ro } rel = f.Dirent if !fs.IsDir(rel.Inode.StableAttr) { + f.DecRef(t) return syserror.ENOTDIR } } @@ -601,19 +603,19 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Shared flags between file and socket. switch request { case linux.FIONCLEX: - t.FDTable().SetFlags(fd, kernel.FDFlags{ + t.FDTable().SetFlags(t, fd, kernel.FDFlags{ CloseOnExec: false, }) return 0, nil, nil case linux.FIOCLEX: - t.FDTable().SetFlags(fd, kernel.FDFlags{ + t.FDTable().SetFlags(t, 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 { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil { return 0, nil, err } flags := file.Flags() @@ -627,7 +629,7 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.FIOASYNC: var set int32 - if _, err := t.CopyIn(args[2].Pointer(), &set); err != nil { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil { return 0, nil, err } flags := file.Flags() @@ -641,15 +643,14 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.FIOSETOWN, linux.SIOCSPGRP: var set int32 - if _, err := t.CopyIn(args[2].Pointer(), &set); err != nil { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil { return 0, nil, err } fSetOwn(t, file, set) return 0, nil, nil case linux.FIOGETOWN, linux.SIOCGPGRP: - who := fGetOwn(t, file) - _, err := t.CopyOut(args[2].Pointer(), &who) + _, err := primitive.CopyInt32Out(t, args[2].Pointer(), fGetOwn(t, file)) return 0, nil, err default: @@ -694,7 +695,7 @@ func Getcwd(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal } // Top it off with a terminator. - _, err = t.CopyOut(addr+usermem.Addr(bytes), []byte("\x00")) + _, err = t.CopyOutBytes(addr+usermem.Addr(bytes), []byte("\x00")) return uintptr(bytes + 1), nil, err } @@ -787,7 +788,7 @@ func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Note that Remove provides a reference on the file that we may use to // flush. It is still active until we drop the final reference below // (and other reference-holding operations complete). - file, _ := t.FDTable().Remove(fd) + file, _ := t.FDTable().Remove(t, fd) if file == nil { return 0, nil, syserror.EBADF } @@ -941,7 +942,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(flags.ToLinuxFDFlags()), nil, nil case linux.F_SETFD: flags := args[2].Uint() - err := t.FDTable().SetFlags(fd, kernel.FDFlags{ + err := t.FDTable().SetFlags(t, fd, kernel.FDFlags{ CloseOnExec: flags&linux.FD_CLOEXEC != 0, }) return 0, nil, err @@ -962,7 +963,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Copy in the lock request. flockAddr := args[2].Pointer() var flock linux.Flock - if _, err := t.CopyIn(flockAddr, &flock); err != nil { + if _, err := flock.CopyIn(t, flockAddr); err != nil { return 0, nil, err } @@ -1052,12 +1053,12 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.F_GETOWN_EX: addr := args[2].Pointer() owner := fGetOwnEx(t, file) - _, err := t.CopyOut(addr, &owner) + _, err := owner.CopyOut(t, addr) return 0, nil, err case linux.F_SETOWN_EX: addr := args[2].Pointer() var owner linux.FOwnerEx - _, err := t.CopyIn(addr, &owner) + _, err := owner.CopyIn(t, addr) if err != nil { return 0, nil, err } @@ -1154,6 +1155,10 @@ func Fadvise64(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, nil } +// LINT.ThenChange(vfs2/fd.go) + +// LINT.IfChange + func mkdirAt(t *kernel.Task, dirFD int32, addr usermem.Addr, mode linux.FileMode) error { path, _, err := copyInPath(t, addr, false /* allowEmpty */) if err != nil { @@ -1918,7 +1923,7 @@ func Utime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall ts := defaultSetToSystemTimeSpec() if timesAddr != 0 { var times linux.Utime - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := times.CopyIn(t, timesAddr); err != nil { return 0, nil, err } ts = fs.TimeSpec{ @@ -1938,7 +1943,7 @@ func Utimes(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal ts := defaultSetToSystemTimeSpec() if timesAddr != 0 { var times [2]linux.Timeval - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := linux.CopyTimevalSliceIn(t, timesAddr, times[:]); err != nil { return 0, nil, err } ts = fs.TimeSpec{ @@ -1966,7 +1971,7 @@ func Utimensat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys ts := defaultSetToSystemTimeSpec() if timesAddr != 0 { var times [2]linux.Timespec - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := linux.CopyTimespecSliceIn(t, timesAddr, times[:]); err != nil { return 0, nil, err } if !timespecIsValid(times[0]) || !timespecIsValid(times[1]) { @@ -2000,7 +2005,7 @@ func Futimesat(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys ts := defaultSetToSystemTimeSpec() if timesAddr != 0 { var times [2]linux.Timeval - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := linux.CopyTimevalSliceIn(t, timesAddr, times[:]); err != nil { return 0, nil, err } if times[0].Usec >= 1e6 || times[0].Usec < 0 || diff --git a/pkg/sentry/syscalls/linux/sys_futex.go b/pkg/sentry/syscalls/linux/sys_futex.go index 12b2fa690..f39ce0639 100644 --- a/pkg/sentry/syscalls/linux/sys_futex.go +++ b/pkg/sentry/syscalls/linux/sys_futex.go @@ -306,8 +306,8 @@ func GetRobustList(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel // Despite the syscall using the name 'pid' for this variable, it is // very much a tid. tid := args[0].Int() - head := args[1].Pointer() - size := args[2].Pointer() + headAddr := args[1].Pointer() + sizeAddr := args[2].Pointer() if tid < 0 { return 0, nil, syserror.EINVAL @@ -321,12 +321,16 @@ func GetRobustList(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel } // Copy out head pointer. - if _, err := t.CopyOut(head, uint64(ot.GetRobustList())); err != nil { + head := t.Arch().Native(uintptr(ot.GetRobustList())) + if _, err := head.CopyOut(t, headAddr); err != nil { return 0, nil, err } - // Copy out size, which is a constant. - if _, err := t.CopyOut(size, uint64(linux.SizeOfRobustListHead)); err != nil { + // Copy out size, which is a constant. Note that while size isn't + // an address, it is defined as the arch-dependent size_t, so it + // needs to be converted to a native-sized int. + size := t.Arch().Native(uintptr(linux.SizeOfRobustListHead)) + if _, err := size.CopyOut(t, sizeAddr); err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_getdents.go b/pkg/sentry/syscalls/linux/sys_getdents.go index 59004cefe..b25f7d881 100644 --- a/pkg/sentry/syscalls/linux/sys_getdents.go +++ b/pkg/sentry/syscalls/linux/sys_getdents.go @@ -19,7 +19,6 @@ import ( "io" "gvisor.dev/gvisor/pkg/abi/linux" - "gvisor.dev/gvisor/pkg/binary" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -93,19 +92,23 @@ func getdents(t *kernel.Task, fd int32, addr usermem.Addr, size int, f func(*dir } } -// oldDirentHdr is a fixed sized header matching the fixed size -// fields found in the old linux dirent struct. +// oldDirentHdr is a fixed sized header matching the fixed size fields found in +// the old linux dirent struct. +// +// +marshal type oldDirentHdr struct { Ino uint64 Off uint64 - Reclen uint16 + Reclen uint16 `marshal:"unaligned"` // Struct ends mid-word. } -// direntHdr is a fixed sized header matching the fixed size -// fields found in the new linux dirent struct. +// direntHdr is a fixed sized header matching the fixed size fields found in the +// new linux dirent struct. +// +// +marshal type direntHdr struct { OldHdr oldDirentHdr - Typ uint8 + Typ uint8 `marshal:"unaligned"` // Struct ends mid-word. } // dirent contains the data pointed to by a new linux dirent struct. @@ -134,20 +137,20 @@ func newDirent(width uint, name string, attr fs.DentAttr, offset uint64) *dirent // the old linux dirent format. func smallestDirent(a arch.Context) uint { d := dirent{} - return uint(binary.Size(d.Hdr.OldHdr)) + a.Width() + 1 + return uint(d.Hdr.OldHdr.SizeBytes()) + 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() + return uint(d.Hdr.SizeBytes()) + a.Width() } // 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) + a := d.Hdr.SizeBytes() + len(d.Name) r := (a + width) &^ (width - 1) padding := r - a d.Name = append(d.Name, make([]byte, padding)...) @@ -157,7 +160,7 @@ func (d *dirent) padRec(width int) uint16 { // 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)) + n1, err := d.Hdr.WriteTo(w) if err != nil { return 0, err } @@ -165,14 +168,14 @@ func (d *dirent) Serialize64(w io.Writer) (int, error) { if err != nil { return 0, err } - return n1 + n2, nil + return int(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)) + n1, err := d.Hdr.OldHdr.WriteTo(w) if err != nil { return 0, err } @@ -184,7 +187,7 @@ func (d *dirent) Serialize(w io.Writer) (int, error) { if err != nil { return 0, err } - return n1 + n2 + n3, nil + return int(n1) + n2 + n3, nil } // direntSerializer implements fs.InodeOperationsInfoSerializer, serializing dirents to an diff --git a/pkg/sentry/syscalls/linux/sys_identity.go b/pkg/sentry/syscalls/linux/sys_identity.go index 715ac45e6..a29d307e5 100644 --- a/pkg/sentry/syscalls/linux/sys_identity.go +++ b/pkg/sentry/syscalls/linux/sys_identity.go @@ -49,13 +49,13 @@ func Getresuid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys 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 { + if _, err := ruid.CopyOut(t, ruidAddr); err != nil { return 0, nil, err } - if _, err := t.CopyOut(euidAddr, euid); err != nil { + if _, err := euid.CopyOut(t, euidAddr); err != nil { return 0, nil, err } - if _, err := t.CopyOut(suidAddr, suid); err != nil { + if _, err := suid.CopyOut(t, suidAddr); err != nil { return 0, nil, err } return 0, nil, nil @@ -84,13 +84,13 @@ func Getresgid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys 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 { + if _, err := rgid.CopyOut(t, rgidAddr); err != nil { return 0, nil, err } - if _, err := t.CopyOut(egidAddr, egid); err != nil { + if _, err := egid.CopyOut(t, egidAddr); err != nil { return 0, nil, err } - if _, err := t.CopyOut(sgidAddr, sgid); err != nil { + if _, err := sgid.CopyOut(t, sgidAddr); err != nil { return 0, nil, err } return 0, nil, nil @@ -157,7 +157,7 @@ func Getgroups(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys for i, kgid := range kgids { gids[i] = kgid.In(t.UserNamespace()).OrOverflow() } - if _, err := t.CopyOut(args[1].Pointer(), gids); err != nil { + if _, err := auth.CopyGIDSliceOut(t, args[1].Pointer(), gids); err != nil { return 0, nil, err } return uintptr(len(gids)), nil, nil @@ -173,7 +173,7 @@ func Setgroups(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, t.SetExtraGIDs(nil) } gids := make([]auth.GID, size) - if _, err := t.CopyIn(args[1].Pointer(), &gids); err != nil { + if _, err := auth.CopyGIDSliceIn(t, args[1].Pointer(), gids); err != nil { return 0, nil, err } return 0, nil, t.SetExtraGIDs(gids) diff --git a/pkg/sentry/syscalls/linux/sys_membarrier.go b/pkg/sentry/syscalls/linux/sys_membarrier.go new file mode 100644 index 000000000..63ee5d435 --- /dev/null +++ b/pkg/sentry/syscalls/linux/sys_membarrier.go @@ -0,0 +1,103 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES 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.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/kernel" + "gvisor.dev/gvisor/pkg/syserror" +) + +// Membarrier implements syscall membarrier(2). +func Membarrier(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + cmd := args[0].Int() + flags := args[1].Uint() + + switch cmd { + case linux.MEMBARRIER_CMD_QUERY: + if flags != 0 { + return 0, nil, syserror.EINVAL + } + var supportedCommands uintptr + if t.Kernel().Platform.HaveGlobalMemoryBarrier() { + supportedCommands |= linux.MEMBARRIER_CMD_GLOBAL | + linux.MEMBARRIER_CMD_GLOBAL_EXPEDITED | + linux.MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED | + linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED | + linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED + } + if t.RSeqAvailable() { + supportedCommands |= linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ | + linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ + } + return supportedCommands, nil, nil + case linux.MEMBARRIER_CMD_GLOBAL, linux.MEMBARRIER_CMD_GLOBAL_EXPEDITED, linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED: + if flags != 0 { + return 0, nil, syserror.EINVAL + } + if !t.Kernel().Platform.HaveGlobalMemoryBarrier() { + return 0, nil, syserror.EINVAL + } + if cmd == linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED && !t.MemoryManager().IsMembarrierPrivateEnabled() { + return 0, nil, syserror.EPERM + } + return 0, nil, t.Kernel().Platform.GlobalMemoryBarrier() + case linux.MEMBARRIER_CMD_REGISTER_GLOBAL_EXPEDITED: + if flags != 0 { + return 0, nil, syserror.EINVAL + } + if !t.Kernel().Platform.HaveGlobalMemoryBarrier() { + return 0, nil, syserror.EINVAL + } + // no-op + return 0, nil, nil + case linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED: + if flags != 0 { + return 0, nil, syserror.EINVAL + } + if !t.Kernel().Platform.HaveGlobalMemoryBarrier() { + return 0, nil, syserror.EINVAL + } + t.MemoryManager().EnableMembarrierPrivate() + return 0, nil, nil + case linux.MEMBARRIER_CMD_PRIVATE_EXPEDITED_RSEQ: + if flags&^linux.MEMBARRIER_CMD_FLAG_CPU != 0 { + return 0, nil, syserror.EINVAL + } + if !t.RSeqAvailable() { + return 0, nil, syserror.EINVAL + } + if !t.MemoryManager().IsMembarrierRSeqEnabled() { + return 0, nil, syserror.EPERM + } + // MEMBARRIER_CMD_FLAG_CPU and cpu_id are ignored since we don't have + // the ability to preempt specific CPUs. + return 0, nil, t.Kernel().Platform.PreemptAllCPUs() + case linux.MEMBARRIER_CMD_REGISTER_PRIVATE_EXPEDITED_RSEQ: + if flags != 0 { + return 0, nil, syserror.EINVAL + } + if !t.RSeqAvailable() { + return 0, nil, syserror.EINVAL + } + t.MemoryManager().EnableMembarrierRSeq() + return 0, nil, nil + default: + // Probably a command we don't implement. + t.Kernel().EmitUnimplementedEvent(t) + return 0, nil, syserror.EINVAL + } +} diff --git a/pkg/sentry/syscalls/linux/sys_mmap.go b/pkg/sentry/syscalls/linux/sys_mmap.go index d0109baa4..cd8dfdfa4 100644 --- a/pkg/sentry/syscalls/linux/sys_mmap.go +++ b/pkg/sentry/syscalls/linux/sys_mmap.go @@ -100,6 +100,15 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC if err := file.ConfigureMMap(t, &opts); err != nil { return 0, nil, err } + } else if shared { + // Back shared anonymous mappings with a special mappable. + opts.Offset = 0 + m, err := mm.NewSharedAnonMappable(opts.Length, t.Kernel()) + if err != nil { + return 0, nil, err + } + opts.MappingIdentity = m // transfers ownership of m to opts + opts.Mappable = m } rv, err := t.MemoryManager().MMap(t, opts) @@ -239,7 +248,7 @@ func Mincore(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca return 0, nil, syserror.ENOMEM } resident := bytes.Repeat([]byte{1}, int(mapped/usermem.PageSize)) - _, err := t.CopyOut(vec, resident) + _, err := t.CopyOutBytes(vec, resident) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_pipe.go b/pkg/sentry/syscalls/linux/sys_pipe.go index 3149e4aad..849a47476 100644 --- a/pkg/sentry/syscalls/linux/sys_pipe.go +++ b/pkg/sentry/syscalls/linux/sys_pipe.go @@ -16,6 +16,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -46,9 +47,9 @@ func pipe2(t *kernel.Task, addr usermem.Addr, flags uint) (uintptr, error) { return 0, err } - if _, err := t.CopyOut(addr, fds); err != nil { + if _, err := primitive.CopyInt32SliceOut(t, addr, fds); err != nil { for _, fd := range fds { - if file, _ := t.FDTable().Remove(fd); file != nil { + if file, _ := t.FDTable().Remove(t, fd); file != nil { file.DecRef(t) } } diff --git a/pkg/sentry/syscalls/linux/sys_poll.go b/pkg/sentry/syscalls/linux/sys_poll.go index 789e2ed5b..254f4c9f9 100644 --- a/pkg/sentry/syscalls/linux/sys_poll.go +++ b/pkg/sentry/syscalls/linux/sys_poll.go @@ -162,7 +162,7 @@ func CopyInPollFDs(t *kernel.Task, addr usermem.Addr, nfds uint) ([]linux.PollFD pfd := make([]linux.PollFD, nfds) if nfds > 0 { - if _, err := t.CopyIn(addr, &pfd); err != nil { + if _, err := linux.CopyPollFDSliceIn(t, addr, pfd); err != nil { return nil, err } } @@ -189,7 +189,7 @@ func doPoll(t *kernel.Task, addr usermem.Addr, nfds uint, timeout time.Duration) // 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(addr, pfd); err != nil { + if _, err := linux.CopyPollFDSliceOut(t, addr, pfd); err != nil { return remainingTimeout, 0, err } } @@ -202,7 +202,7 @@ func CopyInFDSet(t *kernel.Task, addr usermem.Addr, nBytes, nBitsInLastPartialBy set := make([]byte, nBytes) if addr != 0 { - if _, err := t.CopyIn(addr, &set); err != nil { + if _, err := t.CopyInBytes(addr, set); err != nil { return nil, err } // If we only use part of the last byte, mask out the extraneous bits. @@ -329,19 +329,19 @@ func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Add // Copy updated vectors back. if readFDs != 0 { - if _, err := t.CopyOut(readFDs, r); err != nil { + if _, err := t.CopyOutBytes(readFDs, r); err != nil { return 0, err } } if writeFDs != 0 { - if _, err := t.CopyOut(writeFDs, w); err != nil { + if _, err := t.CopyOutBytes(writeFDs, w); err != nil { return 0, err } } if exceptFDs != 0 { - if _, err := t.CopyOut(exceptFDs, e); err != nil { + if _, err := t.CopyOutBytes(exceptFDs, e); err != nil { return 0, err } } diff --git a/pkg/sentry/syscalls/linux/sys_prctl.go b/pkg/sentry/syscalls/linux/sys_prctl.go index 64a725296..a892d2c62 100644 --- a/pkg/sentry/syscalls/linux/sys_prctl.go +++ b/pkg/sentry/syscalls/linux/sys_prctl.go @@ -18,6 +18,7 @@ import ( "fmt" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsbridge" @@ -43,7 +44,7 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, nil case linux.PR_GET_PDEATHSIG: - _, err := t.CopyOut(args[1].Pointer(), int32(t.ParentDeathSignal())) + _, err := primitive.CopyInt32Out(t, args[1].Pointer(), int32(t.ParentDeathSignal())) return 0, nil, err case linux.PR_GET_DUMPABLE: @@ -110,7 +111,7 @@ func Prctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall buf[len] = 0 len++ } - _, err := t.CopyOut(addr, buf[:len]) + _, err := t.CopyOutBytes(addr, buf[:len]) if err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_rlimit.go b/pkg/sentry/syscalls/linux/sys_rlimit.go index d5d5b6959..309c183a3 100644 --- a/pkg/sentry/syscalls/linux/sys_rlimit.go +++ b/pkg/sentry/syscalls/linux/sys_rlimit.go @@ -16,6 +16,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/limits" @@ -26,17 +27,13 @@ import ( // rlimit describes an implementation of 'struct rlimit', which may vary from // system-to-system. type rlimit interface { + marshal.Marshallable + // 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. @@ -50,6 +47,7 @@ func newRlimit(t *kernel.Task) (rlimit, error) { } } +// +marshal type rlimit64 struct { Cur uint64 Max uint64 @@ -70,12 +68,12 @@ func (r *rlimit64) fromLimit(lim limits.Limit) { } func (r *rlimit64) copyIn(t *kernel.Task, addr usermem.Addr) error { - _, err := t.CopyIn(addr, r) + _, err := r.CopyIn(t, addr) return err } func (r *rlimit64) copyOut(t *kernel.Task, addr usermem.Addr) error { - _, err := t.CopyOut(addr, *r) + _, err := r.CopyOut(t, addr) return err } @@ -140,7 +138,8 @@ func Getrlimit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, err } rlim.fromLimit(lim) - return 0, nil, rlim.copyOut(t, addr) + _, err = rlim.CopyOut(t, addr) + return 0, nil, err } // Setrlimit implements linux syscall setrlimit(2). @@ -155,7 +154,7 @@ func Setrlimit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys if err != nil { return 0, nil, err } - if err := rlim.copyIn(t, addr); err != nil { + if _, err := rlim.CopyIn(t, addr); err != nil { return 0, nil, syserror.EFAULT } _, err = prlimit64(t, resource, rlim.toLimit()) diff --git a/pkg/sentry/syscalls/linux/sys_rusage.go b/pkg/sentry/syscalls/linux/sys_rusage.go index 1674c7445..ac5c98a54 100644 --- a/pkg/sentry/syscalls/linux/sys_rusage.go +++ b/pkg/sentry/syscalls/linux/sys_rusage.go @@ -80,7 +80,7 @@ func Getrusage(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys } ru := getrusage(t, which) - _, err := t.CopyOut(addr, &ru) + _, err := ru.CopyOut(t, addr) return 0, nil, err } @@ -104,7 +104,7 @@ func Times(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall CUTime: linux.ClockTFromDuration(cs2.UserTime), CSTime: linux.ClockTFromDuration(cs2.SysTime), } - if _, err := t.CopyOut(addr, &r); err != nil { + if _, err := r.CopyOut(t, addr); err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_sched.go b/pkg/sentry/syscalls/linux/sys_sched.go index 99f6993f5..bfcf44b6f 100644 --- a/pkg/sentry/syscalls/linux/sys_sched.go +++ b/pkg/sentry/syscalls/linux/sys_sched.go @@ -27,8 +27,10 @@ const ( ) // SchedParam replicates struct sched_param in sched.h. +// +// +marshal type SchedParam struct { - schedPriority int64 + schedPriority int32 } // SchedGetparam implements linux syscall sched_getparam(2). @@ -45,7 +47,7 @@ func SchedGetparam(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel return 0, nil, syserror.ESRCH } r := SchedParam{schedPriority: onlyPriority} - if _, err := t.CopyOut(param, r); err != nil { + if _, err := r.CopyOut(t, param); err != nil { return 0, nil, err } @@ -79,7 +81,7 @@ func SchedSetscheduler(t *kernel.Task, args arch.SyscallArguments) (uintptr, *ke return 0, nil, syserror.ESRCH } var r SchedParam - if _, err := t.CopyIn(param, &r); err != nil { + if _, err := r.CopyIn(t, param); err != nil { return 0, nil, syserror.EINVAL } if r.schedPriority != onlyPriority { diff --git a/pkg/sentry/syscalls/linux/sys_seccomp.go b/pkg/sentry/syscalls/linux/sys_seccomp.go index 5b7a66f4d..4fdb4463c 100644 --- a/pkg/sentry/syscalls/linux/sys_seccomp.go +++ b/pkg/sentry/syscalls/linux/sys_seccomp.go @@ -24,6 +24,8 @@ import ( ) // userSockFprog is equivalent to Linux's struct sock_fprog on amd64. +// +// +marshal type userSockFprog struct { // Len is the length of the filter in BPF instructions. Len uint16 @@ -33,7 +35,7 @@ type userSockFprog struct { // 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. + // type. Filter uint64 } @@ -54,11 +56,11 @@ func seccomp(t *kernel.Task, mode, flags uint64, addr usermem.Addr) error { } var fprog userSockFprog - if _, err := t.CopyIn(addr, &fprog); err != nil { + if _, err := fprog.CopyIn(t, addr); err != nil { return err } filter := make([]linux.BPFInstruction, int(fprog.Len)) - if _, err := t.CopyIn(usermem.Addr(fprog.Filter), &filter); err != nil { + if _, err := linux.CopyBPFInstructionSliceIn(t, usermem.Addr(fprog.Filter), filter); err != nil { return err } compiledFilter, err := bpf.Compile(filter) diff --git a/pkg/sentry/syscalls/linux/sys_sem.go b/pkg/sentry/syscalls/linux/sys_sem.go index 5f54f2456..c2d4bf805 100644 --- a/pkg/sentry/syscalls/linux/sys_sem.go +++ b/pkg/sentry/syscalls/linux/sys_sem.go @@ -18,6 +18,7 @@ import ( "math" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -66,7 +67,7 @@ func Semop(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall } ops := make([]linux.Sembuf, nsops) - if _, err := t.CopyIn(sembufAddr, ops); err != nil { + if _, err := linux.CopySembufSliceIn(t, sembufAddr, ops); err != nil { return 0, nil, err } @@ -116,8 +117,8 @@ func Semctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal case linux.IPC_SET: arg := args[3].Pointer() - s := linux.SemidDS{} - if _, err := t.CopyIn(arg, &s); err != nil { + var s linux.SemidDS + if _, err := s.CopyIn(t, arg); err != nil { return 0, nil, err } @@ -128,9 +129,17 @@ func Semctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal v, err := getPID(t, id, num) return uintptr(v), nil, err + case linux.IPC_STAT: + arg := args[3].Pointer() + ds, err := ipcStat(t, id) + if err == nil { + _, err = ds.CopyOut(t, arg) + } + + return 0, nil, err + case linux.IPC_INFO, linux.SEM_INFO, - linux.IPC_STAT, linux.SEM_STAT, linux.SEM_STAT_ANY, linux.GETNCNT, @@ -170,6 +179,16 @@ func ipcSet(t *kernel.Task, id int32, uid auth.UID, gid auth.GID, perms fs.FileP return set.Change(t, creds, owner, perms) } +func ipcStat(t *kernel.Task, id int32) (*linux.SemidDS, error) { + r := t.IPCNamespace().SemaphoreRegistry() + set := r.FindByID(id) + if set == nil { + return nil, syserror.EINVAL + } + creds := auth.CredentialsFromContext(t) + return set.GetStat(creds) +} + func setVal(t *kernel.Task, id int32, num int32, val int16) error { r := t.IPCNamespace().SemaphoreRegistry() set := r.FindByID(id) @@ -188,7 +207,7 @@ func setValAll(t *kernel.Task, id int32, array usermem.Addr) error { return syserror.EINVAL } vals := make([]uint16, set.Size()) - if _, err := t.CopyIn(array, vals); err != nil { + if _, err := primitive.CopyUint16SliceIn(t, array, vals); err != nil { return err } creds := auth.CredentialsFromContext(t) @@ -217,7 +236,7 @@ func getValAll(t *kernel.Task, id int32, array usermem.Addr) error { if err != nil { return err } - _, err = t.CopyOut(array, vals) + _, err = primitive.CopyUint16SliceOut(t, array, vals) return err } diff --git a/pkg/sentry/syscalls/linux/sys_shm.go b/pkg/sentry/syscalls/linux/sys_shm.go index f0ae8fa8e..584064143 100644 --- a/pkg/sentry/syscalls/linux/sys_shm.go +++ b/pkg/sentry/syscalls/linux/sys_shm.go @@ -112,18 +112,18 @@ func Shmctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal stat, err := segment.IPCStat(t) if err == nil { - _, err = t.CopyOut(buf, stat) + _, err = stat.CopyOut(t, buf) } return 0, nil, err case linux.IPC_INFO: params := r.IPCInfo() - _, err := t.CopyOut(buf, params) + _, err := params.CopyOut(t, buf) return 0, nil, err case linux.SHM_INFO: info := r.ShmInfo() - _, err := t.CopyOut(buf, info) + _, err := info.CopyOut(t, buf) return 0, nil, err } @@ -137,11 +137,10 @@ func Shmctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal switch cmd { case linux.IPC_SET: var ds linux.ShmidDS - _, err = t.CopyIn(buf, &ds) - if err != nil { + if _, err = ds.CopyIn(t, buf); err != nil { return 0, nil, err } - err = segment.Set(t, &ds) + err := segment.Set(t, &ds) return 0, nil, err case linux.IPC_RMID: diff --git a/pkg/sentry/syscalls/linux/sys_socket.go b/pkg/sentry/syscalls/linux/sys_socket.go index 38f573c14..9cd052c3d 100644 --- a/pkg/sentry/syscalls/linux/sys_socket.go +++ b/pkg/sentry/syscalls/linux/sys_socket.go @@ -19,6 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -29,8 +31,6 @@ import ( "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) // LINT.IfChange @@ -67,10 +67,10 @@ const flagsOffset = 48 const sizeOfInt32 = 4 // messageHeader64Len is the length of a MessageHeader64 struct. -var messageHeader64Len = uint64(binary.Size(MessageHeader64{})) +var messageHeader64Len = uint64((*MessageHeader64)(nil).SizeBytes()) // multipleMessageHeader64Len is the length of a multipeMessageHeader64 struct. -var multipleMessageHeader64Len = uint64(binary.Size(multipleMessageHeader64{})) +var multipleMessageHeader64Len = uint64((*multipleMessageHeader64)(nil).SizeBytes()) // baseRecvFlags are the flags that are accepted across recvmsg(2), // recvmmsg(2), and recvfrom(2). @@ -78,6 +78,8 @@ const baseRecvFlags = linux.MSG_OOB | linux.MSG_DONTROUTE | linux.MSG_DONTWAIT | // MessageHeader64 is the 64-bit representation of the msghdr struct used in // the recvmsg and sendmsg syscalls. +// +// +marshal type MessageHeader64 struct { // Name is the optional pointer to a network address buffer. Name uint64 @@ -106,30 +108,14 @@ type MessageHeader64 struct { // multipleMessageHeader64 is the 64-bit representation of the mmsghdr struct used in // the recvmmsg and sendmmsg syscalls. +// +// +marshal 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) { @@ -148,10 +134,10 @@ func CaptureAddress(t *kernel.Task, addr usermem.Addr, addrlen uint32) ([]byte, // 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 { +func writeAddress(t *kernel.Task, addr linux.SockAddr, addrLen uint32, addrPtr usermem.Addr, addrLenPtr usermem.Addr) error { // Get the buffer length. var bufLen uint32 - if _, err := t.CopyIn(addrLenPtr, &bufLen); err != nil { + if _, err := primitive.CopyUint32In(t, addrLenPtr, &bufLen); err != nil { return err } @@ -160,7 +146,7 @@ func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr user } // Write the length unconditionally. - if _, err := t.CopyOut(addrLenPtr, addrLen); err != nil { + if _, err := primitive.CopyUint32Out(t, addrLenPtr, addrLen); err != nil { return err } @@ -173,7 +159,8 @@ func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr user } // Copy as much of the address as will fit in the buffer. - encodedAddr := binary.Marshal(nil, usermem.ByteOrder, addr) + encodedAddr := t.CopyScratchBuffer(addr.SizeBytes()) + addr.MarshalUnsafe(encodedAddr) if bufLen > uint32(len(encodedAddr)) { bufLen = uint32(len(encodedAddr)) } @@ -247,9 +234,9 @@ func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } // Copy the file descriptors out. - if _, err := t.CopyOut(socks, fds); err != nil { + if _, err := primitive.CopyInt32SliceOut(t, socks, fds); err != nil { for _, fd := range fds { - if file, _ := t.FDTable().Remove(fd); file != nil { + if file, _ := t.FDTable().Remove(t, fd); file != nil { file.DecRef(t) } } @@ -456,8 +443,8 @@ func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } // Read the length. Reject negative values. - optLen := int32(0) - if _, err := t.CopyIn(optLenAddr, &optLen); err != nil { + var optLen int32 + if _, err := primitive.CopyInt32In(t, optLenAddr, &optLen); err != nil { return 0, nil, err } if optLen < 0 { @@ -471,7 +458,7 @@ func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } vLen := int32(binary.Size(v)) - if _, err := t.CopyOut(optLenAddr, vLen); err != nil { + if _, err := primitive.CopyInt32Out(t, optLenAddr, vLen); err != nil { return 0, nil, err } @@ -733,7 +720,7 @@ func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if !ok { return 0, nil, syserror.EFAULT } - if _, err = t.CopyOut(lp, uint32(n)); err != nil { + if _, err = primitive.CopyUint32Out(t, lp, uint32(n)); err != nil { break } count++ @@ -748,7 +735,7 @@ func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc 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 { + if _, err := msg.CopyIn(t, msgPtr); err != nil { return 0, err } @@ -780,7 +767,7 @@ func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags i if int(msg.Flags) != mflags { // Copy out the flags to the caller. - if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + if _, err := primitive.CopyInt32Out(t, msgPtr+flagsOffset, int32(mflags)); err != nil { return 0, err } } @@ -817,17 +804,17 @@ func recvSingleMsg(t *kernel.Task, s socket.Socket, msgPtr usermem.Addr, flags i } // Copy the control data to the caller. - if _, err := t.CopyOut(msgPtr+controlLenOffset, uint64(len(controlData))); err != nil { + if _, err := primitive.CopyUint64Out(t, 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 { + if _, err := t.CopyOutBytes(usermem.Addr(msg.Control), controlData); err != nil { return 0, err } } // Copy out the flags to the caller. - if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + if _, err := primitive.CopyInt32Out(t, msgPtr+flagsOffset, int32(mflags)); err != nil { return 0, err } @@ -996,7 +983,7 @@ func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if !ok { return 0, nil, syserror.EFAULT } - if _, err = t.CopyOut(lp, uint32(n)); err != nil { + if _, err = primitive.CopyUint32Out(t, lp, uint32(n)); err != nil { break } count++ @@ -1011,7 +998,7 @@ func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc 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 { + if _, err := msg.CopyIn(t, msgPtr); err != nil { return 0, err } @@ -1022,7 +1009,7 @@ func sendSingleMsg(t *kernel.Task, s socket.Socket, file *fs.File, msgPtr userme return 0, syserror.ENOBUFS } controlData = make([]byte, msg.ControlLen) - if _, err := t.CopyIn(usermem.Addr(msg.Control), &controlData); err != nil { + if _, err := t.CopyInBytes(usermem.Addr(msg.Control), controlData); err != nil { return 0, err } } @@ -1065,7 +1052,9 @@ func sendSingleMsg(t *kernel.Task, s socket.Socket, file *fs.File, msgPtr userme // Call the syscall implementation. n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, controlMessages) err = handleIOError(t, n != 0, e.ToError(), syserror.ERESTARTSYS, "sendmsg", file) - if err != nil { + // Control messages should be released on error as well as for zero-length + // messages, which are discarded by the receiver. + if n == 0 || err != nil { controlMessages.Release(t) } return uintptr(n), err diff --git a/pkg/sentry/syscalls/linux/sys_splice.go b/pkg/sentry/syscalls/linux/sys_splice.go index c69941feb..46616c961 100644 --- a/pkg/sentry/syscalls/linux/sys_splice.go +++ b/pkg/sentry/syscalls/linux/sys_splice.go @@ -16,6 +16,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -141,7 +142,7 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc // Copy in the offset. var offset int64 - if _, err := t.CopyIn(offsetAddr, &offset); err != nil { + if _, err := primitive.CopyInt64In(t, offsetAddr, &offset); err != nil { return 0, nil, err } @@ -149,11 +150,11 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc n, err = doSplice(t, outFile, inFile, fs.SpliceOpts{ Length: count, SrcOffset: true, - SrcStart: offset, + SrcStart: int64(offset), }, outFile.Flags().NonBlocking) // Copy out the new offset. - if _, err := t.CopyOut(offsetAddr, n+offset); err != nil { + if _, err := primitive.CopyInt64Out(t, offsetAddr, offset+n); err != nil { return 0, nil, err } } else { @@ -228,7 +229,7 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal } var offset int64 - if _, err := t.CopyIn(outOffset, &offset); err != nil { + if _, err := primitive.CopyInt64In(t, outOffset, &offset); err != nil { return 0, nil, err } @@ -246,7 +247,7 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal } var offset int64 - if _, err := t.CopyIn(inOffset, &offset); err != nil { + if _, err := primitive.CopyInt64In(t, inOffset, &offset); err != nil { return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_stat.go b/pkg/sentry/syscalls/linux/sys_stat.go index a5826f2dd..cda29a8b5 100644 --- a/pkg/sentry/syscalls/linux/sys_stat.go +++ b/pkg/sentry/syscalls/linux/sys_stat.go @@ -221,7 +221,7 @@ func statx(t *kernel.Task, sattr fs.StableAttr, uattr fs.UnstableAttr, statxAddr DevMajor: uint32(devMajor), DevMinor: devMinor, } - _, err := t.CopyOut(statxAddr, &s) + _, err := s.CopyOut(t, statxAddr) return err } @@ -283,7 +283,7 @@ func statfsImpl(t *kernel.Task, d *fs.Dirent, addr usermem.Addr) error { FragmentSize: d.Inode.StableAttr.BlockSize, // Leave other fields 0 like simple_statfs does. } - _, err = t.CopyOut(addr, &statfs) + _, err = statfs.CopyOut(t, addr) return err } diff --git a/pkg/sentry/syscalls/linux/sys_sysinfo.go b/pkg/sentry/syscalls/linux/sys_sysinfo.go index 297de052a..db3d924d9 100644 --- a/pkg/sentry/syscalls/linux/sys_sysinfo.go +++ b/pkg/sentry/syscalls/linux/sys_sysinfo.go @@ -21,13 +21,17 @@ import ( "gvisor.dev/gvisor/pkg/sentry/usage" ) -// Sysinfo implements the sysinfo syscall as described in man 2 sysinfo. +// Sysinfo implements Linux syscall sysinfo(2). func Sysinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { addr := args[0].Pointer() mf := t.Kernel().MemoryFile() - mf.UpdateUsage() - _, totalUsage := usage.MemoryAccounting.Copy() + mfUsage, err := mf.TotalUsage() + if err != nil { + return 0, nil, err + } + memStats, _ := usage.MemoryAccounting.Copy() + totalUsage := mfUsage + memStats.Mapped totalSize := usage.TotalMemory(mf.TotalSize(), totalUsage) memFree := totalSize - totalUsage if memFree > totalSize { @@ -37,12 +41,12 @@ func Sysinfo(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca // Only a subset of the fields in sysinfo_t make sense to return. si := linux.Sysinfo{ - Procs: uint16(len(t.PIDNamespace().Tasks())), + Procs: uint16(t.Kernel().TaskSet().Root.NumTasks()), Uptime: t.Kernel().MonotonicClock().Now().Seconds(), TotalRAM: totalSize, FreeRAM: memFree, Unit: 1, } - _, err := t.CopyOut(addr, si) + _, err = si.CopyOut(t, addr) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_thread.go b/pkg/sentry/syscalls/linux/sys_thread.go index 101096038..39ca9ea97 100644 --- a/pkg/sentry/syscalls/linux/sys_thread.go +++ b/pkg/sentry/syscalls/linux/sys_thread.go @@ -19,6 +19,7 @@ import ( "syscall" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fs" "gvisor.dev/gvisor/pkg/sentry/fsbridge" @@ -311,13 +312,13 @@ func wait4(t *kernel.Task, pid int, statusAddr usermem.Addr, options int, rusage return 0, err } if statusAddr != 0 { - if _, err := t.CopyOut(statusAddr, wr.Status); err != nil { + if _, err := primitive.CopyUint32Out(t, 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 { + if _, err := ru.CopyOut(t, rusageAddr); err != nil { return 0, err } } @@ -395,14 +396,14 @@ func Waitid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal // as well. if infop != 0 { var si arch.SignalInfo - _, err = t.CopyOut(infop, &si) + _, err = si.CopyOut(t, infop) } } return 0, nil, err } if rusageAddr != 0 { ru := getrusage(wr.Task, linux.RUSAGE_BOTH) - if _, err := t.CopyOut(rusageAddr, &ru); err != nil { + if _, err := ru.CopyOut(t, rusageAddr); err != nil { return 0, nil, err } } @@ -441,7 +442,7 @@ func Waitid(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal default: t.Warningf("waitid got incomprehensible wait status %d", s) } - _, err = t.CopyOut(infop, &si) + _, err = si.CopyOut(t, infop) return 0, nil, err } @@ -558,9 +559,7 @@ func Getcpu(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal // 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 { + if _, err := primitive.CopyInt32Out(t, cpu, t.CPU()); err != nil { return 0, nil, err } } diff --git a/pkg/sentry/syscalls/linux/sys_time.go b/pkg/sentry/syscalls/linux/sys_time.go index a2a24a027..c5054d2f1 100644 --- a/pkg/sentry/syscalls/linux/sys_time.go +++ b/pkg/sentry/syscalls/linux/sys_time.go @@ -19,6 +19,7 @@ import ( "time" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -168,7 +169,7 @@ func Time(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC return uintptr(r), nil, nil } - if _, err := t.CopyOut(addr, r); err != nil { + if _, err := r.CopyOut(t, addr); err != nil { return 0, nil, err } return uintptr(r), nil, nil @@ -334,8 +335,8 @@ func Gettimeofday(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. // 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) + timezone := []int32{-int32(offset) / 60, 0} + _, err := primitive.CopyInt32SliceOut(t, 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 index a4c400f87..45eef4feb 100644 --- a/pkg/sentry/syscalls/linux/sys_timer.go +++ b/pkg/sentry/syscalls/linux/sys_timer.go @@ -21,81 +21,63 @@ import ( "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" ) 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{}, syserror.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 syserror.ENOSYS - } -} - // Getitimer implements linux syscall getitimer(2). func Getitimer(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { + if t.Arch().Width() != 8 { + // Definition of linux.ItimerVal assumes 64-bit architecture. + return 0, nil, syserror.ENOSYS + } + timerID := args[0].Int() - val := args[1].Pointer() + addr := args[1].Pointer() olditv, err := t.Getitimer(timerID) if err != nil { return 0, nil, err } - return 0, nil, copyItimerValOut(t, val, &olditv) + // A NULL address is allowed, in which case no copy out takes place. + if addr == 0 { + return 0, nil, nil + } + _, err = olditv.CopyOut(t, addr) + return 0, nil, err } // Setitimer implements linux syscall setitimer(2). func Setitimer(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { - timerID := args[0].Int() - newVal := args[1].Pointer() - oldVal := args[2].Pointer() + if t.Arch().Width() != 8 { + // Definition of linux.ItimerVal assumes 64-bit architecture. + return 0, nil, syserror.ENOSYS + } - newitv, err := copyItimerValIn(t, newVal) - if err != nil { - return 0, nil, err + timerID := args[0].Int() + newAddr := args[1].Pointer() + oldAddr := args[2].Pointer() + + var newitv linux.ItimerVal + // 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. + if newAddr != 0 { + if _, err := newitv.CopyIn(t, newAddr); err != nil { + return 0, nil, err + } } olditv, err := t.Setitimer(timerID, newitv) if err != nil { return 0, nil, err } - return 0, nil, copyItimerValOut(t, oldVal, &olditv) + // A NULL address is allowed, in which case no copy out takes place. + if oldAddr == 0 { + return 0, nil, nil + } + _, err = olditv.CopyOut(t, oldAddr) + return 0, nil, err } // Alarm implements linux syscall alarm(2). @@ -131,7 +113,7 @@ func TimerCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S var sev *linux.Sigevent if sevp != 0 { sev = &linux.Sigevent{} - if _, err = t.CopyIn(sevp, sev); err != nil { + if _, err = sev.CopyIn(t, sevp); err != nil { return 0, nil, err } } @@ -141,7 +123,7 @@ func TimerCreate(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.S return 0, nil, err } - if _, err := t.CopyOut(timerIDp, &id); err != nil { + if _, err := id.CopyOut(t, timerIDp); err != nil { t.IntervalTimerDelete(id) return 0, nil, err } @@ -157,7 +139,7 @@ func TimerSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. oldValAddr := args[3].Pointer() var newVal linux.Itimerspec - if _, err := t.CopyIn(newValAddr, &newVal); err != nil { + if _, err := newVal.CopyIn(t, newValAddr); err != nil { return 0, nil, err } oldVal, err := t.IntervalTimerSettime(timerID, newVal, flags&linux.TIMER_ABSTIME != 0) @@ -165,9 +147,8 @@ func TimerSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. return 0, nil, err } if oldValAddr != 0 { - if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil { - return 0, nil, err - } + _, err = oldVal.CopyOut(t, oldValAddr) + return 0, nil, err } return 0, nil, nil } @@ -181,7 +162,7 @@ func TimerGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. if err != nil { return 0, nil, err } - _, err = t.CopyOut(curValAddr, &curVal) + _, err = curVal.CopyOut(t, curValAddr) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_timerfd.go b/pkg/sentry/syscalls/linux/sys_timerfd.go index 34b03e4ee..cadd9d348 100644 --- a/pkg/sentry/syscalls/linux/sys_timerfd.go +++ b/pkg/sentry/syscalls/linux/sys_timerfd.go @@ -81,7 +81,7 @@ func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne } var newVal linux.Itimerspec - if _, err := t.CopyIn(newValAddr, &newVal); err != nil { + if _, err := newVal.CopyIn(t, newValAddr); err != nil { return 0, nil, err } newS, err := ktime.SettingFromItimerspec(newVal, flags&linux.TFD_TIMER_ABSTIME != 0, tf.Clock()) @@ -91,7 +91,7 @@ func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne tm, oldS := tf.SetTime(newS) if oldValAddr != 0 { oldVal := ktime.ItimerspecFromSetting(tm, oldS) - if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil { + if _, err := oldVal.CopyOut(t, oldValAddr); err != nil { return 0, nil, err } } @@ -116,6 +116,6 @@ func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne tm, s := tf.GetTime() curVal := ktime.ItimerspecFromSetting(tm, s) - _, err := t.CopyOut(curValAddr, &curVal) + _, err := curVal.CopyOut(t, curValAddr) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/sys_tls_amd64.go b/pkg/sentry/syscalls/linux/sys_tls_amd64.go index b3eb96a1c..6ddd30d5c 100644 --- a/pkg/sentry/syscalls/linux/sys_tls_amd64.go +++ b/pkg/sentry/syscalls/linux/sys_tls_amd64.go @@ -18,6 +18,7 @@ package linux import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserror" @@ -30,17 +31,19 @@ func ArchPrctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys case linux.ARCH_GET_FS: addr := args[1].Pointer() fsbase := t.Arch().TLS() - _, err := t.CopyOut(addr, uint64(fsbase)) - if err != nil { - return 0, nil, err + switch t.Arch().Width() { + case 8: + if _, err := primitive.CopyUint64Out(t, addr, uint64(fsbase)); err != nil { + return 0, nil, err + } + default: + return 0, nil, syserror.ENOSYS } - case linux.ARCH_SET_FS: fsbase := args[1].Uint64() if !t.Arch().SetTLS(uintptr(fsbase)) { return 0, nil, syserror.EPERM } - case linux.ARCH_GET_GS, linux.ARCH_SET_GS: t.Kernel().EmitUnimplementedEvent(t) fallthrough diff --git a/pkg/sentry/syscalls/linux/sys_utsname.go b/pkg/sentry/syscalls/linux/sys_utsname.go index e9d702e8e..66c5974f5 100644 --- a/pkg/sentry/syscalls/linux/sys_utsname.go +++ b/pkg/sentry/syscalls/linux/sys_utsname.go @@ -46,7 +46,7 @@ func Uname(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Copy out the result. va := args[0].Pointer() - _, err := t.CopyOut(va, u) + _, err := u.CopyOut(t, va) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/vfs2/BUILD b/pkg/sentry/syscalls/linux/vfs2/BUILD index 64696b438..9ee766552 100644 --- a/pkg/sentry/syscalls/linux/vfs2/BUILD +++ b/pkg/sentry/syscalls/linux/vfs2/BUILD @@ -44,6 +44,9 @@ go_library( "//pkg/context", "//pkg/fspath", "//pkg/gohacks", + "//pkg/log", + "//pkg/marshal", + "//pkg/marshal/primitive", "//pkg/sentry/arch", "//pkg/sentry/fs/lock", "//pkg/sentry/fsbridge", @@ -72,7 +75,5 @@ go_library( "//pkg/syserror", "//pkg/usermem", "//pkg/waiter", - "//tools/go_marshal/marshal", - "//tools/go_marshal/primitive", ], ) diff --git a/pkg/sentry/syscalls/linux/vfs2/aio.go b/pkg/sentry/syscalls/linux/vfs2/aio.go index 42559bf69..6d0a38330 100644 --- a/pkg/sentry/syscalls/linux/vfs2/aio.go +++ b/pkg/sentry/syscalls/linux/vfs2/aio.go @@ -17,6 +17,7 @@ package vfs2 import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fsimpl/eventfd" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -38,21 +39,27 @@ func IoSubmit(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } 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 + // Copy in the callback address. + var cbAddr usermem.Addr + switch t.Arch().Width() { + case 8: + var cbAddrP primitive.Uint64 + if _, err := cbAddrP.CopyIn(t, addr); err != nil { + if i > 0 { + // Some successful. + return uintptr(i), nil, nil + } + // Nothing done. + return 0, nil, err } - // Nothing done. - return 0, nil, err + cbAddr = usermem.Addr(cbAddrP) + default: + return 0, nil, syserror.ENOSYS } // Copy in this callback. var cb linux.IOCallback - cbAddr := usermem.Addr(t.Arch().Value(cbAddrNative)) - if _, err := t.CopyIn(cbAddr, &cb); err != nil { + if _, err := cb.CopyIn(t, cbAddr); err != nil { if i > 0 { // Some have been successful. return uintptr(i), nil, nil diff --git a/pkg/sentry/syscalls/linux/vfs2/epoll.go b/pkg/sentry/syscalls/linux/vfs2/epoll.go index c62f03509..d0cbb77eb 100644 --- a/pkg/sentry/syscalls/linux/vfs2/epoll.go +++ b/pkg/sentry/syscalls/linux/vfs2/epoll.go @@ -24,7 +24,6 @@ import ( ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" - "gvisor.dev/gvisor/pkg/usermem" "gvisor.dev/gvisor/pkg/waiter" ) @@ -141,50 +140,26 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, syserror.EINVAL } - // Use a fixed-size buffer in a loop, instead of make([]linux.EpollEvent, - // maxEvents), so that the buffer can be allocated on the stack. + // Allocate space for a few events on the stack for the common case in + // which we don't have too many events. var ( - events [16]linux.EpollEvent - total int + eventsArr [16]linux.EpollEvent ch chan struct{} haveDeadline bool deadline ktime.Time ) for { - batchEvents := len(events) - if batchEvents > maxEvents { - batchEvents = maxEvents - } - n := ep.ReadEvents(events[:batchEvents]) - maxEvents -= n - if n != 0 { - // Copy what we read out. - copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events[:n]) + events := ep.ReadEvents(eventsArr[:0], maxEvents) + if len(events) != 0 { + copiedBytes, err := linux.CopyEpollEventSliceOut(t, eventsAddr, events) copiedEvents := copiedBytes / sizeofEpollEvent // rounded down - eventsAddr += usermem.Addr(copiedEvents * sizeofEpollEvent) - total += copiedEvents - if err != nil { - if total != 0 { - return uintptr(total), nil, nil - } - return 0, nil, err - } - // If we've filled the application's event buffer, we're done. - if maxEvents == 0 { - return uintptr(total), nil, nil - } - // Loop if we read a full batch, under the expectation that there - // may be more events to read. - if n == batchEvents { - continue + if copiedEvents != 0 { + return uintptr(copiedEvents), nil, nil } + return 0, nil, err } - // We get here if n != batchEvents. If we read any number of events - // (just now, or in a previous iteration of this loop), or if timeout - // is 0 (such that epoll_wait should be non-blocking), return the - // events we've read so far to the application. - if total != 0 || timeout == 0 { - return uintptr(total), nil, nil + if timeout == 0 { + return 0, nil, nil } // In the first iteration of this loop, register with the epoll // instance for readability events, but then immediately continue the @@ -207,8 +182,6 @@ func EpollWait(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys if err == syserror.ETIMEDOUT { err = nil } - // total must be 0 since otherwise we would have returned - // above. return 0, nil, err } } diff --git a/pkg/sentry/syscalls/linux/vfs2/execve.go b/pkg/sentry/syscalls/linux/vfs2/execve.go index 066ee0863..c8ce2aabc 100644 --- a/pkg/sentry/syscalls/linux/vfs2/execve.go +++ b/pkg/sentry/syscalls/linux/vfs2/execve.go @@ -110,8 +110,7 @@ func execveat(t *kernel.Task, dirfd int32, pathnameAddr, argvAddr, envvAddr user } // Load the new TaskContext. - mntns := t.MountNamespaceVFS2() // FIXME(jamieliu): useless refcount change - defer mntns.DecRef(t) + mntns := t.MountNamespaceVFS2() wd := t.FSContext().WorkingDirectoryVFS2() defer wd.DecRef(t) remainingTraversals := uint(linux.MaxSymlinkTraversals) diff --git a/pkg/sentry/syscalls/linux/vfs2/fd.go b/pkg/sentry/syscalls/linux/vfs2/fd.go index 4856554fe..36e89700e 100644 --- a/pkg/sentry/syscalls/linux/vfs2/fd.go +++ b/pkg/sentry/syscalls/linux/vfs2/fd.go @@ -34,7 +34,7 @@ func Close(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Note that Remove provides a reference on the file that we may use to // flush. It is still active until we drop the final reference below // (and other reference-holding operations complete). - _, file := t.FDTable().Remove(fd) + _, file := t.FDTable().Remove(t, fd) if file == nil { return 0, nil, syserror.EBADF } @@ -137,7 +137,7 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(flags.ToLinuxFDFlags()), nil, nil case linux.F_SETFD: flags := args[2].Uint() - err := t.FDTable().SetFlagsVFS2(fd, kernel.FDFlags{ + err := t.FDTable().SetFlagsVFS2(t, fd, kernel.FDFlags{ CloseOnExec: flags&linux.FD_CLOEXEC != 0, }) return 0, nil, err @@ -145,16 +145,6 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return uintptr(file.StatusFlags()), nil, nil case linux.F_SETFL: return 0, nil, file.SetStatusFlags(t, t.Credentials(), args[2].Uint()) - case linux.F_SETPIPE_SZ: - pipefile, ok := file.Impl().(*pipe.VFSPipeFD) - if !ok { - return 0, nil, syserror.EBADF - } - n, err := pipefile.SetPipeSize(int64(args[2].Int())) - if err != nil { - return 0, nil, err - } - return uintptr(n), nil, nil case linux.F_GETOWN: owner, hasOwner := getAsyncOwner(t, file) if !hasOwner { @@ -181,15 +171,25 @@ func Fcntl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall if !hasOwner { return 0, nil, nil } - _, err := t.CopyOut(args[2].Pointer(), &owner) + _, err := owner.CopyOut(t, args[2].Pointer()) return 0, nil, err case linux.F_SETOWN_EX: var owner linux.FOwnerEx - _, err := t.CopyIn(args[2].Pointer(), &owner) + _, err := owner.CopyIn(t, args[2].Pointer()) if err != nil { return 0, nil, err } return 0, nil, setAsyncOwner(t, file, owner.Type, owner.PID) + case linux.F_SETPIPE_SZ: + pipefile, ok := file.Impl().(*pipe.VFSPipeFD) + if !ok { + return 0, nil, syserror.EBADF + } + n, err := pipefile.SetPipeSize(int64(args[2].Int())) + if err != nil { + return 0, nil, err + } + return uintptr(n), nil, nil case linux.F_GETPIPE_SZ: pipefile, ok := file.Impl().(*pipe.VFSPipeFD) if !ok { @@ -286,7 +286,7 @@ func posixLock(t *kernel.Task, args arch.SyscallArguments, file *vfs.FileDescrip // Copy in the lock request. flockAddr := args[2].Pointer() var flock linux.Flock - if _, err := t.CopyIn(flockAddr, &flock); err != nil { + if _, err := flock.CopyIn(t, flockAddr); err != nil { return err } diff --git a/pkg/sentry/syscalls/linux/vfs2/ioctl.go b/pkg/sentry/syscalls/linux/vfs2/ioctl.go index 38778a388..2806c3f6f 100644 --- a/pkg/sentry/syscalls/linux/vfs2/ioctl.go +++ b/pkg/sentry/syscalls/linux/vfs2/ioctl.go @@ -16,6 +16,7 @@ package vfs2 import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/syserror" @@ -34,20 +35,20 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall // Handle ioctls that apply to all FDs. switch args[1].Int() { case linux.FIONCLEX: - t.FDTable().SetFlagsVFS2(fd, kernel.FDFlags{ + t.FDTable().SetFlagsVFS2(t, fd, kernel.FDFlags{ CloseOnExec: false, }) return 0, nil, nil case linux.FIOCLEX: - t.FDTable().SetFlagsVFS2(fd, kernel.FDFlags{ + t.FDTable().SetFlagsVFS2(t, 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 { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil { return 0, nil, err } flags := file.StatusFlags() @@ -60,7 +61,7 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall case linux.FIOASYNC: var set int32 - if _, err := t.CopyIn(args[2].Pointer(), &set); err != nil { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &set); err != nil { return 0, nil, err } flags := file.StatusFlags() @@ -82,12 +83,12 @@ func Ioctl(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall who = owner.PID } } - _, err := t.CopyOut(args[2].Pointer(), &who) + _, err := primitive.CopyInt32Out(t, args[2].Pointer(), who) return 0, nil, err case linux.FIOSETOWN, linux.SIOCSPGRP: var who int32 - if _, err := t.CopyIn(args[2].Pointer(), &who); err != nil { + if _, err := primitive.CopyInt32In(t, args[2].Pointer(), &who); err != nil { return 0, nil, err } ownerType := int32(linux.F_OWNER_PID) diff --git a/pkg/sentry/syscalls/linux/vfs2/mmap.go b/pkg/sentry/syscalls/linux/vfs2/mmap.go index dc05c2994..9d9dbf775 100644 --- a/pkg/sentry/syscalls/linux/vfs2/mmap.go +++ b/pkg/sentry/syscalls/linux/vfs2/mmap.go @@ -17,6 +17,7 @@ package vfs2 import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/sentry/arch" + "gvisor.dev/gvisor/pkg/sentry/fsimpl/tmpfs" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/memmap" "gvisor.dev/gvisor/pkg/syserror" @@ -85,6 +86,17 @@ func Mmap(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallC if err := file.ConfigureMMap(t, &opts); err != nil { return 0, nil, err } + } else if shared { + // Back shared anonymous mappings with an anonymous tmpfs file. + opts.Offset = 0 + file, err := tmpfs.NewZeroFile(t, t.Credentials(), t.Kernel().ShmMount(), opts.Length) + if err != nil { + return 0, nil, err + } + defer file.DecRef(t) + if err := file.ConfigureMMap(t, &opts); err != nil { + return 0, nil, err + } } rv, err := t.MemoryManager().MMap(t, opts) diff --git a/pkg/sentry/syscalls/linux/vfs2/mount.go b/pkg/sentry/syscalls/linux/vfs2/mount.go index 4bd5c7ca2..769c9b92f 100644 --- a/pkg/sentry/syscalls/linux/vfs2/mount.go +++ b/pkg/sentry/syscalls/linux/vfs2/mount.go @@ -109,8 +109,8 @@ func Mount(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscall return 0, nil, err } defer target.Release(t) - - return 0, nil, t.Kernel().VFS().MountAt(t, creds, source, &target.pop, fsType, &opts) + _, err = t.Kernel().VFS().MountAt(t, creds, source, &target.pop, fsType, &opts) + return 0, nil, err } // Umount2 implements Linux syscall umount2(2). diff --git a/pkg/sentry/syscalls/linux/vfs2/pipe.go b/pkg/sentry/syscalls/linux/vfs2/pipe.go index 9b4848d9e..ee38fdca0 100644 --- a/pkg/sentry/syscalls/linux/vfs2/pipe.go +++ b/pkg/sentry/syscalls/linux/vfs2/pipe.go @@ -16,6 +16,7 @@ package vfs2 import ( "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/fsimpl/pipefs" "gvisor.dev/gvisor/pkg/sentry/kernel" @@ -51,9 +52,9 @@ func pipe2(t *kernel.Task, addr usermem.Addr, flags int32) error { if err != nil { return err } - if _, err := t.CopyOut(addr, fds); err != nil { + if _, err := primitive.CopyInt32SliceOut(t, addr, fds); err != nil { for _, fd := range fds { - if _, file := t.FDTable().Remove(fd); file != nil { + if _, file := t.FDTable().Remove(t, fd); file != nil { file.DecRef(t) } } diff --git a/pkg/sentry/syscalls/linux/vfs2/poll.go b/pkg/sentry/syscalls/linux/vfs2/poll.go index 79ad64039..c22e4ce54 100644 --- a/pkg/sentry/syscalls/linux/vfs2/poll.go +++ b/pkg/sentry/syscalls/linux/vfs2/poll.go @@ -165,7 +165,7 @@ func copyInPollFDs(t *kernel.Task, addr usermem.Addr, nfds uint) ([]linux.PollFD pfd := make([]linux.PollFD, nfds) if nfds > 0 { - if _, err := t.CopyIn(addr, &pfd); err != nil { + if _, err := linux.CopyPollFDSliceIn(t, addr, pfd); err != nil { return nil, err } } @@ -192,7 +192,7 @@ func doPoll(t *kernel.Task, addr usermem.Addr, nfds uint, timeout time.Duration) // 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(addr, pfd); err != nil { + if _, err := linux.CopyPollFDSliceOut(t, addr, pfd); err != nil { return remainingTimeout, 0, err } } @@ -205,7 +205,7 @@ func CopyInFDSet(t *kernel.Task, addr usermem.Addr, nBytes, nBitsInLastPartialBy set := make([]byte, nBytes) if addr != 0 { - if _, err := t.CopyIn(addr, &set); err != nil { + if _, err := t.CopyInBytes(addr, set); err != nil { return nil, err } // If we only use part of the last byte, mask out the extraneous bits. @@ -332,19 +332,19 @@ func doSelect(t *kernel.Task, nfds int, readFDs, writeFDs, exceptFDs usermem.Add // Copy updated vectors back. if readFDs != 0 { - if _, err := t.CopyOut(readFDs, r); err != nil { + if _, err := t.CopyOutBytes(readFDs, r); err != nil { return 0, err } } if writeFDs != 0 { - if _, err := t.CopyOut(writeFDs, w); err != nil { + if _, err := t.CopyOutBytes(writeFDs, w); err != nil { return 0, err } } if exceptFDs != 0 { - if _, err := t.CopyOut(exceptFDs, e); err != nil { + if _, err := t.CopyOutBytes(exceptFDs, e); err != nil { return 0, err } } @@ -497,6 +497,12 @@ func Select(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal return n, nil, err } +// +marshal +type sigSetWithSize struct { + sigsetAddr uint64 + sizeofSigset uint64 +} + // 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. @@ -538,12 +544,6 @@ func Pselect(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysca return n, nil, err } -// +marshal -type sigSetWithSize struct { - sigsetAddr uint64 - sizeofSigset uint64 -} - // copyTimespecInToDuration copies a Timespec from the untrusted app range, // validates it and converts it to a Duration. // diff --git a/pkg/sentry/syscalls/linux/vfs2/setstat.go b/pkg/sentry/syscalls/linux/vfs2/setstat.go index 5e6eb13ba..1ee37e5a8 100644 --- a/pkg/sentry/syscalls/linux/vfs2/setstat.go +++ b/pkg/sentry/syscalls/linux/vfs2/setstat.go @@ -346,7 +346,7 @@ func populateSetStatOptionsForUtimes(t *kernel.Task, timesAddr usermem.Addr, opt return nil } var times [2]linux.Timeval - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := linux.CopyTimevalSliceIn(t, timesAddr, times[:]); err != nil { return err } if times[0].Usec < 0 || times[0].Usec > 999999 || times[1].Usec < 0 || times[1].Usec > 999999 { @@ -410,7 +410,7 @@ func populateSetStatOptionsForUtimens(t *kernel.Task, timesAddr usermem.Addr, op return nil } var times [2]linux.Timespec - if _, err := t.CopyIn(timesAddr, ×); err != nil { + if _, err := linux.CopyTimespecSliceIn(t, timesAddr, times[:]); err != nil { return err } if times[0].Nsec != linux.UTIME_OMIT { diff --git a/pkg/sentry/syscalls/linux/vfs2/socket.go b/pkg/sentry/syscalls/linux/vfs2/socket.go index a5032657a..7b33b3f59 100644 --- a/pkg/sentry/syscalls/linux/vfs2/socket.go +++ b/pkg/sentry/syscalls/linux/vfs2/socket.go @@ -19,6 +19,8 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/binary" + "gvisor.dev/gvisor/pkg/marshal" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" ktime "gvisor.dev/gvisor/pkg/sentry/kernel/time" @@ -30,8 +32,6 @@ import ( "gvisor.dev/gvisor/pkg/syserr" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" - "gvisor.dev/gvisor/tools/go_marshal/marshal" - "gvisor.dev/gvisor/tools/go_marshal/primitive" ) // minListenBacklog is the minimum reasonable backlog for listening sockets. @@ -66,10 +66,10 @@ const flagsOffset = 48 const sizeOfInt32 = 4 // messageHeader64Len is the length of a MessageHeader64 struct. -var messageHeader64Len = uint64(binary.Size(MessageHeader64{})) +var messageHeader64Len = uint64((*MessageHeader64)(nil).SizeBytes()) // multipleMessageHeader64Len is the length of a multipeMessageHeader64 struct. -var multipleMessageHeader64Len = uint64(binary.Size(multipleMessageHeader64{})) +var multipleMessageHeader64Len = uint64((*multipleMessageHeader64)(nil).SizeBytes()) // baseRecvFlags are the flags that are accepted across recvmsg(2), // recvmmsg(2), and recvfrom(2). @@ -77,6 +77,8 @@ const baseRecvFlags = linux.MSG_OOB | linux.MSG_DONTROUTE | linux.MSG_DONTWAIT | // MessageHeader64 is the 64-bit representation of the msghdr struct used in // the recvmsg and sendmsg syscalls. +// +// +marshal type MessageHeader64 struct { // Name is the optional pointer to a network address buffer. Name uint64 @@ -105,30 +107,14 @@ type MessageHeader64 struct { // multipleMessageHeader64 is the 64-bit representation of the mmsghdr struct used in // the recvmmsg and sendmmsg syscalls. +// +// +marshal 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) { @@ -147,10 +133,10 @@ func CaptureAddress(t *kernel.Task, addr usermem.Addr, addrlen uint32) ([]byte, // 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 { +func writeAddress(t *kernel.Task, addr linux.SockAddr, addrLen uint32, addrPtr usermem.Addr, addrLenPtr usermem.Addr) error { // Get the buffer length. var bufLen uint32 - if _, err := t.CopyIn(addrLenPtr, &bufLen); err != nil { + if _, err := primitive.CopyUint32In(t, addrLenPtr, &bufLen); err != nil { return err } @@ -159,7 +145,7 @@ func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr user } // Write the length unconditionally. - if _, err := t.CopyOut(addrLenPtr, addrLen); err != nil { + if _, err := primitive.CopyUint32Out(t, addrLenPtr, addrLen); err != nil { return err } @@ -172,7 +158,8 @@ func writeAddress(t *kernel.Task, addr interface{}, addrLen uint32, addrPtr user } // Copy as much of the address as will fit in the buffer. - encodedAddr := binary.Marshal(nil, usermem.ByteOrder, addr) + encodedAddr := t.CopyScratchBuffer(addr.SizeBytes()) + addr.MarshalUnsafe(encodedAddr) if bufLen > uint32(len(encodedAddr)) { bufLen = uint32(len(encodedAddr)) } @@ -250,9 +237,9 @@ func SocketPair(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return 0, nil, err } - if _, err := t.CopyOut(addr, fds); err != nil { + if _, err := primitive.CopyInt32SliceOut(t, addr, fds); err != nil { for _, fd := range fds { - if _, file := t.FDTable().Remove(fd); file != nil { + if _, file := t.FDTable().Remove(t, fd); file != nil { file.DecRef(t) } } @@ -459,8 +446,8 @@ func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } // Read the length. Reject negative values. - optLen := int32(0) - if _, err := t.CopyIn(optLenAddr, &optLen); err != nil { + var optLen int32 + if _, err := primitive.CopyInt32In(t, optLenAddr, &optLen); err != nil { return 0, nil, err } if optLen < 0 { @@ -474,7 +461,7 @@ func GetSockOpt(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } vLen := int32(binary.Size(v)) - if _, err := t.CopyOut(optLenAddr, vLen); err != nil { + if _, err := primitive.CopyInt32Out(t, optLenAddr, vLen); err != nil { return 0, nil, err } @@ -736,7 +723,7 @@ func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if !ok { return 0, nil, syserror.EFAULT } - if _, err = t.CopyOut(lp, uint32(n)); err != nil { + if _, err = primitive.CopyUint32Out(t, lp, uint32(n)); err != nil { break } count++ @@ -751,7 +738,7 @@ func RecvMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, 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 { + if _, err := msg.CopyIn(t, msgPtr); err != nil { return 0, err } @@ -783,7 +770,7 @@ func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, msgPtr usermem.Addr, fla if int(msg.Flags) != mflags { // Copy out the flags to the caller. - if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + if _, err := primitive.CopyInt32Out(t, msgPtr+flagsOffset, int32(mflags)); err != nil { return 0, err } } @@ -820,17 +807,17 @@ func recvSingleMsg(t *kernel.Task, s socket.SocketVFS2, msgPtr usermem.Addr, fla } // Copy the control data to the caller. - if _, err := t.CopyOut(msgPtr+controlLenOffset, uint64(len(controlData))); err != nil { + if _, err := primitive.CopyUint64Out(t, 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 { + if _, err := t.CopyOutBytes(usermem.Addr(msg.Control), controlData); err != nil { return 0, err } } // Copy out the flags to the caller. - if _, err := t.CopyOut(msgPtr+flagsOffset, int32(mflags)); err != nil { + if _, err := primitive.CopyInt32Out(t, msgPtr+flagsOffset, int32(mflags)); err != nil { return 0, err } @@ -999,7 +986,7 @@ func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if !ok { return 0, nil, syserror.EFAULT } - if _, err = t.CopyOut(lp, uint32(n)); err != nil { + if _, err = primitive.CopyUint32Out(t, lp, uint32(n)); err != nil { break } count++ @@ -1014,7 +1001,7 @@ func SendMMsg(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc func sendSingleMsg(t *kernel.Task, s socket.SocketVFS2, file *vfs.FileDescription, msgPtr usermem.Addr, flags int32) (uintptr, error) { // Capture the message header. var msg MessageHeader64 - if err := CopyInMessageHeader64(t, msgPtr, &msg); err != nil { + if _, err := msg.CopyIn(t, msgPtr); err != nil { return 0, err } @@ -1025,7 +1012,7 @@ func sendSingleMsg(t *kernel.Task, s socket.SocketVFS2, file *vfs.FileDescriptio return 0, syserror.ENOBUFS } controlData = make([]byte, msg.ControlLen) - if _, err := t.CopyIn(usermem.Addr(msg.Control), &controlData); err != nil { + if _, err := t.CopyInBytes(usermem.Addr(msg.Control), controlData); err != nil { return 0, err } } @@ -1068,7 +1055,9 @@ func sendSingleMsg(t *kernel.Task, s socket.SocketVFS2, file *vfs.FileDescriptio // Call the syscall implementation. n, e := s.SendMsg(t, src, to, int(flags), haveDeadline, deadline, controlMessages) err = slinux.HandleIOErrorVFS2(t, n != 0, e.ToError(), syserror.ERESTARTSYS, "sendmsg", file) - if err != nil { + // Control messages should be released on error as well as for zero-length + // messages, which are discarded by the receiver. + if n == 0 || err != nil { controlMessages.Release(t) } return uintptr(n), err diff --git a/pkg/sentry/syscalls/linux/vfs2/splice.go b/pkg/sentry/syscalls/linux/vfs2/splice.go index 75bfa2c79..035e2a6b0 100644 --- a/pkg/sentry/syscalls/linux/vfs2/splice.go +++ b/pkg/sentry/syscalls/linux/vfs2/splice.go @@ -18,9 +18,12 @@ import ( "io" "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/log" + "gvisor.dev/gvisor/pkg/marshal/primitive" "gvisor.dev/gvisor/pkg/sentry/arch" "gvisor.dev/gvisor/pkg/sentry/kernel" "gvisor.dev/gvisor/pkg/sentry/kernel/pipe" + slinux "gvisor.dev/gvisor/pkg/sentry/syscalls/linux" "gvisor.dev/gvisor/pkg/sentry/vfs" "gvisor.dev/gvisor/pkg/syserror" "gvisor.dev/gvisor/pkg/usermem" @@ -42,6 +45,9 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if count > int64(kernel.MAX_RW_COUNT) { count = int64(kernel.MAX_RW_COUNT) } + if count < 0 { + return 0, nil, syserror.EINVAL + } // Check for invalid flags. if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { @@ -88,7 +94,7 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if inFile.Options().DenyPRead { return 0, nil, syserror.EINVAL } - if _, err := t.CopyIn(inOffsetPtr, &inOffset); err != nil { + if _, err := primitive.CopyInt64In(t, inOffsetPtr, &inOffset); err != nil { return 0, nil, err } if inOffset < 0 { @@ -103,7 +109,7 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal if outFile.Options().DenyPWrite { return 0, nil, syserror.EINVAL } - if _, err := t.CopyIn(outOffsetPtr, &outOffset); err != nil { + if _, err := primitive.CopyInt64In(t, outOffsetPtr, &outOffset); err != nil { return 0, nil, err } if outOffset < 0 { @@ -131,21 +137,17 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal case inIsPipe && outIsPipe: n, err = pipe.Splice(t, outPipeFD, inPipeFD, count) case inIsPipe: + n, err = inPipeFD.SpliceToNonPipe(t, outFile, outOffset, count) if outOffset != -1 { - n, err = outFile.PWrite(t, inPipeFD.IOSequence(count), outOffset, vfs.WriteOptions{}) outOffset += n - } else { - n, err = outFile.Write(t, inPipeFD.IOSequence(count), vfs.WriteOptions{}) } case outIsPipe: + n, err = outPipeFD.SpliceFromNonPipe(t, inFile, inOffset, count) if inOffset != -1 { - n, err = inFile.PRead(t, outPipeFD.IOSequence(count), inOffset, vfs.ReadOptions{}) inOffset += n - } else { - n, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{}) } default: - panic("not possible") + panic("at least one end of splice must be a pipe") } if n != 0 || err != syserror.ErrWouldBlock || nonBlock { @@ -158,25 +160,26 @@ func Splice(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Syscal // Copy updated offsets out. if inOffsetPtr != 0 { - if _, err := t.CopyOut(inOffsetPtr, &inOffset); err != nil { + if _, err := primitive.CopyInt64Out(t, inOffsetPtr, inOffset); err != nil { return 0, nil, err } } if outOffsetPtr != 0 { - if _, err := t.CopyOut(outOffsetPtr, &outOffset); err != nil { + if _, err := primitive.CopyInt64Out(t, outOffsetPtr, outOffset); err != nil { return 0, nil, err } } - if n == 0 { - return 0, nil, err + if n != 0 { + // On Linux, inotify behavior is not very consistent with splice(2). We try + // our best to emulate Linux for very basic calls to splice, where for some + // reason, events are generated for output files, but not input files. + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) } - // On Linux, inotify behavior is not very consistent with splice(2). We try - // our best to emulate Linux for very basic calls to splice, where for some - // reason, events are generated for output files, but not input files. - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "splice", outFile) } // Tee implements Linux syscall tee(2). @@ -192,6 +195,9 @@ func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo if count > int64(kernel.MAX_RW_COUNT) { count = int64(kernel.MAX_RW_COUNT) } + if count < 0 { + return 0, nil, syserror.EINVAL + } // Check for invalid flags. if flags&^(linux.SPLICE_F_MOVE|linux.SPLICE_F_NONBLOCK|linux.SPLICE_F_MORE|linux.SPLICE_F_GIFT) != 0 { @@ -248,11 +254,20 @@ func Tee(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallCo break } } - if n == 0 { - return 0, nil, err + + if n != 0 { + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) + + // If a partial write is completed, the error is dropped. Log it here. + if err != nil && err != io.EOF && err != syserror.ErrWouldBlock { + log.Debugf("tee completed a partial write with error: %v", err) + err = nil + } } - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "tee", inFile) } // Sendfile implements linux system call sendfile(2). @@ -301,9 +316,12 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if inFile.Options().DenyPRead { return 0, nil, syserror.ESPIPE } - if _, err := t.CopyIn(offsetAddr, &offset); err != nil { + var offsetP primitive.Int64 + if _, err := offsetP.CopyIn(t, offsetAddr); err != nil { return 0, nil, err } + offset = int64(offsetP) + if offset < 0 { return 0, nil, syserror.EINVAL } @@ -341,16 +359,9 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if outIsPipe { for n < count { var spliceN int64 + spliceN, err = outPipeFD.SpliceFromNonPipe(t, inFile, offset, count) if offset != -1 { - spliceN, err = inFile.PRead(t, outPipeFD.IOSequence(count), offset, vfs.ReadOptions{}) offset += spliceN - } else { - spliceN, err = inFile.Read(t, outPipeFD.IOSequence(count), vfs.ReadOptions{}) - } - if spliceN == 0 && err == io.EOF { - // We reached the end of the file. Eat the error and exit the loop. - err = nil - break } n += spliceN if err == syserror.ErrWouldBlock && !nonBlock { @@ -371,19 +382,11 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc } else { readN, err = inFile.Read(t, usermem.BytesIOSequence(buf), vfs.ReadOptions{}) } - if readN == 0 && err == io.EOF { - // We reached the end of the file. Eat the error and exit the loop. - err = nil - break - } n += readN - if err != nil { - break - } // Write all of the bytes that we read. This may need // multiple write calls to complete. - wbuf := buf[:n] + wbuf := buf[:readN] for len(wbuf) > 0 { var writeN int64 writeN, err = outFile.Write(t, usermem.BytesIOSequence(wbuf), vfs.WriteOptions{}) @@ -392,12 +395,21 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc err = dw.waitForOut(t) } if err != nil { - // We didn't complete the write. Only - // report the bytes that were actually - // written, and rewind the offset. + // We didn't complete the write. Only report the bytes that were actually + // written, and rewind offsets as needed. notWritten := int64(len(wbuf)) n -= notWritten - if offset != -1 { + if offset == -1 { + // We modified the offset of the input file itself during the read + // operation. Rewind it. + if _, seekErr := inFile.Seek(t, -notWritten, linux.SEEK_CUR); seekErr != nil { + // Log the error but don't return it, since the write has already + // completed successfully. + log.Warningf("failed to roll back input file offset: %v", seekErr) + } + } else { + // The sendfile call was provided an offset parameter that should be + // adjusted to reflect the number of bytes sent. Rewind it. offset -= notWritten } break @@ -414,18 +426,26 @@ func Sendfile(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sysc if offsetAddr != 0 { // Copy out the new offset. - if _, err := t.CopyOut(offsetAddr, offset); err != nil { + offsetP := primitive.Uint64(offset) + if _, err := offsetP.CopyOut(t, offsetAddr); err != nil { return 0, nil, err } } - if n == 0 { - return 0, nil, err + if n != 0 { + inFile.Dentry().InotifyWithParent(t, linux.IN_ACCESS, 0, vfs.PathEvent) + outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) + + if err != nil && err != io.EOF && err != syserror.ErrWouldBlock { + // If a partial write is completed, the error is dropped. Log it here. + log.Debugf("sendfile completed a partial write with error: %v", err) + err = nil + } } - inFile.Dentry().InotifyWithParent(t, linux.IN_ACCESS, 0, vfs.PathEvent) - outFile.Dentry().InotifyWithParent(t, linux.IN_MODIFY, 0, vfs.PathEvent) - return uintptr(n), nil, nil + // We can only pass a single file to handleIOError, so pick inFile arbitrarily. + // This is used only for debugging purposes. + return uintptr(n), nil, slinux.HandleIOErrorVFS2(t, n != 0, err, syserror.ERESTARTSYS, "sendfile", inFile) } // dualWaiter is used to wait on one or both vfs.FileDescriptions. It is not diff --git a/pkg/sentry/syscalls/linux/vfs2/timerfd.go b/pkg/sentry/syscalls/linux/vfs2/timerfd.go index 7a26890ef..250870c03 100644 --- a/pkg/sentry/syscalls/linux/vfs2/timerfd.go +++ b/pkg/sentry/syscalls/linux/vfs2/timerfd.go @@ -87,7 +87,7 @@ func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne } var newVal linux.Itimerspec - if _, err := t.CopyIn(newValAddr, &newVal); err != nil { + if _, err := newVal.CopyIn(t, newValAddr); err != nil { return 0, nil, err } newS, err := ktime.SettingFromItimerspec(newVal, flags&linux.TFD_TIMER_ABSTIME != 0, tfd.Clock()) @@ -97,7 +97,7 @@ func TimerfdSettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne tm, oldS := tfd.SetTime(newS) if oldValAddr != 0 { oldVal := ktime.ItimerspecFromSetting(tm, oldS) - if _, err := t.CopyOut(oldValAddr, &oldVal); err != nil { + if _, err := oldVal.CopyOut(t, oldValAddr); err != nil { return 0, nil, err } } @@ -122,6 +122,6 @@ func TimerfdGettime(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kerne tm, s := tfd.GetTime() curVal := ktime.ItimerspecFromSetting(tm, s) - _, err := t.CopyOut(curValAddr, &curVal) + _, err := curVal.CopyOut(t, curValAddr) return 0, nil, err } diff --git a/pkg/sentry/syscalls/linux/vfs2/vfs2.go b/pkg/sentry/syscalls/linux/vfs2/vfs2.go index c576d9475..c50fd97eb 100644 --- a/pkg/sentry/syscalls/linux/vfs2/vfs2.go +++ b/pkg/sentry/syscalls/linux/vfs2/vfs2.go @@ -93,16 +93,16 @@ func Override() { s.Table[165] = syscalls.Supported("mount", Mount) s.Table[166] = syscalls.Supported("umount2", Umount2) s.Table[187] = syscalls.Supported("readahead", Readahead) - s.Table[188] = syscalls.Supported("setxattr", Setxattr) + s.Table[188] = syscalls.Supported("setxattr", SetXattr) s.Table[189] = syscalls.Supported("lsetxattr", Lsetxattr) s.Table[190] = syscalls.Supported("fsetxattr", Fsetxattr) - s.Table[191] = syscalls.Supported("getxattr", Getxattr) + s.Table[191] = syscalls.Supported("getxattr", GetXattr) s.Table[192] = syscalls.Supported("lgetxattr", Lgetxattr) s.Table[193] = syscalls.Supported("fgetxattr", Fgetxattr) - s.Table[194] = syscalls.Supported("listxattr", Listxattr) + s.Table[194] = syscalls.Supported("listxattr", ListXattr) s.Table[195] = syscalls.Supported("llistxattr", Llistxattr) s.Table[196] = syscalls.Supported("flistxattr", Flistxattr) - s.Table[197] = syscalls.Supported("removexattr", Removexattr) + s.Table[197] = syscalls.Supported("removexattr", RemoveXattr) s.Table[198] = syscalls.Supported("lremovexattr", Lremovexattr) s.Table[199] = syscalls.Supported("fremovexattr", Fremovexattr) s.Table[209] = syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}) @@ -163,16 +163,17 @@ func Override() { // Override ARM64. s = linux.ARM64 - s.Table[5] = syscalls.Supported("setxattr", Setxattr) + s.Table[2] = syscalls.PartiallySupported("io_submit", IoSubmit, "Generally supported with exceptions. User ring optimizations are not implemented.", []string{"gvisor.dev/issue/204"}) + s.Table[5] = syscalls.Supported("setxattr", SetXattr) s.Table[6] = syscalls.Supported("lsetxattr", Lsetxattr) s.Table[7] = syscalls.Supported("fsetxattr", Fsetxattr) - s.Table[8] = syscalls.Supported("getxattr", Getxattr) + s.Table[8] = syscalls.Supported("getxattr", GetXattr) s.Table[9] = syscalls.Supported("lgetxattr", Lgetxattr) s.Table[10] = syscalls.Supported("fgetxattr", Fgetxattr) - s.Table[11] = syscalls.Supported("listxattr", Listxattr) + s.Table[11] = syscalls.Supported("listxattr", ListXattr) s.Table[12] = syscalls.Supported("llistxattr", Llistxattr) s.Table[13] = syscalls.Supported("flistxattr", Flistxattr) - s.Table[14] = syscalls.Supported("removexattr", Removexattr) + s.Table[14] = syscalls.Supported("removexattr", RemoveXattr) s.Table[15] = syscalls.Supported("lremovexattr", Lremovexattr) s.Table[16] = syscalls.Supported("fremovexattr", Fremovexattr) s.Table[17] = syscalls.Supported("getcwd", Getcwd) @@ -200,6 +201,7 @@ func Override() { s.Table[44] = syscalls.Supported("fstatfs", Fstatfs) s.Table[45] = syscalls.Supported("truncate", Truncate) s.Table[46] = syscalls.Supported("ftruncate", Ftruncate) + s.Table[47] = syscalls.PartiallySupported("fallocate", Fallocate, "Not all options are supported.", nil) s.Table[48] = syscalls.Supported("faccessat", Faccessat) s.Table[49] = syscalls.Supported("chdir", Chdir) s.Table[50] = syscalls.Supported("fchdir", Fchdir) @@ -221,12 +223,14 @@ func Override() { s.Table[68] = syscalls.Supported("pwrite64", Pwrite64) s.Table[69] = syscalls.Supported("preadv", Preadv) s.Table[70] = syscalls.Supported("pwritev", Pwritev) + s.Table[71] = syscalls.Supported("sendfile", Sendfile) s.Table[72] = syscalls.Supported("pselect", Pselect) s.Table[73] = syscalls.Supported("ppoll", Ppoll) s.Table[74] = syscalls.Supported("signalfd4", Signalfd4) s.Table[76] = syscalls.Supported("splice", Splice) s.Table[77] = syscalls.Supported("tee", Tee) s.Table[78] = syscalls.Supported("readlinkat", Readlinkat) + s.Table[79] = syscalls.Supported("newfstatat", Newfstatat) s.Table[80] = syscalls.Supported("fstat", Fstat) s.Table[81] = syscalls.Supported("sync", Sync) s.Table[82] = syscalls.Supported("fsync", Fsync) @@ -251,8 +255,10 @@ func Override() { s.Table[210] = syscalls.Supported("shutdown", Shutdown) s.Table[211] = syscalls.Supported("sendmsg", SendMsg) s.Table[212] = syscalls.Supported("recvmsg", RecvMsg) + s.Table[213] = syscalls.Supported("readahead", Readahead) s.Table[221] = syscalls.Supported("execve", Execve) s.Table[222] = syscalls.Supported("mmap", Mmap) + s.Table[223] = syscalls.PartiallySupported("fadvise64", Fadvise64, "Not all options are supported.", nil) s.Table[242] = syscalls.Supported("accept4", Accept4) s.Table[243] = syscalls.Supported("recvmmsg", RecvMMsg) s.Table[267] = syscalls.Supported("syncfs", Syncfs) diff --git a/pkg/sentry/syscalls/linux/vfs2/xattr.go b/pkg/sentry/syscalls/linux/vfs2/xattr.go index ef99246ed..e05723ef9 100644 --- a/pkg/sentry/syscalls/linux/vfs2/xattr.go +++ b/pkg/sentry/syscalls/linux/vfs2/xattr.go @@ -26,8 +26,8 @@ import ( "gvisor.dev/gvisor/pkg/usermem" ) -// Listxattr implements Linux syscall listxattr(2). -func Listxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { +// ListXattr implements Linux syscall listxattr(2). +func ListXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return listxattr(t, args, followFinalSymlink) } @@ -51,7 +51,7 @@ func listxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSyml } defer tpop.Release(t) - names, err := t.Kernel().VFS().ListxattrAt(t, t.Credentials(), &tpop.pop, uint64(size)) + names, err := t.Kernel().VFS().ListXattrAt(t, t.Credentials(), &tpop.pop, uint64(size)) if err != nil { return 0, nil, err } @@ -74,7 +74,7 @@ func Flistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy } defer file.DecRef(t) - names, err := file.Listxattr(t, uint64(size)) + names, err := file.ListXattr(t, uint64(size)) if err != nil { return 0, nil, err } @@ -85,8 +85,8 @@ func Flistxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sy return uintptr(n), nil, nil } -// Getxattr implements Linux syscall getxattr(2). -func Getxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { +// GetXattr implements Linux syscall getxattr(2). +func GetXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return getxattr(t, args, followFinalSymlink) } @@ -116,7 +116,7 @@ func getxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymli return 0, nil, err } - value, err := t.Kernel().VFS().GetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.GetxattrOptions{ + value, err := t.Kernel().VFS().GetXattrAt(t, t.Credentials(), &tpop.pop, &vfs.GetXattrOptions{ Name: name, Size: uint64(size), }) @@ -148,7 +148,7 @@ func Fgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, err } - value, err := file.Getxattr(t, &vfs.GetxattrOptions{Name: name, Size: uint64(size)}) + value, err := file.GetXattr(t, &vfs.GetXattrOptions{Name: name, Size: uint64(size)}) if err != nil { return 0, nil, err } @@ -159,8 +159,8 @@ func Fgetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return uintptr(n), nil, nil } -// Setxattr implements Linux syscall setxattr(2). -func Setxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { +// SetXattr implements Linux syscall setxattr(2). +func SetXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return 0, nil, setxattr(t, args, followFinalSymlink) } @@ -199,7 +199,7 @@ func setxattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSymli return err } - return t.Kernel().VFS().SetxattrAt(t, t.Credentials(), &tpop.pop, &vfs.SetxattrOptions{ + return t.Kernel().VFS().SetXattrAt(t, t.Credentials(), &tpop.pop, &vfs.SetXattrOptions{ Name: name, Value: value, Flags: uint32(flags), @@ -233,15 +233,15 @@ func Fsetxattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.Sys return 0, nil, err } - return 0, nil, file.Setxattr(t, &vfs.SetxattrOptions{ + return 0, nil, file.SetXattr(t, &vfs.SetXattrOptions{ Name: name, Value: value, Flags: uint32(flags), }) } -// Removexattr implements Linux syscall removexattr(2). -func Removexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { +// RemoveXattr implements Linux syscall removexattr(2). +func RemoveXattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel.SyscallControl, error) { return 0, nil, removexattr(t, args, followFinalSymlink) } @@ -269,7 +269,7 @@ func removexattr(t *kernel.Task, args arch.SyscallArguments, shouldFollowFinalSy return err } - return t.Kernel().VFS().RemovexattrAt(t, t.Credentials(), &tpop.pop, name) + return t.Kernel().VFS().RemoveXattrAt(t, t.Credentials(), &tpop.pop, name) } // Fremovexattr implements Linux syscall fremovexattr(2). @@ -288,7 +288,7 @@ func Fremovexattr(t *kernel.Task, args arch.SyscallArguments) (uintptr, *kernel. return 0, nil, err } - return 0, nil, file.Removexattr(t, name) + return 0, nil, file.RemoveXattr(t, name) } func copyInXattrName(t *kernel.Task, nameAddr usermem.Addr) (string, error) { diff --git a/pkg/sentry/usage/memory.go b/pkg/sentry/usage/memory.go index ab1d140d2..5ed6726ab 100644 --- a/pkg/sentry/usage/memory.go +++ b/pkg/sentry/usage/memory.go @@ -278,7 +278,7 @@ func TotalMemory(memSize, used uint64) uint64 { } if memSize < used { memSize = used - // Bump totalSize to the next largest power of 2, if one exists, so + // Bump memSize 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) diff --git a/pkg/sentry/vfs/BUILD b/pkg/sentry/vfs/BUILD index 642769e7c..440c9307c 100644 --- a/pkg/sentry/vfs/BUILD +++ b/pkg/sentry/vfs/BUILD @@ -27,6 +27,39 @@ go_template_instance( }, ) +go_template_instance( + name = "file_description_refs", + out = "file_description_refs.go", + package = "vfs", + prefix = "FileDescription", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "FileDescription", + }, +) + +go_template_instance( + name = "mount_namespace_refs", + out = "mount_namespace_refs.go", + package = "vfs", + prefix = "MountNamespace", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "MountNamespace", + }, +) + +go_template_instance( + name = "filesystem_refs", + out = "filesystem_refs.go", + package = "vfs", + prefix = "Filesystem", + template = "//pkg/refsvfs2:refs_template", + types = { + "T": "Filesystem", + }, +) + go_library( name = "vfs", srcs = [ @@ -40,29 +73,34 @@ go_library( "event_list.go", "file_description.go", "file_description_impl_util.go", + "file_description_refs.go", "filesystem.go", "filesystem_impl_util.go", + "filesystem_refs.go", "filesystem_type.go", "inotify.go", "lock.go", "mount.go", + "mount_namespace_refs.go", "mount_unsafe.go", "options.go", "pathname.go", "permissions.go", "resolving_path.go", + "save_restore.go", "vfs.go", ], visibility = ["//pkg/sentry:internal"], deps = [ "//pkg/abi/linux", - "//pkg/binary", "//pkg/context", "//pkg/fd", "//pkg/fdnotifier", "//pkg/fspath", "//pkg/gohacks", "//pkg/log", + "//pkg/refs", + "//pkg/refsvfs2", "//pkg/safemem", "//pkg/sentry/arch", "//pkg/sentry/fs", diff --git a/pkg/sentry/vfs/README.md b/pkg/sentry/vfs/README.md index 4b9faf2ea..5aad31b78 100644 --- a/pkg/sentry/vfs/README.md +++ b/pkg/sentry/vfs/README.md @@ -184,12 +184,3 @@ This construction, which is essentially a type-safe analogue to Linux's - File locking - `O_ASYNC` - -- Reference counts in the `vfs` package do not use the `refs` package since - `refs.AtomicRefCount` adds 64 bytes of overhead to each 8-byte reference - count, resulting in considerable cache bloat. 24 bytes of this overhead is - for weak reference support, which have poor performance and will not be used - by VFS2. The remaining 40 bytes is to store a descriptive string and stack - trace for reference leak checking; we can support reference leak checking - without incurring this space overhead by including the applicable - information directly in finalizers for applicable types. diff --git a/pkg/sentry/vfs/anonfs.go b/pkg/sentry/vfs/anonfs.go index 5a0e3e6b5..7ad0eaf86 100644 --- a/pkg/sentry/vfs/anonfs.go +++ b/pkg/sentry/vfs/anonfs.go @@ -52,6 +52,8 @@ const ( ) // anonFilesystemType implements FilesystemType. +// +// +stateify savable type anonFilesystemType struct{} // GetFilesystem implements FilesystemType.GetFilesystem. @@ -59,22 +61,28 @@ func (anonFilesystemType) GetFilesystem(context.Context, *VirtualFilesystem, *au panic("cannot instaniate an anon filesystem") } -// Name implemenents FilesystemType.Name. +// Name implements FilesystemType.Name. func (anonFilesystemType) Name() string { return "none" } +// Release implemenents FilesystemType.Release. +func (anonFilesystemType) Release(ctx context.Context) {} + // anonFilesystem is the implementation of FilesystemImpl that backs // VirtualDentries returned by VirtualFilesystem.NewAnonVirtualDentry(). // // Since all Dentries in anonFilesystem are non-directories, all FilesystemImpl // methods that would require an anonDentry to be a directory return ENOTDIR. +// +// +stateify savable type anonFilesystem struct { vfsfs Filesystem devMinor uint32 } +// +stateify savable type anonDentry struct { vfsd Dentry @@ -245,32 +253,32 @@ func (fs *anonFilesystem) BoundEndpointAt(ctx context.Context, rp *ResolvingPath return nil, syserror.ECONNREFUSED } -// ListxattrAt implements FilesystemImpl.ListxattrAt. -func (fs *anonFilesystem) ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) { +// ListXattrAt implements FilesystemImpl.ListXattrAt. +func (fs *anonFilesystem) ListXattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) { if !rp.Done() { return nil, syserror.ENOTDIR } return nil, nil } -// GetxattrAt implements FilesystemImpl.GetxattrAt. -func (fs *anonFilesystem) GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) { +// GetXattrAt implements FilesystemImpl.GetXattrAt. +func (fs *anonFilesystem) GetXattrAt(ctx context.Context, rp *ResolvingPath, opts GetXattrOptions) (string, error) { if !rp.Done() { return "", syserror.ENOTDIR } return "", syserror.ENOTSUP } -// SetxattrAt implements FilesystemImpl.SetxattrAt. -func (fs *anonFilesystem) SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error { +// SetXattrAt implements FilesystemImpl.SetXattrAt. +func (fs *anonFilesystem) SetXattrAt(ctx context.Context, rp *ResolvingPath, opts SetXattrOptions) error { if !rp.Done() { return syserror.ENOTDIR } return syserror.EPERM } -// RemovexattrAt implements FilesystemImpl.RemovexattrAt. -func (fs *anonFilesystem) RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error { +// RemoveXattrAt implements FilesystemImpl.RemoveXattrAt. +func (fs *anonFilesystem) RemoveXattrAt(ctx context.Context, rp *ResolvingPath, name string) error { if !rp.Done() { return syserror.ENOTDIR } diff --git a/pkg/sentry/vfs/context.go b/pkg/sentry/vfs/context.go index c9e724fef..97018651f 100644 --- a/pkg/sentry/vfs/context.go +++ b/pkg/sentry/vfs/context.go @@ -40,6 +40,30 @@ func MountNamespaceFromContext(ctx context.Context) *MountNamespace { return nil } +type mountNamespaceContext struct { + context.Context + mntns *MountNamespace +} + +// Value implements Context.Value. +func (mc mountNamespaceContext) Value(key interface{}) interface{} { + switch key { + case CtxMountNamespace: + mc.mntns.IncRef() + return mc.mntns + default: + return mc.Context.Value(key) + } +} + +// WithMountNamespace returns a copy of ctx with the given MountNamespace. +func WithMountNamespace(ctx context.Context, mntns *MountNamespace) context.Context { + return &mountNamespaceContext{ + Context: ctx, + mntns: mntns, + } +} + // RootFromContext returns the VFS root used by ctx. It takes a reference on // the returned VirtualDentry. If ctx does not have a specific VFS root, // RootFromContext returns a zero-value VirtualDentry. diff --git a/pkg/sentry/vfs/dentry.go b/pkg/sentry/vfs/dentry.go index bc7ea93ea..320ab7ce1 100644 --- a/pkg/sentry/vfs/dentry.go +++ b/pkg/sentry/vfs/dentry.go @@ -89,6 +89,8 @@ func (d *Dentry) Impl() DentryImpl { // DentryImpl contains implementation details for a Dentry. Implementations of // DentryImpl should contain their associated Dentry by value as their first // field. +// +// +stateify savable type DentryImpl interface { // IncRef increments the Dentry's reference count. A Dentry with a non-zero // reference count must remain coherent with the state of the filesystem. @@ -242,8 +244,9 @@ func (vfs *VirtualFilesystem) InvalidateDentry(ctx context.Context, d *Dentry) { // caller must call AbortRenameDentry, CommitRenameReplaceDentry, or // CommitRenameExchangeDentry depending on the rename's outcome. // -// Preconditions: If to is not nil, it must be a child Dentry from the same -// Filesystem. from != to. +// Preconditions: +// * If to is not nil, it must be a child Dentry from the same Filesystem. +// * from != to. func (vfs *VirtualFilesystem) PrepareRenameDentry(mntns *MountNamespace, from, to *Dentry) error { vfs.mountMu.Lock() if mntns.mountpoints[from] != 0 { diff --git a/pkg/sentry/vfs/device.go b/pkg/sentry/vfs/device.go index 1e9dffc8f..dde2ad79b 100644 --- a/pkg/sentry/vfs/device.go +++ b/pkg/sentry/vfs/device.go @@ -22,6 +22,8 @@ import ( ) // DeviceKind indicates whether a device is a block or character device. +// +// +stateify savable type DeviceKind uint32 const ( @@ -44,6 +46,7 @@ func (kind DeviceKind) String() string { } } +// +stateify savable type devTuple struct { kind DeviceKind major uint32 diff --git a/pkg/sentry/vfs/epoll.go b/pkg/sentry/vfs/epoll.go index 1b5af9f73..a98aac52b 100644 --- a/pkg/sentry/vfs/epoll.go +++ b/pkg/sentry/vfs/epoll.go @@ -27,6 +27,8 @@ import ( var epollCycleMu sync.Mutex // EpollInstance represents an epoll instance, as described by epoll(7). +// +// +stateify savable type EpollInstance struct { vfsfd FileDescription FileDescriptionDefaultImpl @@ -38,11 +40,11 @@ type EpollInstance struct { // interest is the set of file descriptors that are registered with the // EpollInstance for monitoring. interest is protected by interestMu. - interestMu sync.Mutex + interestMu sync.Mutex `state:"nosave"` interest map[epollInterestKey]*epollInterest // mu protects fields in registered epollInterests. - mu sync.Mutex + mu sync.Mutex `state:"nosave"` // ready is the set of file descriptors that may be "ready" for I/O. Note // that this must be an ordered list, not a map: "If more than maxevents @@ -55,6 +57,7 @@ type EpollInstance struct { ready epollInterestList } +// +stateify savable type epollInterestKey struct { // file is the registered FileDescription. No reference is held on file; // instead, when the last reference is dropped, FileDescription.DecRef() @@ -67,9 +70,11 @@ type epollInterestKey struct { } // epollInterest represents an EpollInstance's interest in a file descriptor. +// +// +stateify savable type epollInterest struct { // epoll is the owning EpollInstance. epoll is immutable. - epoll *EpollInstance + epoll *EpollInstance `state:"wait"` // key is the file to which this epollInterest applies. key is immutable. key epollInterestKey @@ -331,11 +336,9 @@ func (ep *EpollInstance) removeLocked(epi *epollInterest) { ep.mu.Unlock() } -// ReadEvents reads up to len(events) ready events into events and returns the -// number of events read. -// -// Preconditions: len(events) != 0. -func (ep *EpollInstance) ReadEvents(events []linux.EpollEvent) int { +// ReadEvents appends up to maxReady events to events and returns the updated +// slice of events. +func (ep *EpollInstance) ReadEvents(events []linux.EpollEvent, maxEvents int) []linux.EpollEvent { i := 0 // Hot path: avoid defer. ep.mu.Lock() @@ -368,16 +371,16 @@ func (ep *EpollInstance) ReadEvents(events []linux.EpollEvent) int { requeue.PushBack(epi) } // Report ievents. - events[i] = linux.EpollEvent{ + events = append(events, linux.EpollEvent{ Events: ievents.ToLinux(), Data: epi.userData, - } + }) i++ - if i == len(events) { + if i == maxEvents { break } } ep.ready.PushBackList(&requeue) ep.mu.Unlock() - return i + return events } diff --git a/pkg/sentry/vfs/file_description.go b/pkg/sentry/vfs/file_description.go index dcafffe57..546e445aa 100644 --- a/pkg/sentry/vfs/file_description.go +++ b/pkg/sentry/vfs/file_description.go @@ -37,13 +37,13 @@ import ( // FileDescription methods require that a reference is held. // // FileDescription is analogous to Linux's struct file. +// +// +stateify savable type FileDescription struct { - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 + FileDescriptionRefs // flagsMu protects statusFlags and asyncHandler below. - flagsMu sync.Mutex + flagsMu sync.Mutex `state:"nosave"` // statusFlags contains status flags, "initialized by open(2) and possibly // modified by fcntl()" - fcntl(2). statusFlags can be read using atomic @@ -58,7 +58,7 @@ type FileDescription struct { // epolls is the set of epollInterests registered for this FileDescription. // epolls is protected by epollMu. - epollMu sync.Mutex + epollMu sync.Mutex `state:"nosave"` epolls map[*epollInterest]struct{} // vd is the filesystem location at which this FileDescription was opened. @@ -90,6 +90,8 @@ type FileDescription struct { } // FileDescriptionOptions contains options to FileDescription.Init(). +// +// +stateify savable type FileDescriptionOptions struct { // If AllowDirectIO is true, allow O_DIRECT to be set on the file. AllowDirectIO bool @@ -103,7 +105,7 @@ type FileDescriptionOptions struct { // If UseDentryMetadata is true, calls to FileDescription methods that // interact with file and filesystem metadata (Stat, SetStat, StatFS, - // Listxattr, Getxattr, Setxattr, Removexattr) are implemented by calling + // ListXattr, GetXattr, SetXattr, RemoveXattr) are implemented by calling // the corresponding FilesystemImpl methods instead of the corresponding // FileDescriptionImpl methods. // @@ -131,7 +133,7 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou } } - fd.refs = 1 + fd.EnableLeakCheck() // Remove "file creation flags" to mirror the behavior from file.f_flags in // fs/open.c:do_dentry_open. @@ -149,30 +151,9 @@ func (fd *FileDescription) Init(impl FileDescriptionImpl, flags uint32, mnt *Mou return nil } -// IncRef increments fd's reference count. -func (fd *FileDescription) IncRef() { - atomic.AddInt64(&fd.refs, 1) -} - -// TryIncRef increments fd's reference count and returns true. If fd's -// reference count is already zero, TryIncRef does nothing and returns false. -// -// TryIncRef does not require that a reference is held on fd. -func (fd *FileDescription) TryIncRef() bool { - for { - refs := atomic.LoadInt64(&fd.refs) - if refs <= 0 { - return false - } - if atomic.CompareAndSwapInt64(&fd.refs, refs, refs+1) { - return true - } - } -} - // DecRef decrements fd's reference count. func (fd *FileDescription) DecRef(ctx context.Context) { - if refs := atomic.AddInt64(&fd.refs, -1); refs == 0 { + fd.FileDescriptionRefs.DecRef(func() { // Unregister fd from all epoll instances. fd.epollMu.Lock() epolls := fd.epolls @@ -202,21 +183,12 @@ func (fd *FileDescription) DecRef(ctx context.Context) { } fd.vd.DecRef(ctx) fd.flagsMu.Lock() - // TODO(gvisor.dev/issue/1663): We may need to unregister during save, as we do in VFS1. if fd.statusFlags&linux.O_ASYNC != 0 && fd.asyncHandler != nil { fd.asyncHandler.Unregister(fd) } fd.asyncHandler = nil fd.flagsMu.Unlock() - } else if refs < 0 { - panic("FileDescription.DecRef() called without holding a reference") - } -} - -// Refs returns the current number of references. The returned count -// is inherently racy and is unsafe to use without external synchronization. -func (fd *FileDescription) Refs() int64 { - return atomic.LoadInt64(&fd.refs) + }) } // Mount returns the mount on which fd was opened. It does not take a reference @@ -357,6 +329,9 @@ type FileDescriptionImpl interface { // Allocate grows the file to offset + length bytes. // Only mode == 0 is supported currently. // + // Allocate should return EISDIR on directories, ESPIPE on pipes, and ENODEV on + // other files where it is not supported. + // // Preconditions: The FileDescription was opened for writing. Allocate(ctx context.Context, mode, offset, length uint64) error @@ -371,8 +346,9 @@ type FileDescriptionImpl interface { // // - If opts.Flags specifies unsupported options, PRead returns EOPNOTSUPP. // - // Preconditions: The FileDescription was opened for reading. - // FileDescriptionOptions.DenyPRead == false. + // Preconditions: + // * The FileDescription was opened for reading. + // * FileDescriptionOptions.DenyPRead == false. PRead(ctx context.Context, dst usermem.IOSequence, offset int64, opts ReadOptions) (int64, error) // Read is similar to PRead, but does not specify an offset. @@ -403,8 +379,9 @@ type FileDescriptionImpl interface { // - If opts.Flags specifies unsupported options, PWrite returns // EOPNOTSUPP. // - // Preconditions: The FileDescription was opened for writing. - // FileDescriptionOptions.DenyPWrite == false. + // Preconditions: + // * The FileDescription was opened for writing. + // * FileDescriptionOptions.DenyPWrite == false. PWrite(ctx context.Context, src usermem.IOSequence, offset int64, opts WriteOptions) (int64, error) // Write is similar to PWrite, but does not specify an offset, which is @@ -449,19 +426,19 @@ type FileDescriptionImpl interface { // Ioctl implements the ioctl(2) syscall. Ioctl(ctx context.Context, uio usermem.IO, args arch.SyscallArguments) (uintptr, error) - // Listxattr returns all extended attribute names for the file. - Listxattr(ctx context.Context, size uint64) ([]string, error) + // ListXattr returns all extended attribute names for the file. + ListXattr(ctx context.Context, size uint64) ([]string, error) - // Getxattr returns the value associated with the given extended attribute + // GetXattr returns the value associated with the given extended attribute // for the file. - Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) + GetXattr(ctx context.Context, opts GetXattrOptions) (string, error) - // Setxattr changes the value associated with the given extended attribute + // SetXattr changes the value associated with the given extended attribute // for the file. - Setxattr(ctx context.Context, opts SetxattrOptions) error + SetXattr(ctx context.Context, opts SetXattrOptions) error - // Removexattr removes the given extended attribute from the file. - Removexattr(ctx context.Context, name string) error + // RemoveXattr removes the given extended attribute from the file. + RemoveXattr(ctx context.Context, name string) error // LockBSD tries to acquire a BSD-style advisory file lock. LockBSD(ctx context.Context, uid lock.UniqueID, t lock.LockType, block lock.Blocker) error @@ -477,6 +454,8 @@ type FileDescriptionImpl interface { } // Dirent holds the information contained in struct linux_dirent64. +// +// +stateify savable type Dirent struct { // Name is the filename. Name string @@ -664,25 +643,25 @@ func (fd *FileDescription) Ioctl(ctx context.Context, uio usermem.IO, args arch. return fd.impl.Ioctl(ctx, uio, args) } -// Listxattr returns all extended attribute names for the file represented by +// ListXattr returns all extended attribute names for the file represented by // fd. // // If the size of the list (including a NUL terminating byte after every entry) // would exceed size, ERANGE may be returned. Note that implementations // are free to ignore size entirely and return without error). In all cases, // if size is 0, the list should be returned without error, regardless of size. -func (fd *FileDescription) Listxattr(ctx context.Context, size uint64) ([]string, error) { +func (fd *FileDescription) ListXattr(ctx context.Context, size uint64) ([]string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - names, err := fd.vd.mount.fs.impl.ListxattrAt(ctx, rp, size) + names, err := fd.vd.mount.fs.impl.ListXattrAt(ctx, rp, size) vfsObj.putResolvingPath(ctx, rp) return names, err } - names, err := fd.impl.Listxattr(ctx, size) + names, err := fd.impl.ListXattr(ctx, size) if err == syserror.ENOTSUP { // Linux doesn't actually return ENOTSUP in this case; instead, // fs/xattr.c:vfs_listxattr() falls back to allowing the security @@ -693,57 +672,57 @@ func (fd *FileDescription) Listxattr(ctx context.Context, size uint64) ([]string return names, err } -// Getxattr returns the value associated with the given extended attribute for +// GetXattr returns the value associated with the given extended attribute for // the file represented by fd. // // If the size of the return value exceeds opts.Size, ERANGE may be returned // (note that implementations are free to ignore opts.Size entirely and return // without error). In all cases, if opts.Size is 0, the value should be // returned without error, regardless of size. -func (fd *FileDescription) Getxattr(ctx context.Context, opts *GetxattrOptions) (string, error) { +func (fd *FileDescription) GetXattr(ctx context.Context, opts *GetXattrOptions) (string, error) { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - val, err := fd.vd.mount.fs.impl.GetxattrAt(ctx, rp, *opts) + val, err := fd.vd.mount.fs.impl.GetXattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(ctx, rp) return val, err } - return fd.impl.Getxattr(ctx, *opts) + return fd.impl.GetXattr(ctx, *opts) } -// Setxattr changes the value associated with the given extended attribute for +// SetXattr changes the value associated with the given extended attribute for // the file represented by fd. -func (fd *FileDescription) Setxattr(ctx context.Context, opts *SetxattrOptions) error { +func (fd *FileDescription) SetXattr(ctx context.Context, opts *SetXattrOptions) error { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - err := fd.vd.mount.fs.impl.SetxattrAt(ctx, rp, *opts) + err := fd.vd.mount.fs.impl.SetXattrAt(ctx, rp, *opts) vfsObj.putResolvingPath(ctx, rp) return err } - return fd.impl.Setxattr(ctx, *opts) + return fd.impl.SetXattr(ctx, *opts) } -// Removexattr removes the given extended attribute from the file represented +// RemoveXattr removes the given extended attribute from the file represented // by fd. -func (fd *FileDescription) Removexattr(ctx context.Context, name string) error { +func (fd *FileDescription) RemoveXattr(ctx context.Context, name string) error { if fd.opts.UseDentryMetadata { vfsObj := fd.vd.mount.vfs rp := vfsObj.getResolvingPath(auth.CredentialsFromContext(ctx), &PathOperation{ Root: fd.vd, Start: fd.vd, }) - err := fd.vd.mount.fs.impl.RemovexattrAt(ctx, rp, name) + err := fd.vd.mount.fs.impl.RemoveXattrAt(ctx, rp, name) vfsObj.putResolvingPath(ctx, rp) return err } - return fd.impl.Removexattr(ctx, name) + return fd.impl.RemoveXattr(ctx, name) } // SyncFS instructs the filesystem containing fd to execute the semantics of @@ -845,3 +824,45 @@ func (fd *FileDescription) SetAsyncHandler(newHandler func() FileAsync) FileAsyn } return fd.asyncHandler } + +// FileReadWriteSeeker is a helper struct to pass a FileDescription as +// io.Reader/io.Writer/io.ReadSeeker/io.ReaderAt/io.WriterAt/etc. +type FileReadWriteSeeker struct { + FD *FileDescription + Ctx context.Context + ROpts ReadOptions + WOpts WriteOptions +} + +// ReadAt implements io.ReaderAt.ReadAt. +func (f *FileReadWriteSeeker) ReadAt(p []byte, off int64) (int, error) { + dst := usermem.BytesIOSequence(p) + n, err := f.FD.PRead(f.Ctx, dst, off, f.ROpts) + return int(n), err +} + +// Read implements io.ReadWriteSeeker.Read. +func (f *FileReadWriteSeeker) Read(p []byte) (int, error) { + dst := usermem.BytesIOSequence(p) + n, err := f.FD.Read(f.Ctx, dst, f.ROpts) + return int(n), err +} + +// Seek implements io.ReadWriteSeeker.Seek. +func (f *FileReadWriteSeeker) Seek(offset int64, whence int) (int64, error) { + return f.FD.Seek(f.Ctx, offset, int32(whence)) +} + +// WriteAt implements io.WriterAt.WriteAt. +func (f *FileReadWriteSeeker) WriteAt(p []byte, off int64) (int, error) { + dst := usermem.BytesIOSequence(p) + n, err := f.FD.PWrite(f.Ctx, dst, off, f.WOpts) + return int(n), err +} + +// Write implements io.ReadWriteSeeker.Write. +func (f *FileReadWriteSeeker) Write(p []byte) (int, error) { + buf := usermem.BytesIOSequence(p) + n, err := f.FD.Write(f.Ctx, buf, f.WOpts) + return int(n), err +} diff --git a/pkg/sentry/vfs/file_description_impl_util.go b/pkg/sentry/vfs/file_description_impl_util.go index 6b8b4ad49..48ca9de44 100644 --- a/pkg/sentry/vfs/file_description_impl_util.go +++ b/pkg/sentry/vfs/file_description_impl_util.go @@ -42,6 +42,8 @@ import ( // FileDescriptionDefaultImpl may be embedded by implementations of // FileDescriptionImpl to obtain implementations of many FileDescriptionImpl // methods with default behavior analogous to Linux's. +// +// +stateify savable type FileDescriptionDefaultImpl struct{} // OnClose implements FileDescriptionImpl.OnClose analogously to @@ -57,7 +59,11 @@ func (FileDescriptionDefaultImpl) StatFS(ctx context.Context) (linux.Statfs, err } // Allocate implements FileDescriptionImpl.Allocate analogously to -// fallocate called on regular file, directory or FIFO in Linux. +// fallocate called on an invalid type of file in Linux. +// +// Note that directories can rely on this implementation even though they +// should technically return EISDIR. Allocate should never be called for a +// directory, because it requires a writable fd. func (FileDescriptionDefaultImpl) Allocate(ctx context.Context, mode, offset, length uint64) error { return syserror.ENODEV } @@ -134,34 +140,36 @@ func (FileDescriptionDefaultImpl) Ioctl(ctx context.Context, uio usermem.IO, arg return 0, syserror.ENOTTY } -// Listxattr implements FileDescriptionImpl.Listxattr analogously to +// ListXattr implements FileDescriptionImpl.ListXattr analogously to // inode_operations::listxattr == NULL in Linux. -func (FileDescriptionDefaultImpl) Listxattr(ctx context.Context, size uint64) ([]string, error) { - // This isn't exactly accurate; see FileDescription.Listxattr. +func (FileDescriptionDefaultImpl) ListXattr(ctx context.Context, size uint64) ([]string, error) { + // This isn't exactly accurate; see FileDescription.ListXattr. return nil, syserror.ENOTSUP } -// Getxattr implements FileDescriptionImpl.Getxattr analogously to +// GetXattr implements FileDescriptionImpl.GetXattr analogously to // inode::i_opflags & IOP_XATTR == 0 in Linux. -func (FileDescriptionDefaultImpl) Getxattr(ctx context.Context, opts GetxattrOptions) (string, error) { +func (FileDescriptionDefaultImpl) GetXattr(ctx context.Context, opts GetXattrOptions) (string, error) { return "", syserror.ENOTSUP } -// Setxattr implements FileDescriptionImpl.Setxattr analogously to +// SetXattr implements FileDescriptionImpl.SetXattr analogously to // inode::i_opflags & IOP_XATTR == 0 in Linux. -func (FileDescriptionDefaultImpl) Setxattr(ctx context.Context, opts SetxattrOptions) error { +func (FileDescriptionDefaultImpl) SetXattr(ctx context.Context, opts SetXattrOptions) error { return syserror.ENOTSUP } -// Removexattr implements FileDescriptionImpl.Removexattr analogously to +// RemoveXattr implements FileDescriptionImpl.RemoveXattr analogously to // inode::i_opflags & IOP_XATTR == 0 in Linux. -func (FileDescriptionDefaultImpl) Removexattr(ctx context.Context, name string) error { +func (FileDescriptionDefaultImpl) RemoveXattr(ctx context.Context, name string) error { return syserror.ENOTSUP } // DirectoryFileDescriptionDefaultImpl may be embedded by implementations of // FileDescriptionImpl that always represent directories to obtain // implementations of non-directory I/O methods that return EISDIR. +// +// +stateify savable type DirectoryFileDescriptionDefaultImpl struct{} // Allocate implements DirectoryFileDescriptionDefaultImpl.Allocate. @@ -192,6 +200,8 @@ func (DirectoryFileDescriptionDefaultImpl) Write(ctx context.Context, src userme // DentryMetadataFileDescriptionImpl may be embedded by implementations of // FileDescriptionImpl for which FileDescriptionOptions.UseDentryMetadata is // true to obtain implementations of Stat and SetStat that panic. +// +// +stateify savable type DentryMetadataFileDescriptionImpl struct{} // Stat implements FileDescriptionImpl.Stat. @@ -206,12 +216,16 @@ func (DentryMetadataFileDescriptionImpl) SetStat(ctx context.Context, opts SetSt // DynamicBytesSource represents a data source for a // DynamicBytesFileDescriptionImpl. +// +// +stateify savable type DynamicBytesSource interface { // Generate writes the file's contents to buf. Generate(ctx context.Context, buf *bytes.Buffer) error } // StaticData implements DynamicBytesSource over a static string. +// +// +stateify savable type StaticData struct { Data string } @@ -238,14 +252,24 @@ type WritableDynamicBytesSource interface { // // DynamicBytesFileDescriptionImpl.SetDataSource() must be called before first // use. +// +// +stateify savable type DynamicBytesFileDescriptionImpl struct { data DynamicBytesSource // immutable - mu sync.Mutex // protects the following fields - buf bytes.Buffer + mu sync.Mutex `state:"nosave"` // protects the following fields + buf bytes.Buffer `state:".([]byte)"` off int64 lastRead int64 // offset at which the last Read, PRead, or Seek ended } +func (fd *DynamicBytesFileDescriptionImpl) saveBuf() []byte { + return fd.buf.Bytes() +} + +func (fd *DynamicBytesFileDescriptionImpl) loadBuf(p []byte) { + fd.buf.Write(p) +} + // SetDataSource must be called exactly once on fd before first use. func (fd *DynamicBytesFileDescriptionImpl) SetDataSource(data DynamicBytesSource) { fd.data = data @@ -378,6 +402,8 @@ func GenericConfigureMMap(fd *FileDescription, m memmap.Mappable, opts *memmap.M // LockFD may be used by most implementations of FileDescriptionImpl.Lock* // functions. Caller must call Init(). +// +// +stateify savable type LockFD struct { locks *FileLocks } @@ -405,6 +431,8 @@ func (fd *LockFD) UnlockBSD(ctx context.Context, uid fslock.UniqueID) error { // NoLockFD implements Lock*/Unlock* portion of FileDescriptionImpl interface // returning ENOLCK. +// +// +stateify savable type NoLockFD struct{} // LockBSD implements vfs.FileDescriptionImpl.LockBSD. diff --git a/pkg/sentry/vfs/filesystem.go b/pkg/sentry/vfs/filesystem.go index df3758fd1..c93d94634 100644 --- a/pkg/sentry/vfs/filesystem.go +++ b/pkg/sentry/vfs/filesystem.go @@ -15,8 +15,6 @@ package vfs import ( - "sync/atomic" - "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" "gvisor.dev/gvisor/pkg/fspath" @@ -34,9 +32,7 @@ import ( // // +stateify savable type Filesystem struct { - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 + FilesystemRefs // vfs is the VirtualFilesystem that uses this Filesystem. vfs is // immutable. @@ -52,7 +48,7 @@ type Filesystem struct { // Init must be called before first use of fs. func (fs *Filesystem) Init(vfsObj *VirtualFilesystem, fsType FilesystemType, impl FilesystemImpl) { - fs.refs = 1 + fs.EnableLeakCheck() fs.vfs = vfsObj fs.fsType = fsType fs.impl = impl @@ -76,39 +72,14 @@ func (fs *Filesystem) Impl() FilesystemImpl { return fs.impl } -// IncRef increments fs' reference count. -func (fs *Filesystem) IncRef() { - if atomic.AddInt64(&fs.refs, 1) <= 1 { - panic("Filesystem.IncRef() called without holding a reference") - } -} - -// TryIncRef increments fs' reference count and returns true. If fs' reference -// count is zero, TryIncRef does nothing and returns false. -// -// TryIncRef does not require that a reference is held on fs. -func (fs *Filesystem) TryIncRef() bool { - for { - refs := atomic.LoadInt64(&fs.refs) - if refs <= 0 { - return false - } - if atomic.CompareAndSwapInt64(&fs.refs, refs, refs+1) { - return true - } - } -} - // DecRef decrements fs' reference count. func (fs *Filesystem) DecRef(ctx context.Context) { - if refs := atomic.AddInt64(&fs.refs, -1); refs == 0 { + fs.FilesystemRefs.DecRef(func() { fs.vfs.filesystemsMu.Lock() delete(fs.vfs.filesystems, fs) fs.vfs.filesystemsMu.Unlock() fs.impl.Release(ctx) - } else if refs < 0 { - panic("Filesystem.decRef() called without holding a reference") - } + }) } // FilesystemImpl contains implementation details for a Filesystem. @@ -212,8 +183,9 @@ type FilesystemImpl interface { // ENOENT. Equivalently, if vd represents a file with a link count of 0 not // created by open(O_TMPFILE) without O_EXCL, LinkAt returns ENOENT. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If LinkAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -231,8 +203,9 @@ type FilesystemImpl interface { // - If the directory in which the new directory would be created has been // removed by RmdirAt or RenameAt, MkdirAt returns ENOENT. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If MkdirAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -253,8 +226,9 @@ type FilesystemImpl interface { // - If the directory in which the file would be created has been removed // by RmdirAt or RenameAt, MknodAt returns ENOENT. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If MknodAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -345,11 +319,12 @@ type FilesystemImpl interface { // - If renaming would replace a non-empty directory, RenameAt returns // ENOTEMPTY. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). oldParentVD.Dentry() was obtained from a - // previous call to - // oldParentVD.Mount().Filesystem().Impl().GetParentDentryAt(). oldName is - // not "." or "..". + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). + // * oldParentVD.Dentry() was obtained from a previous call to + // oldParentVD.Mount().Filesystem().Impl().GetParentDentryAt(). + // * oldName is not "." or "..". // // Postconditions: If RenameAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -372,8 +347,9 @@ type FilesystemImpl interface { // - If the file at rp exists but is not a directory, RmdirAt returns // ENOTDIR. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If RmdirAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -410,8 +386,9 @@ type FilesystemImpl interface { // - If the directory in which the symbolic link would be created has been // removed by RmdirAt or RenameAt, SymlinkAt returns ENOENT. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If SymlinkAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). @@ -431,33 +408,34 @@ type FilesystemImpl interface { // // - If the file at rp exists but is a directory, UnlinkAt returns EISDIR. // - // Preconditions: !rp.Done(). For the final path component in rp, - // !rp.ShouldFollowSymlink(). + // Preconditions: + // * !rp.Done(). + // * For the final path component in rp, !rp.ShouldFollowSymlink(). // // Postconditions: If UnlinkAt returns an error returned by // ResolvingPath.Resolve*(), then !rp.Done(). UnlinkAt(ctx context.Context, rp *ResolvingPath) error - // ListxattrAt returns all extended attribute names for the file at rp. + // ListXattrAt returns all extended attribute names for the file at rp. // // Errors: // // - If extended attributes are not supported by the filesystem, - // ListxattrAt returns ENOTSUP. + // ListXattrAt returns ENOTSUP. // // - If the size of the list (including a NUL terminating byte after every // entry) would exceed size, ERANGE may be returned. Note that // implementations are free to ignore size entirely and return without // error). In all cases, if size is 0, the list should be returned without // error, regardless of size. - ListxattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) + ListXattrAt(ctx context.Context, rp *ResolvingPath, size uint64) ([]string, error) - // GetxattrAt returns the value associated with the given extended + // GetXattrAt returns the value associated with the given extended // attribute for the file at rp. // // Errors: // - // - If extended attributes are not supported by the filesystem, GetxattrAt + // - If extended attributes are not supported by the filesystem, GetXattrAt // returns ENOTSUP. // // - If an extended attribute named opts.Name does not exist, ENODATA is @@ -467,30 +445,30 @@ type FilesystemImpl interface { // returned (note that implementations are free to ignore opts.Size entirely // and return without error). In all cases, if opts.Size is 0, the value // should be returned without error, regardless of size. - GetxattrAt(ctx context.Context, rp *ResolvingPath, opts GetxattrOptions) (string, error) + GetXattrAt(ctx context.Context, rp *ResolvingPath, opts GetXattrOptions) (string, error) - // SetxattrAt changes the value associated with the given extended + // SetXattrAt changes the value associated with the given extended // attribute for the file at rp. // // Errors: // - // - If extended attributes are not supported by the filesystem, SetxattrAt + // - If extended attributes are not supported by the filesystem, SetXattrAt // returns ENOTSUP. // // - If XATTR_CREATE is set in opts.Flag and opts.Name already exists, // EEXIST is returned. If XATTR_REPLACE is set and opts.Name does not exist, // ENODATA is returned. - SetxattrAt(ctx context.Context, rp *ResolvingPath, opts SetxattrOptions) error + SetXattrAt(ctx context.Context, rp *ResolvingPath, opts SetXattrOptions) error - // RemovexattrAt removes the given extended attribute from the file at rp. + // RemoveXattrAt removes the given extended attribute from the file at rp. // // Errors: // // - If extended attributes are not supported by the filesystem, - // RemovexattrAt returns ENOTSUP. + // RemoveXattrAt returns ENOTSUP. // // - If name does not exist, ENODATA is returned. - RemovexattrAt(ctx context.Context, rp *ResolvingPath, name string) error + RemoveXattrAt(ctx context.Context, rp *ResolvingPath, name string) error // BoundEndpointAt returns the Unix socket endpoint bound at the path rp. // @@ -528,6 +506,8 @@ type FilesystemImpl interface { // PrependPathAtVFSRootError is returned by implementations of // FilesystemImpl.PrependPath() when they encounter the contextual VFS root. +// +// +stateify savable type PrependPathAtVFSRootError struct{} // Error implements error.Error. @@ -538,6 +518,8 @@ func (PrependPathAtVFSRootError) Error() string { // PrependPathAtNonMountRootError is returned by implementations of // FilesystemImpl.PrependPath() when they encounter an independent ancestor // Dentry that is not the Mount root. +// +// +stateify savable type PrependPathAtNonMountRootError struct{} // Error implements error.Error. @@ -548,6 +530,8 @@ func (PrependPathAtNonMountRootError) Error() string { // PrependPathSyntheticError is returned by implementations of // FilesystemImpl.PrependPath() for which prepended names do not represent real // paths. +// +// +stateify savable type PrependPathSyntheticError struct{} // Error implements error.Error. diff --git a/pkg/sentry/vfs/filesystem_impl_util.go b/pkg/sentry/vfs/filesystem_impl_util.go index 465e610e0..2620cf975 100644 --- a/pkg/sentry/vfs/filesystem_impl_util.go +++ b/pkg/sentry/vfs/filesystem_impl_util.go @@ -16,6 +16,9 @@ package vfs import ( "strings" + + "gvisor.dev/gvisor/pkg/abi/linux" + "gvisor.dev/gvisor/pkg/usermem" ) // GenericParseMountOptions parses a comma-separated list of options of the @@ -41,3 +44,13 @@ func GenericParseMountOptions(str string) map[string]string { } return m } + +// GenericStatFS returns a statfs struct filled with the common fields for a +// general filesystem. This is analogous to Linux's fs/libfs.cs:simple_statfs(). +func GenericStatFS(fsMagic uint64) linux.Statfs { + return linux.Statfs{ + Type: fsMagic, + BlockSize: usermem.PageSize, + NameLength: linux.NAME_MAX, + } +} diff --git a/pkg/sentry/vfs/filesystem_type.go b/pkg/sentry/vfs/filesystem_type.go index f2298f7f6..9d54cc4ed 100644 --- a/pkg/sentry/vfs/filesystem_type.go +++ b/pkg/sentry/vfs/filesystem_type.go @@ -33,6 +33,9 @@ type FilesystemType interface { // Name returns the name of this FilesystemType. Name() string + + // Release releases all resources held by this FilesystemType. + Release(ctx context.Context) } // GetFilesystemOptions contains options to FilesystemType.GetFilesystem. @@ -55,10 +58,13 @@ type registeredFilesystemType struct { // RegisterFilesystemTypeOptions contains options to // VirtualFilesystem.RegisterFilesystem(). +// +// +stateify savable type RegisterFilesystemTypeOptions struct { - // If AllowUserMount is true, allow calls to VirtualFilesystem.MountAt() - // for which MountOptions.InternalMount == false to use this filesystem - // type. + // AllowUserMount determines whether users are allowed to mount a file system + // of this type, i.e. through mount(2). If AllowUserMount is true, allow calls + // to VirtualFilesystem.MountAt() for which MountOptions.InternalMount == false + // to use this filesystem type. AllowUserMount bool // If AllowUserList is true, make this filesystem type visible in diff --git a/pkg/sentry/vfs/g3doc/inotify.md b/pkg/sentry/vfs/g3doc/inotify.md index e7da49faa..833db213f 100644 --- a/pkg/sentry/vfs/g3doc/inotify.md +++ b/pkg/sentry/vfs/g3doc/inotify.md @@ -28,9 +28,9 @@ The set of all watches held on a single file (i.e., the watch target) is stored in vfs.Watches. Each watch will belong to a different inotify instance (an instance can only have one watch on any watch target). The watches are stored in a map indexed by their vfs.Inotify owner’s id. Hard links and file descriptions -to a single file will all share the same vfs.Watches. Activity on the target -causes its vfs.Watches to generate notifications on its watches’ inotify -instances. +to a single file will all share the same vfs.Watches (with the exception of the +gofer filesystem, described in a later section). Activity on the target causes +its vfs.Watches to generate notifications on its watches’ inotify instances. ### vfs.Watch @@ -103,12 +103,12 @@ inotify: unopened p9 file (and possibly an open FID), through which the Sentry interacts with the gofer. * *Solution:* Because there is no inode structure stored in the sandbox, - inotify watches must be held on the dentry. This would be an issue in - the presence of hard links, where multiple dentries would need to share - the same set of watches, but in VFS2, we do not support the internal - creation of hard links on gofer fs. As a result, we make the assumption - that every dentry corresponds to a unique inode. However, the next point - raises an issue with this assumption: + inotify watches must be held on the dentry. For the purposes of inotify, + we assume that every dentry corresponds to a unique inode, which may + cause unexpected behavior in the presence of hard links, where multiple + dentries should share the same set of watches. Indeed, it is impossible + for us to be absolutely sure whether dentries correspond to the same + file or not, due to the following point: * **The Sentry cannot always be aware of hard links on the remote filesystem.** There is no way for us to confirm whether two files on the remote filesystem are actually links to the same inode. QIDs and inodes are diff --git a/pkg/sentry/vfs/genericfstree/genericfstree.go b/pkg/sentry/vfs/genericfstree/genericfstree.go index 8882fa84a..ba6e6ed49 100644 --- a/pkg/sentry/vfs/genericfstree/genericfstree.go +++ b/pkg/sentry/vfs/genericfstree/genericfstree.go @@ -27,6 +27,8 @@ import ( ) // Dentry is a required type parameter that is a struct with the given fields. +// +// +stateify savable type Dentry struct { // vfsd is the embedded vfs.Dentry corresponding to this vfs.DentryImpl. vfsd vfs.Dentry @@ -69,7 +71,7 @@ func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath if mnt == vfsroot.Mount() && &d.vfsd == vfsroot.Dentry() { return vfs.PrependPathAtVFSRootError{} } - if &d.vfsd == mnt.Root() { + if mnt != nil && &d.vfsd == mnt.Root() { return nil } if d.parent == nil { @@ -79,3 +81,12 @@ func PrependPath(vfsroot vfs.VirtualDentry, mnt *vfs.Mount, d *Dentry, b *fspath d = d.parent } } + +// DebugPathname returns a pathname to d relative to its filesystem root. +// DebugPathname does not correspond to any Linux function; it's used to +// generate dentry pathnames for debugging. +func DebugPathname(d *Dentry) string { + var b fspath.Builder + _ = PrependPath(vfs.VirtualDentry{}, nil, d, &b) + return b.String() +} diff --git a/pkg/sentry/vfs/inotify.go b/pkg/sentry/vfs/inotify.go index aff220a61..3f0b8f45b 100644 --- a/pkg/sentry/vfs/inotify.go +++ b/pkg/sentry/vfs/inotify.go @@ -37,6 +37,8 @@ const inotifyEventBaseSize = 16 // // The way events are labelled appears somewhat arbitrary, but they must match // Linux so that IN_EXCL_UNLINK behaves as it does in Linux. +// +// +stateify savable type EventType uint8 // PathEvent and InodeEvent correspond to FSNOTIFY_EVENT_PATH and diff --git a/pkg/sentry/vfs/lock.go b/pkg/sentry/vfs/lock.go index 42666eebf..1ff202f2a 100644 --- a/pkg/sentry/vfs/lock.go +++ b/pkg/sentry/vfs/lock.go @@ -12,11 +12,6 @@ // See the License for the specific language governing permissions and // limitations under the License. -// Package lock provides POSIX and BSD style file locking for VFS2 file -// implementations. -// -// The actual implementations can be found in the lock package under -// sentry/fs/lock. package vfs import ( @@ -33,6 +28,8 @@ import ( // Note that in Linux these two types of locks are _not_ cooperative, because // race and deadlock conditions make merging them prohibitive. We do the same // and keep them oblivious to each other. +// +// +stateify savable type FileLocks struct { // bsd is a set of BSD-style advisory file wide locks, see flock(2). bsd fslock.Locks diff --git a/pkg/sentry/vfs/memxattr/xattr.go b/pkg/sentry/vfs/memxattr/xattr.go index cc1e7d764..638b5d830 100644 --- a/pkg/sentry/vfs/memxattr/xattr.go +++ b/pkg/sentry/vfs/memxattr/xattr.go @@ -33,8 +33,8 @@ type SimpleExtendedAttributes struct { xattrs map[string]string } -// Getxattr returns the value at 'name'. -func (x *SimpleExtendedAttributes) Getxattr(opts *vfs.GetxattrOptions) (string, error) { +// GetXattr returns the value at 'name'. +func (x *SimpleExtendedAttributes) GetXattr(opts *vfs.GetXattrOptions) (string, error) { x.mu.RLock() value, ok := x.xattrs[opts.Name] x.mu.RUnlock() @@ -49,8 +49,8 @@ func (x *SimpleExtendedAttributes) Getxattr(opts *vfs.GetxattrOptions) (string, return value, nil } -// Setxattr sets 'value' at 'name'. -func (x *SimpleExtendedAttributes) Setxattr(opts *vfs.SetxattrOptions) error { +// SetXattr sets 'value' at 'name'. +func (x *SimpleExtendedAttributes) SetXattr(opts *vfs.SetXattrOptions) error { x.mu.Lock() defer x.mu.Unlock() if x.xattrs == nil { @@ -72,8 +72,8 @@ func (x *SimpleExtendedAttributes) Setxattr(opts *vfs.SetxattrOptions) error { return nil } -// Listxattr returns all names in xattrs. -func (x *SimpleExtendedAttributes) Listxattr(size uint64) ([]string, error) { +// ListXattr returns all names in xattrs. +func (x *SimpleExtendedAttributes) ListXattr(size uint64) ([]string, error) { // Keep track of the size of the buffer needed in listxattr(2) for the list. listSize := 0 x.mu.RLock() @@ -90,8 +90,8 @@ func (x *SimpleExtendedAttributes) Listxattr(size uint64) ([]string, error) { return names, nil } -// Removexattr removes the xattr at 'name'. -func (x *SimpleExtendedAttributes) Removexattr(name string) error { +// RemoveXattr removes the xattr at 'name'. +func (x *SimpleExtendedAttributes) RemoveXattr(name string) error { x.mu.Lock() defer x.mu.Unlock() if _, ok := x.xattrs[name]; !ok { diff --git a/pkg/sentry/vfs/mount.go b/pkg/sentry/vfs/mount.go index 67dfba986..d452d2cda 100644 --- a/pkg/sentry/vfs/mount.go +++ b/pkg/sentry/vfs/mount.go @@ -24,6 +24,7 @@ import ( "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/refsvfs2" "gvisor.dev/gvisor/pkg/sentry/kernel/auth" "gvisor.dev/gvisor/pkg/syserror" ) @@ -46,8 +47,9 @@ import ( // +stateify savable type Mount struct { // vfs, fs, root are immutable. References are held on fs and root. + // Note that for a disconnected mount, root may be nil. // - // Invariant: root belongs to fs. + // Invariant: if not nil, root belongs to fs. vfs *VirtualFilesystem fs *Filesystem root *Dentry @@ -65,7 +67,7 @@ type Mount struct { // // Invariant: key.parent != nil iff key.point != nil. key.point belongs to // key.parent.fs. - key mountKey + key mountKey `state:".(VirtualDentry)"` // ns is the namespace in which this Mount was mounted. ns is protected by // VirtualFilesystem.mountMu. @@ -105,6 +107,9 @@ func newMount(vfs *VirtualFilesystem, fs *Filesystem, root *Dentry, mntns *Mount if opts.ReadOnly { mnt.setReadOnlyLocked(true) } + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Register(mnt, "vfs.Mount") + } return mnt } @@ -126,16 +131,14 @@ func (mnt *Mount) Options() MountOptions { // // +stateify savable type MountNamespace struct { + MountNamespaceRefs + // Owner is the usernamespace that owns this mount namespace. Owner *auth.UserNamespace // root is the MountNamespace's root mount. root is immutable. root *Mount - // refs is the reference count. refs is accessed using atomic memory - // operations. - refs int64 - // mountpoints maps all Dentries which are mount points in this namespace // to the number of Mounts for which they are mount points. mountpoints is // protected by VirtualFilesystem.mountMu. @@ -154,22 +157,22 @@ type MountNamespace struct { // NewMountNamespace returns a new mount namespace with a root filesystem // configured by the given arguments. A reference is taken on the returned // MountNamespace. -func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *GetFilesystemOptions) (*MountNamespace, error) { +func (vfs *VirtualFilesystem) NewMountNamespace(ctx context.Context, creds *auth.Credentials, source, fsTypeName string, opts *MountOptions) (*MountNamespace, error) { rft := vfs.getFilesystemType(fsTypeName) if rft == nil { ctx.Warningf("Unknown filesystem type: %s", fsTypeName) return nil, syserror.ENODEV } - fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, *opts) + fs, root, err := rft.fsType.GetFilesystem(ctx, vfs, creds, source, opts.GetFilesystemOptions) if err != nil { return nil, err } mntns := &MountNamespace{ Owner: creds.UserNamespace, - refs: 1, mountpoints: make(map[*Dentry]uint32), } - mntns.root = newMount(vfs, fs, root, mntns, &MountOptions{}) + mntns.EnableLeakCheck() + mntns.root = newMount(vfs, fs, root, mntns, opts) return mntns, nil } @@ -263,16 +266,20 @@ func (vfs *VirtualFilesystem) ConnectMountAt(ctx context.Context, creds *auth.Cr } // MountAt creates and mounts a Filesystem configured by the given arguments. -func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) error { +// The VirtualFilesystem will hold a reference to the Mount until it is unmounted. +// +// This method returns the mounted Mount without a reference, for convenience +// during VFS setup when there is no chance of racing with unmount. +func (vfs *VirtualFilesystem) MountAt(ctx context.Context, creds *auth.Credentials, source string, target *PathOperation, fsTypeName string, opts *MountOptions) (*Mount, error) { mnt, err := vfs.MountDisconnected(ctx, creds, source, fsTypeName, opts) if err != nil { - return err + return nil, err } defer mnt.DecRef(ctx) if err := vfs.ConnectMountAt(ctx, creds, mnt, target); err != nil { - return err + return nil, err } - return nil + return mnt, nil } // UmountAt removes the Mount at the given path. @@ -343,6 +350,7 @@ func (vfs *VirtualFilesystem) UmountAt(ctx context.Context, creds *auth.Credenti return nil } +// +stateify savable type umountRecursiveOptions struct { // If eager is true, ensure that future calls to Mount.tryIncMountedRef() // on umounted mounts fail. @@ -369,8 +377,9 @@ type umountRecursiveOptions struct { // // umountRecursiveLocked is analogous to Linux's fs/namespace.c:umount_tree(). // -// Preconditions: vfs.mountMu must be locked. vfs.mounts.seq must be in a -// writer critical section. +// Preconditions: +// * vfs.mountMu must be locked. +// * vfs.mounts.seq must be in a writer critical section. func (vfs *VirtualFilesystem) umountRecursiveLocked(mnt *Mount, opts *umountRecursiveOptions, vdsToDecRef []VirtualDentry, mountsToDecRef []*Mount) ([]VirtualDentry, []*Mount) { if !mnt.umounted { mnt.umounted = true @@ -399,9 +408,11 @@ func (vfs *VirtualFilesystem) umountRecursiveLocked(mnt *Mount, opts *umountRecu // connectLocked makes vd the mount parent/point for mnt. It consumes // references held by vd. // -// Preconditions: vfs.mountMu must be locked. vfs.mounts.seq must be in a -// writer critical section. d.mu must be locked. mnt.parent() == nil, i.e. mnt -// must not already be connected. +// Preconditions: +// * vfs.mountMu must be locked. +// * vfs.mounts.seq must be in a writer critical section. +// * d.mu must be locked. +// * mnt.parent() == nil, i.e. mnt must not already be connected. func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns *MountNamespace) { if checkInvariants { if mnt.parent() != nil { @@ -409,7 +420,7 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns } } mnt.IncRef() // dropped by callers of umountRecursiveLocked - mnt.storeKey(vd) + mnt.setKey(vd) if vd.mount.children == nil { vd.mount.children = make(map[*Mount]struct{}) } @@ -429,16 +440,18 @@ func (vfs *VirtualFilesystem) connectLocked(mnt *Mount, vd VirtualDentry, mntns // disconnectLocked makes vd have no mount parent/point and returns its old // mount parent/point with a reference held. // -// Preconditions: vfs.mountMu must be locked. vfs.mounts.seq must be in a -// writer critical section. mnt.parent() != nil. +// Preconditions: +// * vfs.mountMu must be locked. +// * vfs.mounts.seq must be in a writer critical section. +// * mnt.parent() != nil. func (vfs *VirtualFilesystem) disconnectLocked(mnt *Mount) VirtualDentry { - vd := mnt.loadKey() + vd := mnt.getKey() if checkInvariants { if vd.mount != nil { panic("VFS.disconnectLocked called on disconnected mount") } } - mnt.storeKey(VirtualDentry{}) + mnt.loadKey(VirtualDentry{}) delete(vd.mount.children, mnt) atomic.AddUint32(&vd.dentry.mounts, math.MaxUint32) // -1 mnt.ns.mountpoints[vd.dentry]-- @@ -480,35 +493,42 @@ func (mnt *Mount) IncRef() { // DecRef decrements mnt's reference count. func (mnt *Mount) DecRef(ctx context.Context) { - refs := atomic.AddInt64(&mnt.refs, -1) - if refs&^math.MinInt64 == 0 { // mask out MSB - var vd VirtualDentry - if mnt.parent() != nil { - mnt.vfs.mountMu.Lock() - mnt.vfs.mounts.seq.BeginWrite() - vd = mnt.vfs.disconnectLocked(mnt) - mnt.vfs.mounts.seq.EndWrite() - mnt.vfs.mountMu.Unlock() - } - mnt.root.DecRef(ctx) - mnt.fs.DecRef(ctx) - if vd.Ok() { - vd.DecRef(ctx) + r := atomic.AddInt64(&mnt.refs, -1) + if r&^math.MinInt64 == 0 { // mask out MSB + if refsvfs2.LeakCheckEnabled() { + refsvfs2.Unregister(mnt, "vfs.Mount") } + mnt.destroy(ctx) } } -// IncRef increments mntns' reference count. -func (mntns *MountNamespace) IncRef() { - if atomic.AddInt64(&mntns.refs, 1) <= 1 { - panic("MountNamespace.IncRef() called without holding a reference") +func (mnt *Mount) destroy(ctx context.Context) { + var vd VirtualDentry + if mnt.parent() != nil { + mnt.vfs.mountMu.Lock() + mnt.vfs.mounts.seq.BeginWrite() + vd = mnt.vfs.disconnectLocked(mnt) + mnt.vfs.mounts.seq.EndWrite() + mnt.vfs.mountMu.Unlock() + } + if mnt.root != nil { + mnt.root.DecRef(ctx) } + mnt.fs.DecRef(ctx) + if vd.Ok() { + vd.DecRef(ctx) + } +} + +// LeakMessage implements refsvfs2.CheckedObject.LeakMessage. +func (mnt *Mount) LeakMessage() string { + return fmt.Sprintf("[vfs.Mount %p] reference count of %d instead of 0", mnt, atomic.LoadInt64(&mnt.refs)) } // DecRef decrements mntns' reference count. func (mntns *MountNamespace) DecRef(ctx context.Context) { vfs := mntns.root.fs.VirtualFilesystem() - if refs := atomic.AddInt64(&mntns.refs, -1); refs == 0 { + mntns.MountNamespaceRefs.DecRef(func() { vfs.mountMu.Lock() vfs.mounts.seq.BeginWrite() vdsToDecRef, mountsToDecRef := vfs.umountRecursiveLocked(mntns.root, &umountRecursiveOptions{ @@ -522,9 +542,7 @@ func (mntns *MountNamespace) DecRef(ctx context.Context) { for _, mnt := range mountsToDecRef { mnt.DecRef(ctx) } - } else if refs < 0 { - panic("MountNamespace.DecRef() called without holding a reference") - } + }) } // getMountAt returns the last Mount in the stack mounted at (mnt, d). It takes @@ -576,8 +594,9 @@ retryFirst: // mnt. It takes a reference on the returned VirtualDentry. If no such mount // point exists (i.e. mnt is a root mount), getMountpointAt returns (nil, nil). // -// Preconditions: References are held on mnt and root. vfsroot is not (mnt, -// mnt.root). +// Preconditions: +// * References are held on mnt and root. +// * vfsroot is not (mnt, mnt.root). func (vfs *VirtualFilesystem) getMountpointAt(ctx context.Context, mnt *Mount, vfsroot VirtualDentry) VirtualDentry { // The first mount is special-cased: // @@ -651,6 +670,13 @@ retryFirst: return VirtualDentry{mnt, d} } +// SetMountReadOnly sets the mount as ReadOnly. +func (vfs *VirtualFilesystem) SetMountReadOnly(mnt *Mount, ro bool) error { + vfs.mountMu.Lock() + defer vfs.mountMu.Unlock() + return mnt.setReadOnlyLocked(ro) +} + // CheckBeginWrite increments the counter of in-progress write operations on // mnt. If mnt is mounted MS_RDONLY, CheckBeginWrite does nothing and returns // EROFS. @@ -717,14 +743,12 @@ func (mnt *Mount) Root() *Dentry { return mnt.root } -// Root returns mntns' root. A reference is taken on the returned -// VirtualDentry. +// Root returns mntns' root. It does not take a reference on the returned Dentry. func (mntns *MountNamespace) Root() VirtualDentry { vd := VirtualDentry{ mount: mntns.root, dentry: mntns.root.root, } - vd.IncRef() return vd } @@ -732,11 +756,23 @@ func (mntns *MountNamespace) Root() VirtualDentry { // // Preconditions: taskRootDir.Ok(). func (vfs *VirtualFilesystem) GenerateProcMounts(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { - vfs.mountMu.Lock() - defer vfs.mountMu.Unlock() rootMnt := taskRootDir.mount + + vfs.mountMu.Lock() mounts := rootMnt.submountsLocked() + // Take a reference on mounts since we need to drop vfs.mountMu before + // calling vfs.PathnameReachable() (=> FilesystemImpl.PrependPath()). + for _, mnt := range mounts { + mnt.IncRef() + } + vfs.mountMu.Unlock() + defer func() { + for _, mnt := range mounts { + mnt.DecRef(ctx) + } + }() sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + for _, mnt := range mounts { // Get the path to this mount relative to task root. mntRootVD := VirtualDentry{ @@ -747,7 +783,7 @@ func (vfs *VirtualFilesystem) GenerateProcMounts(ctx context.Context, taskRootDi if err != nil { // For some reason we didn't get a path. Log a warning // and run with empty path. - ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + ctx.Warningf("VFS.GenerateProcMounts: error getting pathname for mount root %+v: %v", mnt.root, err) path = "" } if path == "" { @@ -781,11 +817,25 @@ func (vfs *VirtualFilesystem) GenerateProcMounts(ctx context.Context, taskRootDi // // Preconditions: taskRootDir.Ok(). func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRootDir VirtualDentry, buf *bytes.Buffer) { - vfs.mountMu.Lock() - defer vfs.mountMu.Unlock() rootMnt := taskRootDir.mount + + vfs.mountMu.Lock() mounts := rootMnt.submountsLocked() + // Take a reference on mounts since we need to drop vfs.mountMu before + // calling vfs.PathnameReachable() (=> FilesystemImpl.PrependPath()) or + // vfs.StatAt() (=> FilesystemImpl.StatAt()). + for _, mnt := range mounts { + mnt.IncRef() + } + vfs.mountMu.Unlock() + defer func() { + for _, mnt := range mounts { + mnt.DecRef(ctx) + } + }() sort.Slice(mounts, func(i, j int) bool { return mounts[i].ID < mounts[j].ID }) + + creds := auth.CredentialsFromContext(ctx) for _, mnt := range mounts { // Get the path to this mount relative to task root. mntRootVD := VirtualDentry{ @@ -796,7 +846,7 @@ func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRoo if err != nil { // For some reason we didn't get a path. Log a warning // and run with empty path. - ctx.Warningf("Error getting pathname for mount root %+v: %v", mnt.root, err) + ctx.Warningf("VFS.GenerateProcMountInfo: error getting pathname for mount root %+v: %v", mnt.root, err) path = "" } if path == "" { @@ -809,9 +859,10 @@ func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRoo Root: mntRootVD, Start: mntRootVD, } - statx, err := vfs.StatAt(ctx, auth.NewAnonymousCredentials(), pop, &StatOptions{}) + statx, err := vfs.StatAt(ctx, creds, pop, &StatOptions{}) if err != nil { // Well that's not good. Ignore this mount. + ctx.Warningf("VFS.GenerateProcMountInfo: failed to stat mount root %+v: %v", mnt.root, err) break } @@ -823,6 +874,9 @@ func (vfs *VirtualFilesystem) GenerateProcMountInfo(ctx context.Context, taskRoo fmt.Fprintf(buf, "%d ", mnt.ID) // (2) Parent ID (or this ID if there is no parent). + // Note that even if the call to mnt.parent() races with Mount + // destruction (which is possible since we're not holding vfs.mountMu), + // its Mount.ID will still be valid. pID := mnt.ID if p := mnt.parent(); p != nil { pID = p.ID diff --git a/pkg/sentry/vfs/mount_test.go b/pkg/sentry/vfs/mount_test.go index 3335e4057..cb8c56bd3 100644 --- a/pkg/sentry/vfs/mount_test.go +++ b/pkg/sentry/vfs/mount_test.go @@ -38,7 +38,7 @@ func TestMountTableInsertLookup(t *testing.T) { mt.Init() mount := &Mount{} - mount.storeKey(VirtualDentry{&Mount{}, &Dentry{}}) + mount.setKey(VirtualDentry{&Mount{}, &Dentry{}}) mt.Insert(mount) if m := mt.Lookup(mount.parent(), mount.point()); m != mount { @@ -79,7 +79,7 @@ const enableComparativeBenchmarks = false func newBenchMount() *Mount { mount := &Mount{} - mount.storeKey(VirtualDentry{&Mount{}, &Dentry{}}) + mount.loadKey(VirtualDentry{&Mount{}, &Dentry{}}) return mount } @@ -94,7 +94,7 @@ func BenchmarkMountTableParallelLookup(b *testing.B) { for i := 0; i < numMounts; i++ { mount := newBenchMount() mt.Insert(mount) - keys = append(keys, mount.loadKey()) + keys = append(keys, mount.saveKey()) } var ready sync.WaitGroup @@ -146,7 +146,7 @@ func BenchmarkMountMapParallelLookup(b *testing.B) { keys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { mount := newBenchMount() - key := mount.loadKey() + key := mount.saveKey() ms[key] = mount keys = append(keys, key) } @@ -201,7 +201,7 @@ func BenchmarkMountSyncMapParallelLookup(b *testing.B) { keys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { mount := newBenchMount() - key := mount.loadKey() + key := mount.getKey() ms.Store(key, mount) keys = append(keys, key) } @@ -283,7 +283,7 @@ func BenchmarkMountMapNegativeLookup(b *testing.B) { ms := make(map[VirtualDentry]*Mount) for i := 0; i < numMounts; i++ { mount := newBenchMount() - ms[mount.loadKey()] = mount + ms[mount.getKey()] = mount } negkeys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { @@ -318,7 +318,7 @@ func BenchmarkMountSyncMapNegativeLookup(b *testing.B) { var ms sync.Map for i := 0; i < numMounts; i++ { mount := newBenchMount() - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } negkeys := make([]VirtualDentry, 0, numMounts) for i := 0; i < numMounts; i++ { @@ -372,7 +372,7 @@ func BenchmarkMountMapInsert(b *testing.B) { b.ResetTimer() for i := range mounts { mount := mounts[i] - ms[mount.loadKey()] = mount + ms[mount.saveKey()] = mount } } @@ -392,7 +392,7 @@ func BenchmarkMountSyncMapInsert(b *testing.B) { b.ResetTimer() for i := range mounts { mount := mounts[i] - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } } @@ -425,13 +425,13 @@ func BenchmarkMountMapRemove(b *testing.B) { ms := make(map[VirtualDentry]*Mount) for i := range mounts { mount := mounts[i] - ms[mount.loadKey()] = mount + ms[mount.saveKey()] = mount } b.ResetTimer() for i := range mounts { mount := mounts[i] - delete(ms, mount.loadKey()) + delete(ms, mount.saveKey()) } } @@ -447,12 +447,12 @@ func BenchmarkMountSyncMapRemove(b *testing.B) { var ms sync.Map for i := range mounts { mount := mounts[i] - ms.Store(mount.loadKey(), mount) + ms.Store(mount.saveKey(), mount) } b.ResetTimer() for i := range mounts { mount := mounts[i] - ms.Delete(mount.loadKey()) + ms.Delete(mount.saveKey()) } } diff --git a/pkg/sentry/vfs/mount_unsafe.go b/pkg/sentry/vfs/mount_unsafe.go index 70f850ca4..cb48c37a1 100644 --- a/pkg/sentry/vfs/mount_unsafe.go +++ b/pkg/sentry/vfs/mount_unsafe.go @@ -13,7 +13,7 @@ // limitations under the License. // +build go1.12 -// +build !go1.16 +// +build !go1.17 // Check go:linkname function signatures when updating Go version. @@ -34,6 +34,8 @@ import ( // structurally identical to VirtualDentry, but stores its fields as // unsafe.Pointer since mutators synchronize with VFS path traversal using // seqcounts. +// +// This is explicitly not savable. type mountKey struct { parent unsafe.Pointer // *Mount point unsafe.Pointer // *Dentry @@ -47,19 +49,23 @@ func (mnt *Mount) point() *Dentry { return (*Dentry)(atomic.LoadPointer(&mnt.key.point)) } -func (mnt *Mount) loadKey() VirtualDentry { +func (mnt *Mount) getKey() VirtualDentry { return VirtualDentry{ mount: mnt.parent(), dentry: mnt.point(), } } +func (mnt *Mount) saveKey() VirtualDentry { return mnt.getKey() } + // Invariant: mnt.key.parent == nil. vd.Ok(). -func (mnt *Mount) storeKey(vd VirtualDentry) { +func (mnt *Mount) setKey(vd VirtualDentry) { atomic.StorePointer(&mnt.key.parent, unsafe.Pointer(vd.mount)) atomic.StorePointer(&mnt.key.point, unsafe.Pointer(vd.dentry)) } +func (mnt *Mount) loadKey(vd VirtualDentry) { mnt.setKey(vd) } + // mountTable maps (mount parent, mount point) pairs to mounts. It supports // efficient concurrent lookup, even in the presence of concurrent mutators // (provided mutation is sufficiently uncommon). @@ -205,6 +211,26 @@ loop: } } +// Range calls f on each Mount in mt. If f returns false, Range stops iteration +// and returns immediately. +func (mt *mountTable) Range(f func(*Mount) bool) { + tcap := uintptr(1) << (mt.size & mtSizeOrderMask) + slotPtr := mt.slots + last := unsafe.Pointer(uintptr(mt.slots) + ((tcap - 1) * mountSlotBytes)) + for { + slot := (*mountSlot)(slotPtr) + if slot.value != nil { + if !f((*Mount)(slot.value)) { + return + } + } + if slotPtr == last { + return + } + slotPtr = unsafe.Pointer(uintptr(slotPtr) + mountSlotBytes) + } +} + // Insert inserts the given mount into mt. // // Preconditions: mt must not already contain a Mount with the same mount point @@ -217,8 +243,9 @@ func (mt *mountTable) Insert(mount *Mount) { // insertSeqed inserts the given mount into mt. // -// Preconditions: mt.seq must be in a writer critical section. mt must not -// already contain a Mount with the same mount point and parent. +// Preconditions: +// * mt.seq must be in a writer critical section. +// * mt must not already contain a Mount with the same mount point and parent. func (mt *mountTable) insertSeqed(mount *Mount) { hash := memhash(unsafe.Pointer(&mount.key), uintptr(mt.seed), mountKeyBytes) @@ -269,9 +296,11 @@ func (mt *mountTable) insertSeqed(mount *Mount) { atomic.StorePointer(&mt.slots, newSlots) } -// Preconditions: There are no concurrent mutators of the table (slots, cap). -// If the table is visible to readers, then mt.seq must be in a writer critical -// section. cap must be a power of 2. +// Preconditions: +// * There are no concurrent mutators of the table (slots, cap). +// * If the table is visible to readers, then mt.seq must be in a writer +// critical section. +// * cap must be a power of 2. func mtInsertLocked(slots unsafe.Pointer, cap uintptr, value unsafe.Pointer, hash uintptr) { mask := cap - 1 off := (hash & mask) * mountSlotBytes @@ -313,8 +342,9 @@ func (mt *mountTable) Remove(mount *Mount) { // removeSeqed removes the given mount from mt. // -// Preconditions: mt.seq must be in a writer critical section. mt must contain -// mount. +// Preconditions: +// * mt.seq must be in a writer critical section. +// * mt must contain mount. func (mt *mountTable) removeSeqed(mount *Mount) { hash := memhash(unsafe.Pointer(&mount.key), uintptr(mt.seed), mountKeyBytes) tcap := uintptr(1) << (mt.size & mtSizeOrderMask) diff --git a/pkg/sentry/vfs/options.go b/pkg/sentry/vfs/options.go index dfc8573fd..bc79e5ecc 100644 --- a/pkg/sentry/vfs/options.go +++ b/pkg/sentry/vfs/options.go @@ -21,6 +21,8 @@ import ( // GetDentryOptions contains options to VirtualFilesystem.GetDentryAt() and // FilesystemImpl.GetDentryAt(). +// +// +stateify savable type GetDentryOptions struct { // If CheckSearchable is true, FilesystemImpl.GetDentryAt() must check that // the returned Dentry is a directory for which creds has search @@ -30,6 +32,8 @@ type GetDentryOptions struct { // MkdirOptions contains options to VirtualFilesystem.MkdirAt() and // FilesystemImpl.MkdirAt(). +// +// +stateify savable type MkdirOptions struct { // Mode is the file mode bits for the created directory. Mode linux.FileMode @@ -56,6 +60,8 @@ type MkdirOptions struct { // MknodOptions contains options to VirtualFilesystem.MknodAt() and // FilesystemImpl.MknodAt(). +// +// +stateify savable type MknodOptions struct { // Mode is the file type and mode bits for the created file. Mode linux.FileMode @@ -72,6 +78,8 @@ type MknodOptions struct { // MountFlags contains flags as specified for mount(2), e.g. MS_NOEXEC. // MS_RDONLY is not part of MountFlags because it's tracked in Mount.writers. +// +// +stateify savable type MountFlags struct { // NoExec is equivalent to MS_NOEXEC. NoExec bool @@ -93,6 +101,8 @@ type MountFlags struct { } // MountOptions contains options to VirtualFilesystem.MountAt(). +// +// +stateify savable type MountOptions struct { // Flags contains flags as specified for mount(2), e.g. MS_NOEXEC. Flags MountFlags @@ -103,13 +113,17 @@ type MountOptions struct { // GetFilesystemOptions contains options to FilesystemType.GetFilesystem(). GetFilesystemOptions GetFilesystemOptions - // If InternalMount is true, allow the use of filesystem types for which - // RegisterFilesystemTypeOptions.AllowUserMount == false. + // InternalMount indicates whether the mount operation is coming from the + // application, i.e. through mount(2). If InternalMount is true, allow the use + // of filesystem types for which RegisterFilesystemTypeOptions.AllowUserMount + // == false. InternalMount bool } // OpenOptions contains options to VirtualFilesystem.OpenAt() and // FilesystemImpl.OpenAt(). +// +// +stateify savable type OpenOptions struct { // Flags contains access mode and flags as specified for open(2). // @@ -135,6 +149,8 @@ type OpenOptions struct { // ReadOptions contains options to FileDescription.PRead(), // FileDescriptionImpl.PRead(), FileDescription.Read(), and // FileDescriptionImpl.Read(). +// +// +stateify savable type ReadOptions struct { // Flags contains flags as specified for preadv2(2). Flags uint32 @@ -142,6 +158,8 @@ type ReadOptions struct { // RenameOptions contains options to VirtualFilesystem.RenameAt() and // FilesystemImpl.RenameAt(). +// +// +stateify savable type RenameOptions struct { // Flags contains flags as specified for renameat2(2). Flags uint32 @@ -153,6 +171,8 @@ type RenameOptions struct { // SetStatOptions contains options to VirtualFilesystem.SetStatAt(), // FilesystemImpl.SetStatAt(), FileDescription.SetStat(), and // FileDescriptionImpl.SetStat(). +// +// +stateify savable type SetStatOptions struct { // Stat is the metadata that should be set. Only fields indicated by // Stat.Mask should be set. @@ -174,6 +194,8 @@ type SetStatOptions struct { // BoundEndpointOptions contains options to VirtualFilesystem.BoundEndpointAt() // and FilesystemImpl.BoundEndpointAt(). +// +// +stateify savable type BoundEndpointOptions struct { // Addr is the path of the file whose socket endpoint is being retrieved. // It is generally irrelevant: most endpoints are stored at a dentry that @@ -190,10 +212,12 @@ type BoundEndpointOptions struct { Addr string } -// GetxattrOptions contains options to VirtualFilesystem.GetxattrAt(), -// FilesystemImpl.GetxattrAt(), FileDescription.Getxattr(), and -// FileDescriptionImpl.Getxattr(). -type GetxattrOptions struct { +// GetXattrOptions contains options to VirtualFilesystem.GetXattrAt(), +// FilesystemImpl.GetXattrAt(), FileDescription.GetXattr(), and +// FileDescriptionImpl.GetXattr(). +// +// +stateify savable +type GetXattrOptions struct { // Name is the name of the extended attribute to retrieve. Name string @@ -204,10 +228,12 @@ type GetxattrOptions struct { Size uint64 } -// SetxattrOptions contains options to VirtualFilesystem.SetxattrAt(), -// FilesystemImpl.SetxattrAt(), FileDescription.Setxattr(), and -// FileDescriptionImpl.Setxattr(). -type SetxattrOptions struct { +// SetXattrOptions contains options to VirtualFilesystem.SetXattrAt(), +// FilesystemImpl.SetXattrAt(), FileDescription.SetXattr(), and +// FileDescriptionImpl.SetXattr(). +// +// +stateify savable +type SetXattrOptions struct { // Name is the name of the extended attribute being mutated. Name string @@ -221,6 +247,8 @@ type SetxattrOptions struct { // StatOptions contains options to VirtualFilesystem.StatAt(), // FilesystemImpl.StatAt(), FileDescription.Stat(), and // FileDescriptionImpl.Stat(). +// +// +stateify savable type StatOptions struct { // Mask is the set of fields in the returned Statx that the FilesystemImpl // or FileDescriptionImpl should provide. Bits are as in linux.Statx.Mask. @@ -238,6 +266,8 @@ type StatOptions struct { } // UmountOptions contains options to VirtualFilesystem.UmountAt(). +// +// +stateify savable type UmountOptions struct { // Flags contains flags as specified for umount2(2). Flags uint32 @@ -246,6 +276,8 @@ type UmountOptions struct { // WriteOptions contains options to FileDescription.PWrite(), // FileDescriptionImpl.PWrite(), FileDescription.Write(), and // FileDescriptionImpl.Write(). +// +// +stateify savable type WriteOptions struct { // Flags contains flags as specified for pwritev2(2). Flags uint32 diff --git a/pkg/sentry/vfs/permissions.go b/pkg/sentry/vfs/permissions.go index 33389c1df..d48520d58 100644 --- a/pkg/sentry/vfs/permissions.go +++ b/pkg/sentry/vfs/permissions.go @@ -16,6 +16,7 @@ package vfs import ( "math" + "strings" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -25,6 +26,8 @@ import ( ) // AccessTypes is a bitmask of Unix file permissions. +// +// +stateify savable type AccessTypes uint16 // Bits in AccessTypes. @@ -271,7 +274,7 @@ func HasCapabilityOnFile(creds *auth.Credentials, cp linux.Capability, kuid auth // operation must not proceed. Otherwise it returns the max length allowed to // without violating the limit. func CheckLimit(ctx context.Context, offset, size int64) (int64, error) { - fileSizeLimit := limits.FromContext(ctx).Get(limits.FileSize).Cur + fileSizeLimit := limits.FromContextOrDie(ctx).Get(limits.FileSize).Cur if fileSizeLimit > math.MaxInt64 { return size, nil } @@ -284,3 +287,40 @@ func CheckLimit(ctx context.Context, offset, size int64) (int64, error) { } return size, nil } + +// CheckXattrPermissions checks permissions for extended attribute access. +// This is analogous to fs/xattr.c:xattr_permission(). Some key differences: +// * Does not check for read-only filesystem property. +// * Does not check inode immutability or append only mode. In both cases EPERM +// must be returned by filesystem implementations. +// * Does not do inode permission checks. Filesystem implementations should +// handle inode permission checks as they may differ across implementations. +func CheckXattrPermissions(creds *auth.Credentials, ats AccessTypes, mode linux.FileMode, kuid auth.KUID, name string) error { + switch { + case strings.HasPrefix(name, linux.XATTR_TRUSTED_PREFIX): + // The trusted.* namespace can only be accessed by privileged + // users. + if creds.HasCapability(linux.CAP_SYS_ADMIN) { + return nil + } + if ats.MayWrite() { + return syserror.EPERM + } + return syserror.ENODATA + case strings.HasPrefix(name, linux.XATTR_USER_PREFIX): + // In the user.* namespace, only regular files and directories can have + // extended attributes. For sticky directories, only the owner and + // privileged users can write attributes. + filetype := mode.FileType() + if filetype != linux.ModeRegular && filetype != linux.ModeDirectory { + if ats.MayWrite() { + return syserror.EPERM + } + return syserror.ENODATA + } + if filetype == linux.ModeDirectory && mode&linux.ModeSticky != 0 && ats.MayWrite() && !CanActAsOwner(creds, kuid) { + return syserror.EPERM + } + } + return nil +} diff --git a/pkg/sentry/vfs/resolving_path.go b/pkg/sentry/vfs/resolving_path.go index 3304372d9..e4fd55012 100644 --- a/pkg/sentry/vfs/resolving_path.go +++ b/pkg/sentry/vfs/resolving_path.go @@ -35,6 +35,8 @@ import ( // FilesystemImpl methods. // // ResolvingPath is loosely analogous to Linux's struct nameidata. +// +// +stateify savable type ResolvingPath struct { vfs *VirtualFilesystem root VirtualDentry // refs borrowed from PathOperation @@ -88,6 +90,7 @@ func init() { // so error "constants" are really mutable vars, necessitating somewhat // expensive interface object comparisons. +// +stateify savable type resolveMountRootOrJumpError struct{} // Error implements error.Error. @@ -95,6 +98,7 @@ func (resolveMountRootOrJumpError) Error() string { return "resolving mount root or jump" } +// +stateify savable type resolveMountPointError struct{} // Error implements error.Error. @@ -102,6 +106,7 @@ func (resolveMountPointError) Error() string { return "resolving mount point" } +// +stateify savable type resolveAbsSymlinkError struct{} // Error implements error.Error. diff --git a/pkg/sentry/vfs/save_restore.go b/pkg/sentry/vfs/save_restore.go new file mode 100644 index 000000000..46e50d55d --- /dev/null +++ b/pkg/sentry/vfs/save_restore.go @@ -0,0 +1,124 @@ +// Copyright 2020 The gVisor Authors. +// +// Licensed under the Apache License, Version 2.0 (the "License"); +// you may not use this file except in compliance with the License. +// You may obtain a copy of the License at +// +// http://www.apache.org/licenses/LICENSE-2.0 +// +// Unless required by applicable law or agreed to in writing, software +// distributed under the License is distributed on an "AS IS" BASIS, +// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +// See the License for the specific language governing permissions and +// limitations under the License. + +package vfs + +import ( + "fmt" + "sync/atomic" + + "gvisor.dev/gvisor/pkg/context" + "gvisor.dev/gvisor/pkg/refsvfs2" +) + +// FilesystemImplSaveRestoreExtension is an optional extension to +// FilesystemImpl. +type FilesystemImplSaveRestoreExtension interface { + // PrepareSave prepares this filesystem for serialization. + PrepareSave(ctx context.Context) error + + // CompleteRestore completes restoration from checkpoint for this + // filesystem after deserialization. + CompleteRestore(ctx context.Context, opts CompleteRestoreOptions) error +} + +// PrepareSave prepares all filesystems for serialization. +func (vfs *VirtualFilesystem) PrepareSave(ctx context.Context) error { + failures := 0 + for fs := range vfs.getFilesystems() { + if ext, ok := fs.impl.(FilesystemImplSaveRestoreExtension); ok { + if err := ext.PrepareSave(ctx); err != nil { + ctx.Warningf("%T.PrepareSave failed: %v", fs.impl, err) + failures++ + } + } + fs.DecRef(ctx) + } + if failures != 0 { + return fmt.Errorf("%d filesystems failed to prepare for serialization", failures) + } + return nil +} + +// CompleteRestore completes restoration from checkpoint for all filesystems +// after deserialization. +func (vfs *VirtualFilesystem) CompleteRestore(ctx context.Context, opts *CompleteRestoreOptions) error { + failures := 0 + for fs := range vfs.getFilesystems() { + if ext, ok := fs.impl.(FilesystemImplSaveRestoreExtension); ok { + if err := ext.CompleteRestore(ctx, *opts); err != nil { + ctx.Warningf("%T.CompleteRestore failed: %v", fs.impl, err) + failures++ + } + } + fs.DecRef(ctx) + } + if failures != 0 { + return fmt.Errorf("%d filesystems failed to complete restore after deserialization", failures) + } + return nil +} + +// CompleteRestoreOptions contains options to +// VirtualFilesystem.CompleteRestore() and +// FilesystemImplSaveRestoreExtension.CompleteRestore(). +type CompleteRestoreOptions struct { + // If ValidateFileSizes is true, filesystem implementations backed by + // remote filesystems should verify that file sizes have not changed + // between checkpoint and restore. + ValidateFileSizes bool + + // If ValidateFileModificationTimestamps is true, filesystem + // implementations backed by remote filesystems should validate that file + // mtimes have not changed between checkpoint and restore. + ValidateFileModificationTimestamps bool +} + +// saveMounts is called by stateify. +func (vfs *VirtualFilesystem) saveMounts() []*Mount { + if atomic.LoadPointer(&vfs.mounts.slots) == nil { + // vfs.Init() was never called. + return nil + } + var mounts []*Mount + vfs.mounts.Range(func(mount *Mount) bool { + mounts = append(mounts, mount) + return true + }) + return mounts +} + +// loadMounts is called by stateify. +func (vfs *VirtualFilesystem) loadMounts(mounts []*Mount) { + if mounts == nil { + return + } + vfs.mounts.Init() + for _, mount := range mounts { + vfs.mounts.Insert(mount) + } +} + +func (mnt *Mount) afterLoad() { + if refsvfs2.LeakCheckEnabled() && atomic.LoadInt64(&mnt.refs) != 0 { + refsvfs2.Register(mnt, "vfs.Mount") + } +} + +// afterLoad is called by stateify. +func (epi *epollInterest) afterLoad() { + // Mark all epollInterests as ready after restore so that the next call to + // EpollInstance.ReadEvents() rechecks their readiness. + epi.Callback(nil) +} diff --git a/pkg/sentry/vfs/vfs.go b/pkg/sentry/vfs/vfs.go index 9c2420683..48d6252f7 100644 --- a/pkg/sentry/vfs/vfs.go +++ b/pkg/sentry/vfs/vfs.go @@ -24,9 +24,9 @@ // Locks acquired by FilesystemImpls between Prepare{Delete,Rename}Dentry and Commit{Delete,Rename*}Dentry // VirtualFilesystem.filesystemsMu // EpollInstance.mu -// Inotify.mu -// Watches.mu -// Inotify.evMu +// Inotify.mu +// Watches.mu +// Inotify.evMu // VirtualFilesystem.fsTypesMu // // Locking Dentry.mu in multiple Dentries requires holding @@ -36,6 +36,7 @@ package vfs import ( "fmt" + "path" "gvisor.dev/gvisor/pkg/abi/linux" "gvisor.dev/gvisor/pkg/context" @@ -70,7 +71,7 @@ type VirtualFilesystem struct { // points. // // mounts is analogous to Linux's mount_hashtable. - mounts mountTable + mounts mountTable `state:".([]*Mount)"` // mountpoints maps mount points to mounts at those points in all // namespaces. mountpoints is protected by mountMu. @@ -157,11 +158,23 @@ func (vfs *VirtualFilesystem) Init(ctx context.Context) error { return nil } +// Release drops references on filesystem objects held by vfs. +// +// Precondition: This must be called after VFS.Init() has succeeded. +func (vfs *VirtualFilesystem) Release(ctx context.Context) { + vfs.anonMount.DecRef(ctx) + for _, fst := range vfs.fsTypes { + fst.fsType.Release(ctx) + } +} + // PathOperation specifies the path operated on by a VFS method. // // PathOperation is passed to VFS methods by pointer to reduce memory copying: // it's somewhat large and should never escape. (Options structs are passed by // pointer to VFS and FileDescription methods for the same reason.) +// +// +stateify savable type PathOperation struct { // Root is the VFS root. References on Root are borrowed from the provider // of the PathOperation. @@ -296,6 +309,8 @@ func (vfs *VirtualFilesystem) LinkAt(ctx context.Context, creds *auth.Credential // MkdirAt creates a directory at the given path. func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MkdirOptions) error { if !pop.Path.Begin.Ok() { + // pop.Path should not be empty in operations that create/delete files. + // This is consistent with mkdirat(dirfd, "", mode). if pop.Path.Absolute { return syserror.EEXIST } @@ -332,6 +347,8 @@ func (vfs *VirtualFilesystem) MkdirAt(ctx context.Context, creds *auth.Credentia // error from the syserror package. func (vfs *VirtualFilesystem) MknodAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *MknodOptions) error { if !pop.Path.Begin.Ok() { + // pop.Path should not be empty in operations that create/delete files. + // This is consistent with mknodat(dirfd, "", mode, dev). if pop.Path.Absolute { return syserror.EEXIST } @@ -517,6 +534,8 @@ func (vfs *VirtualFilesystem) RenameAt(ctx context.Context, creds *auth.Credenti // RmdirAt removes the directory at the given path. func (vfs *VirtualFilesystem) RmdirAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) error { if !pop.Path.Begin.Ok() { + // pop.Path should not be empty in operations that create/delete files. + // This is consistent with unlinkat(dirfd, "", AT_REMOVEDIR). if pop.Path.Absolute { return syserror.EBUSY } @@ -598,6 +617,8 @@ func (vfs *VirtualFilesystem) StatFSAt(ctx context.Context, creds *auth.Credenti // SymlinkAt creates a symbolic link at the given path with the given target. func (vfs *VirtualFilesystem) SymlinkAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, target string) error { if !pop.Path.Begin.Ok() { + // pop.Path should not be empty in operations that create/delete files. + // This is consistent with symlinkat(oldpath, newdirfd, ""). if pop.Path.Absolute { return syserror.EEXIST } @@ -630,6 +651,8 @@ func (vfs *VirtualFilesystem) SymlinkAt(ctx context.Context, creds *auth.Credent // UnlinkAt deletes the non-directory file at the given path. func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation) error { if !pop.Path.Begin.Ok() { + // pop.Path should not be empty in operations that create/delete files. + // This is consistent with unlinkat(dirfd, "", 0). if pop.Path.Absolute { return syserror.EBUSY } @@ -661,12 +684,6 @@ func (vfs *VirtualFilesystem) UnlinkAt(ctx context.Context, creds *auth.Credenti // BoundEndpointAt gets the bound endpoint at the given path, if one exists. func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *BoundEndpointOptions) (transport.BoundEndpoint, error) { - if !pop.Path.Begin.Ok() { - if pop.Path.Absolute { - return nil, syserror.ECONNREFUSED - } - return nil, syserror.ENOENT - } rp := vfs.getResolvingPath(creds, pop) for { bep, err := rp.mount.fs.impl.BoundEndpointAt(ctx, rp, *opts) @@ -686,12 +703,12 @@ func (vfs *VirtualFilesystem) BoundEndpointAt(ctx context.Context, creds *auth.C } } -// ListxattrAt returns all extended attribute names for the file at the given +// ListXattrAt returns all extended attribute names for the file at the given // path. -func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, size uint64) ([]string, error) { +func (vfs *VirtualFilesystem) ListXattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, size uint64) ([]string, error) { rp := vfs.getResolvingPath(creds, pop) for { - names, err := rp.mount.fs.impl.ListxattrAt(ctx, rp, size) + names, err := rp.mount.fs.impl.ListXattrAt(ctx, rp, size) if err == nil { vfs.putResolvingPath(ctx, rp) return names, nil @@ -711,12 +728,12 @@ func (vfs *VirtualFilesystem) ListxattrAt(ctx context.Context, creds *auth.Crede } } -// GetxattrAt returns the value associated with the given extended attribute +// GetXattrAt returns the value associated with the given extended attribute // for the file at the given path. -func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetxattrOptions) (string, error) { +func (vfs *VirtualFilesystem) GetXattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *GetXattrOptions) (string, error) { rp := vfs.getResolvingPath(creds, pop) for { - val, err := rp.mount.fs.impl.GetxattrAt(ctx, rp, *opts) + val, err := rp.mount.fs.impl.GetXattrAt(ctx, rp, *opts) if err == nil { vfs.putResolvingPath(ctx, rp) return val, nil @@ -728,12 +745,12 @@ func (vfs *VirtualFilesystem) GetxattrAt(ctx context.Context, creds *auth.Creden } } -// SetxattrAt changes the value associated with the given extended attribute +// SetXattrAt changes the value associated with the given extended attribute // for the file at the given path. -func (vfs *VirtualFilesystem) SetxattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *SetxattrOptions) error { +func (vfs *VirtualFilesystem) SetXattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, opts *SetXattrOptions) error { rp := vfs.getResolvingPath(creds, pop) for { - err := rp.mount.fs.impl.SetxattrAt(ctx, rp, *opts) + err := rp.mount.fs.impl.SetXattrAt(ctx, rp, *opts) if err == nil { vfs.putResolvingPath(ctx, rp) return nil @@ -745,11 +762,11 @@ func (vfs *VirtualFilesystem) SetxattrAt(ctx context.Context, creds *auth.Creden } } -// RemovexattrAt removes the given extended attribute from the file at rp. -func (vfs *VirtualFilesystem) RemovexattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) error { +// RemoveXattrAt removes the given extended attribute from the file at rp. +func (vfs *VirtualFilesystem) RemoveXattrAt(ctx context.Context, creds *auth.Credentials, pop *PathOperation, name string) error { rp := vfs.getResolvingPath(creds, pop) for { - err := rp.mount.fs.impl.RemovexattrAt(ctx, rp, name) + err := rp.mount.fs.impl.RemoveXattrAt(ctx, rp, name) if err == nil { vfs.putResolvingPath(ctx, rp) return nil @@ -763,23 +780,83 @@ func (vfs *VirtualFilesystem) RemovexattrAt(ctx context.Context, creds *auth.Cre // SyncAllFilesystems has the semantics of Linux's sync(2). func (vfs *VirtualFilesystem) SyncAllFilesystems(ctx context.Context) error { + var retErr error + for fs := range vfs.getFilesystems() { + if err := fs.impl.Sync(ctx); err != nil && retErr == nil { + retErr = err + } + fs.DecRef(ctx) + } + return retErr +} + +func (vfs *VirtualFilesystem) getFilesystems() map[*Filesystem]struct{} { fss := make(map[*Filesystem]struct{}) vfs.filesystemsMu.Lock() + defer vfs.filesystemsMu.Unlock() for fs := range vfs.filesystems { if !fs.TryIncRef() { continue } fss[fs] = struct{}{} } - vfs.filesystemsMu.Unlock() - var retErr error - for fs := range fss { - if err := fs.impl.Sync(ctx); err != nil && retErr == nil { - retErr = err - } - fs.DecRef(ctx) + return fss +} + +// MkdirAllAt recursively creates non-existent directories on the given path +// (including the last component). +func (vfs *VirtualFilesystem) MkdirAllAt(ctx context.Context, currentPath string, root VirtualDentry, creds *auth.Credentials, mkdirOpts *MkdirOptions) error { + pop := &PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(currentPath), + } + stat, err := vfs.StatAt(ctx, creds, pop, &StatOptions{Mask: linux.STATX_TYPE}) + switch err { + case nil: + if stat.Mask&linux.STATX_TYPE == 0 || stat.Mode&linux.FileTypeMask != linux.ModeDirectory { + return syserror.ENOTDIR + } + // Directory already exists. + return nil + case syserror.ENOENT: + // Expected, we will create the dir. + default: + return fmt.Errorf("stat failed for %q during directory creation: %w", currentPath, err) + } + + // Recurse to ensure parent is created and then create the final directory. + if err := vfs.MkdirAllAt(ctx, path.Dir(currentPath), root, creds, mkdirOpts); err != nil { + return err } - return retErr + if err := vfs.MkdirAt(ctx, creds, pop, mkdirOpts); err != nil { + return fmt.Errorf("failed to create directory %q: %w", currentPath, err) + } + return nil +} + +// MakeSyntheticMountpoint creates parent directories of target if they do not +// exist and attempts to create a directory for the mountpoint. If a +// non-directory file already exists there then we allow it. +func (vfs *VirtualFilesystem) MakeSyntheticMountpoint(ctx context.Context, target string, root VirtualDentry, creds *auth.Credentials) error { + mkdirOpts := &MkdirOptions{Mode: 0777, ForSyntheticMountpoint: true} + + // Make sure the parent directory of target exists. + if err := vfs.MkdirAllAt(ctx, path.Dir(target), root, creds, mkdirOpts); err != nil { + return fmt.Errorf("failed to create parent directory of mountpoint %q: %w", target, err) + } + + // Attempt to mkdir the final component. If a file (of any type) exists + // then we let allow mounting on top of that because we do not require the + // target to be an existing directory, unlike Linux mount(2). + if err := vfs.MkdirAt(ctx, creds, &PathOperation{ + Root: root, + Start: root, + Path: fspath.Parse(target), + }, mkdirOpts); err != nil && err != syserror.EEXIST { + return fmt.Errorf("failed to create mountpoint %q: %w", target, err) + } + return nil } // A VirtualDentry represents a node in a VFS tree, by combining a Dentry diff --git a/pkg/sentry/watchdog/watchdog.go b/pkg/sentry/watchdog/watchdog.go index 748273366..bbafb8b7f 100644 --- a/pkg/sentry/watchdog/watchdog.go +++ b/pkg/sentry/watchdog/watchdog.go @@ -96,15 +96,33 @@ const ( Panic ) +// Set implements flag.Value. +func (a *Action) Set(v string) error { + switch v { + case "log", "logwarning": + *a = LogWarning + case "panic": + *a = Panic + default: + return fmt.Errorf("invalid watchdog action %q", v) + } + return nil +} + +// Get implements flag.Value. +func (a *Action) Get() interface{} { + return *a +} + // String returns Action's string representation. -func (a Action) String() string { - switch a { +func (a *Action) String() string { + switch *a { case LogWarning: - return "LogWarning" + return "logWarning" case Panic: - return "Panic" + return "panic" default: - panic(fmt.Sprintf("Invalid action: %d", a)) + panic(fmt.Sprintf("Invalid watchdog action: %d", *a)) } } |