summaryrefslogtreecommitdiffhomepage
path: root/pkg/sentry/platform
diff options
context:
space:
mode:
Diffstat (limited to 'pkg/sentry/platform')
-rw-r--r--pkg/sentry/platform/BUILD20
-rw-r--r--pkg/sentry/platform/kvm/BUILD5
-rw-r--r--pkg/sentry/platform/kvm/address_space.go82
-rw-r--r--pkg/sentry/platform/kvm/bluepill_allocator.go (renamed from pkg/sentry/platform/kvm/allocator.go)52
-rw-r--r--pkg/sentry/platform/kvm/bluepill_amd64.go12
-rw-r--r--pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go32
-rw-r--r--pkg/sentry/platform/kvm/bluepill_arm64.go19
-rw-r--r--pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go36
-rw-r--r--pkg/sentry/platform/kvm/bluepill_unsafe.go70
-rw-r--r--pkg/sentry/platform/kvm/context.go10
-rw-r--r--pkg/sentry/platform/kvm/kvm_amd64_test.go51
-rw-r--r--pkg/sentry/platform/kvm/kvm_arm64.go15
-rw-r--r--pkg/sentry/platform/kvm/kvm_arm64_unsafe.go4
-rw-r--r--pkg/sentry/platform/kvm/kvm_const.go18
-rw-r--r--pkg/sentry/platform/kvm/kvm_const_arm64.go19
-rw-r--r--pkg/sentry/platform/kvm/kvm_test.go24
-rw-r--r--pkg/sentry/platform/kvm/machine.go52
-rw-r--r--pkg/sentry/platform/kvm/machine_amd64.go34
-rw-r--r--pkg/sentry/platform/kvm/machine_amd64_unsafe.go2
-rw-r--r--pkg/sentry/platform/kvm/machine_arm64.go80
-rw-r--r--pkg/sentry/platform/kvm/machine_arm64_unsafe.go50
-rw-r--r--pkg/sentry/platform/kvm/machine_unsafe.go4
-rw-r--r--pkg/sentry/platform/kvm/testutil/testutil_arm64.go4
-rw-r--r--pkg/sentry/platform/kvm/testutil/testutil_arm64.s12
-rw-r--r--pkg/sentry/platform/platform.go100
-rw-r--r--pkg/sentry/platform/ptrace/BUILD2
-rw-r--r--pkg/sentry/platform/ptrace/ptrace.go10
-rw-r--r--pkg/sentry/platform/ptrace/subprocess.go9
-rw-r--r--pkg/sentry/platform/ptrace/subprocess_unsafe.go2
-rw-r--r--pkg/sentry/platform/ring0/aarch64.go35
-rw-r--r--pkg/sentry/platform/ring0/entry_arm64.s98
-rw-r--r--pkg/sentry/platform/ring0/kernel.go24
-rw-r--r--pkg/sentry/platform/ring0/kernel_amd64.go12
-rw-r--r--pkg/sentry/platform/ring0/kernel_arm64.go12
-rw-r--r--pkg/sentry/platform/ring0/lib_arm64.go6
-rw-r--r--pkg/sentry/platform/ring0/lib_arm64.s86
-rw-r--r--pkg/sentry/platform/ring0/pagetables/allocator.go11
-rw-r--r--pkg/sentry/platform/ring0/pagetables/pagetables.go8
-rw-r--r--pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go9
39 files changed, 753 insertions, 378 deletions
diff --git a/pkg/sentry/platform/BUILD b/pkg/sentry/platform/BUILD
index 453241eca..209b28053 100644
--- a/pkg/sentry/platform/BUILD
+++ b/pkg/sentry/platform/BUILD
@@ -1,39 +1,21 @@
load("//tools:defs.bzl", "go_library")
-load("//tools/go_generics:defs.bzl", "go_template_instance")
package(licenses = ["notice"])
-go_template_instance(
- name = "file_range",
- out = "file_range.go",
- package = "platform",
- prefix = "File",
- template = "//pkg/segment:generic_range",
- types = {
- "T": "uint64",
- },
-)
-
go_library(
name = "platform",
srcs = [
"context.go",
- "file_range.go",
"mmap_min_addr.go",
"platform.go",
],
visibility = ["//pkg/sentry:internal"],
deps = [
"//pkg/abi/linux",
- "//pkg/atomicbitops",
"//pkg/context",
- "//pkg/log",
- "//pkg/safecopy",
- "//pkg/safemem",
"//pkg/seccomp",
"//pkg/sentry/arch",
- "//pkg/sentry/usage",
- "//pkg/syserror",
+ "//pkg/sentry/memmap",
"//pkg/usermem",
],
)
diff --git a/pkg/sentry/platform/kvm/BUILD b/pkg/sentry/platform/kvm/BUILD
index 159f7eafd..3970dd81d 100644
--- a/pkg/sentry/platform/kvm/BUILD
+++ b/pkg/sentry/platform/kvm/BUILD
@@ -6,8 +6,8 @@ go_library(
name = "kvm",
srcs = [
"address_space.go",
- "allocator.go",
"bluepill.go",
+ "bluepill_allocator.go",
"bluepill_amd64.go",
"bluepill_amd64.s",
"bluepill_amd64_unsafe.go",
@@ -41,12 +41,14 @@ go_library(
deps = [
"//pkg/abi/linux",
"//pkg/atomicbitops",
+ "//pkg/context",
"//pkg/cpuid",
"//pkg/log",
"//pkg/procid",
"//pkg/safecopy",
"//pkg/seccomp",
"//pkg/sentry/arch",
+ "//pkg/sentry/memmap",
"//pkg/sentry/platform",
"//pkg/sentry/platform/interrupt",
"//pkg/sentry/platform/ring0",
@@ -60,6 +62,7 @@ go_library(
go_test(
name = "kvm_test",
srcs = [
+ "kvm_amd64_test.go",
"kvm_test.go",
"virtual_map_test.go",
],
diff --git a/pkg/sentry/platform/kvm/address_space.go b/pkg/sentry/platform/kvm/address_space.go
index be213bfe8..af5c5e191 100644
--- a/pkg/sentry/platform/kvm/address_space.go
+++ b/pkg/sentry/platform/kvm/address_space.go
@@ -18,6 +18,7 @@ import (
"sync/atomic"
"gvisor.dev/gvisor/pkg/atomicbitops"
+ "gvisor.dev/gvisor/pkg/sentry/memmap"
"gvisor.dev/gvisor/pkg/sentry/platform"
"gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables"
"gvisor.dev/gvisor/pkg/sync"
@@ -26,16 +27,15 @@ import (
// dirtySet tracks vCPUs for invalidation.
type dirtySet struct {
- vCPUs []uint64
+ vCPUMasks []uint64
}
// forEach iterates over all CPUs in the dirty set.
+//
+//go:nosplit
func (ds *dirtySet) forEach(m *machine, fn func(c *vCPU)) {
- m.mu.RLock()
- defer m.mu.RUnlock()
-
- for index := range ds.vCPUs {
- mask := atomic.SwapUint64(&ds.vCPUs[index], 0)
+ for index := range ds.vCPUMasks {
+ mask := atomic.SwapUint64(&ds.vCPUMasks[index], 0)
if mask != 0 {
for bit := 0; bit < 64; bit++ {
if mask&(1<<uint64(bit)) == 0 {
@@ -54,7 +54,7 @@ func (ds *dirtySet) mark(c *vCPU) bool {
index := uint64(c.id) / 64
bit := uint64(1) << uint(c.id%64)
- oldValue := atomic.LoadUint64(&ds.vCPUs[index])
+ oldValue := atomic.LoadUint64(&ds.vCPUMasks[index])
if oldValue&bit != 0 {
return false // Not clean.
}
@@ -62,7 +62,7 @@ func (ds *dirtySet) mark(c *vCPU) bool {
// Set the bit unilaterally, and ensure that a flush takes place. Note
// that it's possible for races to occur here, but since the flush is
// taking place long after these lines there's no race in practice.
- atomicbitops.OrUint64(&ds.vCPUs[index], bit)
+ atomicbitops.OrUint64(&ds.vCPUMasks[index], bit)
return true // Previously clean.
}
@@ -113,7 +113,12 @@ type hostMapEntry struct {
length uintptr
}
-func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.AccessType) (inv bool) {
+// mapLocked maps the given host entry.
+//
+// +checkescape:hard,stack
+//
+//go:nosplit
+func (as *addressSpace) mapLocked(addr usermem.Addr, m hostMapEntry, at usermem.AccessType) (inv bool) {
for m.length > 0 {
physical, length, ok := translateToPhysical(m.addr)
if !ok {
@@ -133,18 +138,10 @@ func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.Ac
// important; if the pagetable mappings were installed before
// ensuring the physical pages were available, then some other
// thread could theoretically access them.
- //
- // Due to the way KVM's shadow paging implementation works,
- // modifications to the page tables while in host mode may not
- // be trapped, leading to the shadow pages being out of sync.
- // Therefore, we need to ensure that we are in guest mode for
- // page table modifications. See the call to bluepill, below.
- as.machine.retryInGuest(func() {
- inv = as.pageTables.Map(addr, length, pagetables.MapOpts{
- AccessType: at,
- User: true,
- }, physical) || inv
- })
+ inv = as.pageTables.Map(addr, length, pagetables.MapOpts{
+ AccessType: at,
+ User: true,
+ }, physical) || inv
m.addr += length
m.length -= length
addr += usermem.Addr(length)
@@ -154,7 +151,7 @@ func (as *addressSpace) mapHost(addr usermem.Addr, m hostMapEntry, at usermem.Ac
}
// MapFile implements platform.AddressSpace.MapFile.
-func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+func (as *addressSpace) MapFile(addr usermem.Addr, f memmap.File, fr memmap.FileRange, at usermem.AccessType, precommit bool) error {
as.mu.Lock()
defer as.mu.Unlock()
@@ -176,6 +173,10 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform.
return err
}
+ // See block in mapLocked.
+ as.pageTables.Allocator.(*allocator).cpu = as.machine.Get()
+ defer as.machine.Put(as.pageTables.Allocator.(*allocator).cpu)
+
// Map the mappings in the sentry's address space (guest physical memory)
// into the application's address space (guest virtual memory).
inv := false
@@ -190,7 +191,12 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform.
_ = s[i] // Touch to commit.
}
}
- prev := as.mapHost(addr, hostMapEntry{
+
+ // See bluepill_allocator.go.
+ bluepill(as.pageTables.Allocator.(*allocator).cpu)
+
+ // Perform the mapping.
+ prev := as.mapLocked(addr, hostMapEntry{
addr: b.Addr(),
length: uintptr(b.Len()),
}, at)
@@ -204,17 +210,27 @@ func (as *addressSpace) MapFile(addr usermem.Addr, f platform.File, fr platform.
return nil
}
+// unmapLocked is an escape-checked wrapped around Unmap.
+//
+// +checkescape:hard,stack
+//
+//go:nosplit
+func (as *addressSpace) unmapLocked(addr usermem.Addr, length uint64) bool {
+ return as.pageTables.Unmap(addr, uintptr(length))
+}
+
// Unmap unmaps the given range by calling pagetables.PageTables.Unmap.
func (as *addressSpace) Unmap(addr usermem.Addr, length uint64) {
as.mu.Lock()
defer as.mu.Unlock()
- // See above re: retryInGuest.
- var prev bool
- as.machine.retryInGuest(func() {
- prev = as.pageTables.Unmap(addr, uintptr(length)) || prev
- })
- if prev {
+ // See above & bluepill_allocator.go.
+ as.pageTables.Allocator.(*allocator).cpu = as.machine.Get()
+ defer as.machine.Put(as.pageTables.Allocator.(*allocator).cpu)
+ bluepill(as.pageTables.Allocator.(*allocator).cpu)
+
+ if prev := as.unmapLocked(addr, length); prev {
+ // Invalidate all active vCPUs.
as.invalidate()
// Recycle any freed intermediate pages.
@@ -227,8 +243,14 @@ func (as *addressSpace) Release() {
as.Unmap(0, ^uint64(0))
// Free all pages from the allocator.
- as.pageTables.Allocator.(allocator).base.Drain()
+ as.pageTables.Allocator.(*allocator).base.Drain()
// Drop all cached machine references.
as.machine.dropPageTables(as.pageTables)
}
+
+// PreFork implements platform.AddressSpace.PreFork.
+func (as *addressSpace) PreFork() {}
+
+// PostFork implements platform.AddressSpace.PostFork.
+func (as *addressSpace) PostFork() {}
diff --git a/pkg/sentry/platform/kvm/allocator.go b/pkg/sentry/platform/kvm/bluepill_allocator.go
index 3f35414bb..9485e1301 100644
--- a/pkg/sentry/platform/kvm/allocator.go
+++ b/pkg/sentry/platform/kvm/bluepill_allocator.go
@@ -21,56 +21,80 @@ import (
)
type allocator struct {
- base *pagetables.RuntimeAllocator
+ base pagetables.RuntimeAllocator
+
+ // cpu must be set prior to any pagetable operation.
+ //
+ // Due to the way KVM's shadow paging implementation works,
+ // modifications to the page tables while in host mode may not be
+ // trapped, leading to the shadow pages being out of sync. Therefore,
+ // we need to ensure that we are in guest mode for page table
+ // modifications. See the call to bluepill, below.
+ cpu *vCPU
}
// newAllocator is used to define the allocator.
-func newAllocator() allocator {
- return allocator{
- base: pagetables.NewRuntimeAllocator(),
- }
+func newAllocator() *allocator {
+ a := new(allocator)
+ a.base.Init()
+ return a
}
// NewPTEs implements pagetables.Allocator.NewPTEs.
//
+// +checkescape:all
+//
//go:nosplit
-func (a allocator) NewPTEs() *pagetables.PTEs {
- return a.base.NewPTEs()
+func (a *allocator) NewPTEs() *pagetables.PTEs {
+ ptes := a.base.NewPTEs() // escapes: bluepill below.
+ if a.cpu != nil {
+ bluepill(a.cpu)
+ }
+ return ptes
}
// PhysicalFor returns the physical address for a set of PTEs.
//
+// +checkescape:all
+//
//go:nosplit
-func (a allocator) PhysicalFor(ptes *pagetables.PTEs) uintptr {
+func (a *allocator) PhysicalFor(ptes *pagetables.PTEs) uintptr {
virtual := a.base.PhysicalFor(ptes)
physical, _, ok := translateToPhysical(virtual)
if !ok {
- panic(fmt.Sprintf("PhysicalFor failed for %p", ptes))
+ panic(fmt.Sprintf("PhysicalFor failed for %p", ptes)) // escapes: panic.
}
return physical
}
// LookupPTEs implements pagetables.Allocator.LookupPTEs.
//
+// +checkescape:all
+//
//go:nosplit
-func (a allocator) LookupPTEs(physical uintptr) *pagetables.PTEs {
+func (a *allocator) LookupPTEs(physical uintptr) *pagetables.PTEs {
virtualStart, physicalStart, _, ok := calculateBluepillFault(physical, physicalRegions)
if !ok {
- panic(fmt.Sprintf("LookupPTEs failed for 0x%x", physical))
+ panic(fmt.Sprintf("LookupPTEs failed for 0x%x", physical)) // escapes: panic.
}
return a.base.LookupPTEs(virtualStart + (physical - physicalStart))
}
// FreePTEs implements pagetables.Allocator.FreePTEs.
//
+// +checkescape:all
+//
//go:nosplit
-func (a allocator) FreePTEs(ptes *pagetables.PTEs) {
- a.base.FreePTEs(ptes)
+func (a *allocator) FreePTEs(ptes *pagetables.PTEs) {
+ a.base.FreePTEs(ptes) // escapes: bluepill below.
+ if a.cpu != nil {
+ bluepill(a.cpu)
+ }
}
// Recycle implements pagetables.Allocator.Recycle.
//
//go:nosplit
-func (a allocator) Recycle() {
+func (a *allocator) Recycle() {
a.base.Recycle()
}
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64.go b/pkg/sentry/platform/kvm/bluepill_amd64.go
index 133c2203d..ddc1554d5 100644
--- a/pkg/sentry/platform/kvm/bluepill_amd64.go
+++ b/pkg/sentry/platform/kvm/bluepill_amd64.go
@@ -63,6 +63,8 @@ func bluepillArchEnter(context *arch.SignalContext64) *vCPU {
// KernelSyscall handles kernel syscalls.
//
+// +checkescape:all
+//
//go:nosplit
func (c *vCPU) KernelSyscall() {
regs := c.Registers()
@@ -72,13 +74,15 @@ func (c *vCPU) KernelSyscall() {
// We only trigger a bluepill entry in the bluepill function, and can
// therefore be guaranteed that there is no floating point state to be
// loaded on resuming from halt. We only worry about saving on exit.
- ring0.SaveFloatingPoint((*byte)(c.floatingPointState))
+ ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) // escapes: no.
ring0.Halt()
- ring0.WriteFS(uintptr(regs.Fs_base)) // Reload host segment.
+ ring0.WriteFS(uintptr(regs.Fs_base)) // escapes: no, reload host segment.
}
// KernelException handles kernel exceptions.
//
+// +checkescape:all
+//
//go:nosplit
func (c *vCPU) KernelException(vector ring0.Vector) {
regs := c.Registers()
@@ -89,9 +93,9 @@ func (c *vCPU) KernelException(vector ring0.Vector) {
regs.Rip = 0
}
// See above.
- ring0.SaveFloatingPoint((*byte)(c.floatingPointState))
+ ring0.SaveFloatingPoint((*byte)(c.floatingPointState)) // escapes: no.
ring0.Halt()
- ring0.WriteFS(uintptr(regs.Fs_base)) // Reload host segment.
+ ring0.WriteFS(uintptr(regs.Fs_base)) // escapes: no; reload host segment.
}
// bluepillArchExit is called during bluepillEnter.
diff --git a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
index 99cac665d..03a98512e 100644
--- a/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
+++ b/pkg/sentry/platform/kvm/bluepill_amd64_unsafe.go
@@ -17,6 +17,7 @@
package kvm
import (
+ "syscall"
"unsafe"
"gvisor.dev/gvisor/pkg/sentry/arch"
@@ -53,3 +54,34 @@ func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) {
context.Rbx = uint64(uintptr(unsafe.Pointer(c)))
context.Rip = uint64(dieTrampolineAddr)
}
+
+// getHypercallID returns hypercall ID.
+//
+//go:nosplit
+func getHypercallID(addr uintptr) int {
+ return _KVM_HYPERCALL_MAX
+}
+
+// bluepillStopGuest is reponsible for injecting interrupt.
+//
+//go:nosplit
+func bluepillStopGuest(c *vCPU) {
+ // Interrupt: we must have requested an interrupt
+ // window; set the interrupt line.
+ if _, _, errno := syscall.RawSyscall(
+ syscall.SYS_IOCTL,
+ uintptr(c.fd),
+ _KVM_INTERRUPT,
+ uintptr(unsafe.Pointer(&bounce))); errno != 0 {
+ throw("interrupt injection failed")
+ }
+ // Clear previous injection request.
+ c.runData.requestInterruptWindow = 0
+}
+
+// bluepillReadyStopGuest checks whether the current vCPU is ready for interrupt injection.
+//
+//go:nosplit
+func bluepillReadyStopGuest(c *vCPU) bool {
+ return c.runData.readyForInterruptInjection != 0
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_arm64.go b/pkg/sentry/platform/kvm/bluepill_arm64.go
index c215d443c..ed5ae03d3 100644
--- a/pkg/sentry/platform/kvm/bluepill_arm64.go
+++ b/pkg/sentry/platform/kvm/bluepill_arm64.go
@@ -26,6 +26,17 @@ import (
var (
// The action for bluepillSignal is changed by sigaction().
bluepillSignal = syscall.SIGILL
+
+ // vcpuSErr is the event of system error.
+ vcpuSErr = kvmVcpuEvents{
+ exception: exception{
+ sErrPending: 1,
+ sErrHasEsr: 0,
+ pad: [6]uint8{0, 0, 0, 0, 0, 0},
+ sErrEsr: 1,
+ },
+ rsvd: [12]uint32{0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0},
+ }
)
// bluepillArchEnter is called during bluepillEnter.
@@ -38,7 +49,7 @@ func bluepillArchEnter(context *arch.SignalContext64) (c *vCPU) {
regs.Sp = context.Sp
regs.Pc = context.Pc
regs.Pstate = context.Pstate
- regs.Pstate &^= uint64(ring0.KernelFlagsClear)
+ regs.Pstate &^= uint64(ring0.PsrFlagsClear)
regs.Pstate |= ring0.KernelFlagsSet
return
}
@@ -52,7 +63,7 @@ func bluepillArchExit(c *vCPU, context *arch.SignalContext64) {
context.Sp = regs.Sp
context.Pc = regs.Pc
context.Pstate = regs.Pstate
- context.Pstate &^= uint64(ring0.UserFlagsClear)
+ context.Pstate &^= uint64(ring0.PsrFlagsClear)
context.Pstate |= ring0.UserFlagsSet
lazyVfp := c.GetLazyVFP()
@@ -66,6 +77,8 @@ func bluepillArchExit(c *vCPU, context *arch.SignalContext64) {
// KernelSyscall handles kernel syscalls.
//
+// +checkescape:all
+//
//go:nosplit
func (c *vCPU) KernelSyscall() {
regs := c.Registers()
@@ -88,6 +101,8 @@ func (c *vCPU) KernelSyscall() {
// KernelException handles kernel exceptions.
//
+// +checkescape:all
+//
//go:nosplit
func (c *vCPU) KernelException(vector ring0.Vector) {
regs := c.Registers()
diff --git a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
index 4ca2b7717..b35c930e2 100644
--- a/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
+++ b/pkg/sentry/platform/kvm/bluepill_arm64_unsafe.go
@@ -17,6 +17,7 @@
package kvm
import (
+ "syscall"
"unsafe"
"gvisor.dev/gvisor/pkg/sentry/arch"
@@ -40,7 +41,7 @@ func fpsimdPtr(addr *byte) *arch.FpsimdContext {
func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) {
// If the vCPU is in user mode, we set the stack to the stored stack
// value in the vCPU itself. We don't want to unwind the user stack.
- if guestRegs.Regs.Pstate&ring0.PSR_MODE_MASK == ring0.PSR_MODE_EL0t {
+ if guestRegs.Regs.Pstate&ring0.PsrModeMask == ring0.UserFlagsSet {
regs := c.CPU.Registers()
context.Regs[0] = regs.Regs[0]
context.Sp = regs.Sp
@@ -61,3 +62,36 @@ func dieArchSetup(c *vCPU, context *arch.SignalContext64, guestRegs *userRegs) {
func bluepillArchFpContext(context unsafe.Pointer) *arch.FpsimdContext {
return &((*arch.SignalContext64)(context).Fpsimd64)
}
+
+// getHypercallID returns hypercall ID.
+//
+// On Arm64, the MMIO address should be 64-bit aligned.
+//
+//go:nosplit
+func getHypercallID(addr uintptr) int {
+ if addr < arm64HypercallMMIOBase || addr >= (arm64HypercallMMIOBase+_AARCH64_HYPERCALL_MMIO_SIZE) {
+ return _KVM_HYPERCALL_MAX
+ } else {
+ return int(((addr) - arm64HypercallMMIOBase) >> 3)
+ }
+}
+
+// bluepillStopGuest is reponsible for injecting sError.
+//
+//go:nosplit
+func bluepillStopGuest(c *vCPU) {
+ if _, _, errno := syscall.RawSyscall(
+ syscall.SYS_IOCTL,
+ uintptr(c.fd),
+ _KVM_SET_VCPU_EVENTS,
+ uintptr(unsafe.Pointer(&vcpuSErr))); errno != 0 {
+ throw("sErr injection failed")
+ }
+}
+
+// bluepillReadyStopGuest checks whether the current vCPU is ready for sError injection.
+//
+//go:nosplit
+func bluepillReadyStopGuest(c *vCPU) bool {
+ return true
+}
diff --git a/pkg/sentry/platform/kvm/bluepill_unsafe.go b/pkg/sentry/platform/kvm/bluepill_unsafe.go
index 9add7c944..bf357de1a 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 !go1.15
+// +build !go1.16
// Check go:linkname function signatures when updating Go version.
@@ -58,12 +58,32 @@ func bluepillArchContext(context unsafe.Pointer) *arch.SignalContext64 {
return &((*arch.UContext64)(context).MContext)
}
+// bluepillHandleHlt is reponsible for handling VM-Exit.
+//
+//go:nosplit
+func bluepillGuestExit(c *vCPU, context unsafe.Pointer) {
+ // Copy out registers.
+ bluepillArchExit(c, bluepillArchContext(context))
+
+ // Return to the vCPUReady state; notify any waiters.
+ user := atomic.LoadUint32(&c.state) & vCPUUser
+ switch atomic.SwapUint32(&c.state, user) {
+ case user | vCPUGuest: // Expected case.
+ case user | vCPUGuest | vCPUWaiter:
+ c.notify()
+ default:
+ throw("invalid state")
+ }
+}
+
// bluepillHandler is called from the signal stub.
//
// The world may be stopped while this is executing, and it executes on the
// signal stack. It should only execute raw system calls and functions that are
// explicitly marked go:nosplit.
//
+// +checkescape:all
+//
//go:nosplit
func bluepillHandler(context unsafe.Pointer) {
// Sanitize the registers; interrupts must always be disabled.
@@ -82,7 +102,8 @@ func bluepillHandler(context unsafe.Pointer) {
}
for {
- switch _, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(c.fd), _KVM_RUN, 0); errno {
+ _, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(c.fd), _KVM_RUN, 0) // escapes: no.
+ switch errno {
case 0: // Expected case.
case syscall.EINTR:
// First, we process whatever pending signal
@@ -90,7 +111,7 @@ func bluepillHandler(context unsafe.Pointer) {
// currently, all signals are masked and the signal
// must have been delivered directly to this thread.
timeout := syscall.Timespec{}
- sig, _, errno := syscall.RawSyscall6(
+ sig, _, errno := syscall.RawSyscall6( // escapes: no.
syscall.SYS_RT_SIGTIMEDWAIT,
uintptr(unsafe.Pointer(&bounceSignalMask)),
0, // siginfo.
@@ -112,12 +133,12 @@ func bluepillHandler(context unsafe.Pointer) {
// PIC, we can't inject an interrupt while they are
// masked. We need to request a window if it's not
// ready.
- if c.runData.readyForInterruptInjection == 0 {
- c.runData.requestInterruptWindow = 1
- continue // Rerun vCPU.
- } else {
+ if bluepillReadyStopGuest(c) {
// Force injection below; the vCPU is ready.
c.runData.exitReason = _KVM_EXIT_IRQ_WINDOW_OPEN
+ } else {
+ c.runData.requestInterruptWindow = 1
+ continue // Rerun vCPU.
}
case syscall.EFAULT:
// If a fault is not serviceable due to the host
@@ -125,7 +146,7 @@ func bluepillHandler(context unsafe.Pointer) {
// MMIO exit we receive EFAULT from the run ioctl. We
// always inject an NMI here since we may be in kernel
// mode and have interrupts disabled.
- if _, _, errno := syscall.RawSyscall(
+ if _, _, errno := syscall.RawSyscall( // escapes: no.
syscall.SYS_IOCTL,
uintptr(c.fd),
_KVM_NMI, 0); errno != 0 {
@@ -156,25 +177,20 @@ func bluepillHandler(context unsafe.Pointer) {
c.die(bluepillArchContext(context), "debug")
return
case _KVM_EXIT_HLT:
- // Copy out registers.
- bluepillArchExit(c, bluepillArchContext(context))
-
- // Return to the vCPUReady state; notify any waiters.
- user := atomic.LoadUint32(&c.state) & vCPUUser
- switch atomic.SwapUint32(&c.state, user) {
- case user | vCPUGuest: // Expected case.
- case user | vCPUGuest | vCPUWaiter:
- c.notify()
- default:
- throw("invalid state")
- }
+ bluepillGuestExit(c, context)
return
case _KVM_EXIT_MMIO:
+ physical := uintptr(c.runData.data[0])
+ if getHypercallID(physical) == _KVM_HYPERCALL_VMEXIT {
+ bluepillGuestExit(c, context)
+ return
+ }
+
// Increment the fault count.
atomic.AddUint32(&c.faults, 1)
// For MMIO, the physical address is the first data item.
- physical := uintptr(c.runData.data[0])
+ physical = uintptr(c.runData.data[0])
virtual, ok := handleBluepillFault(c.machine, physical, physicalRegions, _KVM_MEM_FLAGS_NONE)
if !ok {
c.die(bluepillArchContext(context), "invalid physical address")
@@ -201,17 +217,7 @@ func bluepillHandler(context unsafe.Pointer) {
}
}
case _KVM_EXIT_IRQ_WINDOW_OPEN:
- // Interrupt: we must have requested an interrupt
- // window; set the interrupt line.
- if _, _, errno := syscall.RawSyscall(
- syscall.SYS_IOCTL,
- uintptr(c.fd),
- _KVM_INTERRUPT,
- uintptr(unsafe.Pointer(&bounce))); errno != 0 {
- throw("interrupt injection failed")
- }
- // Clear previous injection request.
- c.runData.requestInterruptWindow = 0
+ bluepillStopGuest(c)
case _KVM_EXIT_SHUTDOWN:
c.die(bluepillArchContext(context), "shutdown")
return
diff --git a/pkg/sentry/platform/kvm/context.go b/pkg/sentry/platform/kvm/context.go
index 6507121ea..6e6b76416 100644
--- a/pkg/sentry/platform/kvm/context.go
+++ b/pkg/sentry/platform/kvm/context.go
@@ -15,6 +15,7 @@
package kvm
import (
+ pkgcontext "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/sentry/arch"
"gvisor.dev/gvisor/pkg/sentry/platform"
"gvisor.dev/gvisor/pkg/sentry/platform/interrupt"
@@ -37,7 +38,8 @@ type context struct {
}
// Switch runs the provided context in the given address space.
-func (c *context) Switch(as platform.AddressSpace, ac arch.Context, _ int32) (*arch.SignalInfo, usermem.AccessType, error) {
+func (c *context) Switch(ctx pkgcontext.Context, mm platform.MemoryManager, ac arch.Context, _ int32) (*arch.SignalInfo, usermem.AccessType, error) {
+ as := mm.AddressSpace()
localAS := as.(*addressSpace)
// Grab a vCPU.
@@ -88,3 +90,9 @@ func (c *context) Interrupt() {
// Release implements platform.Context.Release().
func (c *context) Release() {}
+
+// FullStateChanged implements platform.Context.FullStateChanged.
+func (c *context) FullStateChanged() {}
+
+// PullFullState implements platform.Context.PullFullState.
+func (c *context) PullFullState(as platform.AddressSpace, ac arch.Context) {}
diff --git a/pkg/sentry/platform/kvm/kvm_amd64_test.go b/pkg/sentry/platform/kvm/kvm_amd64_test.go
new file mode 100644
index 000000000..c0b4fd374
--- /dev/null
+++ b/pkg/sentry/platform/kvm/kvm_amd64_test.go
@@ -0,0 +1,51 @@
+// 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 amd64
+
+package kvm
+
+import (
+ "testing"
+
+ "gvisor.dev/gvisor/pkg/sentry/arch"
+ "gvisor.dev/gvisor/pkg/sentry/platform"
+ "gvisor.dev/gvisor/pkg/sentry/platform/kvm/testutil"
+ "gvisor.dev/gvisor/pkg/sentry/platform/ring0"
+ "gvisor.dev/gvisor/pkg/sentry/platform/ring0/pagetables"
+)
+
+func TestSegments(t *testing.T) {
+ applicationTest(t, true, testutil.TwiddleSegments, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool {
+ testutil.SetTestSegments(regs)
+ for {
+ var si arch.SignalInfo
+ if _, err := c.SwitchToUser(ring0.SwitchOpts{
+ Registers: regs,
+ FloatingPointState: dummyFPState,
+ PageTables: pt,
+ FullRestore: true,
+ }, &si); err == platform.ErrContextInterrupt {
+ continue // Retry.
+ } else if err != nil {
+ t.Errorf("application segment check with full restore got unexpected error: %v", err)
+ }
+ if err := testutil.CheckTestSegments(regs); err != nil {
+ t.Errorf("application segment check with full restore failed: %v", err)
+ }
+ break // Done.
+ }
+ return false
+ })
+}
diff --git a/pkg/sentry/platform/kvm/kvm_arm64.go b/pkg/sentry/platform/kvm/kvm_arm64.go
index 29d457a7e..0b06a923a 100644
--- a/pkg/sentry/platform/kvm/kvm_arm64.go
+++ b/pkg/sentry/platform/kvm/kvm_arm64.go
@@ -26,6 +26,9 @@ type kvmOneReg struct {
addr uint64
}
+// arm64HypercallMMIOBase is MMIO base address used to dispatch hypercalls.
+var arm64HypercallMMIOBase uintptr
+
const KVM_NR_SPSR = 5
type userFpsimdState struct {
@@ -43,6 +46,18 @@ type userRegs struct {
fpRegs userFpsimdState
}
+type exception struct {
+ sErrPending uint8
+ sErrHasEsr uint8
+ pad [6]uint8
+ sErrEsr uint64
+}
+
+type kvmVcpuEvents struct {
+ exception
+ rsvd [12]uint32
+}
+
// updateGlobalOnce does global initialization. It has to be called only once.
func updateGlobalOnce(fd int) error {
physicalInit()
diff --git a/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go b/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go
index 6531bae1d..48ccf8474 100644
--- a/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go
+++ b/pkg/sentry/platform/kvm/kvm_arm64_unsafe.go
@@ -22,7 +22,8 @@ import (
)
var (
- runDataSize int
+ runDataSize int
+ hasGuestPCID bool
)
func updateSystemValues(fd int) error {
@@ -33,6 +34,7 @@ func updateSystemValues(fd int) error {
}
// Save the data.
runDataSize = int(sz)
+ hasGuestPCID = true
// Success.
return nil
diff --git a/pkg/sentry/platform/kvm/kvm_const.go b/pkg/sentry/platform/kvm/kvm_const.go
index 1d5c77ff4..3bf918446 100644
--- a/pkg/sentry/platform/kvm/kvm_const.go
+++ b/pkg/sentry/platform/kvm/kvm_const.go
@@ -35,6 +35,8 @@ const (
_KVM_GET_SUPPORTED_CPUID = 0xc008ae05
_KVM_SET_CPUID2 = 0x4008ae90
_KVM_SET_SIGNAL_MASK = 0x4004ae8b
+ _KVM_GET_VCPU_EVENTS = 0x8040ae9f
+ _KVM_SET_VCPU_EVENTS = 0x4040aea0
)
// KVM exit reasons.
@@ -54,8 +56,10 @@ const (
// KVM capability options.
const (
- _KVM_CAP_MAX_VCPUS = 0x42
- _KVM_CAP_ARM_VM_IPA_SIZE = 0xa5
+ _KVM_CAP_MAX_VCPUS = 0x42
+ _KVM_CAP_ARM_VM_IPA_SIZE = 0xa5
+ _KVM_CAP_VCPU_EVENTS = 0x29
+ _KVM_CAP_ARM_INJECT_SERROR_ESR = 0x9e
)
// KVM limits.
@@ -71,3 +75,13 @@ const (
_KVM_MEM_READONLY = uint32(1) << 1
_KVM_MEM_FLAGS_NONE = 0
)
+
+// 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 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 531ae8b1e..9a7be3655 100644
--- a/pkg/sentry/platform/kvm/kvm_const_arm64.go
+++ b/pkg/sentry/platform/kvm/kvm_const_arm64.go
@@ -72,6 +72,7 @@ const (
_TCR_T0SZ_VA48 = 64 - 48 // VA=48
_TCR_T1SZ_VA48 = 64 - 48 // VA=48
+ _TCR_A1 = 1 << 22
_TCR_ASID16 = 1 << 36
_TCR_TBI0 = 1 << 37
@@ -116,6 +117,17 @@ const (
// Arm64: Exception Syndrome Register EL1.
const (
+ _ESR_ELx_EC_SHIFT = 26
+ _ESR_ELx_EC_MASK = 0x3F << _ESR_ELx_EC_SHIFT
+
+ _ESR_ELx_EC_IMP_DEF = 0x1f
+ _ESR_ELx_EC_IABT_LOW = 0x20
+ _ESR_ELx_EC_IABT_CUR = 0x21
+ _ESR_ELx_EC_PC_ALIGN = 0x22
+
+ _ESR_ELx_CM = 1 << 8
+ _ESR_ELx_WNR = 1 << 6
+
_ESR_ELx_FSC = 0x3F
_ESR_SEGV_MAPERR_L0 = 0x4
@@ -131,3 +143,10 @@ const (
_ESR_SEGV_PEMERR_L2 = 0xe
_ESR_SEGV_PEMERR_L3 = 0xf
)
+
+// Arm64: MMIO base address used to dispatch hypercalls.
+const (
+ // on Arm64, the MMIO address must be 64-bit aligned.
+ // Currently, we only need 1 hypercall: hypercall_vmexit.
+ _AARCH64_HYPERCALL_MMIO_SIZE = 1 << 3
+)
diff --git a/pkg/sentry/platform/kvm/kvm_test.go b/pkg/sentry/platform/kvm/kvm_test.go
index 6c8f4fa28..45b3180f1 100644
--- a/pkg/sentry/platform/kvm/kvm_test.go
+++ b/pkg/sentry/platform/kvm/kvm_test.go
@@ -262,30 +262,6 @@ func TestRegistersFault(t *testing.T) {
})
}
-func TestSegments(t *testing.T) {
- applicationTest(t, true, testutil.TwiddleSegments, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool {
- testutil.SetTestSegments(regs)
- for {
- var si arch.SignalInfo
- if _, err := c.SwitchToUser(ring0.SwitchOpts{
- Registers: regs,
- FloatingPointState: dummyFPState,
- PageTables: pt,
- FullRestore: true,
- }, &si); err == platform.ErrContextInterrupt {
- continue // Retry.
- } else if err != nil {
- t.Errorf("application segment check with full restore got unexpected error: %v", err)
- }
- if err := testutil.CheckTestSegments(regs); err != nil {
- t.Errorf("application segment check with full restore failed: %v", err)
- }
- break // Done.
- }
- return false
- })
-}
-
func TestBounce(t *testing.T) {
applicationTest(t, true, testutil.SpinLoop, func(c *vCPU, regs *arch.Registers, pt *pagetables.PageTables) bool {
go func() {
diff --git a/pkg/sentry/platform/kvm/machine.go b/pkg/sentry/platform/kvm/machine.go
index f1afc74dc..6c54712d1 100644
--- a/pkg/sentry/platform/kvm/machine.go
+++ b/pkg/sentry/platform/kvm/machine.go
@@ -52,16 +52,19 @@ type machine struct {
// available is notified when vCPUs are available.
available sync.Cond
- // vCPUs are the machine vCPUs.
+ // vCPUsByTID are the machine vCPUs.
//
// These are populated dynamically.
- vCPUs map[uint64]*vCPU
+ vCPUsByTID map[uint64]*vCPU
// vCPUsByID are the machine vCPUs, can be indexed by the vCPU's ID.
- vCPUsByID map[int]*vCPU
+ vCPUsByID []*vCPU
// maxVCPUs is the maximum number of vCPUs supported by the machine.
maxVCPUs int
+
+ // nextID is the next vCPU ID.
+ nextID uint32
}
const (
@@ -137,9 +140,8 @@ type dieState struct {
//
// Precondition: mu must be held.
func (m *machine) newVCPU() *vCPU {
- id := len(m.vCPUs)
-
// Create the vCPU.
+ id := int(atomic.AddUint32(&m.nextID, 1) - 1)
fd, _, errno := syscall.RawSyscall(syscall.SYS_IOCTL, uintptr(m.fd), _KVM_CREATE_VCPU, uintptr(id))
if errno != 0 {
panic(fmt.Sprintf("error creating new vCPU: %v", errno))
@@ -176,11 +178,7 @@ func (m *machine) newVCPU() *vCPU {
// newMachine returns a new VM context.
func newMachine(vm int) (*machine, error) {
// Create the machine.
- m := &machine{
- fd: vm,
- vCPUs: make(map[uint64]*vCPU),
- vCPUsByID: make(map[int]*vCPU),
- }
+ m := &machine{fd: vm}
m.available.L = &m.mu
m.kernel.Init(ring0.KernelOpts{
PageTables: pagetables.New(newAllocator()),
@@ -194,6 +192,10 @@ func newMachine(vm int) (*machine, error) {
}
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)
+
// Apply the physical mappings. Note that these mappings may point to
// guest physical addresses that are not actually available. These
// physical pages are mapped on demand, see kernel_unsafe.go.
@@ -274,6 +276,8 @@ func newMachine(vm int) (*machine, error) {
// not available. This attempts to be efficient for calls in the hot path.
//
// This panics on error.
+//
+//go:nosplit
func (m *machine) mapPhysical(physical, length uintptr, phyRegions []physicalRegion, flags uint32) {
for end := physical + length; physical < end; {
_, physicalStart, length, ok := calculateBluepillFault(physical, phyRegions)
@@ -304,7 +308,11 @@ func (m *machine) Destroy() {
runtime.SetFinalizer(m, nil)
// Destroy vCPUs.
- for _, c := range m.vCPUs {
+ for _, c := range m.vCPUsByID {
+ if c == nil {
+ continue
+ }
+
// Ensure the vCPU is not still running in guest mode. This is
// possible iff teardown has been done by other threads, and
// somehow a single thread has not executed any system calls.
@@ -337,7 +345,7 @@ func (m *machine) Get() *vCPU {
tid := procid.Current()
// Check for an exact match.
- if c := m.vCPUs[tid]; c != nil {
+ if c := m.vCPUsByTID[tid]; c != nil {
c.lock()
m.mu.RUnlock()
return c
@@ -356,7 +364,7 @@ func (m *machine) Get() *vCPU {
tid = procid.Current()
// Recheck for an exact match.
- if c := m.vCPUs[tid]; c != nil {
+ if c := m.vCPUsByTID[tid]; c != nil {
c.lock()
m.mu.Unlock()
return c
@@ -364,10 +372,10 @@ func (m *machine) Get() *vCPU {
for {
// Scan for an available vCPU.
- for origTID, c := range m.vCPUs {
+ for origTID, c := range m.vCPUsByTID {
if atomic.CompareAndSwapUint32(&c.state, vCPUReady, vCPUUser) {
- delete(m.vCPUs, origTID)
- m.vCPUs[tid] = c
+ delete(m.vCPUsByTID, origTID)
+ m.vCPUsByTID[tid] = c
m.mu.Unlock()
c.loadSegments(tid)
return c
@@ -375,17 +383,17 @@ func (m *machine) Get() *vCPU {
}
// Create a new vCPU (maybe).
- if len(m.vCPUs) < m.maxVCPUs {
+ if int(m.nextID) < m.maxVCPUs {
c := m.newVCPU()
c.lock()
- m.vCPUs[tid] = c
+ m.vCPUsByTID[tid] = c
m.mu.Unlock()
c.loadSegments(tid)
return c
}
// Scan for something not in user mode.
- for origTID, c := range m.vCPUs {
+ for origTID, c := range m.vCPUsByTID {
if !atomic.CompareAndSwapUint32(&c.state, vCPUGuest, vCPUGuest|vCPUWaiter) {
continue
}
@@ -403,8 +411,8 @@ func (m *machine) Get() *vCPU {
}
// Steal the vCPU.
- delete(m.vCPUs, origTID)
- m.vCPUs[tid] = c
+ delete(m.vCPUsByTID, origTID)
+ m.vCPUsByTID[tid] = c
m.mu.Unlock()
c.loadSegments(tid)
return c
@@ -431,7 +439,7 @@ func (m *machine) Put(c *vCPU) {
// newDirtySet returns a new dirty set.
func (m *machine) newDirtySet() *dirtySet {
return &dirtySet{
- vCPUs: make([]uint64, (m.maxVCPUs+63)/64, (m.maxVCPUs+63)/64),
+ vCPUMasks: make([]uint64, (m.maxVCPUs+63)/64, (m.maxVCPUs+63)/64),
}
}
diff --git a/pkg/sentry/platform/kvm/machine_amd64.go b/pkg/sentry/platform/kvm/machine_amd64.go
index 923ce3909..acc823ba6 100644
--- a/pkg/sentry/platform/kvm/machine_amd64.go
+++ b/pkg/sentry/platform/kvm/machine_amd64.go
@@ -51,9 +51,10 @@ func (m *machine) initArchState() error {
recover()
debug.SetPanicOnFault(old)
}()
- m.retryInGuest(func() {
- ring0.SetCPUIDFaulting(true)
- })
+ c := m.Get()
+ defer m.Put(c)
+ bluepill(c)
+ ring0.SetCPUIDFaulting(true)
return nil
}
@@ -89,8 +90,8 @@ func (m *machine) dropPageTables(pt *pagetables.PageTables) {
defer m.mu.Unlock()
// Clear from all PCIDs.
- for _, c := range m.vCPUs {
- if c.PCIDs != nil {
+ for _, c := range m.vCPUsByID {
+ if c != nil && c.PCIDs != nil {
c.PCIDs.Drop(pt)
}
}
@@ -335,29 +336,6 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo)
}
}
-// retryInGuest runs the given function in guest mode.
-//
-// If the function does not complete in guest mode (due to execution of a
-// system call due to a GC stall, for example), then it will be retried. The
-// given function must be idempotent as a result of the retry mechanism.
-func (m *machine) retryInGuest(fn func()) {
- c := m.Get()
- defer m.Put(c)
- for {
- c.ClearErrorCode() // See below.
- bluepill(c) // Force guest mode.
- fn() // Execute the given function.
- _, user := c.ErrorCode()
- if user {
- // If user is set, then we haven't bailed back to host
- // mode via a kernel exception or system call. We
- // consider the full function to have executed in guest
- // mode and we can return.
- break
- }
- }
-}
-
// On x86 platform, the flags for "setMemoryRegion" can always be set as 0.
// There is no need to return read-only physicalRegions.
func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) {
diff --git a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go
index 7156c245f..290f035dd 100644
--- a/pkg/sentry/platform/kvm/machine_amd64_unsafe.go
+++ b/pkg/sentry/platform/kvm/machine_amd64_unsafe.go
@@ -154,7 +154,7 @@ func (c *vCPU) setUserRegisters(uregs *userRegs) error {
//
//go:nosplit
func (c *vCPU) getUserRegisters(uregs *userRegs) syscall.Errno {
- if _, _, errno := syscall.RawSyscall(
+ if _, _, errno := syscall.RawSyscall( // escapes: no.
syscall.SYS_IOCTL,
uintptr(c.fd),
_KVM_GET_REGS,
diff --git a/pkg/sentry/platform/kvm/machine_arm64.go b/pkg/sentry/platform/kvm/machine_arm64.go
index e42505542..9db171af9 100644
--- a/pkg/sentry/platform/kvm/machine_arm64.go
+++ b/pkg/sentry/platform/kvm/machine_arm64.go
@@ -60,6 +60,12 @@ func rdonlyRegionsForSetMem() (phyRegions []physicalRegion) {
if !vr.accessType.Write && vr.accessType.Read {
rdonlyRegions = append(rdonlyRegions, vr.region)
}
+
+ // TODO(gvisor.dev/issue/2686): PROT_NONE should be specially treated.
+ // Workaround: treated as rdonly temporarily.
+ if !vr.accessType.Write && !vr.accessType.Read && !vr.accessType.Execute {
+ rdonlyRegions = append(rdonlyRegions, vr.region)
+ }
})
for _, r := range rdonlyRegions {
@@ -100,7 +106,7 @@ func (m *machine) dropPageTables(pt *pagetables.PageTables) {
defer m.mu.Unlock()
// Clear from all PCIDs.
- for _, c := range m.vCPUs {
+ for _, c := range m.vCPUsByID {
if c.PCIDs != nil {
c.PCIDs.Drop(pt)
}
@@ -119,71 +125,59 @@ func nonCanonical(addr uint64, signal int32, info *arch.SignalInfo) (usermem.Acc
return usermem.NoAccess, platform.ErrContextSignal
}
+// isInstructionAbort returns true if it is an instruction abort.
+//
+//go:nosplit
+func isInstructionAbort(code uint64) bool {
+ value := (code & _ESR_ELx_EC_MASK) >> _ESR_ELx_EC_SHIFT
+ return value == _ESR_ELx_EC_IABT_LOW
+}
+
+// isWriteFault returns whether it is a write fault.
+//
+//go:nosplit
+func isWriteFault(code uint64) bool {
+ if isInstructionAbort(code) {
+ return false
+ }
+
+ return (code & _ESR_ELx_WNR) != 0
+}
+
// fault generates an appropriate fault return.
//
//go:nosplit
func (c *vCPU) fault(signal int32, info *arch.SignalInfo) (usermem.AccessType, error) {
+ bluepill(c) // Probably no-op, but may not be.
faultAddr := c.GetFaultAddr()
code, user := c.ErrorCode()
+ if !user {
+ // The last fault serviced by this CPU was not a user
+ // fault, so we can't reliably trust the faultAddr or
+ // the code provided here. We need to re-execute.
+ return usermem.NoAccess, platform.ErrContextInterrupt
+ }
+
// Reset the pointed SignalInfo.
*info = arch.SignalInfo{Signo: signal}
info.SetAddr(uint64(faultAddr))
- read := true
- write := false
- execute := true
-
ret := code & _ESR_ELx_FSC
switch ret {
case _ESR_SEGV_MAPERR_L0, _ESR_SEGV_MAPERR_L1, _ESR_SEGV_MAPERR_L2, _ESR_SEGV_MAPERR_L3:
info.Code = 1 //SEGV_MAPERR
- read = false
- write = true
- execute = false
case _ESR_SEGV_ACCERR_L1, _ESR_SEGV_ACCERR_L2, _ESR_SEGV_ACCERR_L3, _ESR_SEGV_PEMERR_L1, _ESR_SEGV_PEMERR_L2, _ESR_SEGV_PEMERR_L3:
info.Code = 2 // SEGV_ACCERR.
- read = true
- write = false
- execute = false
default:
info.Code = 2
}
- if !user {
- read = true
- write = false
- execute = true
-
- }
accessType := usermem.AccessType{
- Read: read,
- Write: write,
- Execute: execute,
+ Read: !isWriteFault(uint64(code)),
+ Write: isWriteFault(uint64(code)),
+ Execute: isInstructionAbort(uint64(code)),
}
return accessType, platform.ErrContextSignal
}
-
-// retryInGuest runs the given function in guest mode.
-//
-// If the function does not complete in guest mode (due to execution of a
-// system call due to a GC stall, for example), then it will be retried. The
-// given function must be idempotent as a result of the retry mechanism.
-func (m *machine) retryInGuest(fn func()) {
- c := m.Get()
- defer m.Put(c)
- for {
- c.ClearErrorCode() // See below.
- bluepill(c) // Force guest mode.
- fn() // Execute the given function.
- _, user := c.ErrorCode()
- if user {
- // If user is set, then we haven't bailed back to host
- // mode via a kernel exception or system call. We
- // consider the full function to have executed in guest
- // mode and we can return.
- break
- }
- }
-}
diff --git a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
index 3c02cef7c..905712076 100644
--- a/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
+++ b/pkg/sentry/platform/kvm/machine_arm64_unsafe.go
@@ -26,6 +26,7 @@ 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"
)
@@ -60,7 +61,6 @@ func (c *vCPU) initArchState() error {
reg.addr = uint64(reflect.ValueOf(&data).Pointer())
regGet.addr = uint64(reflect.ValueOf(&dataGet).Pointer())
- vcpuInit.target = _KVM_ARM_TARGET_GENERIC_V8
vcpuInit.features[0] |= (1 << _KVM_ARM_VCPU_PSCI_0_2)
if _, _, errno := syscall.RawSyscall(
syscall.SYS_IOCTL,
@@ -78,19 +78,6 @@ func (c *vCPU) initArchState() error {
return err
}
- // sctlr_el1
- regGet.id = _KVM_ARM64_REGS_SCTLR_EL1
- if err := c.getOneRegister(&regGet); err != nil {
- return err
- }
-
- dataGet |= (_SCTLR_M | _SCTLR_C | _SCTLR_I)
- data = dataGet
- reg.id = _KVM_ARM64_REGS_SCTLR_EL1
- if err := c.setOneRegister(&reg); err != nil {
- return err
- }
-
// tcr_el1
data = _TCR_TXSZ_VA48 | _TCR_CACHE_FLAGS | _TCR_SHARED | _TCR_TG_FLAGS | _TCR_ASID16 | _TCR_IPS_40BITS
reg.id = _KVM_ARM64_REGS_TCR_EL1
@@ -159,10 +146,16 @@ func (c *vCPU) initArchState() error {
return err
}
- data = ring0.PsrDefaultSet | ring0.KernelFlagsSet
- reg.id = _KVM_ARM64_REGS_PSTATE
- if err := c.setOneRegister(&reg); err != nil {
- return err
+ // Use the address of the exception vector table as
+ // the MMIO address base.
+ arm64HypercallMMIOBase = toLocation
+
+ // Initialize the PCID database.
+ if hasGuestPCID {
+ // Note that NewPCIDs may return a nil table here, in which
+ // case we simply don't use PCID support (see below). In
+ // practice, this should not happen, however.
+ c.PCIDs = pagetables.NewPCIDs(fixedKernelPCID+1, poolPCIDs)
}
c.floatingPointState = arch.NewFloatingPointData()
@@ -243,6 +236,13 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo)
return nonCanonical(regs.Sp, int32(syscall.SIGBUS), info)
}
+ // Assign PCIDs.
+ if c.PCIDs != nil {
+ var requireFlushPCID bool // Force a flush?
+ switchOpts.UserASID, requireFlushPCID = c.PCIDs.Assign(switchOpts.PageTables)
+ switchOpts.Flush = switchOpts.Flush || requireFlushPCID
+ }
+
var vector ring0.Vector
ttbr0App := switchOpts.PageTables.TTBR0_EL1(false, 0)
c.SetTtbr0App(uintptr(ttbr0App))
@@ -269,10 +269,18 @@ func (c *vCPU) SwitchToUser(switchOpts ring0.SwitchOpts, info *arch.SignalInfo)
case ring0.PageFault:
return c.fault(int32(syscall.SIGSEGV), info)
- case 0xaa:
- return usermem.NoAccess, nil
+ case ring0.Vector(bounce): // ring0.VirtualizationException
+ return usermem.NoAccess, platform.ErrContextInterrupt
+ case ring0.El0Sync_undef,
+ ring0.El1Sync_undef:
+ *info = arch.SignalInfo{
+ Signo: int32(syscall.SIGILL),
+ Code: 1, // ILL_ILLOPC (illegal opcode).
+ }
+ info.SetAddr(switchOpts.Registers.Pc) // Include address.
+ return usermem.AccessType{}, platform.ErrContextSignal
default:
- return usermem.NoAccess, platform.ErrContextSignal
+ panic(fmt.Sprintf("unexpected vector: 0x%x", vector))
}
}
diff --git a/pkg/sentry/platform/kvm/machine_unsafe.go b/pkg/sentry/platform/kvm/machine_unsafe.go
index f04be2ab5..9f86f6a7a 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 !go1.15
+// +build !go1.16
// Check go:linkname function signatures when updating Go version.
@@ -115,7 +115,7 @@ func (a *atomicAddressSpace) get() *addressSpace {
//
//go:nosplit
func (c *vCPU) notify() {
- _, _, errno := syscall.RawSyscall6(
+ _, _, errno := syscall.RawSyscall6( // escapes: no.
syscall.SYS_FUTEX,
uintptr(unsafe.Pointer(&c.state)),
linux.FUTEX_WAKE|linux.FUTEX_PRIVATE_FLAG,
diff --git a/pkg/sentry/platform/kvm/testutil/testutil_arm64.go b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go
index ca902c8c1..4dad877ba 100644
--- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.go
+++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.go
@@ -56,5 +56,9 @@ func CheckTestRegs(regs *arch.Registers, full bool) (err error) {
err = addRegisterMismatch(err, fmt.Sprintf("R%d", i), regs.Regs[i], need)
}
}
+ // Check tls.
+ if need := ^uint64(11); regs.TPIDR_EL0 != need {
+ err = addRegisterMismatch(err, "tpdir_el0", regs.TPIDR_EL0, need)
+ }
return
}
diff --git a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s
index 0bebee852..6caf7282d 100644
--- a/pkg/sentry/platform/kvm/testutil/testutil_arm64.s
+++ b/pkg/sentry/platform/kvm/testutil/testutil_arm64.s
@@ -52,6 +52,8 @@ start:
TEXT ·FloatingPointWorks(SB),NOSPLIT,$0-8
NO_LOCAL_POINTERS
+ // gc will touch fpsimd, so we should test it.
+ // such as in <runtime.deductSweepCredit>.
FMOVD $(9.9), F0
MOVD $SYS_GETPID, R8 // getpid
SVC
@@ -102,5 +104,15 @@ isNaN:
TEXT ·TwiddleRegsSyscall(SB),NOSPLIT,$0
TWIDDLE_REGS()
+ MSR R10, TPIDR_EL0
+ // Trapped in el0_svc.
SVC
RET // never reached
+
+TEXT ·TwiddleRegsFault(SB),NOSPLIT,$0
+ TWIDDLE_REGS()
+ MSR R10, TPIDR_EL0
+ // Trapped in el0_ia.
+ // Branch to Register branches unconditionally to an address in <Rn>.
+ JMP (R6) // <=> br x6, must fault
+ RET // never reached
diff --git a/pkg/sentry/platform/platform.go b/pkg/sentry/platform/platform.go
index 171513f3f..ba031516a 100644
--- a/pkg/sentry/platform/platform.go
+++ b/pkg/sentry/platform/platform.go
@@ -22,9 +22,10 @@ import (
"os"
"gvisor.dev/gvisor/pkg/abi/linux"
- "gvisor.dev/gvisor/pkg/safemem"
+ "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/seccomp"
"gvisor.dev/gvisor/pkg/sentry/arch"
+ "gvisor.dev/gvisor/pkg/sentry/memmap"
"gvisor.dev/gvisor/pkg/usermem"
)
@@ -114,6 +115,17 @@ func (NoCPUPreemptionDetection) PreemptAllCPUs() error {
panic("This platform does not support CPU preemption detection")
}
+// MemoryManager represents an abstraction above the platform address space
+// which manages memory mappings and their contents.
+type MemoryManager interface {
+ //usermem.IO provides access to the contents of a virtual memory space.
+ usermem.IO
+ // MMap establishes a memory mapping.
+ MMap(ctx context.Context, opts memmap.MMapOpts) (usermem.Addr, error)
+ // AddressSpace returns the AddressSpace bound to mm.
+ AddressSpace() AddressSpace
+}
+
// Context represents the execution context for a single thread.
type Context interface {
// Switch resumes execution of the thread specified by the arch.Context
@@ -143,7 +155,36 @@ type Context interface {
// concurrent call to Switch().
//
// - ErrContextCPUPreempted: See the definition of that error for details.
- Switch(as AddressSpace, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error)
+ Switch(ctx context.Context, mm MemoryManager, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error)
+
+ // PullFullState() pulls a full state of the application thread.
+ //
+ // A platform can support lazy loading/restoring of a thread state
+ // which includes registers and a floating point state.
+ //
+ // For example, when the Sentry handles a system call, it may have only
+ // syscall arguments without other registers and a floating point
+ // state. And in this case, if the Sentry will need to construct a
+ // signal frame to call a signal handler, it will need to call
+ // PullFullState() to load all registers and FPU state.
+ //
+ // Preconditions: The caller must be running on the task goroutine.
+ PullFullState(as AddressSpace, ac arch.Context)
+
+ // FullStateChanged() indicates that a thread state has been changed by
+ // the Sentry. This happens in case of the rt_sigreturn, execve, etc.
+ //
+ // First, it indicates that the Sentry has the full state of the thread
+ // and PullFullState() has to do nothing if it is called after
+ // FullStateChanged().
+ //
+ // Second, it forces restoring the full state of the application
+ // thread. A platform can support lazy loading/restoring of a thread
+ // state. This means that if the Sentry has not changed a thread state,
+ // the platform may not restore it.
+ //
+ // Preconditions: The caller must be running on the task goroutine.
+ FullStateChanged()
// Interrupt interrupts a concurrent call to Switch(), causing it to return
// ErrContextInterrupt.
@@ -207,7 +248,7 @@ type AddressSpace interface {
// 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 File, fr FileRange, at usermem.AccessType, precommit bool) error
+ MapFile(addr usermem.Addr, f memmap.File, fr memmap.FileRange, at usermem.AccessType, precommit bool) error
// Unmap unmaps the given range.
//
@@ -218,6 +259,13 @@ type AddressSpace interface {
// must be acquired via platform.NewAddressSpace().
Release()
+ // PreFork() is called before creating a copy of AddressSpace. This
+ // guarantees that this address space will be in a consistent state.
+ PreFork()
+
+ // PostFork() is called after creating a copy of AddressSpace.
+ PostFork()
+
// AddressSpaceIO methods are supported iff the associated platform's
// Platform.SupportsAddressSpaceIO() == true. AddressSpaces for which this
// does not hold may panic if AddressSpaceIO methods are invoked.
@@ -310,52 +358,6 @@ func (f SegmentationFault) Error() string {
return fmt.Sprintf("segmentation fault at %#x", f.Addr)
}
-// File represents a host file that may be mapped into an AddressSpace.
-type File interface {
- // All pages in a File are reference-counted.
-
- // 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.)
- 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.
- DecRef(fr FileRange)
-
- // MapInternal returns a mapping of the given file offsets in the invoking
- // process' address space for reading and writing.
- //
- // 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.
- //
- // Postconditions: The returned mapping is valid as long as at least one
- // reference is held on the mapped pages.
- MapInternal(fr FileRange, at usermem.AccessType) (safemem.BlockSeq, error)
-
- // FD returns the file descriptor represented by the File.
- //
- // The only permitted operation on the returned file descriptor is to map
- // pages from it consistent with the requirements of AddressSpace.MapFile.
- FD() int
-}
-
-// FileRange represents a range of uint64 offsets into a File.
-//
-// type FileRange <generated using go_generics>
-
-// String implements fmt.Stringer.String.
-func (fr FileRange) String() string {
- return fmt.Sprintf("[%#x, %#x)", fr.Start, fr.End)
-}
-
// Requirements is used to specify platform specific requirements.
type Requirements struct {
// RequiresCurrentPIDNS indicates that the sandbox has to be started in the
diff --git a/pkg/sentry/platform/ptrace/BUILD b/pkg/sentry/platform/ptrace/BUILD
index 30402c2df..e04165fbf 100644
--- a/pkg/sentry/platform/ptrace/BUILD
+++ b/pkg/sentry/platform/ptrace/BUILD
@@ -24,12 +24,14 @@ go_library(
visibility = ["//:sandbox"],
deps = [
"//pkg/abi/linux",
+ "//pkg/context",
"//pkg/log",
"//pkg/procid",
"//pkg/safecopy",
"//pkg/seccomp",
"//pkg/sentry/arch",
"//pkg/sentry/hostcpu",
+ "//pkg/sentry/memmap",
"//pkg/sentry/platform",
"//pkg/sentry/platform/interrupt",
"//pkg/sync",
diff --git a/pkg/sentry/platform/ptrace/ptrace.go b/pkg/sentry/platform/ptrace/ptrace.go
index 08d055e05..b52d0fbd8 100644
--- a/pkg/sentry/platform/ptrace/ptrace.go
+++ b/pkg/sentry/platform/ptrace/ptrace.go
@@ -48,6 +48,7 @@ import (
"os"
"gvisor.dev/gvisor/pkg/abi/linux"
+ pkgcontext "gvisor.dev/gvisor/pkg/context"
"gvisor.dev/gvisor/pkg/sentry/arch"
"gvisor.dev/gvisor/pkg/sentry/platform"
"gvisor.dev/gvisor/pkg/sentry/platform/interrupt"
@@ -95,7 +96,8 @@ type context struct {
}
// Switch runs the provided context in the given address space.
-func (c *context) Switch(as platform.AddressSpace, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error) {
+func (c *context) Switch(ctx pkgcontext.Context, mm platform.MemoryManager, ac arch.Context, cpu int32) (*arch.SignalInfo, usermem.AccessType, error) {
+ as := mm.AddressSpace()
s := as.(*subprocess)
isSyscall := s.switchToApp(c, ac)
@@ -180,6 +182,12 @@ func (c *context) Interrupt() {
// Release implements platform.Context.Release().
func (c *context) Release() {}
+// FullStateChanged implements platform.Context.FullStateChanged.
+func (c *context) FullStateChanged() {}
+
+// PullFullState implements platform.Context.PullFullState.
+func (c *context) PullFullState(as platform.AddressSpace, ac arch.Context) {}
+
// PTrace represents a collection of ptrace subprocesses.
type PTrace struct {
platform.MMapMinAddr
diff --git a/pkg/sentry/platform/ptrace/subprocess.go b/pkg/sentry/platform/ptrace/subprocess.go
index 2389423b0..e1d54d8a2 100644
--- a/pkg/sentry/platform/ptrace/subprocess.go
+++ b/pkg/sentry/platform/ptrace/subprocess.go
@@ -24,6 +24,7 @@ import (
"gvisor.dev/gvisor/pkg/log"
"gvisor.dev/gvisor/pkg/procid"
"gvisor.dev/gvisor/pkg/sentry/arch"
+ "gvisor.dev/gvisor/pkg/sentry/memmap"
"gvisor.dev/gvisor/pkg/sentry/platform"
"gvisor.dev/gvisor/pkg/sync"
"gvisor.dev/gvisor/pkg/usermem"
@@ -616,7 +617,7 @@ func (s *subprocess) syscall(sysno uintptr, args ...arch.SyscallArgument) (uintp
}
// MapFile implements platform.AddressSpace.MapFile.
-func (s *subprocess) MapFile(addr usermem.Addr, f platform.File, fr platform.FileRange, at usermem.AccessType, precommit bool) error {
+func (s *subprocess) MapFile(addr usermem.Addr, f memmap.File, fr memmap.FileRange, at usermem.AccessType, precommit bool) error {
var flags int
if precommit {
flags |= syscall.MAP_POPULATE
@@ -661,3 +662,9 @@ func (s *subprocess) Unmap(addr usermem.Addr, length uint64) {
panic(fmt.Sprintf("munmap(%x, %x)) failed: %v", addr, length, err))
}
}
+
+// PreFork implements platform.AddressSpace.PreFork.
+func (s *subprocess) PreFork() {}
+
+// PostFork implements platform.AddressSpace.PostFork.
+func (s *subprocess) PostFork() {}
diff --git a/pkg/sentry/platform/ptrace/subprocess_unsafe.go b/pkg/sentry/platform/ptrace/subprocess_unsafe.go
index 2ae6b9f9d..0bee995e4 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 !go1.15
+// +build !go1.16
// Check go:linkname function signatures when updating Go version.
diff --git a/pkg/sentry/platform/ring0/aarch64.go b/pkg/sentry/platform/ring0/aarch64.go
index 8122ac6e2..87a573cc4 100644
--- a/pkg/sentry/platform/ring0/aarch64.go
+++ b/pkg/sentry/platform/ring0/aarch64.go
@@ -26,30 +26,31 @@ const (
_PMD_PGT_SIZE = 0x4000
_PTE_PGT_BASE = 0x7000
_PTE_PGT_SIZE = 0x1000
-
- _PSR_D_BIT = 0x00000200
- _PSR_A_BIT = 0x00000100
- _PSR_I_BIT = 0x00000080
- _PSR_F_BIT = 0x00000040
)
const (
- // PSR bits
- PSR_MODE_EL0t = 0x00000000
- PSR_MODE_EL1t = 0x00000004
- PSR_MODE_EL1h = 0x00000005
- PSR_MODE_MASK = 0x0000000f
+ // DAIF bits:debug, sError, IRQ, FIQ.
+ _PSR_D_BIT = 0x00000200
+ _PSR_A_BIT = 0x00000100
+ _PSR_I_BIT = 0x00000080
+ _PSR_F_BIT = 0x00000040
+ _PSR_DAIF_SHIFT = 6
+ _PSR_DAIF_MASK = 0xf << _PSR_DAIF_SHIFT
- // KernelFlagsSet should always be set in the kernel.
- KernelFlagsSet = PSR_MODE_EL1h
+ // PSR bits.
+ _PSR_MODE_EL0t = 0x00000000
+ _PSR_MODE_EL1t = 0x00000004
+ _PSR_MODE_EL1h = 0x00000005
+ _PSR_MODE_MASK = 0x0000000f
- // UserFlagsSet are always set in userspace.
- UserFlagsSet = PSR_MODE_EL0t
+ PsrFlagsClear = _PSR_MODE_MASK | _PSR_DAIF_MASK
+ PsrModeMask = _PSR_MODE_MASK
- KernelFlagsClear = PSR_MODE_MASK
- UserFlagsClear = PSR_MODE_MASK
+ // KernelFlagsSet should always be set in the kernel.
+ KernelFlagsSet = _PSR_MODE_EL1h | _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT
- PsrDefaultSet = _PSR_D_BIT | _PSR_A_BIT | _PSR_I_BIT | _PSR_F_BIT
+ // UserFlagsSet are always set in userspace.
+ UserFlagsSet = _PSR_MODE_EL0t
)
// Vector is an exception vector.
diff --git a/pkg/sentry/platform/ring0/entry_arm64.s b/pkg/sentry/platform/ring0/entry_arm64.s
index db6465663..9d29b7168 100644
--- a/pkg/sentry/platform/ring0/entry_arm64.s
+++ b/pkg/sentry/platform/ring0/entry_arm64.s
@@ -40,6 +40,20 @@
#define FPEN_ENABLE (FPEN_NOTRAP << FPEN_SHIFT)
+// sctlr_el1: system control register el1.
+#define SCTLR_M 1 << 0
+#define SCTLR_C 1 << 2
+#define SCTLR_I 1 << 12
+#define SCTLR_UCT 1 << 15
+
+#define SCTLR_EL1_DEFAULT (SCTLR_M | SCTLR_C | SCTLR_I | SCTLR_UCT)
+
+// cntkctl_el1: counter-timer kernel control register el1.
+#define CNTKCTL_EL0PCTEN 1 << 0
+#define CNTKCTL_EL0VCTEN 1 << 1
+
+#define CNTKCTL_EL1_DEFAULT (CNTKCTL_EL0PCTEN | CNTKCTL_EL0VCTEN)
+
// Saves a register set.
//
// This is a macro because it may need to executed in contents where a stack is
@@ -298,12 +312,6 @@
ISB $15; \
DSB $15;
-#define IRQ_ENABLE \
- MSR $2, DAIFSet;
-
-#define IRQ_DISABLE \
- MSR $2, DAIFClr;
-
#define VFP_ENABLE \
MOVD $FPEN_ENABLE, R0; \
WORD $0xd5181040; \ //MSR R0, CPACR_EL1
@@ -356,15 +364,29 @@ TEXT ·Halt(SB),NOSPLIT,$0
CMP RSV_REG, R9
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
MOVD R1, CPU_LAZY_VFP(RSV_REG)
VFP_DISABLE
- // MMIO_EXIT.
- MOVD $0, R9
- MOVD R0, 0xffff000000001000(R9)
+ // Trigger MMIO_EXIT/_KVM_HYPERCALL_VMEXIT.
+ //
+ // To keep it simple, I used the address of exception table as the
+ // MMIO base address, so that I can trigger a MMIO-EXIT by forcibly writing
+ // a read-only space.
+ // Also, the length is engough to match a sufficient number of hypercall ID.
+ // Then, in host user space, I can calculate this address to find out
+ // which hypercall.
+ 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.
@@ -398,6 +420,7 @@ TEXT ·Current(SB),NOSPLIT,$0-8
// Prepare the vcpu environment for container application.
TEXT ·kernelExitToEl0(SB),NOSPLIT,$0
// Step1, save sentry context into memory.
+ MRS TPIDR_EL1, RSV_REG
REGISTERS_SAVE(RSV_REG, CPU_REGISTERS)
MOVD RSV_REG_APP, CPU_REGISTERS+PTRACE_R9(RSV_REG)
@@ -409,34 +432,13 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0
MOVD CPU_REGISTERS+PTRACE_R3(RSV_REG), R3
- // Step2, save SP_EL1, PSTATE into kernel temporary stack.
- // switch to temporary stack.
+ // Step2, switch to temporary stack.
LOAD_KERNEL_STACK(RSV_REG)
- WORD $0xd538d092 //MRS TPIDR_EL1, R18
-
- SUB $STACK_FRAME_SIZE, RSP, RSP
- MOVD CPU_REGISTERS+PTRACE_SP(RSV_REG), R11
- MOVD CPU_REGISTERS+PTRACE_PSTATE(RSV_REG), R12
- STP (R11, R12), 16*0(RSP)
-
- MOVD CPU_REGISTERS+PTRACE_R11(RSV_REG), R11
- MOVD CPU_REGISTERS+PTRACE_R12(RSV_REG), R12
-
- // Step3, test user pagetable.
- // If user pagetable is empty, trapped in el1_ia.
- WORD $0xd538d092 //MRS TPIDR_EL1, R18
- SWITCH_TO_APP_PAGETABLE(RSV_REG)
- WORD $0xd538d092 //MRS TPIDR_EL1, R18
- SWITCH_TO_KVM_PAGETABLE(RSV_REG)
- WORD $0xd538d092 //MRS TPIDR_EL1, R18
-
- // If pagetable is not empty, recovery kernel temporary stack.
- ADD $STACK_FRAME_SIZE, RSP, RSP
- // Step4, load app context pointer.
+ // Step3, load app context pointer.
MOVD CPU_APP_ADDR(RSV_REG), RSV_REG_APP
- // Step5, prepare the environment for container application.
+ // Step4, prepare the environment for container application.
// set sp_el0.
MOVD PTRACE_SP(RSV_REG_APP), R1
WORD $0xd5184101 //MSR R1, SP_EL0
@@ -464,13 +466,13 @@ TEXT ·kernelExitToEl0(SB),NOSPLIT,$0
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.
// Prepare the vcpu environment for sentry.
TEXT ·kernelExitToEl1(SB),NOSPLIT,$0
WORD $0xd538d092 //MRS TPIDR_EL1, R18
-
MOVD CPU_REGISTERS+PTRACE_PSTATE(RSV_REG), R1
WORD $0xd5184001 //MSR R1, SPSR_EL1
@@ -487,7 +489,15 @@ TEXT ·kernelExitToEl1(SB),NOSPLIT,$0
// Start is the CPU entrypoint.
TEXT ·Start(SB),NOSPLIT,$0
- IRQ_DISABLE
+ // Flush dcache.
+ WORD $0xd5087e52 // DC CISW
+ // Init.
+ MOVD $SCTLR_EL1_DEFAULT, R1
+ MSR R1, SCTLR_EL1
+
+ MOVD $CNTKCTL_EL1_DEFAULT, R1
+ MSR R1, CNTKCTL_EL1
+
MOVD R8, RSV_REG
ORR $0xffff000000000000, RSV_REG, RSV_REG
WORD $0xd518d092 //MSR R18, TPIDR_EL1
@@ -536,6 +546,7 @@ TEXT ·El1_sync(SB),NOSPLIT,$0
B el1_invalid
el1_da:
+el1_ia:
WORD $0xd538d092 //MRS TPIDR_EL1, R18
WORD $0xd538601a //MRS FAR_EL1, R26
@@ -548,9 +559,6 @@ el1_da:
B ·HaltAndResume(SB)
-el1_ia:
- B ·HaltAndResume(SB)
-
el1_sp_pc:
B ·Shutdown(SB)
@@ -622,9 +630,10 @@ el0_svc:
MOVD $Syscall, R3
MOVD R3, CPU_VECTOR_CODE(RSV_REG)
- B ·HaltAndResume(SB)
+ B ·kernelExitToEl1(SB)
el0_da:
+el0_ia:
WORD $0xd538d092 //MRS TPIDR_EL1, R18
WORD $0xd538601a //MRS FAR_EL1, R26
@@ -636,10 +645,10 @@ el0_da:
MOVD $PageFault, R3
MOVD R3, CPU_VECTOR_CODE(RSV_REG)
- B ·HaltAndResume(SB)
+ MRS ESR_EL1, R3
+ MOVD R3, CPU_ERROR_CODE(RSV_REG)
-el0_ia:
- B ·Shutdown(SB)
+ B ·kernelExitToEl1(SB)
el0_fpsimd_acc:
B ·Shutdown(SB)
@@ -654,7 +663,10 @@ el0_sp_pc:
B ·Shutdown(SB)
el0_undef:
- B ·Shutdown(SB)
+ MOVD $El0Sync_undef, R3
+ MOVD R3, CPU_VECTOR_CODE(RSV_REG)
+
+ B ·kernelExitToEl1(SB)
el0_dbg:
B ·Shutdown(SB)
diff --git a/pkg/sentry/platform/ring0/kernel.go b/pkg/sentry/platform/ring0/kernel.go
index 900c0bba7..021693791 100644
--- a/pkg/sentry/platform/ring0/kernel.go
+++ b/pkg/sentry/platform/ring0/kernel.go
@@ -31,23 +31,39 @@ type defaultHooks struct{}
// KernelSyscall implements Hooks.KernelSyscall.
//
+// +checkescape:all
+//
//go:nosplit
-func (defaultHooks) KernelSyscall() { Halt() }
+func (defaultHooks) KernelSyscall() {
+ Halt()
+}
// KernelException implements Hooks.KernelException.
//
+// +checkescape:all
+//
//go:nosplit
-func (defaultHooks) KernelException(Vector) { Halt() }
+func (defaultHooks) KernelException(Vector) {
+ Halt()
+}
// kernelSyscall is a trampoline.
//
+// +checkescape:hard,stack
+//
//go:nosplit
-func kernelSyscall(c *CPU) { c.hooks.KernelSyscall() }
+func kernelSyscall(c *CPU) {
+ c.hooks.KernelSyscall()
+}
// kernelException is a trampoline.
//
+// +checkescape:hard,stack
+//
//go:nosplit
-func kernelException(c *CPU, vector Vector) { c.hooks.KernelException(vector) }
+func kernelException(c *CPU, vector Vector) {
+ c.hooks.KernelException(vector)
+}
// Init initializes a new CPU.
//
diff --git a/pkg/sentry/platform/ring0/kernel_amd64.go b/pkg/sentry/platform/ring0/kernel_amd64.go
index 0feff8778..d37981dbf 100644
--- a/pkg/sentry/platform/ring0/kernel_amd64.go
+++ b/pkg/sentry/platform/ring0/kernel_amd64.go
@@ -178,6 +178,8 @@ func IsCanonical(addr uint64) bool {
//
// Precondition: the Rip, Rsp, Fs and Gs registers must be canonical.
//
+// +checkescape:all
+//
//go:nosplit
func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) {
userCR3 := switchOpts.PageTables.CR3(!switchOpts.Flush, switchOpts.UserPCID)
@@ -192,9 +194,9 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) {
// Perform the switch.
swapgs() // GS will be swapped on return.
- WriteFS(uintptr(regs.Fs_base)) // Set application FS.
- WriteGS(uintptr(regs.Gs_base)) // Set application GS.
- LoadFloatingPoint(switchOpts.FloatingPointState) // Copy in floating point.
+ 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 {
@@ -204,8 +206,8 @@ func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) {
}
writeCR3(uintptr(kernelCR3)) // Return to kernel address space.
jumpToUser() // Return to lower half.
- SaveFloatingPoint(switchOpts.FloatingPointState) // Copy out floating point.
- WriteFS(uintptr(c.registers.Fs_base)) // Restore kernel FS.
+ SaveFloatingPoint(switchOpts.FloatingPointState) // escapes: no. Copy out floating point.
+ WriteFS(uintptr(c.registers.Fs_base)) // escapes: no. Restore kernel FS.
return
}
diff --git a/pkg/sentry/platform/ring0/kernel_arm64.go b/pkg/sentry/platform/ring0/kernel_arm64.go
index ccacaea6b..d0afa1aaa 100644
--- a/pkg/sentry/platform/ring0/kernel_arm64.go
+++ b/pkg/sentry/platform/ring0/kernel_arm64.go
@@ -53,14 +53,20 @@ func IsCanonical(addr uint64) bool {
//go:nosplit
func (c *CPU) SwitchToUser(switchOpts SwitchOpts) (vector Vector) {
- // Sanitize registers.
regs := switchOpts.Registers
- regs.Pstate &= ^uint64(UserFlagsClear)
+ 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
- // Perform the switch.
return
}
diff --git a/pkg/sentry/platform/ring0/lib_arm64.go b/pkg/sentry/platform/ring0/lib_arm64.go
index a6345010d..00e52c8af 100644
--- a/pkg/sentry/platform/ring0/lib_arm64.go
+++ b/pkg/sentry/platform/ring0/lib_arm64.go
@@ -38,6 +38,12 @@ func SaveVRegs(*byte)
// LoadVRegs loads V0-V31 registers.
func LoadVRegs(*byte)
+// LoadFloatingPoint loads floating point state.
+func LoadFloatingPoint(*byte)
+
+// SaveFloatingPoint saves floating point state.
+func SaveFloatingPoint(*byte)
+
// GetTLS returns the value of TPIDR_EL0 register.
func GetTLS() (value uint64)
diff --git a/pkg/sentry/platform/ring0/lib_arm64.s b/pkg/sentry/platform/ring0/lib_arm64.s
index b63e14b41..86bfbe46f 100644
--- a/pkg/sentry/platform/ring0/lib_arm64.s
+++ b/pkg/sentry/platform/ring0/lib_arm64.s
@@ -129,3 +129,89 @@ TEXT ·LoadVRegs(SB),NOSPLIT,$0-8
ISB $15
RET
+
+TEXT ·LoadFloatingPoint(SB),NOSPLIT,$0-8
+ MOVD addr+0(FP), R0
+
+ MOVD 0(R0), R1
+ MOVD R1, FPSR
+ MOVD 8(R0), R1
+ MOVD R1, NZCV
+
+ FMOVD 16*1(R0), F0
+ FMOVD 16*2(R0), F1
+ FMOVD 16*3(R0), F2
+ FMOVD 16*4(R0), F3
+ FMOVD 16*5(R0), F4
+ FMOVD 16*6(R0), F5
+ FMOVD 16*7(R0), F6
+ FMOVD 16*8(R0), F7
+ FMOVD 16*9(R0), F8
+ FMOVD 16*10(R0), F9
+ FMOVD 16*11(R0), F10
+ FMOVD 16*12(R0), F11
+ FMOVD 16*13(R0), F12
+ FMOVD 16*14(R0), F13
+ FMOVD 16*15(R0), F14
+ FMOVD 16*16(R0), F15
+ FMOVD 16*17(R0), F16
+ FMOVD 16*18(R0), F17
+ FMOVD 16*19(R0), F18
+ FMOVD 16*20(R0), F19
+ FMOVD 16*21(R0), F20
+ FMOVD 16*22(R0), F21
+ FMOVD 16*23(R0), F22
+ FMOVD 16*24(R0), F23
+ FMOVD 16*25(R0), F24
+ FMOVD 16*26(R0), F25
+ FMOVD 16*27(R0), F26
+ FMOVD 16*28(R0), F27
+ FMOVD 16*29(R0), F28
+ FMOVD 16*30(R0), F29
+ FMOVD 16*31(R0), F30
+ FMOVD 16*32(R0), F31
+
+ RET
+
+TEXT ·SaveFloatingPoint(SB),NOSPLIT,$0-8
+ MOVD addr+0(FP), R0
+
+ MOVD FPSR, R1
+ MOVD R1, 0(R0)
+ MOVD NZCV, R1
+ MOVD R1, 8(R0)
+
+ FMOVD F0, 16*1(R0)
+ FMOVD F1, 16*2(R0)
+ FMOVD F2, 16*3(R0)
+ FMOVD F3, 16*4(R0)
+ FMOVD F4, 16*5(R0)
+ FMOVD F5, 16*6(R0)
+ FMOVD F6, 16*7(R0)
+ FMOVD F7, 16*8(R0)
+ FMOVD F8, 16*9(R0)
+ FMOVD F9, 16*10(R0)
+ FMOVD F10, 16*11(R0)
+ FMOVD F11, 16*12(R0)
+ FMOVD F12, 16*13(R0)
+ FMOVD F13, 16*14(R0)
+ FMOVD F14, 16*15(R0)
+ FMOVD F15, 16*16(R0)
+ FMOVD F16, 16*17(R0)
+ FMOVD F17, 16*18(R0)
+ FMOVD F18, 16*19(R0)
+ FMOVD F19, 16*20(R0)
+ FMOVD F20, 16*21(R0)
+ FMOVD F21, 16*22(R0)
+ FMOVD F22, 16*23(R0)
+ FMOVD F23, 16*24(R0)
+ FMOVD F24, 16*25(R0)
+ FMOVD F25, 16*26(R0)
+ FMOVD F26, 16*27(R0)
+ FMOVD F27, 16*28(R0)
+ FMOVD F28, 16*29(R0)
+ FMOVD F29, 16*30(R0)
+ FMOVD F30, 16*31(R0)
+ FMOVD F31, 16*32(R0)
+
+ RET
diff --git a/pkg/sentry/platform/ring0/pagetables/allocator.go b/pkg/sentry/platform/ring0/pagetables/allocator.go
index 23fd5c352..8d75b7599 100644
--- a/pkg/sentry/platform/ring0/pagetables/allocator.go
+++ b/pkg/sentry/platform/ring0/pagetables/allocator.go
@@ -53,9 +53,14 @@ type RuntimeAllocator struct {
// NewRuntimeAllocator returns an allocator that uses runtime allocation.
func NewRuntimeAllocator() *RuntimeAllocator {
- return &RuntimeAllocator{
- used: make(map[*PTEs]struct{}),
- }
+ r := new(RuntimeAllocator)
+ r.Init()
+ return r
+}
+
+// Init initializes a RuntimeAllocator.
+func (r *RuntimeAllocator) Init() {
+ r.used = make(map[*PTEs]struct{})
}
// Recycle returns freed pages to the pool.
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables.go b/pkg/sentry/platform/ring0/pagetables/pagetables.go
index 87e88e97d..7f18ac296 100644
--- a/pkg/sentry/platform/ring0/pagetables/pagetables.go
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables.go
@@ -86,6 +86,8 @@ func (*mapVisitor) requiresSplit() bool { return true }
//
// Precondition: addr & length must be page-aligned, their sum must not overflow.
//
+// +checkescape:hard,stack
+//
//go:nosplit
func (p *PageTables) Map(addr usermem.Addr, length uintptr, opts MapOpts, physical uintptr) bool {
if !opts.AccessType.Any() {
@@ -128,6 +130,8 @@ func (v *unmapVisitor) visit(start uintptr, pte *PTE, align uintptr) {
//
// Precondition: addr & length must be page-aligned.
//
+// +checkescape:hard,stack
+//
//go:nosplit
func (p *PageTables) Unmap(addr usermem.Addr, length uintptr) bool {
w := unmapWalker{
@@ -162,6 +166,8 @@ func (v *emptyVisitor) visit(start uintptr, pte *PTE, align uintptr) {
//
// Precondition: addr & length must be page-aligned.
//
+// +checkescape:hard,stack
+//
//go:nosplit
func (p *PageTables) IsEmpty(addr usermem.Addr, length uintptr) bool {
w := emptyWalker{
@@ -197,6 +203,8 @@ func (*lookupVisitor) requiresSplit() bool { return false }
// Lookup returns the physical address for the given virtual address.
//
+// +checkescape:hard,stack
+//
//go:nosplit
func (p *PageTables) Lookup(addr usermem.Addr) (physical uintptr, opts MapOpts) {
mask := uintptr(usermem.PageSize - 1)
diff --git a/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go b/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go
index 78510ebed..6409d1d91 100644
--- a/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go
+++ b/pkg/sentry/platform/ring0/pagetables/pagetables_aarch64.go
@@ -72,13 +72,14 @@ const (
)
const (
- mtNormal = 0x4 << 2
+ mtDevicenGnRE = 0x1 << 2
+ mtNormal = 0x4 << 2
)
const (
executeDisable = xn
optionMask = 0xfff | 0xfff<<48
- protDefault = accessed | shared | mtNormal
+ protDefault = accessed | shared
)
// MapOpts are x86 options.
@@ -184,8 +185,10 @@ func (p *PTE) Set(addr uintptr, opts MapOpts) {
if opts.User {
v |= user
+ v |= mtNormal
} else {
v = v &^ user
+ v |= mtDevicenGnRE // Strong order for the addresses with ring0.KernelStartAddress.
}
atomic.StoreUintptr((*uintptr)(p), v)
}
@@ -200,7 +203,7 @@ func (p *PTE) setPageTable(pt *PageTables, ptes *PTEs) {
// This should never happen.
panic("unaligned physical address!")
}
- v := addr | typeTable | protDefault
+ v := addr | typeTable | protDefault | mtNormal
atomic.StoreUintptr((*uintptr)(p), v)
}