Add callbacks to support lazy loading/restoring thread states

PiperOrigin-RevId: 324748508

pkg/sentry/kernel/kernel.go

@@ -1263,6 +1263,13 @@ func (k *Kernel) Pause() {
 	k.tasks.aioGoroutines.Wait()
 }
 
+// ReceiveTaskStates receives full states for all tasks.
+func (k *Kernel) ReceiveTaskStates() {
+	k.extMu.Lock()
+	k.tasks.PullFullState()
+	k.extMu.Unlock()
+}
+
 // Unpause ends the effect of a previous call to Pause. If Unpause is called
 // without a matching preceding call to Pause, Unpause may panic.
 func (k *Kernel) Unpause() {

pkg/sentry/kernel/ptrace.go

@@ -1018,6 +1018,9 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error {
 		if err != nil {
 			return err
 		}
+
+		t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch())
+
 		ar := ars.Head()
 		n, err := target.Arch().PtraceGetRegSet(uintptr(addr), &usermem.IOReadWriter{
 			Ctx:  t,
@@ -1044,10 +1047,14 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error {
 		if err != nil {
 			return err
 		}
+
+		mm := t.MemoryManager()
+		t.p.PullFullState(mm.AddressSpace(), t.Arch())
+
 		ar := ars.Head()
 		n, err := target.Arch().PtraceSetRegSet(uintptr(addr), &usermem.IOReadWriter{
 			Ctx:  t,
-			IO:   t.MemoryManager(),
+			IO:   mm,
 			Addr: ar.Start,
 			Opts: usermem.IOOpts{
 				AddressSpaceActive: true,
@@ -1056,6 +1063,7 @@ func (t *Task) Ptrace(req int64, pid ThreadID, addr, data usermem.Addr) error {
 		if err != nil {
 			return err
 		}
+		t.p.FloatingPointStateChanged()
 		ar.End -= usermem.Addr(n)
 		return t.CopyOutIovecs(data, usermem.AddrRangeSeqOf(ar))
 

pkg/sentry/kernel/task_clone.go

@@ -161,6 +161,10 @@ func (t *Task) Clone(opts *CloneOptions) (ThreadID, *SyscallControl, error) {
 		return 0, nil, syserror.EINVAL
 	}
 
+	// Pull task registers and FPU state, a cloned task will inherit the
+	// state of the current task.
+	t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch())
+
 	// "If CLONE_NEWUSER is specified along with other CLONE_NEW* flags in a
 	// single clone(2) or unshare(2) call, the user namespace is guaranteed to
 	// be created first, giving the child (clone(2)) or caller (unshare(2))

pkg/sentry/kernel/task_log.go

@@ -27,6 +27,9 @@ const (
 	// maxStackDebugBytes is the maximum number of user stack bytes that may be
 	// printed by debugDumpStack.
 	maxStackDebugBytes = 1024
+	// maxCodeDebugBytes is the maximum number of user code bytes that may be
+	// printed by debugDumpCode.
+	maxCodeDebugBytes = 128
 )
 
 // Infof logs an formatted info message by calling log.Infof.
@@ -61,6 +64,7 @@ func (t *Task) IsLogging(level log.Level) bool {
 func (t *Task) DebugDumpState() {
 	t.debugDumpRegisters()
 	t.debugDumpStack()
+	t.debugDumpCode()
 	if mm := t.MemoryManager(); mm != nil {
 		t.Debugf("Mappings:\n%s", mm)
 	}
@@ -128,6 +132,45 @@ func (t *Task) debugDumpStack() {
 	}
 }
 
+// debugDumpCode logs user code contents at log level debug.
+//
+// Preconditions: The caller must be running on the task goroutine.
+func (t *Task) debugDumpCode() {
+	if !t.IsLogging(log.Debug) {
+		return
+	}
+	m := t.MemoryManager()
+	if m == nil {
+		t.Debugf("Memory manager for task is gone, skipping application code dump.")
+		return
+	}
+	t.Debugf("Code:")
+	// Print code on both sides of the instruction register.
+	start := usermem.Addr(t.Arch().IP()) - maxCodeDebugBytes/2
+	// Round addr down to a 16-byte boundary.
+	start &= ^usermem.Addr(15)
+	// Print 16 bytes per line, one byte at a time.
+	for offset := uint64(0); offset < maxCodeDebugBytes; offset += 16 {
+		addr, ok := start.AddLength(offset)
+		if !ok {
+			break
+		}
+		var data [16]byte
+		n, err := m.CopyIn(t, addr, data[:], usermem.IOOpts{
+			IgnorePermissions: true,
+		})
+		// Print as much of the line as we can, even if an error was
+		// encountered.
+		if n > 0 {
+			t.Debugf("%x: % x", addr, data[:n])
+		}
+		if err != nil {
+			t.Debugf("Error reading stack at address %x: %v", addr+usermem.Addr(n), err)
+			break
+		}
+	}
+}
+
 // trace definitions.
 //
 // Note that all region names are prefixed by ':' in order to ensure that they

pkg/sentry/kernel/task_run.go

@@ -260,7 +260,7 @@ func (app *runApp) execute(t *Task) taskRunState {
 
 	region := trace.StartRegion(t.traceContext, runRegion)
 	t.accountTaskGoroutineEnter(TaskGoroutineRunningApp)
-	info, at, err := t.p.Switch(t.MemoryManager().AddressSpace(), t.Arch(), t.rseqCPU)
+	info, at, err := t.p.Switch(t, t.MemoryManager(), t.Arch(), t.rseqCPU)
 	t.accountTaskGoroutineLeave(TaskGoroutineRunningApp)
 	region.End()
 

pkg/sentry/kernel/task_signals.go

@@ -255,10 +255,11 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct)
 		}
 	}
 
+	mm := t.MemoryManager()
 	// Set up the signal handler. If we have a saved signal mask, the signal
 	// handler should run with the current mask, but sigreturn should restore
 	// the saved one.
-	st := &arch.Stack{t.Arch(), t.MemoryManager(), sp}
+	st := &arch.Stack{t.Arch(), mm, sp}
 	mask := t.signalMask
 	if t.haveSavedSignalMask {
 		mask = t.savedSignalMask
@@ -273,12 +274,13 @@ func (t *Task) deliverSignalToHandler(info *arch.SignalInfo, act arch.SignalAct)
 	// Please see the linux code as reference:
 	// linux/arch/arm64/kernel/signal.c:setup_return()
 	if act.Flags&linux.SA_RESTORER == 0 {
-		act.Restorer = t.MemoryManager().VDSOSigReturn()
+		act.Restorer = mm.VDSOSigReturn()
 	}
 
 	if err := t.Arch().SignalSetup(st, &act, info, &alt, mask); err != nil {
 		return err
 	}
+	t.p.FloatingPointStateChanged()
 	t.haveSavedSignalMask = false
 
 	// Add our signal mask.
@@ -310,6 +312,7 @@ func (t *Task) SignalReturn(rt bool) (*SyscallControl, error) {
 
 	// Restore our signal mask. SIGKILL and SIGSTOP should not be blocked.
 	t.SetSignalMask(sigset &^ UnblockableSignals)
+	t.p.FloatingPointStateChanged()
 
 	return ctrlResume, nil
 }
@@ -636,6 +639,7 @@ func (t *Task) SetSavedSignalMask(mask linux.SignalSet) {
 
 // SignalStack returns the task-private signal stack.
 func (t *Task) SignalStack() arch.SignalStack {
+	t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch())
 	alt := t.signalStack
 	if t.onSignalStack(alt) {
 		alt.Flags |= arch.SignalStackFlagOnStack
@@ -1050,6 +1054,8 @@ func (*runInterrupt) execute(t *Task) taskRunState {
 
 	// Are there signals pending?
 	if info := t.dequeueSignalLocked(t.signalMask); info != nil {
+		t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch())
+
 		if linux.SignalSetOf(linux.Signal(info.Signo))&StopSignals != 0 {
 			// Indicate that we've dequeued a stop signal before unlocking the
 			// signal mutex; initiateGroupStop will check for races with

pkg/sentry/kernel/task_stop.go

@@ -205,6 +205,22 @@ func (ts *TaskSet) BeginExternalStop() {
 	}
 }
 
+// PullFullState receives full states for all tasks.
+func (ts *TaskSet) PullFullState() {
+	ts.mu.Lock()
+	defer ts.mu.Unlock()
+	if ts.Root == nil {
+		return
+	}
+	for t := range ts.Root.tids {
+		t.Activate()
+		if mm := t.MemoryManager(); mm != nil {
+			t.p.PullFullState(t.MemoryManager().AddressSpace(), t.Arch())
+		}
+		t.Deactivate()
+	}
+}
+
 // EndExternalStop indicates the end of an external stop started by a previous
 // call to TaskSet.BeginExternalStop. EndExternalStop does not wait for task
 // goroutines to resume.

pkg/sentry/memmap/memmap.go

@@ -360,6 +360,13 @@ type MMapOpts struct {
 	//
 	// TODO(jamieliu): Replace entirely with MappingIdentity?
 	Hint string
+
+	// Force means to skip validation checks of Addr and Length. It can be
+	// used to create special mappings below mm.layout.MinAddr and
+	// mm.layout.MaxAddr. It has to be used with caution.
+	//
+	// If Force is true, Unmap and Fixed must be true.
+	Force bool
 }
 
 // File represents a host file that may be mapped into an platform.AddressSpace.

pkg/sentry/mm/lifecycle.go

@@ -57,6 +57,8 @@ func (mm *MemoryManager) SetMmapLayout(ac arch.Context, r *limits.LimitSet) (arc
 // Fork creates a copy of mm with 1 user, as for Linux syscalls fork() or
 // clone() (without CLONE_VM).
 func (mm *MemoryManager) Fork(ctx context.Context) (*MemoryManager, error) {
+	mm.AddressSpace().PreFork()
+	defer mm.AddressSpace().PostFork()
 	mm.metadataMu.Lock()
 	defer mm.metadataMu.Unlock()
 	mm.mappingMu.RLock()

pkg/sentry/mm/vma.go

@@ -42,7 +42,12 @@ func (mm *MemoryManager) createVMALocked(ctx context.Context, opts memmap.MMapOp
 		Map32Bit: opts.Map32Bit,
 	})
 	if err != nil {
-		return vmaIterator{}, usermem.AddrRange{}, err
+		// Can't force without opts.Unmap and opts.Fixed.
+		if opts.Force && opts.Unmap && opts.Fixed {
+			addr = opts.Addr
+		} else {
+			return vmaIterator{}, usermem.AddrRange{}, err
+		}
 	}
 	ar, _ := addr.ToRange(opts.Length)
 

pkg/sentry/platform/kvm/BUILD

@@ -41,6 +41,7 @@ go_library(
     deps = [
         "//pkg/abi/linux",
         "//pkg/atomicbitops",
+        "//pkg/context",
         "//pkg/cpuid",
         "//pkg/log",
         "//pkg/procid",

pkg/sentry/platform/kvm/address_space.go

@@ -248,3 +248,9 @@ func (as *addressSpace) Release() {
 	// 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() {}

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() {}
+
+// FloatingPointStateChanged implements platform.Context.FloatingPointStateChanged.
+func (c *context) FloatingPointStateChanged() {}
+
+// PullFullState implements platform.Context.PullFullState.
+func (c *context) PullFullState(as platform.AddressSpace, ac arch.Context) {}

pkg/sentry/platform/platform.go

@@ -22,6 +22,7 @@ import (
 	"os"
 
 	"gvisor.dev/gvisor/pkg/abi/linux"
+	"gvisor.dev/gvisor/pkg/context"
 	"gvisor.dev/gvisor/pkg/seccomp"
 	"gvisor.dev/gvisor/pkg/sentry/arch"
 	"gvisor.dev/gvisor/pkg/sentry/memmap"
@@ -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 +155,30 @@ 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)
+
+	// FloatingPointStateChanged forces restoring a 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.
+	FloatingPointStateChanged()
 
 	// Interrupt interrupts a concurrent call to Switch(), causing it to return
 	// ErrContextInterrupt.
@@ -218,6 +253,13 @@ 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.

pkg/sentry/platform/ptrace/BUILD

@@ -24,6 +24,7 @@ go_library(
     visibility = ["//:sandbox"],
     deps = [
         "//pkg/abi/linux",
+        "//pkg/context",
         "//pkg/log",
         "//pkg/procid",
         "//pkg/safecopy",

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() {}
 
+// FloatingPointStateChanged implements platform.Context.FloatingPointStateChanged.
+func (c *context) FloatingPointStateChanged() {}
+
+// 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

pkg/sentry/platform/ptrace/subprocess.go

@@ -662,3 +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() {}

pkg/sentry/state/state.go

@@ -60,6 +60,7 @@ type SaveOpts struct {
 func (opts SaveOpts) Save(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.")