Merge release-20200921.0-102-g4f462b0ed (automated)

6 files changed, 172 insertions, 42 deletions

diff --git a/pkg/sentry/fsimpl/signalfd/signalfd.go b/pkg/sentry/fsimpl/signalfd/signalfd.go
index bf11b425a..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"
@@ -95,8 +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,
@@ -105,9 +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.
diff --git a/pkg/sentry/kernel/kernel_abi_autogen_unsafe.go b/pkg/sentry/kernel/kernel_abi_autogen_unsafe.go
index eaa2ebdb3..9ac2a552a 100644
--- a/pkg/sentry/kernel/kernel_abi_autogen_unsafe.go
+++ b/pkg/sentry/kernel/kernel_abi_autogen_unsafe.go
@@ -15,6 +15,7 @@ import (
 
 // Marshallable types used by this file.
 var _ marshal.Marshallable = (*ThreadID)(nil)
+var _ marshal.Marshallable = (*vdsoParams)(nil)
 
 // SizeBytes implements marshal.Marshallable.SizeBytes.
 //go:nosplit
@@ -105,3 +106,120 @@ func (t *ThreadID) WriteTo(w io.Writer) (int64, error) {
     return int64(length), err
 }
 
+// SizeBytes implements marshal.Marshallable.SizeBytes.
+func (v *vdsoParams) SizeBytes() int {
+    return 64
+}
+
+// MarshalBytes implements marshal.Marshallable.MarshalBytes.
+func (v *vdsoParams) MarshalBytes(dst []byte) {
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.monotonicReady))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.monotonicBaseCycles))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.monotonicBaseRef))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.monotonicFrequency))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.realtimeReady))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.realtimeBaseCycles))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.realtimeBaseRef))
+    dst = dst[8:]
+    usermem.ByteOrder.PutUint64(dst[:8], uint64(v.realtimeFrequency))
+    dst = dst[8:]
+}
+
+// UnmarshalBytes implements marshal.Marshallable.UnmarshalBytes.
+func (v *vdsoParams) UnmarshalBytes(src []byte) {
+    v.monotonicReady = uint64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.monotonicBaseCycles = int64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.monotonicBaseRef = int64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.monotonicFrequency = uint64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.realtimeReady = uint64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.realtimeBaseCycles = int64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.realtimeBaseRef = int64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+    v.realtimeFrequency = uint64(usermem.ByteOrder.Uint64(src[:8]))
+    src = src[8:]
+}
+
+// Packed implements marshal.Marshallable.Packed.
+//go:nosplit
+func (v *vdsoParams) Packed() bool {
+    return true
+}
+
+// MarshalUnsafe implements marshal.Marshallable.MarshalUnsafe.
+func (v *vdsoParams) MarshalUnsafe(dst []byte) {
+    safecopy.CopyIn(dst, unsafe.Pointer(v))
+}
+
+// UnmarshalUnsafe implements marshal.Marshallable.UnmarshalUnsafe.
+func (v *vdsoParams) UnmarshalUnsafe(src []byte) {
+    safecopy.CopyOut(unsafe.Pointer(v), src)
+}
+
+// CopyOutN implements marshal.Marshallable.CopyOutN.
+//go:nosplit
+func (v *vdsoParams) CopyOutN(cc marshal.CopyContext, addr usermem.Addr, limit int) (int, error) {
+    // Construct a slice backed by dst's underlying memory.
+    var buf []byte
+    hdr := (*reflect.SliceHeader)(unsafe.Pointer(&buf))
+    hdr.Data = uintptr(gohacks.Noescape(unsafe.Pointer(v)))
+    hdr.Len = v.SizeBytes()
+    hdr.Cap = v.SizeBytes()
+
+    length, err := cc.CopyOutBytes(addr, buf[:limit]) // escapes: okay.
+    // Since we bypassed the compiler's escape analysis, indicate that v
+    // must live until the use above.
+    runtime.KeepAlive(v) // escapes: replaced by intrinsic.
+    return length, err
+}
+
+// CopyOut implements marshal.Marshallable.CopyOut.
+//go:nosplit
+func (v *vdsoParams) CopyOut(cc marshal.CopyContext, addr usermem.Addr) (int, error) {
+    return v.CopyOutN(cc, addr, v.SizeBytes())
+}
+
+// CopyIn implements marshal.Marshallable.CopyIn.
+//go:nosplit
+func (v *vdsoParams) CopyIn(cc marshal.CopyContext, addr usermem.Addr) (int, error) {
+    // Construct a slice backed by dst's underlying memory.
+    var buf []byte
+    hdr := (*reflect.SliceHeader)(unsafe.Pointer(&buf))
+    hdr.Data = uintptr(gohacks.Noescape(unsafe.Pointer(v)))
+    hdr.Len = v.SizeBytes()
+    hdr.Cap = v.SizeBytes()
+
+    length, err := cc.CopyInBytes(addr, buf) // escapes: okay.
+    // Since we bypassed the compiler's escape analysis, indicate that v
+    // must live until the use above.
+    runtime.KeepAlive(v) // escapes: replaced by intrinsic.
+    return length, err
+}
+
+// WriteTo implements io.WriterTo.WriteTo.
+func (v *vdsoParams) WriteTo(writer io.Writer) (int64, error) {
+    // Construct a slice backed by dst's underlying memory.
+    var buf []byte
+    hdr := (*reflect.SliceHeader)(unsafe.Pointer(&buf))
+    hdr.Data = uintptr(gohacks.Noescape(unsafe.Pointer(v)))
+    hdr.Len = v.SizeBytes()
+    hdr.Cap = v.SizeBytes()
+
+    length, err := writer.Write(buf)
+    // Since we bypassed the compiler's escape analysis, indicate that v
+    // must live until the use above.
+    runtime.KeepAlive(v) // escapes: replaced by intrinsic.
+    return int64(length), err
+}
+
diff --git a/pkg/sentry/kernel/kernel_state_autogen.go b/pkg/sentry/kernel/kernel_state_autogen.go
index f20800960..7ff5e1eb9 100644
--- a/pkg/sentry/kernel/kernel_state_autogen.go
+++ b/pkg/sentry/kernel/kernel_state_autogen.go
@@ -2259,6 +2259,7 @@ func (x *VDSOParamPage) StateFields() []string {
 		"mfp",
 		"fr",
 		"seq",
+		"copyScratchBuffer",
 	}
 }
 
@@ -2269,6 +2270,7 @@ func (x *VDSOParamPage) StateSave(m state.Sink) {
 	m.Save(0, &x.mfp)
 	m.Save(1, &x.fr)
 	m.Save(2, &x.seq)
+	m.Save(3, &x.copyScratchBuffer)
 }
 
 func (x *VDSOParamPage) afterLoad() {}
@@ -2277,6 +2279,7 @@ func (x *VDSOParamPage) StateLoad(m state.Source) {
 	m.Load(0, &x.mfp)
 	m.Load(1, &x.fr)
 	m.Load(2, &x.seq)
+	m.Load(3, &x.copyScratchBuffer)
 }
 
 func init() {
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/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/vdso.go b/pkg/sentry/kernel/vdso.go
index e44a139b3..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
@@ -68,6 +69,13 @@ 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.
@@ -79,7 +87,11 @@ type VDSOParamPage struct {
 // * 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.
@@ -133,7 +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 {