Implement fallocate(2)

Closes #225

PiperOrigin-RevId: 247508791
Change-Id: I04f47cf2770b30043e5a272aba4ba6e11d0476cc

pkg/abi/linux/file.go

@@ -254,3 +254,14 @@ const (
 	F_SEAL_GROW   = 0x0004 // Prevent file from growing.
 	F_SEAL_WRITE  = 0x0008 // Prevent writes.
 )
+
+// Constants related to fallocate(2). Source: include/uapi/linux/falloc.h
+const (
+	FALLOC_FL_KEEP_SIZE      = 0x01
+	FALLOC_FL_PUNCH_HOLE     = 0x02
+	FALLOC_FL_NO_HIDE_STALE  = 0x04
+	FALLOC_FL_COLLAPSE_RANGE = 0x08
+	FALLOC_FL_ZERO_RANGE     = 0x10
+	FALLOC_FL_INSERT_RANGE   = 0x20
+	FALLOC_FL_UNSHARE_RANGE  = 0x40
+)

pkg/p9/BUILD

@@ -26,6 +26,7 @@ go_library(
         "//pkg/fd",
         "//pkg/log",
         "//pkg/unet",
+        "@org_golang_x_sys//unix:go_default_library",
     ],
 )
 

pkg/p9/client_file.go

@@ -171,6 +171,18 @@ func (c *clientFile) SetAttr(valid SetAttrMask, attr SetAttr) error {
 	return c.client.sendRecv(&Tsetattr{FID: c.fid, Valid: valid, SetAttr: attr}, &Rsetattr{})
 }
 
+// Allocate implements File.Allocate.
+func (c *clientFile) Allocate(mode AllocateMode, offset, length uint64) error {
+	if atomic.LoadUint32(&c.closed) != 0 {
+		return syscall.EBADF
+	}
+	if !versionSupportsTallocate(c.client.version) {
+		return syscall.EOPNOTSUPP
+	}
+
+	return c.client.sendRecv(&Tallocate{FID: c.fid, Mode: mode, Offset: offset, Length: length}, &Rallocate{})
+}
+
 // Remove implements File.Remove.
 //
 // N.B. This method is no longer part of the file interface and should be

pkg/p9/file.go

@@ -89,6 +89,10 @@ type File interface {
 	// On the server, SetAttr has a write concurrency guarantee.
 	SetAttr(valid SetAttrMask, attr SetAttr) error
 
+	// Allocate allows the caller to directly manipulate the allocated disk space
+	// for the file. See fallocate(2) for more details.
+	Allocate(mode AllocateMode, offset, length uint64) error
+
 	// Close is called when all references are dropped on the server side,
 	// and Close should be called by the client to drop all references.
 	//

pkg/p9/handlers.go

@@ -877,6 +877,40 @@ func (t *Tsetattr) handle(cs *connState) message {
 	return &Rsetattr{}
 }
 
+// handle implements handler.handle.
+func (t *Tallocate) handle(cs *connState) message {
+	// Lookup the FID.
+	ref, ok := cs.LookupFID(t.FID)
+	if !ok {
+		return newErr(syscall.EBADF)
+	}
+	defer ref.DecRef()
+
+	if err := ref.safelyWrite(func() error {
+		// Has it been opened already?
+		openFlags, opened := ref.OpenFlags()
+		if !opened {
+			return syscall.EINVAL
+		}
+
+		// Can it be written? Check permissions.
+		if openFlags&OpenFlagsModeMask == ReadOnly {
+			return syscall.EBADF
+		}
+
+		// We don't allow allocate on files that have been deleted.
+		if ref.isDeleted() {
+			return syscall.EINVAL
+		}
+
+		return ref.file.Allocate(t.Mode, t.Offset, t.Length)
+	}); err != nil {
+		return newErr(err)
+	}
+
+	return &Rallocate{}
+}
+
 // handle implements handler.handle.
 func (t *Txattrwalk) handle(cs *connState) message {
 	// Lookup the FID.

pkg/p9/local_server/local_server.go

@@ -323,6 +323,11 @@ func (l *local) Renamed(parent p9.File, newName string) {
 	l.path = path.Join(parent.(*local).path, newName)
 }
 
+// Allocate implements p9.File.Allocate.
+func (l *local) Allocate(mode p9.AllocateMode, offset, length uint64) error {
+	return syscall.Fallocate(int(l.file.Fd()), mode.ToLinux(), int64(offset), int64(length))
+}
+
 func main() {
 	log.SetLevel(log.Debug)
 

pkg/p9/messages.go

@@ -1424,6 +1424,63 @@ func (r *Rsetattr) String() string {
 	return fmt.Sprintf("Rsetattr{}")
 }
 
+// Tallocate is an allocate request. This is an extension to 9P protocol, not
+// present in the 9P2000.L standard.
+type Tallocate struct {
+	FID    FID
+	Mode   AllocateMode
+	Offset uint64
+	Length uint64
+}
+
+// Decode implements encoder.Decode.
+func (t *Tallocate) Decode(b *buffer) {
+	t.FID = b.ReadFID()
+	t.Mode.Decode(b)
+	t.Offset = b.Read64()
+	t.Length = b.Read64()
+}
+
+// Encode implements encoder.Encode.
+func (t *Tallocate) Encode(b *buffer) {
+	b.WriteFID(t.FID)
+	t.Mode.Encode(b)
+	b.Write64(t.Offset)
+	b.Write64(t.Length)
+}
+
+// Type implements message.Type.
+func (*Tallocate) Type() MsgType {
+	return MsgTallocate
+}
+
+// String implements fmt.Stringer.
+func (t *Tallocate) String() string {
+	return fmt.Sprintf("Tallocate{FID: %d, Offset: %d, Length: %d}", t.FID, t.Offset, t.Length)
+}
+
+// Rallocate is an allocate response.
+type Rallocate struct {
+}
+
+// Decode implements encoder.Decode.
+func (*Rallocate) Decode(b *buffer) {
+}
+
+// Encode implements encoder.Encode.
+func (*Rallocate) Encode(b *buffer) {
+}
+
+// Type implements message.Type.
+func (*Rallocate) Type() MsgType {
+	return MsgRallocate
+}
+
+// String implements fmt.Stringer.
+func (r *Rallocate) String() string {
+	return fmt.Sprintf("Rallocate{}")
+}
+
 // Txattrwalk walks extended attributes.
 type Txattrwalk struct {
 	// FID is the FID to check for attributes.
@@ -2297,4 +2354,6 @@ func init() {
 	msgRegistry.register(MsgRusymlink, func() message { return &Rusymlink{} })
 	msgRegistry.register(MsgTlconnect, func() message { return &Tlconnect{} })
 	msgRegistry.register(MsgRlconnect, func() message { return &Rlconnect{} })
+	msgRegistry.register(MsgTallocate, func() message { return &Tallocate{} })
+	msgRegistry.register(MsgRallocate, func() message { return &Rallocate{} })
 }

pkg/p9/p9.go

@@ -22,6 +22,8 @@ import (
 	"strings"
 	"sync/atomic"
 	"syscall"
+
+	"golang.org/x/sys/unix"
 )
 
 // OpenFlags is the mode passed to Open and Create operations.
@@ -374,6 +376,8 @@ const (
 	MsgRusymlink            = 135
 	MsgTlconnect            = 136
 	MsgRlconnect            = 137
+	MsgTallocate            = 138
+	MsgRallocate            = 139
 )
 
 // QIDType represents the file type for QIDs.
@@ -1058,3 +1062,80 @@ func (d *Dirent) Encode(b *buffer) {
 	b.WriteQIDType(d.Type)
 	b.WriteString(d.Name)
 }
+
+// AllocateMode are possible modes to p9.File.Allocate().
+type AllocateMode struct {
+	KeepSize      bool
+	PunchHole     bool
+	NoHideStale   bool
+	CollapseRange bool
+	ZeroRange     bool
+	InsertRange   bool
+	Unshare       bool
+}
+
+// ToLinux converts to a value compatible with fallocate(2)'s mode.
+func (a *AllocateMode) ToLinux() uint32 {
+	rv := uint32(0)
+	if a.KeepSize {
+		rv |= unix.FALLOC_FL_KEEP_SIZE
+	}
+	if a.PunchHole {
+		rv |= unix.FALLOC_FL_PUNCH_HOLE
+	}
+	if a.NoHideStale {
+		rv |= unix.FALLOC_FL_NO_HIDE_STALE
+	}
+	if a.CollapseRange {
+		rv |= unix.FALLOC_FL_COLLAPSE_RANGE
+	}
+	if a.ZeroRange {
+		rv |= unix.FALLOC_FL_ZERO_RANGE
+	}
+	if a.InsertRange {
+		rv |= unix.FALLOC_FL_INSERT_RANGE
+	}
+	if a.Unshare {
+		rv |= unix.FALLOC_FL_UNSHARE_RANGE
+	}
+	return rv
+}
+
+// Decode implements encoder.Decode.
+func (a *AllocateMode) Decode(b *buffer) {
+	mask := b.Read32()
+	a.KeepSize = mask&0x01 != 0
+	a.PunchHole = mask&0x02 != 0
+	a.NoHideStale = mask&0x04 != 0
+	a.CollapseRange = mask&0x08 != 0
+	a.ZeroRange = mask&0x10 != 0
+	a.InsertRange = mask&0x20 != 0
+	a.Unshare = mask&0x40 != 0
+}
+
+// Encode implements encoder.Encode.
+func (a *AllocateMode) Encode(b *buffer) {
+	mask := uint32(0)
+	if a.KeepSize {
+		mask |= 0x01
+	}
+	if a.PunchHole {
+		mask |= 0x02
+	}
+	if a.NoHideStale {
+		mask |= 0x04
+	}
+	if a.CollapseRange {
+		mask |= 0x08
+	}
+	if a.ZeroRange {
+		mask |= 0x10
+	}
+	if a.InsertRange {
+		mask |= 0x20
+	}
+	if a.Unshare {
+		mask |= 0x40
+	}
+	b.Write32(mask)
+}

pkg/p9/version.go

@@ -26,7 +26,7 @@ const (
 	//
 	// Clients are expected to start requesting this version number and
 	// to continuously decrement it until a Tversion request succeeds.
-	highestSupportedVersion uint32 = 6
+	highestSupportedVersion uint32 = 7
 
 	// lowestSupportedVersion is the lowest supported version X in a
 	// version string of the format 9P2000.L.Google.X.
@@ -143,3 +143,8 @@ func VersionSupportsAnonymous(v uint32) bool {
 func VersionSupportsMultiUser(v uint32) bool {
 	return v >= 6
 }
+
+// versionSupportsTallocate returns true if version v supports Allocate().
+func versionSupportsTallocate(v uint32) bool {
+	return v >= 7
+}

pkg/sentry/fs/ashmem/device.go

@@ -29,6 +29,7 @@ import (
 type Device struct {
 	fsutil.InodeGenericChecker       `state:"nosave"`
 	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNoopAllocate         `state:"nosave"`
 	fsutil.InodeNoopRelease          `state:"nosave"`
 	fsutil.InodeNoopTruncate         `state:"nosave"`
 	fsutil.InodeNoopWriteOut         `state:"nosave"`

pkg/sentry/fs/binder/binder.go

@@ -46,6 +46,7 @@ const (
 type Device struct {
 	fsutil.InodeGenericChecker       `state:"nosave"`
 	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNoopAllocate         `state:"nosave"`
 	fsutil.InodeNoopRelease          `state:"nosave"`
 	fsutil.InodeNoopTruncate         `state:"nosave"`
 	fsutil.InodeNoopWriteOut         `state:"nosave"`

pkg/sentry/fs/dev/full.go

@@ -30,6 +30,7 @@ import (
 type fullDevice struct {
 	fsutil.InodeGenericChecker       `state:"nosave"`
 	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNoopAllocate         `state:"nosave"`
 	fsutil.InodeNoopRelease          `state:"nosave"`
 	fsutil.InodeNoopTruncate         `state:"nosave"`
 	fsutil.InodeNoopWriteOut         `state:"nosave"`
@@ -59,7 +60,6 @@ func (f *fullDevice) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.Fi
 
 // +stateify savable
 type fullFileOperations struct {
-	waiter.AlwaysReady              `state:"nosave"`
 	fsutil.FileGenericSeek          `state:"nosave"`
 	fsutil.FileNoIoctl              `state:"nosave"`
 	fsutil.FileNoMMap               `state:"nosave"`
@@ -69,6 +69,7 @@ type fullFileOperations struct {
 	fsutil.FileNotDirReaddir        `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
 	readZeros                       `state:"nosave"`
+	waiter.AlwaysReady              `state:"nosave"`
 }
 
 var _ fs.FileOperations = (*fullFileOperations)(nil)

pkg/sentry/fs/dev/null.go

@@ -29,6 +29,7 @@ import (
 type nullDevice struct {
 	fsutil.InodeGenericChecker       `state:"nosave"`
 	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNoopAllocate         `state:"nosave"`
 	fsutil.InodeNoopRelease          `state:"nosave"`
 	fsutil.InodeNoopTruncate         `state:"nosave"`
 	fsutil.InodeNoopWriteOut         `state:"nosave"`
@@ -60,17 +61,17 @@ func (n *nullDevice) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.Fi
 
 // +stateify savable
 type nullFileOperations struct {
-	waiter.AlwaysReady              `state:"nosave"`
 	fsutil.FileGenericSeek          `state:"nosave"`
 	fsutil.FileNoIoctl              `state:"nosave"`
 	fsutil.FileNoMMap               `state:"nosave"`
 	fsutil.FileNoopFlush            `state:"nosave"`
 	fsutil.FileNoopFsync            `state:"nosave"`
 	fsutil.FileNoopRead             `state:"nosave"`
-	fsutil.FileNoopWrite            `state:"nosave"`
 	fsutil.FileNoopRelease          `state:"nosave"`
+	fsutil.FileNoopWrite            `state:"nosave"`
 	fsutil.FileNotDirReaddir        `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
+	waiter.AlwaysReady              `state:"nosave"`
 }
 
 var _ fs.FileOperations = (*nullFileOperations)(nil)
@@ -101,16 +102,16 @@ func (zd *zeroDevice) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.F
 
 // +stateify savable
 type zeroFileOperations struct {
-	waiter.AlwaysReady              `state:"nosave"`
 	fsutil.FileGenericSeek          `state:"nosave"`
+	fsutil.FileNoIoctl              `state:"nosave"`
 	fsutil.FileNoopFlush            `state:"nosave"`
 	fsutil.FileNoopFsync            `state:"nosave"`
 	fsutil.FileNoopRelease          `state:"nosave"`
 	fsutil.FileNoopWrite            `state:"nosave"`
 	fsutil.FileNotDirReaddir        `state:"nosave"`
-	fsutil.FileNoIoctl              `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
 	readZeros                       `state:"nosave"`
+	waiter.AlwaysReady              `state:"nosave"`
 }
 
 var _ fs.FileOperations = (*zeroFileOperations)(nil)

pkg/sentry/fs/dev/random.go

@@ -29,6 +29,7 @@ import (
 type randomDevice struct {
 	fsutil.InodeGenericChecker       `state:"nosave"`
 	fsutil.InodeNoExtendedAttributes `state:"nosave"`
+	fsutil.InodeNoopAllocate         `state:"nosave"`
 	fsutil.InodeNoopRelease          `state:"nosave"`
 	fsutil.InodeNoopTruncate         `state:"nosave"`
 	fsutil.InodeNoopWriteOut         `state:"nosave"`
@@ -57,16 +58,16 @@ func (*randomDevice) GetFile(ctx context.Context, dirent *fs.Dirent, flags fs.Fi
 
 // +stateify savable
 type randomFileOperations struct {
-	waiter.AlwaysReady              `state:"nosave"`
 	fsutil.FileGenericSeek          `state:"nosave"`
-	fsutil.FileNotDirReaddir        `state:"nosave"`
-	fsutil.FileNoMMap               `state:"nosave"`
-	fsutil.FileNoopFsync            `state:"nosave"`
-	fsutil.FileNoopFlush            `state:"nosave"`
 	fsutil.FileNoIoctl              `state:"nosave"`
+	fsutil.FileNoMMap               `state:"nosave"`
+	fsutil.FileNoopFlush            `state:"nosave"`
+	fsutil.FileNoopFsync            `state:"nosave"`
 	fsutil.FileNoopRelease          `state:"nosave"`
 	fsutil.FileNoopWrite            `state:"nosave"`
+	fsutil.FileNotDirReaddir        `state:"nosave"`
 	fsutil.FileUseInodeUnstableAttr `state:"nosave"`
+	waiter.AlwaysReady              `state:"nosave"`
 }
 
 var _ fs.FileOperations = (*randomFileOperations)(nil)

pkg/sentry/fs/fsutil/BUILD

@@ -113,5 +113,6 @@ go_test(
         "//pkg/sentry/memmap",
         "//pkg/sentry/safemem",
         "//pkg/sentry/usermem",
+        "//pkg/syserror",
     ],
 )

pkg/sentry/fs/fsutil/host_mappable.go

@@ -149,7 +149,7 @@ func (h *HostMappable) Truncate(ctx context.Context, newSize int64) error {
 	}
 
 	// Invalidate COW mappings that may exist beyond the new size in case the file
-	// is being shrunk. Other mappinsg don't need to be invalidated because
+	// is being shrunk. Other mappings don't need to be invalidated because
 	// translate will just return identical mappings after invalidation anyway,
 	// and SIGBUS will be raised and handled when the mappings are touched.
 	//
@@ -167,6 +167,14 @@ func (h *HostMappable) Truncate(ctx context.Context, newSize int64) error {
 	return nil
 }
 
+// Allocate reserves space in the backing file.
+func (h *HostMappable) Allocate(ctx context.Context, offset int64, length int64) error {
+	h.truncateMu.RLock()
+	err := h.backingFile.Allocate(ctx, offset, length)
+	h.truncateMu.RUnlock()
+	return err
+}
+
 // Write writes to the file backing this mappable.
 func (h *HostMappable) Write(ctx context.Context, src usermem.IOSequence, offset int64) (int64, error) {
 	h.truncateMu.RLock()

pkg/sentry/fs/fsutil/inode.go

@@ -34,6 +34,7 @@ type SimpleFileInode struct {
 	InodeNoExtendedAttributes `state:"nosave"`
 	InodeNoopRelease          `state:"nosave"`
 	InodeNoopWriteOut         `state:"nosave"`
+	InodeNotAllocatable       `state:"nosave"`
 	InodeNotDirectory         `state:"nosave"`
 	InodeNotMappable          `state:"nosave"`
 	InodeNotOpenable          `state:"nosave"`
@@ -61,6 +62,7 @@ type NoReadWriteFileInode struct {
 	InodeNoExtendedAttributes `state:"nosave"`
 	InodeNoopRelease          `state:"nosave"`
 	InodeNoopWriteOut         `state:"nosave"`
+	InodeNotAllocatable       `state:"nosave"`
 	InodeNotDirectory         `state:"nosave"`
 	InodeNotMappable          `state:"nosave"`
 	InodeNotSocket            `state:"nosave"`
@@ -465,3 +467,26 @@ func (InodeDenyWriteChecker) Check(ctx context.Context, inode *fs.Inode, p fs.Pe
 	}
 	return fs.ContextCanAccessFile(ctx, inode, p)
 }
+
+//InodeNotAllocatable can be used by Inodes that do not support Allocate().
+type InodeNotAllocatable struct{}
+
+func (InodeNotAllocatable) Allocate(_ context.Context, _ *fs.Inode, _, _ int64) error {
+	return syserror.EOPNOTSUPP
+}
+
+// InodeNoopAllocate implements fs.InodeOperations.Allocate as a noop.
+type InodeNoopAllocate struct{}
+
+// Allocate implements fs.InodeOperations.Allocate.
+func (InodeNoopAllocate) Allocate(_ context.Context, _ *fs.Inode, _, _ int64) error {
+	return nil
+}
+
+// InodeIsDirAllocate implements fs.InodeOperations.Allocate for directories.
+type InodeIsDirAllocate struct{}
+
+// Allocate implements fs.InodeOperations.Allocate.
+func (InodeIsDirAllocate) Allocate(_ context.Context, _ *fs.Inode, _, _ int64) error {
+	return syserror.EISDIR
+}

pkg/sentry/fs/fsutil/inode_cached.go

@@ -135,6 +135,10 @@ type CachedFileObject interface {
 	// the file was opened.
 	SetMaskedAttributes(ctx context.Context, mask fs.AttrMask, attr fs.UnstableAttr) error
 
+	// Allocate allows the caller to reserve disk space for the inode.
+	// It's equivalent to fallocate(2) with 'mode=0'.
+	Allocate(ctx context.Context, offset int64, length int64) error
+
 	// Sync instructs the remote filesystem to sync the file to stable storage.
 	Sync(ctx context.Context) error
 
@@ -336,6 +340,30 @@ func (c *CachingInodeOperations) Truncate(ctx context.Context, inode *fs.Inode,
 	return nil
 }
 
+// Allocate implements fs.InodeOperations.Allocate.
+func (c *CachingInodeOperations) Allocate(ctx context.Context, offset, length int64) error {
+	newSize := offset + length
+
+	// c.attr.Size is protected by both c.attrMu and c.dataMu.
+	c.attrMu.Lock()
+	defer c.attrMu.Unlock()
+	c.dataMu.Lock()
+	defer c.dataMu.Unlock()
+
+	if newSize <= c.attr.Size {
+		return nil
+	}
+
+	now := ktime.NowFromContext(ctx)
+	if err := c.backingFile.Allocate(ctx, offset, length); err != nil {
+		return err
+	}
+
+	c.attr.Size = newSize
+	c.touchModificationTimeLocked(now)
+	return nil
+}
+
 // WriteOut implements fs.InodeOperations.WriteOut.
 func (c *CachingInodeOperations) WriteOut(ctx context.Context, inode *fs.Inode) error {
 	c.attrMu.Lock()

pkg/sentry/fs/fsutil/inode_cached_test.go

@@ -26,6 +26,7 @@ import (
 	"gvisor.googlesource.com/gvisor/pkg/sentry/memmap"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/safemem"
 	"gvisor.googlesource.com/gvisor/pkg/sentry/usermem"
+	"gvisor.googlesource.com/gvisor/pkg/syserror"
 )
 
 type noopBackingFile struct{}
@@ -50,6 +51,10 @@ func (noopBackingFile) FD() int {
 	return -1
 }
 
+func (noopBackingFile) Allocate(ctx context.Context, offset int64, length int64) error {
+	return nil
+}
+
 func TestSetPermissions(t *testing.T) {
 	ctx := contexttest.Context(t)
 
@@ -237,6 +242,10 @@ func (*sliceBackingFile) FD() int {
 	return -1
 }
 
+func (f *sliceBackingFile) Allocate(ctx context.Context, offset int64, length int64) error {
+	return syserror.EOPNOTSUPP
+}
+
 type noopMappingSpace struct{}
 
 // Invalidate implements memmap.MappingSpace.Invalidate.

pkg/sentry/fs/gofer/context_file.go

@@ -59,6 +59,13 @@ func (c *contextFile) setAttr(ctx context.Context, valid p9.SetAttrMask, attr p9
 	return err
 }
 
+func (c *contextFile) allocate(ctx context.Context, mode p9.AllocateMode, offset, length uint64) error {
+	ctx.UninterruptibleSleepStart(false)
+	err := c.file.Allocate(mode, offset, length)
+	ctx.UninterruptibleSleepFinish(false)
+	return err
+}
+
 func (c *contextFile) rename(ctx context.Context, directory contextFile, name string) error {
 	ctx.UninterruptibleSleepStart(false)
 	err := c.file.Rename(directory.file, name)