Eliminated an allocation and a copy for each ReadFileOp.

This adds up to a significant performance gain for gcsfuse, in its sequential
read benchmark.

Also ReadDirOp.

connection.go

@@ -87,6 +87,7 @@ type Connection struct {
 // context that the user uses to reply to the op.
 type opState struct {
 	inMsg  *buffer.InMessage
+	outMsg *buffer.OutMessage
 	op     interface{}
 	opID   uint32 // For logging
 }
@@ -370,15 +371,17 @@ func (c *Connection) ReadOp() (ctx context.Context, op interface{}, err error) {
 	// Keep going until we find a request we know how to convert.
 	for {
 		// Read the next message from the kernel.
-		var m *buffer.InMessage
-		m, err = c.readMessage()
+		var inMsg *buffer.InMessage
+		inMsg, err = c.readMessage()
 		if err != nil {
 			return
 		}
 
 		// Convert the message to an op.
-		op, err = convertInMessage(m, c.protocol)
+		outMsg := c.getOutMessage()
+		op, err = convertInMessage(inMsg, outMsg, c.protocol)
 		if err != nil {
+			c.putOutMessage(outMsg)
 			err = fmt.Errorf("convertInMessage: %v", err)
 			return
 		}
@@ -396,8 +399,8 @@ func (c *Connection) ReadOp() (ctx context.Context, op interface{}, err error) {
 		}
 
 		// Set up a context that remembers information about this op.
-		ctx = c.beginOp(m.Header().Opcode, m.Header().Unique)
-		ctx = context.WithValue(ctx, contextKey, opState{m, op, opID})
+		ctx = c.beginOp(inMsg.Header().Opcode, inMsg.Header().Unique)
+		ctx = context.WithValue(ctx, contextKey, opState{inMsg, outMsg, op, opID})
 
 		// Special case: responding to statfs is required to make mounting work on
 		// OS X. We don't currently expose the capability for the file system to
@@ -426,14 +429,16 @@ func (c *Connection) Reply(ctx context.Context, opErr error) {
 	}
 
 	op := state.op
-	m := state.inMsg
+	inMsg := state.inMsg
+	outMsg := state.outMsg
 	opID := state.opID
 
-	// Make sure we destroy the message when we're done.
-	defer c.putInMessage(m)
+	// Make sure we destroy the messages when we're done.
+	defer c.putInMessage(inMsg)
+	defer c.putOutMessage(outMsg)
 
 	// Clean up state for this op.
-	c.finishOp(m.Header().Opcode, m.Header().Unique)
+	c.finishOp(inMsg.Header().Opcode, inMsg.Header().Unique)
 
 	// Debug logging
 	if c.debugLogger != nil {
@@ -450,11 +455,10 @@ func (c *Connection) Reply(ctx context.Context, opErr error) {
 	}
 
 	// Send the reply to the kernel, if one is required.
-	outMsg := c.kernelResponse(m.Header().Unique, op, opErr)
-	if outMsg != nil {
-		err := c.writeMessage(outMsg.Bytes())
-		c.putOutMessage(outMsg)
+	noResponse := c.kernelResponse(outMsg, inMsg.Header().Unique, op, opErr)
 
+	if !noResponse {
+		err := c.writeMessage(outMsg.Bytes())
 		if err != nil && c.errorLogger != nil {
 			c.errorLogger.Printf("writeMessage: %v", err)
 		}

fuseops/ops.go

@@ -384,25 +384,29 @@ type ReadDirOp struct {
 	// offset, and return array offsets into that cached listing.
 	Offset DirOffset
 
-	// The maximum number of bytes to return in ReadDirResponse.Data. A smaller
-	// number is acceptable.
-	Size int
-
-	// Set by the file system: a buffer consisting of a sequence of FUSE
-	// directory entries in the format generated by fuse_add_direntry
-	// (http://goo.gl/qCcHCV), which is consumed by parse_dirfile
-	// (http://goo.gl/2WUmD2). Use fuseutil.AppendDirent to generate this data.
+	// The destination buffer, whose length gives the size of the read.
 	//
-	// The buffer must not exceed the length specified in ReadDirRequest.Size. It
-	// is okay for the final entry to be truncated; parse_dirfile copes with this
-	// by ignoring the partial record.
+	// The output data should consist of a sequence of FUSE directory entries in
+	// the format generated by fuse_add_direntry (http://goo.gl/qCcHCV), which is
+	// consumed by parse_dirfile (http://goo.gl/2WUmD2). Use fuseutil.WriteDirent
+	// to generate this data.
 	//
 	// Each entry returned exposes a directory offset to the user that may later
 	// show up in ReadDirRequest.Offset. See notes on that field for more
 	// information.
+	Dst []byte
+
+	// Set by the file system: the number of bytes read into Dst.
 	//
-	// An empty buffer indicates the end of the directory has been reached.
-	Data []byte
+	// It is okay for this to be less than len(Dst) if there are not enough
+	// entries available or the final entry would not fit.
+	//
+	// Zero means that the end of the directory has been reached. This is
+	// unambiguous because NAME_MAX (https://goo.gl/ZxzKaE) plus the size of
+	// fuse_dirent (https://goo.gl/WO8s3F) plus the 8-byte alignment of
+	// FUSE_DIRENT_ALIGN (http://goo.gl/UziWvH) is less than the read size of
+	// PAGE_SIZE used by fuse_readdir (cf. https://goo.gl/VajtS2).
+	BytesRead int
 }
 
 // Release a previously-minted directory handle. The kernel sends this when
@@ -455,20 +459,21 @@ type ReadFileOp struct {
 	Inode  InodeID
 	Handle HandleID
 
-	// The range of the file to read.
+	// The offset within the file at which to read.
+	Offset int64
+
+	// The destination buffer, whose length gives the size of the read.
+	Dst []byte
+
+	// Set by the file system: the number of bytes read.
 	//
-	// The FUSE documentation requires that exactly the number of bytes be
-	// returned, except in the case of EOF or error (http://goo.gl/ZgfBkF). This
-	// appears to be because it uses file mmapping machinery
+	// The FUSE documentation requires that exactly the requested number of bytes
+	// be returned, except in the case of EOF or error (http://goo.gl/ZgfBkF).
+	// This appears to be because it uses file mmapping machinery
 	// (http://goo.gl/SGxnaN) to read a page at a time. It appears to understand
 	// where EOF is by checking the inode size (http://goo.gl/0BkqKD), returned
 	// by a previous call to LookUpInode, GetInodeAttributes, etc.
-	Offset int64
-	Size   int
-
-	// Set by the file system: the data read. If this is less than the requested
-	// size, it indicates EOF. An error should not be returned in this case.
-	Data []byte
+	BytesRead int
 }
 
 // Write data to a file previously opened with CreateFile or OpenFile.

fuseutil/dirent.go

@@ -35,7 +35,7 @@ const (
 )
 
 // A struct representing an entry within a directory file, describing a child.
-// See notes on fuseops.ReadDirOp and on AppendDirent for details.
+// See notes on fuseops.ReadDirOp and on WriteDirent for details.
 type Dirent struct {
 	// The (opaque) offset within the directory file of the entry following this
 	// one. See notes on fuseops.ReadDirOp.Offset for details.
@@ -50,10 +50,11 @@ type Dirent struct {
 	Type DirentType
 }
 
-// Append the supplied directory entry to the given buffer in the format
-// expected in fuseops.ReadFileOp.Data, returning the resulting buffer.
-func AppendDirent(input []byte, d Dirent) (output []byte) {
-	// We want to append bytes with the layout of fuse_dirent
+// Write the supplied directory entry intto the given buffer in the format
+// expected in fuseops.ReadFileOp.Data, returning the number of bytes written.
+// Return zero if the entry would not fit.
+func WriteDirent(buf []byte, d Dirent) (n int) {
+	// We want to write bytes with the layout of fuse_dirent
 	// (http://goo.gl/BmFxob) in host order. The struct must be aligned according
 	// to FUSE_DIRENT_ALIGN (http://goo.gl/UziWvH), which dictates 8-byte
 	// alignment.
@@ -65,10 +66,23 @@ func AppendDirent(input []byte, d Dirent) (output []byte) {
 		name    [0]byte
 	}
 
-	const alignment = 8
-	const nameOffset = 8 + 8 + 4 + 4
+	const direntAlignment = 8
+	const direntSize = 8 + 8 + 4 + 4
 
-	// Write the header into the buffer.
+	// Compute the number of bytes of padding we'll need to maintain alignment
+	// for the next entry.
+	var padLen int
+	if len(d.Name)%direntAlignment != 0 {
+		padLen = direntAlignment - (len(d.Name) % direntAlignment)
+	}
+
+	// Do we have enough room?
+	totalLen := direntSize + len(d.Name) + padLen
+	if totalLen > len(buf) {
+		return
+	}
+
+	// Write the header.
 	de := fuse_dirent{
 		ino:     uint64(d.Inode),
 		off:     uint64(d.Offset),
@@ -76,17 +90,15 @@ func AppendDirent(input []byte, d Dirent) (output []byte) {
 		type_:   uint32(d.Type),
 	}
 
-	output = append(input, (*[nameOffset]byte)(unsafe.Pointer(&de))[:]...)
+	n += copy(buf[n:], (*[direntSize]byte)(unsafe.Pointer(&de))[:])
 
 	// Write the name afterward.
-	output = append(output, d.Name...)
+	n += copy(buf[n:], d.Name)
 
 	// Add any necessary padding.
-	if len(d.Name)%alignment != 0 {
-		padLen := alignment - (len(d.Name) % alignment)
-
-		var padding [alignment]byte
-		output = append(output, padding[:padLen]...)
+	if padLen != 0 {
+		var padding [direntAlignment]byte
+		n += copy(buf[n:], padding[:padLen])
 	}
 
 	return

internal/buffer/out_message.go

@@ -25,6 +25,9 @@ import (
 
 const outHeaderSize = unsafe.Sizeof(fusekernel.OutHeader{})
 
+// OutMessage structs begin life with Len() == OutMessageInitialSize.
+const OutMessageInitialSize = outHeaderSize
+
 // We size out messages to be large enough to hold a header for the response
 // plus the largest read that may come in.
 const outMessageSize = outHeaderSize + MaxReadSize
@@ -53,8 +56,8 @@ func init() {
 // Reset the message so that it is ready to be used again. Afterward, the
 // contents are solely a zeroed header.
 func (m *OutMessage) Reset() {
-	m.offset = outHeaderSize
-	memclr(unsafe.Pointer(&m.storage), outHeaderSize)
+	m.offset = OutMessageInitialSize
+	memclr(unsafe.Pointer(&m.storage), OutMessageInitialSize)
 }
 
 // Return a pointer to the header at the start of the message.
@@ -87,6 +90,15 @@ func (b *OutMessage) GrowNoZero(size uintptr) (p unsafe.Pointer) {
 	return
 }
 
+// Throw away the last n bytes. Panics if n is out of range.
+func (b *OutMessage) Shrink(n uintptr) {
+	if n > b.offset-OutMessageInitialSize {
+		panic(fmt.Sprintf("Shrink(%d) out of range for offset %d", n, b.offset))
+	}
+
+	b.offset -= n
+}
+
 // Equivalent to growing by the length of p, then copying p over the new
 // segment. Panics if there is not enough room available.
 func (b *OutMessage) Append(src []byte) {

samples/flushfs/flush_fs.go

@@ -180,8 +180,7 @@ func (fs *flushFS) ReadFile(
 	}
 
 	// Read what we can.
-	op.Data = make([]byte, op.Size)
-	copy(op.Data, fs.fooContents[op.Offset:])
+	op.BytesRead = copy(op.Dst, fs.fooContents[op.Offset:])
 
 	return
 }
@@ -298,14 +297,16 @@ func (fs *flushFS) ReadDir(
 
 	// Fill in the listing.
 	for _, de := range dirents {
-		op.Data = fuseutil.AppendDirent(op.Data, de)
-	}
+		n := fuseutil.WriteDirent(op.Dst[op.BytesRead:], de)
 
 		// We don't support doing this in anything more than one shot.
-	if len(op.Data) > op.Size {
-		err = fmt.Errorf("Couldn't fit listing in %v bytes", op.Size)
+		if n == 0 {
+			err = fmt.Errorf("Couldn't fit listing in %v bytes", len(op.Dst))
 			return
 		}
 
+		op.BytesRead += n
+	}
+
 	return
 }

samples/hellofs/hello_fs.go

@@ -228,11 +228,12 @@ func (fs *helloFS) ReadDir(
 
 	// Resume at the specified offset into the array.
 	for _, e := range entries {
-		op.Data = fuseutil.AppendDirent(op.Data, e)
-		if len(op.Data) > op.Size {
-			op.Data = op.Data[:op.Size]
+		n := fuseutil.WriteDirent(op.Dst[op.BytesRead:], e)
+		if n == 0 {
 			break
 		}
+
+		op.BytesRead += n
 	}
 
 	return
@@ -251,9 +252,7 @@ func (fs *helloFS) ReadFile(
 	// Let io.ReaderAt deal with the semantics.
 	reader := strings.NewReader("Hello, world!")
 
-	op.Data = make([]byte, op.Size)
-	n, err := reader.ReadAt(op.Data, op.Offset)
-	op.Data = op.Data[:n]
+	op.BytesRead, err = reader.ReadAt(op.Dst, op.Offset)
 
 	// Special case: FUSE doesn't expect us to return io.EOF.
 	if err == io.EOF {

samples/memfs/inode.go

@@ -278,7 +278,7 @@ func (in *inode) RemoveChild(name string) {
 // Serve a ReadDir request.
 //
 // REQUIRES: in.isDir()
-func (in *inode) ReadDir(offset int, size int) (data []byte) {
+func (in *inode) ReadDir(p []byte, offset int) (n int) {
 	if !in.isDir() {
 		panic("ReadDir called on non-directory.")
 	}
@@ -291,13 +291,12 @@ func (in *inode) ReadDir(offset int, size int) (data []byte) {
 			continue
 		}
 
-		data = fuseutil.AppendDirent(data, in.entries[i])
-
-		// Trim and stop early if we've exceeded the requested size.
-		if len(data) > size {
-			data = data[:size]
+		tmp := fuseutil.WriteDirent(p[n:], in.entries[i])
+		if tmp == 0 {
 			break
 		}
+
+		n += tmp
 	}
 
 	return

samples/memfs/memfs.go

@@ -428,7 +428,9 @@ func (fs *memFS) Rename(
 	existingID, _, ok := newParent.LookUpChild(op.NewName)
 	if ok {
 		existing := fs.getInodeOrDie(existingID)
-		if existing.isDir() && len(existing.ReadDir(0, 1024)) > 0 {
+
+		var buf [4096]byte
+		if existing.isDir() && existing.ReadDir(buf[:], 0) > 0 {
 			err = fuse.ENOTEMPTY
 			return
 		}
@@ -538,7 +540,7 @@ func (fs *memFS) ReadDir(
 	inode := fs.getInodeOrDie(op.Inode)
 
 	// Serve the request.
-	op.Data = inode.ReadDir(int(op.Offset), op.Size)
+	op.BytesRead = inode.ReadDir(op.Dst, int(op.Offset))
 
 	return
 }
@@ -571,9 +573,7 @@ func (fs *memFS) ReadFile(
 	inode := fs.getInodeOrDie(op.Inode)
 
 	// Serve the request.
-	op.Data = make([]byte, op.Size)
-	n, err := inode.ReadAt(op.Data, op.Offset)
-	op.Data = op.Data[:n]
+	op.BytesRead, err = inode.ReadAt(op.Dst, op.Offset)
 
 	// Don't return EOF errors; we just indicate EOF to fuse using a short read.
 	if err == io.EOF {