Remove default os.ModeDevice in ConvertFileMode

connection.go

@@ -151,6 +151,7 @@ func (c *Connection) Init() error {
 	cacheSymlinks := initOp.Flags&fusekernel.InitCacheSymlinks > 0
 	noOpenSupport := initOp.Flags&fusekernel.InitNoOpenSupport > 0
 	noOpendirSupport := initOp.Flags&fusekernel.InitNoOpendirSupport > 0
+	readdirplusSupport := initOp.Flags&fusekernel.InitDoReaddirplus > 0
 
 	// Respond to the init op.
 	initOp.Library = c.protocol
@@ -193,6 +194,11 @@ func (c *Connection) Init() error {
 		initOp.Flags |= fusekernel.InitNoOpendirSupport
 	}
 
+	// Tell the kernel to do readdirplus (readdir+lookup in one call)
+	if c.cfg.UseReadDirPlus && readdirplusSupport {
+		initOp.Flags |= fusekernel.InitDoReaddirplus
+	}
+
 	c.Reply(ctx, nil)
 	return nil
 }
@@ -369,18 +375,24 @@ func (c *Connection) readMessage() (*buffer.InMessage, error) {
 }
 
 // Write the supplied message to the kernel.
-func (c *Connection) writeMessage(msg []byte) error {
-	// Avoid the retry loop in os.File.Write.
-	n, err := syscall.Write(int(c.dev.Fd()), msg)
-	if err != nil {
-		return err
+func (c *Connection) writeMessage(outMsg *buffer.OutMessage) error {
+	var err error
+	var n int
+	expectedLen := outMsg.Len()
+	if outMsg.Sglist != nil {
+		n, err = writev(int(c.dev.Fd()), outMsg.Sglist)
+	} else {
+		// Avoid the retry loop in os.File.Write.
+		n, err = syscall.Write(int(c.dev.Fd()), outMsg.OutHeaderBytes())
 	}
-
-	if n != len(msg) {
-		return fmt.Errorf("Wrote %d bytes; expected %d", n, len(msg))
+	if err == nil && n != expectedLen {
+		err = fmt.Errorf("Wrote %d bytes; expected %d", n, expectedLen)
 	}
-
-	return nil
+	if err != nil && c.errorLogger != nil {
+		c.errorLogger.Printf("writeMessage: %v %v", err, outMsg.OutHeaderBytes())
+	}
+	outMsg.Sglist = nil
+	return err
 }
 
 // ReadOp consumes the next op from the kernel process, returning the op and a
@@ -485,8 +497,15 @@ func (c *Connection) Reply(ctx context.Context, opErr error) {
 	outMsg := state.outMsg
 	fuseID := inMsg.Header().Unique
 
+	suppressReuse := false
+	if wr, ok := op.(*fuseops.WriteFileOp); ok {
+		suppressReuse = wr.SuppressReuse
+	}
+
 	// Make sure we destroy the messages when we're done.
-	defer c.putInMessage(inMsg)
+	if !suppressReuse {
+		defer c.putInMessage(inMsg)
+	}
 	defer c.putOutMessage(outMsg)
 
 	// Clean up state for this op.
@@ -510,19 +529,21 @@ func (c *Connection) Reply(ctx context.Context, opErr error) {
 	noResponse := c.kernelResponse(outMsg, inMsg.Header().Unique, op, opErr)
 
 	if !noResponse {
-		var err error
-		if outMsg.Sglist != nil {
-			_, err = writev(int(c.dev.Fd()), outMsg.Sglist)
-		} else {
-			err = c.writeMessage(outMsg.OutHeaderBytes())
-		}
-		if err != nil && c.errorLogger != nil {
-			c.errorLogger.Printf("writeMessage: %v %v", err, outMsg.OutHeaderBytes())
-		}
-		outMsg.Sglist = nil
+		c.writeMessage(outMsg)
 	}
 }
 
+// Send a notification to the kernel
+// notification must be a pointer to one of fuseops.NotifyXXX structures
+// To avoid a deadlock notifications must not be called in the execution path of a related filesytem operation or within any code that could hold a lock that could be needed to execute such an operation. As of kernel 4.18, a "related operation" is a lookup(), symlink(), mknod(), mkdir(), unlink(), rename(), link() or create() request for the parent, and a setattr(), unlink(), rmdir(), rename(), setxattr(), removexattr(), readdir() or readdirplus() request for the inode itself.
+func (c *Connection) Notify(notification interface{}) error {
+	outMsg := c.getOutMessage()
+	defer c.putOutMessage(outMsg)
+	c.kernelNotification(outMsg, notification)
+	outMsg.OutHeader().Len = uint32(outMsg.Len())
+	return c.writeMessage(outMsg)
+}
+
 // Close the connection. Must not be called until operations that were read
 // from the connection have been responded to.
 func (c *Connection) close() error {

conversions.go

@@ -89,7 +89,7 @@ func convertInMessage(
 		}
 
 		if valid&fusekernel.SetattrMode != 0 {
-			mode := convertFileMode(in.Mode)
+			mode := fuseops.ConvertFileMode(in.Mode)
 			to.Mode = &mode
 		}
 
@@ -170,7 +170,7 @@ func convertInMessage(
 			// the fact that this is a directory is implicit in the fact that the
 			// opcode is mkdir. But we want the correct mode to go through, so ensure
 			// that os.ModeDir is set.
-			Mode:      convertFileMode(in.Mode) | os.ModeDir,
+			Mode:      fuseops.ConvertFileMode(in.Mode) | os.ModeDir,
 			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
 		}
 
@@ -190,7 +190,8 @@ func convertInMessage(
 		o = &fuseops.MkNodeOp{
 			Parent:    fuseops.InodeID(inMsg.Header().Nodeid),
 			Name:      string(name),
-			Mode:      convertFileMode(in.Mode),
+			Mode:      fuseops.ConvertFileMode(in.Mode),
+			Rdev:      in.Rdev,
 			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
 		}
 
@@ -210,7 +211,7 @@ func convertInMessage(
 		o = &fuseops.CreateFileOp{
 			Parent:    fuseops.InodeID(inMsg.Header().Nodeid),
 			Name:      string(name),
-			Mode:      convertFileMode(in.Mode),
+			Mode:      fuseops.ConvertFileMode(in.Mode),
 			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
 		}
 
@@ -340,6 +341,8 @@ func convertInMessage(
 		}
 		o = to
 
+	case fusekernel.OpReaddirplus:
+		fallthrough
 	case fusekernel.OpReaddir:
 		in := (*fusekernel.ReadIn)(inMsg.Consume(fusekernel.ReadInSize(protocol)))
 		if in == nil {
@@ -350,6 +353,7 @@ func convertInMessage(
 			Inode:     fuseops.InodeID(inMsg.Header().Nodeid),
 			Handle:    fuseops.HandleID(in.Fh),
 			Offset:    fuseops.DirOffset(in.Offset),
+			Plus:      inMsg.Header().Opcode == fusekernel.OpReaddirplus,
 			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
 		}
 		o = to
@@ -606,6 +610,42 @@ func convertInMessage(
 			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
 		}
 
+	case fusekernel.OpPoll:
+		type input fusekernel.PollIn
+		in := (*input)(inMsg.Consume(unsafe.Sizeof(input{})))
+		if in == nil {
+			return nil, errors.New("Corrupt OpPoll")
+		}
+
+		o = &fuseops.PollOp{
+			Inode:     fuseops.InodeID(inMsg.Header().Nodeid),
+			Handle:    fuseops.HandleID(in.Fh),
+			Kh:        in.Kh,
+			Flags:     fusekernel.PollFlags(in.Flags),
+			Events:    fusekernel.PollEvents(in.Events),
+			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
+		}
+
+	case fusekernel.OpNotifyReply:
+		type input fusekernel.NotifyRetrieveIn
+		in := (*input)(inMsg.Consume(unsafe.Sizeof(input{})))
+		if in == nil {
+			return nil, errors.New("Corrupt OpNotifyReply")
+		}
+
+		buf := inMsg.ConsumeBytes(inMsg.Len())
+		if len(buf) < int(in.Size) {
+			return nil, errors.New("Corrupt OpNotifyReply")
+		}
+
+		o = &fuseops.NotifyRetrieveReplyOp{
+			Inode:     fuseops.InodeID(inMsg.Header().Nodeid),
+			Unique:    inMsg.Header().Unique,
+			Offset:    in.Offset,
+			Length:    in.Size,
+			OpContext: fuseops.OpContext{Pid: inMsg.Header().Pid},
+		}
+
 	default:
 		o = &unknownOp{
 			OpCode: inMsg.Header().Opcode,
@@ -639,6 +679,9 @@ func (c *Connection) kernelResponse(
 	case *fuseops.BatchForgetOp:
 		return true
 
+	case *fuseops.NotifyRetrieveReplyOp:
+		return true
+
 	case *interruptOp:
 		return true
 	}
@@ -682,37 +725,37 @@ func (c *Connection) kernelResponseForOp(
 	case *fuseops.LookUpInodeOp:
 		size := int(fusekernel.EntryOutSize(c.protocol))
 		out := (*fusekernel.EntryOut)(m.Grow(size))
-		convertChildInodeEntry(&o.Entry, out)
+		fuseops.ConvertChildInodeEntry(&o.Entry, out)
 
 	case *fuseops.GetInodeAttributesOp:
 		size := int(fusekernel.AttrOutSize(c.protocol))
 		out := (*fusekernel.AttrOut)(m.Grow(size))
-		out.AttrValid, out.AttrValidNsec = convertExpirationTime(
+		out.AttrValid, out.AttrValidNsec = fuseops.ConvertExpirationTime(
 			o.AttributesExpiration)
-		convertAttributes(o.Inode, &o.Attributes, &out.Attr)
+		fuseops.ConvertAttributes(o.Inode, &o.Attributes, &out.Attr)
 
 	case *fuseops.SetInodeAttributesOp:
 		size := int(fusekernel.AttrOutSize(c.protocol))
 		out := (*fusekernel.AttrOut)(m.Grow(size))
-		out.AttrValid, out.AttrValidNsec = convertExpirationTime(
+		out.AttrValid, out.AttrValidNsec = fuseops.ConvertExpirationTime(
 			o.AttributesExpiration)
-		convertAttributes(o.Inode, &o.Attributes, &out.Attr)
+		fuseops.ConvertAttributes(o.Inode, &o.Attributes, &out.Attr)
 
 	case *fuseops.MkDirOp:
 		size := int(fusekernel.EntryOutSize(c.protocol))
 		out := (*fusekernel.EntryOut)(m.Grow(size))
-		convertChildInodeEntry(&o.Entry, out)
+		fuseops.ConvertChildInodeEntry(&o.Entry, out)
 
 	case *fuseops.MkNodeOp:
 		size := int(fusekernel.EntryOutSize(c.protocol))
 		out := (*fusekernel.EntryOut)(m.Grow(size))
-		convertChildInodeEntry(&o.Entry, out)
+		fuseops.ConvertChildInodeEntry(&o.Entry, out)
 
 	case *fuseops.CreateFileOp:
 		eSize := int(fusekernel.EntryOutSize(c.protocol))
 
 		e := (*fusekernel.EntryOut)(m.Grow(eSize))
-		convertChildInodeEntry(&o.Entry, e)
+		fuseops.ConvertChildInodeEntry(&o.Entry, e)
 
 		oo := (*fusekernel.OpenOut)(m.Grow(int(unsafe.Sizeof(fusekernel.OpenOut{}))))
 		oo.Fh = uint64(o.Handle)
@@ -720,12 +763,12 @@ func (c *Connection) kernelResponseForOp(
 	case *fuseops.CreateSymlinkOp:
 		size := int(fusekernel.EntryOutSize(c.protocol))
 		out := (*fusekernel.EntryOut)(m.Grow(size))
-		convertChildInodeEntry(&o.Entry, out)
+		fuseops.ConvertChildInodeEntry(&o.Entry, out)
 
 	case *fuseops.CreateLinkOp:
 		size := int(fusekernel.EntryOutSize(c.protocol))
 		out := (*fusekernel.EntryOut)(m.Grow(size))
-		convertChildInodeEntry(&o.Entry, out)
+		fuseops.ConvertChildInodeEntry(&o.Entry, out)
 
 	case *fuseops.RenameOp:
 		// Empty response
@@ -862,6 +905,13 @@ func (c *Connection) kernelResponseForOp(
 		out.TimeGran = 1
 		out.MaxPages = o.MaxPages
 
+	case *fuseops.PollOp:
+		out := (*fusekernel.PollOut)(m.Grow(int(unsafe.Sizeof(fusekernel.PollOut{}))))
+		out.Revents = uint32(o.Revents)
+
+	case *fuseops.NotifyRetrieveReplyOp:
+		// Empty response
+
 	default:
 		panic(fmt.Sprintf("Unexpected op: %#v", op))
 	}
@@ -869,124 +919,73 @@ func (c *Connection) kernelResponseForOp(
 	return
 }
 
+// Like kernelResponse, but assumes the user replied with a nil error to the op.
+func (c *Connection) kernelNotification(
+	m *buffer.OutMessage,
+	op interface{}) {
+
+	h := m.OutHeader()
+	h.Unique = 0
+
+	// Create the appropriate output message
+	switch o := op.(type) {
+	case *fuseops.NotifyPollWakeup:
+		h.Error = fusekernel.NotifyCodePoll
+		out := (*fusekernel.NotifyPollWakeupOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyPollWakeupOut{}))))
+		out.Kh = o.Kh
+
+	case *fuseops.NotifyInvalInode:
+		h.Error = fusekernel.NotifyCodeInvalInode
+		out := (*fusekernel.NotifyInvalInodeOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyInvalInodeOut{}))))
+		out.Ino = uint64(o.Inode)
+		out.Off = o.Offset
+		out.Len = o.Length
+
+	case *fuseops.NotifyInvalEntry:
+		h.Error = fusekernel.NotifyCodeInvalEntry
+		out := (*fusekernel.NotifyInvalEntryOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyInvalEntryOut{}))))
+		out.Parent = uint64(o.Parent)
+		out.Namelen = uint32(len(o.Name))
+		m.AppendString(o.Name)
+		m.AppendString("\x00")
+
+	case *fuseops.NotifyDelete:
+		h.Error = fusekernel.NotifyCodeDelete
+		out := (*fusekernel.NotifyDeleteOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyDeleteOut{}))))
+		out.Parent = uint64(o.Parent)
+		out.Child = uint64(o.Child)
+		out.Namelen = uint32(len(o.Name))
+		m.AppendString(o.Name)
+		m.AppendString("\x00")
+
+	case *fuseops.NotifyStore:
+		h.Error = fusekernel.NotifyCodeStore
+		out := (*fusekernel.NotifyStoreOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyStoreOut{}))))
+		out.Nodeid = uint64(o.Inode)
+		out.Offset = o.Offset
+		out.Size = o.Length
+		m.Append(o.Data...)
+		m.ShrinkTo(buffer.OutMessageHeaderSize + int(unsafe.Sizeof(fusekernel.NotifyStoreOut{})) + int(o.Length))
+
+	case *fuseops.NotifyRetrieve:
+		h.Error = fusekernel.NotifyCodeRetrieve
+		out := (*fusekernel.NotifyRetrieveOut)(m.Grow(int(unsafe.Sizeof(fusekernel.NotifyRetrieveOut{}))))
+		out.Unique = o.Unique
+		out.Nodeid = uint64(o.Inode)
+		out.Offset = o.Offset
+		out.Size = o.Length
+
+	default:
+		panic(fmt.Sprintf("Unexpected notification: %#v", op))
+	}
+
+	return
+}
+
 ////////////////////////////////////////////////////////////////////////
 // General conversions
 ////////////////////////////////////////////////////////////////////////
 
-func convertTime(t time.Time) (secs uint64, nsec uint32) {
-	totalNano := t.UnixNano()
-	secs = uint64(totalNano / 1e9)
-	nsec = uint32(totalNano % 1e9)
-	return secs, nsec
-}
-
-func convertAttributes(
-	inodeID fuseops.InodeID,
-	in *fuseops.InodeAttributes,
-	out *fusekernel.Attr) {
-	out.Ino = uint64(inodeID)
-	out.Size = in.Size
-	out.Atime, out.AtimeNsec = convertTime(in.Atime)
-	out.Mtime, out.MtimeNsec = convertTime(in.Mtime)
-	out.Ctime, out.CtimeNsec = convertTime(in.Ctime)
-	out.SetCrtime(convertTime(in.Crtime))
-	out.Nlink = in.Nlink
-	out.Uid = in.Uid
-	out.Gid = in.Gid
-	// round up to the nearest 512 boundary
-	out.Blocks = (in.Size + 512 - 1) / 512
-
-	// Set the mode.
-	out.Mode = uint32(in.Mode) & 0777
-	switch {
-	default:
-		out.Mode |= syscall.S_IFREG
-	case in.Mode&os.ModeDir != 0:
-		out.Mode |= syscall.S_IFDIR
-	case in.Mode&os.ModeDevice != 0:
-		if in.Mode&os.ModeCharDevice != 0 {
-			out.Mode |= syscall.S_IFCHR
-		} else {
-			out.Mode |= syscall.S_IFBLK
-		}
-	case in.Mode&os.ModeNamedPipe != 0:
-		out.Mode |= syscall.S_IFIFO
-	case in.Mode&os.ModeSymlink != 0:
-		out.Mode |= syscall.S_IFLNK
-	case in.Mode&os.ModeSocket != 0:
-		out.Mode |= syscall.S_IFSOCK
-	}
-	if in.Mode&os.ModeSetuid != 0 {
-		out.Mode |= syscall.S_ISUID
-	}
-	if in.Mode&os.ModeSetgid != 0 {
-		out.Mode |= syscall.S_ISGID
-	}
-	if in.Mode&os.ModeSticky != 0 {
-		out.Mode |= syscall.S_ISVTX
-	}
-}
-
-// Convert an absolute cache expiration time to a relative time from now for
-// consumption by the fuse kernel module.
-func convertExpirationTime(t time.Time) (secs uint64, nsecs uint32) {
-	// Fuse represents durations as unsigned 64-bit counts of seconds and 32-bit
-	// counts of nanoseconds (cf. http://goo.gl/EJupJV). So negative durations
-	// are right out. There is no need to cap the positive magnitude, because
-	// 2^64 seconds is well longer than the 2^63 ns range of time.Duration.
-	d := t.Sub(time.Now())
-	if d > 0 {
-		secs = uint64(d / time.Second)
-		nsecs = uint32((d % time.Second) / time.Nanosecond)
-	}
-
-	return secs, nsecs
-}
-
-func convertChildInodeEntry(
-	in *fuseops.ChildInodeEntry,
-	out *fusekernel.EntryOut) {
-	out.Nodeid = uint64(in.Child)
-	out.Generation = uint64(in.Generation)
-	out.EntryValid, out.EntryValidNsec = convertExpirationTime(in.EntryExpiration)
-	out.AttrValid, out.AttrValidNsec = convertExpirationTime(in.AttributesExpiration)
-
-	convertAttributes(in.Child, &in.Attributes, &out.Attr)
-}
-
-func convertFileMode(unixMode uint32) os.FileMode {
-	mode := os.FileMode(unixMode & 0777)
-	switch unixMode & syscall.S_IFMT {
-	case syscall.S_IFREG:
-		// nothing
-	case syscall.S_IFDIR:
-		mode |= os.ModeDir
-	case syscall.S_IFCHR:
-		mode |= os.ModeCharDevice | os.ModeDevice
-	case syscall.S_IFBLK:
-		mode |= os.ModeDevice
-	case syscall.S_IFIFO:
-		mode |= os.ModeNamedPipe
-	case syscall.S_IFLNK:
-		mode |= os.ModeSymlink
-	case syscall.S_IFSOCK:
-		mode |= os.ModeSocket
-	default:
-		// no idea
-		mode |= os.ModeDevice
-	}
-	if unixMode&syscall.S_ISUID != 0 {
-		mode |= os.ModeSetuid
-	}
-	if unixMode&syscall.S_ISGID != 0 {
-		mode |= os.ModeSetgid
-	}
-	if unixMode&syscall.S_ISVTX != 0 {
-		mode |= os.ModeSticky
-	}
-	return mode
-}
-
 func writeXattrSize(m *buffer.OutMessage, size uint32) {
 	out := (*fusekernel.GetxattrOut)(m.Grow(int(unsafe.Sizeof(fusekernel.GetxattrOut{}))))
 	out.Size = size

fuseops/conv.go

@@ -0,0 +1,84 @@
+// Copyright 2023 Vitaliy Filippov
+//
+// 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.
+
+package fuseops
+
+import (
+	"time"
+	"syscall"
+
+	"github.com/jacobsa/fuse/internal/fusekernel"
+)
+
+////////////////////////////////////////////////////////////////////////
+// General conversions
+////////////////////////////////////////////////////////////////////////
+
+func ConvertTime(t time.Time) (secs uint64, nsec uint32) {
+	totalNano := t.UnixNano()
+	secs = uint64(totalNano / 1e9)
+	nsec = uint32(totalNano % 1e9)
+	return secs, nsec
+}
+
+func ConvertAttributes(
+	inodeID InodeID,
+	in *InodeAttributes,
+	out *fusekernel.Attr) {
+	out.Ino = uint64(inodeID)
+	out.Size = in.Size
+	out.Atime, out.AtimeNsec = ConvertTime(in.Atime)
+	out.Mtime, out.MtimeNsec = ConvertTime(in.Mtime)
+	out.Ctime, out.CtimeNsec = ConvertTime(in.Ctime)
+	out.SetCrtime(ConvertTime(in.Crtime))
+	out.Nlink = in.Nlink
+	out.Uid = in.Uid
+	out.Gid = in.Gid
+	// round up to the nearest 512 boundary
+	out.Blocks = (in.Size + 512 - 1) / 512
+
+	// Set the mode.
+	out.Mode = ConvertGolangMode(in.Mode)
+
+	if out.Mode & (syscall.S_IFCHR | syscall.S_IFBLK) != 0 {
+		out.Rdev = in.Rdev
+	}
+}
+
+// Convert an absolute cache expiration time to a relative time from now for
+// consumption by the fuse kernel module.
+func ConvertExpirationTime(t time.Time) (secs uint64, nsecs uint32) {
+	// Fuse represents durations as unsigned 64-bit counts of seconds and 32-bit
+	// counts of nanoseconds (cf. http://goo.gl/EJupJV). So negative durations
+	// are right out. There is no need to cap the positive magnitude, because
+	// 2^64 seconds is well longer than the 2^63 ns range of time.Duration.
+	d := t.Sub(time.Now())
+	if d > 0 {
+		secs = uint64(d / time.Second)
+		nsecs = uint32((d % time.Second) / time.Nanosecond)
+	}
+
+	return secs, nsecs
+}
+
+func ConvertChildInodeEntry(
+	in *ChildInodeEntry,
+	out *fusekernel.EntryOut) {
+	out.Nodeid = uint64(in.Child)
+	out.Generation = uint64(in.Generation)
+	out.EntryValid, out.EntryValidNsec = ConvertExpirationTime(in.EntryExpiration)
+	out.AttrValid, out.AttrValidNsec = ConvertExpirationTime(in.AttributesExpiration)
+
+	ConvertAttributes(in.Child, &in.Attributes, &out.Attr)
+}

fuseops/filemode.go

@@ -0,0 +1,70 @@
+package fuseops
+
+import (
+	"os"
+	"syscall"
+)
+
+func ConvertFileMode(unixMode uint32) os.FileMode {
+	mode := os.FileMode(unixMode & 0777)
+	switch unixMode & syscall.S_IFMT {
+	case syscall.S_IFREG:
+		// nothing
+	case syscall.S_IFDIR:
+		mode |= os.ModeDir
+	case syscall.S_IFCHR:
+		mode |= os.ModeCharDevice | os.ModeDevice
+	case syscall.S_IFBLK:
+		mode |= os.ModeDevice
+	case syscall.S_IFIFO:
+		mode |= os.ModeNamedPipe
+	case syscall.S_IFLNK:
+		mode |= os.ModeSymlink
+	case syscall.S_IFSOCK:
+		mode |= os.ModeSocket
+	default:
+		// no idea
+	}
+	if unixMode&syscall.S_ISUID != 0 {
+		mode |= os.ModeSetuid
+	}
+	if unixMode&syscall.S_ISGID != 0 {
+		mode |= os.ModeSetgid
+	}
+	if unixMode&syscall.S_ISVTX != 0 {
+		mode |= os.ModeSticky
+	}
+	return mode
+}
+
+func ConvertGolangMode(inMode os.FileMode) uint32 {
+	outMode := uint32(inMode) & 0777
+	switch {
+	default:
+		outMode |= syscall.S_IFREG
+	case inMode&os.ModeDir != 0:
+		outMode |= syscall.S_IFDIR
+	case inMode&os.ModeDevice != 0:
+		if inMode&os.ModeCharDevice != 0 {
+			outMode |= syscall.S_IFCHR
+		} else {
+			outMode |= syscall.S_IFBLK
+		}
+	case inMode&os.ModeNamedPipe != 0:
+		outMode |= syscall.S_IFIFO
+	case inMode&os.ModeSymlink != 0:
+		outMode |= syscall.S_IFLNK
+	case inMode&os.ModeSocket != 0:
+		outMode |= syscall.S_IFSOCK
+	}
+	if inMode&os.ModeSetuid != 0 {
+		outMode |= syscall.S_ISUID
+	}
+	if inMode&os.ModeSetgid != 0 {
+		outMode |= syscall.S_ISGID
+	}
+	if inMode&os.ModeSticky != 0 {
+		outMode |= syscall.S_ISVTX
+	}
+	return outMode
+}

fuseops/ops.go

@@ -301,6 +301,9 @@ type MkNodeOp struct {
 	Name string
 	Mode os.FileMode
 
+	// The device number (only valid if created file is a device)
+	Rdev uint32
+
 	// Set by the file system: information about the inode that was created.
 	//
 	// The lookup count for the inode is implicitly incremented. See notes on
@@ -559,12 +562,18 @@ type ReadDirOp struct {
 	// offset, and return array offsets into that cached listing.
 	Offset DirOffset
 
+	// Whether this operation is a READDIRPLUS
+	//
+	// If true, then the FS must return inode attributes and expiration time
+	// along with each directory entry and increment its reference count.
+	Plus bool
+
 	// The destination buffer, whose length gives the size of the read.
 	//
 	// 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.
+	// or fuseutil.WriteDirentPlus 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
@@ -749,7 +758,18 @@ type WriteFileOp struct {
 	// be written, except on error (http://goo.gl/KUpwwn). This appears to be
 	// because it uses file mmapping machinery (http://goo.gl/SGxnaN) to write a
 	// page at a time.
-	Data      []byte
+	Data []byte
+
+	// Set by the file system: "no reuse" flag.
+	//
+	// By default, the Data buffer is reused by the library, so the file system
+	// must copy the data if it wants to use it later.
+	//
+	// However, if the file system sets this flag to true, the library doesn't
+	// reuse this buffer, so the file system can safely store and use Data slice
+	// without copying memory.
+	SuppressReuse bool
+
 	OpContext OpContext
 }
 
@@ -962,3 +982,127 @@ type FallocateOp struct {
 	Mode      uint32
 	OpContext OpContext
 }
+
+// Request notifications when the file system user calls poll/select or
+// similar operations on a file.
+type PollOp struct {
+	// The inode and handle the user wants to poll
+	Inode  InodeID
+	Handle HandleID
+
+	// Kh is the "kernel handle". The reason behind it is that it's allocated
+	// by the kernel on file allocation and guaranteed to be unique as opposed
+	// to regular file handles (HandleID) generated by the userland server
+	// (by us). Kh has to be used in NotifyPollWakeupOut replies.
+	Kh uint64
+
+	// Poll flags
+	Flags fusekernel.PollFlags
+
+	// Requested events
+	Events fusekernel.PollEvents
+
+	// Set by the file system: the actual events that have happened
+	// since the last poll
+	Revents fusekernel.PollEvents
+	OpContext OpContext
+}
+
+// Notify consumers waiting for poll/epoll that events are incoming
+// for the specified kernel handle. The kernel will send a PollOp request
+// to get the event mask after receiving this notification
+type NotifyPollWakeup struct {
+	Kh uint64
+}
+
+// Notify to invalidate cache for an inode.
+//
+// If the filesystem has writeback caching enabled, invalidating an inode
+// will first trigger a writeback of all dirty pages. The call will block
+// until all writeback requests have completed and the inode has been
+// invalidated. It will, however, not wait for completion of pending writeback
+// requests that have been issued before.
+type NotifyInvalInode struct {
+	Inode  InodeID
+	Offset int64
+	Length int64
+}
+
+// Notify to invalidate parent attributes and the dentry matching parent/name
+//
+// To avoid a deadlock this request must not be sent in the execution path
+// of a related filesytem operation or within any code that could hold a lock
+// that could be needed to execute such an operation. As of kernel 4.18, a
+// "related operation" is a lookup(), symlink(), mknod(), mkdir(), unlink(),
+// rename(), link() or create() request for the parent, and a setattr(),
+// unlink(), rmdir(), rename(), setxattr(), removexattr(), readdir() or
+// readdirplus() request for the inode itself.
+//
+// When called correctly, it will never block.
+type NotifyInvalEntry struct {
+	Parent InodeID
+	Name   string
+}
+
+// This request behaves like NotifyInvalEntry with the following additional
+// effect (at least as of Linux kernel 4.8):
+//
+// If the provided child inode matches the inode that is currently associated
+// with the cached dentry, and if there are any inotify watches registered for
+// the dentry, then the watchers are informed that the dentry has been deleted.
+//
+// To avoid a deadlock this request must not be sent while executing a
+// related filesytem operation or while holding a lock that could be needed to
+// execute such an operation.
+type NotifyDelete struct {
+	Parent InodeID
+	Child  InodeID
+	Name   string
+}
+
+// Store data to the kernel buffers
+//
+// Synchronously store data in the kernel buffers belonging to the given inode.
+// The stored data is marked up-to-date (no read will be performed against it,
+// unless it's invalidated or evicted from the cache).
+//
+// If the stored data overflows the current file size, then the size is extended,
+// similarly to a write(2) on the filesystem.
+//
+// If this request returns an error, then the store wasn't fully completed, but
+// it may have been partially completed.
+type NotifyStore struct {
+	Inode  InodeID
+	Offset uint64
+	Length uint32
+	Data   [][]byte
+}
+
+// Retrieve data from the kernel buffers belonging to the given inode
+//
+// If successful then the kernel will send a NotifyRetrieveReplyOp as a reply.
+// Only present pages are returned in the retrieve reply. Retrieving stops when it
+// finds a non-present page and only data prior to that is returned.
+//
+// If this request returns an error, then the retrieve will not be completed and
+// no reply will be sent.
+//
+// This request doesn't change the dirty state of pages in the kernel buffer. For
+// dirty pages the write() method will be called regardless of having been retrieved
+// previously.
+type NotifyRetrieve struct {
+	Inode  InodeID
+	Unique uint64
+	Offset uint64
+	Length uint32
+}
+
+// Matches the size of WriteIn
+type NotifyRetrieveReplyOp struct {
+	Inode  InodeID
+	Unique uint64
+	Offset uint64
+	Length uint32
+
+	OpContext OpContext
+}

fuseops/simple_types.go

@@ -87,6 +87,9 @@ type InodeAttributes struct {
 	//
 	Mode os.FileMode
 
+	// The device number. Only valid if the file is a device
+	Rdev uint32
+
 	// Time information. See `man 2 stat` for full details.
 	Atime  time.Time // Time of last access
 	Mtime  time.Time // Time of last modification

fuseutil/dirent.go

@@ -19,6 +19,7 @@ import (
 	"unsafe"
 
 	"github.com/jacobsa/fuse/fuseops"
+	"github.com/jacobsa/fuse/internal/fusekernel"
 )
 
 type DirentType uint32
@@ -50,10 +51,18 @@ type Dirent struct {
 	Type DirentType
 }
 
-// Write the supplied directory entry intto the given buffer in the format
-// expected in fuseops.ReadFileOp.Data, returning the number of bytes written.
+// Write the supplied directory entry into the given buffer in the format
+// expected in fuseops.ReadDirOp.Data, returning the number of bytes written.
 // Return zero if the entry would not fit.
 func WriteDirent(buf []byte, d Dirent) (n int) {
+	return WriteDirentPlus(buf, nil, d)
+}
+
+// Write the supplied directory entry and, optionally, inode entry into the
+// given buffer in the format expected in fuseops.ReadDirOp.Data with enabled
+// READDIRPLUS capability, returning the number of bytes written.
+// Returns zero if the entry would not fit.
+func WriteDirentPlus(buf []byte, e *fuseops.ChildInodeEntry, 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
@@ -78,10 +87,21 @@ func WriteDirent(buf []byte, d Dirent) (n int) {
 
 	// Do we have enough room?
 	totalLen := direntSize + len(d.Name) + padLen
+	if e != nil {
+		// READDIRPLUS was added in protocol 7.21, entry attributes were added in 7.9
+		// So here EntryOut is always full-length
+		totalLen += int(unsafe.Sizeof(fusekernel.EntryOut{}))
+	}
 	if totalLen > len(buf) {
 		return n
 	}
 
+	if e != nil {
+		out := (*fusekernel.EntryOut)(unsafe.Pointer(&buf[n]))
+		fuseops.ConvertChildInodeEntry(e, out)
+		n += int(unsafe.Sizeof(fusekernel.EntryOut{}))
+	}
+
 	// Write the header.
 	de := fuse_dirent{
 		ino:     uint64(d.Inode),

fuseutil/file_system.go

@@ -63,6 +63,9 @@ type FileSystem interface {
 	ListXattr(context.Context, *fuseops.ListXattrOp) error
 	SetXattr(context.Context, *fuseops.SetXattrOp) error
 	Fallocate(context.Context, *fuseops.FallocateOp) error
+	Poll(context.Context, *fuseops.PollOp) error
+
+	SetConnection(*fuse.Connection)
 
 	// Regard all inodes (including the root inode) as having their lookup counts
 	// decremented to zero, and clean up any resources associated with the file
@@ -95,6 +98,8 @@ type fileSystemServer struct {
 }
 
 func (s *fileSystemServer) ServeOps(c *fuse.Connection) {
+	s.fs.SetConnection(c)
+
 	// When we are done, we clean up by waiting for all in-flight ops then
 	// destroying the file system.
 	defer func() {
@@ -236,6 +241,9 @@ func (s *fileSystemServer) handleOp(
 
 	case *fuseops.FallocateOp:
 		err = s.fs.Fallocate(ctx, typed)
+
+	case *fuseops.PollOp:
+		err = s.fs.Poll(ctx, typed)
 	}
 
 	c.Reply(ctx, err)

fuseutil/not_implemented_file_system.go

@@ -204,5 +204,14 @@ func (fs *NotImplementedFileSystem) Fallocate(
 	return fuse.ENOSYS
 }
 
+func (fs *NotImplementedFileSystem) Poll(
+	ctx context.Context,
+	op *fuseops.PollOp) error {
+	return fuse.ENOSYS
+}
+
+func (fs *NotImplementedFileSystem) SetConnection(*fuse.Connection) {
+}
+
 func (fs *NotImplementedFileSystem) Destroy() {
 }

internal/buffer/in_message_linux.go

@@ -17,4 +17,4 @@ package buffer
 // The maximum fuse write request size that InMessage can acommodate.
 //
 // As of kernel 4.20 Linux accepts writes up to 256 pages or 1MiB
-const MaxWriteSize = 1 << 20
+const MaxWriteSize = 1 << 17

internal/fusekernel/fuse_kernel.go

@@ -168,7 +168,7 @@ const (
 
 // OpenAccessModeMask is a bitmask that separates the access mode
 // from the other flags in OpenFlags.
-const OpenAccessModeMask OpenFlags = syscall.O_ACCMODE
+const OpenAccessModeMask OpenFlags = OpenReadOnly | OpenWriteOnly | OpenReadWrite
 
 // OpenFlags are the O_FOO flags passed to open/create/etc calls. For
 // example, os.O_WRONLY | os.O_APPEND.
@@ -346,6 +346,34 @@ var releaseFlagNames = []flagName{
 	{uint32(ReleaseFlush), "ReleaseFlush"},
 }
 
+// Poll flags and events are used in the Poll exchange.
+type PollFlags uint32
+
+const (
+	// From the kernel source:
+	// Ask for notification if there's someone waiting for it.
+	// The client may ignore the flag and always notify.
+	PollScheduleNotify PollFlags = 1 << 0
+)
+
+type PollEvents uint32
+
+const (
+	PollInEvent     PollEvents = 0x0001
+	PollPriEvent    PollEvents = 0x0002
+	PollOutEvent    PollEvents = 0x0004
+	PollErrEvent    PollEvents = 0x0008
+	PollHupEvent    PollEvents = 0x0010
+	PollNvalEvent   PollEvents = 0x0020
+	PollRdNormEvent PollEvents = 0x0040
+	PollRdBandEvent PollEvents = 0x0080
+	PollWrNormEvent PollEvents = 0x0100
+	PollWrBandEvent PollEvents = 0x0200
+	PollMsgEvent    PollEvents = 0x0400
+	PollRemoveEvent PollEvents = 0x1000
+	PollRdHupEvent  PollEvents = 0x2000
+)
+
 // Opcodes
 const (
 	OpLookup      = 1
@@ -386,8 +414,10 @@ const (
 	OpDestroy     = 38
 	OpIoctl       = 39 // Linux?
 	OpPoll        = 40 // Linux?
+	OpNotifyReply = 41
 	OpBatchForget = 42
 	OpFallocate   = 43
+	OpReaddirplus = 44
 
 	// OS X
 	OpSetvolname = 61
@@ -552,6 +582,18 @@ func CreateInSize(p Protocol) uintptr {
 	}
 }
 
+type PollIn struct {
+	Fh     uint64
+	Kh     uint64
+	Flags  uint32
+	Events uint32
+}
+
+type PollOut struct {
+	Revents uint32
+	padding uint32
+}
+
 type ReleaseIn struct {
 	Fh           uint64
 	Flags        uint32
@@ -787,8 +829,15 @@ const (
 	NotifyCodePoll       int32 = 1
 	NotifyCodeInvalInode int32 = 2
 	NotifyCodeInvalEntry int32 = 3
+	NotifyCodeStore      int32 = 4
+	NotifyCodeRetrieve   int32 = 5
+	NotifyCodeDelete     int32 = 6
 )
 
+type NotifyPollWakeupOut struct {
+	Kh uint64
+}
+
 type NotifyInvalInodeOut struct {
 	Ino uint64
 	Off int64
@@ -800,3 +849,35 @@ type NotifyInvalEntryOut struct {
 	Namelen uint32
 	padding uint32
 }
+
+type NotifyDeleteOut struct {
+	Parent  uint64
+	Child   uint64
+	Namelen uint32
+	padding uint32
+}
+
+type NotifyStoreOut struct {
+	Nodeid  uint64
+	Offset  uint64
+	Size    uint32
+	padding uint32
+}
+
+type NotifyRetrieveOut struct {
+	Unique  uint64
+	Nodeid  uint64
+	Offset  uint64
+	Size    uint32
+	padding uint32
+}
+
+// Matches the size of WriteIn
+type NotifyRetrieveIn struct {
+	dummy1  uint64
+	Offset  uint64
+	Size    uint32
+	dummy2  uint32
+	dummy3  uint64
+	dummy4  uint64
+}

internal/fusekernel/fuse_kernel_default.go

@@ -1,3 +1,5 @@
+// +build linux windows
+
 package fusekernel
 
 import "time"

internal/fusekernel/fuse_kernel_std.go

@@ -1 +0,0 @@
-package fusekernel

mount_config.go

@@ -151,6 +151,11 @@ type MountConfig struct {
 	// OpenDir calls at all (Linux >= 5.1):
 	EnableNoOpendirSupport bool
 
+	// Tell the kernel to use READDIRPLUS.
+	// Note that the implementation may still fall back to READDIR if the running
+	// kernel doesn't have support for READDIRPLUS.
+	UseReadDirPlus bool
+
 	// Disable FUSE default permissions.
 	// This is useful for situations where the backing data store (e.g., S3) doesn't
 	// actually utilise any form of qualifiable UNIX permissions.