serve.go

// FUSE service loop, for servers that wish to use it.

package fs // import "github.com/KyleSanderson/fuse/fs"

import (
	"bytes"
	"context"
	"encoding/binary"
	"errors"
	"fmt"
	"hash/fnv"
	"io"
	"log"
	"reflect"
	"runtime"
	"strings"
	"sync"
	"syscall"
	"time"

	"github.com/KyleSanderson/fuse"
	"github.com/KyleSanderson/fuse/fuseutil"
	"golang.org/x/sys/unix"
)

const (
	attrValidTime  = 1 * time.Minute
	entryValidTime = 1 * time.Minute
)

// TODO: FINISH DOCS

// An FS is the interface required of a file system.
//
// Other FUSE requests can be handled by implementing methods from the
// FS* interfaces, for example FSStatfser.
type FS interface {
	// Root is called to obtain the Node for the file system root.
	Root() (Node, error)
}

type FSStatfser interface {
	// Statfs is called to obtain file system metadata.
	// It should write that data to resp.
	Statfs(ctx context.Context, req *fuse.StatfsRequest, resp *fuse.StatfsResponse) error
}

type FSDestroyer interface {
	// Destroy is called when the file system is shutting down.
	//
	// Linux only sends this request for block device backed (fuseblk)
	// filesystems, to allow them to flush writes to disk before the
	// unmount completes.
	Destroy()
}

type FSInodeGenerator interface {
	// GenerateInode is called to pick a dynamic inode number when it
	// would otherwise be 0.
	//
	// Not all filesystems bother tracking inodes, but FUSE requires
	// the inode to be set, and fewer duplicates in general makes UNIX
	// tools work better.
	//
	// Operations where the nodes may return 0 inodes include Getattr,
	// Setattr and ReadDir.
	//
	// If FS does not implement FSInodeGenerator, GenerateDynamicInode
	// is used.
	//
	// Implementing this is useful to e.g. constrain the range of
	// inode values used for dynamic inodes.
	//
	// Non-zero return values should be greater than 1, as that is
	// always used for the root inode.
	GenerateInode(parentInode uint64, name string) uint64
}

// A Node is the interface required of a file or directory.
// See the documentation for type FS for general information
// pertaining to all methods.
//
// A Node must be usable as a map key, that is, it cannot be a
// function, map or slice.
//
// Other FUSE requests can be handled by implementing methods from the
// Node* interfaces, for example NodeOpener.
//
// Methods returning Node should take care to return the same Node
// when the result is logically the same instance. Without this, each
// Node will get a new NodeID, causing spurious cache invalidations,
// extra lookups and aliasing anomalies. This may not matter for a
// simple, read-only filesystem.
type Node interface {
	// Attr fills attr with the standard metadata for the node.
	//
	// Fields with reasonable defaults are prepopulated. For example,
	// all times are set to a fixed moment when the program started.
	//
	// If Inode is left as 0, a dynamic inode number is chosen.
	//
	// The result may be cached for the duration set in Valid.
	Attr(ctx context.Context, attr *fuse.Attr) error
}

type NodeGetattrer interface {
	// Getattr obtains the standard metadata for the receiver.
	// It should store that metadata in resp.
	//
	// If this method is not implemented, the attributes will be
	// generated based on Attr(), with zero values filled in.
	Getattr(ctx context.Context, req *fuse.GetattrRequest, resp *fuse.GetattrResponse) error
}

type HandleGetattrer interface {
	Getattr(ctx context.Context, req *fuse.GetattrRequest, resp *fuse.GetattrResponse) error
}

type NodeSetattrer interface {
	// Setattr sets the standard metadata for the receiver.
	//
	// Note, this is also used to communicate changes in the size of
	// the file, outside of Writes.
	//
	// req.Valid is a bitmask of what fields are actually being set.
	// For example, the method should not change the mode of the file
	// unless req.Valid.Mode() is true.
	Setattr(ctx context.Context, req *fuse.SetattrRequest, resp *fuse.SetattrResponse) error
}

type HandleSetattrer interface {
	Setattr(ctx context.Context, req *fuse.SetattrRequest, resp *fuse.SetattrResponse) error
}

type NodeSymlinker interface {
	// Symlink creates a new symbolic link in the receiver, which must be a directory.
	//
	// TODO is the above true about directories?
	Symlink(ctx context.Context, req *fuse.SymlinkRequest) (Node, error)
}

// This optional request will be called only for symbolic link nodes.
type NodeReadlinker interface {
	// Readlink reads a symbolic link.
	Readlink(ctx context.Context, req *fuse.ReadlinkRequest) (string, error)
}

type NodeLinker interface {
	// Link creates a new directory entry in the receiver based on an
	// existing Node. Receiver must be a directory.
	Link(ctx context.Context, req *fuse.LinkRequest, old Node) (Node, error)
}

type NodeRemover interface {
	// Remove removes the entry with the given name from
	// the receiver, which must be a directory.  The entry to be removed
	// may correspond to a file (unlink) or to a directory (rmdir).
	Remove(ctx context.Context, req *fuse.RemoveRequest) error
}

type NodeAccesser interface {
	// Access checks whether the calling context has permission for
	// the given operations on the receiver. If so, Access should
	// return nil. If not, Access should return EPERM.
	//
	// Note that this call affects the result of the access(2) system
	// call but not the open(2) system call. If Access is not
	// implemented, the Node behaves as if it always returns nil
	// (permission granted), relying on checks in Open instead.
	Access(ctx context.Context, req *fuse.AccessRequest) error
}

type NodeStringLookuper interface {
	// Lookup looks up a specific entry in the receiver,
	// which must be a directory.  Lookup should return a Node
	// corresponding to the entry.  If the name does not exist in
	// the directory, Lookup should return ENOENT.
	//
	// Lookup need not to handle the names "." and "..".
	Lookup(ctx context.Context, name string) (Node, error)
}

type NodeRequestLookuper interface {
	// Lookup looks up a specific entry in the receiver.
	// See NodeStringLookuper for more.
	Lookup(ctx context.Context, req *fuse.LookupRequest, resp *fuse.LookupResponse) (Node, error)
}

type NodeMkdirer interface {
	Mkdir(ctx context.Context, req *fuse.MkdirRequest) (Node, error)
}

type NodeOpener interface {
	// Open opens the receiver. After a successful open, a client
	// process has a file descriptor referring to this Handle.
	//
	// Open can also be also called on non-files. For example,
	// directories are Opened for ReadDir or fchdir(2).
	//
	// If this method is not implemented, the open will always
	// succeed, and the Node itself will be used as the Handle.
	//
	// XXX note about access.  XXX OpenFlags.
	Open(ctx context.Context, req *fuse.OpenRequest, resp *fuse.OpenResponse) (Handle, error)
}

type NodeCreater interface {
	// Create creates a new directory entry in the receiver, which
	// must be a directory.
	Create(ctx context.Context, req *fuse.CreateRequest, resp *fuse.CreateResponse) (Node, Handle, error)
}

type NodeForgetter interface {
	// Forget about this node. This node will not receive further
	// method calls.
	//
	// Forget is not necessarily seen on unmount, as all nodes are
	// implicitly forgotten as part of the unmount.
	Forget()
}

type NodeRenamer interface {
	Rename(ctx context.Context, req *fuse.RenameRequest, newDir Node) error
}

type NodeMknoder interface {
	Mknod(ctx context.Context, req *fuse.MknodRequest) (Node, error)
}

// TODO this should be on Handle not Node
type NodeFsyncer interface {
	Fsync(ctx context.Context, req *fuse.FsyncRequest) error
}

type HandleFsyncer interface {
	Fsync(ctx context.Context, req *fuse.FsyncRequest) error
}

type NodeGetxattrer interface {
	// Getxattr gets an extended attribute by the given name from the
	// node.
	//
	// If there is no xattr by that name, returns fuse.ErrNoXattr.
	Getxattr(ctx context.Context, req *fuse.GetxattrRequest, resp *fuse.GetxattrResponse) error
}

type NodeListxattrer interface {
	// Listxattr lists the extended attributes recorded for the node.
	Listxattr(ctx context.Context, req *fuse.ListxattrRequest, resp *fuse.ListxattrResponse) error
}

type NodeSetxattrer interface {
	// Setxattr sets an extended attribute with the given name and
	// value for the node.
	Setxattr(ctx context.Context, req *fuse.SetxattrRequest) error
}

type NodeRemovexattrer interface {
	// Removexattr removes an extended attribute for the name.
	//
	// If there is no xattr by that name, returns fuse.ErrNoXattr.
	Removexattr(ctx context.Context, req *fuse.RemovexattrRequest) error
}

var startTime = time.Now()

func nodeAttr(ctx context.Context, n Node, attr *fuse.Attr) error {
	attr.Valid = attrValidTime
	attr.Nlink = 1
	attr.Atime = startTime
	attr.Mtime = startTime
	attr.Ctime = startTime
	if err := n.Attr(ctx, attr); err != nil {
		return err
	}
	return nil
}

// A Handle is the interface required of an opened file or directory.
// See the documentation for type FS for general information
// pertaining to all methods.
//
// Other FUSE requests can be handled by implementing methods from the
// Handle* interfaces. The most common to implement are HandleReader,
// HandleReadDirer, and HandleWriter.
//
// TODO implement methods: Getlk, Setlk, Setlkw
type Handle interface {
}

type HandleFlusher interface {
	// Flush is called each time the file or directory is closed.
	// Because there can be multiple file descriptors referring to a
	// single opened file, Flush can be called multiple times.
	Flush(ctx context.Context, req *fuse.FlushRequest) error
}

type HandleReadAller interface {
	ReadAll(ctx context.Context) ([]byte, error)
}

type HandleReadDirAller interface {
	ReadDirAll(ctx context.Context) ([]fuse.Dirent, error)
}

type HandleReader interface {
	// Read requests to read data from the handle.
	//
	// There is a page cache in the kernel that normally submits only
	// page-aligned reads spanning one or more pages. However, you
	// should not rely on this. To see individual requests as
	// submitted by the file system clients, set OpenDirectIO.
	//
	// Note that reads beyond the size of the file as reported by Attr
	// are not even attempted (except in OpenDirectIO mode).
	Read(ctx context.Context, req *fuse.ReadRequest, resp *fuse.ReadResponse) error
}

type HandleWriter interface {
	// Write requests to write data into the handle at the given offset.
	// Store the amount of data written in resp.Size.
	//
	// There is a writeback page cache in the kernel that normally submits
	// only page-aligned writes spanning one or more pages. However,
	// you should not rely on this. To see individual requests as
	// submitted by the file system clients, set OpenDirectIO.
	//
	// Writes that grow the file are expected to update the file size
	// (as seen through Attr). Note that file size changes are
	// communicated also through Setattr.
	Write(ctx context.Context, req *fuse.WriteRequest, resp *fuse.WriteResponse) error
}

type HandleReleaser interface {
	Release(ctx context.Context, req *fuse.ReleaseRequest) error
}

type HandlePoller interface {
	// Poll checks whether the handle is currently ready for I/O, and
	// may request a wakeup when it is.
	//
	// Poll should always return quickly. Clients waiting for
	// readiness can be woken up by passing the return value of
	// PollRequest.Wakeup to fs.Server.NotifyPollWakeup or
	// fuse.Conn.NotifyPollWakeup.
	//
	// To allow supporting poll for only some of your Nodes/Handles,
	// the default behavior is to report immediate readiness. If your
	// FS does not support polling and you want to minimize needless
	// requests and log noise, implement NodePoller and return
	// syscall.ENOSYS.
	//
	// The Go runtime uses epoll-based I/O whenever possible, even for
	// regular files.
	Poll(ctx context.Context, req *fuse.PollRequest, resp *fuse.PollResponse) error
}

type NodePoller interface {
	// Poll checks whether the node is currently ready for I/O, and
	// may request a wakeup when it is. See HandlePoller.
	Poll(ctx context.Context, req *fuse.PollRequest, resp *fuse.PollResponse) error
}

// HandleLocker contains the common operations for all kinds of file
// locks. See also lock family specific interfaces: HandleFlockLocker,
// HandlePOSIXLocker.
type HandleLocker interface {
	// Lock tries to acquire a lock on a byte range of the node. If a
	// conflicting lock is already held, returns syscall.EAGAIN.
	//
	// LockRequest.LockOwner is a file-unique identifier for this
	// lock, and will be seen in calls releasing this lock
	// (UnlockRequest, ReleaseRequest, FlushRequest) and also
	// in e.g. ReadRequest, WriteRequest.
	Lock(ctx context.Context, req *fuse.LockRequest) error

	// LockWait acquires a lock on a byte range of the node, waiting
	// until the lock can be obtained (or context is canceled).
	LockWait(ctx context.Context, req *fuse.LockWaitRequest) error

	// Unlock releases the lock on a byte range of the node. Locks can
	// be released also implicitly, see HandleFlockLocker and
	// HandlePOSIXLocker.
	Unlock(ctx context.Context, req *fuse.UnlockRequest) error

	// QueryLock returns the current state of locks held for the byte
	// range of the node.
	//
	// See QueryLockRequest for details on how to respond.
	//
	// To simplify implementing this method, resp.Lock is prefilled to
	// have Lock.Type F_UNLCK, and the whole struct should be
	// overwritten for in case of conflicting locks.
	QueryLock(ctx context.Context, req *fuse.QueryLockRequest, resp *fuse.QueryLockResponse) error
}

// HandleFlockLocker describes locking behavior unique to flock (BSD)
// locks. See HandleLocker.
type HandleFlockLocker interface {
	HandleLocker

	// Flock unlocking can also happen implicitly as part of Release,
	// in which case Unlock is not called, and Release will have
	// ReleaseFlags bit ReleaseFlockUnlock set.
	HandleReleaser
}

// HandlePOSIXLocker describes locking behavior unique to POSIX (fcntl
// F_SETLK) locks. See HandleLocker.
type HandlePOSIXLocker interface {
	HandleLocker

	// POSIX unlocking can also happen implicitly as part of Flush,
	// in which case Unlock is not called.
	HandleFlusher
}

type Config struct {
	// Function to send debug log messages to. If nil, use fuse.Debug.
	// Note that changing this or fuse.Debug may not affect existing
	// calls to Serve.
	//
	// See fuse.Debug for the rules that log functions must follow.
	Debug func(msg interface{})

	// Function to put things into context for processing the request.
	// The returned context must have ctx as its parent.
	//
	// Note that changing this may not affect existing calls to Serve.
	//
	// Must not retain req.
	WithContext func(ctx context.Context, req fuse.Request) context.Context
}

// New returns a new FUSE server ready to serve this kernel FUSE
// connection.
//
// Config may be nil.
func New(conn *fuse.Conn, config *Config) *Server {
	s := &Server{
		conn:         conn,
		req:          map[fuse.RequestID]*serveRequest{},
		nodeRef:      map[Node]fuse.NodeID{},
		notifyWait:   map[fuse.RequestID]chan<- *fuse.NotifyReply{},
		dynamicInode: GenerateDynamicInode,
	}
	if config != nil {
		s.debug = config.Debug
		s.context = config.WithContext
	}
	if s.debug == nil {
		s.debug = fuse.Debug
	}
	return s
}

type Server struct {
	// set in New
	conn    *fuse.Conn
	debug   func(msg interface{})
	context func(ctx context.Context, req fuse.Request) context.Context

	// set once at Serve time
	fs           FS
	dynamicInode func(parent uint64, name string) uint64

	// state, protected by meta
	meta       sync.Mutex
	req        map[fuse.RequestID]*serveRequest
	node       []*serveNode
	nodeRef    map[Node]fuse.NodeID
	handle     []*serveHandle
	freeNode   []fuse.NodeID
	freeHandle []fuse.HandleID
	nodeGen    uint64

	// pending notify upcalls to kernel
	notifyMu   sync.Mutex
	notifySeq  fuse.RequestID
	notifyWait map[fuse.RequestID]chan<- *fuse.NotifyReply

	// Used to ensure worker goroutines finish before Serve returns
	wg sync.WaitGroup
}

// Serve serves the FUSE connection by making calls to the methods
// of fs and the Nodes and Handles it makes available.  It returns only
// when the connection has been closed or an unexpected error occurs.
func (s *Server) Serve(fs FS) error {
	defer s.wg.Wait() // Wait for worker goroutines to complete before return

	s.fs = fs
	if dyn, ok := fs.(FSInodeGenerator); ok {
		s.dynamicInode = dyn.GenerateInode
	}

	root, err := fs.Root()
	if err != nil {
		return fmt.Errorf("cannot obtain root node: %v", err)
	}
	// Recognize the root node if it's ever returned from Lookup,
	// passed to Invalidate, etc.
	s.nodeRef[root] = 1
	s.node = append(s.node, nil, &serveNode{
		inode:      1,
		generation: s.nodeGen,
		node:       root,
		refs:       1,
	})
	s.handle = append(s.handle, nil)

	for {
		req, err := s.conn.ReadRequest()
		if err != nil {
			if err == io.EOF {
				break
			}
			return err
		}

		s.wg.Add(1)
		go func() {
			defer s.wg.Done()
			s.serve(req)
		}()
	}
	return nil
}

// Serve serves a FUSE connection with the default settings. See
// Server.Serve.
func Serve(c *fuse.Conn, fs FS) error {
	server := New(c, nil)
	return server.Serve(fs)
}

type serveRequest struct {
	Request fuse.Request
	cancel  func()
}

type serveNode struct {
	inode      uint64
	generation uint64
	node       Node
	refs       uint64

	// Delay freeing the NodeID until waitgroup is done. This allows
	// using the NodeID for short periods of time without holding the
	// Server.meta lock.
	//
	// Rules:
	//
	//     - hold Server.meta while calling wg.Add, then unlock
	//     - do NOT try to reacquire Server.meta
	wg sync.WaitGroup
}

func (sn *serveNode) attr(ctx context.Context, attr *fuse.Attr) error {
	err := nodeAttr(ctx, sn.node, attr)
	if attr.Inode == 0 {
		attr.Inode = sn.inode
	}
	return err
}

type serveHandle struct {
	handle   Handle
	readData []byte
}

func (c *Server) saveNode(inode uint64, node Node) (id fuse.NodeID, gen uint64) {
	c.meta.Lock()
	defer c.meta.Unlock()

	if id, ok := c.nodeRef[node]; ok {
		sn := c.node[id]
		sn.refs++
		return id, sn.generation
	}

	sn := &serveNode{inode: inode, node: node, refs: 1}
	if n := len(c.freeNode); n > 0 {
		id = c.freeNode[n-1]
		c.freeNode = c.freeNode[:n-1]
		c.node[id] = sn
		c.nodeGen++
	} else {
		id = fuse.NodeID(len(c.node))
		c.node = append(c.node, sn)
	}
	sn.generation = c.nodeGen
	c.nodeRef[node] = id
	return id, sn.generation
}

func (c *Server) saveHandle(handle Handle) (id fuse.HandleID) {
	c.meta.Lock()
	shandle := &serveHandle{handle: handle}
	if n := len(c.freeHandle); n > 0 {
		id = c.freeHandle[n-1]
		c.freeHandle = c.freeHandle[:n-1]
		c.handle[id] = shandle
	} else {
		id = fuse.HandleID(len(c.handle))
		c.handle = append(c.handle, shandle)
	}
	c.meta.Unlock()
	return
}

type nodeRefcountDropBug struct {
	N    uint64
	Refs uint64
	Node fuse.NodeID
}

func (n nodeRefcountDropBug) String() string {
	return fmt.Sprintf("bug: trying to drop %d of %d references to %v", n.N, n.Refs, n.Node)
}

// dropNode decreases reference count for node with id by n.
// If reference count dropped to zero, returns true.
// Note that node is not guaranteed to be non-nil.
func (c *Server) dropNode(id fuse.NodeID, n uint64) (node Node, forget bool) {
	c.meta.Lock()
	defer c.meta.Unlock()
	snode := c.node[id]

	if snode == nil {
		// this should only happen if refcounts kernel<->us disagree
		// *and* two ForgetRequests for the same node race each other;
		// this indicates a bug somewhere
		c.debug(nodeRefcountDropBug{N: n, Node: id})

		// we may end up triggering Forget twice, but that's better
		// than not even once, and that's the best we can do
		return nil, true
	}

	if n > snode.refs {
		c.debug(nodeRefcountDropBug{N: n, Refs: snode.refs, Node: id})
		n = snode.refs
	}

	snode.refs -= n
	if snode.refs == 0 {
		snode.wg.Wait()
		c.node[id] = nil
		delete(c.nodeRef, snode.node)
		c.freeNode = append(c.freeNode, id)
		return snode.node, true
	}
	return nil, false
}

func (c *Server) dropHandle(id fuse.HandleID) {
	c.meta.Lock()
	c.handle[id] = nil
	c.freeHandle = append(c.freeHandle, id)
	c.meta.Unlock()
}

type missingHandle struct {
	Handle    fuse.HandleID
	MaxHandle fuse.HandleID
}

func (m missingHandle) String() string {
	return fmt.Sprint("missing handle: ", m.Handle, m.MaxHandle)
}

// Returns nil for invalid handles.
func (c *Server) getHandle(id fuse.HandleID) (shandle *serveHandle) {
	c.meta.Lock()
	defer c.meta.Unlock()
	if id < fuse.HandleID(len(c.handle)) {
		shandle = c.handle[uint(id)]
	}
	if shandle == nil {
		c.debug(missingHandle{
			Handle:    id,
			MaxHandle: fuse.HandleID(len(c.handle)),
		})
	}
	return
}

type request struct {
	In interface{} `json:",omitempty"`
}

func (r request) String() string {
	return fmt.Sprintf("<- %s", r.In)
}

type logResponseHeader struct {
	ID fuse.RequestID
}

func (m logResponseHeader) String() string {
	return fmt.Sprintf("ID=%v", m.ID)
}

type response struct {
	Op      string
	Request logResponseHeader
	Out     interface{} `json:",omitempty"`
	// Errno contains the errno value as a string, for example "EPERM".
	Errno string `json:",omitempty"`
	// Error may contain a free form error message.
	Error string `json:",omitempty"`
}

func (r response) errstr() string {
	s := r.Errno
	if r.Error != "" {
		// prefix the errno constant to the long form message
		s = s + ": " + r.Error
	}
	return s
}

func (r response) String() string {
	switch {
	case r.Errno != "" && r.Out != nil:
		return fmt.Sprintf("-> [%v] %v error=%s", r.Request, r.Out, r.errstr())
	case r.Errno != "":
		return fmt.Sprintf("-> [%v] %s error=%s", r.Request, r.Op, r.errstr())
	case r.Out != nil:
		// make sure (seemingly) empty values are readable
		switch r.Out.(type) {
		case string:
			return fmt.Sprintf("-> [%v] %s %q", r.Request, r.Op, r.Out)
		case []byte:
			return fmt.Sprintf("-> [%v] %s [% x]", r.Request, r.Op, r.Out)
		default:
			return fmt.Sprintf("-> [%v] %v", r.Request, r.Out)
		}
	default:
		return fmt.Sprintf("-> [%v] %s", r.Request, r.Op)
	}
}

type notification struct {
	Op   string
	Node fuse.NodeID
	Out  interface{} `json:",omitempty"`
	Err  string      `json:",omitempty"`
}

func (n notification) String() string {
	var buf bytes.Buffer
	fmt.Fprintf(&buf, "=> %s %v", n.Op, n.Node)
	if n.Out != nil {
		// make sure (seemingly) empty values are readable
		switch n.Out.(type) {
		case string:
			fmt.Fprintf(&buf, " %q", n.Out)
		case []byte:
			fmt.Fprintf(&buf, " [% x]", n.Out)
		default:
			fmt.Fprintf(&buf, " %s", n.Out)
		}
	}
	if n.Err != "" {
		fmt.Fprintf(&buf, " Err:%v", n.Err)
	}
	return buf.String()
}

type notificationRequest struct {
	ID   fuse.RequestID
	Op   string
	Node fuse.NodeID
	Out  interface{} `json:",omitempty"`
}

func (n notificationRequest) String() string {
	var buf bytes.Buffer
	fmt.Fprintf(&buf, ">> %s [ID=%d] %v", n.Op, n.ID, n.Node)
	if n.Out != nil {
		// make sure (seemingly) empty values are readable
		switch n.Out.(type) {
		case string:
			fmt.Fprintf(&buf, " %q", n.Out)
		case []byte:
			fmt.Fprintf(&buf, " [% x]", n.Out)
		default:
			fmt.Fprintf(&buf, " %s", n.Out)
		}
	}
	return buf.String()
}

type notificationResponse struct {
	ID  fuse.RequestID
	Op  string
	In  interface{} `json:",omitempty"`
	Err string      `json:",omitempty"`
}

func (n notificationResponse) String() string {
	var buf bytes.Buffer
	fmt.Fprintf(&buf, "<< [ID=%d] %s", n.ID, n.Op)
	if n.In != nil {
		// make sure (seemingly) empty values are readable
		switch n.In.(type) {
		case string:
			fmt.Fprintf(&buf, " %q", n.In)
		case []byte:
			fmt.Fprintf(&buf, " [% x]", n.In)
		default:
			fmt.Fprintf(&buf, " %s", n.In)
		}
	}
	if n.Err != "" {
		fmt.Fprintf(&buf, " Err:%v", n.Err)
	}
	return buf.String()
}

type logMissingNode struct {
	MaxNode fuse.NodeID
}

func opName(req fuse.Request) string {
	t := reflect.Indirect(reflect.ValueOf(req)).Type()
	s := t.Name()
	s = strings.TrimSuffix(s, "Request")
	return s
}

type logLinkRequestOldNodeNotFound struct {
	Request *fuse.Header
	In      *fuse.LinkRequest
}

func (m *logLinkRequestOldNodeNotFound) String() string {
	return fmt.Sprintf("In LinkRequest (request %v), node %d not found", m.Request.Hdr().ID, m.In.OldNode)
}

type renameNewDirNodeNotFound struct {
	Request *fuse.Header
	In      *fuse.RenameRequest
}

func (m *renameNewDirNodeNotFound) String() string {
	return fmt.Sprintf("In RenameRequest (request %v), node %d not found", m.Request.Hdr().ID, m.In.NewDir)
}

type handlerPanickedError struct {
	Request interface{}
	Err     interface{}
}

var _ error = handlerPanickedError{}

func (h handlerPanickedError) Error() string {
	return fmt.Sprintf("handler panicked: %v", h.Err)
}

var _ fuse.ErrorNumber = handlerPanickedError{}

func (h handlerPanickedError) Errno() fuse.Errno {
	if err, ok := h.Err.(fuse.ErrorNumber); ok {
		return err.Errno()
	}
	return fuse.DefaultErrno
}

// handlerTerminatedError happens when a handler terminates itself
// with runtime.Goexit. This is most commonly because of incorrect use
// of testing.TB.FailNow, typically via t.Fatal.
type handlerTerminatedError struct {
	Request interface{}
}

var _ error = handlerTerminatedError{}

func (h handlerTerminatedError) Error() string {
	return fmt.Sprintf("handler terminated (called runtime.Goexit)")
}

var _ fuse.ErrorNumber = handlerTerminatedError{}

func (h handlerTerminatedError) Errno() fuse.Errno {
	return fuse.DefaultErrno
}

type handleNotReaderError struct {
	handle Handle
}

var _ error = handleNotReaderError{}

func (e handleNotReaderError) Error() string {
	return fmt.Sprintf("handle has no Read: %T", e.handle)
}

var _ fuse.ErrorNumber = handleNotReaderError{}

func (e handleNotReaderError) Errno() fuse.Errno {
	return fuse.Errno(syscall.ENOTSUP)
}

func initLookupResponse(s *fuse.LookupResponse) {
	s.EntryValid = entryValidTime
}

type logDuplicateRequestID struct {
	New fuse.Request
	Old fuse.Request
}

func (m *logDuplicateRequestID) String() string {
	return fmt.Sprintf("Duplicate request: new %v, old %v", m.New, m.Old)
}

func (c *Server) serve(r fuse.Request) {
	ctx, cancel := context.WithCancel(context.Background())
	defer cancel()
	parentCtx := ctx
	if c.context != nil {
		ctx = c.context(ctx, r)
	}

	req := &serveRequest{Request: r, cancel: cancel}

	switch r.(type) {
	case *fuse.NotifyReply:
		// don't log NotifyReply here, they're logged by the recipient
		// as soon as we have decoded them to the right types
	default:
		c.debug(request{
			In: r,
		})
	}
	var node Node
	var snode *serveNode
	c.meta.Lock()
	hdr := r.Hdr()
	if id := hdr.Node; id != 0 {
		if id < fuse.NodeID(len(c.node)) {
			snode = c.node[uint(id)]
		}
		if snode == nil {
			c.meta.Unlock()
			err := syscall.ESTALE
			c.debug(response{
				Op:      opName(r),
				Request: logResponseHeader{ID: hdr.ID},
				Error:   fuse.Errno(err).ErrnoName(),
				// this is the only place that sets both Error and
				// Out; not sure if i want to do that; might get rid
				// of len(c.node) things altogether
				Out: logMissingNode{
					MaxNode: fuse.NodeID(len(c.node)),
				},
			})
			r.RespondError(err)
			return
		}
		node = snode.node
	}
	if old, found := c.req[hdr.ID]; found {
		c.debug(logDuplicateRequestID{
			New: req.Request,
			Old: old.Request,
		})
	}
	c.req[hdr.ID] = req
	c.meta.Unlock()

	// Call this before responding.
	// After responding is too late: we might get another request
	// with the same ID and be very confused.
	done := func(resp interface{}) {
		msg := response{
			Op:      opName(r),
			Request: logResponseHeader{ID: hdr.ID},
		}
		if err, ok := resp.(error); ok {
			errno := fuse.ToErrno(err)
			msg.Errno = errno.ErrnoName()
			if errno != err && syscall.Errno(errno) != err {
				// if it's more than just a fuse.Errno or a
				// syscall.Errno, log extra detail
				msg.Error = err.Error()
			}
		} else {
			msg.Out = resp
		}
		c.debug(msg)

		c.meta.Lock()
		delete(c.req, hdr.ID)
		c.meta.Unlock()
	}

	var responded bool
	defer func() {
		if rec := recover(); rec != nil {
			const size = 1 << 16
			buf := make([]byte, size)
			n := runtime.Stack(buf, false)
			buf = buf[:n]
			log.Printf("fuse: panic in handler for %v: %v\n%s", r, rec, buf)
			err := handlerPanickedError{
				Request: r,
				Err:     rec,
			}
			done(err)
			r.RespondError(err)
			return
		}

		if !responded {
			err := handlerTerminatedError{
				Request: r,
			}
			done(err)
			r.RespondError(err)
		}
	}()

	if err := c.handleRequest(ctx, node, snode, r, done); err != nil {
		if err == context.Canceled {
			select {
			case <-parentCtx.Done():
				// We canceled the parent context because of an
				// incoming interrupt request, so return EINTR
				// to trigger the right behavior in the client app.
				//
				// Only do this when it's the parent context that was
				// canceled, not a context controlled by the program
				// using this library, so we don't return EINTR too
				// eagerly -- it might cause busy loops.
				//
				// Decent write-up on role of EINTR:
				// http://250bpm.com/blog:12
				err = syscall.EINTR
			default:
				// nothing
			}
		}
		done(err)
		r.RespondError(err)
	}

	// disarm runtime.Goexit protection
	responded = true
}

// handleRequest will either a) call done(s) and r.Respond(s) OR b) return an error.
func (c *Server) handleRequest(ctx context.Context, node Node, snode *serveNode, r fuse.Request, done func(resp interface{})) error {
	switch r := r.(type) {
	default:
		// Note: To FUSE, ENOSYS means "this server never implements this request."
		// It would be inappropriate to return ENOSYS for other operations in this
		// switch that might only be unavailable in some contexts, not all.
		return syscall.ENOSYS

	case *fuse.StatfsRequest:
		s := &fuse.StatfsResponse{}
		if fs, ok := c.fs.(FSStatfser); ok {
			if err := fs.Statfs(ctx, r, s); err != nil {
				return err
			}
		}
		done(s)
		r.Respond(s)
		return nil

	// Node operations.
	case *fuse.GetattrRequest:
		s := &fuse.GetattrResponse{}
		shandle := c.getHandle(r.Handle)
		if shandle != nil {
			if h, ok := shandle.handle.(HandleGetattrer); ok {
				if err := h.Getattr(ctx, r, s); err != nil {
					return err
				}
				if s.Attr.Inode == 0 {
					s.Attr.Inode = c.dynamicInode(snode.inode, r.Request)
				}
				done(s)
				r.Respond(s)
				return nil
			}
		}
		if n, ok := node.(NodeGetattrer); ok {
			if err := n.Getattr(ctx, r, s); err != nil {
				return err
			}
		} else {
			if err := snode.attr(ctx, &s.Attr); err != nil {
				return err
			}
		}
		if s.Attr.Inode == 0 {
			s.Attr.Inode = c.dynamicInode(snode.inode, r.Request)
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.SetattrRequest:
		s := &fuse.SetattrResponse{}
		shandle := c.getHandle(r.Handle)
		if shandle != nil {
			if h, ok := shandle.handle.(HandleSetattrer); ok {
				if err := h.Setattr(ctx, r, s); err != nil {
					return err
				}
				if err := snode.attr(ctx, &s.Attr); err != nil {
					return err
				}

				if s.Attr.Inode == 0 {
					s.Attr.Inode = c.dynamicInode(snode.inode, r.Request)
				}
				done(s)
				r.Respond(s)
				return nil
			}
		}

		if n, ok := node.(NodeSetattrer); ok {
			if err := n.Setattr(ctx, r, s); err != nil {
				return err
			}
		}

		if err := snode.attr(ctx, &s.Attr); err != nil {
			return err
		}

		if s.Attr.Inode == 0 {
			s.Attr.Inode = c.dynamicInode(snode.inode, r.Request)
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.SymlinkRequest:
		s := &fuse.SymlinkResponse{}
		initLookupResponse(&s.LookupResponse)
		n, ok := node.(NodeSymlinker)
		if !ok {
			return syscall.EIO // XXX or EPERM like Mkdir?
		}
		n2, err := n.Symlink(ctx, r)
		if err != nil {
			return err
		}
		if err := c.saveLookup(ctx, &s.LookupResponse, snode, r.NewName, n2); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.ReadlinkRequest:
		n, ok := node.(NodeReadlinker)
		if !ok {
			return syscall.EIO /// XXX or EPERM?
		}
		target, err := n.Readlink(ctx, r)
		if err != nil {
			return err
		}
		done(target)
		r.Respond(target)
		return nil

	case *fuse.LinkRequest:
		n, ok := node.(NodeLinker)
		if !ok {
			return syscall.EIO /// XXX or EPERM?
		}
		c.meta.Lock()
		var oldNode *serveNode
		if int(r.OldNode) < len(c.node) {
			oldNode = c.node[r.OldNode]
		}
		c.meta.Unlock()
		if oldNode == nil {
			c.debug(logLinkRequestOldNodeNotFound{
				Request: r.Hdr(),
				In:      r,
			})
			return syscall.EIO
		}
		n2, err := n.Link(ctx, r, oldNode.node)
		if err != nil {
			return err
		}
		s := &fuse.LookupResponse{}
		initLookupResponse(s)
		if err := c.saveLookup(ctx, s, snode, r.NewName, n2); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.RemoveRequest:
		n, ok := node.(NodeRemover)
		if !ok {
			return syscall.EIO /// XXX or EPERM?
		}
		err := n.Remove(ctx, r)
		if err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.AccessRequest:
		if n, ok := node.(NodeAccesser); ok {
			if err := n.Access(ctx, r); err != nil {
				return err
			}
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.LookupRequest:
		var n2 Node
		var err error
		s := &fuse.LookupResponse{}
		initLookupResponse(s)
		if n, ok := node.(NodeStringLookuper); ok {
			n2, err = n.Lookup(ctx, r.Name)
		} else if n, ok := node.(NodeRequestLookuper); ok {
			n2, err = n.Lookup(ctx, r, s)
		} else {
			return syscall.ENOENT
		}
		if err != nil {
			return err
		}
		if err := c.saveLookup(ctx, s, snode, r.Name, n2); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.MkdirRequest:
		s := &fuse.MkdirResponse{}
		initLookupResponse(&s.LookupResponse)
		n, ok := node.(NodeMkdirer)
		if !ok {
			return syscall.EPERM
		}
		n2, err := n.Mkdir(ctx, r)
		if err != nil {
			return err
		}
		if err := c.saveLookup(ctx, &s.LookupResponse, snode, r.Name, n2); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.OpenRequest:
		s := &fuse.OpenResponse{}
		var h2 Handle
		if n, ok := node.(NodeOpener); ok {
			hh, err := n.Open(ctx, r, s)
			if err != nil {
				return err
			}
			h2 = hh
		} else {
			h2 = node
		}
		s.Handle = c.saveHandle(h2)
		done(s)
		r.Respond(s)
		return nil

	case *fuse.CreateRequest:
		n, ok := node.(NodeCreater)
		if !ok {
			// If we send back ENOSYS, FUSE will try mknod+open.
			return syscall.EPERM
		}
		s := &fuse.CreateResponse{OpenResponse: fuse.OpenResponse{}}
		initLookupResponse(&s.LookupResponse)
		n2, h2, err := n.Create(ctx, r, s)
		if err != nil {
			return err
		}
		if err := c.saveLookup(ctx, &s.LookupResponse, snode, r.Name, n2); err != nil {
			return err
		}
		s.Handle = c.saveHandle(h2)
		done(s)
		r.Respond(s)
		return nil

	case *fuse.GetxattrRequest:
		n, ok := node.(NodeGetxattrer)
		if !ok {
			return syscall.ENOTSUP
		}
		s := &fuse.GetxattrResponse{}
		err := n.Getxattr(ctx, r, s)
		if err != nil {
			return err
		}
		if r.Size != 0 && uint64(len(s.Xattr)) > uint64(r.Size) {
			return syscall.ERANGE
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.ListxattrRequest:
		n, ok := node.(NodeListxattrer)
		if !ok {
			return syscall.ENOTSUP
		}
		s := &fuse.ListxattrResponse{}
		err := n.Listxattr(ctx, r, s)
		if err != nil {
			return err
		}
		if r.Size != 0 && uint64(len(s.Xattr)) > uint64(r.Size) {
			return syscall.ERANGE
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.SetxattrRequest:
		n, ok := node.(NodeSetxattrer)
		if !ok {
			return syscall.ENOTSUP
		}
		err := n.Setxattr(ctx, r)
		if err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.RemovexattrRequest:
		n, ok := node.(NodeRemovexattrer)
		if !ok {
			return syscall.ENOTSUP
		}
		err := n.Removexattr(ctx, r)
		if err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.ForgetRequest:
		_, forget := c.dropNode(r.Hdr().Node, r.N)
		if forget {
			n, ok := node.(NodeForgetter)
			if ok {
				n.Forget()
			}
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.BatchForgetRequest:
		// BatchForgetRequest is hard to unit test, as it
		// fundamentally relies on something unprivileged userspace
		// has little control over. A root-only, Linux-only test could
		// be written with `echo 2 >/proc/sys/vm/drop_caches`, but
		// that would still rely on timing, the number of batches and
		// operation spread over them could vary, it wouldn't run in a
		// typical container regardless of privileges, and it would
		// degrade performance for the rest of the machine. It would
		// still probably be worth doing, just not the most fun.

		// node is nil here because BatchForget as a message is not
		// aimed at a any one node
		for _, item := range r.Forget {
			node, forget := c.dropNode(item.NodeID, item.N)
			// node can be nil here if kernel vs our refcount were out
			// of sync and multiple Forgets raced each other
			if node == nil {
				// nothing we can do about that
				continue
			}
			if forget {
				n, ok := node.(NodeForgetter)
				if ok {
					n.Forget()
				}
			}
		}
		done(nil)
		r.Respond()
		return nil

	// Handle operations.
	case *fuse.ReadRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		handle := shandle.handle

		s := &fuse.ReadResponse{Data: make([]byte, 0, r.Size)}
		if r.Dir {
			if h, ok := handle.(HandleReadDirAller); ok {
				// detect rewinddir(3) or similar seek and refresh
				// contents
				if r.Offset == 0 {
					shandle.readData = nil
				}

				if shandle.readData == nil {
					dirs, err := h.ReadDirAll(ctx)
					if err != nil {
						return err
					}
					var data []byte
					for _, dir := range dirs {
						if dir.Inode == 0 {
							dir.Inode = c.dynamicInode(snode.inode, dir.Name)
						}
						data = fuse.AppendDirent(data, dir)
					}
					shandle.readData = data
				}
				fuseutil.HandleRead(r, s, shandle.readData)
				done(s)
				r.Respond(s)
				return nil
			}
		} else {
			if h, ok := handle.(HandleReadAller); ok {
				if shandle.readData == nil {
					data, err := h.ReadAll(ctx)
					if err != nil {
						return err
					}
					if data == nil {
						data = []byte{}
					}
					shandle.readData = data
				}
				fuseutil.HandleRead(r, s, shandle.readData)
				done(s)
				r.Respond(s)
				return nil
			}
			h, ok := handle.(HandleReader)
			if !ok {
				err := handleNotReaderError{handle: handle}
				return err
			}
			if err := h.Read(ctx, r, s); err != nil {
				return err
			}
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.WriteRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}

		s := &fuse.WriteResponse{}
		if h, ok := shandle.handle.(HandleWriter); ok {
			if err := h.Write(ctx, r, s); err != nil {
				return err
			}
			done(s)
			r.Respond(s)
			return nil
		}
		return syscall.EIO

	case *fuse.FlushRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		handle := shandle.handle

		if h, ok := handle.(HandleFlusher); ok {
			if err := h.Flush(ctx, r); err != nil {
				return err
			}
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.ReleaseRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		handle := shandle.handle

		// No matter what, release the handle.
		c.dropHandle(r.Handle)

		if h, ok := handle.(HandleReleaser); ok {
			if err := h.Release(ctx, r); err != nil {
				return err
			}
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.DestroyRequest:
		if fs, ok := c.fs.(FSDestroyer); ok {
			fs.Destroy()
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.RenameRequest:
		c.meta.Lock()
		var newDirNode *serveNode
		if int(r.NewDir) < len(c.node) {
			newDirNode = c.node[r.NewDir]
		}
		c.meta.Unlock()
		if newDirNode == nil {
			c.debug(renameNewDirNodeNotFound{
				Request: r.Hdr(),
				In:      r,
			})
			return syscall.EIO
		}
		n, ok := node.(NodeRenamer)
		if !ok {
			return syscall.EIO // XXX or EPERM like Mkdir?
		}
		err := n.Rename(ctx, r, newDirNode.node)
		if err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.MknodRequest:
		n, ok := node.(NodeMknoder)
		if !ok {
			return syscall.EIO
		}
		n2, err := n.Mknod(ctx, r)
		if err != nil {
			return err
		}
		s := &fuse.LookupResponse{}
		initLookupResponse(s)
		if err := c.saveLookup(ctx, s, snode, r.Name, n2); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

	case *fuse.FsyncRequest:
		if shandle := c.getHandle(r.Handle); shandle != nil {
			if h, ok := shandle.handle.(HandleFsyncer); ok {
				err := h.Fsync(ctx, r)
				if err != nil {
					return err
				}
				done(nil)
				r.Respond()
				return nil
			}
		}

		n, ok := node.(NodeFsyncer)
		if !ok {
			return syscall.EIO
		}
		err := n.Fsync(ctx, r)
		if err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.InterruptRequest:
		c.meta.Lock()
		ireq := c.req[r.IntrID]
		if ireq != nil && ireq.cancel != nil {
			ireq.cancel()
			ireq.cancel = nil
		}
		c.meta.Unlock()
		done(nil)
		r.Respond()
		return nil

	case *fuse.PollRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		s := &fuse.PollResponse{}

		if h, ok := shandle.handle.(HandlePoller); ok {
			if err := h.Poll(ctx, r, s); err != nil {
				return err
			}
			done(s)
			r.Respond(s)
			return nil
		}

		if n, ok := node.(NodePoller); ok {
			if err := n.Poll(ctx, r, s); err != nil {
				return err
			}
			done(s)
			r.Respond(s)
			return nil
		}

		// fallback to always claim ready
		s.REvents = fuse.DefaultPollMask
		done(s)
		r.Respond(s)
		return nil

	case *fuse.NotifyReply:
		c.notifyMu.Lock()
		w, ok := c.notifyWait[r.Hdr().ID]
		if ok {
			delete(c.notifyWait, r.Hdr().ID)
		}
		c.notifyMu.Unlock()
		if !ok {
			c.debug(notificationResponse{
				ID:  r.Hdr().ID,
				Op:  "NotifyReply",
				Err: "unknown ID",
			})
			return nil
		}
		w <- r
		return nil

	case *fuse.LockRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		h, ok := shandle.handle.(HandleLocker)
		if !ok {
			return syscall.ENOTSUP
		}
		if err := h.Lock(ctx, r); err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.LockWaitRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		h, ok := shandle.handle.(HandleLocker)
		if !ok {
			return syscall.ENOTSUP
		}
		if err := h.LockWait(ctx, r); err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.UnlockRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		h, ok := shandle.handle.(HandleLocker)
		if !ok {
			return syscall.ENOTSUP
		}
		if err := h.Unlock(ctx, r); err != nil {
			return err
		}
		done(nil)
		r.Respond()
		return nil

	case *fuse.QueryLockRequest:
		shandle := c.getHandle(r.Handle)
		if shandle == nil {
			return syscall.ESTALE
		}
		h, ok := shandle.handle.(HandleLocker)
		if !ok {
			return syscall.ENOTSUP
		}
		s := &fuse.QueryLockResponse{
			Lock: fuse.FileLock{
				Type: unix.F_UNLCK,
			},
		}
		if err := h.QueryLock(ctx, r, s); err != nil {
			return err
		}
		done(s)
		r.Respond(s)
		return nil

		/*	case *FsyncdirRequest:
				return ENOSYS

			case *BmapRequest:
				return ENOSYS
		*/
	}

	panic("not reached")
}

func (c *Server) saveLookup(ctx context.Context, s *fuse.LookupResponse, snode *serveNode, elem string, n2 Node) error {
	if err := nodeAttr(ctx, n2, &s.Attr); err != nil {
		return err
	}
	if s.Attr.Inode == 0 {
		s.Attr.Inode = c.dynamicInode(snode.inode, elem)
	}

	s.Node, s.Generation = c.saveNode(s.Attr.Inode, n2)
	return nil
}

type invalidateNodeDetail struct {
	Off  int64
	Size int64
}

func (i invalidateNodeDetail) String() string {
	return fmt.Sprintf("Off:%d Size:%d", i.Off, i.Size)
}

func errstr(err error) string {
	if err == nil {
		return ""
	}
	return err.Error()
}

func (s *Server) invalidateNode(node Node, off int64, size int64) error {
	s.meta.Lock()
	id, ok := s.nodeRef[node]
	if ok {
		snode := s.node[id]
		snode.wg.Add(1)
		defer snode.wg.Done()
	}
	s.meta.Unlock()
	if !ok {
		// This is what the kernel would have said, if we had been
		// able to send this message; it's not cached.
		return fuse.ErrNotCached
	}
	// Delay logging until after we can record the error too. We
	// consider a /dev/fuse write to be instantaneous enough to not
	// need separate before and after messages.
	err := s.conn.InvalidateNode(id, off, size)
	s.debug(notification{
		Op:   "InvalidateNode",
		Node: id,
		Out: invalidateNodeDetail{
			Off:  off,
			Size: size,
		},
		Err: errstr(err),
	})
	return err
}

// InvalidateNodeAttr invalidates the kernel cache of the attributes
// of node.
//
// Returns fuse.ErrNotCached if the kernel is not currently caching
// the node.
func (s *Server) InvalidateNodeAttr(node Node) error {
	return s.invalidateNode(node, 0, 0)
}

// InvalidateNodeData invalidates the kernel cache of the attributes
// and data of node.
//
// Returns fuse.ErrNotCached if the kernel is not currently caching
// the node.
func (s *Server) InvalidateNodeData(node Node) error {
	return s.invalidateNode(node, 0, -1)
}

// InvalidateNodeDataRange invalidates the kernel cache of the
// attributes and a range of the data of node.
//
// Returns fuse.ErrNotCached if the kernel is not currently caching
// the node.
func (s *Server) InvalidateNodeDataRange(node Node, off int64, size int64) error {
	return s.invalidateNode(node, off, size)
}

type invalidateEntryDetail struct {
	Name string
}

func (i invalidateEntryDetail) String() string {
	return fmt.Sprintf("%q", i.Name)
}

// InvalidateEntry invalidates the kernel cache of the directory entry
// identified by parent node and entry basename.
//
// Kernel may or may not cache directory listings. To invalidate
// those, use InvalidateNode to invalidate all of the data for a
// directory. (As of 2015-06, Linux FUSE does not cache directory
// listings.)
//
// Returns ErrNotCached if the kernel is not currently caching the
// node.
func (s *Server) InvalidateEntry(parent Node, name string) error {
	s.meta.Lock()
	id, ok := s.nodeRef[parent]
	if ok {
		snode := s.node[id]
		snode.wg.Add(1)
		defer snode.wg.Done()
	}
	s.meta.Unlock()
	if !ok {
		// This is what the kernel would have said, if we had been
		// able to send this message; it's not cached.
		return fuse.ErrNotCached
	}
	err := s.conn.InvalidateEntry(id, name)
	s.debug(notification{
		Op:   "InvalidateEntry",
		Node: id,
		Out: invalidateEntryDetail{
			Name: name,
		},
		Err: errstr(err),
	})
	return err
}

type notifyStoreRetrieveDetail struct {
	Off  uint64
	Size uint64
}

func (i notifyStoreRetrieveDetail) String() string {
	return fmt.Sprintf("Off:%d Size:%d", i.Off, i.Size)
}

type notifyRetrieveReplyDetail struct {
	Size uint64
}

func (i notifyRetrieveReplyDetail) String() string {
	return fmt.Sprintf("Size:%d", i.Size)
}

// NotifyStore puts data into the kernel page cache.
//
// Returns fuse.ErrNotCached if the kernel is not currently caching
// the node.
func (s *Server) NotifyStore(node Node, offset uint64, data []byte) error {
	s.meta.Lock()
	id, ok := s.nodeRef[node]
	if ok {
		snode := s.node[id]
		snode.wg.Add(1)
		defer snode.wg.Done()
	}
	s.meta.Unlock()
	if !ok {
		// This is what the kernel would have said, if we had been
		// able to send this message; it's not cached.
		return fuse.ErrNotCached
	}
	// Delay logging until after we can record the error too. We
	// consider a /dev/fuse write to be instantaneous enough to not
	// need separate before and after messages.
	err := s.conn.NotifyStore(id, offset, data)
	s.debug(notification{
		Op:   "NotifyStore",
		Node: id,
		Out: notifyStoreRetrieveDetail{
			Off:  offset,
			Size: uint64(len(data)),
		},
		Err: errstr(err),
	})
	return err
}

// NotifyRetrieve gets data from the kernel page cache.
//
// Returns fuse.ErrNotCached if the kernel is not currently caching
// the node.
func (s *Server) NotifyRetrieve(node Node, offset uint64, size uint32) ([]byte, error) {
	s.meta.Lock()
	id, ok := s.nodeRef[node]
	if ok {
		snode := s.node[id]
		snode.wg.Add(1)
		defer snode.wg.Done()
	}
	s.meta.Unlock()
	if !ok {
		// This is what the kernel would have said, if we had been
		// able to send this message; it's not cached.
		return nil, fuse.ErrNotCached
	}

	ch := make(chan *fuse.NotifyReply, 1)
	s.notifyMu.Lock()
	const wraparoundThreshold = 1 << 63
	if s.notifySeq > wraparoundThreshold {
		s.notifyMu.Unlock()
		return nil, errors.New("running out of notify sequence numbers")
	}
	s.notifySeq++
	seq := s.notifySeq
	s.notifyWait[seq] = ch
	s.notifyMu.Unlock()

	s.debug(notificationRequest{
		ID:   seq,
		Op:   "NotifyRetrieve",
		Node: id,
		Out: notifyStoreRetrieveDetail{
			Off:  offset,
			Size: uint64(size),
		},
	})
	retrieval, err := s.conn.NotifyRetrieve(seq, id, offset, size)
	if err != nil {
		s.debug(notificationResponse{
			ID:  seq,
			Op:  "NotifyRetrieve",
			Err: errstr(err),
		})
		return nil, err
	}

	reply := <-ch
	data := retrieval.Finish(reply)
	s.debug(notificationResponse{
		ID: seq,
		Op: "NotifyRetrieve",
		In: notifyRetrieveReplyDetail{
			Size: uint64(len(data)),
		},
	})
	return data, nil
}

func (s *Server) NotifyPollWakeup(wakeup fuse.PollWakeup) error {
	// Delay logging until after we can record the error too. We
	// consider a /dev/fuse write to be instantaneous enough to not
	// need separate before and after messages.
	err := s.conn.NotifyPollWakeup(wakeup)
	s.debug(notification{
		Op:  "NotifyPollWakeup",
		Out: wakeup,
		Err: errstr(err),
	})
	return err
}

// DataHandle returns a read-only Handle that satisfies reads
// using the given data.
func DataHandle(data []byte) Handle {
	return &dataHandle{data}
}

type dataHandle struct {
	data []byte
}

func (d *dataHandle) ReadAll(ctx context.Context) ([]byte, error) {
	return d.data, nil
}

// GenerateDynamicInode returns a dynamic inode.
//
// The parent inode and current entry name are used as the criteria
// for choosing a pseudorandom inode. This makes it likely the same
// entry will get the same inode on multiple runs.
func GenerateDynamicInode(parent uint64, name string) uint64 {
	h := fnv.New64a()
	var buf [8]byte
	binary.LittleEndian.PutUint64(buf[:], parent)
	_, _ = h.Write(buf[:])
	_, _ = h.Write([]byte(name))
	var inode uint64
	for {
		inode = h.Sum64()
		if inode > 1 {
			break
		}
		// there's a tiny probability that result is zero or the
		// hardcoded root inode 1; change the input a little and try
		// again
		_, _ = h.Write([]byte{'x'})
	}
	return inode
}