fusego/ops.go

// Copyright 2015 Google Inc. All Rights Reserved.
//
// 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 (
	"fmt"
	"os"
	"time"
	"unsafe"

	"github.com/jacobsa/fuse/internal/buffer"
	"github.com/jacobsa/fuse/internal/fusekernel"
	"golang.org/x/net/context"
)

// A common interface implemented by all ops in this package. Use a type switch
// to find particular concrete types, responding with fuse.ENOSYS if a type is
// not supported.
type Op interface {
	// A short description of the op, to be used in logging.
	ShortDesc() string

	// A long description of the op, to be used in debug logging.
	DebugString() string

	// A context that can be used for long-running operations.
	Context() context.Context

	// Repond to the operation with the supplied error. If there is no error, set
	// any necessary output fields and then call Respond(nil). The user must not
	// call with a nil error for unrecognized ops; instead, use ENOSYS.
	//
	// Once this is invoked, the user must exclude any further calls to any
	// method of this op.
	Respond(error)

	// Log information tied to this operation, with semantics equivalent to
	// log.Printf, except that the format is different and logging is suppressed
	// if no debug logger was set when mounting.
	Logf(format string, v ...interface{})
}

////////////////////////////////////////////////////////////////////////
// Inodes
////////////////////////////////////////////////////////////////////////

// Look up a child by name within a parent directory. The kernel sends this
// when resolving user paths to dentry structs, which are then cached.
type LookUpInodeOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The ID of the directory inode to which the child belongs.
	Parent InodeID

	// The name of the child of interest, relative to the parent. For example, in
	// this directory structure:
	//
	//     foo/
	//         bar/
	//             baz
	//
	// the file system may receive a request to look up the child named "bar" for
	// the parent foo/.
	Name string

	// The resulting entry. Must be filled out by the file system.
	//
	// The lookup count for the inode is implicitly incremented. See notes on
	// ForgetInodeOp for more information.
	Entry ChildInodeEntry
}

func (o *LookUpInodeOp) ShortDesc() (desc string) {
	desc = fmt.Sprintf("LookUpInode(parent=%v, name=%q)", o.Parent, o.Name)
	return
}

func (o *LookUpInodeOp) kernelResponse() (b buffer.OutMessage) {
	size := fusekernel.EntryOutSize(o.protocol)
	b = buffer.NewOutMessage(size)
	out := (*fusekernel.EntryOut)(b.Grow(size))
	convertChildInodeEntry(&o.Entry, out)

	return
}

// Refresh the attributes for an inode whose ID was previously returned in a
// LookUpInodeOp. The kernel sends this when the FUSE VFS layer's cache of
// inode attributes is stale. This is controlled by the AttributesExpiration
// field of ChildInodeEntry, etc.
type GetInodeAttributesOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The inode of interest.
	Inode InodeID

	// Set by the file system: attributes for the inode, and the time at which
	// they should expire. See notes on ChildInodeEntry.AttributesExpiration for
	// more.
	Attributes           InodeAttributes
	AttributesExpiration time.Time
}

func (o *GetInodeAttributesOp) DebugString() string {
	return fmt.Sprintf(
		"Inode: %d, Exp: %v, Attr: %s",
		o.Inode,
		o.AttributesExpiration,
		o.Attributes.DebugString())
}

func (o *GetInodeAttributesOp) kernelResponse() (b buffer.OutMessage) {
	size := fusekernel.AttrOutSize(o.protocol)
	b = buffer.NewOutMessage(size)
	out := (*fusekernel.AttrOut)(b.Grow(size))
	out.AttrValid, out.AttrValidNsec = convertExpirationTime(o.AttributesExpiration)
	convertAttributes(o.Inode, &o.Attributes, &out.Attr)

	return
}

// Change attributes for an inode.
//
// The kernel sends this for obvious cases like chmod(2), and for less obvious
// cases like ftrunctate(2).
type SetInodeAttributesOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The inode of interest.
	Inode InodeID

	// The attributes to modify, or nil for attributes that don't need a change.
	Size  *uint64
	Mode  *os.FileMode
	Atime *time.Time
	Mtime *time.Time

	// Set by the file system: the new attributes for the inode, and the time at
	// which they should expire. See notes on
	// ChildInodeEntry.AttributesExpiration for more.
	Attributes           InodeAttributes
	AttributesExpiration time.Time
}

func (o *SetInodeAttributesOp) kernelResponse() (b buffer.OutMessage) {
	size := fusekernel.AttrOutSize(o.protocol)
	b = buffer.NewOutMessage(size)
	out := (*fusekernel.AttrOut)(b.Grow(size))
	out.AttrValid, out.AttrValidNsec = convertExpirationTime(o.AttributesExpiration)
	convertAttributes(o.Inode, &o.Attributes, &out.Attr)

	return
}

// Decrement the reference count for an inode ID previously issued by the file
// system.
//
// The comments for the ops that implicitly increment the reference count
// contain a note of this (but see also the note about the root inode below).
// For example, LookUpInodeOp and MkDirOp. The authoritative source is the
// libfuse documentation, which states that any op that returns
// fuse_reply_entry fuse_reply_create implicitly increments (cf.
// http://goo.gl/o5C7Dx).
//
// If the reference count hits zero, the file system can forget about that ID
// entirely, and even re-use it in future responses. The kernel guarantees that
// it will not otherwise use it again.
//
// The reference count corresponds to fuse_inode::nlookup
// (http://goo.gl/ut48S4). Some examples of where the kernel manipulates it:
//
//  *  (http://goo.gl/vPD9Oh) Any caller to fuse_iget increases the count.
//  *  (http://goo.gl/B6tTTC) fuse_lookup_name calls fuse_iget.
//  *  (http://goo.gl/IlcxWv) fuse_create_open calls fuse_iget.
//  *  (http://goo.gl/VQMQul) fuse_dentry_revalidate increments after
//     revalidating.
//
// In contrast to all other inodes, RootInodeID begins with an implicit
// lookup count of one, without a corresponding op to increase it. (There
// could be no such op, because the root cannot be referred to by name.) Code
// walk:
//
//  *  (http://goo.gl/gWAheU) fuse_fill_super calls fuse_get_root_inode.
//
//  *  (http://goo.gl/AoLsbb) fuse_get_root_inode calls fuse_iget without
//     sending any particular request.
//
//  *  (http://goo.gl/vPD9Oh) fuse_iget increments nlookup.
//
// File systems should tolerate but not rely on receiving forget ops for
// remaining inodes when the file system unmounts, including the root inode.
// Rather they should take fuse.Connection.ReadOp returning io.EOF as
// implicitly decrementing all lookup counts to zero.
type ForgetInodeOp struct {
	commonOp

	// The inode whose reference count should be decremented.
	Inode InodeID

	// The amount to decrement the reference count.
	N uint64
}

func (o *ForgetInodeOp) kernelResponse() (b buffer.OutMessage) {
	// No response.
	return
}

////////////////////////////////////////////////////////////////////////
// Inode creation
////////////////////////////////////////////////////////////////////////

// Create a directory inode as a child of an existing directory inode. The
// kernel sends this in response to a mkdir(2) call.
//
// The Linux kernel appears to verify the name doesn't already exist (mkdir
// calls mkdirat calls user_path_create calls filename_create, which verifies:
// http://goo.gl/FZpLu5). Indeed, the tests in samples/memfs that call in
// parallel appear to bear this out. But osxfuse does not appear to guarantee
// this (cf. https://goo.gl/PqzZDv). And if names may be created outside of the
// kernel's control, it doesn't matter what the kernel does anyway.
//
// Therefore the file system should return EEXIST if the name already exists.
type MkDirOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The ID of parent directory inode within which to create the child.
	Parent InodeID

	// The name of the child to create, and the mode with which to create it.
	Name string
	Mode os.FileMode

	// Set by the file system: information about the inode that was created.
	//
	// The lookup count for the inode is implicitly incremented. See notes on
	// ForgetInodeOp for more information.
	Entry ChildInodeEntry
}

func (o *MkDirOp) ShortDesc() (desc string) {
	desc = fmt.Sprintf("MkDir(parent=%v, name=%q)", o.Parent, o.Name)
	return
}

func (o *MkDirOp) kernelResponse() (b buffer.OutMessage) {
	size := fusekernel.EntryOutSize(o.protocol)
	b = buffer.NewOutMessage(size)
	out := (*fusekernel.EntryOut)(b.Grow(size))
	convertChildInodeEntry(&o.Entry, out)

	return
}

// Create a file inode and open it.
//
// The kernel sends this when the user asks to open a file with the O_CREAT
// flag and the kernel has observed that the file doesn't exist. (See for
// example lookup_open, http://goo.gl/PlqE9d). However, osxfuse doesn't appear
// to make this check atomically (cf. https://goo.gl/PqzZDv). And if names may
// be created outside of the kernel's control, it doesn't matter what the
// kernel does anyway.
//
// Therefore the file system should return EEXIST if the name already exists.
type CreateFileOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The ID of parent directory inode within which to create the child file.
	Parent InodeID

	// The name of the child to create, and the mode with which to create it.
	Name string
	Mode os.FileMode

	// Set by the file system: information about the inode that was created.
	//
	// The lookup count for the inode is implicitly incremented. See notes on
	// ForgetInodeOp for more information.
	Entry ChildInodeEntry

	// Set by the file system: an opaque ID that will be echoed in follow-up
	// calls for this file using the same struct file in the kernel. In practice
	// this usually means follow-up calls using the file descriptor returned by
	// open(2).
	//
	// The handle may be supplied in future ops like ReadFileOp that contain a
	// file handle. The file system must ensure this ID remains valid until a
	// later call to ReleaseFileHandle.
	Handle HandleID
}

func (o *CreateFileOp) ShortDesc() (desc string) {
	desc = fmt.Sprintf("CreateFile(parent=%v, name=%q)", o.Parent, o.Name)
	return
}

func (o *CreateFileOp) kernelResponse() (b buffer.OutMessage) {
	eSize := fusekernel.EntryOutSize(o.protocol)
	b = buffer.NewOutMessage(eSize + unsafe.Sizeof(fusekernel.OpenOut{}))

	e := (*fusekernel.EntryOut)(b.Grow(eSize))
	convertChildInodeEntry(&o.Entry, e)

	oo := (*fusekernel.OpenOut)(b.Grow(unsafe.Sizeof(fusekernel.OpenOut{})))
	oo.Fh = uint64(o.Handle)

	return
}

// Create a symlink inode. If the name already exists, the file system should
// return EEXIST (cf. the notes on CreateFileOp and MkDirOp).
type CreateSymlinkOp struct {
	commonOp
	protocol fusekernel.Protocol

	// The ID of parent directory inode within which to create the child symlink.
	Parent InodeID

	// The name of the symlink to create.
	Name string

	// The target of the symlink.
	Target string

	// Set by the file system: information about the symlink inode that was
	// created.
	//
	// The lookup count for the inode is implicitly incremented. See notes on
	// ForgetInodeOp for more information.
	Entry ChildInodeEntry
}

func (o *CreateSymlinkOp) ShortDesc() (desc string) {
	desc = fmt.Sprintf(
		"CreateSymlink(parent=%v, name=%q, target=%q)",
		o.Parent,
		o.Name,
		o.Target)

	return
}

func (o *CreateSymlinkOp) kernelResponse() (b buffer.OutMessage) {
	size := fusekernel.EntryOutSize(o.protocol)
	b = buffer.NewOutMessage(size)
	out := (*fusekernel.EntryOut)(b.Grow(size))
	convertChildInodeEntry(&o.Entry, out)

	return
}

////////////////////////////////////////////////////////////////////////
// Unlinking
////////////////////////////////////////////////////////////////////////

// Rename a file or directory, given the IDs of the original parent directory
// and the new one (which may be the same).
//
// In Linux, this is called by vfs_rename (https://goo.gl/eERItT), which is
// called by sys_renameat2 (https://goo.gl/fCC9qC).
//
// The kernel takes care of ensuring that the source and destination are not
// identical (in which case it does nothing), that the rename is not across
// file system boundaries, and that the destination doesn't already exist with
// the wrong type. Some subtleties that the file system must care about:
//
//  *  If the new name is an existing directory, the file system must ensure it
//     is empty before replacing it, returning ENOTEMPTY otherwise. (This is
//     per the posix spec: http://goo.gl/4XtT79)
//
//  *  The rename must be atomic from the point of view of an observer of the
//     new name. That is, if the new name already exists, there must be no
//     point at which it doesn't exist.
//
//  *  It is okay for the new name to be modified before the old name is
//     removed; these need not be atomic. In fact, the Linux man page
//     explicitly says this is likely (cf. https://goo.gl/Y1wVZc).
//
//  *  Linux bends over backwards (https://goo.gl/pLDn3r) to ensure that
//     neither the old nor the new parent can be concurrently modified. But
//     it's not clear whether OS X does this, and in any case it doesn't matter
//     for file systems that may be modified remotely. Therefore a careful file
//     system implementor should probably ensure if possible that the unlink
//     step in the "link new name, unlink old name" process doesn't unlink a
//     different inode than the one that was linked to the new name. Still,
//     posix and the man pages are imprecise about the actual semantics of a
//     rename if it's not atomic, so it is probably not disastrous to be loose
//     about this.
//
type RenameOp struct {
	commonOp

	// The old parent directory, and the name of the entry within it to be
	// relocated.
	OldParent InodeID
	OldName   string

	// The new parent directory, and the name of the entry to be created or
	// overwritten within it.
	NewParent InodeID
	NewName   string
}

func (o *RenameOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

// Unlink a directory from its parent. Because directories cannot have a link
// count above one, this means the directory inode should be deleted as well
// once the kernel sends ForgetInodeOp.
//
// The file system is responsible for checking that the directory is empty.
//
// Sample implementation in ext2: ext2_rmdir (http://goo.gl/B9QmFf)
type RmDirOp struct {
	commonOp

	// The ID of parent directory inode, and the name of the directory being
	// removed within it.
	Parent InodeID
	Name   string
}

func (o *RmDirOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

// Unlink a file or symlink from its parent. If this brings the inode's link
// count to zero, the inode should be deleted once the kernel sends
// ForgetInodeOp. It may still be referenced before then if a user still has
// the file open.
//
// Sample implementation in ext2: ext2_unlink (http://goo.gl/hY6r6C)
type UnlinkOp struct {
	commonOp

	// The ID of parent directory inode, and the name of the entry being removed
	// within it.
	Parent InodeID
	Name   string
}

func (o *UnlinkOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

////////////////////////////////////////////////////////////////////////
// Directory handles
////////////////////////////////////////////////////////////////////////

// Open a directory inode.
//
// On Linux the sends this when setting up a struct file for a particular inode
// with type directory, usually in response to an open(2) call from a
// user-space process. On OS X it may not be sent for every open(2) (cf.
// https://github.com/osxfuse/osxfuse/issues/199).
type OpenDirOp struct {
	commonOp

	// The ID of the inode to be opened.
	Inode InodeID

	// Set by the file system: an opaque ID that will be echoed in follow-up
	// calls for this directory using the same struct file in the kernel. In
	// practice this usually means follow-up calls using the file descriptor
	// returned by open(2).
	//
	// The handle may be supplied in future ops like ReadDirOp that contain a
	// directory handle. The file system must ensure this ID remains valid until
	// a later call to ReleaseDirHandle.
	Handle HandleID
}

func (o *OpenDirOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(unsafe.Sizeof(fusekernel.OpenOut{}))
	out := (*fusekernel.OpenOut)(b.Grow(unsafe.Sizeof(fusekernel.OpenOut{})))
	out.Fh = uint64(o.Handle)

	return
}

// Read entries from a directory previously opened with OpenDir.
type ReadDirOp struct {
	commonOp

	// The directory inode that we are reading, and the handle previously
	// returned by OpenDir when opening that inode.
	Inode  InodeID
	Handle HandleID

	// The offset within the directory at which to read.
	//
	// Warning: this field is not necessarily a count of bytes. Its legal values
	// are defined by the results returned in ReadDirResponse. See the notes
	// below and the notes on that struct.
	//
	// In the Linux kernel this ultimately comes from file::f_pos, which starts
	// at zero and is set by llseek and by the final consumed result returned by
	// each call to ReadDir:
	//
	//  *  (http://goo.gl/2nWJPL) iterate_dir, which is called by getdents(2) and
	//     readdir(2), sets dir_context::pos to file::f_pos before calling
	//     f_op->iterate, and then does the opposite assignment afterward.
	//
	//  *  (http://goo.gl/rTQVSL) fuse_readdir, which implements iterate for fuse
	//     directories, passes dir_context::pos as the offset to fuse_read_fill,
	//     which passes it on to user-space. fuse_readdir later calls
	//     parse_dirfile with the same context.
	//
	//  *  (http://goo.gl/vU5ukv) For each returned result (except perhaps the
	//     last, which may be truncated by the page boundary), parse_dirfile
	//     updates dir_context::pos with fuse_dirent::off.
	//
	// It is affected by the Posix directory stream interfaces in the following
	// manner:
	//
	//  *  (http://goo.gl/fQhbyn, http://goo.gl/ns1kDF) opendir initially causes
	//     filepos to be set to zero.
	//
	//  *  (http://goo.gl/ezNKyR, http://goo.gl/xOmDv0) readdir allows the user
	//     to iterate through the directory one entry at a time. As each entry is
	//     consumed, its d_off field is stored in __dirstream::filepos.
	//
	//  *  (http://goo.gl/WEOXG8, http://goo.gl/rjSXl3) telldir allows the user
	//     to obtain the d_off field from the most recently returned entry.
	//
	//  *  (http://goo.gl/WG3nDZ, http://goo.gl/Lp0U6W) seekdir allows the user
	//     to seek backward to an offset previously returned by telldir. It
	//     stores the new offset in filepos, and calls llseek to update the
	//     kernel's struct file.
	//
	//  *  (http://goo.gl/gONQhz, http://goo.gl/VlrQkc) rewinddir allows the user
	//     to go back to the beginning of the directory, obtaining a fresh view.
	//     It updates filepos and calls llseek to update the kernel's struct
	//     file.
	//
	// Unfortunately, FUSE offers no way to intercept seeks
	// (http://goo.gl/H6gEXa), so there is no way to cause seekdir or rewinddir
	// to fail. Additionally, there is no way to distinguish an explicit
	// rewinddir followed by readdir from the initial readdir, or a rewinddir
	// from a seekdir to the value returned by telldir just after opendir.
	//
	// Luckily, Posix is vague about what the user will see if they seek
	// backwards, and requires the user not to seek to an old offset after a
	// rewind. The only requirement on freshness is that rewinddir results in
	// something that looks like a newly-opened directory. So FUSE file systems
	// may e.g. cache an entire fresh listing for each ReadDir with a zero
	// 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 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.
	//
	// 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.
	//
	// An empty buffer indicates the end of the directory has been reached.
	Data []byte
}

func (o *ReadDirOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(uintptr(len(o.Data)))
	b.Append(o.Data)
	return
}

// Release a previously-minted directory handle. The kernel sends this when
// there are no more references to an open directory: all file descriptors are
// closed and all memory mappings are unmapped.
//
// The kernel guarantees that the handle ID will not be used in further ops
// sent to the file system (unless it is reissued by the file system).
//
// Errors from this op are ignored by the kernel (cf. http://goo.gl/RL38Do).
type ReleaseDirHandleOp struct {
	commonOp

	// The handle ID to be released. The kernel guarantees that this ID will not
	// be used in further calls to the file system (unless it is reissued by the
	// file system).
	Handle HandleID
}

func (o *ReleaseDirHandleOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

////////////////////////////////////////////////////////////////////////
// File handles
////////////////////////////////////////////////////////////////////////

// Open a file inode.
//
// On Linux the sends this when setting up a struct file for a particular inode
// with type file, usually in response to an open(2) call from a user-space
// process. On OS X it may not be sent for every open(2)
// (cf.https://github.com/osxfuse/osxfuse/issues/199).
type OpenFileOp struct {
	commonOp

	// The ID of the inode to be opened.
	Inode InodeID

	// An opaque ID that will be echoed in follow-up calls for this file using
	// the same struct file in the kernel. In practice this usually means
	// follow-up calls using the file descriptor returned by open(2).
	//
	// The handle may be supplied in future ops like ReadFileOp that contain a
	// file handle. The file system must ensure this ID remains valid until a
	// later call to ReleaseFileHandle.
	Handle HandleID
}

func (o *OpenFileOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(unsafe.Sizeof(fusekernel.OpenOut{}))
	out := (*fusekernel.OpenOut)(b.Grow(unsafe.Sizeof(fusekernel.OpenOut{})))
	out.Fh = uint64(o.Handle)

	return
}

// Read data from a file previously opened with CreateFile or OpenFile.
//
// Note that this op is not sent for every call to read(2) by the end user;
// some reads may be served by the page cache. See notes on WriteFileOp for
// more.
type ReadFileOp struct {
	commonOp

	// The file inode that we are reading, and the handle previously returned by
	// CreateFile or OpenFile when opening that inode.
	Inode  InodeID
	Handle HandleID

	// The range of the file to 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
	// (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
}

func (o *ReadFileOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(uintptr(len(o.Data)))
	b.Append(o.Data)
	return
}

// Write data to a file previously opened with CreateFile or OpenFile.
//
// When the user writes data using write(2), the write goes into the page
// cache and the page is marked dirty. Later the kernel may write back the
// page via the FUSE VFS layer, causing this op to be sent:
//
//  *  The kernel calls address_space_operations::writepage when a dirty page
//     needs to be written to backing store (cf. http://goo.gl/Ezbewg). Fuse
//     sets this to fuse_writepage (cf. http://goo.gl/IeNvLT).
//
//  *  (http://goo.gl/Eestuy) fuse_writepage calls fuse_writepage_locked.
//
//  *  (http://goo.gl/RqYIxY) fuse_writepage_locked makes a write request to
//     the userspace server.
//
// Note that the kernel *will* ensure that writes are received and acknowledged
// by the file system before sending a FlushFileOp when closing the file
// descriptor to which they were written:
//
//  *  (http://goo.gl/PheZjf) fuse_flush calls write_inode_now, which appears
//     to start a writeback in the background (it talks about a "flusher
//     thread").
//
//  *  (http://goo.gl/1IiepM) fuse_flush then calls fuse_sync_writes, which
//     "[waits] for all pending writepages on the inode to finish".
//
//  *  (http://goo.gl/zzvxWv) Only then does fuse_flush finally send the
//     flush request.
//
// (See also http://goo.gl/ocdTdM, fuse-devel thread "Fuse guarantees on
// concurrent requests".)
type WriteFileOp struct {
	commonOp

	// The file inode that we are modifying, and the handle previously returned
	// by CreateFile or OpenFile when opening that inode.
	Inode  InodeID
	Handle HandleID

	// The offset at which to write the data below.
	//
	// The man page for pwrite(2) implies that aside from changing the file
	// handle's offset, using pwrite is equivalent to using lseek(2) and then
	// write(2). The man page for lseek(2) says the following:
	//
	// "The lseek() function allows the file offset to be set beyond the end of
	// the file (but this does not change the size of the file). If data is later
	// written at this point, subsequent reads of the data in the gap (a "hole")
	// return null bytes (aq\0aq) until data is actually written into the gap."
	//
	// It is therefore reasonable to assume that the kernel is looking for
	// the following semantics:
	//
	// *   If the offset is less than or equal to the current size, extend the
	//     file as necessary to fit any data that goes past the end of the file.
	//
	// *   If the offset is greater than the current size, extend the file
	//     with null bytes until it is not, then do the above.
	//
	Offset int64

	// The data to write.
	//
	// The FUSE documentation requires that exactly the number of bytes supplied
	// 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
}

func (o *WriteFileOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(unsafe.Sizeof(fusekernel.WriteOut{}))
	out := (*fusekernel.WriteOut)(b.Grow(unsafe.Sizeof(fusekernel.WriteOut{})))
	out.Size = uint32(len(o.Data))

	return
}

// Synchronize the current contents of an open file to storage.
//
// vfs.txt documents this as being called for by the fsync(2) system call
// (cf. http://goo.gl/j9X8nB). Code walk for that case:
//
//  *  (http://goo.gl/IQkWZa) sys_fsync calls do_fsync, calls vfs_fsync, calls
//     vfs_fsync_range.
//
//  *  (http://goo.gl/5L2SMy) vfs_fsync_range calls f_op->fsync.
//
// Note that this is also sent by fdatasync(2) (cf. http://goo.gl/01R7rF), and
// may be sent for msync(2) with the MS_SYNC flag (see the notes on
// FlushFileOp).
//
// See also: FlushFileOp, which may perform a similar function when closing a
// file (but which is not used in "real" file systems).
type SyncFileOp struct {
	commonOp

	// The file and handle being sync'd.
	Inode  InodeID
	Handle HandleID
}

func (o *SyncFileOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

// Flush the current state of an open file to storage upon closing a file
// descriptor.
//
// vfs.txt documents this as being sent for each close(2) system call (cf.
// http://goo.gl/FSkbrq). Code walk for that case:
//
//  *  (http://goo.gl/e3lv0e) sys_close calls __close_fd, calls filp_close.
//  *  (http://goo.gl/nI8fxD) filp_close calls f_op->flush (fuse_flush).
//
// But note that this is also sent in other contexts where a file descriptor is
// closed, such as dup2(2) (cf. http://goo.gl/NQDvFS). In the case of close(2),
// a flush error is returned to the user. For dup2(2), it is not.
//
// One potentially significant case where this may not be sent is mmap'd files,
// where the behavior is complicated:
//
//  *  munmap(2) does not cause flushes (cf. http://goo.gl/j8B9g0).
//
//  *  On OS X, if a user modifies a mapped file via the mapping before
//     closing the file with close(2), the WriteFileOps for the modifications
//     may not be received before the FlushFileOp for the close(2) (cf.
//     http://goo.gl/kVmNcx).
//
//  *  However, even on OS X you can arrange for writes via a mapping to be
//     flushed by calling msync(2) followed by close(2). On OS X msync(2)
//     will cause a WriteFileOps to go through and close(2) will cause a
//     FlushFile as usual (cf. http://goo.gl/kVmNcx). On Linux, msync(2) does
//     nothing unless you set the MS_SYNC flag, in which case it causes a
//     SyncFileOp to be sent (cf. http://goo.gl/P3mErk).
//
// In summary: if you make data durable in both FlushFile and SyncFile, then
// your users can get safe behavior from mapped files on both operating systems
// by calling msync(2) with MS_SYNC, followed by munmap(2), followed by
// close(2). On Linux, the msync(2) is optional (cf. http://goo.gl/EIhAxv and
// the notes on WriteFileOp).
//
// Because of cases like dup2(2), FlushFileOps are not necessarily one to one
// with OpenFileOps. They should not be used for reference counting, and the
// handle must remain valid even after the flush op is received (use
// ReleaseFileHandleOp for disposing of it).
//
// Typical "real" file systems do not implement this, presumably relying on
// the kernel to write out the page cache to the block device eventually.
// They can get away with this because a later open(2) will see the same
// data. A file system that writes to remote storage however probably wants
// to at least schedule a real flush, and maybe do it immediately in order to
// return any errors that occur.
type FlushFileOp struct {
	commonOp

	// The file and handle being flushed.
	Inode  InodeID
	Handle HandleID
}

func (o *FlushFileOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

// Release a previously-minted file handle. The kernel calls this when there
// are no more references to an open file: all file descriptors are closed
// and all memory mappings are unmapped.
//
// The kernel guarantees that the handle ID will not be used in further calls
// to the file system (unless it is reissued by the file system).
//
// Errors from this op are ignored by the kernel (cf. http://goo.gl/RL38Do).
type ReleaseFileHandleOp struct {
	commonOp

	// The handle ID to be released. The kernel guarantees that this ID will not
	// be used in further calls to the file system (unless it is reissued by the
	// file system).
	Handle HandleID
}

func (o *ReleaseFileHandleOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(0)
	return
}

// A sentinel used for unknown ops. The user is expected to respond with a
// non-nil error.
type unknownOp struct {
	commonOp
	opCode uint32
	inode  InodeID
}

func (o *unknownOp) ShortDesc() (desc string) {
	desc = fmt.Sprintf("<opcode %d>(inode=%v)", o.opCode, o.inode)
	return
}

func (o *unknownOp) kernelResponse() (b buffer.OutMessage) {
	panic(fmt.Sprintf("Should never get here for unknown op: %s", o.ShortDesc()))
}

////////////////////////////////////////////////////////////////////////
// Reading symlinks
////////////////////////////////////////////////////////////////////////

// Read the target of a symlink inode.
type ReadSymlinkOp struct {
	commonOp

	// The symlink inode that we are reading.
	Inode InodeID

	// Set by the file system: the target of the symlink.
	Target string
}

func (o *ReadSymlinkOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(uintptr(len(o.Target)))
	b.AppendString(o.Target)
	return
}

////////////////////////////////////////////////////////////////////////
// Internal
////////////////////////////////////////////////////////////////////////

// TODO(jacobsa): Untangle the way ops work and move these to an internal
// package, along with Convert. I think all of the behavior wants to be on
// Connection. Ops have only String methods. Connection.ReadOp returns an
// interace{} and a context. If we must restore debug logging later, we can
// stuff an op ID in that context and add a Connection.Logf method. Connection
// has a Reply method that takes a descendent context and an error.

// Do not use this struct directly. See the TODO in fuseops/ops.go.
type InternalStatFSOp struct {
	commonOp
}

func (o *InternalStatFSOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(unsafe.Sizeof(fusekernel.StatfsOut{}))
	b.Grow(unsafe.Sizeof(fusekernel.StatfsOut{}))

	return
}

// Do not use this struct directly. See the TODO in fuseops/ops.go.
type InternalInterruptOp struct {
	commonOp
	FuseID uint64
}

func (o *InternalInterruptOp) kernelResponse() (b buffer.OutMessage) {
	panic("Shouldn't get here.")
}

// Do not use this struct directly. See the TODO in fuseops/ops.go.
type InternalInitOp struct {
	commonOp

	// In
	Kernel fusekernel.Protocol

	// Out
	Library      fusekernel.Protocol
	MaxReadahead uint32
	Flags        fusekernel.InitFlags
	MaxWrite     uint32
}

func (o *InternalInitOp) kernelResponse() (b buffer.OutMessage) {
	b = buffer.NewOutMessage(unsafe.Sizeof(fusekernel.InitOut{}))
	out := (*fusekernel.InitOut)(b.Grow(unsafe.Sizeof(fusekernel.InitOut{})))

	out.Major = o.Library.Major
	out.Minor = o.Library.Minor
	out.MaxReadahead = o.MaxReadahead
	out.Flags = uint32(o.Flags)
	out.MaxWrite = o.MaxWrite

	return
}