1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
21 ErrReadOnlyFile = errors.New("read-only file")
22 ErrNegativeOffset = errors.New("cannot seek to negative offset")
23 ErrFileExists = errors.New("file exists")
24 ErrInvalidOperation = errors.New("invalid operation")
25 ErrInvalidArgument = errors.New("invalid argument")
26 ErrDirectoryNotEmpty = errors.New("directory not empty")
27 ErrWriteOnlyMode = errors.New("file is O_WRONLY")
28 ErrSyncNotSupported = errors.New("O_SYNC flag is not supported")
29 ErrIsDirectory = errors.New("cannot rename file to overwrite existing directory")
30 ErrNotADirectory = errors.New("not a directory")
31 ErrPermission = os.ErrPermission
34 // A File is an *os.File-like interface for reading and writing files
42 Readdir(int) ([]os.FileInfo, error)
43 Stat() (os.FileInfo, error)
48 // A FileSystem is an http.Filesystem plus Stat() and support for
49 // opening writable files. All methods are safe to call from multiple
51 type FileSystem interface {
57 // filesystem-wide lock: used by Rename() to prevent deadlock
58 // while locking multiple inodes.
61 // create a new node with nil parent.
62 newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error)
64 // analogous to os.Stat()
65 Stat(name string) (os.FileInfo, error)
67 // analogous to os.Create(): create/truncate a file and open it O_RDWR.
68 Create(name string) (File, error)
70 // Like os.OpenFile(): create or open a file or directory.
72 // If flag&os.O_EXCL==0, it opens an existing file or
73 // directory if one exists. If flag&os.O_CREATE!=0, it creates
74 // a new empty file or directory if one does not already
77 // When creating a new item, perm&os.ModeDir determines
78 // whether it is a file or a directory.
80 // A file can be opened multiple times and used concurrently
81 // from multiple goroutines. However, each File object should
82 // be used by only one goroutine at a time.
83 OpenFile(name string, flag int, perm os.FileMode) (File, error)
85 Mkdir(name string, perm os.FileMode) error
86 Remove(name string) error
87 RemoveAll(name string) error
88 Rename(oldname, newname string) error
90 // Write buffered data from memory to storage, returning when
91 // all updates have been saved to persistent storage.
94 // Write buffered data from memory to storage, but don't wait
95 // for all writes to finish before returning. If shortBlocks
96 // is true, flush everything; otherwise, if there's less than
97 // a full block of buffered data at the end of a stream, leave
98 // it buffered in memory in case more data can be appended.
99 Flush(shortBlocks bool) error
102 type inode interface {
103 SetParent(parent inode, name string)
106 Read([]byte, filenodePtr) (int, filenodePtr, error)
107 Write([]byte, filenodePtr) (int, filenodePtr, error)
108 Truncate(int64) error
110 Readdir() ([]os.FileInfo, error)
112 FileInfo() os.FileInfo
114 // Child() performs lookups and updates of named child nodes.
116 // (The term "child" here is used strictly. This means name is
117 // not "." or "..", and name does not contain "/".)
119 // If replace is non-nil, Child calls replace(x) where x is
120 // the current child inode with the given name. If possible,
121 // the child inode is replaced with the one returned by
124 // If replace(x) returns an inode (besides x or nil) that is
125 // subsequently returned by Child(), then Child()'s caller
126 // must ensure the new child's name and parent are set/updated
127 // to Child()'s name argument and its receiver respectively.
128 // This is not necessarily done before replace(x) returns, but
129 // it must be done before Child()'s caller releases the
132 // Nil represents "no child". replace(nil) signifies that no
133 // child with this name exists yet. If replace() returns nil,
134 // the existing child should be deleted if possible.
136 // An implementation of Child() is permitted to ignore
137 // replace() or its return value. For example, a regular file
138 // inode does not have children, so Child() always returns
141 // Child() returns the child, if any, with the given name: if
142 // a child was added or changed, the new child is returned.
144 // Caller must have lock (or rlock if replace is nil).
145 Child(name string, replace func(inode) (inode, error)) (inode, error)
152 type fileinfo struct {
159 // Name implements os.FileInfo.
160 func (fi fileinfo) Name() string {
164 // ModTime implements os.FileInfo.
165 func (fi fileinfo) ModTime() time.Time {
169 // Mode implements os.FileInfo.
170 func (fi fileinfo) Mode() os.FileMode {
174 // IsDir implements os.FileInfo.
175 func (fi fileinfo) IsDir() bool {
176 return fi.mode&os.ModeDir != 0
179 // Size implements os.FileInfo.
180 func (fi fileinfo) Size() int64 {
184 // Sys implements os.FileInfo.
185 func (fi fileinfo) Sys() interface{} {
189 type nullnode struct{}
191 func (*nullnode) Mkdir(string, os.FileMode) error {
192 return ErrInvalidOperation
195 func (*nullnode) Read([]byte, filenodePtr) (int, filenodePtr, error) {
196 return 0, filenodePtr{}, ErrInvalidOperation
199 func (*nullnode) Write([]byte, filenodePtr) (int, filenodePtr, error) {
200 return 0, filenodePtr{}, ErrInvalidOperation
203 func (*nullnode) Truncate(int64) error {
204 return ErrInvalidOperation
207 func (*nullnode) FileInfo() os.FileInfo {
211 func (*nullnode) IsDir() bool {
215 func (*nullnode) Readdir() ([]os.FileInfo, error) {
216 return nil, ErrInvalidOperation
219 func (*nullnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
220 return nil, ErrNotADirectory
223 type treenode struct {
226 inodes map[string]inode
232 func (n *treenode) FS() FileSystem {
236 func (n *treenode) SetParent(p inode, name string) {
240 n.fileinfo.name = name
243 func (n *treenode) Parent() inode {
249 func (n *treenode) IsDir() bool {
253 func (n *treenode) Child(name string, replace func(inode) (inode, error)) (child inode, err error) {
254 child = n.inodes[name]
255 if name == "" || name == "." || name == ".." {
256 err = ErrInvalidArgument
262 newchild, err := replace(child)
267 delete(n.inodes, name)
268 } else if newchild != child {
269 n.inodes[name] = newchild
270 n.fileinfo.modTime = time.Now()
276 func (n *treenode) Size() int64 {
277 return n.FileInfo().Size()
280 func (n *treenode) FileInfo() os.FileInfo {
283 n.fileinfo.size = int64(len(n.inodes))
287 func (n *treenode) Readdir() (fi []os.FileInfo, err error) {
290 fi = make([]os.FileInfo, 0, len(n.inodes))
291 for _, inode := range n.inodes {
292 fi = append(fi, inode.FileInfo())
297 type fileSystem struct {
303 func (fs *fileSystem) rootnode() inode {
307 func (fs *fileSystem) locker() sync.Locker {
311 // OpenFile is analogous to os.OpenFile().
312 func (fs *fileSystem) OpenFile(name string, flag int, perm os.FileMode) (File, error) {
313 return fs.openFile(name, flag, perm)
316 func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*filehandle, error) {
317 if flag&os.O_SYNC != 0 {
318 return nil, ErrSyncNotSupported
320 dirname, name := path.Split(name)
321 parent, err := rlookup(fs.root, dirname)
325 var readable, writable bool
326 switch flag & (os.O_RDWR | os.O_RDONLY | os.O_WRONLY) {
335 return nil, fmt.Errorf("invalid flags 0x%x", flag)
337 if !writable && parent.IsDir() {
338 // A directory can be opened via "foo/", "foo/.", or
342 return &filehandle{inode: parent}, nil
344 return &filehandle{inode: parent.Parent()}, nil
347 createMode := flag&os.O_CREATE != 0
350 defer parent.Unlock()
353 defer parent.RUnlock()
355 n, err := parent.Child(name, nil)
360 return nil, os.ErrNotExist
362 n, err = parent.Child(name, func(inode) (repl inode, err error) {
363 repl, err = parent.FS().newNode(name, perm|0755, time.Now())
367 repl.SetParent(parent, name)
373 // Parent rejected new child, but returned no error
374 return nil, ErrInvalidArgument
376 } else if flag&os.O_EXCL != 0 {
377 return nil, ErrFileExists
378 } else if flag&os.O_TRUNC != 0 {
380 return nil, fmt.Errorf("invalid flag O_TRUNC in read-only mode")
381 } else if n.IsDir() {
382 return nil, fmt.Errorf("invalid flag O_TRUNC when opening directory")
383 } else if err := n.Truncate(0); err != nil {
389 append: flag&os.O_APPEND != 0,
395 func (fs *fileSystem) Open(name string) (http.File, error) {
396 return fs.OpenFile(name, os.O_RDONLY, 0)
399 func (fs *fileSystem) Create(name string) (File, error) {
400 return fs.OpenFile(name, os.O_CREATE|os.O_RDWR|os.O_TRUNC, 0)
403 func (fs *fileSystem) Mkdir(name string, perm os.FileMode) error {
404 dirname, name := path.Split(name)
405 n, err := rlookup(fs.root, dirname)
411 if child, err := n.Child(name, nil); err != nil {
413 } else if child != nil {
417 _, err = n.Child(name, func(inode) (repl inode, err error) {
418 repl, err = n.FS().newNode(name, perm|os.ModeDir, time.Now())
422 repl.SetParent(n, name)
428 func (fs *fileSystem) Stat(name string) (os.FileInfo, error) {
429 node, err := rlookup(fs.root, name)
433 return node.FileInfo(), nil
436 func (fs *fileSystem) Rename(oldname, newname string) error {
437 olddir, oldname := path.Split(oldname)
438 if oldname == "" || oldname == "." || oldname == ".." {
439 return ErrInvalidArgument
441 olddirf, err := fs.openFile(olddir+".", os.O_RDONLY, 0)
443 return fmt.Errorf("%q: %s", olddir, err)
445 defer olddirf.Close()
447 newdir, newname := path.Split(newname)
448 if newname == "." || newname == ".." {
449 return ErrInvalidArgument
450 } else if newname == "" {
451 // Rename("a/b", "c/") means Rename("a/b", "c/b")
454 newdirf, err := fs.openFile(newdir+".", os.O_RDONLY, 0)
456 return fmt.Errorf("%q: %s", newdir, err)
458 defer newdirf.Close()
460 // TODO: If the nearest common ancestor ("nca") of olddirf and
461 // newdirf is on a different filesystem than fs, we should
462 // call nca.FS().Rename() instead of proceeding. Until then
463 // it's awkward for filesystems to implement their own Rename
464 // methods effectively: the only one that runs is the one on
465 // the root FileSystem exposed to the caller (webdav, fuse,
468 // When acquiring locks on multiple inodes, avoid deadlock by
469 // locking the entire containing filesystem first.
470 cfs := olddirf.inode.FS()
472 defer cfs.locker().Unlock()
474 if cfs != newdirf.inode.FS() {
475 // Moving inodes across filesystems is not (yet)
476 // supported. Locking inodes from different
477 // filesystems could deadlock, so we must error out
479 return ErrInvalidArgument
482 // To ensure we can test reliably whether we're about to move
483 // a directory into itself, lock all potential common
484 // ancestors of olddir and newdir.
485 needLock := []sync.Locker{}
486 for _, node := range []inode{olddirf.inode, newdirf.inode} {
487 needLock = append(needLock, node)
488 for node.Parent() != node && node.Parent().FS() == node.FS() {
490 needLock = append(needLock, node)
493 locked := map[sync.Locker]bool{}
494 for i := len(needLock) - 1; i >= 0; i-- {
495 if n := needLock[i]; !locked[n] {
502 _, err = olddirf.inode.Child(oldname, func(oldinode inode) (inode, error) {
504 return oldinode, os.ErrNotExist
506 if locked[oldinode] {
507 // oldinode cannot become a descendant of itself.
508 return oldinode, ErrInvalidArgument
510 if oldinode.FS() != cfs && newdirf.inode != olddirf.inode {
511 // moving a mount point to a different parent
512 // is not (yet) supported.
513 return oldinode, ErrInvalidArgument
515 accepted, err := newdirf.inode.Child(newname, func(existing inode) (inode, error) {
516 if existing != nil && existing.IsDir() {
517 return existing, ErrIsDirectory
522 // Leave oldinode in olddir.
525 accepted.SetParent(newdirf.inode, newname)
531 func (fs *fileSystem) Remove(name string) error {
532 return fs.remove(strings.TrimRight(name, "/"), false)
535 func (fs *fileSystem) RemoveAll(name string) error {
536 err := fs.remove(strings.TrimRight(name, "/"), true)
537 if os.IsNotExist(err) {
538 // "If the path does not exist, RemoveAll returns
539 // nil." (see "os" pkg)
545 func (fs *fileSystem) remove(name string, recursive bool) error {
546 dirname, name := path.Split(name)
547 if name == "" || name == "." || name == ".." {
548 return ErrInvalidArgument
550 dir, err := rlookup(fs.root, dirname)
556 _, err = dir.Child(name, func(node inode) (inode, error) {
558 return nil, os.ErrNotExist
560 if !recursive && node.IsDir() && node.Size() > 0 {
561 return node, ErrDirectoryNotEmpty
568 func (fs *fileSystem) Sync() error {
569 log.Printf("TODO: sync fileSystem")
570 return ErrInvalidOperation
573 func (fs *fileSystem) Flush(bool) error {
574 log.Printf("TODO: flush fileSystem")
575 return ErrInvalidOperation
578 // rlookup (recursive lookup) returns the inode for the file/directory
579 // with the given name (which may contain "/" separators). If no such
580 // file/directory exists, the returned node is nil.
581 func rlookup(start inode, path string) (node inode, err error) {
583 for _, name := range strings.Split(path, "/") {
585 if name == "." || name == "" {
593 node, err = func() (inode, error) {
596 return node.Child(name, nil)
598 if node == nil || err != nil {
602 if node == nil && err == nil {
608 func permittedName(name string) bool {
609 return name != "" && name != "." && name != ".." && !strings.Contains(name, "/")