1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
23 ErrReadOnlyFile = errors.New("read-only file")
24 ErrNegativeOffset = errors.New("cannot seek to negative offset")
25 ErrFileExists = errors.New("file exists")
26 ErrInvalidOperation = errors.New("invalid operation")
27 ErrInvalidArgument = errors.New("invalid argument")
28 ErrDirectoryNotEmpty = errors.New("directory not empty")
29 ErrWriteOnlyMode = errors.New("file is O_WRONLY")
30 ErrSyncNotSupported = errors.New("O_SYNC flag is not supported")
31 ErrIsDirectory = errors.New("cannot rename file to overwrite existing directory")
32 ErrPermission = os.ErrPermission
34 maxBlockSize = 1 << 26
37 // A File is an *os.File-like interface for reading and writing files
38 // in a CollectionFileSystem.
45 Readdir(int) ([]os.FileInfo, error)
46 Stat() (os.FileInfo, error)
50 type keepClient interface {
51 ReadAt(locator string, p []byte, off int) (int, error)
52 PutB(p []byte) (string, int, error)
55 type fileinfo struct {
62 // Name implements os.FileInfo.
63 func (fi fileinfo) Name() string {
67 // ModTime implements os.FileInfo.
68 func (fi fileinfo) ModTime() time.Time {
72 // Mode implements os.FileInfo.
73 func (fi fileinfo) Mode() os.FileMode {
77 // IsDir implements os.FileInfo.
78 func (fi fileinfo) IsDir() bool {
79 return fi.mode&os.ModeDir != 0
82 // Size implements os.FileInfo.
83 func (fi fileinfo) Size() int64 {
87 // Sys implements os.FileInfo.
88 func (fi fileinfo) Sys() interface{} {
92 // A FileSystem is an http.Filesystem plus Stat() and support for
93 // opening writable files. All methods are safe to call from multiple
95 type FileSystem interface {
100 // analogous to os.Stat()
101 Stat(name string) (os.FileInfo, error)
103 // analogous to os.Create(): create/truncate a file and open it O_RDWR.
104 Create(name string) (File, error)
106 // Like os.OpenFile(): create or open a file or directory.
108 // If flag&os.O_EXCL==0, it opens an existing file or
109 // directory if one exists. If flag&os.O_CREATE!=0, it creates
110 // a new empty file or directory if one does not already
113 // When creating a new item, perm&os.ModeDir determines
114 // whether it is a file or a directory.
116 // A file can be opened multiple times and used concurrently
117 // from multiple goroutines. However, each File object should
118 // be used by only one goroutine at a time.
119 OpenFile(name string, flag int, perm os.FileMode) (File, error)
121 Mkdir(name string, perm os.FileMode) error
122 Remove(name string) error
123 RemoveAll(name string) error
124 Rename(oldname, newname string) error
127 // A CollectionFileSystem is a FileSystem that can be serialized as a
128 // manifest and stored as a collection.
129 type CollectionFileSystem interface {
132 // Flush all file data to Keep and return a snapshot of the
133 // filesystem suitable for saving as (Collection)ManifestText.
134 // Prefix (normally ".") is a top level directory, effectively
135 // prepended to all paths in the returned manifest.
136 MarshalManifest(prefix string) (string, error)
139 type fileSystem struct {
143 type collectionFileSystem struct {
147 func (fs collectionFileSystem) MarshalManifest(prefix string) (string, error) {
148 fs.fileSystem.inode.Lock()
149 defer fs.fileSystem.inode.Unlock()
150 return fs.fileSystem.inode.(*dirnode).marshalManifest(prefix)
153 // OpenFile is analogous to os.OpenFile().
154 func (fs *fileSystem) OpenFile(name string, flag int, perm os.FileMode) (File, error) {
155 return fs.openFile(name, flag, perm)
158 func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*filehandle, error) {
159 var dn inode = fs.inode
160 if flag&os.O_SYNC != 0 {
161 return nil, ErrSyncNotSupported
163 dirname, name := path.Split(name)
164 parent := rlookup(dn, dirname)
166 return nil, os.ErrNotExist
168 var readable, writable bool
169 switch flag & (os.O_RDWR | os.O_RDONLY | os.O_WRONLY) {
178 return nil, fmt.Errorf("invalid flags 0x%x", flag)
180 if !writable && parent.IsDir() {
181 // A directory can be opened via "foo/", "foo/.", or
185 return &filehandle{inode: parent}, nil
187 return &filehandle{inode: parent.Parent()}, nil
190 createMode := flag&os.O_CREATE != 0
193 defer parent.Unlock()
196 defer parent.RUnlock()
198 n := parent.Child(name, nil)
201 return nil, os.ErrNotExist
204 n = parent.Child(name, func(inode) inode {
206 switch parent := parent.(type) {
209 case *collectionFileSystem:
210 dn = parent.inode.(*dirnode)
212 err = ErrInvalidArgument
216 n, err = dn.newDirnode(dn, name, perm|0755, time.Now())
218 n, err = dn.newFilenode(dn, name, perm|0755, time.Now())
225 // parent rejected new child
226 return nil, ErrInvalidOperation
228 } else if flag&os.O_EXCL != 0 {
229 return nil, ErrFileExists
230 } else if flag&os.O_TRUNC != 0 {
232 return nil, fmt.Errorf("invalid flag O_TRUNC in read-only mode")
233 } else if fn, ok := n.(*filenode); !ok {
234 return nil, fmt.Errorf("invalid flag O_TRUNC when opening directory")
241 append: flag&os.O_APPEND != 0,
247 func (fs *fileSystem) Open(name string) (http.File, error) {
248 return fs.OpenFile(name, os.O_RDONLY, 0)
251 func (fs *fileSystem) Create(name string) (File, error) {
252 return fs.OpenFile(name, os.O_CREATE|os.O_RDWR|os.O_TRUNC, 0)
255 func (fs *fileSystem) Mkdir(name string, perm os.FileMode) (err error) {
256 dirname, name := path.Split(name)
257 n := rlookup(fs.inode, dirname)
259 return os.ErrNotExist
263 if n.Child(name, nil) != nil {
266 dn, ok := n.(*dirnode)
268 return ErrInvalidArgument
270 child := n.Child(name, func(inode) (child inode) {
271 child, err = dn.newDirnode(dn, name, perm, time.Now())
276 } else if child == nil {
277 return ErrInvalidArgument
282 func (fs *fileSystem) Stat(name string) (fi os.FileInfo, err error) {
283 node := rlookup(fs.inode, name)
292 func (fs *fileSystem) Rename(oldname, newname string) error {
293 olddir, oldname := path.Split(oldname)
294 if oldname == "" || oldname == "." || oldname == ".." {
295 return ErrInvalidArgument
297 olddirf, err := fs.openFile(olddir+".", os.O_RDONLY, 0)
299 return fmt.Errorf("%q: %s", olddir, err)
301 defer olddirf.Close()
303 newdir, newname := path.Split(newname)
304 if newname == "." || newname == ".." {
305 return ErrInvalidArgument
306 } else if newname == "" {
307 // Rename("a/b", "c/") means Rename("a/b", "c/b")
310 newdirf, err := fs.openFile(newdir+".", os.O_RDONLY, 0)
312 return fmt.Errorf("%q: %s", newdir, err)
314 defer newdirf.Close()
316 // When acquiring locks on multiple nodes, all common
317 // ancestors must be locked first in order to avoid
318 // deadlock. This is assured by locking the path from root to
319 // newdir, then locking the path from root to olddir, skipping
320 // any already-locked nodes.
321 needLock := []sync.Locker{}
322 for _, f := range []*filehandle{olddirf, newdirf} {
324 needLock = append(needLock, node)
325 for node.Parent() != node {
327 needLock = append(needLock, node)
330 locked := map[sync.Locker]bool{}
331 for i := len(needLock) - 1; i >= 0; i-- {
332 if n := needLock[i]; !locked[n] {
339 if _, ok := newdirf.inode.(*dirnode); !ok {
340 return ErrInvalidOperation
344 olddirf.inode.Child(oldname, func(oldinode inode) inode {
349 newdirf.inode.Child(newname, func(existing inode) inode {
350 if existing != nil && existing.IsDir() {
360 defer oldinode.Unlock()
361 olddn := olddirf.inode.(*dirnode)
362 newdn := newdirf.inode.(*dirnode)
363 switch n := oldinode.(type) {
365 n.parent = newdirf.inode
366 n.treenode.fileinfo.name = newname
369 n.fileinfo.name = newname
371 panic(fmt.Sprintf("bad inode type %T", n))
373 olddn.treenode.fileinfo.modTime = time.Now()
374 newdn.treenode.fileinfo.modTime = time.Now()
380 func (fs *fileSystem) Remove(name string) error {
381 return fs.remove(strings.TrimRight(name, "/"), false)
384 func (fs *fileSystem) RemoveAll(name string) error {
385 err := fs.remove(strings.TrimRight(name, "/"), true)
386 if os.IsNotExist(err) {
387 // "If the path does not exist, RemoveAll returns
388 // nil." (see "os" pkg)
394 func (fs *fileSystem) remove(name string, recursive bool) (err error) {
395 dirname, name := path.Split(name)
396 if name == "" || name == "." || name == ".." {
397 return ErrInvalidArgument
399 dir := rlookup(fs, dirname)
401 return os.ErrNotExist
405 dir.Child(name, func(node inode) inode {
410 if !recursive && node.IsDir() && node.Size() > 0 {
411 err = ErrDirectoryNotEmpty
419 type inode interface {
421 Read([]byte, filenodePtr) (int, filenodePtr, error)
422 Write([]byte, filenodePtr) (int, filenodePtr, error)
423 Truncate(int64) error
425 Readdir() []os.FileInfo
427 FileInfo() os.FileInfo
428 // Caller must have lock (or rlock if func is nil)
429 Child(string, func(inode) inode) inode
435 // filenode implements inode.
436 type filenode struct {
440 // number of times `segments` has changed in a
441 // way that might invalidate a filenodePtr
443 memsize int64 // bytes in memSegments
448 // filenodePtr is an offset into a file that is (usually) efficient to
449 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
451 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
452 // corresponds to file offset filenodePtr.off. Otherwise, it is
453 // necessary to reexamine len(filenode.segments[0]) etc. to find the
454 // correct segment and offset.
455 type filenodePtr struct {
462 // seek returns a ptr that is consistent with both startPtr.off and
463 // the current state of fn. The caller must already hold fn.RLock() or
466 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
471 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
473 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
474 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
477 // meaningless anyway
479 } else if ptr.off >= fn.fileinfo.size {
480 ptr.segmentIdx = len(fn.segments)
482 ptr.repacked = fn.repacked
484 } else if ptr.repacked == fn.repacked {
485 // segmentIdx and segmentOff accurately reflect
486 // ptr.off, but might have fallen off the end of a
488 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
495 ptr.repacked = fn.repacked
497 if ptr.off >= fn.fileinfo.size {
498 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
501 // Recompute segmentIdx and segmentOff. We have already
502 // established fn.fileinfo.size > ptr.off >= 0, so we don't
503 // have to deal with edge cases here.
505 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
506 // This would panic (index out of range) if
507 // fn.fileinfo.size were larger than
508 // sum(fn.segments[i].Len()) -- but that can't happen
509 // because we have ensured fn.fileinfo.size is always
511 segLen := int64(fn.segments[ptr.segmentIdx].Len())
512 if off+segLen > ptr.off {
513 ptr.segmentOff = int(ptr.off - off)
521 // caller must have lock
522 func (fn *filenode) appendSegment(e segment) {
523 fn.segments = append(fn.segments, e)
524 fn.fileinfo.size += int64(e.Len())
527 func (fn *filenode) Parent() inode {
533 // Read reads file data from a single segment, starting at startPtr,
534 // into p. startPtr is assumed not to be up-to-date. Caller must have
536 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
537 ptr = fn.seek(startPtr)
539 err = ErrNegativeOffset
542 if ptr.segmentIdx >= len(fn.segments) {
546 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
550 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
553 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
561 func (fn *filenode) Size() int64 {
564 return fn.fileinfo.Size()
567 func (fn *filenode) FileInfo() os.FileInfo {
573 func (fn *filenode) Truncate(size int64) error {
576 return fn.truncate(size)
579 func (fn *filenode) truncate(size int64) error {
580 if size == fn.fileinfo.size {
584 if size < fn.fileinfo.size {
585 ptr := fn.seek(filenodePtr{off: size})
586 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
587 if seg, ok := fn.segments[i].(*memSegment); ok {
588 fn.memsize -= int64(seg.Len())
591 if ptr.segmentOff == 0 {
592 fn.segments = fn.segments[:ptr.segmentIdx]
594 fn.segments = fn.segments[:ptr.segmentIdx+1]
595 switch seg := fn.segments[ptr.segmentIdx].(type) {
597 seg.Truncate(ptr.segmentOff)
598 fn.memsize += int64(seg.Len())
600 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
603 fn.fileinfo.size = size
606 for size > fn.fileinfo.size {
607 grow := size - fn.fileinfo.size
610 if len(fn.segments) == 0 {
612 fn.segments = append(fn.segments, seg)
613 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
615 fn.segments = append(fn.segments, seg)
617 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
620 seg.Truncate(seg.Len() + int(grow))
621 fn.fileinfo.size += grow
627 // Write writes data from p to the file, starting at startPtr,
628 // extending the file size if necessary. Caller must have Lock.
629 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
630 if startPtr.off > fn.fileinfo.size {
631 if err = fn.truncate(startPtr.off); err != nil {
632 return 0, startPtr, err
635 ptr = fn.seek(startPtr)
637 err = ErrNegativeOffset
640 for len(p) > 0 && err == nil {
642 if len(cando) > maxBlockSize {
643 cando = cando[:maxBlockSize]
645 // Rearrange/grow fn.segments (and shrink cando if
646 // needed) such that cando can be copied to
647 // fn.segments[ptr.segmentIdx] at offset
649 cur := ptr.segmentIdx
650 prev := ptr.segmentIdx - 1
652 if cur < len(fn.segments) {
653 _, curWritable = fn.segments[cur].(*memSegment)
655 var prevAppendable bool
656 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
657 _, prevAppendable = fn.segments[prev].(*memSegment)
659 if ptr.segmentOff > 0 && !curWritable {
660 // Split a non-writable block.
661 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
662 // Truncate cur, and insert a new
665 fn.segments = append(fn.segments, nil)
666 copy(fn.segments[cur+1:], fn.segments[cur:])
668 // Split cur into two copies, truncate
669 // the one on the left, shift the one
670 // on the right, and insert a new
671 // segment between them.
672 fn.segments = append(fn.segments, nil, nil)
673 copy(fn.segments[cur+2:], fn.segments[cur:])
674 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
679 seg.Truncate(len(cando))
680 fn.memsize += int64(len(cando))
681 fn.segments[cur] = seg
682 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
687 } else if curWritable {
688 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
693 // Shrink cando if needed to fit in
695 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
696 cando = cando[:cangrow]
700 if cur == len(fn.segments) {
701 // ptr is at EOF, filesize is changing.
702 fn.fileinfo.size += int64(len(cando))
703 } else if el := fn.segments[cur].Len(); el <= len(cando) {
704 // cando is long enough that we won't
705 // need cur any more. shrink cando to
706 // be exactly as long as cur
707 // (otherwise we'd accidentally shift
708 // the effective position of all
709 // segments after cur).
711 copy(fn.segments[cur:], fn.segments[cur+1:])
712 fn.segments = fn.segments[:len(fn.segments)-1]
714 // shrink cur by the same #bytes we're growing prev
715 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
721 ptr.segmentOff = fn.segments[prev].Len()
722 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
723 fn.memsize += int64(len(cando))
727 // Insert a segment between prev and
728 // cur, and advance prev/cur.
729 fn.segments = append(fn.segments, nil)
730 if cur < len(fn.segments) {
731 copy(fn.segments[cur+1:], fn.segments[cur:])
735 // appending a new segment does
736 // not invalidate any ptrs
739 seg.Truncate(len(cando))
740 fn.memsize += int64(len(cando))
741 fn.segments[cur] = seg
747 // Finally we can copy bytes from cando to the current segment.
748 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
752 ptr.off += int64(len(cando))
753 ptr.segmentOff += len(cando)
754 if ptr.segmentOff >= maxBlockSize {
755 fn.pruneMemSegments()
757 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
762 fn.fileinfo.modTime = time.Now()
767 // Write some data out to disk to reduce memory use. Caller must have
769 func (fn *filenode) pruneMemSegments() {
770 // TODO: async (don't hold Lock() while waiting for Keep)
771 // TODO: share code with (*dirnode)sync()
772 // TODO: pack/flush small blocks too, when fragmented
773 for idx, seg := range fn.segments {
774 seg, ok := seg.(*memSegment)
775 if !ok || seg.Len() < maxBlockSize {
778 locator, _, err := fn.parent.kc.PutB(seg.buf)
780 // TODO: stall (or return errors from)
781 // subsequent writes until flushing
785 fn.memsize -= int64(seg.Len())
786 fn.segments[idx] = storedSegment{
796 // FileSystem returns a CollectionFileSystem for the collection.
797 func (c *Collection) FileSystem(client *Client, kc keepClient) (CollectionFileSystem, error) {
798 var modTime time.Time
799 if c.ModifiedAt == nil {
802 modTime = *c.ModifiedAt
810 mode: os.ModeDir | 0755,
814 inodes: make(map[string]inode),
818 fs := &collectionFileSystem{fileSystem: fileSystem{inode: dn}}
819 if err := dn.loadManifest(c.ManifestText); err != nil {
825 type filehandle struct {
831 unreaddirs []os.FileInfo
834 func (f *filehandle) Read(p []byte) (n int, err error) {
836 return 0, ErrWriteOnlyMode
839 defer f.inode.RUnlock()
840 n, f.ptr, err = f.inode.Read(p, f.ptr)
844 func (f *filehandle) Seek(off int64, whence int) (pos int64, err error) {
845 size := f.inode.Size()
856 return f.ptr.off, ErrNegativeOffset
858 if ptr.off != f.ptr.off {
860 // force filenode to recompute f.ptr fields on next
864 return f.ptr.off, nil
867 func (f *filehandle) Truncate(size int64) error {
868 return f.inode.Truncate(size)
871 func (f *filehandle) Write(p []byte) (n int, err error) {
873 return 0, ErrReadOnlyFile
876 defer f.inode.Unlock()
877 if fn, ok := f.inode.(*filenode); ok && f.append {
879 off: fn.fileinfo.size,
880 segmentIdx: len(fn.segments),
882 repacked: fn.repacked,
885 n, f.ptr, err = f.inode.Write(p, f.ptr)
889 func (f *filehandle) Readdir(count int) ([]os.FileInfo, error) {
890 if !f.inode.IsDir() {
891 return nil, ErrInvalidOperation
894 return f.inode.Readdir(), nil
896 if f.unreaddirs == nil {
897 f.unreaddirs = f.inode.Readdir()
899 if len(f.unreaddirs) == 0 {
902 if count > len(f.unreaddirs) {
903 count = len(f.unreaddirs)
905 ret := f.unreaddirs[:count]
906 f.unreaddirs = f.unreaddirs[count:]
910 func (f *filehandle) Stat() (os.FileInfo, error) {
911 return f.inode.FileInfo(), nil
914 func (f *filehandle) Close() error {
918 type dirnode struct {
924 // sync flushes in-memory data (for all files in the tree rooted at
925 // dn) to persistent storage. Caller must hold dn.Lock().
926 func (dn *dirnode) sync() error {
927 type shortBlock struct {
931 var pending []shortBlock
934 flush := func(sbs []shortBlock) error {
938 block := make([]byte, 0, maxBlockSize)
939 for _, sb := range sbs {
940 block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
942 locator, _, err := dn.kc.PutB(block)
947 for _, sb := range sbs {
948 data := sb.fn.segments[sb.idx].(*memSegment).buf
949 sb.fn.segments[sb.idx] = storedSegment{
957 sb.fn.memsize -= int64(len(data))
962 names := make([]string, 0, len(dn.inodes))
963 for name := range dn.inodes {
964 names = append(names, name)
968 for _, name := range names {
969 fn, ok := dn.inodes[name].(*filenode)
975 for idx, seg := range fn.segments {
976 seg, ok := seg.(*memSegment)
980 if seg.Len() > maxBlockSize/2 {
981 if err := flush([]shortBlock{{fn, idx}}); err != nil {
986 if pendingLen+seg.Len() > maxBlockSize {
987 if err := flush(pending); err != nil {
993 pending = append(pending, shortBlock{fn, idx})
994 pendingLen += seg.Len()
997 return flush(pending)
1000 // caller must have read lock.
1001 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
1003 type filepart struct {
1008 var fileparts []filepart
1012 if err := dn.sync(); err != nil {
1016 names := make([]string, 0, len(dn.inodes))
1017 for name, node := range dn.inodes {
1018 names = append(names, name)
1024 for _, name := range names {
1025 switch node := dn.inodes[name].(type) {
1027 subdir, err := node.marshalManifest(prefix + "/" + name)
1031 subdirs = subdirs + subdir
1033 if len(node.segments) == 0 {
1034 fileparts = append(fileparts, filepart{name: name})
1037 for _, seg := range node.segments {
1038 switch seg := seg.(type) {
1040 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
1041 streamLen -= int64(seg.size)
1043 blocks = append(blocks, seg.locator)
1047 offset: streamLen + int64(seg.offset),
1048 length: int64(seg.length),
1050 if prev := len(fileparts) - 1; prev >= 0 &&
1051 fileparts[prev].name == name &&
1052 fileparts[prev].offset+fileparts[prev].length == next.offset {
1053 fileparts[prev].length += next.length
1055 fileparts = append(fileparts, next)
1057 streamLen += int64(seg.size)
1059 // This can't happen: we
1060 // haven't unlocked since
1062 panic(fmt.Sprintf("can't marshal segment type %T", seg))
1066 panic(fmt.Sprintf("can't marshal inode type %T", node))
1069 var filetokens []string
1070 for _, s := range fileparts {
1071 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
1073 if len(filetokens) == 0 {
1075 } else if len(blocks) == 0 {
1076 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
1078 return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
1081 func (dn *dirnode) loadManifest(txt string) error {
1083 streams := strings.Split(txt, "\n")
1084 if streams[len(streams)-1] != "" {
1085 return fmt.Errorf("line %d: no trailing newline", len(streams))
1087 streams = streams[:len(streams)-1]
1088 segments := []storedSegment{}
1089 for i, stream := range streams {
1091 var anyFileTokens bool
1094 segments = segments[:0]
1095 for i, token := range strings.Split(stream, " ") {
1097 dirname = manifestUnescape(token)
1100 if !strings.Contains(token, ":") {
1102 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1104 toks := strings.SplitN(token, "+", 3)
1106 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1108 length, err := strconv.ParseInt(toks[1], 10, 32)
1109 if err != nil || length < 0 {
1110 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1112 segments = append(segments, storedSegment{
1116 length: int(length),
1119 } else if len(segments) == 0 {
1120 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1123 toks := strings.SplitN(token, ":", 3)
1125 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1127 anyFileTokens = true
1129 offset, err := strconv.ParseInt(toks[0], 10, 64)
1130 if err != nil || offset < 0 {
1131 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1133 length, err := strconv.ParseInt(toks[1], 10, 64)
1134 if err != nil || length < 0 {
1135 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1137 name := dirname + "/" + manifestUnescape(toks[2])
1138 fnode, err := dn.createFileAndParents(name)
1140 return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
1142 // Map the stream offset/range coordinates to
1143 // block/offset/range coordinates and add
1144 // corresponding storedSegments to the filenode
1146 // Can't continue where we left off.
1147 // TODO: binary search instead of
1148 // rewinding all the way (but this
1149 // situation might be rare anyway)
1152 for next := int64(0); segIdx < len(segments); segIdx++ {
1153 seg := segments[segIdx]
1154 next = pos + int64(seg.Len())
1155 if next <= offset || seg.Len() == 0 {
1159 if pos >= offset+length {
1164 blkOff = int(offset - pos)
1166 blkLen := seg.Len() - blkOff
1167 if pos+int64(blkOff+blkLen) > offset+length {
1168 blkLen = int(offset + length - pos - int64(blkOff))
1170 fnode.appendSegment(storedSegment{
1172 locator: seg.locator,
1177 if next > offset+length {
1183 if segIdx == len(segments) && pos < offset+length {
1184 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1188 return fmt.Errorf("line %d: no file segments", lineno)
1189 } else if len(segments) == 0 {
1190 return fmt.Errorf("line %d: no locators", lineno)
1191 } else if dirname == "" {
1192 return fmt.Errorf("line %d: no stream name", lineno)
1198 // only safe to call from loadManifest -- no locking
1199 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
1201 names := strings.Split(path, "/")
1202 basename := names[len(names)-1]
1203 if basename == "" || basename == "." || basename == ".." {
1204 err = fmt.Errorf("invalid filename")
1207 for _, name := range names[:len(names)-1] {
1213 // can't be sure parent will be a *dirnode
1214 return nil, ErrInvalidArgument
1216 node = node.Parent().(*dirnode)
1219 node.Child(name, func(child inode) inode {
1220 switch child.(type) {
1222 node, err = dn.newDirnode(node, name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
1225 node = child.(*dirnode)
1229 err = ErrInvalidOperation
1237 node.Child(basename, func(child inode) inode {
1238 switch child := child.(type) {
1240 fn, err = dn.newFilenode(node, basename, 0755, node.FileInfo().ModTime())
1246 err = ErrIsDirectory
1249 err = ErrInvalidOperation
1256 // rlookup (recursive lookup) returns the inode for the file/directory
1257 // with the given name (which may contain "/" separators). If no such
1258 // file/directory exists, the returned node is nil.
1259 func rlookup(start inode, path string) (node inode) {
1261 for _, name := range strings.Split(path, "/") {
1266 if name == "." || name == "" {
1270 node = node.Parent()
1274 node = func() inode {
1276 defer node.RUnlock()
1277 return node.Child(name, nil)
1283 // Caller must have lock, and must have already ensured
1284 // Children(name,nil) is nil.
1285 func (dn *dirnode) newDirnode(parent *dirnode, name string, perm os.FileMode, modTime time.Time) (node *dirnode, err error) {
1286 if name == "" || name == "." || name == ".." {
1287 return nil, ErrInvalidArgument
1296 mode: perm | os.ModeDir,
1299 inodes: make(map[string]inode),
1304 func (dn *dirnode) newFilenode(parent *dirnode, name string, perm os.FileMode, modTime time.Time) (node *filenode, err error) {
1305 if name == "" || name == "." || name == ".." {
1306 return nil, ErrInvalidArgument
1312 mode: perm & ^os.ModeDir,
1318 type segment interface {
1321 // Return a new segment with a subsection of the data from this
1322 // one. length<0 means length=Len()-off.
1323 Slice(off int, length int) segment
1326 type memSegment struct {
1330 func (me *memSegment) Len() int {
1334 func (me *memSegment) Slice(off, length int) segment {
1336 length = len(me.buf) - off
1338 buf := make([]byte, length)
1339 copy(buf, me.buf[off:])
1340 return &memSegment{buf: buf}
1343 func (me *memSegment) Truncate(n int) {
1344 if n > cap(me.buf) {
1347 newsize = newsize << 2
1349 newbuf := make([]byte, n, newsize)
1350 copy(newbuf, me.buf)
1353 // Zero unused part when shrinking, in case we grow
1354 // and start using it again later.
1355 for i := n; i < len(me.buf); i++ {
1362 func (me *memSegment) WriteAt(p []byte, off int) {
1363 if off+len(p) > len(me.buf) {
1364 panic("overflowed segment")
1366 copy(me.buf[off:], p)
1369 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1370 if off > int64(me.Len()) {
1374 n = copy(p, me.buf[int(off):])
1381 type storedSegment struct {
1384 size int // size of stored block (also encoded in locator)
1385 offset int // position of segment within the stored block
1386 length int // bytes in this segment (offset + length <= size)
1389 func (se storedSegment) Len() int {
1393 func (se storedSegment) Slice(n, size int) segment {
1396 if size >= 0 && se.length > size {
1402 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1403 if off > int64(se.length) {
1406 maxlen := se.length - int(off)
1407 if len(p) > maxlen {
1409 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1415 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1418 func canonicalName(name string) string {
1419 name = path.Clean("/" + name)
1420 if name == "/" || name == "./" {
1422 } else if strings.HasPrefix(name, "/") {
1428 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1430 func manifestUnescapeFunc(seq string) string {
1434 i, err := strconv.ParseUint(seq[1:], 8, 8)
1436 // Invalid escape sequence: can't unescape.
1439 return string([]byte{byte(i)})
1442 func manifestUnescape(s string) string {
1443 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1446 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1448 func manifestEscapeFunc(seq string) string {
1449 return fmt.Sprintf("\\%03o", byte(seq[0]))
1452 func manifestEscape(s string) string {
1453 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)