1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
24 ErrReadOnlyFile = errors.New("read-only file")
25 ErrNegativeOffset = errors.New("cannot seek to negative offset")
26 ErrFileExists = errors.New("file exists")
27 ErrInvalidOperation = errors.New("invalid operation")
28 ErrInvalidArgument = errors.New("invalid argument")
29 ErrDirectoryNotEmpty = errors.New("directory not empty")
30 ErrWriteOnlyMode = errors.New("file is O_WRONLY")
31 ErrSyncNotSupported = errors.New("O_SYNC flag is not supported")
32 ErrIsDirectory = errors.New("cannot rename file to overwrite existing directory")
33 ErrPermission = os.ErrPermission
35 maxBlockSize = 1 << 26
38 // A File is an *os.File-like interface for reading and writing files
39 // in a CollectionFileSystem.
46 Readdir(int) ([]os.FileInfo, error)
47 Stat() (os.FileInfo, error)
51 type keepClient interface {
52 ReadAt(locator string, p []byte, off int) (int, error)
53 PutB(p []byte) (string, int, error)
56 type fileinfo struct {
63 // Name implements os.FileInfo.
64 func (fi fileinfo) Name() string {
68 // ModTime implements os.FileInfo.
69 func (fi fileinfo) ModTime() time.Time {
73 // Mode implements os.FileInfo.
74 func (fi fileinfo) Mode() os.FileMode {
78 // IsDir implements os.FileInfo.
79 func (fi fileinfo) IsDir() bool {
80 return fi.mode&os.ModeDir != 0
83 // Size implements os.FileInfo.
84 func (fi fileinfo) Size() int64 {
88 // Sys implements os.FileInfo.
89 func (fi fileinfo) Sys() interface{} {
93 // A FileSystem is an http.Filesystem plus Stat() and support for
94 // opening writable files. All methods are safe to call from multiple
96 type FileSystem interface {
101 // analogous to os.Stat()
102 Stat(name string) (os.FileInfo, error)
104 // analogous to os.Create(): create/truncate a file and open it O_RDWR.
105 Create(name string) (File, error)
107 // Like os.OpenFile(): create or open a file or directory.
109 // If flag&os.O_EXCL==0, it opens an existing file or
110 // directory if one exists. If flag&os.O_CREATE!=0, it creates
111 // a new empty file or directory if one does not already
114 // When creating a new item, perm&os.ModeDir determines
115 // whether it is a file or a directory.
117 // A file can be opened multiple times and used concurrently
118 // from multiple goroutines. However, each File object should
119 // be used by only one goroutine at a time.
120 OpenFile(name string, flag int, perm os.FileMode) (File, error)
122 Mkdir(name string, perm os.FileMode) error
123 Remove(name string) error
124 RemoveAll(name string) error
125 Rename(oldname, newname string) error
128 // A CollectionFileSystem is a FileSystem that can be serialized as a
129 // manifest and stored as a collection.
130 type CollectionFileSystem interface {
133 // Flush all file data to Keep and return a snapshot of the
134 // filesystem suitable for saving as (Collection)ManifestText.
135 // Prefix (normally ".") is a top level directory, effectively
136 // prepended to all paths in the returned manifest.
137 MarshalManifest(prefix string) (string, error)
140 type fileSystem struct {
144 type collectionFileSystem struct {
148 func (fs collectionFileSystem) Child(name string, replace func(inode) inode) inode {
149 if name == ".arvados#collection" {
150 return &getternode{Getter: func() ([]byte, error) {
153 coll.ManifestText, err = fs.MarshalManifest(".")
157 data, err := json.Marshal(&coll)
159 data = append(data, 10)
164 return fs.fileSystem.Child(name, replace)
167 func (fs collectionFileSystem) MarshalManifest(prefix string) (string, error) {
168 fs.fileSystem.inode.Lock()
169 defer fs.fileSystem.inode.Unlock()
170 return fs.fileSystem.inode.(*dirnode).marshalManifest(prefix)
173 // OpenFile is analogous to os.OpenFile().
174 func (fs *fileSystem) OpenFile(name string, flag int, perm os.FileMode) (File, error) {
175 return fs.openFile(name, flag, perm)
178 func (fs *fileSystem) openFile(name string, flag int, perm os.FileMode) (*filehandle, error) {
179 var dn inode = fs.inode
180 if flag&os.O_SYNC != 0 {
181 return nil, ErrSyncNotSupported
183 dirname, name := path.Split(name)
184 parent := rlookup(dn, dirname)
186 return nil, os.ErrNotExist
188 var readable, writable bool
189 switch flag & (os.O_RDWR | os.O_RDONLY | os.O_WRONLY) {
198 return nil, fmt.Errorf("invalid flags 0x%x", flag)
200 if !writable && parent.IsDir() {
201 // A directory can be opened via "foo/", "foo/.", or
205 return &filehandle{inode: parent}, nil
207 return &filehandle{inode: parent.Parent()}, nil
210 createMode := flag&os.O_CREATE != 0
213 defer parent.Unlock()
216 defer parent.RUnlock()
218 n := parent.Child(name, nil)
221 return nil, os.ErrNotExist
224 n = parent.Child(name, func(inode) inode {
226 switch parent := parent.(type) {
229 case *collectionFileSystem:
230 dn = parent.inode.(*dirnode)
232 err = ErrInvalidArgument
236 n, err = dn.newDirnode(dn, name, perm|0755, time.Now())
238 n, err = dn.newFilenode(dn, name, perm|0755, time.Now())
245 // parent rejected new child
246 return nil, ErrInvalidOperation
248 } else if flag&os.O_EXCL != 0 {
249 return nil, ErrFileExists
250 } else if flag&os.O_TRUNC != 0 {
252 return nil, fmt.Errorf("invalid flag O_TRUNC in read-only mode")
253 } else if fn, ok := n.(*filenode); !ok {
254 return nil, fmt.Errorf("invalid flag O_TRUNC when opening directory")
261 append: flag&os.O_APPEND != 0,
267 func (fs *fileSystem) Open(name string) (http.File, error) {
268 return fs.OpenFile(name, os.O_RDONLY, 0)
271 func (fs *fileSystem) Create(name string) (File, error) {
272 return fs.OpenFile(name, os.O_CREATE|os.O_RDWR|os.O_TRUNC, 0)
275 func (fs *fileSystem) Mkdir(name string, perm os.FileMode) (err error) {
276 dirname, name := path.Split(name)
277 n := rlookup(fs.inode, dirname)
279 return os.ErrNotExist
283 if n.Child(name, nil) != nil {
286 dn, ok := n.(*dirnode)
288 return ErrInvalidArgument
290 child := n.Child(name, func(inode) (child inode) {
291 child, err = dn.newDirnode(dn, name, perm, time.Now())
296 } else if child == nil {
297 return ErrInvalidArgument
302 func (fs *fileSystem) Stat(name string) (fi os.FileInfo, err error) {
303 node := rlookup(fs.inode, name)
312 func (fs *fileSystem) Rename(oldname, newname string) error {
313 olddir, oldname := path.Split(oldname)
314 if oldname == "" || oldname == "." || oldname == ".." {
315 return ErrInvalidArgument
317 olddirf, err := fs.openFile(olddir+".", os.O_RDONLY, 0)
319 return fmt.Errorf("%q: %s", olddir, err)
321 defer olddirf.Close()
323 newdir, newname := path.Split(newname)
324 if newname == "." || newname == ".." {
325 return ErrInvalidArgument
326 } else if newname == "" {
327 // Rename("a/b", "c/") means Rename("a/b", "c/b")
330 newdirf, err := fs.openFile(newdir+".", os.O_RDONLY, 0)
332 return fmt.Errorf("%q: %s", newdir, err)
334 defer newdirf.Close()
336 // When acquiring locks on multiple nodes, all common
337 // ancestors must be locked first in order to avoid
338 // deadlock. This is assured by locking the path from root to
339 // newdir, then locking the path from root to olddir, skipping
340 // any already-locked nodes.
341 needLock := []sync.Locker{}
342 for _, f := range []*filehandle{olddirf, newdirf} {
344 needLock = append(needLock, node)
345 for node.Parent() != node {
347 needLock = append(needLock, node)
350 locked := map[sync.Locker]bool{}
351 for i := len(needLock) - 1; i >= 0; i-- {
352 if n := needLock[i]; !locked[n] {
359 if _, ok := newdirf.inode.(*dirnode); !ok {
360 return ErrInvalidOperation
364 olddirf.inode.Child(oldname, func(oldinode inode) inode {
369 newdirf.inode.Child(newname, func(existing inode) inode {
370 if existing != nil && existing.IsDir() {
380 defer oldinode.Unlock()
381 olddn := olddirf.inode.(*dirnode)
382 newdn := newdirf.inode.(*dirnode)
383 switch n := oldinode.(type) {
385 n.parent = newdirf.inode
386 n.treenode.fileinfo.name = newname
389 n.fileinfo.name = newname
391 panic(fmt.Sprintf("bad inode type %T", n))
393 olddn.treenode.fileinfo.modTime = time.Now()
394 newdn.treenode.fileinfo.modTime = time.Now()
400 func (fs *fileSystem) Remove(name string) error {
401 return fs.remove(strings.TrimRight(name, "/"), false)
404 func (fs *fileSystem) RemoveAll(name string) error {
405 err := fs.remove(strings.TrimRight(name, "/"), true)
406 if os.IsNotExist(err) {
407 // "If the path does not exist, RemoveAll returns
408 // nil." (see "os" pkg)
414 func (fs *fileSystem) remove(name string, recursive bool) (err error) {
415 dirname, name := path.Split(name)
416 if name == "" || name == "." || name == ".." {
417 return ErrInvalidArgument
419 dir := rlookup(fs, dirname)
421 return os.ErrNotExist
425 dir.Child(name, func(node inode) inode {
430 if !recursive && node.IsDir() && node.Size() > 0 {
431 err = ErrDirectoryNotEmpty
439 type inode interface {
441 Read([]byte, filenodePtr) (int, filenodePtr, error)
442 Write([]byte, filenodePtr) (int, filenodePtr, error)
443 Truncate(int64) error
445 Readdir() []os.FileInfo
447 FileInfo() os.FileInfo
448 // Caller must have lock (or rlock if func is nil)
449 Child(string, func(inode) inode) inode
455 // filenode implements inode.
456 type filenode struct {
460 // number of times `segments` has changed in a
461 // way that might invalidate a filenodePtr
463 memsize int64 // bytes in memSegments
468 // filenodePtr is an offset into a file that is (usually) efficient to
469 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
471 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
472 // corresponds to file offset filenodePtr.off. Otherwise, it is
473 // necessary to reexamine len(filenode.segments[0]) etc. to find the
474 // correct segment and offset.
475 type filenodePtr struct {
482 // seek returns a ptr that is consistent with both startPtr.off and
483 // the current state of fn. The caller must already hold fn.RLock() or
486 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
491 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
493 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
494 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
497 // meaningless anyway
499 } else if ptr.off >= fn.fileinfo.size {
500 ptr.segmentIdx = len(fn.segments)
502 ptr.repacked = fn.repacked
504 } else if ptr.repacked == fn.repacked {
505 // segmentIdx and segmentOff accurately reflect
506 // ptr.off, but might have fallen off the end of a
508 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
515 ptr.repacked = fn.repacked
517 if ptr.off >= fn.fileinfo.size {
518 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
521 // Recompute segmentIdx and segmentOff. We have already
522 // established fn.fileinfo.size > ptr.off >= 0, so we don't
523 // have to deal with edge cases here.
525 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
526 // This would panic (index out of range) if
527 // fn.fileinfo.size were larger than
528 // sum(fn.segments[i].Len()) -- but that can't happen
529 // because we have ensured fn.fileinfo.size is always
531 segLen := int64(fn.segments[ptr.segmentIdx].Len())
532 if off+segLen > ptr.off {
533 ptr.segmentOff = int(ptr.off - off)
541 // caller must have lock
542 func (fn *filenode) appendSegment(e segment) {
543 fn.segments = append(fn.segments, e)
544 fn.fileinfo.size += int64(e.Len())
547 func (fn *filenode) Parent() inode {
553 // Read reads file data from a single segment, starting at startPtr,
554 // into p. startPtr is assumed not to be up-to-date. Caller must have
556 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
557 ptr = fn.seek(startPtr)
559 err = ErrNegativeOffset
562 if ptr.segmentIdx >= len(fn.segments) {
566 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
570 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
573 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
581 func (fn *filenode) Size() int64 {
584 return fn.fileinfo.Size()
587 func (fn *filenode) FileInfo() os.FileInfo {
593 func (fn *filenode) Truncate(size int64) error {
596 return fn.truncate(size)
599 func (fn *filenode) truncate(size int64) error {
600 if size == fn.fileinfo.size {
604 if size < fn.fileinfo.size {
605 ptr := fn.seek(filenodePtr{off: size})
606 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
607 if seg, ok := fn.segments[i].(*memSegment); ok {
608 fn.memsize -= int64(seg.Len())
611 if ptr.segmentOff == 0 {
612 fn.segments = fn.segments[:ptr.segmentIdx]
614 fn.segments = fn.segments[:ptr.segmentIdx+1]
615 switch seg := fn.segments[ptr.segmentIdx].(type) {
617 seg.Truncate(ptr.segmentOff)
618 fn.memsize += int64(seg.Len())
620 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
623 fn.fileinfo.size = size
626 for size > fn.fileinfo.size {
627 grow := size - fn.fileinfo.size
630 if len(fn.segments) == 0 {
632 fn.segments = append(fn.segments, seg)
633 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
635 fn.segments = append(fn.segments, seg)
637 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
640 seg.Truncate(seg.Len() + int(grow))
641 fn.fileinfo.size += grow
647 // Write writes data from p to the file, starting at startPtr,
648 // extending the file size if necessary. Caller must have Lock.
649 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
650 if startPtr.off > fn.fileinfo.size {
651 if err = fn.truncate(startPtr.off); err != nil {
652 return 0, startPtr, err
655 ptr = fn.seek(startPtr)
657 err = ErrNegativeOffset
660 for len(p) > 0 && err == nil {
662 if len(cando) > maxBlockSize {
663 cando = cando[:maxBlockSize]
665 // Rearrange/grow fn.segments (and shrink cando if
666 // needed) such that cando can be copied to
667 // fn.segments[ptr.segmentIdx] at offset
669 cur := ptr.segmentIdx
670 prev := ptr.segmentIdx - 1
672 if cur < len(fn.segments) {
673 _, curWritable = fn.segments[cur].(*memSegment)
675 var prevAppendable bool
676 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
677 _, prevAppendable = fn.segments[prev].(*memSegment)
679 if ptr.segmentOff > 0 && !curWritable {
680 // Split a non-writable block.
681 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
682 // Truncate cur, and insert a new
685 fn.segments = append(fn.segments, nil)
686 copy(fn.segments[cur+1:], fn.segments[cur:])
688 // Split cur into two copies, truncate
689 // the one on the left, shift the one
690 // on the right, and insert a new
691 // segment between them.
692 fn.segments = append(fn.segments, nil, nil)
693 copy(fn.segments[cur+2:], fn.segments[cur:])
694 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
699 seg.Truncate(len(cando))
700 fn.memsize += int64(len(cando))
701 fn.segments[cur] = seg
702 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
707 } else if curWritable {
708 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
713 // Shrink cando if needed to fit in
715 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
716 cando = cando[:cangrow]
720 if cur == len(fn.segments) {
721 // ptr is at EOF, filesize is changing.
722 fn.fileinfo.size += int64(len(cando))
723 } else if el := fn.segments[cur].Len(); el <= len(cando) {
724 // cando is long enough that we won't
725 // need cur any more. shrink cando to
726 // be exactly as long as cur
727 // (otherwise we'd accidentally shift
728 // the effective position of all
729 // segments after cur).
731 copy(fn.segments[cur:], fn.segments[cur+1:])
732 fn.segments = fn.segments[:len(fn.segments)-1]
734 // shrink cur by the same #bytes we're growing prev
735 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
741 ptr.segmentOff = fn.segments[prev].Len()
742 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
743 fn.memsize += int64(len(cando))
747 // Insert a segment between prev and
748 // cur, and advance prev/cur.
749 fn.segments = append(fn.segments, nil)
750 if cur < len(fn.segments) {
751 copy(fn.segments[cur+1:], fn.segments[cur:])
755 // appending a new segment does
756 // not invalidate any ptrs
759 seg.Truncate(len(cando))
760 fn.memsize += int64(len(cando))
761 fn.segments[cur] = seg
767 // Finally we can copy bytes from cando to the current segment.
768 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
772 ptr.off += int64(len(cando))
773 ptr.segmentOff += len(cando)
774 if ptr.segmentOff >= maxBlockSize {
775 fn.pruneMemSegments()
777 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
782 fn.fileinfo.modTime = time.Now()
787 // Write some data out to disk to reduce memory use. Caller must have
789 func (fn *filenode) pruneMemSegments() {
790 // TODO: async (don't hold Lock() while waiting for Keep)
791 // TODO: share code with (*dirnode)sync()
792 // TODO: pack/flush small blocks too, when fragmented
793 for idx, seg := range fn.segments {
794 seg, ok := seg.(*memSegment)
795 if !ok || seg.Len() < maxBlockSize {
798 locator, _, err := fn.parent.kc.PutB(seg.buf)
800 // TODO: stall (or return errors from)
801 // subsequent writes until flushing
805 fn.memsize -= int64(seg.Len())
806 fn.segments[idx] = storedSegment{
816 // FileSystem returns a CollectionFileSystem for the collection.
817 func (c *Collection) FileSystem(client *Client, kc keepClient) (CollectionFileSystem, error) {
818 var modTime time.Time
819 if c.ModifiedAt == nil {
822 modTime = *c.ModifiedAt
830 mode: os.ModeDir | 0755,
834 inodes: make(map[string]inode),
838 fs := &collectionFileSystem{fileSystem: fileSystem{inode: dn}}
839 if err := dn.loadManifest(c.ManifestText); err != nil {
845 type filehandle struct {
851 unreaddirs []os.FileInfo
854 func (f *filehandle) Read(p []byte) (n int, err error) {
856 return 0, ErrWriteOnlyMode
859 defer f.inode.RUnlock()
860 n, f.ptr, err = f.inode.Read(p, f.ptr)
864 func (f *filehandle) Seek(off int64, whence int) (pos int64, err error) {
865 size := f.inode.Size()
876 return f.ptr.off, ErrNegativeOffset
878 if ptr.off != f.ptr.off {
880 // force filenode to recompute f.ptr fields on next
884 return f.ptr.off, nil
887 func (f *filehandle) Truncate(size int64) error {
888 return f.inode.Truncate(size)
891 func (f *filehandle) Write(p []byte) (n int, err error) {
893 return 0, ErrReadOnlyFile
896 defer f.inode.Unlock()
897 if fn, ok := f.inode.(*filenode); ok && f.append {
899 off: fn.fileinfo.size,
900 segmentIdx: len(fn.segments),
902 repacked: fn.repacked,
905 n, f.ptr, err = f.inode.Write(p, f.ptr)
909 func (f *filehandle) Readdir(count int) ([]os.FileInfo, error) {
910 if !f.inode.IsDir() {
911 return nil, ErrInvalidOperation
914 return f.inode.Readdir(), nil
916 if f.unreaddirs == nil {
917 f.unreaddirs = f.inode.Readdir()
919 if len(f.unreaddirs) == 0 {
922 if count > len(f.unreaddirs) {
923 count = len(f.unreaddirs)
925 ret := f.unreaddirs[:count]
926 f.unreaddirs = f.unreaddirs[count:]
930 func (f *filehandle) Stat() (os.FileInfo, error) {
931 return f.inode.FileInfo(), nil
934 func (f *filehandle) Close() error {
938 type dirnode struct {
944 // sync flushes in-memory data (for all files in the tree rooted at
945 // dn) to persistent storage. Caller must hold dn.Lock().
946 func (dn *dirnode) sync() error {
947 type shortBlock struct {
951 var pending []shortBlock
954 flush := func(sbs []shortBlock) error {
958 block := make([]byte, 0, maxBlockSize)
959 for _, sb := range sbs {
960 block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
962 locator, _, err := dn.kc.PutB(block)
967 for _, sb := range sbs {
968 data := sb.fn.segments[sb.idx].(*memSegment).buf
969 sb.fn.segments[sb.idx] = storedSegment{
977 sb.fn.memsize -= int64(len(data))
982 names := make([]string, 0, len(dn.inodes))
983 for name := range dn.inodes {
984 names = append(names, name)
988 for _, name := range names {
989 fn, ok := dn.inodes[name].(*filenode)
995 for idx, seg := range fn.segments {
996 seg, ok := seg.(*memSegment)
1000 if seg.Len() > maxBlockSize/2 {
1001 if err := flush([]shortBlock{{fn, idx}}); err != nil {
1006 if pendingLen+seg.Len() > maxBlockSize {
1007 if err := flush(pending); err != nil {
1013 pending = append(pending, shortBlock{fn, idx})
1014 pendingLen += seg.Len()
1017 return flush(pending)
1020 // caller must have read lock.
1021 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
1023 type filepart struct {
1028 var fileparts []filepart
1032 if err := dn.sync(); err != nil {
1036 names := make([]string, 0, len(dn.inodes))
1037 for name, node := range dn.inodes {
1038 names = append(names, name)
1044 for _, name := range names {
1045 switch node := dn.inodes[name].(type) {
1047 subdir, err := node.marshalManifest(prefix + "/" + name)
1051 subdirs = subdirs + subdir
1053 if len(node.segments) == 0 {
1054 fileparts = append(fileparts, filepart{name: name})
1057 for _, seg := range node.segments {
1058 switch seg := seg.(type) {
1060 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
1061 streamLen -= int64(seg.size)
1063 blocks = append(blocks, seg.locator)
1067 offset: streamLen + int64(seg.offset),
1068 length: int64(seg.length),
1070 if prev := len(fileparts) - 1; prev >= 0 &&
1071 fileparts[prev].name == name &&
1072 fileparts[prev].offset+fileparts[prev].length == next.offset {
1073 fileparts[prev].length += next.length
1075 fileparts = append(fileparts, next)
1077 streamLen += int64(seg.size)
1079 // This can't happen: we
1080 // haven't unlocked since
1082 panic(fmt.Sprintf("can't marshal segment type %T", seg))
1086 panic(fmt.Sprintf("can't marshal inode type %T", node))
1089 var filetokens []string
1090 for _, s := range fileparts {
1091 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
1093 if len(filetokens) == 0 {
1095 } else if len(blocks) == 0 {
1096 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
1098 return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
1101 func (dn *dirnode) loadManifest(txt string) error {
1103 streams := strings.Split(txt, "\n")
1104 if streams[len(streams)-1] != "" {
1105 return fmt.Errorf("line %d: no trailing newline", len(streams))
1107 streams = streams[:len(streams)-1]
1108 segments := []storedSegment{}
1109 for i, stream := range streams {
1111 var anyFileTokens bool
1114 segments = segments[:0]
1115 for i, token := range strings.Split(stream, " ") {
1117 dirname = manifestUnescape(token)
1120 if !strings.Contains(token, ":") {
1122 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1124 toks := strings.SplitN(token, "+", 3)
1126 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1128 length, err := strconv.ParseInt(toks[1], 10, 32)
1129 if err != nil || length < 0 {
1130 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1132 segments = append(segments, storedSegment{
1136 length: int(length),
1139 } else if len(segments) == 0 {
1140 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1143 toks := strings.SplitN(token, ":", 3)
1145 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1147 anyFileTokens = true
1149 offset, err := strconv.ParseInt(toks[0], 10, 64)
1150 if err != nil || offset < 0 {
1151 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1153 length, err := strconv.ParseInt(toks[1], 10, 64)
1154 if err != nil || length < 0 {
1155 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1157 name := dirname + "/" + manifestUnescape(toks[2])
1158 fnode, err := dn.createFileAndParents(name)
1160 return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
1162 // Map the stream offset/range coordinates to
1163 // block/offset/range coordinates and add
1164 // corresponding storedSegments to the filenode
1166 // Can't continue where we left off.
1167 // TODO: binary search instead of
1168 // rewinding all the way (but this
1169 // situation might be rare anyway)
1172 for next := int64(0); segIdx < len(segments); segIdx++ {
1173 seg := segments[segIdx]
1174 next = pos + int64(seg.Len())
1175 if next <= offset || seg.Len() == 0 {
1179 if pos >= offset+length {
1184 blkOff = int(offset - pos)
1186 blkLen := seg.Len() - blkOff
1187 if pos+int64(blkOff+blkLen) > offset+length {
1188 blkLen = int(offset + length - pos - int64(blkOff))
1190 fnode.appendSegment(storedSegment{
1192 locator: seg.locator,
1197 if next > offset+length {
1203 if segIdx == len(segments) && pos < offset+length {
1204 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1208 return fmt.Errorf("line %d: no file segments", lineno)
1209 } else if len(segments) == 0 {
1210 return fmt.Errorf("line %d: no locators", lineno)
1211 } else if dirname == "" {
1212 return fmt.Errorf("line %d: no stream name", lineno)
1218 // only safe to call from loadManifest -- no locking
1219 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
1221 names := strings.Split(path, "/")
1222 basename := names[len(names)-1]
1223 if basename == "" || basename == "." || basename == ".." {
1224 err = fmt.Errorf("invalid filename")
1227 for _, name := range names[:len(names)-1] {
1233 // can't be sure parent will be a *dirnode
1234 return nil, ErrInvalidArgument
1236 node = node.Parent().(*dirnode)
1239 node.Child(name, func(child inode) inode {
1240 switch child.(type) {
1242 node, err = dn.newDirnode(node, name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
1245 node = child.(*dirnode)
1249 err = ErrInvalidOperation
1257 node.Child(basename, func(child inode) inode {
1258 switch child := child.(type) {
1260 fn, err = dn.newFilenode(node, basename, 0755, node.FileInfo().ModTime())
1266 err = ErrIsDirectory
1269 err = ErrInvalidOperation
1276 // rlookup (recursive lookup) returns the inode for the file/directory
1277 // with the given name (which may contain "/" separators). If no such
1278 // file/directory exists, the returned node is nil.
1279 func rlookup(start inode, path string) (node inode) {
1281 for _, name := range strings.Split(path, "/") {
1286 if name == "." || name == "" {
1290 node = node.Parent()
1294 node = func() inode {
1296 defer node.RUnlock()
1297 return node.Child(name, nil)
1303 // Caller must have lock, and must have already ensured
1304 // Children(name,nil) is nil.
1305 func (dn *dirnode) newDirnode(parent *dirnode, name string, perm os.FileMode, modTime time.Time) (node *dirnode, err error) {
1306 if name == "" || name == "." || name == ".." {
1307 return nil, ErrInvalidArgument
1316 mode: perm | os.ModeDir,
1319 inodes: make(map[string]inode),
1324 func (dn *dirnode) newFilenode(parent *dirnode, name string, perm os.FileMode, modTime time.Time) (node *filenode, err error) {
1325 if name == "" || name == "." || name == ".." {
1326 return nil, ErrInvalidArgument
1332 mode: perm & ^os.ModeDir,
1338 type segment interface {
1341 // Return a new segment with a subsection of the data from this
1342 // one. length<0 means length=Len()-off.
1343 Slice(off int, length int) segment
1346 type memSegment struct {
1350 func (me *memSegment) Len() int {
1354 func (me *memSegment) Slice(off, length int) segment {
1356 length = len(me.buf) - off
1358 buf := make([]byte, length)
1359 copy(buf, me.buf[off:])
1360 return &memSegment{buf: buf}
1363 func (me *memSegment) Truncate(n int) {
1364 if n > cap(me.buf) {
1367 newsize = newsize << 2
1369 newbuf := make([]byte, n, newsize)
1370 copy(newbuf, me.buf)
1373 // Zero unused part when shrinking, in case we grow
1374 // and start using it again later.
1375 for i := n; i < len(me.buf); i++ {
1382 func (me *memSegment) WriteAt(p []byte, off int) {
1383 if off+len(p) > len(me.buf) {
1384 panic("overflowed segment")
1386 copy(me.buf[off:], p)
1389 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1390 if off > int64(me.Len()) {
1394 n = copy(p, me.buf[int(off):])
1401 type storedSegment struct {
1404 size int // size of stored block (also encoded in locator)
1405 offset int // position of segment within the stored block
1406 length int // bytes in this segment (offset + length <= size)
1409 func (se storedSegment) Len() int {
1413 func (se storedSegment) Slice(n, size int) segment {
1416 if size >= 0 && se.length > size {
1422 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1423 if off > int64(se.length) {
1426 maxlen := se.length - int(off)
1427 if len(p) > maxlen {
1429 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1435 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1438 func canonicalName(name string) string {
1439 name = path.Clean("/" + name)
1440 if name == "/" || name == "./" {
1442 } else if strings.HasPrefix(name, "/") {
1448 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1450 func manifestUnescapeFunc(seq string) string {
1454 i, err := strconv.ParseUint(seq[1:], 8, 8)
1456 // Invalid escape sequence: can't unescape.
1459 return string([]byte{byte(i)})
1462 func manifestUnescape(s string) string {
1463 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1466 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1468 func manifestEscapeFunc(seq string) string {
1469 return fmt.Sprintf("\\%03o", byte(seq[0]))
1472 func manifestEscape(s string) string {
1473 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)