1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
24 maxBlockSize = 1 << 26
25 concurrentWriters = 4 // max goroutines writing to Keep in background and during flush()
28 // A CollectionFileSystem is a FileSystem that can be serialized as a
29 // manifest and stored as a collection.
30 type CollectionFileSystem interface {
33 // Flush all file data to Keep and return a snapshot of the
34 // filesystem suitable for saving as (Collection)ManifestText.
35 // Prefix (normally ".") is a top level directory, effectively
36 // prepended to all paths in the returned manifest.
37 MarshalManifest(prefix string) (string, error)
39 // Total data bytes in all files.
42 // Memory consumed by buffered file data.
46 type collectionFileSystem struct {
51 // FileSystem returns a CollectionFileSystem for the collection.
52 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
53 modTime := c.ModifiedAt
57 fs := &collectionFileSystem{
59 fileSystem: fileSystem{
60 fsBackend: keepBackend{apiClient: client, keepClient: kc},
61 thr: newThrottle(concurrentWriters),
69 mode: os.ModeDir | 0755,
72 inodes: make(map[string]inode),
75 root.SetParent(root, ".")
76 if err := root.loadManifest(c.ManifestText); err != nil {
79 backdateTree(root, modTime)
84 func backdateTree(n inode, modTime time.Time) {
85 switch n := n.(type) {
87 n.fileinfo.modTime = modTime
89 n.fileinfo.modTime = modTime
90 for _, n := range n.inodes {
91 backdateTree(n, modTime)
96 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
97 if name == "" || name == "." || name == ".." {
98 return nil, ErrInvalidArgument
106 mode: perm | os.ModeDir,
109 inodes: make(map[string]inode),
117 mode: perm & ^os.ModeDir,
124 func (fs *collectionFileSystem) Sync() error {
128 txt, err := fs.MarshalManifest(".")
130 return fmt.Errorf("sync failed: %s", err)
136 err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
137 "collection": map[string]string{
138 "manifest_text": coll.ManifestText,
140 "select": []string{"uuid"},
143 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
148 func (fs *collectionFileSystem) Flush(path string, shortBlocks bool) error {
149 node, err := rlookup(fs.fileSystem.root, path)
153 dn, ok := node.(*dirnode)
155 return ErrNotADirectory
159 names := dn.sortedNames()
161 // Caller only wants to flush the specified dir,
162 // non-recursively. Drop subdirs from the list of
164 var filenames []string
165 for _, name := range names {
166 if _, ok := dn.inodes[name].(*filenode); ok {
167 filenames = append(filenames, name)
172 for _, name := range names {
173 child := dn.inodes[name]
177 return dn.flush(context.TODO(), names, flushOpts{sync: false, shortBlocks: shortBlocks})
180 func (fs *collectionFileSystem) memorySize() int64 {
181 fs.fileSystem.root.Lock()
182 defer fs.fileSystem.root.Unlock()
183 return fs.fileSystem.root.(*dirnode).memorySize()
186 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
187 fs.fileSystem.root.Lock()
188 defer fs.fileSystem.root.Unlock()
189 return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix)
192 func (fs *collectionFileSystem) Size() int64 {
193 return fs.fileSystem.root.(*dirnode).TreeSize()
196 // filenodePtr is an offset into a file that is (usually) efficient to
197 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
199 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
200 // corresponds to file offset filenodePtr.off. Otherwise, it is
201 // necessary to reexamine len(filenode.segments[0]) etc. to find the
202 // correct segment and offset.
203 type filenodePtr struct {
210 // seek returns a ptr that is consistent with both startPtr.off and
211 // the current state of fn. The caller must already hold fn.RLock() or
214 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
219 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
221 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
222 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
225 // meaningless anyway
227 } else if ptr.off >= fn.fileinfo.size {
228 ptr.segmentIdx = len(fn.segments)
230 ptr.repacked = fn.repacked
232 } else if ptr.repacked == fn.repacked {
233 // segmentIdx and segmentOff accurately reflect
234 // ptr.off, but might have fallen off the end of a
236 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
243 ptr.repacked = fn.repacked
245 if ptr.off >= fn.fileinfo.size {
246 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
249 // Recompute segmentIdx and segmentOff. We have already
250 // established fn.fileinfo.size > ptr.off >= 0, so we don't
251 // have to deal with edge cases here.
253 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
254 // This would panic (index out of range) if
255 // fn.fileinfo.size were larger than
256 // sum(fn.segments[i].Len()) -- but that can't happen
257 // because we have ensured fn.fileinfo.size is always
259 segLen := int64(fn.segments[ptr.segmentIdx].Len())
260 if off+segLen > ptr.off {
261 ptr.segmentOff = int(ptr.off - off)
269 // filenode implements inode.
270 type filenode struct {
275 // number of times `segments` has changed in a
276 // way that might invalidate a filenodePtr
278 memsize int64 // bytes in memSegments
283 // caller must have lock
284 func (fn *filenode) appendSegment(e segment) {
285 fn.segments = append(fn.segments, e)
286 fn.fileinfo.size += int64(e.Len())
289 func (fn *filenode) SetParent(p inode, name string) {
293 fn.fileinfo.name = name
296 func (fn *filenode) Parent() inode {
302 func (fn *filenode) FS() FileSystem {
306 // Read reads file data from a single segment, starting at startPtr,
307 // into p. startPtr is assumed not to be up-to-date. Caller must have
309 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
310 ptr = fn.seek(startPtr)
312 err = ErrNegativeOffset
315 if ptr.segmentIdx >= len(fn.segments) {
319 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
323 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
326 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
334 func (fn *filenode) Size() int64 {
337 return fn.fileinfo.Size()
340 func (fn *filenode) FileInfo() os.FileInfo {
346 func (fn *filenode) Truncate(size int64) error {
349 return fn.truncate(size)
352 func (fn *filenode) truncate(size int64) error {
353 if size == fn.fileinfo.size {
357 if size < fn.fileinfo.size {
358 ptr := fn.seek(filenodePtr{off: size})
359 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
360 if seg, ok := fn.segments[i].(*memSegment); ok {
361 fn.memsize -= int64(seg.Len())
364 if ptr.segmentOff == 0 {
365 fn.segments = fn.segments[:ptr.segmentIdx]
367 fn.segments = fn.segments[:ptr.segmentIdx+1]
368 switch seg := fn.segments[ptr.segmentIdx].(type) {
370 seg.Truncate(ptr.segmentOff)
371 fn.memsize += int64(seg.Len())
373 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
376 fn.fileinfo.size = size
379 for size > fn.fileinfo.size {
380 grow := size - fn.fileinfo.size
383 if len(fn.segments) == 0 {
385 fn.segments = append(fn.segments, seg)
386 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
388 fn.segments = append(fn.segments, seg)
390 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
393 seg.Truncate(seg.Len() + int(grow))
394 fn.fileinfo.size += grow
400 // Write writes data from p to the file, starting at startPtr,
401 // extending the file size if necessary. Caller must have Lock.
402 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
403 if startPtr.off > fn.fileinfo.size {
404 if err = fn.truncate(startPtr.off); err != nil {
405 return 0, startPtr, err
408 ptr = fn.seek(startPtr)
410 err = ErrNegativeOffset
413 for len(p) > 0 && err == nil {
415 if len(cando) > maxBlockSize {
416 cando = cando[:maxBlockSize]
418 // Rearrange/grow fn.segments (and shrink cando if
419 // needed) such that cando can be copied to
420 // fn.segments[ptr.segmentIdx] at offset
422 cur := ptr.segmentIdx
423 prev := ptr.segmentIdx - 1
425 if cur < len(fn.segments) {
426 _, curWritable = fn.segments[cur].(*memSegment)
428 var prevAppendable bool
429 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
430 _, prevAppendable = fn.segments[prev].(*memSegment)
432 if ptr.segmentOff > 0 && !curWritable {
433 // Split a non-writable block.
434 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
435 // Truncate cur, and insert a new
438 fn.segments = append(fn.segments, nil)
439 copy(fn.segments[cur+1:], fn.segments[cur:])
441 // Split cur into two copies, truncate
442 // the one on the left, shift the one
443 // on the right, and insert a new
444 // segment between them.
445 fn.segments = append(fn.segments, nil, nil)
446 copy(fn.segments[cur+2:], fn.segments[cur:])
447 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
452 seg.Truncate(len(cando))
453 fn.memsize += int64(len(cando))
454 fn.segments[cur] = seg
455 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
460 } else if curWritable {
461 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
466 // Shrink cando if needed to fit in
468 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
469 cando = cando[:cangrow]
473 if cur == len(fn.segments) {
474 // ptr is at EOF, filesize is changing.
475 fn.fileinfo.size += int64(len(cando))
476 } else if el := fn.segments[cur].Len(); el <= len(cando) {
477 // cando is long enough that we won't
478 // need cur any more. shrink cando to
479 // be exactly as long as cur
480 // (otherwise we'd accidentally shift
481 // the effective position of all
482 // segments after cur).
484 copy(fn.segments[cur:], fn.segments[cur+1:])
485 fn.segments = fn.segments[:len(fn.segments)-1]
487 // shrink cur by the same #bytes we're growing prev
488 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
494 ptr.segmentOff = fn.segments[prev].Len()
495 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
496 fn.memsize += int64(len(cando))
500 // Insert a segment between prev and
501 // cur, and advance prev/cur.
502 fn.segments = append(fn.segments, nil)
503 if cur < len(fn.segments) {
504 copy(fn.segments[cur+1:], fn.segments[cur:])
508 // appending a new segment does
509 // not invalidate any ptrs
512 seg.Truncate(len(cando))
513 fn.memsize += int64(len(cando))
514 fn.segments[cur] = seg
520 // Finally we can copy bytes from cando to the current segment.
521 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
525 ptr.off += int64(len(cando))
526 ptr.segmentOff += len(cando)
527 if ptr.segmentOff >= maxBlockSize {
528 fn.pruneMemSegments()
530 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
535 fn.fileinfo.modTime = time.Now()
540 // Write some data out to disk to reduce memory use. Caller must have
542 func (fn *filenode) pruneMemSegments() {
543 // TODO: share code with (*dirnode)flush()
544 // TODO: pack/flush small blocks too, when fragmented
545 for idx, seg := range fn.segments {
546 seg, ok := seg.(*memSegment)
547 if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
550 // Setting seg.flushing guarantees seg.buf will not be
551 // modified in place: WriteAt and Truncate will
552 // allocate a new buf instead, if necessary.
553 idx, buf := idx, seg.buf
554 done := make(chan struct{})
556 // If lots of background writes are already in
557 // progress, block here until one finishes, rather
558 // than pile up an unlimited number of buffered writes
559 // and network flush operations.
560 fn.fs.throttle().Acquire()
563 locator, _, err := fn.FS().PutB(buf)
564 fn.fs.throttle().Release()
567 if seg.flushing != done {
568 // A new seg.buf has been allocated.
573 // TODO: stall (or return errors from)
574 // subsequent writes until flushing
575 // starts to succeed.
578 if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
579 // Segment has been dropped/moved/resized.
582 fn.memsize -= int64(len(buf))
583 fn.segments[idx] = storedSegment{
594 // Block until all pending pruneMemSegments/flush work is
595 // finished. Caller must NOT have lock.
596 func (fn *filenode) waitPrune() {
597 var pending []<-chan struct{}
599 for _, seg := range fn.segments {
600 if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
601 pending = append(pending, seg.flushing)
605 for _, p := range pending {
610 type dirnode struct {
611 fs *collectionFileSystem
615 func (dn *dirnode) FS() FileSystem {
619 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
620 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
621 gn := &getternode{Getter: func() ([]byte, error) {
624 coll.ManifestText, err = dn.fs.MarshalManifest(".")
628 data, err := json.Marshal(&coll)
630 data = append(data, '\n')
634 gn.SetParent(dn, name)
637 return dn.treenode.Child(name, replace)
640 type fnSegmentRef struct {
645 // commitBlock concatenates the data from the given filenode segments
646 // (which must be *memSegments), writes the data out to Keep as a
647 // single block, and replaces the filenodes' *memSegments with
648 // storedSegments that reference the relevant portions of the new
651 // bufsize is the total data size in refs. It is used to preallocate
652 // the correct amount of memory when len(refs)>1.
654 // If sync is false, commitBlock returns right away, after starting a
655 // goroutine to do the writes, reacquire the filenodes' locks, and
656 // swap out the *memSegments. Some filenodes' segments might get
657 // modified/rearranged in the meantime, in which case commitBlock
658 // won't replace them.
660 // Caller must have write lock.
661 func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize int, sync bool) error {
665 if err := ctx.Err(); err != nil {
668 done := make(chan struct{})
670 segs := make([]*memSegment, 0, len(refs))
671 offsets := make([]int, 0, len(refs)) // location of segment's data within block
672 for _, ref := range refs {
673 seg := ref.fn.segments[ref.idx].(*memSegment)
674 if seg.flushing != nil && !sync {
675 // Let the other flushing goroutine finish. If
676 // it fails, we'll try again next time.
679 // In sync mode, we proceed regardless of
680 // whether another flush is in progress: It
681 // can't finish before we do, because we hold
682 // fn's lock until we finish our own writes.
685 offsets = append(offsets, len(block))
688 } else if block == nil {
689 block = append(make([]byte, 0, bufsize), seg.buf...)
691 block = append(block, seg.buf...)
693 segs = append(segs, seg)
695 blocksize := len(block)
696 dn.fs.throttle().Acquire()
697 errs := make(chan error, 1)
701 locator, _, err := dn.fs.PutB(block)
702 dn.fs.throttle().Release()
705 for _, seg := range segs {
706 if seg.flushing == done {
716 for idx, ref := range refs {
719 // In async mode, fn's lock was
720 // released while we were waiting for
721 // PutB(); lots of things might have
723 if len(ref.fn.segments) <= ref.idx {
724 // file segments have
725 // rearranged or changed in
729 } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] {
730 // segment has been replaced
733 } else if seg.flushing != done {
734 // seg.buf has been replaced
739 data := ref.fn.segments[ref.idx].(*memSegment).buf
740 ref.fn.segments[ref.idx] = storedSegment{
744 offset: offsets[idx],
747 // atomic is needed here despite caller having
748 // lock: caller might be running concurrent
749 // commitBlock() goroutines using the same
750 // lock, writing different segments from the
752 atomic.AddInt64(&ref.fn.memsize, -int64(len(data)))
765 type flushOpts struct {
770 // flush in-memory data and remote-cluster block references (for the
771 // children with the given names, which must be children of dn) to
772 // local-cluster persistent storage.
774 // Caller must have write lock on dn and the named children.
776 // If any children are dirs, they will be flushed recursively.
777 func (dn *dirnode) flush(ctx context.Context, names []string, opts flushOpts) error {
778 cg := newContextGroup(ctx)
781 goCommit := func(refs []fnSegmentRef, bufsize int) {
783 return dn.commitBlock(cg.Context(), refs, bufsize, opts.sync)
787 var pending []fnSegmentRef
788 var pendingLen int = 0
789 localLocator := map[string]string{}
790 for _, name := range names {
791 switch node := dn.inodes[name].(type) {
793 grandchildNames := node.sortedNames()
794 for _, grandchildName := range grandchildNames {
795 grandchild := node.inodes[grandchildName]
797 defer grandchild.Unlock()
799 cg.Go(func() error { return node.flush(cg.Context(), grandchildNames, opts) })
801 for idx, seg := range node.segments {
802 switch seg := seg.(type) {
804 loc, ok := localLocator[seg.locator]
807 loc, err = dn.fs.LocalLocator(seg.locator)
811 localLocator[seg.locator] = loc
814 node.segments[idx] = seg
816 if seg.Len() > maxBlockSize/2 {
817 goCommit([]fnSegmentRef{{node, idx}}, seg.Len())
820 if pendingLen+seg.Len() > maxBlockSize {
821 goCommit(pending, pendingLen)
825 pending = append(pending, fnSegmentRef{node, idx})
826 pendingLen += seg.Len()
828 panic(fmt.Sprintf("can't sync segment type %T", seg))
833 if opts.shortBlocks {
834 goCommit(pending, pendingLen)
839 // caller must have write lock.
840 func (dn *dirnode) memorySize() (size int64) {
841 for _, name := range dn.sortedNames() {
842 node := dn.inodes[name]
845 switch node := node.(type) {
847 size += node.memorySize()
849 for _, seg := range node.segments {
850 switch seg := seg.(type) {
852 size += int64(seg.Len())
860 // caller must have write lock.
861 func (dn *dirnode) sortedNames() []string {
862 names := make([]string, 0, len(dn.inodes))
863 for name := range dn.inodes {
864 names = append(names, name)
870 // caller must have write lock.
871 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string) (string, error) {
872 cg := newContextGroup(ctx)
875 if len(dn.inodes) == 0 {
879 // Express the existence of an empty directory by
880 // adding an empty file named `\056`, which (unlike
881 // the more obvious spelling `.`) is accepted by the
882 // API's manifest validator.
883 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
886 names := dn.sortedNames()
888 // Wait for children to finish any pending write operations
889 // before locking them.
890 for _, name := range names {
891 node := dn.inodes[name]
892 if fn, ok := node.(*filenode); ok {
897 var dirnames []string
898 var filenames []string
899 for _, name := range names {
900 node := dn.inodes[name]
903 switch node := node.(type) {
905 dirnames = append(dirnames, name)
907 filenames = append(filenames, name)
909 panic(fmt.Sprintf("can't marshal inode type %T", node))
913 subdirs := make([]string, len(dirnames))
915 for i, name := range dirnames {
918 txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name)
926 type filepart struct {
932 var fileparts []filepart
934 if err := dn.flush(cg.Context(), filenames, flushOpts{sync: true, shortBlocks: true}); err != nil {
937 for _, name := range filenames {
938 node := dn.inodes[name].(*filenode)
939 if len(node.segments) == 0 {
940 fileparts = append(fileparts, filepart{name: name})
943 for _, seg := range node.segments {
944 switch seg := seg.(type) {
946 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
947 streamLen -= int64(seg.size)
949 blocks = append(blocks, seg.locator)
953 offset: streamLen + int64(seg.offset),
954 length: int64(seg.length),
956 if prev := len(fileparts) - 1; prev >= 0 &&
957 fileparts[prev].name == name &&
958 fileparts[prev].offset+fileparts[prev].length == next.offset {
959 fileparts[prev].length += next.length
961 fileparts = append(fileparts, next)
963 streamLen += int64(seg.size)
965 // This can't happen: we
966 // haven't unlocked since
967 // calling flush(sync=true).
968 panic(fmt.Sprintf("can't marshal segment type %T", seg))
972 var filetokens []string
973 for _, s := range fileparts {
974 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
976 if len(filetokens) == 0 {
978 } else if len(blocks) == 0 {
979 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
981 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
985 return rootdir + strings.Join(subdirs, ""), err
988 func (dn *dirnode) loadManifest(txt string) error {
990 streams := strings.Split(txt, "\n")
991 if streams[len(streams)-1] != "" {
992 return fmt.Errorf("line %d: no trailing newline", len(streams))
994 streams = streams[:len(streams)-1]
995 segments := []storedSegment{}
996 for i, stream := range streams {
998 var anyFileTokens bool
1001 segments = segments[:0]
1002 for i, token := range strings.Split(stream, " ") {
1004 dirname = manifestUnescape(token)
1007 if !strings.Contains(token, ":") {
1009 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1011 toks := strings.SplitN(token, "+", 3)
1013 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1015 length, err := strconv.ParseInt(toks[1], 10, 32)
1016 if err != nil || length < 0 {
1017 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1019 segments = append(segments, storedSegment{
1023 length: int(length),
1026 } else if len(segments) == 0 {
1027 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1030 toks := strings.SplitN(token, ":", 3)
1032 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1034 anyFileTokens = true
1036 offset, err := strconv.ParseInt(toks[0], 10, 64)
1037 if err != nil || offset < 0 {
1038 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1040 length, err := strconv.ParseInt(toks[1], 10, 64)
1041 if err != nil || length < 0 {
1042 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1044 name := dirname + "/" + manifestUnescape(toks[2])
1045 fnode, err := dn.createFileAndParents(name)
1046 if fnode == nil && err == nil && length == 0 {
1047 // Special case: an empty file used as
1048 // a marker to preserve an otherwise
1049 // empty directory in a manifest.
1052 if err != nil || (fnode == nil && length != 0) {
1053 return fmt.Errorf("line %d: cannot use path %q with length %d: %s", lineno, name, length, err)
1055 // Map the stream offset/range coordinates to
1056 // block/offset/range coordinates and add
1057 // corresponding storedSegments to the filenode
1059 // Can't continue where we left off.
1060 // TODO: binary search instead of
1061 // rewinding all the way (but this
1062 // situation might be rare anyway)
1065 for next := int64(0); segIdx < len(segments); segIdx++ {
1066 seg := segments[segIdx]
1067 next = pos + int64(seg.Len())
1068 if next <= offset || seg.Len() == 0 {
1072 if pos >= offset+length {
1077 blkOff = int(offset - pos)
1079 blkLen := seg.Len() - blkOff
1080 if pos+int64(blkOff+blkLen) > offset+length {
1081 blkLen = int(offset + length - pos - int64(blkOff))
1083 fnode.appendSegment(storedSegment{
1085 locator: seg.locator,
1090 if next > offset+length {
1096 if segIdx == len(segments) && pos < offset+length {
1097 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1101 return fmt.Errorf("line %d: no file segments", lineno)
1102 } else if len(segments) == 0 {
1103 return fmt.Errorf("line %d: no locators", lineno)
1104 } else if dirname == "" {
1105 return fmt.Errorf("line %d: no stream name", lineno)
1111 // only safe to call from loadManifest -- no locking.
1113 // If path is a "parent directory exists" marker (the last path
1114 // component is "."), the returned values are both nil.
1115 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
1117 names := strings.Split(path, "/")
1118 basename := names[len(names)-1]
1119 for _, name := range names[:len(names)-1] {
1125 // can't be sure parent will be a *dirnode
1126 return nil, ErrInvalidArgument
1128 node = node.Parent()
1131 node, err = node.Child(name, func(child inode) (inode, error) {
1133 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
1137 child.SetParent(node, name)
1139 } else if !child.IsDir() {
1140 return child, ErrFileExists
1149 if basename == "." {
1151 } else if !permittedName(basename) {
1152 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
1155 _, err = node.Child(basename, func(child inode) (inode, error) {
1156 switch child := child.(type) {
1158 child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
1162 child.SetParent(node, basename)
1163 fn = child.(*filenode)
1169 return child, ErrIsDirectory
1171 return child, ErrInvalidArgument
1177 func (dn *dirnode) TreeSize() (bytes int64) {
1180 for _, i := range dn.inodes {
1181 switch i := i.(type) {
1185 bytes += i.TreeSize()
1191 type segment interface {
1194 // Return a new segment with a subsection of the data from this
1195 // one. length<0 means length=Len()-off.
1196 Slice(off int, length int) segment
1199 type memSegment struct {
1201 // If flushing is not nil, then a) buf is being shared by a
1202 // pruneMemSegments goroutine, and must be copied on write;
1203 // and b) the flushing channel will close when the goroutine
1204 // finishes, whether it succeeds or not.
1205 flushing <-chan struct{}
1208 func (me *memSegment) Len() int {
1212 func (me *memSegment) Slice(off, length int) segment {
1214 length = len(me.buf) - off
1216 buf := make([]byte, length)
1217 copy(buf, me.buf[off:])
1218 return &memSegment{buf: buf}
1221 func (me *memSegment) Truncate(n int) {
1222 if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
1225 newsize = newsize << 2
1227 newbuf := make([]byte, n, newsize)
1228 copy(newbuf, me.buf)
1229 me.buf, me.flushing = newbuf, nil
1231 // reclaim existing capacity, and zero reclaimed part
1232 oldlen := len(me.buf)
1234 for i := oldlen; i < n; i++ {
1240 func (me *memSegment) WriteAt(p []byte, off int) {
1241 if off+len(p) > len(me.buf) {
1242 panic("overflowed segment")
1244 if me.flushing != nil {
1245 me.buf, me.flushing = append([]byte(nil), me.buf...), nil
1247 copy(me.buf[off:], p)
1250 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1251 if off > int64(me.Len()) {
1255 n = copy(p, me.buf[int(off):])
1262 type storedSegment struct {
1265 size int // size of stored block (also encoded in locator)
1266 offset int // position of segment within the stored block
1267 length int // bytes in this segment (offset + length <= size)
1270 func (se storedSegment) Len() int {
1274 func (se storedSegment) Slice(n, size int) segment {
1277 if size >= 0 && se.length > size {
1283 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1284 if off > int64(se.length) {
1287 maxlen := se.length - int(off)
1288 if len(p) > maxlen {
1290 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1296 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1299 func canonicalName(name string) string {
1300 name = path.Clean("/" + name)
1301 if name == "/" || name == "./" {
1303 } else if strings.HasPrefix(name, "/") {
1309 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1311 func manifestUnescapeFunc(seq string) string {
1315 i, err := strconv.ParseUint(seq[1:], 8, 8)
1317 // Invalid escape sequence: can't unescape.
1320 return string([]byte{byte(i)})
1323 func manifestUnescape(s string) string {
1324 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1327 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1329 func manifestEscapeFunc(seq string) string {
1330 return fmt.Sprintf("\\%03o", byte(seq[0]))
1333 func manifestEscape(s string) string {
1334 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)