1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
21 var maxBlockSize = 1 << 26
23 // A CollectionFileSystem is a FileSystem that can be serialized as a
24 // manifest and stored as a collection.
25 type CollectionFileSystem interface {
28 // Flush all file data to Keep and return a snapshot of the
29 // filesystem suitable for saving as (Collection)ManifestText.
30 // Prefix (normally ".") is a top level directory, effectively
31 // prepended to all paths in the returned manifest.
32 MarshalManifest(prefix string) (string, error)
34 // Total data bytes in all files.
38 type collectionFileSystem struct {
43 // FileSystem returns a CollectionFileSystem for the collection.
44 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
46 if c.ModifiedAt == nil {
49 modTime = *c.ModifiedAt
51 fs := &collectionFileSystem{
53 fileSystem: fileSystem{
54 fsBackend: keepBackend{apiClient: client, keepClient: kc},
62 mode: os.ModeDir | 0755,
65 inodes: make(map[string]inode),
68 root.SetParent(root, ".")
69 if err := root.loadManifest(c.ManifestText); err != nil {
72 backdateTree(root, modTime)
77 func backdateTree(n inode, modTime time.Time) {
78 switch n := n.(type) {
80 n.fileinfo.modTime = modTime
82 n.fileinfo.modTime = modTime
83 for _, n := range n.inodes {
84 backdateTree(n, modTime)
89 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
90 if name == "" || name == "." || name == ".." {
91 return nil, ErrInvalidArgument
99 mode: perm | os.ModeDir,
102 inodes: make(map[string]inode),
110 mode: perm & ^os.ModeDir,
117 func (fs *collectionFileSystem) Sync() error {
121 txt, err := fs.MarshalManifest(".")
123 return fmt.Errorf("sync failed: %s", err)
129 err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
131 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
136 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
137 fs.fileSystem.root.Lock()
138 defer fs.fileSystem.root.Unlock()
139 return fs.fileSystem.root.(*dirnode).marshalManifest(prefix)
142 func (fs *collectionFileSystem) Size() int64 {
143 return fs.fileSystem.root.(*dirnode).TreeSize()
146 // filenodePtr is an offset into a file that is (usually) efficient to
147 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
149 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
150 // corresponds to file offset filenodePtr.off. Otherwise, it is
151 // necessary to reexamine len(filenode.segments[0]) etc. to find the
152 // correct segment and offset.
153 type filenodePtr struct {
160 // seek returns a ptr that is consistent with both startPtr.off and
161 // the current state of fn. The caller must already hold fn.RLock() or
164 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
169 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
171 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
172 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
175 // meaningless anyway
177 } else if ptr.off >= fn.fileinfo.size {
178 ptr.segmentIdx = len(fn.segments)
180 ptr.repacked = fn.repacked
182 } else if ptr.repacked == fn.repacked {
183 // segmentIdx and segmentOff accurately reflect
184 // ptr.off, but might have fallen off the end of a
186 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
193 ptr.repacked = fn.repacked
195 if ptr.off >= fn.fileinfo.size {
196 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
199 // Recompute segmentIdx and segmentOff. We have already
200 // established fn.fileinfo.size > ptr.off >= 0, so we don't
201 // have to deal with edge cases here.
203 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
204 // This would panic (index out of range) if
205 // fn.fileinfo.size were larger than
206 // sum(fn.segments[i].Len()) -- but that can't happen
207 // because we have ensured fn.fileinfo.size is always
209 segLen := int64(fn.segments[ptr.segmentIdx].Len())
210 if off+segLen > ptr.off {
211 ptr.segmentOff = int(ptr.off - off)
219 // filenode implements inode.
220 type filenode struct {
225 // number of times `segments` has changed in a
226 // way that might invalidate a filenodePtr
228 memsize int64 // bytes in memSegments
233 // caller must have lock
234 func (fn *filenode) appendSegment(e segment) {
235 fn.segments = append(fn.segments, e)
236 fn.fileinfo.size += int64(e.Len())
239 func (fn *filenode) SetParent(p inode, name string) {
243 fn.fileinfo.name = name
246 func (fn *filenode) Parent() inode {
252 func (fn *filenode) FS() FileSystem {
256 // Read reads file data from a single segment, starting at startPtr,
257 // into p. startPtr is assumed not to be up-to-date. Caller must have
259 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
260 ptr = fn.seek(startPtr)
262 err = ErrNegativeOffset
265 if ptr.segmentIdx >= len(fn.segments) {
269 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
273 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
276 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
284 func (fn *filenode) Size() int64 {
287 return fn.fileinfo.Size()
290 func (fn *filenode) FileInfo() os.FileInfo {
296 func (fn *filenode) Truncate(size int64) error {
299 return fn.truncate(size)
302 func (fn *filenode) truncate(size int64) error {
303 if size == fn.fileinfo.size {
307 if size < fn.fileinfo.size {
308 ptr := fn.seek(filenodePtr{off: size})
309 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
310 if seg, ok := fn.segments[i].(*memSegment); ok {
311 fn.memsize -= int64(seg.Len())
314 if ptr.segmentOff == 0 {
315 fn.segments = fn.segments[:ptr.segmentIdx]
317 fn.segments = fn.segments[:ptr.segmentIdx+1]
318 switch seg := fn.segments[ptr.segmentIdx].(type) {
320 seg.Truncate(ptr.segmentOff)
321 fn.memsize += int64(seg.Len())
323 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
326 fn.fileinfo.size = size
329 for size > fn.fileinfo.size {
330 grow := size - fn.fileinfo.size
333 if len(fn.segments) == 0 {
335 fn.segments = append(fn.segments, seg)
336 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
338 fn.segments = append(fn.segments, seg)
340 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
343 seg.Truncate(seg.Len() + int(grow))
344 fn.fileinfo.size += grow
350 // Write writes data from p to the file, starting at startPtr,
351 // extending the file size if necessary. Caller must have Lock.
352 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
353 if startPtr.off > fn.fileinfo.size {
354 if err = fn.truncate(startPtr.off); err != nil {
355 return 0, startPtr, err
358 ptr = fn.seek(startPtr)
360 err = ErrNegativeOffset
363 for len(p) > 0 && err == nil {
365 if len(cando) > maxBlockSize {
366 cando = cando[:maxBlockSize]
368 // Rearrange/grow fn.segments (and shrink cando if
369 // needed) such that cando can be copied to
370 // fn.segments[ptr.segmentIdx] at offset
372 cur := ptr.segmentIdx
373 prev := ptr.segmentIdx - 1
375 if cur < len(fn.segments) {
376 _, curWritable = fn.segments[cur].(*memSegment)
378 var prevAppendable bool
379 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
380 _, prevAppendable = fn.segments[prev].(*memSegment)
382 if ptr.segmentOff > 0 && !curWritable {
383 // Split a non-writable block.
384 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
385 // Truncate cur, and insert a new
388 fn.segments = append(fn.segments, nil)
389 copy(fn.segments[cur+1:], fn.segments[cur:])
391 // Split cur into two copies, truncate
392 // the one on the left, shift the one
393 // on the right, and insert a new
394 // segment between them.
395 fn.segments = append(fn.segments, nil, nil)
396 copy(fn.segments[cur+2:], fn.segments[cur:])
397 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
402 seg.Truncate(len(cando))
403 fn.memsize += int64(len(cando))
404 fn.segments[cur] = seg
405 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
410 } else if curWritable {
411 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
416 // Shrink cando if needed to fit in
418 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
419 cando = cando[:cangrow]
423 if cur == len(fn.segments) {
424 // ptr is at EOF, filesize is changing.
425 fn.fileinfo.size += int64(len(cando))
426 } else if el := fn.segments[cur].Len(); el <= len(cando) {
427 // cando is long enough that we won't
428 // need cur any more. shrink cando to
429 // be exactly as long as cur
430 // (otherwise we'd accidentally shift
431 // the effective position of all
432 // segments after cur).
434 copy(fn.segments[cur:], fn.segments[cur+1:])
435 fn.segments = fn.segments[:len(fn.segments)-1]
437 // shrink cur by the same #bytes we're growing prev
438 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
444 ptr.segmentOff = fn.segments[prev].Len()
445 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
446 fn.memsize += int64(len(cando))
450 // Insert a segment between prev and
451 // cur, and advance prev/cur.
452 fn.segments = append(fn.segments, nil)
453 if cur < len(fn.segments) {
454 copy(fn.segments[cur+1:], fn.segments[cur:])
458 // appending a new segment does
459 // not invalidate any ptrs
462 seg.Truncate(len(cando))
463 fn.memsize += int64(len(cando))
464 fn.segments[cur] = seg
470 // Finally we can copy bytes from cando to the current segment.
471 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
475 ptr.off += int64(len(cando))
476 ptr.segmentOff += len(cando)
477 if ptr.segmentOff >= maxBlockSize {
478 fn.pruneMemSegments()
480 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
485 fn.fileinfo.modTime = time.Now()
490 // Write some data out to disk to reduce memory use. Caller must have
492 func (fn *filenode) pruneMemSegments() {
493 // TODO: async (don't hold Lock() while waiting for Keep)
494 // TODO: share code with (*dirnode)sync()
495 // TODO: pack/flush small blocks too, when fragmented
496 for idx, seg := range fn.segments {
497 seg, ok := seg.(*memSegment)
498 if !ok || seg.Len() < maxBlockSize {
501 locator, _, err := fn.FS().PutB(seg.buf)
503 // TODO: stall (or return errors from)
504 // subsequent writes until flushing
508 fn.memsize -= int64(seg.Len())
509 fn.segments[idx] = storedSegment{
519 type dirnode struct {
520 fs *collectionFileSystem
524 func (dn *dirnode) FS() FileSystem {
528 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
529 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
530 gn := &getternode{Getter: func() ([]byte, error) {
533 coll.ManifestText, err = dn.fs.MarshalManifest(".")
537 data, err := json.Marshal(&coll)
539 data = append(data, '\n')
543 gn.SetParent(dn, name)
546 return dn.treenode.Child(name, replace)
549 // sync flushes in-memory data and remote block references (for the
550 // children with the given names, which must be children of dn) to
551 // local persistent storage. Caller must have write lock on dn and the
553 func (dn *dirnode) sync(names []string) error {
554 type shortBlock struct {
558 var pending []shortBlock
561 flush := func(sbs []shortBlock) error {
565 block := make([]byte, 0, maxBlockSize)
566 for _, sb := range sbs {
567 block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
569 locator, _, err := dn.fs.PutB(block)
574 for _, sb := range sbs {
575 data := sb.fn.segments[sb.idx].(*memSegment).buf
576 sb.fn.segments[sb.idx] = storedSegment{
584 sb.fn.memsize -= int64(len(data))
589 localLocator := map[string]string{}
590 for _, name := range names {
591 fn, ok := dn.inodes[name].(*filenode)
595 for idx, seg := range fn.segments {
596 switch seg := seg.(type) {
598 loc, ok := localLocator[seg.locator]
601 loc, err = dn.fs.LocalLocator(seg.locator)
605 localLocator[seg.locator] = loc
608 fn.segments[idx] = seg
610 if seg.Len() > maxBlockSize/2 {
611 if err := flush([]shortBlock{{fn, idx}}); err != nil {
616 if pendingLen+seg.Len() > maxBlockSize {
617 if err := flush(pending); err != nil {
623 pending = append(pending, shortBlock{fn, idx})
624 pendingLen += seg.Len()
626 panic(fmt.Sprintf("can't sync segment type %T", seg))
630 return flush(pending)
633 // caller must have write lock.
634 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
636 type filepart struct {
641 var fileparts []filepart
645 names := make([]string, 0, len(dn.inodes))
646 for name := range dn.inodes {
647 names = append(names, name)
650 for _, name := range names {
651 node := dn.inodes[name]
655 if err := dn.sync(names); err != nil {
658 for _, name := range names {
659 switch node := dn.inodes[name].(type) {
661 subdir, err := node.marshalManifest(prefix + "/" + name)
665 subdirs = subdirs + subdir
667 if len(node.segments) == 0 {
668 fileparts = append(fileparts, filepart{name: name})
671 for _, seg := range node.segments {
672 switch seg := seg.(type) {
674 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
675 streamLen -= int64(seg.size)
677 blocks = append(blocks, seg.locator)
681 offset: streamLen + int64(seg.offset),
682 length: int64(seg.length),
684 if prev := len(fileparts) - 1; prev >= 0 &&
685 fileparts[prev].name == name &&
686 fileparts[prev].offset+fileparts[prev].length == next.offset {
687 fileparts[prev].length += next.length
689 fileparts = append(fileparts, next)
691 streamLen += int64(seg.size)
693 // This can't happen: we
694 // haven't unlocked since
696 panic(fmt.Sprintf("can't marshal segment type %T", seg))
700 panic(fmt.Sprintf("can't marshal inode type %T", node))
703 var filetokens []string
704 for _, s := range fileparts {
705 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
707 if len(filetokens) == 0 {
709 } else if len(blocks) == 0 {
710 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
712 return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
715 func (dn *dirnode) loadManifest(txt string) error {
717 streams := strings.Split(txt, "\n")
718 if streams[len(streams)-1] != "" {
719 return fmt.Errorf("line %d: no trailing newline", len(streams))
721 streams = streams[:len(streams)-1]
722 segments := []storedSegment{}
723 for i, stream := range streams {
725 var anyFileTokens bool
728 segments = segments[:0]
729 for i, token := range strings.Split(stream, " ") {
731 dirname = manifestUnescape(token)
734 if !strings.Contains(token, ":") {
736 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
738 toks := strings.SplitN(token, "+", 3)
740 return fmt.Errorf("line %d: bad locator %q", lineno, token)
742 length, err := strconv.ParseInt(toks[1], 10, 32)
743 if err != nil || length < 0 {
744 return fmt.Errorf("line %d: bad locator %q", lineno, token)
746 segments = append(segments, storedSegment{
753 } else if len(segments) == 0 {
754 return fmt.Errorf("line %d: bad locator %q", lineno, token)
757 toks := strings.SplitN(token, ":", 3)
759 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
763 offset, err := strconv.ParseInt(toks[0], 10, 64)
764 if err != nil || offset < 0 {
765 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
767 length, err := strconv.ParseInt(toks[1], 10, 64)
768 if err != nil || length < 0 {
769 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
771 name := dirname + "/" + manifestUnescape(toks[2])
772 fnode, err := dn.createFileAndParents(name)
774 return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
776 // Map the stream offset/range coordinates to
777 // block/offset/range coordinates and add
778 // corresponding storedSegments to the filenode
780 // Can't continue where we left off.
781 // TODO: binary search instead of
782 // rewinding all the way (but this
783 // situation might be rare anyway)
786 for next := int64(0); segIdx < len(segments); segIdx++ {
787 seg := segments[segIdx]
788 next = pos + int64(seg.Len())
789 if next <= offset || seg.Len() == 0 {
793 if pos >= offset+length {
798 blkOff = int(offset - pos)
800 blkLen := seg.Len() - blkOff
801 if pos+int64(blkOff+blkLen) > offset+length {
802 blkLen = int(offset + length - pos - int64(blkOff))
804 fnode.appendSegment(storedSegment{
806 locator: seg.locator,
811 if next > offset+length {
817 if segIdx == len(segments) && pos < offset+length {
818 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
822 return fmt.Errorf("line %d: no file segments", lineno)
823 } else if len(segments) == 0 {
824 return fmt.Errorf("line %d: no locators", lineno)
825 } else if dirname == "" {
826 return fmt.Errorf("line %d: no stream name", lineno)
832 // only safe to call from loadManifest -- no locking
833 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
835 names := strings.Split(path, "/")
836 basename := names[len(names)-1]
837 if !permittedName(basename) {
838 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
841 for _, name := range names[:len(names)-1] {
847 // can't be sure parent will be a *dirnode
848 return nil, ErrInvalidArgument
853 node, err = node.Child(name, func(child inode) (inode, error) {
855 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
859 child.SetParent(node, name)
861 } else if !child.IsDir() {
862 return child, ErrFileExists
871 _, err = node.Child(basename, func(child inode) (inode, error) {
872 switch child := child.(type) {
874 child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
878 child.SetParent(node, basename)
879 fn = child.(*filenode)
885 return child, ErrIsDirectory
887 return child, ErrInvalidArgument
893 func (dn *dirnode) TreeSize() (bytes int64) {
896 for _, i := range dn.inodes {
897 switch i := i.(type) {
901 bytes += i.TreeSize()
907 type segment interface {
910 // Return a new segment with a subsection of the data from this
911 // one. length<0 means length=Len()-off.
912 Slice(off int, length int) segment
915 type memSegment struct {
919 func (me *memSegment) Len() int {
923 func (me *memSegment) Slice(off, length int) segment {
925 length = len(me.buf) - off
927 buf := make([]byte, length)
928 copy(buf, me.buf[off:])
929 return &memSegment{buf: buf}
932 func (me *memSegment) Truncate(n int) {
936 newsize = newsize << 2
938 newbuf := make([]byte, n, newsize)
942 // Zero unused part when shrinking, in case we grow
943 // and start using it again later.
944 for i := n; i < len(me.buf); i++ {
951 func (me *memSegment) WriteAt(p []byte, off int) {
952 if off+len(p) > len(me.buf) {
953 panic("overflowed segment")
955 copy(me.buf[off:], p)
958 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
959 if off > int64(me.Len()) {
963 n = copy(p, me.buf[int(off):])
970 type storedSegment struct {
973 size int // size of stored block (also encoded in locator)
974 offset int // position of segment within the stored block
975 length int // bytes in this segment (offset + length <= size)
978 func (se storedSegment) Len() int {
982 func (se storedSegment) Slice(n, size int) segment {
985 if size >= 0 && se.length > size {
991 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
992 if off > int64(se.length) {
995 maxlen := se.length - int(off)
998 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1004 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1007 func canonicalName(name string) string {
1008 name = path.Clean("/" + name)
1009 if name == "/" || name == "./" {
1011 } else if strings.HasPrefix(name, "/") {
1017 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1019 func manifestUnescapeFunc(seq string) string {
1023 i, err := strconv.ParseUint(seq[1:], 8, 8)
1025 // Invalid escape sequence: can't unescape.
1028 return string([]byte{byte(i)})
1031 func manifestUnescape(s string) string {
1032 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1035 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1037 func manifestEscapeFunc(seq string) string {
1038 return fmt.Sprintf("\\%03o", byte(seq[0]))
1041 func manifestEscape(s string) string {
1042 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)