1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
22 var maxBlockSize = 1 << 26
24 // A CollectionFileSystem is a FileSystem that can be serialized as a
25 // manifest and stored as a collection.
26 type CollectionFileSystem interface {
29 // Flush all file data to Keep and return a snapshot of the
30 // filesystem suitable for saving as (Collection)ManifestText.
31 // Prefix (normally ".") is a top level directory, effectively
32 // prepended to all paths in the returned manifest.
33 MarshalManifest(prefix string) (string, error)
35 // Total data bytes in all files.
39 type collectionFileSystem struct {
44 // FileSystem returns a CollectionFileSystem for the collection.
45 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
47 if c.ModifiedAt == nil {
50 modTime = *c.ModifiedAt
52 fs := &collectionFileSystem{
54 fileSystem: fileSystem{
55 fsBackend: keepBackend{apiClient: client, keepClient: kc},
63 mode: os.ModeDir | 0755,
66 inodes: make(map[string]inode),
69 root.SetParent(root, ".")
70 if err := root.loadManifest(c.ManifestText); err != nil {
73 backdateTree(root, modTime)
78 func backdateTree(n inode, modTime time.Time) {
79 switch n := n.(type) {
81 n.fileinfo.modTime = modTime
83 n.fileinfo.modTime = modTime
84 for _, n := range n.inodes {
85 backdateTree(n, modTime)
90 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
91 if name == "" || name == "." || name == ".." {
92 return nil, ErrInvalidArgument
100 mode: perm | os.ModeDir,
103 inodes: make(map[string]inode),
111 mode: perm & ^os.ModeDir,
118 func (fs *collectionFileSystem) Sync() error {
119 log.Printf("cfs.Sync()")
123 txt, err := fs.MarshalManifest(".")
125 log.Printf("WARNING: (collectionFileSystem)Sync() failed: %s", err)
132 err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
134 log.Printf("WARNING: (collectionFileSystem)Sync() failed: %s", err)
139 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
140 fs.fileSystem.root.Lock()
141 defer fs.fileSystem.root.Unlock()
142 return fs.fileSystem.root.(*dirnode).marshalManifest(prefix)
145 func (fs *collectionFileSystem) Size() int64 {
146 return fs.fileSystem.root.(*dirnode).TreeSize()
149 // filenodePtr is an offset into a file that is (usually) efficient to
150 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
152 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
153 // corresponds to file offset filenodePtr.off. Otherwise, it is
154 // necessary to reexamine len(filenode.segments[0]) etc. to find the
155 // correct segment and offset.
156 type filenodePtr struct {
163 // seek returns a ptr that is consistent with both startPtr.off and
164 // the current state of fn. The caller must already hold fn.RLock() or
167 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
172 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
174 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
175 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
178 // meaningless anyway
180 } else if ptr.off >= fn.fileinfo.size {
181 ptr.segmentIdx = len(fn.segments)
183 ptr.repacked = fn.repacked
185 } else if ptr.repacked == fn.repacked {
186 // segmentIdx and segmentOff accurately reflect
187 // ptr.off, but might have fallen off the end of a
189 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
196 ptr.repacked = fn.repacked
198 if ptr.off >= fn.fileinfo.size {
199 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
202 // Recompute segmentIdx and segmentOff. We have already
203 // established fn.fileinfo.size > ptr.off >= 0, so we don't
204 // have to deal with edge cases here.
206 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
207 // This would panic (index out of range) if
208 // fn.fileinfo.size were larger than
209 // sum(fn.segments[i].Len()) -- but that can't happen
210 // because we have ensured fn.fileinfo.size is always
212 segLen := int64(fn.segments[ptr.segmentIdx].Len())
213 if off+segLen > ptr.off {
214 ptr.segmentOff = int(ptr.off - off)
222 // filenode implements inode.
223 type filenode struct {
228 // number of times `segments` has changed in a
229 // way that might invalidate a filenodePtr
231 memsize int64 // bytes in memSegments
236 // caller must have lock
237 func (fn *filenode) appendSegment(e segment) {
238 fn.segments = append(fn.segments, e)
239 fn.fileinfo.size += int64(e.Len())
242 func (fn *filenode) SetParent(p inode, name string) {
246 fn.fileinfo.name = name
249 func (fn *filenode) Parent() inode {
255 func (fn *filenode) FS() FileSystem {
259 // Read reads file data from a single segment, starting at startPtr,
260 // into p. startPtr is assumed not to be up-to-date. Caller must have
262 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
263 ptr = fn.seek(startPtr)
265 err = ErrNegativeOffset
268 if ptr.segmentIdx >= len(fn.segments) {
272 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
276 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
279 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
287 func (fn *filenode) Size() int64 {
290 return fn.fileinfo.Size()
293 func (fn *filenode) FileInfo() os.FileInfo {
299 func (fn *filenode) Truncate(size int64) error {
302 return fn.truncate(size)
305 func (fn *filenode) truncate(size int64) error {
306 if size == fn.fileinfo.size {
310 if size < fn.fileinfo.size {
311 ptr := fn.seek(filenodePtr{off: size})
312 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
313 if seg, ok := fn.segments[i].(*memSegment); ok {
314 fn.memsize -= int64(seg.Len())
317 if ptr.segmentOff == 0 {
318 fn.segments = fn.segments[:ptr.segmentIdx]
320 fn.segments = fn.segments[:ptr.segmentIdx+1]
321 switch seg := fn.segments[ptr.segmentIdx].(type) {
323 seg.Truncate(ptr.segmentOff)
324 fn.memsize += int64(seg.Len())
326 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
329 fn.fileinfo.size = size
332 for size > fn.fileinfo.size {
333 grow := size - fn.fileinfo.size
336 if len(fn.segments) == 0 {
338 fn.segments = append(fn.segments, seg)
339 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
341 fn.segments = append(fn.segments, seg)
343 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
346 seg.Truncate(seg.Len() + int(grow))
347 fn.fileinfo.size += grow
353 // Write writes data from p to the file, starting at startPtr,
354 // extending the file size if necessary. Caller must have Lock.
355 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
356 if startPtr.off > fn.fileinfo.size {
357 if err = fn.truncate(startPtr.off); err != nil {
358 return 0, startPtr, err
361 ptr = fn.seek(startPtr)
363 err = ErrNegativeOffset
366 for len(p) > 0 && err == nil {
368 if len(cando) > maxBlockSize {
369 cando = cando[:maxBlockSize]
371 // Rearrange/grow fn.segments (and shrink cando if
372 // needed) such that cando can be copied to
373 // fn.segments[ptr.segmentIdx] at offset
375 cur := ptr.segmentIdx
376 prev := ptr.segmentIdx - 1
378 if cur < len(fn.segments) {
379 _, curWritable = fn.segments[cur].(*memSegment)
381 var prevAppendable bool
382 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
383 _, prevAppendable = fn.segments[prev].(*memSegment)
385 if ptr.segmentOff > 0 && !curWritable {
386 // Split a non-writable block.
387 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
388 // Truncate cur, and insert a new
391 fn.segments = append(fn.segments, nil)
392 copy(fn.segments[cur+1:], fn.segments[cur:])
394 // Split cur into two copies, truncate
395 // the one on the left, shift the one
396 // on the right, and insert a new
397 // segment between them.
398 fn.segments = append(fn.segments, nil, nil)
399 copy(fn.segments[cur+2:], fn.segments[cur:])
400 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
405 seg.Truncate(len(cando))
406 fn.memsize += int64(len(cando))
407 fn.segments[cur] = seg
408 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
413 } else if curWritable {
414 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
419 // Shrink cando if needed to fit in
421 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
422 cando = cando[:cangrow]
426 if cur == len(fn.segments) {
427 // ptr is at EOF, filesize is changing.
428 fn.fileinfo.size += int64(len(cando))
429 } else if el := fn.segments[cur].Len(); el <= len(cando) {
430 // cando is long enough that we won't
431 // need cur any more. shrink cando to
432 // be exactly as long as cur
433 // (otherwise we'd accidentally shift
434 // the effective position of all
435 // segments after cur).
437 copy(fn.segments[cur:], fn.segments[cur+1:])
438 fn.segments = fn.segments[:len(fn.segments)-1]
440 // shrink cur by the same #bytes we're growing prev
441 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
447 ptr.segmentOff = fn.segments[prev].Len()
448 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
449 fn.memsize += int64(len(cando))
453 // Insert a segment between prev and
454 // cur, and advance prev/cur.
455 fn.segments = append(fn.segments, nil)
456 if cur < len(fn.segments) {
457 copy(fn.segments[cur+1:], fn.segments[cur:])
461 // appending a new segment does
462 // not invalidate any ptrs
465 seg.Truncate(len(cando))
466 fn.memsize += int64(len(cando))
467 fn.segments[cur] = seg
473 // Finally we can copy bytes from cando to the current segment.
474 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
478 ptr.off += int64(len(cando))
479 ptr.segmentOff += len(cando)
480 if ptr.segmentOff >= maxBlockSize {
481 fn.pruneMemSegments()
483 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
488 fn.fileinfo.modTime = time.Now()
493 // Write some data out to disk to reduce memory use. Caller must have
495 func (fn *filenode) pruneMemSegments() {
496 // TODO: async (don't hold Lock() while waiting for Keep)
497 // TODO: share code with (*dirnode)sync()
498 // TODO: pack/flush small blocks too, when fragmented
499 for idx, seg := range fn.segments {
500 seg, ok := seg.(*memSegment)
501 if !ok || seg.Len() < maxBlockSize {
504 locator, _, err := fn.FS().PutB(seg.buf)
506 // TODO: stall (or return errors from)
507 // subsequent writes until flushing
511 fn.memsize -= int64(seg.Len())
512 fn.segments[idx] = storedSegment{
522 type dirnode struct {
523 fs *collectionFileSystem
527 func (dn *dirnode) FS() FileSystem {
531 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
532 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
533 gn := &getternode{Getter: func() ([]byte, error) {
536 coll.ManifestText, err = dn.fs.MarshalManifest(".")
540 data, err := json.Marshal(&coll)
542 data = append(data, '\n')
546 gn.SetParent(dn, name)
549 return dn.treenode.Child(name, replace)
552 // sync flushes in-memory data (for the children with the given names,
553 // which must be children of dn) to persistent storage. Caller must
554 // have write lock on dn and the named children.
555 func (dn *dirnode) sync(names []string) error {
556 type shortBlock struct {
560 var pending []shortBlock
563 flush := func(sbs []shortBlock) error {
567 block := make([]byte, 0, maxBlockSize)
568 for _, sb := range sbs {
569 block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
571 locator, _, err := dn.fs.PutB(block)
576 for _, sb := range sbs {
577 data := sb.fn.segments[sb.idx].(*memSegment).buf
578 sb.fn.segments[sb.idx] = storedSegment{
586 sb.fn.memsize -= int64(len(data))
591 for _, name := range names {
592 fn, ok := dn.inodes[name].(*filenode)
596 for idx, seg := range fn.segments {
597 seg, ok := seg.(*memSegment)
601 if seg.Len() > maxBlockSize/2 {
602 if err := flush([]shortBlock{{fn, idx}}); err != nil {
607 if pendingLen+seg.Len() > maxBlockSize {
608 if err := flush(pending); err != nil {
614 pending = append(pending, shortBlock{fn, idx})
615 pendingLen += seg.Len()
618 return flush(pending)
621 // caller must have read lock.
622 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
624 type filepart struct {
629 var fileparts []filepart
633 names := make([]string, 0, len(dn.inodes))
634 for name := range dn.inodes {
635 names = append(names, name)
638 for _, name := range names {
639 node := dn.inodes[name]
643 if err := dn.sync(names); err != nil {
646 for _, name := range names {
647 switch node := dn.inodes[name].(type) {
649 subdir, err := node.marshalManifest(prefix + "/" + name)
653 subdirs = subdirs + subdir
655 if len(node.segments) == 0 {
656 fileparts = append(fileparts, filepart{name: name})
659 for _, seg := range node.segments {
660 switch seg := seg.(type) {
662 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
663 streamLen -= int64(seg.size)
665 blocks = append(blocks, seg.locator)
669 offset: streamLen + int64(seg.offset),
670 length: int64(seg.length),
672 if prev := len(fileparts) - 1; prev >= 0 &&
673 fileparts[prev].name == name &&
674 fileparts[prev].offset+fileparts[prev].length == next.offset {
675 fileparts[prev].length += next.length
677 fileparts = append(fileparts, next)
679 streamLen += int64(seg.size)
681 // This can't happen: we
682 // haven't unlocked since
684 panic(fmt.Sprintf("can't marshal segment type %T", seg))
688 panic(fmt.Sprintf("can't marshal inode type %T", node))
691 var filetokens []string
692 for _, s := range fileparts {
693 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
695 if len(filetokens) == 0 {
697 } else if len(blocks) == 0 {
698 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
700 return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
703 func (dn *dirnode) loadManifest(txt string) error {
705 streams := strings.Split(txt, "\n")
706 if streams[len(streams)-1] != "" {
707 return fmt.Errorf("line %d: no trailing newline", len(streams))
709 streams = streams[:len(streams)-1]
710 segments := []storedSegment{}
711 for i, stream := range streams {
713 var anyFileTokens bool
716 segments = segments[:0]
717 for i, token := range strings.Split(stream, " ") {
719 dirname = manifestUnescape(token)
722 if !strings.Contains(token, ":") {
724 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
726 toks := strings.SplitN(token, "+", 3)
728 return fmt.Errorf("line %d: bad locator %q", lineno, token)
730 length, err := strconv.ParseInt(toks[1], 10, 32)
731 if err != nil || length < 0 {
732 return fmt.Errorf("line %d: bad locator %q", lineno, token)
734 segments = append(segments, storedSegment{
741 } else if len(segments) == 0 {
742 return fmt.Errorf("line %d: bad locator %q", lineno, token)
745 toks := strings.SplitN(token, ":", 3)
747 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
751 offset, err := strconv.ParseInt(toks[0], 10, 64)
752 if err != nil || offset < 0 {
753 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
755 length, err := strconv.ParseInt(toks[1], 10, 64)
756 if err != nil || length < 0 {
757 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
759 name := dirname + "/" + manifestUnescape(toks[2])
760 fnode, err := dn.createFileAndParents(name)
762 return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
764 // Map the stream offset/range coordinates to
765 // block/offset/range coordinates and add
766 // corresponding storedSegments to the filenode
768 // Can't continue where we left off.
769 // TODO: binary search instead of
770 // rewinding all the way (but this
771 // situation might be rare anyway)
774 for next := int64(0); segIdx < len(segments); segIdx++ {
775 seg := segments[segIdx]
776 next = pos + int64(seg.Len())
777 if next <= offset || seg.Len() == 0 {
781 if pos >= offset+length {
786 blkOff = int(offset - pos)
788 blkLen := seg.Len() - blkOff
789 if pos+int64(blkOff+blkLen) > offset+length {
790 blkLen = int(offset + length - pos - int64(blkOff))
792 fnode.appendSegment(storedSegment{
794 locator: seg.locator,
799 if next > offset+length {
805 if segIdx == len(segments) && pos < offset+length {
806 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
810 return fmt.Errorf("line %d: no file segments", lineno)
811 } else if len(segments) == 0 {
812 return fmt.Errorf("line %d: no locators", lineno)
813 } else if dirname == "" {
814 return fmt.Errorf("line %d: no stream name", lineno)
820 // only safe to call from loadManifest -- no locking
821 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
823 names := strings.Split(path, "/")
824 basename := names[len(names)-1]
825 if !permittedName(basename) {
826 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
829 for _, name := range names[:len(names)-1] {
835 // can't be sure parent will be a *dirnode
836 return nil, ErrInvalidArgument
841 node, err = node.Child(name, func(child inode) (inode, error) {
843 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
847 child.SetParent(node, name)
849 } else if !child.IsDir() {
850 return child, ErrFileExists
859 _, err = node.Child(basename, func(child inode) (inode, error) {
860 switch child := child.(type) {
862 child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
866 child.SetParent(node, basename)
867 fn = child.(*filenode)
873 return child, ErrIsDirectory
875 return child, ErrInvalidArgument
881 func (dn *dirnode) TreeSize() (bytes int64) {
884 for _, i := range dn.inodes {
885 switch i := i.(type) {
889 bytes += i.TreeSize()
895 type segment interface {
898 // Return a new segment with a subsection of the data from this
899 // one. length<0 means length=Len()-off.
900 Slice(off int, length int) segment
903 type memSegment struct {
907 func (me *memSegment) Len() int {
911 func (me *memSegment) Slice(off, length int) segment {
913 length = len(me.buf) - off
915 buf := make([]byte, length)
916 copy(buf, me.buf[off:])
917 return &memSegment{buf: buf}
920 func (me *memSegment) Truncate(n int) {
924 newsize = newsize << 2
926 newbuf := make([]byte, n, newsize)
930 // Zero unused part when shrinking, in case we grow
931 // and start using it again later.
932 for i := n; i < len(me.buf); i++ {
939 func (me *memSegment) WriteAt(p []byte, off int) {
940 if off+len(p) > len(me.buf) {
941 panic("overflowed segment")
943 copy(me.buf[off:], p)
946 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
947 if off > int64(me.Len()) {
951 n = copy(p, me.buf[int(off):])
958 type storedSegment struct {
961 size int // size of stored block (also encoded in locator)
962 offset int // position of segment within the stored block
963 length int // bytes in this segment (offset + length <= size)
966 func (se storedSegment) Len() int {
970 func (se storedSegment) Slice(n, size int) segment {
973 if size >= 0 && se.length > size {
979 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
980 if off > int64(se.length) {
983 maxlen := se.length - int(off)
986 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
992 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
995 func canonicalName(name string) string {
996 name = path.Clean("/" + name)
997 if name == "/" || name == "./" {
999 } else if strings.HasPrefix(name, "/") {
1005 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1007 func manifestUnescapeFunc(seq string) string {
1011 i, err := strconv.ParseUint(seq[1:], 8, 8)
1013 // Invalid escape sequence: can't unescape.
1016 return string([]byte{byte(i)})
1019 func manifestUnescape(s string) string {
1020 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1023 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1025 func manifestEscapeFunc(seq string) string {
1026 return fmt.Sprintf("\\%03o", byte(seq[0]))
1029 func manifestEscape(s string) string {
1030 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)