1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
22 var maxBlockSize = 1 << 26
24 type fsBackend interface {
29 // Ideally *Client would do everything; meanwhile keepBackend
30 // implements fsBackend by merging the two kinds of arvados client.
31 type keepBackend struct {
36 type keepClient interface {
37 ReadAt(locator string, p []byte, off int) (int, error)
38 PutB(p []byte) (string, int, error)
41 type apiClient interface {
42 RequestAndDecode(dst interface{}, method, path string, body io.Reader, params interface{}) error
43 UpdateBody(rsc resource) io.Reader
46 // A CollectionFileSystem is a FileSystem that can be serialized as a
47 // manifest and stored as a collection.
48 type CollectionFileSystem interface {
51 // Flush all file data to Keep and return a snapshot of the
52 // filesystem suitable for saving as (Collection)ManifestText.
53 // Prefix (normally ".") is a top level directory, effectively
54 // prepended to all paths in the returned manifest.
55 MarshalManifest(prefix string) (string, error)
58 // FileSystem returns a CollectionFileSystem for the collection.
59 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
61 if c.ModifiedAt == nil {
64 modTime = *c.ModifiedAt
66 fs := &collectionFileSystem{
68 fileSystem: fileSystem{
69 fsBackend: keepBackend{apiClient: client, keepClient: kc},
77 mode: os.ModeDir | 0755,
80 inodes: make(map[string]inode),
84 if err := root.loadManifest(c.ManifestText); err != nil {
91 type collectionFileSystem struct {
96 // Caller must have parent lock, and must have already ensured
97 // parent.Child(name,nil) is nil.
98 func (fs *collectionFileSystem) newDirnode(parent inode, name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
99 if name == "" || name == "." || name == ".." {
100 return nil, ErrInvalidArgument
107 mode: perm | os.ModeDir,
110 inodes: make(map[string]inode),
115 func (fs *collectionFileSystem) newFilenode(parent inode, name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
116 if name == "" || name == "." || name == ".." {
117 return nil, ErrInvalidArgument
123 mode: perm & ^os.ModeDir,
129 func (fs *collectionFileSystem) Sync() error {
130 log.Printf("cfs.Sync()")
134 txt, err := fs.MarshalManifest(".")
136 log.Printf("WARNING: (collectionFileSystem)Sync() failed: %s", err)
143 err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
145 log.Printf("WARNING: (collectionFileSystem)Sync() failed: %s", err)
150 func (dn *dirnode) Child(name string, replace func(inode) inode) inode {
151 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
152 gn := &getternode{Getter: func() ([]byte, error) {
155 coll.ManifestText, err = dn.fs.MarshalManifest(".")
159 data, err := json.Marshal(&coll)
161 data = append(data, 10)
168 return dn.treenode.Child(name, replace)
171 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
172 fs.fileSystem.root.Lock()
173 defer fs.fileSystem.root.Unlock()
174 return fs.fileSystem.root.(*dirnode).marshalManifest(prefix)
177 // filenodePtr is an offset into a file that is (usually) efficient to
178 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
180 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
181 // corresponds to file offset filenodePtr.off. Otherwise, it is
182 // necessary to reexamine len(filenode.segments[0]) etc. to find the
183 // correct segment and offset.
184 type filenodePtr struct {
191 // seek returns a ptr that is consistent with both startPtr.off and
192 // the current state of fn. The caller must already hold fn.RLock() or
195 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
200 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
202 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
203 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
206 // meaningless anyway
208 } else if ptr.off >= fn.fileinfo.size {
209 ptr.segmentIdx = len(fn.segments)
211 ptr.repacked = fn.repacked
213 } else if ptr.repacked == fn.repacked {
214 // segmentIdx and segmentOff accurately reflect
215 // ptr.off, but might have fallen off the end of a
217 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
224 ptr.repacked = fn.repacked
226 if ptr.off >= fn.fileinfo.size {
227 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
230 // Recompute segmentIdx and segmentOff. We have already
231 // established fn.fileinfo.size > ptr.off >= 0, so we don't
232 // have to deal with edge cases here.
234 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
235 // This would panic (index out of range) if
236 // fn.fileinfo.size were larger than
237 // sum(fn.segments[i].Len()) -- but that can't happen
238 // because we have ensured fn.fileinfo.size is always
240 segLen := int64(fn.segments[ptr.segmentIdx].Len())
241 if off+segLen > ptr.off {
242 ptr.segmentOff = int(ptr.off - off)
250 // filenode implements inode.
251 type filenode struct {
256 // number of times `segments` has changed in a
257 // way that might invalidate a filenodePtr
259 memsize int64 // bytes in memSegments
264 // caller must have lock
265 func (fn *filenode) appendSegment(e segment) {
266 fn.segments = append(fn.segments, e)
267 fn.fileinfo.size += int64(e.Len())
270 func (fn *filenode) SetParent(p inode) {
276 func (fn *filenode) Parent() inode {
282 func (fn *filenode) FS() FileSystem {
286 // Read reads file data from a single segment, starting at startPtr,
287 // into p. startPtr is assumed not to be up-to-date. Caller must have
289 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
290 ptr = fn.seek(startPtr)
292 err = ErrNegativeOffset
295 if ptr.segmentIdx >= len(fn.segments) {
299 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
303 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
306 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
314 func (fn *filenode) Size() int64 {
317 return fn.fileinfo.Size()
320 func (fn *filenode) FileInfo() os.FileInfo {
326 func (fn *filenode) Truncate(size int64) error {
329 return fn.truncate(size)
332 func (fn *filenode) truncate(size int64) error {
333 if size == fn.fileinfo.size {
337 if size < fn.fileinfo.size {
338 ptr := fn.seek(filenodePtr{off: size})
339 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
340 if seg, ok := fn.segments[i].(*memSegment); ok {
341 fn.memsize -= int64(seg.Len())
344 if ptr.segmentOff == 0 {
345 fn.segments = fn.segments[:ptr.segmentIdx]
347 fn.segments = fn.segments[:ptr.segmentIdx+1]
348 switch seg := fn.segments[ptr.segmentIdx].(type) {
350 seg.Truncate(ptr.segmentOff)
351 fn.memsize += int64(seg.Len())
353 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
356 fn.fileinfo.size = size
359 for size > fn.fileinfo.size {
360 grow := size - fn.fileinfo.size
363 if len(fn.segments) == 0 {
365 fn.segments = append(fn.segments, seg)
366 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
368 fn.segments = append(fn.segments, seg)
370 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
373 seg.Truncate(seg.Len() + int(grow))
374 fn.fileinfo.size += grow
380 // Write writes data from p to the file, starting at startPtr,
381 // extending the file size if necessary. Caller must have Lock.
382 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
383 if startPtr.off > fn.fileinfo.size {
384 if err = fn.truncate(startPtr.off); err != nil {
385 return 0, startPtr, err
388 ptr = fn.seek(startPtr)
390 err = ErrNegativeOffset
393 for len(p) > 0 && err == nil {
395 if len(cando) > maxBlockSize {
396 cando = cando[:maxBlockSize]
398 // Rearrange/grow fn.segments (and shrink cando if
399 // needed) such that cando can be copied to
400 // fn.segments[ptr.segmentIdx] at offset
402 cur := ptr.segmentIdx
403 prev := ptr.segmentIdx - 1
405 if cur < len(fn.segments) {
406 _, curWritable = fn.segments[cur].(*memSegment)
408 var prevAppendable bool
409 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
410 _, prevAppendable = fn.segments[prev].(*memSegment)
412 if ptr.segmentOff > 0 && !curWritable {
413 // Split a non-writable block.
414 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
415 // Truncate cur, and insert a new
418 fn.segments = append(fn.segments, nil)
419 copy(fn.segments[cur+1:], fn.segments[cur:])
421 // Split cur into two copies, truncate
422 // the one on the left, shift the one
423 // on the right, and insert a new
424 // segment between them.
425 fn.segments = append(fn.segments, nil, nil)
426 copy(fn.segments[cur+2:], fn.segments[cur:])
427 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
432 seg.Truncate(len(cando))
433 fn.memsize += int64(len(cando))
434 fn.segments[cur] = seg
435 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
440 } else if curWritable {
441 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
446 // Shrink cando if needed to fit in
448 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
449 cando = cando[:cangrow]
453 if cur == len(fn.segments) {
454 // ptr is at EOF, filesize is changing.
455 fn.fileinfo.size += int64(len(cando))
456 } else if el := fn.segments[cur].Len(); el <= len(cando) {
457 // cando is long enough that we won't
458 // need cur any more. shrink cando to
459 // be exactly as long as cur
460 // (otherwise we'd accidentally shift
461 // the effective position of all
462 // segments after cur).
464 copy(fn.segments[cur:], fn.segments[cur+1:])
465 fn.segments = fn.segments[:len(fn.segments)-1]
467 // shrink cur by the same #bytes we're growing prev
468 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
474 ptr.segmentOff = fn.segments[prev].Len()
475 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
476 fn.memsize += int64(len(cando))
480 // Insert a segment between prev and
481 // cur, and advance prev/cur.
482 fn.segments = append(fn.segments, nil)
483 if cur < len(fn.segments) {
484 copy(fn.segments[cur+1:], fn.segments[cur:])
488 // appending a new segment does
489 // not invalidate any ptrs
492 seg.Truncate(len(cando))
493 fn.memsize += int64(len(cando))
494 fn.segments[cur] = seg
500 // Finally we can copy bytes from cando to the current segment.
501 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
505 ptr.off += int64(len(cando))
506 ptr.segmentOff += len(cando)
507 if ptr.segmentOff >= maxBlockSize {
508 fn.pruneMemSegments()
510 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
515 fn.fileinfo.modTime = time.Now()
520 // Write some data out to disk to reduce memory use. Caller must have
522 func (fn *filenode) pruneMemSegments() {
523 // TODO: async (don't hold Lock() while waiting for Keep)
524 // TODO: share code with (*dirnode)sync()
525 // TODO: pack/flush small blocks too, when fragmented
526 for idx, seg := range fn.segments {
527 seg, ok := seg.(*memSegment)
528 if !ok || seg.Len() < maxBlockSize {
531 locator, _, err := fn.parent.(fsBackend).PutB(seg.buf)
533 // TODO: stall (or return errors from)
534 // subsequent writes until flushing
538 fn.memsize -= int64(seg.Len())
539 fn.segments[idx] = storedSegment{
549 type dirnode struct {
550 fs *collectionFileSystem
554 func (dn *dirnode) FS() FileSystem {
558 // sync flushes in-memory data (for all files in the tree rooted at
559 // dn) to persistent storage. Caller must hold dn.Lock().
560 func (dn *dirnode) sync() error {
561 type shortBlock struct {
565 var pending []shortBlock
568 flush := func(sbs []shortBlock) error {
572 block := make([]byte, 0, maxBlockSize)
573 for _, sb := range sbs {
574 block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
576 locator, _, err := dn.fs.PutB(block)
581 for _, sb := range sbs {
582 data := sb.fn.segments[sb.idx].(*memSegment).buf
583 sb.fn.segments[sb.idx] = storedSegment{
591 sb.fn.memsize -= int64(len(data))
596 names := make([]string, 0, len(dn.inodes))
597 for name := range dn.inodes {
598 names = append(names, name)
602 for _, name := range names {
603 fn, ok := dn.inodes[name].(*filenode)
609 for idx, seg := range fn.segments {
610 seg, ok := seg.(*memSegment)
614 if seg.Len() > maxBlockSize/2 {
615 if err := flush([]shortBlock{{fn, idx}}); err != nil {
620 if pendingLen+seg.Len() > maxBlockSize {
621 if err := flush(pending); err != nil {
627 pending = append(pending, shortBlock{fn, idx})
628 pendingLen += seg.Len()
631 return flush(pending)
634 // caller must have read lock.
635 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
637 type filepart struct {
642 var fileparts []filepart
646 if err := dn.sync(); err != nil {
650 names := make([]string, 0, len(dn.inodes))
651 for name, node := range dn.inodes {
652 names = append(names, name)
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 basename == "" || basename == "." || basename == ".." {
838 err = fmt.Errorf("invalid filename")
841 for _, name := range names[:len(names)-1] {
847 // can't be sure parent will be a *dirnode
848 return nil, ErrInvalidArgument
853 node.Child(name, func(child inode) inode {
855 node, err = node.FS().newDirnode(node, name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
857 } else if !child.IsDir() {
868 node.Child(basename, func(child inode) inode {
869 switch child := child.(type) {
871 child, err = node.FS().newFilenode(node, basename, 0755, node.FileInfo().ModTime())
872 fn = child.(*filenode)
881 err = ErrInvalidArgument
888 type segment interface {
891 // Return a new segment with a subsection of the data from this
892 // one. length<0 means length=Len()-off.
893 Slice(off int, length int) segment
896 type memSegment struct {
900 func (me *memSegment) Len() int {
904 func (me *memSegment) Slice(off, length int) segment {
906 length = len(me.buf) - off
908 buf := make([]byte, length)
909 copy(buf, me.buf[off:])
910 return &memSegment{buf: buf}
913 func (me *memSegment) Truncate(n int) {
917 newsize = newsize << 2
919 newbuf := make([]byte, n, newsize)
923 // Zero unused part when shrinking, in case we grow
924 // and start using it again later.
925 for i := n; i < len(me.buf); i++ {
932 func (me *memSegment) WriteAt(p []byte, off int) {
933 if off+len(p) > len(me.buf) {
934 panic("overflowed segment")
936 copy(me.buf[off:], p)
939 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
940 if off > int64(me.Len()) {
944 n = copy(p, me.buf[int(off):])
951 type storedSegment struct {
954 size int // size of stored block (also encoded in locator)
955 offset int // position of segment within the stored block
956 length int // bytes in this segment (offset + length <= size)
959 func (se storedSegment) Len() int {
963 func (se storedSegment) Slice(n, size int) segment {
966 if size >= 0 && se.length > size {
972 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
973 if off > int64(se.length) {
976 maxlen := se.length - int(off)
979 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
985 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
988 func canonicalName(name string) string {
989 name = path.Clean("/" + name)
990 if name == "/" || name == "./" {
992 } else if strings.HasPrefix(name, "/") {
998 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1000 func manifestUnescapeFunc(seq string) string {
1004 i, err := strconv.ParseUint(seq[1:], 8, 8)
1006 // Invalid escape sequence: can't unescape.
1009 return string([]byte{byte(i)})
1012 func manifestUnescape(s string) string {
1013 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1016 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1018 func manifestEscapeFunc(seq string) string {
1019 return fmt.Sprintf("\\%03o", byte(seq[0]))
1022 func manifestEscape(s string) string {
1023 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)