1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: Apache-2.0
25 maxBlockSize = 1 << 26
26 concurrentWriters = 4 // max goroutines writing to Keep in background and during flush()
29 // A CollectionFileSystem is a FileSystem that can be serialized as a
30 // manifest and stored as a collection.
31 type CollectionFileSystem interface {
34 // Flush all file data to Keep and return a snapshot of the
35 // filesystem suitable for saving as (Collection)ManifestText.
36 // Prefix (normally ".") is a top level directory, effectively
37 // prepended to all paths in the returned manifest.
38 MarshalManifest(prefix string) (string, error)
40 // Total data bytes in all files.
44 type collectionFileSystem struct {
49 storageClasses []string
50 // guessSignatureTTL tracks a lower bound for the server's
51 // configured BlobSigningTTL. The guess is initially zero, and
52 // increases when we come across a signature with an expiry
53 // time further in the future than the previous guess.
55 // When the guessed TTL is much smaller than the real TTL,
56 // preemptive signature refresh is delayed or missed entirely,
58 guessSignatureTTL time.Duration
59 holdCheckChanges time.Time
60 lockCheckChanges sync.Mutex
63 // FileSystem returns a CollectionFileSystem for the collection.
64 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
65 modTime := c.ModifiedAt
69 fs := &collectionFileSystem{
71 storageClasses: c.StorageClassesDesired,
72 fileSystem: fileSystem{
73 fsBackend: keepBackend{apiClient: client, keepClient: kc},
74 thr: newThrottle(concurrentWriters),
77 fs.savedPDH.Store(c.PortableDataHash)
78 if r := c.ReplicationDesired; r != nil {
86 mode: os.ModeDir | 0755,
88 sys: func() interface{} { return c },
90 inodes: make(map[string]inode),
93 root.SetParent(root, ".")
94 if err := root.loadManifest(c.ManifestText); err != nil {
97 backdateTree(root, modTime)
102 // caller must have lock (or guarantee no concurrent accesses somehow)
103 func eachNode(n inode, ffunc func(*filenode), dfunc func(*dirnode)) {
104 switch n := n.(type) {
113 for _, n := range n.inodes {
114 eachNode(n, ffunc, dfunc)
119 // caller must have lock (or guarantee no concurrent accesses somehow)
120 func backdateTree(n inode, modTime time.Time) {
121 eachNode(n, func(fn *filenode) {
122 fn.fileinfo.modTime = modTime
123 }, func(dn *dirnode) {
124 dn.fileinfo.modTime = modTime
128 // Approximate portion of signature TTL remaining, usually between 0
129 // and 1, or negative if some signatures have expired.
130 func (fs *collectionFileSystem) signatureTimeLeft() (float64, time.Duration) {
133 earliest = now.Add(time.Hour * 24 * 7 * 365)
136 fs.fileSystem.root.RLock()
137 eachNode(fs.root, func(fn *filenode) {
140 for _, seg := range fn.segments {
141 seg, ok := seg.(storedSegment)
145 expiryTime, err := signatureExpiryTime(seg.locator)
149 if expiryTime.Before(earliest) {
150 earliest = expiryTime
152 if expiryTime.After(latest) {
157 fs.fileSystem.root.RUnlock()
160 // No signatures == 100% of TTL remaining.
164 ttl := latest.Sub(now)
165 fs.fileSystem.root.Lock()
167 if ttl > fs.guessSignatureTTL {
168 // ttl is closer to the real TTL than
169 // guessSignatureTTL.
170 fs.guessSignatureTTL = ttl
172 // Use the previous best guess to compute the
173 // portion remaining (below, after unlocking
175 ttl = fs.guessSignatureTTL
178 fs.fileSystem.root.Unlock()
180 return earliest.Sub(now).Seconds() / ttl.Seconds(), ttl
183 func (fs *collectionFileSystem) updateSignatures(newmanifest string) {
184 newLoc := map[string]string{}
185 for _, tok := range regexp.MustCompile(`\S+`).FindAllString(newmanifest, -1) {
186 if mBlkRe.MatchString(tok) {
187 newLoc[stripAllHints(tok)] = tok
190 fs.fileSystem.root.Lock()
191 defer fs.fileSystem.root.Unlock()
192 eachNode(fs.root, func(fn *filenode) {
195 for idx, seg := range fn.segments {
196 seg, ok := seg.(storedSegment)
200 loc, ok := newLoc[stripAllHints(seg.locator)]
205 fn.segments[idx] = seg
210 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
211 if name == "" || name == "." || name == ".." {
212 return nil, ErrInvalidArgument
220 mode: perm | os.ModeDir,
223 inodes: make(map[string]inode),
231 mode: perm & ^os.ModeDir,
237 func (fs *collectionFileSystem) Child(name string, replace func(inode) (inode, error)) (inode, error) {
238 return fs.rootnode().Child(name, replace)
241 func (fs *collectionFileSystem) FS() FileSystem {
245 func (fs *collectionFileSystem) FileInfo() os.FileInfo {
246 return fs.rootnode().FileInfo()
249 func (fs *collectionFileSystem) IsDir() bool {
253 func (fs *collectionFileSystem) Lock() {
257 func (fs *collectionFileSystem) Unlock() {
258 fs.rootnode().Unlock()
261 func (fs *collectionFileSystem) RLock() {
262 fs.rootnode().RLock()
265 func (fs *collectionFileSystem) RUnlock() {
266 fs.rootnode().RUnlock()
269 func (fs *collectionFileSystem) Parent() inode {
270 return fs.rootnode().Parent()
273 func (fs *collectionFileSystem) Read(_ []byte, ptr filenodePtr) (int, filenodePtr, error) {
274 return 0, ptr, ErrInvalidOperation
277 func (fs *collectionFileSystem) Write(_ []byte, ptr filenodePtr) (int, filenodePtr, error) {
278 return 0, ptr, ErrInvalidOperation
281 func (fs *collectionFileSystem) Readdir() ([]os.FileInfo, error) {
282 return fs.rootnode().Readdir()
285 func (fs *collectionFileSystem) SetParent(parent inode, name string) {
286 fs.rootnode().SetParent(parent, name)
289 func (fs *collectionFileSystem) Truncate(int64) error {
290 return ErrInvalidOperation
293 // Check for and incorporate upstream changes. If force==false, this
294 // is a no-op except once every ttl/100 or so.
296 // Return value is true if new content was loaded from upstream and
297 // any unsaved local changes have been discarded.
298 func (fs *collectionFileSystem) checkChangesOnServer(force bool) (bool, error) {
299 if fs.uuid == "" && fs.savedPDH.Load() == "" {
303 fs.lockCheckChanges.Lock()
304 if !force && fs.holdCheckChanges.After(time.Now()) {
305 fs.lockCheckChanges.Unlock()
308 remain, ttl := fs.signatureTimeLeft()
310 fs.holdCheckChanges = time.Now().Add(ttl / 100)
312 fs.lockCheckChanges.Unlock()
314 if !force && remain >= 0.5 {
315 // plenty of time left on current signatures
319 getparams := map[string]interface{}{"select": []string{"portable_data_hash", "manifest_text"}}
322 err := fs.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+fs.uuid, nil, getparams)
326 if coll.PortableDataHash != fs.savedPDH.Load().(string) {
327 // collection has changed upstream since we
328 // last loaded or saved. Refresh local data,
329 // losing any unsaved local changes.
330 newfs, err := coll.FileSystem(fs.fileSystem.fsBackend, fs.fileSystem.fsBackend)
334 snap, err := Snapshot(newfs, "/")
338 err = Splice(fs, "/", snap)
342 fs.savedPDH.Store(coll.PortableDataHash)
345 fs.updateSignatures(coll.ManifestText)
348 if pdh := fs.savedPDH.Load().(string); pdh != "" {
350 err := fs.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+pdh, nil, getparams)
354 fs.updateSignatures(coll.ManifestText)
359 // Refresh signature on a single locator, if necessary. Assume caller
360 // has lock. If an update is needed, and there are any storedSegments
361 // whose signatures can be updated, start a background task to update
362 // them asynchronously when the caller releases locks.
363 func (fs *collectionFileSystem) refreshSignature(locator string) string {
364 exp, err := signatureExpiryTime(locator)
365 if err != nil || exp.Sub(time.Now()) > time.Minute {
366 // Synchronous update is not needed. Start an
367 // asynchronous update if needed.
368 go fs.checkChangesOnServer(false)
372 for _, id := range []string{fs.uuid, fs.savedPDH.Load().(string)} {
377 err := fs.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+id, nil, map[string]interface{}{"select": []string{"portable_data_hash", "manifest_text"}})
381 manifests += coll.ManifestText
383 hash := stripAllHints(locator)
384 for _, tok := range regexp.MustCompile(`\S+`).FindAllString(manifests, -1) {
385 if mBlkRe.MatchString(tok) {
386 if stripAllHints(tok) == hash {
392 go fs.updateSignatures(manifests)
396 func (fs *collectionFileSystem) Sync() error {
397 refreshed, err := fs.checkChangesOnServer(true)
401 if refreshed || fs.uuid == "" {
404 txt, err := fs.MarshalManifest(".")
406 return fmt.Errorf("sync failed: %s", err)
408 if PortableDataHash(txt) == fs.savedPDH.Load() {
409 // No local changes since last save or initial load.
417 selectFields := []string{"uuid", "portable_data_hash"}
418 fs.lockCheckChanges.Lock()
419 remain, _ := fs.signatureTimeLeft()
420 fs.lockCheckChanges.Unlock()
422 selectFields = append(selectFields, "manifest_text")
425 err = fs.RequestAndDecode(&coll, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
426 "collection": map[string]string{
427 "manifest_text": coll.ManifestText,
429 "select": selectFields,
432 return fmt.Errorf("sync failed: update %s: %w", fs.uuid, err)
434 fs.updateSignatures(coll.ManifestText)
435 fs.savedPDH.Store(coll.PortableDataHash)
439 func (fs *collectionFileSystem) Flush(path string, shortBlocks bool) error {
440 node, err := rlookup(fs.fileSystem.root, path)
444 dn, ok := node.(*dirnode)
446 return ErrNotADirectory
450 names := dn.sortedNames()
452 // Caller only wants to flush the specified dir,
453 // non-recursively. Drop subdirs from the list of
455 var filenames []string
456 for _, name := range names {
457 if _, ok := dn.inodes[name].(*filenode); ok {
458 filenames = append(filenames, name)
463 for _, name := range names {
464 child := dn.inodes[name]
468 return dn.flush(context.TODO(), names, flushOpts{sync: false, shortBlocks: shortBlocks})
471 func (fs *collectionFileSystem) MemorySize() int64 {
472 return fs.fileSystem.root.(*dirnode).MemorySize()
475 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
476 fs.fileSystem.root.Lock()
477 defer fs.fileSystem.root.Unlock()
478 return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix)
481 func (fs *collectionFileSystem) Size() int64 {
482 return fs.fileSystem.root.(*dirnode).TreeSize()
485 func (fs *collectionFileSystem) Snapshot() (inode, error) {
486 return fs.fileSystem.root.Snapshot()
489 func (fs *collectionFileSystem) Splice(r inode) error {
490 return fs.fileSystem.root.Splice(r)
493 // filenodePtr is an offset into a file that is (usually) efficient to
494 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
496 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
497 // corresponds to file offset filenodePtr.off. Otherwise, it is
498 // necessary to reexamine len(filenode.segments[0]) etc. to find the
499 // correct segment and offset.
500 type filenodePtr struct {
507 // seek returns a ptr that is consistent with both startPtr.off and
508 // the current state of fn. The caller must already hold fn.RLock() or
511 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
516 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
518 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
519 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
522 // meaningless anyway
524 } else if ptr.off >= fn.fileinfo.size {
525 ptr.segmentIdx = len(fn.segments)
527 ptr.repacked = fn.repacked
529 } else if ptr.repacked == fn.repacked {
530 // segmentIdx and segmentOff accurately reflect
531 // ptr.off, but might have fallen off the end of a
533 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
540 ptr.repacked = fn.repacked
542 if ptr.off >= fn.fileinfo.size {
543 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
546 // Recompute segmentIdx and segmentOff. We have already
547 // established fn.fileinfo.size > ptr.off >= 0, so we don't
548 // have to deal with edge cases here.
550 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
551 // This would panic (index out of range) if
552 // fn.fileinfo.size were larger than
553 // sum(fn.segments[i].Len()) -- but that can't happen
554 // because we have ensured fn.fileinfo.size is always
556 segLen := int64(fn.segments[ptr.segmentIdx].Len())
557 if off+segLen > ptr.off {
558 ptr.segmentOff = int(ptr.off - off)
566 // filenode implements inode.
567 type filenode struct {
569 fs *collectionFileSystem
572 // number of times `segments` has changed in a
573 // way that might invalidate a filenodePtr
575 memsize int64 // bytes in memSegments
580 // caller must have lock
581 func (fn *filenode) appendSegment(e segment) {
582 fn.segments = append(fn.segments, e)
583 fn.fileinfo.size += int64(e.Len())
586 func (fn *filenode) SetParent(p inode, name string) {
590 fn.fileinfo.name = name
593 func (fn *filenode) Parent() inode {
599 func (fn *filenode) FS() FileSystem {
603 func (fn *filenode) MemorySize() (size int64) {
607 for _, seg := range fn.segments {
608 size += seg.memorySize()
613 // Read reads file data from a single segment, starting at startPtr,
614 // into p. startPtr is assumed not to be up-to-date. Caller must have
616 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
617 ptr = fn.seek(startPtr)
619 err = ErrNegativeOffset
622 if ptr.segmentIdx >= len(fn.segments) {
626 if ss, ok := fn.segments[ptr.segmentIdx].(storedSegment); ok {
627 ss.locator = fn.fs.refreshSignature(ss.locator)
628 fn.segments[ptr.segmentIdx] = ss
630 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
634 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
637 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
645 func (fn *filenode) Size() int64 {
648 return fn.fileinfo.Size()
651 func (fn *filenode) FileInfo() os.FileInfo {
657 func (fn *filenode) Truncate(size int64) error {
660 return fn.truncate(size)
663 func (fn *filenode) truncate(size int64) error {
664 if size == fn.fileinfo.size {
668 if size < fn.fileinfo.size {
669 ptr := fn.seek(filenodePtr{off: size})
670 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
671 if seg, ok := fn.segments[i].(*memSegment); ok {
672 fn.memsize -= int64(seg.Len())
675 if ptr.segmentOff == 0 {
676 fn.segments = fn.segments[:ptr.segmentIdx]
678 fn.segments = fn.segments[:ptr.segmentIdx+1]
679 switch seg := fn.segments[ptr.segmentIdx].(type) {
681 seg.Truncate(ptr.segmentOff)
682 fn.memsize += int64(seg.Len())
684 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
687 fn.fileinfo.size = size
690 for size > fn.fileinfo.size {
691 grow := size - fn.fileinfo.size
694 if len(fn.segments) == 0 {
696 fn.segments = append(fn.segments, seg)
697 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
699 fn.segments = append(fn.segments, seg)
701 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
704 seg.Truncate(seg.Len() + int(grow))
705 fn.fileinfo.size += grow
711 // Write writes data from p to the file, starting at startPtr,
712 // extending the file size if necessary. Caller must have Lock.
713 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
714 if startPtr.off > fn.fileinfo.size {
715 if err = fn.truncate(startPtr.off); err != nil {
716 return 0, startPtr, err
719 ptr = fn.seek(startPtr)
721 err = ErrNegativeOffset
724 for len(p) > 0 && err == nil {
726 if len(cando) > maxBlockSize {
727 cando = cando[:maxBlockSize]
729 // Rearrange/grow fn.segments (and shrink cando if
730 // needed) such that cando can be copied to
731 // fn.segments[ptr.segmentIdx] at offset
733 cur := ptr.segmentIdx
734 prev := ptr.segmentIdx - 1
736 if cur < len(fn.segments) {
737 _, curWritable = fn.segments[cur].(*memSegment)
739 var prevAppendable bool
740 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
741 _, prevAppendable = fn.segments[prev].(*memSegment)
743 if ptr.segmentOff > 0 && !curWritable {
744 // Split a non-writable block.
745 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
746 // Truncate cur, and insert a new
749 fn.segments = append(fn.segments, nil)
750 copy(fn.segments[cur+1:], fn.segments[cur:])
752 // Split cur into two copies, truncate
753 // the one on the left, shift the one
754 // on the right, and insert a new
755 // segment between them.
756 fn.segments = append(fn.segments, nil, nil)
757 copy(fn.segments[cur+2:], fn.segments[cur:])
758 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
763 seg.Truncate(len(cando))
764 fn.memsize += int64(len(cando))
765 fn.segments[cur] = seg
766 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
771 } else if curWritable {
772 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
777 // Shrink cando if needed to fit in
779 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
780 cando = cando[:cangrow]
784 if cur == len(fn.segments) {
785 // ptr is at EOF, filesize is changing.
786 fn.fileinfo.size += int64(len(cando))
787 } else if el := fn.segments[cur].Len(); el <= len(cando) {
788 // cando is long enough that we won't
789 // need cur any more. shrink cando to
790 // be exactly as long as cur
791 // (otherwise we'd accidentally shift
792 // the effective position of all
793 // segments after cur).
795 copy(fn.segments[cur:], fn.segments[cur+1:])
796 fn.segments = fn.segments[:len(fn.segments)-1]
798 // shrink cur by the same #bytes we're growing prev
799 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
805 ptr.segmentOff = fn.segments[prev].Len()
806 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
807 fn.memsize += int64(len(cando))
811 // Insert a segment between prev and
812 // cur, and advance prev/cur.
813 fn.segments = append(fn.segments, nil)
814 if cur < len(fn.segments) {
815 copy(fn.segments[cur+1:], fn.segments[cur:])
819 // appending a new segment does
820 // not invalidate any ptrs
823 seg.Truncate(len(cando))
824 fn.memsize += int64(len(cando))
825 fn.segments[cur] = seg
829 // Finally we can copy bytes from cando to the current segment.
830 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
834 ptr.off += int64(len(cando))
835 ptr.segmentOff += len(cando)
836 if ptr.segmentOff >= maxBlockSize {
837 fn.pruneMemSegments()
839 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
844 fn.fileinfo.modTime = time.Now()
849 // Write some data out to disk to reduce memory use. Caller must have
851 func (fn *filenode) pruneMemSegments() {
852 // TODO: share code with (*dirnode)flush()
853 // TODO: pack/flush small blocks too, when fragmented
854 for idx, seg := range fn.segments {
855 seg, ok := seg.(*memSegment)
856 if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
859 // Setting seg.flushing guarantees seg.buf will not be
860 // modified in place: WriteAt and Truncate will
861 // allocate a new buf instead, if necessary.
862 idx, buf := idx, seg.buf
863 done := make(chan struct{})
865 // If lots of background writes are already in
866 // progress, block here until one finishes, rather
867 // than pile up an unlimited number of buffered writes
868 // and network flush operations.
869 fn.fs.throttle().Acquire()
872 resp, err := fn.FS().BlockWrite(context.Background(), BlockWriteOptions{
874 Replicas: fn.fs.replicas,
875 StorageClasses: fn.fs.storageClasses,
877 fn.fs.throttle().Release()
880 if seg.flushing != done {
881 // A new seg.buf has been allocated.
885 // TODO: stall (or return errors from)
886 // subsequent writes until flushing
887 // starts to succeed.
890 if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
891 // Segment has been dropped/moved/resized.
894 fn.memsize -= int64(len(buf))
895 fn.segments[idx] = storedSegment{
897 locator: resp.Locator,
906 // Block until all pending pruneMemSegments/flush work is
907 // finished. Caller must NOT have lock.
908 func (fn *filenode) waitPrune() {
909 var pending []<-chan struct{}
911 for _, seg := range fn.segments {
912 if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
913 pending = append(pending, seg.flushing)
917 for _, p := range pending {
922 func (fn *filenode) Snapshot() (inode, error) {
925 segments := make([]segment, 0, len(fn.segments))
926 for _, seg := range fn.segments {
927 segments = append(segments, seg.Slice(0, seg.Len()))
930 fileinfo: fn.fileinfo,
935 func (fn *filenode) Splice(repl inode) error {
936 repl, err := repl.Snapshot()
941 defer fn.parent.Unlock()
944 _, err = fn.parent.Child(fn.fileinfo.name, func(inode) (inode, error) { return repl, nil })
948 switch repl := repl.(type) {
950 repl.parent = fn.parent
951 repl.fileinfo.name = fn.fileinfo.name
952 repl.setTreeFS(fn.fs)
954 repl.parent = fn.parent
955 repl.fileinfo.name = fn.fileinfo.name
958 return fmt.Errorf("cannot splice snapshot containing %T: %w", repl, ErrInvalidArgument)
963 type dirnode struct {
964 fs *collectionFileSystem
968 func (dn *dirnode) FS() FileSystem {
972 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
973 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
974 gn := &getternode{Getter: func() ([]byte, error) {
977 coll.ManifestText, err = dn.fs.MarshalManifest(".")
981 coll.UUID = dn.fs.uuid
982 data, err := json.Marshal(&coll)
984 data = append(data, '\n')
988 gn.SetParent(dn, name)
991 return dn.treenode.Child(name, replace)
994 type fnSegmentRef struct {
999 // commitBlock concatenates the data from the given filenode segments
1000 // (which must be *memSegments), writes the data out to Keep as a
1001 // single block, and replaces the filenodes' *memSegments with
1002 // storedSegments that reference the relevant portions of the new
1005 // bufsize is the total data size in refs. It is used to preallocate
1006 // the correct amount of memory when len(refs)>1.
1008 // If sync is false, commitBlock returns right away, after starting a
1009 // goroutine to do the writes, reacquire the filenodes' locks, and
1010 // swap out the *memSegments. Some filenodes' segments might get
1011 // modified/rearranged in the meantime, in which case commitBlock
1012 // won't replace them.
1014 // Caller must have write lock.
1015 func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize int, sync bool) error {
1019 if err := ctx.Err(); err != nil {
1022 done := make(chan struct{})
1024 segs := make([]*memSegment, 0, len(refs))
1025 offsets := make([]int, 0, len(refs)) // location of segment's data within block
1026 for _, ref := range refs {
1027 seg := ref.fn.segments[ref.idx].(*memSegment)
1028 if !sync && seg.flushingUnfinished() {
1029 // Let the other flushing goroutine finish. If
1030 // it fails, we'll try again next time.
1034 // In sync mode, we proceed regardless of
1035 // whether another flush is in progress: It
1036 // can't finish before we do, because we hold
1037 // fn's lock until we finish our own writes.
1039 offsets = append(offsets, len(block))
1042 } else if block == nil {
1043 block = append(make([]byte, 0, bufsize), seg.buf...)
1045 block = append(block, seg.buf...)
1047 segs = append(segs, seg)
1049 blocksize := len(block)
1050 dn.fs.throttle().Acquire()
1051 errs := make(chan error, 1)
1055 resp, err := dn.fs.BlockWrite(context.Background(), BlockWriteOptions{
1057 Replicas: dn.fs.replicas,
1058 StorageClasses: dn.fs.storageClasses,
1060 dn.fs.throttle().Release()
1065 for idx, ref := range refs {
1068 // In async mode, fn's lock was
1069 // released while we were waiting for
1070 // PutB(); lots of things might have
1072 if len(ref.fn.segments) <= ref.idx {
1073 // file segments have
1074 // rearranged or changed in
1078 } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] {
1079 // segment has been replaced
1082 } else if seg.flushing != done {
1083 // seg.buf has been replaced
1088 data := ref.fn.segments[ref.idx].(*memSegment).buf
1089 ref.fn.segments[ref.idx] = storedSegment{
1091 locator: resp.Locator,
1093 offset: offsets[idx],
1096 // atomic is needed here despite caller having
1097 // lock: caller might be running concurrent
1098 // commitBlock() goroutines using the same
1099 // lock, writing different segments from the
1101 atomic.AddInt64(&ref.fn.memsize, -int64(len(data)))
1113 type flushOpts struct {
1118 // flush in-memory data and remote-cluster block references (for the
1119 // children with the given names, which must be children of dn) to
1120 // local-cluster persistent storage.
1122 // Caller must have write lock on dn and the named children.
1124 // If any children are dirs, they will be flushed recursively.
1125 func (dn *dirnode) flush(ctx context.Context, names []string, opts flushOpts) error {
1126 cg := newContextGroup(ctx)
1129 goCommit := func(refs []fnSegmentRef, bufsize int) {
1130 cg.Go(func() error {
1131 return dn.commitBlock(cg.Context(), refs, bufsize, opts.sync)
1135 var pending []fnSegmentRef
1136 var pendingLen int = 0
1137 localLocator := map[string]string{}
1138 for _, name := range names {
1139 switch node := dn.inodes[name].(type) {
1141 grandchildNames := node.sortedNames()
1142 for _, grandchildName := range grandchildNames {
1143 grandchild := node.inodes[grandchildName]
1145 defer grandchild.Unlock()
1147 cg.Go(func() error { return node.flush(cg.Context(), grandchildNames, opts) })
1149 for idx, seg := range node.segments {
1150 switch seg := seg.(type) {
1152 loc, ok := localLocator[seg.locator]
1155 loc, err = dn.fs.LocalLocator(seg.locator)
1159 localLocator[seg.locator] = loc
1162 node.segments[idx] = seg
1164 if seg.Len() > maxBlockSize/2 {
1165 goCommit([]fnSegmentRef{{node, idx}}, seg.Len())
1168 if pendingLen+seg.Len() > maxBlockSize {
1169 goCommit(pending, pendingLen)
1173 pending = append(pending, fnSegmentRef{node, idx})
1174 pendingLen += seg.Len()
1176 panic(fmt.Sprintf("can't sync segment type %T", seg))
1181 if opts.shortBlocks {
1182 goCommit(pending, pendingLen)
1187 func (dn *dirnode) MemorySize() (size int64) {
1189 todo := make([]inode, 0, len(dn.inodes))
1190 for _, node := range dn.inodes {
1191 todo = append(todo, node)
1195 for _, node := range todo {
1196 size += node.MemorySize()
1201 // caller must have write lock.
1202 func (dn *dirnode) sortedNames() []string {
1203 names := make([]string, 0, len(dn.inodes))
1204 for name := range dn.inodes {
1205 names = append(names, name)
1211 // caller must have write lock.
1212 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string) (string, error) {
1213 cg := newContextGroup(ctx)
1216 if len(dn.inodes) == 0 {
1220 // Express the existence of an empty directory by
1221 // adding an empty file named `\056`, which (unlike
1222 // the more obvious spelling `.`) is accepted by the
1223 // API's manifest validator.
1224 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
1227 names := dn.sortedNames()
1229 // Wait for children to finish any pending write operations
1230 // before locking them.
1231 for _, name := range names {
1232 node := dn.inodes[name]
1233 if fn, ok := node.(*filenode); ok {
1238 var dirnames []string
1239 var filenames []string
1240 for _, name := range names {
1241 node := dn.inodes[name]
1244 switch node := node.(type) {
1246 dirnames = append(dirnames, name)
1248 filenames = append(filenames, name)
1250 panic(fmt.Sprintf("can't marshal inode type %T", node))
1254 subdirs := make([]string, len(dirnames))
1256 for i, name := range dirnames {
1258 cg.Go(func() error {
1259 txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name)
1265 cg.Go(func() error {
1267 type filepart struct {
1273 var fileparts []filepart
1275 if err := dn.flush(cg.Context(), filenames, flushOpts{sync: true, shortBlocks: true}); err != nil {
1278 for _, name := range filenames {
1279 node := dn.inodes[name].(*filenode)
1280 if len(node.segments) == 0 {
1281 fileparts = append(fileparts, filepart{name: name})
1284 for _, seg := range node.segments {
1285 switch seg := seg.(type) {
1287 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
1288 streamLen -= int64(seg.size)
1290 blocks = append(blocks, seg.locator)
1294 offset: streamLen + int64(seg.offset),
1295 length: int64(seg.length),
1297 if prev := len(fileparts) - 1; prev >= 0 &&
1298 fileparts[prev].name == name &&
1299 fileparts[prev].offset+fileparts[prev].length == next.offset {
1300 fileparts[prev].length += next.length
1302 fileparts = append(fileparts, next)
1304 streamLen += int64(seg.size)
1306 // This can't happen: we
1307 // haven't unlocked since
1308 // calling flush(sync=true).
1309 panic(fmt.Sprintf("can't marshal segment type %T", seg))
1313 var filetokens []string
1314 for _, s := range fileparts {
1315 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
1317 if len(filetokens) == 0 {
1319 } else if len(blocks) == 0 {
1320 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
1322 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
1326 return rootdir + strings.Join(subdirs, ""), err
1329 func (dn *dirnode) loadManifest(txt string) error {
1330 streams := bytes.Split([]byte(txt), []byte{'\n'})
1331 if len(streams[len(streams)-1]) != 0 {
1332 return fmt.Errorf("line %d: no trailing newline", len(streams))
1334 streams = streams[:len(streams)-1]
1335 segments := []storedSegment{}
1336 // To reduce allocs, we reuse a single "pathparts" slice
1337 // (pre-split on "/" separators) for the duration of this
1339 var pathparts []string
1340 // To reduce allocs, we reuse a single "toks" slice of 3 byte
1342 var toks = make([][]byte, 3)
1343 // Similar to bytes.SplitN(token, []byte{c}, 3), but splits
1344 // into the toks slice rather than allocating a new one, and
1345 // returns the number of toks (1, 2, or 3).
1346 splitToToks := func(src []byte, c rune) int {
1347 c1 := bytes.IndexRune(src, c)
1352 toks[0], src = src[:c1], src[c1+1:]
1353 c2 := bytes.IndexRune(src, c)
1358 toks[1], toks[2] = src[:c2], src[c2+1:]
1361 for i, stream := range streams {
1363 var anyFileTokens bool
1366 segments = segments[:0]
1369 for i, token := range bytes.Split(stream, []byte{' '}) {
1371 pathparts = strings.Split(manifestUnescape(string(token)), "/")
1372 streamparts = len(pathparts)
1375 if !bytes.ContainsRune(token, ':') {
1377 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1379 if splitToToks(token, '+') < 2 {
1380 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1382 length, err := strconv.ParseInt(string(toks[1]), 10, 32)
1383 if err != nil || length < 0 {
1384 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1386 segments = append(segments, storedSegment{
1387 locator: string(token),
1390 length: int(length),
1393 } else if len(segments) == 0 {
1394 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1396 if splitToToks(token, ':') != 3 {
1397 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1399 anyFileTokens = true
1401 offset, err := strconv.ParseInt(string(toks[0]), 10, 64)
1402 if err != nil || offset < 0 {
1403 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1405 length, err := strconv.ParseInt(string(toks[1]), 10, 64)
1406 if err != nil || length < 0 {
1407 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1409 if !bytes.ContainsAny(toks[2], `\/`) {
1410 // optimization for a common case
1411 pathparts = append(pathparts[:streamparts], string(toks[2]))
1413 pathparts = append(pathparts[:streamparts], strings.Split(manifestUnescape(string(toks[2])), "/")...)
1415 fnode, err := dn.createFileAndParents(pathparts)
1416 if fnode == nil && err == nil && length == 0 {
1417 // Special case: an empty file used as
1418 // a marker to preserve an otherwise
1419 // empty directory in a manifest.
1422 if err != nil || (fnode == nil && length != 0) {
1423 return fmt.Errorf("line %d: cannot use name %q with length %d: %s", lineno, toks[2], length, err)
1425 // Map the stream offset/range coordinates to
1426 // block/offset/range coordinates and add
1427 // corresponding storedSegments to the filenode
1429 // Can't continue where we left off.
1430 // TODO: binary search instead of
1431 // rewinding all the way (but this
1432 // situation might be rare anyway)
1435 for ; segIdx < len(segments); segIdx++ {
1436 seg := segments[segIdx]
1437 next := pos + int64(seg.Len())
1438 if next <= offset || seg.Len() == 0 {
1442 if pos >= offset+length {
1447 blkOff = int(offset - pos)
1449 blkLen := seg.Len() - blkOff
1450 if pos+int64(blkOff+blkLen) > offset+length {
1451 blkLen = int(offset + length - pos - int64(blkOff))
1453 fnode.appendSegment(storedSegment{
1455 locator: seg.locator,
1460 if next > offset+length {
1466 if segIdx == len(segments) && pos < offset+length {
1467 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1471 return fmt.Errorf("line %d: no file segments", lineno)
1472 } else if len(segments) == 0 {
1473 return fmt.Errorf("line %d: no locators", lineno)
1474 } else if streamparts == 0 {
1475 return fmt.Errorf("line %d: no stream name", lineno)
1481 // only safe to call from loadManifest -- no locking.
1483 // If path is a "parent directory exists" marker (the last path
1484 // component is "."), the returned values are both nil.
1486 // Newly added nodes have modtime==0. Caller is responsible for fixing
1487 // them with backdateTree.
1488 func (dn *dirnode) createFileAndParents(names []string) (fn *filenode, err error) {
1490 basename := names[len(names)-1]
1491 for _, name := range names[:len(names)-1] {
1497 // can't be sure parent will be a *dirnode
1498 return nil, ErrInvalidArgument
1500 node = node.Parent()
1504 unlock := node.Unlock
1505 node, err = node.Child(name, func(child inode) (inode, error) {
1507 // note modtime will be fixed later in backdateTree()
1508 child, err := node.FS().newNode(name, 0755|os.ModeDir, time.Time{})
1512 child.SetParent(node, name)
1514 } else if !child.IsDir() {
1515 return child, ErrFileExists
1525 if basename == "." {
1527 } else if !permittedName(basename) {
1528 err = fmt.Errorf("invalid file part %q in path %q", basename, names)
1533 _, err = node.Child(basename, func(child inode) (inode, error) {
1534 switch child := child.(type) {
1536 child, err = node.FS().newNode(basename, 0755, time.Time{})
1540 child.SetParent(node, basename)
1541 fn = child.(*filenode)
1547 return child, ErrIsDirectory
1549 return child, ErrInvalidArgument
1555 func (dn *dirnode) TreeSize() (bytes int64) {
1558 for _, i := range dn.inodes {
1559 switch i := i.(type) {
1563 bytes += i.TreeSize()
1569 func (dn *dirnode) Snapshot() (inode, error) {
1570 return dn.snapshot()
1573 func (dn *dirnode) snapshot() (*dirnode, error) {
1578 inodes: make(map[string]inode, len(dn.inodes)),
1579 fileinfo: dn.fileinfo,
1582 for name, child := range dn.inodes {
1583 dupchild, err := child.Snapshot()
1587 snap.inodes[name] = dupchild
1588 dupchild.SetParent(snap, name)
1593 func (dn *dirnode) Splice(repl inode) error {
1594 repl, err := repl.Snapshot()
1596 return fmt.Errorf("cannot copy snapshot: %w", err)
1598 switch repl := repl.(type) {
1600 return fmt.Errorf("cannot splice snapshot containing %T: %w", repl, ErrInvalidArgument)
1604 dn.inodes = repl.inodes
1608 defer dn.parent.Unlock()
1609 removing, err := dn.parent.Child(dn.fileinfo.name, nil)
1611 return fmt.Errorf("cannot use Splice to replace a top-level directory with a file: %w", ErrInvalidOperation)
1612 } else if removing != dn {
1613 // If ../thisdirname is not this dirnode, it
1614 // must be an inode that wraps a dirnode, like
1615 // a collectionFileSystem or deferrednode.
1616 if deferred, ok := removing.(*deferrednode); ok {
1617 // More useful to report the type of
1618 // the wrapped node rather than just
1619 // *deferrednode. (We know the real
1620 // inode is already loaded because dn
1622 removing = deferred.realinode()
1624 return fmt.Errorf("cannot use Splice to attach a file at top level of %T: %w", removing, ErrInvalidOperation)
1628 _, err = dn.parent.Child(dn.fileinfo.name, func(inode) (inode, error) { return repl, nil })
1630 return fmt.Errorf("error replacing filenode: dn.parent.Child(): %w", err)
1637 func (dn *dirnode) setTreeFS(fs *collectionFileSystem) {
1639 for _, child := range dn.inodes {
1640 switch child := child.(type) {
1649 type segment interface {
1652 // Return a new segment with a subsection of the data from this
1653 // one. length<0 means length=Len()-off.
1654 Slice(off int, length int) segment
1658 type memSegment struct {
1660 // If flushing is not nil and not ready/closed, then a) buf is
1661 // being shared by a pruneMemSegments goroutine, and must be
1662 // copied on write; and b) the flushing channel will close
1663 // when the goroutine finishes, whether it succeeds or not.
1664 flushing <-chan struct{}
1667 func (me *memSegment) flushingUnfinished() bool {
1668 if me.flushing == nil {
1680 func (me *memSegment) Len() int {
1684 func (me *memSegment) Slice(off, length int) segment {
1686 length = len(me.buf) - off
1688 buf := make([]byte, length)
1689 copy(buf, me.buf[off:])
1690 return &memSegment{buf: buf}
1693 func (me *memSegment) Truncate(n int) {
1694 if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
1697 newsize = newsize << 2
1699 newbuf := make([]byte, n, newsize)
1700 copy(newbuf, me.buf)
1701 me.buf, me.flushing = newbuf, nil
1703 // reclaim existing capacity, and zero reclaimed part
1704 oldlen := len(me.buf)
1706 for i := oldlen; i < n; i++ {
1712 func (me *memSegment) WriteAt(p []byte, off int) {
1713 if off+len(p) > len(me.buf) {
1714 panic("overflowed segment")
1716 if me.flushing != nil {
1717 me.buf, me.flushing = append([]byte(nil), me.buf...), nil
1719 copy(me.buf[off:], p)
1722 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1723 if off > int64(me.Len()) {
1727 n = copy(p, me.buf[int(off):])
1734 func (me *memSegment) memorySize() int64 {
1735 return 64 + int64(len(me.buf))
1738 type storedSegment struct {
1741 size int // size of stored block (also encoded in locator)
1742 offset int // position of segment within the stored block
1743 length int // bytes in this segment (offset + length <= size)
1746 func (se storedSegment) Len() int {
1750 func (se storedSegment) Slice(n, size int) segment {
1753 if size >= 0 && se.length > size {
1759 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1760 if off > int64(se.length) {
1763 maxlen := se.length - int(off)
1764 if len(p) > maxlen {
1766 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1772 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1775 func (se storedSegment) memorySize() int64 {
1776 return 64 + int64(len(se.locator))
1779 func canonicalName(name string) string {
1780 name = path.Clean("/" + name)
1781 if name == "/" || name == "./" {
1783 } else if strings.HasPrefix(name, "/") {
1789 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1791 func manifestUnescapeFunc(seq string) string {
1795 i, err := strconv.ParseUint(seq[1:], 8, 8)
1797 // Invalid escape sequence: can't unescape.
1800 return string([]byte{byte(i)})
1803 func manifestUnescape(s string) string {
1804 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1807 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1809 func manifestEscapeFunc(seq string) string {
1810 return fmt.Sprintf("\\%03o", byte(seq[0]))
1813 func manifestEscape(s string) string {
1814 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)