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 -- unless that has
294 // already been done recently, in which case this func is a no-op.
295 func (fs *collectionFileSystem) checkChangesOnServer() error {
296 if fs.uuid == "" && fs.savedPDH.Load() == "" {
300 // First try UUID if any, then last known PDH. Stop if all
301 // signatures are new enough.
303 for _, id := range []string{fs.uuid, fs.savedPDH.Load().(string)} {
308 fs.lockCheckChanges.Lock()
309 if !checkingAll && fs.holdCheckChanges.After(time.Now()) {
310 fs.lockCheckChanges.Unlock()
313 remain, ttl := fs.signatureTimeLeft()
314 if remain > 0.01 && !checkingAll {
315 fs.holdCheckChanges = time.Now().Add(ttl / 100)
317 fs.lockCheckChanges.Unlock()
324 err := fs.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+id, nil, map[string]interface{}{"select": []string{"portable_data_hash", "manifest_text"}})
328 fs.updateSignatures(coll.ManifestText)
333 // Refresh signature on a single locator, if necessary. Assume caller
334 // has lock. If an update is needed, and there are any storedSegments
335 // whose signatures can be updated, start a background task to update
336 // them asynchronously when the caller releases locks.
337 func (fs *collectionFileSystem) refreshSignature(locator string) string {
338 exp, err := signatureExpiryTime(locator)
339 if err != nil || exp.Sub(time.Now()) > time.Minute {
340 // Synchronous update is not needed. Start an
341 // asynchronous update if needed.
342 go fs.checkChangesOnServer()
346 for _, id := range []string{fs.uuid, fs.savedPDH.Load().(string)} {
351 err := fs.RequestAndDecode(&coll, "GET", "arvados/v1/collections/"+id, nil, map[string]interface{}{"select": []string{"portable_data_hash", "manifest_text"}})
355 manifests += coll.ManifestText
357 hash := stripAllHints(locator)
358 for _, tok := range regexp.MustCompile(`\S+`).FindAllString(manifests, -1) {
359 if mBlkRe.MatchString(tok) {
360 if stripAllHints(tok) == hash {
366 go fs.updateSignatures(manifests)
370 func (fs *collectionFileSystem) Sync() error {
371 err := fs.checkChangesOnServer()
378 txt, err := fs.MarshalManifest(".")
380 return fmt.Errorf("sync failed: %s", err)
382 if PortableDataHash(txt) == fs.savedPDH.Load() {
383 // No local changes since last save or initial load.
391 selectFields := []string{"uuid", "portable_data_hash"}
392 fs.lockCheckChanges.Lock()
393 remain, _ := fs.signatureTimeLeft()
394 fs.lockCheckChanges.Unlock()
396 selectFields = append(selectFields, "manifest_text")
399 err = fs.RequestAndDecode(&coll, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
400 "collection": map[string]string{
401 "manifest_text": coll.ManifestText,
403 "select": selectFields,
406 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
408 fs.updateSignatures(coll.ManifestText)
409 fs.savedPDH.Store(coll.PortableDataHash)
413 func (fs *collectionFileSystem) Flush(path string, shortBlocks bool) error {
414 node, err := rlookup(fs.fileSystem.root, path)
418 dn, ok := node.(*dirnode)
420 return ErrNotADirectory
424 names := dn.sortedNames()
426 // Caller only wants to flush the specified dir,
427 // non-recursively. Drop subdirs from the list of
429 var filenames []string
430 for _, name := range names {
431 if _, ok := dn.inodes[name].(*filenode); ok {
432 filenames = append(filenames, name)
437 for _, name := range names {
438 child := dn.inodes[name]
442 return dn.flush(context.TODO(), names, flushOpts{sync: false, shortBlocks: shortBlocks})
445 func (fs *collectionFileSystem) MemorySize() int64 {
446 fs.fileSystem.root.Lock()
447 defer fs.fileSystem.root.Unlock()
448 return fs.fileSystem.root.(*dirnode).MemorySize()
451 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
452 fs.fileSystem.root.Lock()
453 defer fs.fileSystem.root.Unlock()
454 return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix)
457 func (fs *collectionFileSystem) Size() int64 {
458 return fs.fileSystem.root.(*dirnode).TreeSize()
461 func (fs *collectionFileSystem) Snapshot() (inode, error) {
462 return fs.fileSystem.root.Snapshot()
465 func (fs *collectionFileSystem) Splice(r inode) error {
466 return fs.fileSystem.root.Splice(r)
469 // filenodePtr is an offset into a file that is (usually) efficient to
470 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
472 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
473 // corresponds to file offset filenodePtr.off. Otherwise, it is
474 // necessary to reexamine len(filenode.segments[0]) etc. to find the
475 // correct segment and offset.
476 type filenodePtr struct {
483 // seek returns a ptr that is consistent with both startPtr.off and
484 // the current state of fn. The caller must already hold fn.RLock() or
487 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
492 // ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
494 // filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
495 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
498 // meaningless anyway
500 } else if ptr.off >= fn.fileinfo.size {
501 ptr.segmentIdx = len(fn.segments)
503 ptr.repacked = fn.repacked
505 } else if ptr.repacked == fn.repacked {
506 // segmentIdx and segmentOff accurately reflect
507 // ptr.off, but might have fallen off the end of a
509 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
516 ptr.repacked = fn.repacked
518 if ptr.off >= fn.fileinfo.size {
519 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
522 // Recompute segmentIdx and segmentOff. We have already
523 // established fn.fileinfo.size > ptr.off >= 0, so we don't
524 // have to deal with edge cases here.
526 for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
527 // This would panic (index out of range) if
528 // fn.fileinfo.size were larger than
529 // sum(fn.segments[i].Len()) -- but that can't happen
530 // because we have ensured fn.fileinfo.size is always
532 segLen := int64(fn.segments[ptr.segmentIdx].Len())
533 if off+segLen > ptr.off {
534 ptr.segmentOff = int(ptr.off - off)
542 // filenode implements inode.
543 type filenode struct {
545 fs *collectionFileSystem
548 // number of times `segments` has changed in a
549 // way that might invalidate a filenodePtr
551 memsize int64 // bytes in memSegments
556 // caller must have lock
557 func (fn *filenode) appendSegment(e segment) {
558 fn.segments = append(fn.segments, e)
559 fn.fileinfo.size += int64(e.Len())
562 func (fn *filenode) SetParent(p inode, name string) {
566 fn.fileinfo.name = name
569 func (fn *filenode) Parent() inode {
575 func (fn *filenode) FS() FileSystem {
579 // Read reads file data from a single segment, starting at startPtr,
580 // into p. startPtr is assumed not to be up-to-date. Caller must have
582 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
583 ptr = fn.seek(startPtr)
585 err = ErrNegativeOffset
588 if ptr.segmentIdx >= len(fn.segments) {
592 if ss, ok := fn.segments[ptr.segmentIdx].(storedSegment); ok {
593 ss.locator = fn.fs.refreshSignature(ss.locator)
594 fn.segments[ptr.segmentIdx] = ss
596 n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
600 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
603 if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
611 func (fn *filenode) Size() int64 {
614 return fn.fileinfo.Size()
617 func (fn *filenode) FileInfo() os.FileInfo {
623 func (fn *filenode) Truncate(size int64) error {
626 return fn.truncate(size)
629 func (fn *filenode) truncate(size int64) error {
630 if size == fn.fileinfo.size {
634 if size < fn.fileinfo.size {
635 ptr := fn.seek(filenodePtr{off: size})
636 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
637 if seg, ok := fn.segments[i].(*memSegment); ok {
638 fn.memsize -= int64(seg.Len())
641 if ptr.segmentOff == 0 {
642 fn.segments = fn.segments[:ptr.segmentIdx]
644 fn.segments = fn.segments[:ptr.segmentIdx+1]
645 switch seg := fn.segments[ptr.segmentIdx].(type) {
647 seg.Truncate(ptr.segmentOff)
648 fn.memsize += int64(seg.Len())
650 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
653 fn.fileinfo.size = size
656 for size > fn.fileinfo.size {
657 grow := size - fn.fileinfo.size
660 if len(fn.segments) == 0 {
662 fn.segments = append(fn.segments, seg)
663 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
665 fn.segments = append(fn.segments, seg)
667 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
670 seg.Truncate(seg.Len() + int(grow))
671 fn.fileinfo.size += grow
677 // Write writes data from p to the file, starting at startPtr,
678 // extending the file size if necessary. Caller must have Lock.
679 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
680 if startPtr.off > fn.fileinfo.size {
681 if err = fn.truncate(startPtr.off); err != nil {
682 return 0, startPtr, err
685 ptr = fn.seek(startPtr)
687 err = ErrNegativeOffset
690 for len(p) > 0 && err == nil {
692 if len(cando) > maxBlockSize {
693 cando = cando[:maxBlockSize]
695 // Rearrange/grow fn.segments (and shrink cando if
696 // needed) such that cando can be copied to
697 // fn.segments[ptr.segmentIdx] at offset
699 cur := ptr.segmentIdx
700 prev := ptr.segmentIdx - 1
702 if cur < len(fn.segments) {
703 _, curWritable = fn.segments[cur].(*memSegment)
705 var prevAppendable bool
706 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
707 _, prevAppendable = fn.segments[prev].(*memSegment)
709 if ptr.segmentOff > 0 && !curWritable {
710 // Split a non-writable block.
711 if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
712 // Truncate cur, and insert a new
715 fn.segments = append(fn.segments, nil)
716 copy(fn.segments[cur+1:], fn.segments[cur:])
718 // Split cur into two copies, truncate
719 // the one on the left, shift the one
720 // on the right, and insert a new
721 // segment between them.
722 fn.segments = append(fn.segments, nil, nil)
723 copy(fn.segments[cur+2:], fn.segments[cur:])
724 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
729 seg.Truncate(len(cando))
730 fn.memsize += int64(len(cando))
731 fn.segments[cur] = seg
732 fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
737 } else if curWritable {
738 if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
743 // Shrink cando if needed to fit in
745 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
746 cando = cando[:cangrow]
750 if cur == len(fn.segments) {
751 // ptr is at EOF, filesize is changing.
752 fn.fileinfo.size += int64(len(cando))
753 } else if el := fn.segments[cur].Len(); el <= len(cando) {
754 // cando is long enough that we won't
755 // need cur any more. shrink cando to
756 // be exactly as long as cur
757 // (otherwise we'd accidentally shift
758 // the effective position of all
759 // segments after cur).
761 copy(fn.segments[cur:], fn.segments[cur+1:])
762 fn.segments = fn.segments[:len(fn.segments)-1]
764 // shrink cur by the same #bytes we're growing prev
765 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
771 ptr.segmentOff = fn.segments[prev].Len()
772 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
773 fn.memsize += int64(len(cando))
777 // Insert a segment between prev and
778 // cur, and advance prev/cur.
779 fn.segments = append(fn.segments, nil)
780 if cur < len(fn.segments) {
781 copy(fn.segments[cur+1:], fn.segments[cur:])
785 // appending a new segment does
786 // not invalidate any ptrs
789 seg.Truncate(len(cando))
790 fn.memsize += int64(len(cando))
791 fn.segments[cur] = seg
795 // Finally we can copy bytes from cando to the current segment.
796 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
800 ptr.off += int64(len(cando))
801 ptr.segmentOff += len(cando)
802 if ptr.segmentOff >= maxBlockSize {
803 fn.pruneMemSegments()
805 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
810 fn.fileinfo.modTime = time.Now()
815 // Write some data out to disk to reduce memory use. Caller must have
817 func (fn *filenode) pruneMemSegments() {
818 // TODO: share code with (*dirnode)flush()
819 // TODO: pack/flush small blocks too, when fragmented
820 for idx, seg := range fn.segments {
821 seg, ok := seg.(*memSegment)
822 if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
825 // Setting seg.flushing guarantees seg.buf will not be
826 // modified in place: WriteAt and Truncate will
827 // allocate a new buf instead, if necessary.
828 idx, buf := idx, seg.buf
829 done := make(chan struct{})
831 // If lots of background writes are already in
832 // progress, block here until one finishes, rather
833 // than pile up an unlimited number of buffered writes
834 // and network flush operations.
835 fn.fs.throttle().Acquire()
838 resp, err := fn.FS().BlockWrite(context.Background(), BlockWriteOptions{
840 Replicas: fn.fs.replicas,
841 StorageClasses: fn.fs.storageClasses,
843 fn.fs.throttle().Release()
846 if seg.flushing != done {
847 // A new seg.buf has been allocated.
851 // TODO: stall (or return errors from)
852 // subsequent writes until flushing
853 // starts to succeed.
856 if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
857 // Segment has been dropped/moved/resized.
860 fn.memsize -= int64(len(buf))
861 fn.segments[idx] = storedSegment{
863 locator: resp.Locator,
872 // Block until all pending pruneMemSegments/flush work is
873 // finished. Caller must NOT have lock.
874 func (fn *filenode) waitPrune() {
875 var pending []<-chan struct{}
877 for _, seg := range fn.segments {
878 if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
879 pending = append(pending, seg.flushing)
883 for _, p := range pending {
888 func (fn *filenode) Snapshot() (inode, error) {
891 segments := make([]segment, 0, len(fn.segments))
892 for _, seg := range fn.segments {
893 segments = append(segments, seg.Slice(0, seg.Len()))
896 fileinfo: fn.fileinfo,
901 func (fn *filenode) Splice(repl inode) error {
902 repl, err := repl.Snapshot()
907 defer fn.parent.Unlock()
910 _, err = fn.parent.Child(fn.fileinfo.name, func(inode) (inode, error) { return repl, nil })
914 switch repl := repl.(type) {
916 repl.parent = fn.parent
917 repl.fileinfo.name = fn.fileinfo.name
918 repl.setTreeFS(fn.fs)
920 repl.parent = fn.parent
921 repl.fileinfo.name = fn.fileinfo.name
924 return fmt.Errorf("cannot splice snapshot containing %T: %w", repl, ErrInvalidArgument)
929 type dirnode struct {
930 fs *collectionFileSystem
934 func (dn *dirnode) FS() FileSystem {
938 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
939 if dn == dn.fs.rootnode() && name == ".arvados#collection" {
940 gn := &getternode{Getter: func() ([]byte, error) {
943 coll.ManifestText, err = dn.fs.MarshalManifest(".")
947 coll.UUID = dn.fs.uuid
948 data, err := json.Marshal(&coll)
950 data = append(data, '\n')
954 gn.SetParent(dn, name)
957 return dn.treenode.Child(name, replace)
960 type fnSegmentRef struct {
965 // commitBlock concatenates the data from the given filenode segments
966 // (which must be *memSegments), writes the data out to Keep as a
967 // single block, and replaces the filenodes' *memSegments with
968 // storedSegments that reference the relevant portions of the new
971 // bufsize is the total data size in refs. It is used to preallocate
972 // the correct amount of memory when len(refs)>1.
974 // If sync is false, commitBlock returns right away, after starting a
975 // goroutine to do the writes, reacquire the filenodes' locks, and
976 // swap out the *memSegments. Some filenodes' segments might get
977 // modified/rearranged in the meantime, in which case commitBlock
978 // won't replace them.
980 // Caller must have write lock.
981 func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize int, sync bool) error {
985 if err := ctx.Err(); err != nil {
988 done := make(chan struct{})
990 segs := make([]*memSegment, 0, len(refs))
991 offsets := make([]int, 0, len(refs)) // location of segment's data within block
992 for _, ref := range refs {
993 seg := ref.fn.segments[ref.idx].(*memSegment)
994 if !sync && seg.flushingUnfinished() {
995 // Let the other flushing goroutine finish. If
996 // it fails, we'll try again next time.
1000 // In sync mode, we proceed regardless of
1001 // whether another flush is in progress: It
1002 // can't finish before we do, because we hold
1003 // fn's lock until we finish our own writes.
1005 offsets = append(offsets, len(block))
1008 } else if block == nil {
1009 block = append(make([]byte, 0, bufsize), seg.buf...)
1011 block = append(block, seg.buf...)
1013 segs = append(segs, seg)
1015 blocksize := len(block)
1016 dn.fs.throttle().Acquire()
1017 errs := make(chan error, 1)
1021 resp, err := dn.fs.BlockWrite(context.Background(), BlockWriteOptions{
1023 Replicas: dn.fs.replicas,
1024 StorageClasses: dn.fs.storageClasses,
1026 dn.fs.throttle().Release()
1031 for idx, ref := range refs {
1034 // In async mode, fn's lock was
1035 // released while we were waiting for
1036 // PutB(); lots of things might have
1038 if len(ref.fn.segments) <= ref.idx {
1039 // file segments have
1040 // rearranged or changed in
1044 } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] {
1045 // segment has been replaced
1048 } else if seg.flushing != done {
1049 // seg.buf has been replaced
1054 data := ref.fn.segments[ref.idx].(*memSegment).buf
1055 ref.fn.segments[ref.idx] = storedSegment{
1057 locator: resp.Locator,
1059 offset: offsets[idx],
1062 // atomic is needed here despite caller having
1063 // lock: caller might be running concurrent
1064 // commitBlock() goroutines using the same
1065 // lock, writing different segments from the
1067 atomic.AddInt64(&ref.fn.memsize, -int64(len(data)))
1079 type flushOpts struct {
1084 // flush in-memory data and remote-cluster block references (for the
1085 // children with the given names, which must be children of dn) to
1086 // local-cluster persistent storage.
1088 // Caller must have write lock on dn and the named children.
1090 // If any children are dirs, they will be flushed recursively.
1091 func (dn *dirnode) flush(ctx context.Context, names []string, opts flushOpts) error {
1092 cg := newContextGroup(ctx)
1095 goCommit := func(refs []fnSegmentRef, bufsize int) {
1096 cg.Go(func() error {
1097 return dn.commitBlock(cg.Context(), refs, bufsize, opts.sync)
1101 var pending []fnSegmentRef
1102 var pendingLen int = 0
1103 localLocator := map[string]string{}
1104 for _, name := range names {
1105 switch node := dn.inodes[name].(type) {
1107 grandchildNames := node.sortedNames()
1108 for _, grandchildName := range grandchildNames {
1109 grandchild := node.inodes[grandchildName]
1111 defer grandchild.Unlock()
1113 cg.Go(func() error { return node.flush(cg.Context(), grandchildNames, opts) })
1115 for idx, seg := range node.segments {
1116 switch seg := seg.(type) {
1118 loc, ok := localLocator[seg.locator]
1121 loc, err = dn.fs.LocalLocator(seg.locator)
1125 localLocator[seg.locator] = loc
1128 node.segments[idx] = seg
1130 if seg.Len() > maxBlockSize/2 {
1131 goCommit([]fnSegmentRef{{node, idx}}, seg.Len())
1134 if pendingLen+seg.Len() > maxBlockSize {
1135 goCommit(pending, pendingLen)
1139 pending = append(pending, fnSegmentRef{node, idx})
1140 pendingLen += seg.Len()
1142 panic(fmt.Sprintf("can't sync segment type %T", seg))
1147 if opts.shortBlocks {
1148 goCommit(pending, pendingLen)
1153 // caller must have write lock.
1154 func (dn *dirnode) MemorySize() (size int64) {
1155 for _, name := range dn.sortedNames() {
1156 node := dn.inodes[name]
1159 switch node := node.(type) {
1161 size += node.MemorySize()
1164 for _, seg := range node.segments {
1165 switch seg := seg.(type) {
1167 size += int64(seg.Len())
1176 // caller must have write lock.
1177 func (dn *dirnode) sortedNames() []string {
1178 names := make([]string, 0, len(dn.inodes))
1179 for name := range dn.inodes {
1180 names = append(names, name)
1186 // caller must have write lock.
1187 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string) (string, error) {
1188 cg := newContextGroup(ctx)
1191 if len(dn.inodes) == 0 {
1195 // Express the existence of an empty directory by
1196 // adding an empty file named `\056`, which (unlike
1197 // the more obvious spelling `.`) is accepted by the
1198 // API's manifest validator.
1199 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
1202 names := dn.sortedNames()
1204 // Wait for children to finish any pending write operations
1205 // before locking them.
1206 for _, name := range names {
1207 node := dn.inodes[name]
1208 if fn, ok := node.(*filenode); ok {
1213 var dirnames []string
1214 var filenames []string
1215 for _, name := range names {
1216 node := dn.inodes[name]
1219 switch node := node.(type) {
1221 dirnames = append(dirnames, name)
1223 filenames = append(filenames, name)
1225 panic(fmt.Sprintf("can't marshal inode type %T", node))
1229 subdirs := make([]string, len(dirnames))
1231 for i, name := range dirnames {
1233 cg.Go(func() error {
1234 txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name)
1240 cg.Go(func() error {
1242 type filepart struct {
1248 var fileparts []filepart
1250 if err := dn.flush(cg.Context(), filenames, flushOpts{sync: true, shortBlocks: true}); err != nil {
1253 for _, name := range filenames {
1254 node := dn.inodes[name].(*filenode)
1255 if len(node.segments) == 0 {
1256 fileparts = append(fileparts, filepart{name: name})
1259 for _, seg := range node.segments {
1260 switch seg := seg.(type) {
1262 if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
1263 streamLen -= int64(seg.size)
1265 blocks = append(blocks, seg.locator)
1269 offset: streamLen + int64(seg.offset),
1270 length: int64(seg.length),
1272 if prev := len(fileparts) - 1; prev >= 0 &&
1273 fileparts[prev].name == name &&
1274 fileparts[prev].offset+fileparts[prev].length == next.offset {
1275 fileparts[prev].length += next.length
1277 fileparts = append(fileparts, next)
1279 streamLen += int64(seg.size)
1281 // This can't happen: we
1282 // haven't unlocked since
1283 // calling flush(sync=true).
1284 panic(fmt.Sprintf("can't marshal segment type %T", seg))
1288 var filetokens []string
1289 for _, s := range fileparts {
1290 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
1292 if len(filetokens) == 0 {
1294 } else if len(blocks) == 0 {
1295 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
1297 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
1301 return rootdir + strings.Join(subdirs, ""), err
1304 func (dn *dirnode) loadManifest(txt string) error {
1305 streams := bytes.Split([]byte(txt), []byte{'\n'})
1306 if len(streams[len(streams)-1]) != 0 {
1307 return fmt.Errorf("line %d: no trailing newline", len(streams))
1309 streams = streams[:len(streams)-1]
1310 segments := []storedSegment{}
1311 // To reduce allocs, we reuse a single "pathparts" slice
1312 // (pre-split on "/" separators) for the duration of this
1314 var pathparts []string
1315 // To reduce allocs, we reuse a single "toks" slice of 3 byte
1317 var toks = make([][]byte, 3)
1318 // Similar to bytes.SplitN(token, []byte{c}, 3), but splits
1319 // into the toks slice rather than allocating a new one, and
1320 // returns the number of toks (1, 2, or 3).
1321 splitToToks := func(src []byte, c rune) int {
1322 c1 := bytes.IndexRune(src, c)
1327 toks[0], src = src[:c1], src[c1+1:]
1328 c2 := bytes.IndexRune(src, c)
1333 toks[1], toks[2] = src[:c2], src[c2+1:]
1336 for i, stream := range streams {
1338 var anyFileTokens bool
1341 segments = segments[:0]
1344 for i, token := range bytes.Split(stream, []byte{' '}) {
1346 pathparts = strings.Split(manifestUnescape(string(token)), "/")
1347 streamparts = len(pathparts)
1350 if !bytes.ContainsRune(token, ':') {
1352 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1354 if splitToToks(token, '+') < 2 {
1355 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1357 length, err := strconv.ParseInt(string(toks[1]), 10, 32)
1358 if err != nil || length < 0 {
1359 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1361 segments = append(segments, storedSegment{
1362 locator: string(token),
1365 length: int(length),
1368 } else if len(segments) == 0 {
1369 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1371 if splitToToks(token, ':') != 3 {
1372 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1374 anyFileTokens = true
1376 offset, err := strconv.ParseInt(string(toks[0]), 10, 64)
1377 if err != nil || offset < 0 {
1378 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1380 length, err := strconv.ParseInt(string(toks[1]), 10, 64)
1381 if err != nil || length < 0 {
1382 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1384 if !bytes.ContainsAny(toks[2], `\/`) {
1385 // optimization for a common case
1386 pathparts = append(pathparts[:streamparts], string(toks[2]))
1388 pathparts = append(pathparts[:streamparts], strings.Split(manifestUnescape(string(toks[2])), "/")...)
1390 fnode, err := dn.createFileAndParents(pathparts)
1391 if fnode == nil && err == nil && length == 0 {
1392 // Special case: an empty file used as
1393 // a marker to preserve an otherwise
1394 // empty directory in a manifest.
1397 if err != nil || (fnode == nil && length != 0) {
1398 return fmt.Errorf("line %d: cannot use name %q with length %d: %s", lineno, toks[2], length, err)
1400 // Map the stream offset/range coordinates to
1401 // block/offset/range coordinates and add
1402 // corresponding storedSegments to the filenode
1404 // Can't continue where we left off.
1405 // TODO: binary search instead of
1406 // rewinding all the way (but this
1407 // situation might be rare anyway)
1410 for ; segIdx < len(segments); segIdx++ {
1411 seg := segments[segIdx]
1412 next := pos + int64(seg.Len())
1413 if next <= offset || seg.Len() == 0 {
1417 if pos >= offset+length {
1422 blkOff = int(offset - pos)
1424 blkLen := seg.Len() - blkOff
1425 if pos+int64(blkOff+blkLen) > offset+length {
1426 blkLen = int(offset + length - pos - int64(blkOff))
1428 fnode.appendSegment(storedSegment{
1430 locator: seg.locator,
1435 if next > offset+length {
1441 if segIdx == len(segments) && pos < offset+length {
1442 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1446 return fmt.Errorf("line %d: no file segments", lineno)
1447 } else if len(segments) == 0 {
1448 return fmt.Errorf("line %d: no locators", lineno)
1449 } else if streamparts == 0 {
1450 return fmt.Errorf("line %d: no stream name", lineno)
1456 // only safe to call from loadManifest -- no locking.
1458 // If path is a "parent directory exists" marker (the last path
1459 // component is "."), the returned values are both nil.
1461 // Newly added nodes have modtime==0. Caller is responsible for fixing
1462 // them with backdateTree.
1463 func (dn *dirnode) createFileAndParents(names []string) (fn *filenode, err error) {
1465 basename := names[len(names)-1]
1466 for _, name := range names[:len(names)-1] {
1472 // can't be sure parent will be a *dirnode
1473 return nil, ErrInvalidArgument
1475 node = node.Parent()
1479 unlock := node.Unlock
1480 node, err = node.Child(name, func(child inode) (inode, error) {
1482 // note modtime will be fixed later in backdateTree()
1483 child, err := node.FS().newNode(name, 0755|os.ModeDir, time.Time{})
1487 child.SetParent(node, name)
1489 } else if !child.IsDir() {
1490 return child, ErrFileExists
1500 if basename == "." {
1502 } else if !permittedName(basename) {
1503 err = fmt.Errorf("invalid file part %q in path %q", basename, names)
1508 _, err = node.Child(basename, func(child inode) (inode, error) {
1509 switch child := child.(type) {
1511 child, err = node.FS().newNode(basename, 0755, time.Time{})
1515 child.SetParent(node, basename)
1516 fn = child.(*filenode)
1522 return child, ErrIsDirectory
1524 return child, ErrInvalidArgument
1530 func (dn *dirnode) TreeSize() (bytes int64) {
1533 for _, i := range dn.inodes {
1534 switch i := i.(type) {
1538 bytes += i.TreeSize()
1544 func (dn *dirnode) Snapshot() (inode, error) {
1545 return dn.snapshot()
1548 func (dn *dirnode) snapshot() (*dirnode, error) {
1553 inodes: make(map[string]inode, len(dn.inodes)),
1554 fileinfo: dn.fileinfo,
1557 for name, child := range dn.inodes {
1558 dupchild, err := child.Snapshot()
1562 snap.inodes[name] = dupchild
1563 dupchild.SetParent(snap, name)
1568 func (dn *dirnode) Splice(repl inode) error {
1569 repl, err := repl.Snapshot()
1571 return fmt.Errorf("cannot copy snapshot: %w", err)
1573 switch repl := repl.(type) {
1575 return fmt.Errorf("cannot splice snapshot containing %T: %w", repl, ErrInvalidArgument)
1579 dn.inodes = repl.inodes
1583 defer dn.parent.Unlock()
1584 removing, err := dn.parent.Child(dn.fileinfo.name, nil)
1586 return fmt.Errorf("cannot use Splice to replace a top-level directory with a file: %w", ErrInvalidOperation)
1587 } else if removing != dn {
1588 // If ../thisdirname is not this dirnode, it
1589 // must be an inode that wraps a dirnode, like
1590 // a collectionFileSystem or deferrednode.
1591 if deferred, ok := removing.(*deferrednode); ok {
1592 // More useful to report the type of
1593 // the wrapped node rather than just
1594 // *deferrednode. (We know the real
1595 // inode is already loaded because dn
1597 removing = deferred.realinode()
1599 return fmt.Errorf("cannot use Splice to attach a file at top level of %T: %w", removing, ErrInvalidOperation)
1603 _, err = dn.parent.Child(dn.fileinfo.name, func(inode) (inode, error) { return repl, nil })
1605 return fmt.Errorf("error replacing filenode: dn.parent.Child(): %w", err)
1612 func (dn *dirnode) setTreeFS(fs *collectionFileSystem) {
1614 for _, child := range dn.inodes {
1615 switch child := child.(type) {
1624 type segment interface {
1627 // Return a new segment with a subsection of the data from this
1628 // one. length<0 means length=Len()-off.
1629 Slice(off int, length int) segment
1632 type memSegment struct {
1634 // If flushing is not nil and not ready/closed, then a) buf is
1635 // being shared by a pruneMemSegments goroutine, and must be
1636 // copied on write; and b) the flushing channel will close
1637 // when the goroutine finishes, whether it succeeds or not.
1638 flushing <-chan struct{}
1641 func (me *memSegment) flushingUnfinished() bool {
1642 if me.flushing == nil {
1654 func (me *memSegment) Len() int {
1658 func (me *memSegment) Slice(off, length int) segment {
1660 length = len(me.buf) - off
1662 buf := make([]byte, length)
1663 copy(buf, me.buf[off:])
1664 return &memSegment{buf: buf}
1667 func (me *memSegment) Truncate(n int) {
1668 if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
1671 newsize = newsize << 2
1673 newbuf := make([]byte, n, newsize)
1674 copy(newbuf, me.buf)
1675 me.buf, me.flushing = newbuf, nil
1677 // reclaim existing capacity, and zero reclaimed part
1678 oldlen := len(me.buf)
1680 for i := oldlen; i < n; i++ {
1686 func (me *memSegment) WriteAt(p []byte, off int) {
1687 if off+len(p) > len(me.buf) {
1688 panic("overflowed segment")
1690 if me.flushing != nil {
1691 me.buf, me.flushing = append([]byte(nil), me.buf...), nil
1693 copy(me.buf[off:], p)
1696 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1697 if off > int64(me.Len()) {
1701 n = copy(p, me.buf[int(off):])
1708 type storedSegment struct {
1711 size int // size of stored block (also encoded in locator)
1712 offset int // position of segment within the stored block
1713 length int // bytes in this segment (offset + length <= size)
1716 func (se storedSegment) Len() int {
1720 func (se storedSegment) Slice(n, size int) segment {
1723 if size >= 0 && se.length > size {
1729 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1730 if off > int64(se.length) {
1733 maxlen := se.length - int(off)
1734 if len(p) > maxlen {
1736 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1742 return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1745 func canonicalName(name string) string {
1746 name = path.Clean("/" + name)
1747 if name == "/" || name == "./" {
1749 } else if strings.HasPrefix(name, "/") {
1755 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1757 func manifestUnescapeFunc(seq string) string {
1761 i, err := strconv.ParseUint(seq[1:], 8, 8)
1763 // Invalid escape sequence: can't unescape.
1766 return string([]byte{byte(i)})
1769 func manifestUnescape(s string) string {
1770 return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1773 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1775 func manifestEscapeFunc(seq string) string {
1776 return fmt.Sprintf("\\%03o", byte(seq[0]))
1779 func manifestEscape(s string) string {
1780 return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)