sdk/go/arvados/fs_collection.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: Apache-2.0
   4
   5 package arvados
   6
   7 import (
   8         "encoding/json"
   9         "fmt"
  10         "io"
  11         "os"
  12         "path"
  13         "regexp"
  14         "sort"
  15         "strconv"
  16         "strings"
  17         "sync"
  18         "time"
  19 )
  20
  21 var maxBlockSize = 1 << 26
  22
  23 var concurrentWriters = 4
  24
  25 // A CollectionFileSystem is a FileSystem that can be serialized as a
  26 // manifest and stored as a collection.
  27 type CollectionFileSystem interface {
  28         FileSystem
  29
  30         // Flush all file data to Keep and return a snapshot of the
  31         // filesystem suitable for saving as (Collection)ManifestText.
  32         // Prefix (normally ".") is a top level directory, effectively
  33         // prepended to all paths in the returned manifest.
  34         MarshalManifest(prefix string) (string, error)
  35
  36         // Total data bytes in all files.
  37         Size() int64
  38 }
  39
  40 type collectionFileSystem struct {
  41         fileSystem
  42         uuid string
  43 }
  44
  45 // FileSystem returns a CollectionFileSystem for the collection.
  46 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
  47         var modTime time.Time
  48         if c.ModifiedAt == nil {
  49                 modTime = time.Now()
  50         } else {
  51                 modTime = *c.ModifiedAt
  52         }
  53         fs := &collectionFileSystem{
  54                 uuid: c.UUID,
  55                 fileSystem: fileSystem{
  56                         fsBackend: keepBackend{apiClient: client, keepClient: kc},
  57                 },
  58         }
  59         root := &dirnode{
  60                 fs: fs,
  61                 treenode: treenode{
  62                         fileinfo: fileinfo{
  63                                 name:    ".",
  64                                 mode:    os.ModeDir | 0755,
  65                                 modTime: modTime,
  66                         },
  67                         inodes: make(map[string]inode),
  68                 },
  69         }
  70         root.SetParent(root, ".")
  71         if err := root.loadManifest(c.ManifestText); err != nil {
  72                 return nil, err
  73         }
  74         backdateTree(root, modTime)
  75         fs.root = root
  76         return fs, nil
  77 }
  78
  79 func backdateTree(n inode, modTime time.Time) {
  80         switch n := n.(type) {
  81         case *filenode:
  82                 n.fileinfo.modTime = modTime
  83         case *dirnode:
  84                 n.fileinfo.modTime = modTime
  85                 for _, n := range n.inodes {
  86                         backdateTree(n, modTime)
  87                 }
  88         }
  89 }
  90
  91 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
  92         if name == "" || name == "." || name == ".." {
  93                 return nil, ErrInvalidArgument
  94         }
  95         if perm.IsDir() {
  96                 return &dirnode{
  97                         fs: fs,
  98                         treenode: treenode{
  99                                 fileinfo: fileinfo{
 100                                         name:    name,
 101                                         mode:    perm | os.ModeDir,
 102                                         modTime: modTime,
 103                                 },
 104                                 inodes: make(map[string]inode),
 105                         },
 106                 }, nil
 107         } else {
 108                 return &filenode{
 109                         fs: fs,
 110                         fileinfo: fileinfo{
 111                                 name:    name,
 112                                 mode:    perm & ^os.ModeDir,
 113                                 modTime: modTime,
 114                         },
 115                 }, nil
 116         }
 117 }
 118
 119 func (fs *collectionFileSystem) Sync() error {
 120         if fs.uuid == "" {
 121                 return nil
 122         }
 123         txt, err := fs.MarshalManifest(".")
 124         if err != nil {
 125                 return fmt.Errorf("sync failed: %s", err)
 126         }
 127         coll := &Collection{
 128                 UUID:         fs.uuid,
 129                 ManifestText: txt,
 130         }
 131         err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
 132         if err != nil {
 133                 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
 134         }
 135         return nil
 136 }
 137
 138 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
 139         fs.fileSystem.root.Lock()
 140         defer fs.fileSystem.root.Unlock()
 141         return fs.fileSystem.root.(*dirnode).marshalManifest(prefix, newThrottle(concurrentWriters))
 142 }
 143
 144 func (fs *collectionFileSystem) Size() int64 {
 145         return fs.fileSystem.root.(*dirnode).TreeSize()
 146 }
 147
 148 // filenodePtr is an offset into a file that is (usually) efficient to
 149 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
 150 // then
 151 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
 152 // corresponds to file offset filenodePtr.off. Otherwise, it is
 153 // necessary to reexamine len(filenode.segments[0]) etc. to find the
 154 // correct segment and offset.
 155 type filenodePtr struct {
 156         off        int64
 157         segmentIdx int
 158         segmentOff int
 159         repacked   int64
 160 }
 161
 162 // seek returns a ptr that is consistent with both startPtr.off and
 163 // the current state of fn. The caller must already hold fn.RLock() or
 164 // fn.Lock().
 165 //
 166 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
 167 // EOF.
 168 //
 169 // After seeking:
 170 //
 171 //     ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
 172 //     ||
 173 //     filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
 174 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
 175         ptr = startPtr
 176         if ptr.off < 0 {
 177                 // meaningless anyway
 178                 return
 179         } else if ptr.off >= fn.fileinfo.size {
 180                 ptr.segmentIdx = len(fn.segments)
 181                 ptr.segmentOff = 0
 182                 ptr.repacked = fn.repacked
 183                 return
 184         } else if ptr.repacked == fn.repacked {
 185                 // segmentIdx and segmentOff accurately reflect
 186                 // ptr.off, but might have fallen off the end of a
 187                 // segment
 188                 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
 189                         ptr.segmentIdx++
 190                         ptr.segmentOff = 0
 191                 }
 192                 return
 193         }
 194         defer func() {
 195                 ptr.repacked = fn.repacked
 196         }()
 197         if ptr.off >= fn.fileinfo.size {
 198                 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
 199                 return
 200         }
 201         // Recompute segmentIdx and segmentOff.  We have already
 202         // established fn.fileinfo.size > ptr.off >= 0, so we don't
 203         // have to deal with edge cases here.
 204         var off int64
 205         for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
 206                 // This would panic (index out of range) if
 207                 // fn.fileinfo.size were larger than
 208                 // sum(fn.segments[i].Len()) -- but that can't happen
 209                 // because we have ensured fn.fileinfo.size is always
 210                 // accurate.
 211                 segLen := int64(fn.segments[ptr.segmentIdx].Len())
 212                 if off+segLen > ptr.off {
 213                         ptr.segmentOff = int(ptr.off - off)
 214                         break
 215                 }
 216                 off += segLen
 217         }
 218         return
 219 }
 220
 221 // filenode implements inode.
 222 type filenode struct {
 223         parent   inode
 224         fs       FileSystem
 225         fileinfo fileinfo
 226         segments []segment
 227         // number of times `segments` has changed in a
 228         // way that might invalidate a filenodePtr
 229         repacked int64
 230         memsize  int64 // bytes in memSegments
 231         sync.RWMutex
 232         nullnode
 233 }
 234
 235 // caller must have lock
 236 func (fn *filenode) appendSegment(e segment) {
 237         fn.segments = append(fn.segments, e)
 238         fn.fileinfo.size += int64(e.Len())
 239 }
 240
 241 func (fn *filenode) SetParent(p inode, name string) {
 242         fn.Lock()
 243         defer fn.Unlock()
 244         fn.parent = p
 245         fn.fileinfo.name = name
 246 }
 247
 248 func (fn *filenode) Parent() inode {
 249         fn.RLock()
 250         defer fn.RUnlock()
 251         return fn.parent
 252 }
 253
 254 func (fn *filenode) FS() FileSystem {
 255         return fn.fs
 256 }
 257
 258 // Read reads file data from a single segment, starting at startPtr,
 259 // into p. startPtr is assumed not to be up-to-date. Caller must have
 260 // RLock or Lock.
 261 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 262         ptr = fn.seek(startPtr)
 263         if ptr.off < 0 {
 264                 err = ErrNegativeOffset
 265                 return
 266         }
 267         if ptr.segmentIdx >= len(fn.segments) {
 268                 err = io.EOF
 269                 return
 270         }
 271         n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
 272         if n > 0 {
 273                 ptr.off += int64(n)
 274                 ptr.segmentOff += n
 275                 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
 276                         ptr.segmentIdx++
 277                         ptr.segmentOff = 0
 278                         if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
 279                                 err = nil
 280                         }
 281                 }
 282         }
 283         return
 284 }
 285
 286 func (fn *filenode) Size() int64 {
 287         fn.RLock()
 288         defer fn.RUnlock()
 289         return fn.fileinfo.Size()
 290 }
 291
 292 func (fn *filenode) FileInfo() os.FileInfo {
 293         fn.RLock()
 294         defer fn.RUnlock()
 295         return fn.fileinfo
 296 }
 297
 298 func (fn *filenode) Truncate(size int64) error {
 299         fn.Lock()
 300         defer fn.Unlock()
 301         return fn.truncate(size)
 302 }
 303
 304 func (fn *filenode) truncate(size int64) error {
 305         if size == fn.fileinfo.size {
 306                 return nil
 307         }
 308         fn.repacked++
 309         if size < fn.fileinfo.size {
 310                 ptr := fn.seek(filenodePtr{off: size})
 311                 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
 312                         if seg, ok := fn.segments[i].(*memSegment); ok {
 313                                 fn.memsize -= int64(seg.Len())
 314                         }
 315                 }
 316                 if ptr.segmentOff == 0 {
 317                         fn.segments = fn.segments[:ptr.segmentIdx]
 318                 } else {
 319                         fn.segments = fn.segments[:ptr.segmentIdx+1]
 320                         switch seg := fn.segments[ptr.segmentIdx].(type) {
 321                         case *memSegment:
 322                                 seg.Truncate(ptr.segmentOff)
 323                                 fn.memsize += int64(seg.Len())
 324                         default:
 325                                 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
 326                         }
 327                 }
 328                 fn.fileinfo.size = size
 329                 return nil
 330         }
 331         for size > fn.fileinfo.size {
 332                 grow := size - fn.fileinfo.size
 333                 var seg *memSegment
 334                 var ok bool
 335                 if len(fn.segments) == 0 {
 336                         seg = &memSegment{}
 337                         fn.segments = append(fn.segments, seg)
 338                 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
 339                         seg = &memSegment{}
 340                         fn.segments = append(fn.segments, seg)
 341                 }
 342                 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
 343                         grow = maxgrow
 344                 }
 345                 seg.Truncate(seg.Len() + int(grow))
 346                 fn.fileinfo.size += grow
 347                 fn.memsize += grow
 348         }
 349         return nil
 350 }
 351
 352 // Write writes data from p to the file, starting at startPtr,
 353 // extending the file size if necessary. Caller must have Lock.
 354 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 355         if startPtr.off > fn.fileinfo.size {
 356                 if err = fn.truncate(startPtr.off); err != nil {
 357                         return 0, startPtr, err
 358                 }
 359         }
 360         ptr = fn.seek(startPtr)
 361         if ptr.off < 0 {
 362                 err = ErrNegativeOffset
 363                 return
 364         }
 365         for len(p) > 0 && err == nil {
 366                 cando := p
 367                 if len(cando) > maxBlockSize {
 368                         cando = cando[:maxBlockSize]
 369                 }
 370                 // Rearrange/grow fn.segments (and shrink cando if
 371                 // needed) such that cando can be copied to
 372                 // fn.segments[ptr.segmentIdx] at offset
 373                 // ptr.segmentOff.
 374                 cur := ptr.segmentIdx
 375                 prev := ptr.segmentIdx - 1
 376                 var curWritable bool
 377                 if cur < len(fn.segments) {
 378                         _, curWritable = fn.segments[cur].(*memSegment)
 379                 }
 380                 var prevAppendable bool
 381                 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
 382                         _, prevAppendable = fn.segments[prev].(*memSegment)
 383                 }
 384                 if ptr.segmentOff > 0 && !curWritable {
 385                         // Split a non-writable block.
 386                         if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
 387                                 // Truncate cur, and insert a new
 388                                 // segment after it.
 389                                 cando = cando[:max]
 390                                 fn.segments = append(fn.segments, nil)
 391                                 copy(fn.segments[cur+1:], fn.segments[cur:])
 392                         } else {
 393                                 // Split cur into two copies, truncate
 394                                 // the one on the left, shift the one
 395                                 // on the right, and insert a new
 396                                 // segment between them.
 397                                 fn.segments = append(fn.segments, nil, nil)
 398                                 copy(fn.segments[cur+2:], fn.segments[cur:])
 399                                 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
 400                         }
 401                         cur++
 402                         prev++
 403                         seg := &memSegment{}
 404                         seg.Truncate(len(cando))
 405                         fn.memsize += int64(len(cando))
 406                         fn.segments[cur] = seg
 407                         fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
 408                         ptr.segmentIdx++
 409                         ptr.segmentOff = 0
 410                         fn.repacked++
 411                         ptr.repacked++
 412                 } else if curWritable {
 413                         if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
 414                                 cando = cando[:fit]
 415                         }
 416                 } else {
 417                         if prevAppendable {
 418                                 // Shrink cando if needed to fit in
 419                                 // prev segment.
 420                                 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
 421                                         cando = cando[:cangrow]
 422                                 }
 423                         }
 424
 425                         if cur == len(fn.segments) {
 426                                 // ptr is at EOF, filesize is changing.
 427                                 fn.fileinfo.size += int64(len(cando))
 428                         } else if el := fn.segments[cur].Len(); el <= len(cando) {
 429                                 // cando is long enough that we won't
 430                                 // need cur any more. shrink cando to
 431                                 // be exactly as long as cur
 432                                 // (otherwise we'd accidentally shift
 433                                 // the effective position of all
 434                                 // segments after cur).
 435                                 cando = cando[:el]
 436                                 copy(fn.segments[cur:], fn.segments[cur+1:])
 437                                 fn.segments = fn.segments[:len(fn.segments)-1]
 438                         } else {
 439                                 // shrink cur by the same #bytes we're growing prev
 440                                 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
 441                         }
 442
 443                         if prevAppendable {
 444                                 // Grow prev.
 445                                 ptr.segmentIdx--
 446                                 ptr.segmentOff = fn.segments[prev].Len()
 447                                 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
 448                                 fn.memsize += int64(len(cando))
 449                                 ptr.repacked++
 450                                 fn.repacked++
 451                         } else {
 452                                 // Insert a segment between prev and
 453                                 // cur, and advance prev/cur.
 454                                 fn.segments = append(fn.segments, nil)
 455                                 if cur < len(fn.segments) {
 456                                         copy(fn.segments[cur+1:], fn.segments[cur:])
 457                                         ptr.repacked++
 458                                         fn.repacked++
 459                                 } else {
 460                                         // appending a new segment does
 461                                         // not invalidate any ptrs
 462                                 }
 463                                 seg := &memSegment{}
 464                                 seg.Truncate(len(cando))
 465                                 fn.memsize += int64(len(cando))
 466                                 fn.segments[cur] = seg
 467                                 cur++
 468                                 prev++
 469                         }
 470                 }
 471
 472                 // Finally we can copy bytes from cando to the current segment.
 473                 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
 474                 n += len(cando)
 475                 p = p[len(cando):]
 476
 477                 ptr.off += int64(len(cando))
 478                 ptr.segmentOff += len(cando)
 479                 if ptr.segmentOff >= maxBlockSize {
 480                         fn.pruneMemSegments()
 481                 }
 482                 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
 483                         ptr.segmentOff = 0
 484                         ptr.segmentIdx++
 485                 }
 486
 487                 fn.fileinfo.modTime = time.Now()
 488         }
 489         return
 490 }
 491
 492 // Write some data out to disk to reduce memory use. Caller must have
 493 // write lock.
 494 func (fn *filenode) pruneMemSegments() {
 495         // TODO: async (don't hold Lock() while waiting for Keep)
 496         // TODO: share code with (*dirnode)sync()
 497         // TODO: pack/flush small blocks too, when fragmented
 498         for idx, seg := range fn.segments {
 499                 seg, ok := seg.(*memSegment)
 500                 if !ok || seg.Len() < maxBlockSize {
 501                         continue
 502                 }
 503                 locator, _, err := fn.FS().PutB(seg.buf)
 504                 if err != nil {
 505                         // TODO: stall (or return errors from)
 506                         // subsequent writes until flushing
 507                         // starts to succeed
 508                         continue
 509                 }
 510                 fn.memsize -= int64(seg.Len())
 511                 fn.segments[idx] = storedSegment{
 512                         kc:      fn.FS(),
 513                         locator: locator,
 514                         size:    seg.Len(),
 515                         offset:  0,
 516                         length:  seg.Len(),
 517                 }
 518         }
 519 }
 520
 521 type dirnode struct {
 522         fs *collectionFileSystem
 523         treenode
 524 }
 525
 526 func (dn *dirnode) FS() FileSystem {
 527         return dn.fs
 528 }
 529
 530 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
 531         if dn == dn.fs.rootnode() && name == ".arvados#collection" {
 532                 gn := &getternode{Getter: func() ([]byte, error) {
 533                         var coll Collection
 534                         var err error
 535                         coll.ManifestText, err = dn.fs.MarshalManifest(".")
 536                         if err != nil {
 537                                 return nil, err
 538                         }
 539                         data, err := json.Marshal(&coll)
 540                         if err == nil {
 541                                 data = append(data, '\n')
 542                         }
 543                         return data, err
 544                 }}
 545                 gn.SetParent(dn, name)
 546                 return gn, nil
 547         }
 548         return dn.treenode.Child(name, replace)
 549 }
 550
 551 // sync flushes in-memory data and remote block references (for the
 552 // children with the given names, which must be children of dn) to
 553 // local persistent storage. Caller must have write lock on dn and the
 554 // named children.
 555 func (dn *dirnode) sync(names []string, throttle *throttle) error {
 556         type shortBlock struct {
 557                 fn  *filenode
 558                 idx int
 559         }
 560         var pending []shortBlock
 561         var pendingLen int
 562
 563         var wg sync.WaitGroup
 564         errors := make(chan error, 1)
 565         flush := func(sbs []shortBlock) {
 566                 defer wg.Done()
 567                 if len(sbs) == 0 {
 568                         return
 569                 }
 570                 throttle.Acquire()
 571                 defer throttle.Release()
 572                 block := make([]byte, 0, maxBlockSize)
 573                 for _, sb := range sbs {
 574                         block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
 575                 }
 576                 locator, _, err := dn.fs.PutB(block)
 577                 if err != nil {
 578                         select {
 579                         case errors <- err:
 580                         default:
 581                         }
 582                 }
 583                 off := 0
 584                 for _, sb := range sbs {
 585                         data := sb.fn.segments[sb.idx].(*memSegment).buf
 586                         sb.fn.segments[sb.idx] = storedSegment{
 587                                 kc:      dn.fs,
 588                                 locator: locator,
 589                                 size:    len(block),
 590                                 offset:  off,
 591                                 length:  len(data),
 592                         }
 593                         off += len(data)
 594                         sb.fn.memsize -= int64(len(data))
 595                 }
 596         }
 597
 598         localLocator := map[string]string{}
 599         for _, name := range names {
 600                 fn, ok := dn.inodes[name].(*filenode)
 601                 if !ok {
 602                         continue
 603                 }
 604                 for idx, seg := range fn.segments {
 605                         switch seg := seg.(type) {
 606                         case storedSegment:
 607                                 loc, ok := localLocator[seg.locator]
 608                                 if !ok {
 609                                         var err error
 610                                         loc, err = dn.fs.LocalLocator(seg.locator)
 611                                         if err != nil {
 612                                                 return err
 613                                         }
 614                                         localLocator[seg.locator] = loc
 615                                 }
 616                                 seg.locator = loc
 617                                 fn.segments[idx] = seg
 618                         case *memSegment:
 619                                 if seg.Len() > maxBlockSize/2 {
 620                                         wg.Add(1)
 621                                         go flush([]shortBlock{{fn, idx}})
 622                                         continue
 623                                 }
 624                                 if pendingLen+seg.Len() > maxBlockSize {
 625                                         wg.Add(1)
 626                                         go flush(pending)
 627                                         pending = nil
 628                                         pendingLen = 0
 629                                 }
 630                                 pending = append(pending, shortBlock{fn, idx})
 631                                 pendingLen += seg.Len()
 632                         default:
 633                                 panic(fmt.Sprintf("can't sync segment type %T", seg))
 634                         }
 635                 }
 636         }
 637         wg.Add(1)
 638         flush(pending)
 639         wg.Wait()
 640         close(errors)
 641         return <-errors
 642 }
 643
 644 // caller must have write lock.
 645 func (dn *dirnode) marshalManifest(prefix string, throttle *throttle) (string, error) {
 646         var streamLen int64
 647         type filepart struct {
 648                 name   string
 649                 offset int64
 650                 length int64
 651         }
 652         var fileparts []filepart
 653         var subdirs string
 654         var blocks []string
 655
 656         if len(dn.inodes) == 0 {
 657                 if prefix == "." {
 658                         return "", nil
 659                 }
 660                 // Express the existence of an empty directory by
 661                 // adding an empty file named `\056`, which (unlike
 662                 // the more obvious spelling `.`) is accepted by the
 663                 // API's manifest validator.
 664                 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
 665         }
 666
 667         names := make([]string, 0, len(dn.inodes))
 668         for name := range dn.inodes {
 669                 names = append(names, name)
 670         }
 671         sort.Strings(names)
 672         for _, name := range names {
 673                 node := dn.inodes[name]
 674                 node.Lock()
 675                 defer node.Unlock()
 676         }
 677         if err := dn.sync(names, throttle); err != nil {
 678                 return "", err
 679         }
 680         for _, name := range names {
 681                 switch node := dn.inodes[name].(type) {
 682                 case *dirnode:
 683                         subdir, err := node.marshalManifest(prefix+"/"+name, throttle)
 684                         if err != nil {
 685                                 return "", err
 686                         }
 687                         subdirs = subdirs + subdir
 688                 case *filenode:
 689                         if len(node.segments) == 0 {
 690                                 fileparts = append(fileparts, filepart{name: name})
 691                                 break
 692                         }
 693                         for _, seg := range node.segments {
 694                                 switch seg := seg.(type) {
 695                                 case storedSegment:
 696                                         if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
 697                                                 streamLen -= int64(seg.size)
 698                                         } else {
 699                                                 blocks = append(blocks, seg.locator)
 700                                         }
 701                                         next := filepart{
 702                                                 name:   name,
 703                                                 offset: streamLen + int64(seg.offset),
 704                                                 length: int64(seg.length),
 705                                         }
 706                                         if prev := len(fileparts) - 1; prev >= 0 &&
 707                                                 fileparts[prev].name == name &&
 708                                                 fileparts[prev].offset+fileparts[prev].length == next.offset {
 709                                                 fileparts[prev].length += next.length
 710                                         } else {
 711                                                 fileparts = append(fileparts, next)
 712                                         }
 713                                         streamLen += int64(seg.size)
 714                                 default:
 715                                         // This can't happen: we
 716                                         // haven't unlocked since
 717                                         // calling sync().
 718                                         panic(fmt.Sprintf("can't marshal segment type %T", seg))
 719                                 }
 720                         }
 721                 default:
 722                         panic(fmt.Sprintf("can't marshal inode type %T", node))
 723                 }
 724         }
 725         var filetokens []string
 726         for _, s := range fileparts {
 727                 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
 728         }
 729         if len(filetokens) == 0 {
 730                 return subdirs, nil
 731         } else if len(blocks) == 0 {
 732                 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
 733         }
 734         return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
 735 }
 736
 737 func (dn *dirnode) loadManifest(txt string) error {
 738         var dirname string
 739         streams := strings.Split(txt, "\n")
 740         if streams[len(streams)-1] != "" {
 741                 return fmt.Errorf("line %d: no trailing newline", len(streams))
 742         }
 743         streams = streams[:len(streams)-1]
 744         segments := []storedSegment{}
 745         for i, stream := range streams {
 746                 lineno := i + 1
 747                 var anyFileTokens bool
 748                 var pos int64
 749                 var segIdx int
 750                 segments = segments[:0]
 751                 for i, token := range strings.Split(stream, " ") {
 752                         if i == 0 {
 753                                 dirname = manifestUnescape(token)
 754                                 continue
 755                         }
 756                         if !strings.Contains(token, ":") {
 757                                 if anyFileTokens {
 758                                         return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 759                                 }
 760                                 toks := strings.SplitN(token, "+", 3)
 761                                 if len(toks) < 2 {
 762                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 763                                 }
 764                                 length, err := strconv.ParseInt(toks[1], 10, 32)
 765                                 if err != nil || length < 0 {
 766                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 767                                 }
 768                                 segments = append(segments, storedSegment{
 769                                         locator: token,
 770                                         size:    int(length),
 771                                         offset:  0,
 772                                         length:  int(length),
 773                                 })
 774                                 continue
 775                         } else if len(segments) == 0 {
 776                                 return fmt.Errorf("line %d: bad locator %q", lineno, token)
 777                         }
 778
 779                         toks := strings.SplitN(token, ":", 3)
 780                         if len(toks) != 3 {
 781                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 782                         }
 783                         anyFileTokens = true
 784
 785                         offset, err := strconv.ParseInt(toks[0], 10, 64)
 786                         if err != nil || offset < 0 {
 787                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 788                         }
 789                         length, err := strconv.ParseInt(toks[1], 10, 64)
 790                         if err != nil || length < 0 {
 791                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 792                         }
 793                         name := dirname + "/" + manifestUnescape(toks[2])
 794                         fnode, err := dn.createFileAndParents(name)
 795                         if fnode == nil && err == nil && length == 0 {
 796                                 // Special case: an empty file used as
 797                                 // a marker to preserve an otherwise
 798                                 // empty directory in a manifest.
 799                                 continue
 800                         }
 801                         if err != nil || (fnode == nil && length != 0) {
 802                                 return fmt.Errorf("line %d: cannot use path %q with length %d: %s", lineno, name, length, err)
 803                         }
 804                         // Map the stream offset/range coordinates to
 805                         // block/offset/range coordinates and add
 806                         // corresponding storedSegments to the filenode
 807                         if pos > offset {
 808                                 // Can't continue where we left off.
 809                                 // TODO: binary search instead of
 810                                 // rewinding all the way (but this
 811                                 // situation might be rare anyway)
 812                                 segIdx, pos = 0, 0
 813                         }
 814                         for next := int64(0); segIdx < len(segments); segIdx++ {
 815                                 seg := segments[segIdx]
 816                                 next = pos + int64(seg.Len())
 817                                 if next <= offset || seg.Len() == 0 {
 818                                         pos = next
 819                                         continue
 820                                 }
 821                                 if pos >= offset+length {
 822                                         break
 823                                 }
 824                                 var blkOff int
 825                                 if pos < offset {
 826                                         blkOff = int(offset - pos)
 827                                 }
 828                                 blkLen := seg.Len() - blkOff
 829                                 if pos+int64(blkOff+blkLen) > offset+length {
 830                                         blkLen = int(offset + length - pos - int64(blkOff))
 831                                 }
 832                                 fnode.appendSegment(storedSegment{
 833                                         kc:      dn.fs,
 834                                         locator: seg.locator,
 835                                         size:    seg.size,
 836                                         offset:  blkOff,
 837                                         length:  blkLen,
 838                                 })
 839                                 if next > offset+length {
 840                                         break
 841                                 } else {
 842                                         pos = next
 843                                 }
 844                         }
 845                         if segIdx == len(segments) && pos < offset+length {
 846                                 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
 847                         }
 848                 }
 849                 if !anyFileTokens {
 850                         return fmt.Errorf("line %d: no file segments", lineno)
 851                 } else if len(segments) == 0 {
 852                         return fmt.Errorf("line %d: no locators", lineno)
 853                 } else if dirname == "" {
 854                         return fmt.Errorf("line %d: no stream name", lineno)
 855                 }
 856         }
 857         return nil
 858 }
 859
 860 // only safe to call from loadManifest -- no locking.
 861 //
 862 // If path is a "parent directory exists" marker (the last path
 863 // component is "."), the returned values are both nil.
 864 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
 865         var node inode = dn
 866         names := strings.Split(path, "/")
 867         basename := names[len(names)-1]
 868         for _, name := range names[:len(names)-1] {
 869                 switch name {
 870                 case "", ".":
 871                         continue
 872                 case "..":
 873                         if node == dn {
 874                                 // can't be sure parent will be a *dirnode
 875                                 return nil, ErrInvalidArgument
 876                         }
 877                         node = node.Parent()
 878                         continue
 879                 }
 880                 node, err = node.Child(name, func(child inode) (inode, error) {
 881                         if child == nil {
 882                                 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
 883                                 if err != nil {
 884                                         return nil, err
 885                                 }
 886                                 child.SetParent(node, name)
 887                                 return child, nil
 888                         } else if !child.IsDir() {
 889                                 return child, ErrFileExists
 890                         } else {
 891                                 return child, nil
 892                         }
 893                 })
 894                 if err != nil {
 895                         return
 896                 }
 897         }
 898         if basename == "." {
 899                 return
 900         } else if !permittedName(basename) {
 901                 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
 902                 return
 903         }
 904         _, err = node.Child(basename, func(child inode) (inode, error) {
 905                 switch child := child.(type) {
 906                 case nil:
 907                         child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
 908                         if err != nil {
 909                                 return nil, err
 910                         }
 911                         child.SetParent(node, basename)
 912                         fn = child.(*filenode)
 913                         return child, nil
 914                 case *filenode:
 915                         fn = child
 916                         return child, nil
 917                 case *dirnode:
 918                         return child, ErrIsDirectory
 919                 default:
 920                         return child, ErrInvalidArgument
 921                 }
 922         })
 923         return
 924 }
 925
 926 func (dn *dirnode) TreeSize() (bytes int64) {
 927         dn.RLock()
 928         defer dn.RUnlock()
 929         for _, i := range dn.inodes {
 930                 switch i := i.(type) {
 931                 case *filenode:
 932                         bytes += i.Size()
 933                 case *dirnode:
 934                         bytes += i.TreeSize()
 935                 }
 936         }
 937         return
 938 }
 939
 940 type segment interface {
 941         io.ReaderAt
 942         Len() int
 943         // Return a new segment with a subsection of the data from this
 944         // one. length<0 means length=Len()-off.
 945         Slice(off int, length int) segment
 946 }
 947
 948 type memSegment struct {
 949         buf []byte
 950 }
 951
 952 func (me *memSegment) Len() int {
 953         return len(me.buf)
 954 }
 955
 956 func (me *memSegment) Slice(off, length int) segment {
 957         if length < 0 {
 958                 length = len(me.buf) - off
 959         }
 960         buf := make([]byte, length)
 961         copy(buf, me.buf[off:])
 962         return &memSegment{buf: buf}
 963 }
 964
 965 func (me *memSegment) Truncate(n int) {
 966         if n > cap(me.buf) {
 967                 newsize := 1024
 968                 for newsize < n {
 969                         newsize = newsize << 2
 970                 }
 971                 newbuf := make([]byte, n, newsize)
 972                 copy(newbuf, me.buf)
 973                 me.buf = newbuf
 974         } else {
 975                 // Zero unused part when shrinking, in case we grow
 976                 // and start using it again later.
 977                 for i := n; i < len(me.buf); i++ {
 978                         me.buf[i] = 0
 979                 }
 980         }
 981         me.buf = me.buf[:n]
 982 }
 983
 984 func (me *memSegment) WriteAt(p []byte, off int) {
 985         if off+len(p) > len(me.buf) {
 986                 panic("overflowed segment")
 987         }
 988         copy(me.buf[off:], p)
 989 }
 990
 991 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
 992         if off > int64(me.Len()) {
 993                 err = io.EOF
 994                 return
 995         }
 996         n = copy(p, me.buf[int(off):])
 997         if n < len(p) {
 998                 err = io.EOF
 999         }
1000         return
1001 }
1002
1003 type storedSegment struct {
1004         kc      fsBackend
1005         locator string
1006         size    int // size of stored block (also encoded in locator)
1007         offset  int // position of segment within the stored block
1008         length  int // bytes in this segment (offset + length <= size)
1009 }
1010
1011 func (se storedSegment) Len() int {
1012         return se.length
1013 }
1014
1015 func (se storedSegment) Slice(n, size int) segment {
1016         se.offset += n
1017         se.length -= n
1018         if size >= 0 && se.length > size {
1019                 se.length = size
1020         }
1021         return se
1022 }
1023
1024 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1025         if off > int64(se.length) {
1026                 return 0, io.EOF
1027         }
1028         maxlen := se.length - int(off)
1029         if len(p) > maxlen {
1030                 p = p[:maxlen]
1031                 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1032                 if err == nil {
1033                         err = io.EOF
1034                 }
1035                 return
1036         }
1037         return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1038 }
1039
1040 func canonicalName(name string) string {
1041         name = path.Clean("/" + name)
1042         if name == "/" || name == "./" {
1043                 name = "."
1044         } else if strings.HasPrefix(name, "/") {
1045                 name = "." + name
1046         }
1047         return name
1048 }
1049
1050 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1051
1052 func manifestUnescapeFunc(seq string) string {
1053         if seq == `\\` {
1054                 return `\`
1055         }
1056         i, err := strconv.ParseUint(seq[1:], 8, 8)
1057         if err != nil {
1058                 // Invalid escape sequence: can't unescape.
1059                 return seq
1060         }
1061         return string([]byte{byte(i)})
1062 }
1063
1064 func manifestUnescape(s string) string {
1065         return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1066 }
1067
1068 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1069
1070 func manifestEscapeFunc(seq string) string {
1071         return fmt.Sprintf("\\%03o", byte(seq[0]))
1072 }
1073
1074 func manifestEscape(s string) string {
1075         return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)
1076 }