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         "context"
   9         "encoding/json"
  10         "fmt"
  11         "io"
  12         "os"
  13         "path"
  14         "regexp"
  15         "sort"
  16         "strconv"
  17         "strings"
  18         "sync"
  19         "time"
  20 )
  21
  22 var maxBlockSize = 1 << 26
  23
  24 var concurrentWriters = 4
  25
  26 // A CollectionFileSystem is a FileSystem that can be serialized as a
  27 // manifest and stored as a collection.
  28 type CollectionFileSystem interface {
  29         FileSystem
  30
  31         // Flush all file data to Keep and return a snapshot of the
  32         // filesystem suitable for saving as (Collection)ManifestText.
  33         // Prefix (normally ".") is a top level directory, effectively
  34         // prepended to all paths in the returned manifest.
  35         MarshalManifest(prefix string) (string, error)
  36
  37         // Total data bytes in all files.
  38         Size() int64
  39 }
  40
  41 type collectionFileSystem struct {
  42         fileSystem
  43         uuid string
  44 }
  45
  46 // FileSystem returns a CollectionFileSystem for the collection.
  47 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
  48         var modTime time.Time
  49         if c.ModifiedAt == nil {
  50                 modTime = time.Now()
  51         } else {
  52                 modTime = *c.ModifiedAt
  53         }
  54         fs := &collectionFileSystem{
  55                 uuid: c.UUID,
  56                 fileSystem: fileSystem{
  57                         fsBackend: keepBackend{apiClient: client, keepClient: kc},
  58                 },
  59         }
  60         root := &dirnode{
  61                 fs: fs,
  62                 treenode: treenode{
  63                         fileinfo: fileinfo{
  64                                 name:    ".",
  65                                 mode:    os.ModeDir | 0755,
  66                                 modTime: modTime,
  67                         },
  68                         inodes: make(map[string]inode),
  69                 },
  70         }
  71         root.SetParent(root, ".")
  72         if err := root.loadManifest(c.ManifestText); err != nil {
  73                 return nil, err
  74         }
  75         backdateTree(root, modTime)
  76         fs.root = root
  77         return fs, nil
  78 }
  79
  80 func backdateTree(n inode, modTime time.Time) {
  81         switch n := n.(type) {
  82         case *filenode:
  83                 n.fileinfo.modTime = modTime
  84         case *dirnode:
  85                 n.fileinfo.modTime = modTime
  86                 for _, n := range n.inodes {
  87                         backdateTree(n, modTime)
  88                 }
  89         }
  90 }
  91
  92 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
  93         if name == "" || name == "." || name == ".." {
  94                 return nil, ErrInvalidArgument
  95         }
  96         if perm.IsDir() {
  97                 return &dirnode{
  98                         fs: fs,
  99                         treenode: treenode{
 100                                 fileinfo: fileinfo{
 101                                         name:    name,
 102                                         mode:    perm | os.ModeDir,
 103                                         modTime: modTime,
 104                                 },
 105                                 inodes: make(map[string]inode),
 106                         },
 107                 }, nil
 108         } else {
 109                 return &filenode{
 110                         fs: fs,
 111                         fileinfo: fileinfo{
 112                                 name:    name,
 113                                 mode:    perm & ^os.ModeDir,
 114                                 modTime: modTime,
 115                         },
 116                 }, nil
 117         }
 118 }
 119
 120 func (fs *collectionFileSystem) Sync() error {
 121         if fs.uuid == "" {
 122                 return nil
 123         }
 124         txt, err := fs.MarshalManifest(".")
 125         if err != nil {
 126                 return fmt.Errorf("sync failed: %s", err)
 127         }
 128         coll := &Collection{
 129                 UUID:         fs.uuid,
 130                 ManifestText: txt,
 131         }
 132         err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
 133         if err != nil {
 134                 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
 135         }
 136         return nil
 137 }
 138
 139 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
 140         fs.fileSystem.root.Lock()
 141         defer fs.fileSystem.root.Unlock()
 142         return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix, newThrottle(concurrentWriters))
 143 }
 144
 145 func (fs *collectionFileSystem) Size() int64 {
 146         return fs.fileSystem.root.(*dirnode).TreeSize()
 147 }
 148
 149 // filenodePtr is an offset into a file that is (usually) efficient to
 150 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
 151 // then
 152 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
 153 // corresponds to file offset filenodePtr.off. Otherwise, it is
 154 // necessary to reexamine len(filenode.segments[0]) etc. to find the
 155 // correct segment and offset.
 156 type filenodePtr struct {
 157         off        int64
 158         segmentIdx int
 159         segmentOff int
 160         repacked   int64
 161 }
 162
 163 // seek returns a ptr that is consistent with both startPtr.off and
 164 // the current state of fn. The caller must already hold fn.RLock() or
 165 // fn.Lock().
 166 //
 167 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
 168 // EOF.
 169 //
 170 // After seeking:
 171 //
 172 //     ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
 173 //     ||
 174 //     filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
 175 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
 176         ptr = startPtr
 177         if ptr.off < 0 {
 178                 // meaningless anyway
 179                 return
 180         } else if ptr.off >= fn.fileinfo.size {
 181                 ptr.segmentIdx = len(fn.segments)
 182                 ptr.segmentOff = 0
 183                 ptr.repacked = fn.repacked
 184                 return
 185         } else if ptr.repacked == fn.repacked {
 186                 // segmentIdx and segmentOff accurately reflect
 187                 // ptr.off, but might have fallen off the end of a
 188                 // segment
 189                 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
 190                         ptr.segmentIdx++
 191                         ptr.segmentOff = 0
 192                 }
 193                 return
 194         }
 195         defer func() {
 196                 ptr.repacked = fn.repacked
 197         }()
 198         if ptr.off >= fn.fileinfo.size {
 199                 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
 200                 return
 201         }
 202         // Recompute segmentIdx and segmentOff.  We have already
 203         // established fn.fileinfo.size > ptr.off >= 0, so we don't
 204         // have to deal with edge cases here.
 205         var off int64
 206         for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
 207                 // This would panic (index out of range) if
 208                 // fn.fileinfo.size were larger than
 209                 // sum(fn.segments[i].Len()) -- but that can't happen
 210                 // because we have ensured fn.fileinfo.size is always
 211                 // accurate.
 212                 segLen := int64(fn.segments[ptr.segmentIdx].Len())
 213                 if off+segLen > ptr.off {
 214                         ptr.segmentOff = int(ptr.off - off)
 215                         break
 216                 }
 217                 off += segLen
 218         }
 219         return
 220 }
 221
 222 // filenode implements inode.
 223 type filenode struct {
 224         parent   inode
 225         fs       FileSystem
 226         fileinfo fileinfo
 227         segments []segment
 228         // number of times `segments` has changed in a
 229         // way that might invalidate a filenodePtr
 230         repacked int64
 231         memsize  int64 // bytes in memSegments
 232         sync.RWMutex
 233         nullnode
 234 }
 235
 236 // caller must have lock
 237 func (fn *filenode) appendSegment(e segment) {
 238         fn.segments = append(fn.segments, e)
 239         fn.fileinfo.size += int64(e.Len())
 240 }
 241
 242 func (fn *filenode) SetParent(p inode, name string) {
 243         fn.Lock()
 244         defer fn.Unlock()
 245         fn.parent = p
 246         fn.fileinfo.name = name
 247 }
 248
 249 func (fn *filenode) Parent() inode {
 250         fn.RLock()
 251         defer fn.RUnlock()
 252         return fn.parent
 253 }
 254
 255 func (fn *filenode) FS() FileSystem {
 256         return fn.fs
 257 }
 258
 259 // Read reads file data from a single segment, starting at startPtr,
 260 // into p. startPtr is assumed not to be up-to-date. Caller must have
 261 // RLock or Lock.
 262 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 263         ptr = fn.seek(startPtr)
 264         if ptr.off < 0 {
 265                 err = ErrNegativeOffset
 266                 return
 267         }
 268         if ptr.segmentIdx >= len(fn.segments) {
 269                 err = io.EOF
 270                 return
 271         }
 272         n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
 273         if n > 0 {
 274                 ptr.off += int64(n)
 275                 ptr.segmentOff += n
 276                 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
 277                         ptr.segmentIdx++
 278                         ptr.segmentOff = 0
 279                         if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
 280                                 err = nil
 281                         }
 282                 }
 283         }
 284         return
 285 }
 286
 287 func (fn *filenode) Size() int64 {
 288         fn.RLock()
 289         defer fn.RUnlock()
 290         return fn.fileinfo.Size()
 291 }
 292
 293 func (fn *filenode) FileInfo() os.FileInfo {
 294         fn.RLock()
 295         defer fn.RUnlock()
 296         return fn.fileinfo
 297 }
 298
 299 func (fn *filenode) Truncate(size int64) error {
 300         fn.Lock()
 301         defer fn.Unlock()
 302         return fn.truncate(size)
 303 }
 304
 305 func (fn *filenode) truncate(size int64) error {
 306         if size == fn.fileinfo.size {
 307                 return nil
 308         }
 309         fn.repacked++
 310         if size < fn.fileinfo.size {
 311                 ptr := fn.seek(filenodePtr{off: size})
 312                 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
 313                         if seg, ok := fn.segments[i].(*memSegment); ok {
 314                                 fn.memsize -= int64(seg.Len())
 315                         }
 316                 }
 317                 if ptr.segmentOff == 0 {
 318                         fn.segments = fn.segments[:ptr.segmentIdx]
 319                 } else {
 320                         fn.segments = fn.segments[:ptr.segmentIdx+1]
 321                         switch seg := fn.segments[ptr.segmentIdx].(type) {
 322                         case *memSegment:
 323                                 seg.Truncate(ptr.segmentOff)
 324                                 fn.memsize += int64(seg.Len())
 325                         default:
 326                                 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
 327                         }
 328                 }
 329                 fn.fileinfo.size = size
 330                 return nil
 331         }
 332         for size > fn.fileinfo.size {
 333                 grow := size - fn.fileinfo.size
 334                 var seg *memSegment
 335                 var ok bool
 336                 if len(fn.segments) == 0 {
 337                         seg = &memSegment{}
 338                         fn.segments = append(fn.segments, seg)
 339                 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
 340                         seg = &memSegment{}
 341                         fn.segments = append(fn.segments, seg)
 342                 }
 343                 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
 344                         grow = maxgrow
 345                 }
 346                 seg.Truncate(seg.Len() + int(grow))
 347                 fn.fileinfo.size += grow
 348                 fn.memsize += grow
 349         }
 350         return nil
 351 }
 352
 353 // Write writes data from p to the file, starting at startPtr,
 354 // extending the file size if necessary. Caller must have Lock.
 355 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 356         if startPtr.off > fn.fileinfo.size {
 357                 if err = fn.truncate(startPtr.off); err != nil {
 358                         return 0, startPtr, err
 359                 }
 360         }
 361         ptr = fn.seek(startPtr)
 362         if ptr.off < 0 {
 363                 err = ErrNegativeOffset
 364                 return
 365         }
 366         for len(p) > 0 && err == nil {
 367                 cando := p
 368                 if len(cando) > maxBlockSize {
 369                         cando = cando[:maxBlockSize]
 370                 }
 371                 // Rearrange/grow fn.segments (and shrink cando if
 372                 // needed) such that cando can be copied to
 373                 // fn.segments[ptr.segmentIdx] at offset
 374                 // ptr.segmentOff.
 375                 cur := ptr.segmentIdx
 376                 prev := ptr.segmentIdx - 1
 377                 var curWritable bool
 378                 if cur < len(fn.segments) {
 379                         _, curWritable = fn.segments[cur].(*memSegment)
 380                 }
 381                 var prevAppendable bool
 382                 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
 383                         _, prevAppendable = fn.segments[prev].(*memSegment)
 384                 }
 385                 if ptr.segmentOff > 0 && !curWritable {
 386                         // Split a non-writable block.
 387                         if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
 388                                 // Truncate cur, and insert a new
 389                                 // segment after it.
 390                                 cando = cando[:max]
 391                                 fn.segments = append(fn.segments, nil)
 392                                 copy(fn.segments[cur+1:], fn.segments[cur:])
 393                         } else {
 394                                 // Split cur into two copies, truncate
 395                                 // the one on the left, shift the one
 396                                 // on the right, and insert a new
 397                                 // segment between them.
 398                                 fn.segments = append(fn.segments, nil, nil)
 399                                 copy(fn.segments[cur+2:], fn.segments[cur:])
 400                                 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
 401                         }
 402                         cur++
 403                         prev++
 404                         seg := &memSegment{}
 405                         seg.Truncate(len(cando))
 406                         fn.memsize += int64(len(cando))
 407                         fn.segments[cur] = seg
 408                         fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
 409                         ptr.segmentIdx++
 410                         ptr.segmentOff = 0
 411                         fn.repacked++
 412                         ptr.repacked++
 413                 } else if curWritable {
 414                         if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
 415                                 cando = cando[:fit]
 416                         }
 417                 } else {
 418                         if prevAppendable {
 419                                 // Shrink cando if needed to fit in
 420                                 // prev segment.
 421                                 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
 422                                         cando = cando[:cangrow]
 423                                 }
 424                         }
 425
 426                         if cur == len(fn.segments) {
 427                                 // ptr is at EOF, filesize is changing.
 428                                 fn.fileinfo.size += int64(len(cando))
 429                         } else if el := fn.segments[cur].Len(); el <= len(cando) {
 430                                 // cando is long enough that we won't
 431                                 // need cur any more. shrink cando to
 432                                 // be exactly as long as cur
 433                                 // (otherwise we'd accidentally shift
 434                                 // the effective position of all
 435                                 // segments after cur).
 436                                 cando = cando[:el]
 437                                 copy(fn.segments[cur:], fn.segments[cur+1:])
 438                                 fn.segments = fn.segments[:len(fn.segments)-1]
 439                         } else {
 440                                 // shrink cur by the same #bytes we're growing prev
 441                                 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
 442                         }
 443
 444                         if prevAppendable {
 445                                 // Grow prev.
 446                                 ptr.segmentIdx--
 447                                 ptr.segmentOff = fn.segments[prev].Len()
 448                                 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
 449                                 fn.memsize += int64(len(cando))
 450                                 ptr.repacked++
 451                                 fn.repacked++
 452                         } else {
 453                                 // Insert a segment between prev and
 454                                 // cur, and advance prev/cur.
 455                                 fn.segments = append(fn.segments, nil)
 456                                 if cur < len(fn.segments) {
 457                                         copy(fn.segments[cur+1:], fn.segments[cur:])
 458                                         ptr.repacked++
 459                                         fn.repacked++
 460                                 } else {
 461                                         // appending a new segment does
 462                                         // not invalidate any ptrs
 463                                 }
 464                                 seg := &memSegment{}
 465                                 seg.Truncate(len(cando))
 466                                 fn.memsize += int64(len(cando))
 467                                 fn.segments[cur] = seg
 468                                 cur++
 469                                 prev++
 470                         }
 471                 }
 472
 473                 // Finally we can copy bytes from cando to the current segment.
 474                 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
 475                 n += len(cando)
 476                 p = p[len(cando):]
 477
 478                 ptr.off += int64(len(cando))
 479                 ptr.segmentOff += len(cando)
 480                 if ptr.segmentOff >= maxBlockSize {
 481                         fn.pruneMemSegments()
 482                 }
 483                 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
 484                         ptr.segmentOff = 0
 485                         ptr.segmentIdx++
 486                 }
 487
 488                 fn.fileinfo.modTime = time.Now()
 489         }
 490         return
 491 }
 492
 493 // Write some data out to disk to reduce memory use. Caller must have
 494 // write lock.
 495 func (fn *filenode) pruneMemSegments() {
 496         // TODO: async (don't hold Lock() while waiting for Keep)
 497         // TODO: share code with (*dirnode)sync()
 498         // TODO: pack/flush small blocks too, when fragmented
 499         for idx, seg := range fn.segments {
 500                 seg, ok := seg.(*memSegment)
 501                 if !ok || seg.Len() < maxBlockSize {
 502                         continue
 503                 }
 504                 locator, _, err := fn.FS().PutB(seg.buf)
 505                 if err != nil {
 506                         // TODO: stall (or return errors from)
 507                         // subsequent writes until flushing
 508                         // starts to succeed
 509                         continue
 510                 }
 511                 fn.memsize -= int64(seg.Len())
 512                 fn.segments[idx] = storedSegment{
 513                         kc:      fn.FS(),
 514                         locator: locator,
 515                         size:    seg.Len(),
 516                         offset:  0,
 517                         length:  seg.Len(),
 518                 }
 519         }
 520 }
 521
 522 type dirnode struct {
 523         fs *collectionFileSystem
 524         treenode
 525 }
 526
 527 func (dn *dirnode) FS() FileSystem {
 528         return dn.fs
 529 }
 530
 531 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
 532         if dn == dn.fs.rootnode() && name == ".arvados#collection" {
 533                 gn := &getternode{Getter: func() ([]byte, error) {
 534                         var coll Collection
 535                         var err error
 536                         coll.ManifestText, err = dn.fs.MarshalManifest(".")
 537                         if err != nil {
 538                                 return nil, err
 539                         }
 540                         data, err := json.Marshal(&coll)
 541                         if err == nil {
 542                                 data = append(data, '\n')
 543                         }
 544                         return data, err
 545                 }}
 546                 gn.SetParent(dn, name)
 547                 return gn, nil
 548         }
 549         return dn.treenode.Child(name, replace)
 550 }
 551
 552 // sync flushes in-memory data and remote block references (for the
 553 // children with the given names, which must be children of dn) to
 554 // local persistent storage. Caller must have write lock on dn and the
 555 // named children.
 556 func (dn *dirnode) sync(ctx context.Context, names []string, throttle *throttle) error {
 557         cg := newContextGroup(ctx)
 558         defer cg.Cancel()
 559
 560         type shortBlock struct {
 561                 fn  *filenode
 562                 idx int
 563         }
 564
 565         flush := func(sbs []shortBlock) error {
 566                 if len(sbs) == 0 {
 567                         return nil
 568                 }
 569                 throttle.Acquire()
 570                 defer throttle.Release()
 571                 if err := cg.Context().Err(); err != nil {
 572                         return err
 573                 }
 574                 block := make([]byte, 0, maxBlockSize)
 575                 for _, sb := range sbs {
 576                         block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
 577                 }
 578                 locator, _, err := dn.fs.PutB(block)
 579                 if err != nil {
 580                         return err
 581                 }
 582                 off := 0
 583                 for _, sb := range sbs {
 584                         data := sb.fn.segments[sb.idx].(*memSegment).buf
 585                         sb.fn.segments[sb.idx] = storedSegment{
 586                                 kc:      dn.fs,
 587                                 locator: locator,
 588                                 size:    len(block),
 589                                 offset:  off,
 590                                 length:  len(data),
 591                         }
 592                         off += len(data)
 593                         sb.fn.memsize -= int64(len(data))
 594                 }
 595                 return nil
 596         }
 597
 598         goFlush := func(sbs []shortBlock) {
 599                 cg.Go(func() error {
 600                         return flush(sbs)
 601                 })
 602         }
 603
 604         var pending []shortBlock
 605         var pendingLen int
 606         localLocator := map[string]string{}
 607         for _, name := range names {
 608                 fn, ok := dn.inodes[name].(*filenode)
 609                 if !ok {
 610                         continue
 611                 }
 612                 for idx, seg := range fn.segments {
 613                         switch seg := seg.(type) {
 614                         case storedSegment:
 615                                 loc, ok := localLocator[seg.locator]
 616                                 if !ok {
 617                                         var err error
 618                                         loc, err = dn.fs.LocalLocator(seg.locator)
 619                                         if err != nil {
 620                                                 return err
 621                                         }
 622                                         localLocator[seg.locator] = loc
 623                                 }
 624                                 seg.locator = loc
 625                                 fn.segments[idx] = seg
 626                         case *memSegment:
 627                                 if seg.Len() > maxBlockSize/2 {
 628                                         goFlush([]shortBlock{{fn, idx}})
 629                                         continue
 630                                 }
 631                                 if pendingLen+seg.Len() > maxBlockSize {
 632                                         goFlush(pending)
 633                                         pending = nil
 634                                         pendingLen = 0
 635                                 }
 636                                 pending = append(pending, shortBlock{fn, idx})
 637                                 pendingLen += seg.Len()
 638                         default:
 639                                 panic(fmt.Sprintf("can't sync segment type %T", seg))
 640                         }
 641                 }
 642         }
 643         goFlush(pending)
 644         return cg.Wait()
 645 }
 646
 647 // caller must have write lock.
 648 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string, throttle *throttle) (string, error) {
 649         cg := newContextGroup(ctx)
 650         defer cg.Cancel()
 651
 652         if len(dn.inodes) == 0 {
 653                 if prefix == "." {
 654                         return "", nil
 655                 }
 656                 // Express the existence of an empty directory by
 657                 // adding an empty file named `\056`, which (unlike
 658                 // the more obvious spelling `.`) is accepted by the
 659                 // API's manifest validator.
 660                 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
 661         }
 662
 663         names := make([]string, 0, len(dn.inodes))
 664         for name := range dn.inodes {
 665                 names = append(names, name)
 666         }
 667         sort.Strings(names)
 668         var dirnames []string
 669         var filenames []string
 670         for _, name := range names {
 671                 node := dn.inodes[name]
 672                 node.Lock()
 673                 defer node.Unlock()
 674                 switch node := node.(type) {
 675                 case *dirnode:
 676                         dirnames = append(dirnames, name)
 677                 case *filenode:
 678                         filenames = append(filenames, name)
 679                 default:
 680                         panic(fmt.Sprintf("can't marshal inode type %T", node))
 681                 }
 682         }
 683
 684         subdirs := make([]string, len(dirnames))
 685         rootdir := ""
 686         for i, name := range dirnames {
 687                 i, name := i, name
 688                 cg.Go(func() error {
 689                         txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name, throttle)
 690                         subdirs[i] = txt
 691                         return err
 692                 })
 693         }
 694
 695         cg.Go(func() error {
 696                 var streamLen int64
 697                 type filepart struct {
 698                         name   string
 699                         offset int64
 700                         length int64
 701                 }
 702
 703                 var fileparts []filepart
 704                 var blocks []string
 705                 if err := dn.sync(cg.Context(), names, throttle); err != nil {
 706                         return err
 707                 }
 708                 for _, name := range filenames {
 709                         node := dn.inodes[name].(*filenode)
 710                         if len(node.segments) == 0 {
 711                                 fileparts = append(fileparts, filepart{name: name})
 712                                 continue
 713                         }
 714                         for _, seg := range node.segments {
 715                                 switch seg := seg.(type) {
 716                                 case storedSegment:
 717                                         if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
 718                                                 streamLen -= int64(seg.size)
 719                                         } else {
 720                                                 blocks = append(blocks, seg.locator)
 721                                         }
 722                                         next := filepart{
 723                                                 name:   name,
 724                                                 offset: streamLen + int64(seg.offset),
 725                                                 length: int64(seg.length),
 726                                         }
 727                                         if prev := len(fileparts) - 1; prev >= 0 &&
 728                                                 fileparts[prev].name == name &&
 729                                                 fileparts[prev].offset+fileparts[prev].length == next.offset {
 730                                                 fileparts[prev].length += next.length
 731                                         } else {
 732                                                 fileparts = append(fileparts, next)
 733                                         }
 734                                         streamLen += int64(seg.size)
 735                                 default:
 736                                         // This can't happen: we
 737                                         // haven't unlocked since
 738                                         // calling sync().
 739                                         panic(fmt.Sprintf("can't marshal segment type %T", seg))
 740                                 }
 741                         }
 742                 }
 743                 var filetokens []string
 744                 for _, s := range fileparts {
 745                         filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
 746                 }
 747                 if len(filetokens) == 0 {
 748                         return nil
 749                 } else if len(blocks) == 0 {
 750                         blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
 751                 }
 752                 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
 753                 return nil
 754         })
 755         err := cg.Wait()
 756         return rootdir + strings.Join(subdirs, ""), err
 757 }
 758
 759 func (dn *dirnode) loadManifest(txt string) error {
 760         var dirname string
 761         streams := strings.Split(txt, "\n")
 762         if streams[len(streams)-1] != "" {
 763                 return fmt.Errorf("line %d: no trailing newline", len(streams))
 764         }
 765         streams = streams[:len(streams)-1]
 766         segments := []storedSegment{}
 767         for i, stream := range streams {
 768                 lineno := i + 1
 769                 var anyFileTokens bool
 770                 var pos int64
 771                 var segIdx int
 772                 segments = segments[:0]
 773                 for i, token := range strings.Split(stream, " ") {
 774                         if i == 0 {
 775                                 dirname = manifestUnescape(token)
 776                                 continue
 777                         }
 778                         if !strings.Contains(token, ":") {
 779                                 if anyFileTokens {
 780                                         return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 781                                 }
 782                                 toks := strings.SplitN(token, "+", 3)
 783                                 if len(toks) < 2 {
 784                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 785                                 }
 786                                 length, err := strconv.ParseInt(toks[1], 10, 32)
 787                                 if err != nil || length < 0 {
 788                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 789                                 }
 790                                 segments = append(segments, storedSegment{
 791                                         locator: token,
 792                                         size:    int(length),
 793                                         offset:  0,
 794                                         length:  int(length),
 795                                 })
 796                                 continue
 797                         } else if len(segments) == 0 {
 798                                 return fmt.Errorf("line %d: bad locator %q", lineno, token)
 799                         }
 800
 801                         toks := strings.SplitN(token, ":", 3)
 802                         if len(toks) != 3 {
 803                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 804                         }
 805                         anyFileTokens = true
 806
 807                         offset, err := strconv.ParseInt(toks[0], 10, 64)
 808                         if err != nil || offset < 0 {
 809                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 810                         }
 811                         length, err := strconv.ParseInt(toks[1], 10, 64)
 812                         if err != nil || length < 0 {
 813                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 814                         }
 815                         name := dirname + "/" + manifestUnescape(toks[2])
 816                         fnode, err := dn.createFileAndParents(name)
 817                         if fnode == nil && err == nil && length == 0 {
 818                                 // Special case: an empty file used as
 819                                 // a marker to preserve an otherwise
 820                                 // empty directory in a manifest.
 821                                 continue
 822                         }
 823                         if err != nil || (fnode == nil && length != 0) {
 824                                 return fmt.Errorf("line %d: cannot use path %q with length %d: %s", lineno, name, length, err)
 825                         }
 826                         // Map the stream offset/range coordinates to
 827                         // block/offset/range coordinates and add
 828                         // corresponding storedSegments to the filenode
 829                         if pos > offset {
 830                                 // Can't continue where we left off.
 831                                 // TODO: binary search instead of
 832                                 // rewinding all the way (but this
 833                                 // situation might be rare anyway)
 834                                 segIdx, pos = 0, 0
 835                         }
 836                         for next := int64(0); segIdx < len(segments); segIdx++ {
 837                                 seg := segments[segIdx]
 838                                 next = pos + int64(seg.Len())
 839                                 if next <= offset || seg.Len() == 0 {
 840                                         pos = next
 841                                         continue
 842                                 }
 843                                 if pos >= offset+length {
 844                                         break
 845                                 }
 846                                 var blkOff int
 847                                 if pos < offset {
 848                                         blkOff = int(offset - pos)
 849                                 }
 850                                 blkLen := seg.Len() - blkOff
 851                                 if pos+int64(blkOff+blkLen) > offset+length {
 852                                         blkLen = int(offset + length - pos - int64(blkOff))
 853                                 }
 854                                 fnode.appendSegment(storedSegment{
 855                                         kc:      dn.fs,
 856                                         locator: seg.locator,
 857                                         size:    seg.size,
 858                                         offset:  blkOff,
 859                                         length:  blkLen,
 860                                 })
 861                                 if next > offset+length {
 862                                         break
 863                                 } else {
 864                                         pos = next
 865                                 }
 866                         }
 867                         if segIdx == len(segments) && pos < offset+length {
 868                                 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
 869                         }
 870                 }
 871                 if !anyFileTokens {
 872                         return fmt.Errorf("line %d: no file segments", lineno)
 873                 } else if len(segments) == 0 {
 874                         return fmt.Errorf("line %d: no locators", lineno)
 875                 } else if dirname == "" {
 876                         return fmt.Errorf("line %d: no stream name", lineno)
 877                 }
 878         }
 879         return nil
 880 }
 881
 882 // only safe to call from loadManifest -- no locking.
 883 //
 884 // If path is a "parent directory exists" marker (the last path
 885 // component is "."), the returned values are both nil.
 886 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
 887         var node inode = dn
 888         names := strings.Split(path, "/")
 889         basename := names[len(names)-1]
 890         for _, name := range names[:len(names)-1] {
 891                 switch name {
 892                 case "", ".":
 893                         continue
 894                 case "..":
 895                         if node == dn {
 896                                 // can't be sure parent will be a *dirnode
 897                                 return nil, ErrInvalidArgument
 898                         }
 899                         node = node.Parent()
 900                         continue
 901                 }
 902                 node, err = node.Child(name, func(child inode) (inode, error) {
 903                         if child == nil {
 904                                 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
 905                                 if err != nil {
 906                                         return nil, err
 907                                 }
 908                                 child.SetParent(node, name)
 909                                 return child, nil
 910                         } else if !child.IsDir() {
 911                                 return child, ErrFileExists
 912                         } else {
 913                                 return child, nil
 914                         }
 915                 })
 916                 if err != nil {
 917                         return
 918                 }
 919         }
 920         if basename == "." {
 921                 return
 922         } else if !permittedName(basename) {
 923                 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
 924                 return
 925         }
 926         _, err = node.Child(basename, func(child inode) (inode, error) {
 927                 switch child := child.(type) {
 928                 case nil:
 929                         child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
 930                         if err != nil {
 931                                 return nil, err
 932                         }
 933                         child.SetParent(node, basename)
 934                         fn = child.(*filenode)
 935                         return child, nil
 936                 case *filenode:
 937                         fn = child
 938                         return child, nil
 939                 case *dirnode:
 940                         return child, ErrIsDirectory
 941                 default:
 942                         return child, ErrInvalidArgument
 943                 }
 944         })
 945         return
 946 }
 947
 948 func (dn *dirnode) TreeSize() (bytes int64) {
 949         dn.RLock()
 950         defer dn.RUnlock()
 951         for _, i := range dn.inodes {
 952                 switch i := i.(type) {
 953                 case *filenode:
 954                         bytes += i.Size()
 955                 case *dirnode:
 956                         bytes += i.TreeSize()
 957                 }
 958         }
 959         return
 960 }
 961
 962 type segment interface {
 963         io.ReaderAt
 964         Len() int
 965         // Return a new segment with a subsection of the data from this
 966         // one. length<0 means length=Len()-off.
 967         Slice(off int, length int) segment
 968 }
 969
 970 type memSegment struct {
 971         buf []byte
 972 }
 973
 974 func (me *memSegment) Len() int {
 975         return len(me.buf)
 976 }
 977
 978 func (me *memSegment) Slice(off, length int) segment {
 979         if length < 0 {
 980                 length = len(me.buf) - off
 981         }
 982         buf := make([]byte, length)
 983         copy(buf, me.buf[off:])
 984         return &memSegment{buf: buf}
 985 }
 986
 987 func (me *memSegment) Truncate(n int) {
 988         if n > cap(me.buf) {
 989                 newsize := 1024
 990                 for newsize < n {
 991                         newsize = newsize << 2
 992                 }
 993                 newbuf := make([]byte, n, newsize)
 994                 copy(newbuf, me.buf)
 995                 me.buf = newbuf
 996         } else {
 997                 // Zero unused part when shrinking, in case we grow
 998                 // and start using it again later.
 999                 for i := n; i < len(me.buf); i++ {
1000                         me.buf[i] = 0
1001                 }
1002         }
1003         me.buf = me.buf[:n]
1004 }
1005
1006 func (me *memSegment) WriteAt(p []byte, off int) {
1007         if off+len(p) > len(me.buf) {
1008                 panic("overflowed segment")
1009         }
1010         copy(me.buf[off:], p)
1011 }
1012
1013 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1014         if off > int64(me.Len()) {
1015                 err = io.EOF
1016                 return
1017         }
1018         n = copy(p, me.buf[int(off):])
1019         if n < len(p) {
1020                 err = io.EOF
1021         }
1022         return
1023 }
1024
1025 type storedSegment struct {
1026         kc      fsBackend
1027         locator string
1028         size    int // size of stored block (also encoded in locator)
1029         offset  int // position of segment within the stored block
1030         length  int // bytes in this segment (offset + length <= size)
1031 }
1032
1033 func (se storedSegment) Len() int {
1034         return se.length
1035 }
1036
1037 func (se storedSegment) Slice(n, size int) segment {
1038         se.offset += n
1039         se.length -= n
1040         if size >= 0 && se.length > size {
1041                 se.length = size
1042         }
1043         return se
1044 }
1045
1046 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1047         if off > int64(se.length) {
1048                 return 0, io.EOF
1049         }
1050         maxlen := se.length - int(off)
1051         if len(p) > maxlen {
1052                 p = p[:maxlen]
1053                 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1054                 if err == nil {
1055                         err = io.EOF
1056                 }
1057                 return
1058         }
1059         return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1060 }
1061
1062 func canonicalName(name string) string {
1063         name = path.Clean("/" + name)
1064         if name == "/" || name == "./" {
1065                 name = "."
1066         } else if strings.HasPrefix(name, "/") {
1067                 name = "." + name
1068         }
1069         return name
1070 }
1071
1072 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1073
1074 func manifestUnescapeFunc(seq string) string {
1075         if seq == `\\` {
1076                 return `\`
1077         }
1078         i, err := strconv.ParseUint(seq[1:], 8, 8)
1079         if err != nil {
1080                 // Invalid escape sequence: can't unescape.
1081                 return seq
1082         }
1083         return string([]byte{byte(i)})
1084 }
1085
1086 func manifestUnescape(s string) string {
1087         return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1088 }
1089
1090 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1091
1092 func manifestEscapeFunc(seq string) string {
1093         return fmt.Sprintf("\\%03o", byte(seq[0]))
1094 }
1095
1096 func manifestEscape(s string) string {
1097         return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)
1098 }