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