sdk/go/arvados/fs_collection.go

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