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, name string) {
 223         fn.Lock()
 224         defer fn.Unlock()
 225         fn.parent = p
 226         fn.fileinfo.name = name
 227 }
 228
 229 func (fn *filenode) Parent() inode {
 230         fn.RLock()
 231         defer fn.RUnlock()
 232         return fn.parent
 233 }
 234
 235 func (fn *filenode) FS() FileSystem {
 236         return fn.fs
 237 }
 238
 239 // Read reads file data from a single segment, starting at startPtr,
 240 // into p. startPtr is assumed not to be up-to-date. Caller must have
 241 // RLock or Lock.
 242 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 243         ptr = fn.seek(startPtr)
 244         if ptr.off < 0 {
 245                 err = ErrNegativeOffset
 246                 return
 247         }
 248         if ptr.segmentIdx >= len(fn.segments) {
 249                 err = io.EOF
 250                 return
 251         }
 252         n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
 253         if n > 0 {
 254                 ptr.off += int64(n)
 255                 ptr.segmentOff += n
 256                 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
 257                         ptr.segmentIdx++
 258                         ptr.segmentOff = 0
 259                         if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
 260                                 err = nil
 261                         }
 262                 }
 263         }
 264         return
 265 }
 266
 267 func (fn *filenode) Size() int64 {
 268         fn.RLock()
 269         defer fn.RUnlock()
 270         return fn.fileinfo.Size()
 271 }
 272
 273 func (fn *filenode) FileInfo() os.FileInfo {
 274         fn.RLock()
 275         defer fn.RUnlock()
 276         return fn.fileinfo
 277 }
 278
 279 func (fn *filenode) Truncate(size int64) error {
 280         fn.Lock()
 281         defer fn.Unlock()
 282         return fn.truncate(size)
 283 }
 284
 285 func (fn *filenode) truncate(size int64) error {
 286         if size == fn.fileinfo.size {
 287                 return nil
 288         }
 289         fn.repacked++
 290         if size < fn.fileinfo.size {
 291                 ptr := fn.seek(filenodePtr{off: size})
 292                 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
 293                         if seg, ok := fn.segments[i].(*memSegment); ok {
 294                                 fn.memsize -= int64(seg.Len())
 295                         }
 296                 }
 297                 if ptr.segmentOff == 0 {
 298                         fn.segments = fn.segments[:ptr.segmentIdx]
 299                 } else {
 300                         fn.segments = fn.segments[:ptr.segmentIdx+1]
 301                         switch seg := fn.segments[ptr.segmentIdx].(type) {
 302                         case *memSegment:
 303                                 seg.Truncate(ptr.segmentOff)
 304                                 fn.memsize += int64(seg.Len())
 305                         default:
 306                                 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
 307                         }
 308                 }
 309                 fn.fileinfo.size = size
 310                 return nil
 311         }
 312         for size > fn.fileinfo.size {
 313                 grow := size - fn.fileinfo.size
 314                 var seg *memSegment
 315                 var ok bool
 316                 if len(fn.segments) == 0 {
 317                         seg = &memSegment{}
 318                         fn.segments = append(fn.segments, seg)
 319                 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
 320                         seg = &memSegment{}
 321                         fn.segments = append(fn.segments, seg)
 322                 }
 323                 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
 324                         grow = maxgrow
 325                 }
 326                 seg.Truncate(seg.Len() + int(grow))
 327                 fn.fileinfo.size += grow
 328                 fn.memsize += grow
 329         }
 330         return nil
 331 }
 332
 333 // Write writes data from p to the file, starting at startPtr,
 334 // extending the file size if necessary. Caller must have Lock.
 335 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 336         if startPtr.off > fn.fileinfo.size {
 337                 if err = fn.truncate(startPtr.off); err != nil {
 338                         return 0, startPtr, err
 339                 }
 340         }
 341         ptr = fn.seek(startPtr)
 342         if ptr.off < 0 {
 343                 err = ErrNegativeOffset
 344                 return
 345         }
 346         for len(p) > 0 && err == nil {
 347                 cando := p
 348                 if len(cando) > maxBlockSize {
 349                         cando = cando[:maxBlockSize]
 350                 }
 351                 // Rearrange/grow fn.segments (and shrink cando if
 352                 // needed) such that cando can be copied to
 353                 // fn.segments[ptr.segmentIdx] at offset
 354                 // ptr.segmentOff.
 355                 cur := ptr.segmentIdx
 356                 prev := ptr.segmentIdx - 1
 357                 var curWritable bool
 358                 if cur < len(fn.segments) {
 359                         _, curWritable = fn.segments[cur].(*memSegment)
 360                 }
 361                 var prevAppendable bool
 362                 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
 363                         _, prevAppendable = fn.segments[prev].(*memSegment)
 364                 }
 365                 if ptr.segmentOff > 0 && !curWritable {
 366                         // Split a non-writable block.
 367                         if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
 368                                 // Truncate cur, and insert a new
 369                                 // segment after it.
 370                                 cando = cando[:max]
 371                                 fn.segments = append(fn.segments, nil)
 372                                 copy(fn.segments[cur+1:], fn.segments[cur:])
 373                         } else {
 374                                 // Split cur into two copies, truncate
 375                                 // the one on the left, shift the one
 376                                 // on the right, and insert a new
 377                                 // segment between them.
 378                                 fn.segments = append(fn.segments, nil, nil)
 379                                 copy(fn.segments[cur+2:], fn.segments[cur:])
 380                                 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
 381                         }
 382                         cur++
 383                         prev++
 384                         seg := &memSegment{}
 385                         seg.Truncate(len(cando))
 386                         fn.memsize += int64(len(cando))
 387                         fn.segments[cur] = seg
 388                         fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
 389                         ptr.segmentIdx++
 390                         ptr.segmentOff = 0
 391                         fn.repacked++
 392                         ptr.repacked++
 393                 } else if curWritable {
 394                         if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
 395                                 cando = cando[:fit]
 396                         }
 397                 } else {
 398                         if prevAppendable {
 399                                 // Shrink cando if needed to fit in
 400                                 // prev segment.
 401                                 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
 402                                         cando = cando[:cangrow]
 403                                 }
 404                         }
 405
 406                         if cur == len(fn.segments) {
 407                                 // ptr is at EOF, filesize is changing.
 408                                 fn.fileinfo.size += int64(len(cando))
 409                         } else if el := fn.segments[cur].Len(); el <= len(cando) {
 410                                 // cando is long enough that we won't
 411                                 // need cur any more. shrink cando to
 412                                 // be exactly as long as cur
 413                                 // (otherwise we'd accidentally shift
 414                                 // the effective position of all
 415                                 // segments after cur).
 416                                 cando = cando[:el]
 417                                 copy(fn.segments[cur:], fn.segments[cur+1:])
 418                                 fn.segments = fn.segments[:len(fn.segments)-1]
 419                         } else {
 420                                 // shrink cur by the same #bytes we're growing prev
 421                                 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
 422                         }
 423
 424                         if prevAppendable {
 425                                 // Grow prev.
 426                                 ptr.segmentIdx--
 427                                 ptr.segmentOff = fn.segments[prev].Len()
 428                                 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
 429                                 fn.memsize += int64(len(cando))
 430                                 ptr.repacked++
 431                                 fn.repacked++
 432                         } else {
 433                                 // Insert a segment between prev and
 434                                 // cur, and advance prev/cur.
 435                                 fn.segments = append(fn.segments, nil)
 436                                 if cur < len(fn.segments) {
 437                                         copy(fn.segments[cur+1:], fn.segments[cur:])
 438                                         ptr.repacked++
 439                                         fn.repacked++
 440                                 } else {
 441                                         // appending a new segment does
 442                                         // not invalidate any ptrs
 443                                 }
 444                                 seg := &memSegment{}
 445                                 seg.Truncate(len(cando))
 446                                 fn.memsize += int64(len(cando))
 447                                 fn.segments[cur] = seg
 448                                 cur++
 449                                 prev++
 450                         }
 451                 }
 452
 453                 // Finally we can copy bytes from cando to the current segment.
 454                 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
 455                 n += len(cando)
 456                 p = p[len(cando):]
 457
 458                 ptr.off += int64(len(cando))
 459                 ptr.segmentOff += len(cando)
 460                 if ptr.segmentOff >= maxBlockSize {
 461                         fn.pruneMemSegments()
 462                 }
 463                 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
 464                         ptr.segmentOff = 0
 465                         ptr.segmentIdx++
 466                 }
 467
 468                 fn.fileinfo.modTime = time.Now()
 469         }
 470         return
 471 }
 472
 473 // Write some data out to disk to reduce memory use. Caller must have
 474 // write lock.
 475 func (fn *filenode) pruneMemSegments() {
 476         // TODO: async (don't hold Lock() while waiting for Keep)
 477         // TODO: share code with (*dirnode)sync()
 478         // TODO: pack/flush small blocks too, when fragmented
 479         for idx, seg := range fn.segments {
 480                 seg, ok := seg.(*memSegment)
 481                 if !ok || seg.Len() < maxBlockSize {
 482                         continue
 483                 }
 484                 locator, _, err := fn.FS().PutB(seg.buf)
 485                 if err != nil {
 486                         // TODO: stall (or return errors from)
 487                         // subsequent writes until flushing
 488                         // starts to succeed
 489                         continue
 490                 }
 491                 fn.memsize -= int64(seg.Len())
 492                 fn.segments[idx] = storedSegment{
 493                         kc:      fn.FS(),
 494                         locator: locator,
 495                         size:    seg.Len(),
 496                         offset:  0,
 497                         length:  seg.Len(),
 498                 }
 499         }
 500 }
 501
 502 type dirnode struct {
 503         fs *collectionFileSystem
 504         treenode
 505 }
 506
 507 func (dn *dirnode) FS() FileSystem {
 508         return dn.fs
 509 }
 510
 511 func (dn *dirnode) Child(name string, replace func(inode) inode) inode {
 512         if dn == dn.fs.rootnode() && name == ".arvados#collection" {
 513                 gn := &getternode{Getter: func() ([]byte, error) {
 514                         var coll Collection
 515                         var err error
 516                         coll.ManifestText, err = dn.fs.MarshalManifest(".")
 517                         if err != nil {
 518                                 return nil, err
 519                         }
 520                         data, err := json.Marshal(&coll)
 521                         if err == nil {
 522                                 data = append(data, 10)
 523                         }
 524                         return data, err
 525                 }}
 526                 gn.SetParent(dn, name)
 527                 return gn
 528         }
 529         return dn.treenode.Child(name, replace)
 530 }
 531
 532 // sync flushes in-memory data (for all files in the tree rooted at
 533 // dn) to persistent storage. Caller must hold dn.Lock().
 534 func (dn *dirnode) sync() error {
 535         type shortBlock struct {
 536                 fn  *filenode
 537                 idx int
 538         }
 539         var pending []shortBlock
 540         var pendingLen int
 541
 542         flush := func(sbs []shortBlock) error {
 543                 if len(sbs) == 0 {
 544                         return nil
 545                 }
 546                 block := make([]byte, 0, maxBlockSize)
 547                 for _, sb := range sbs {
 548                         block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
 549                 }
 550                 locator, _, err := dn.fs.PutB(block)
 551                 if err != nil {
 552                         return err
 553                 }
 554                 off := 0
 555                 for _, sb := range sbs {
 556                         data := sb.fn.segments[sb.idx].(*memSegment).buf
 557                         sb.fn.segments[sb.idx] = storedSegment{
 558                                 kc:      dn.fs,
 559                                 locator: locator,
 560                                 size:    len(block),
 561                                 offset:  off,
 562                                 length:  len(data),
 563                         }
 564                         off += len(data)
 565                         sb.fn.memsize -= int64(len(data))
 566                 }
 567                 return nil
 568         }
 569
 570         names := make([]string, 0, len(dn.inodes))
 571         for name := range dn.inodes {
 572                 names = append(names, name)
 573         }
 574         sort.Strings(names)
 575
 576         for _, name := range names {
 577                 fn, ok := dn.inodes[name].(*filenode)
 578                 if !ok {
 579                         continue
 580                 }
 581                 fn.Lock()
 582                 defer fn.Unlock()
 583                 for idx, seg := range fn.segments {
 584                         seg, ok := seg.(*memSegment)
 585                         if !ok {
 586                                 continue
 587                         }
 588                         if seg.Len() > maxBlockSize/2 {
 589                                 if err := flush([]shortBlock{{fn, idx}}); err != nil {
 590                                         return err
 591                                 }
 592                                 continue
 593                         }
 594                         if pendingLen+seg.Len() > maxBlockSize {
 595                                 if err := flush(pending); err != nil {
 596                                         return err
 597                                 }
 598                                 pending = nil
 599                                 pendingLen = 0
 600                         }
 601                         pending = append(pending, shortBlock{fn, idx})
 602                         pendingLen += seg.Len()
 603                 }
 604         }
 605         return flush(pending)
 606 }
 607
 608 // caller must have read lock.
 609 func (dn *dirnode) marshalManifest(prefix string) (string, error) {
 610         var streamLen int64
 611         type filepart struct {
 612                 name   string
 613                 offset int64
 614                 length int64
 615         }
 616         var fileparts []filepart
 617         var subdirs string
 618         var blocks []string
 619
 620         if err := dn.sync(); err != nil {
 621                 return "", err
 622         }
 623
 624         names := make([]string, 0, len(dn.inodes))
 625         for name, node := range dn.inodes {
 626                 names = append(names, name)
 627                 node.Lock()
 628                 defer node.Unlock()
 629         }
 630         sort.Strings(names)
 631
 632         for _, name := range names {
 633                 switch node := dn.inodes[name].(type) {
 634                 case *dirnode:
 635                         subdir, err := node.marshalManifest(prefix + "/" + name)
 636                         if err != nil {
 637                                 return "", err
 638                         }
 639                         subdirs = subdirs + subdir
 640                 case *filenode:
 641                         if len(node.segments) == 0 {
 642                                 fileparts = append(fileparts, filepart{name: name})
 643                                 break
 644                         }
 645                         for _, seg := range node.segments {
 646                                 switch seg := seg.(type) {
 647                                 case storedSegment:
 648                                         if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
 649                                                 streamLen -= int64(seg.size)
 650                                         } else {
 651                                                 blocks = append(blocks, seg.locator)
 652                                         }
 653                                         next := filepart{
 654                                                 name:   name,
 655                                                 offset: streamLen + int64(seg.offset),
 656                                                 length: int64(seg.length),
 657                                         }
 658                                         if prev := len(fileparts) - 1; prev >= 0 &&
 659                                                 fileparts[prev].name == name &&
 660                                                 fileparts[prev].offset+fileparts[prev].length == next.offset {
 661                                                 fileparts[prev].length += next.length
 662                                         } else {
 663                                                 fileparts = append(fileparts, next)
 664                                         }
 665                                         streamLen += int64(seg.size)
 666                                 default:
 667                                         // This can't happen: we
 668                                         // haven't unlocked since
 669                                         // calling sync().
 670                                         panic(fmt.Sprintf("can't marshal segment type %T", seg))
 671                                 }
 672                         }
 673                 default:
 674                         panic(fmt.Sprintf("can't marshal inode type %T", node))
 675                 }
 676         }
 677         var filetokens []string
 678         for _, s := range fileparts {
 679                 filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
 680         }
 681         if len(filetokens) == 0 {
 682                 return subdirs, nil
 683         } else if len(blocks) == 0 {
 684                 blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
 685         }
 686         return manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n" + subdirs, nil
 687 }
 688
 689 func (dn *dirnode) loadManifest(txt string) error {
 690         var dirname string
 691         streams := strings.Split(txt, "\n")
 692         if streams[len(streams)-1] != "" {
 693                 return fmt.Errorf("line %d: no trailing newline", len(streams))
 694         }
 695         streams = streams[:len(streams)-1]
 696         segments := []storedSegment{}
 697         for i, stream := range streams {
 698                 lineno := i + 1
 699                 var anyFileTokens bool
 700                 var pos int64
 701                 var segIdx int
 702                 segments = segments[:0]
 703                 for i, token := range strings.Split(stream, " ") {
 704                         if i == 0 {
 705                                 dirname = manifestUnescape(token)
 706                                 continue
 707                         }
 708                         if !strings.Contains(token, ":") {
 709                                 if anyFileTokens {
 710                                         return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 711                                 }
 712                                 toks := strings.SplitN(token, "+", 3)
 713                                 if len(toks) < 2 {
 714                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 715                                 }
 716                                 length, err := strconv.ParseInt(toks[1], 10, 32)
 717                                 if err != nil || length < 0 {
 718                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
 719                                 }
 720                                 segments = append(segments, storedSegment{
 721                                         locator: token,
 722                                         size:    int(length),
 723                                         offset:  0,
 724                                         length:  int(length),
 725                                 })
 726                                 continue
 727                         } else if len(segments) == 0 {
 728                                 return fmt.Errorf("line %d: bad locator %q", lineno, token)
 729                         }
 730
 731                         toks := strings.SplitN(token, ":", 3)
 732                         if len(toks) != 3 {
 733                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 734                         }
 735                         anyFileTokens = true
 736
 737                         offset, err := strconv.ParseInt(toks[0], 10, 64)
 738                         if err != nil || offset < 0 {
 739                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 740                         }
 741                         length, err := strconv.ParseInt(toks[1], 10, 64)
 742                         if err != nil || length < 0 {
 743                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 744                         }
 745                         name := dirname + "/" + manifestUnescape(toks[2])
 746                         fnode, err := dn.createFileAndParents(name)
 747                         if err != nil {
 748                                 return fmt.Errorf("line %d: cannot use path %q: %s", lineno, name, err)
 749                         }
 750                         // Map the stream offset/range coordinates to
 751                         // block/offset/range coordinates and add
 752                         // corresponding storedSegments to the filenode
 753                         if pos > offset {
 754                                 // Can't continue where we left off.
 755                                 // TODO: binary search instead of
 756                                 // rewinding all the way (but this
 757                                 // situation might be rare anyway)
 758                                 segIdx, pos = 0, 0
 759                         }
 760                         for next := int64(0); segIdx < len(segments); segIdx++ {
 761                                 seg := segments[segIdx]
 762                                 next = pos + int64(seg.Len())
 763                                 if next <= offset || seg.Len() == 0 {
 764                                         pos = next
 765                                         continue
 766                                 }
 767                                 if pos >= offset+length {
 768                                         break
 769                                 }
 770                                 var blkOff int
 771                                 if pos < offset {
 772                                         blkOff = int(offset - pos)
 773                                 }
 774                                 blkLen := seg.Len() - blkOff
 775                                 if pos+int64(blkOff+blkLen) > offset+length {
 776                                         blkLen = int(offset + length - pos - int64(blkOff))
 777                                 }
 778                                 fnode.appendSegment(storedSegment{
 779                                         kc:      dn.fs,
 780                                         locator: seg.locator,
 781                                         size:    seg.size,
 782                                         offset:  blkOff,
 783                                         length:  blkLen,
 784                                 })
 785                                 if next > offset+length {
 786                                         break
 787                                 } else {
 788                                         pos = next
 789                                 }
 790                         }
 791                         if segIdx == len(segments) && pos < offset+length {
 792                                 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
 793                         }
 794                 }
 795                 if !anyFileTokens {
 796                         return fmt.Errorf("line %d: no file segments", lineno)
 797                 } else if len(segments) == 0 {
 798                         return fmt.Errorf("line %d: no locators", lineno)
 799                 } else if dirname == "" {
 800                         return fmt.Errorf("line %d: no stream name", lineno)
 801                 }
 802         }
 803         return nil
 804 }
 805
 806 // only safe to call from loadManifest -- no locking
 807 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
 808         var node inode = dn
 809         names := strings.Split(path, "/")
 810         basename := names[len(names)-1]
 811         if basename == "" || basename == "." || basename == ".." {
 812                 err = fmt.Errorf("invalid filename")
 813                 return
 814         }
 815         for _, name := range names[:len(names)-1] {
 816                 switch name {
 817                 case "", ".":
 818                         continue
 819                 case "..":
 820                         if node == dn {
 821                                 // can't be sure parent will be a *dirnode
 822                                 return nil, ErrInvalidArgument
 823                         }
 824                         node = node.Parent()
 825                         continue
 826                 }
 827                 node.Child(name, func(child inode) inode {
 828                         if child == nil {
 829                                 node, err = node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
 830                                 child = node
 831                         } else if !child.IsDir() {
 832                                 err = ErrFileExists
 833                         } else {
 834                                 node = child
 835                         }
 836                         return child
 837                 })
 838                 if err != nil {
 839                         return
 840                 }
 841         }
 842         node.Child(basename, func(child inode) inode {
 843                 switch child := child.(type) {
 844                 case nil:
 845                         child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
 846                         fn = child.(*filenode)
 847                         return child
 848                 case *filenode:
 849                         fn = child
 850                         return child
 851                 case *dirnode:
 852                         err = ErrIsDirectory
 853                         return child
 854                 default:
 855                         err = ErrInvalidArgument
 856                         return child
 857                 }
 858         })
 859         return
 860 }
 861
 862 type segment interface {
 863         io.ReaderAt
 864         Len() int
 865         // Return a new segment with a subsection of the data from this
 866         // one. length<0 means length=Len()-off.
 867         Slice(off int, length int) segment
 868 }
 869
 870 type memSegment struct {
 871         buf []byte
 872 }
 873
 874 func (me *memSegment) Len() int {
 875         return len(me.buf)
 876 }
 877
 878 func (me *memSegment) Slice(off, length int) segment {
 879         if length < 0 {
 880                 length = len(me.buf) - off
 881         }
 882         buf := make([]byte, length)
 883         copy(buf, me.buf[off:])
 884         return &memSegment{buf: buf}
 885 }
 886
 887 func (me *memSegment) Truncate(n int) {
 888         if n > cap(me.buf) {
 889                 newsize := 1024
 890                 for newsize < n {
 891                         newsize = newsize << 2
 892                 }
 893                 newbuf := make([]byte, n, newsize)
 894                 copy(newbuf, me.buf)
 895                 me.buf = newbuf
 896         } else {
 897                 // Zero unused part when shrinking, in case we grow
 898                 // and start using it again later.
 899                 for i := n; i < len(me.buf); i++ {
 900                         me.buf[i] = 0
 901                 }
 902         }
 903         me.buf = me.buf[:n]
 904 }
 905
 906 func (me *memSegment) WriteAt(p []byte, off int) {
 907         if off+len(p) > len(me.buf) {
 908                 panic("overflowed segment")
 909         }
 910         copy(me.buf[off:], p)
 911 }
 912
 913 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
 914         if off > int64(me.Len()) {
 915                 err = io.EOF
 916                 return
 917         }
 918         n = copy(p, me.buf[int(off):])
 919         if n < len(p) {
 920                 err = io.EOF
 921         }
 922         return
 923 }
 924
 925 type storedSegment struct {
 926         kc      fsBackend
 927         locator string
 928         size    int // size of stored block (also encoded in locator)
 929         offset  int // position of segment within the stored block
 930         length  int // bytes in this segment (offset + length <= size)
 931 }
 932
 933 func (se storedSegment) Len() int {
 934         return se.length
 935 }
 936
 937 func (se storedSegment) Slice(n, size int) segment {
 938         se.offset += n
 939         se.length -= n
 940         if size >= 0 && se.length > size {
 941                 se.length = size
 942         }
 943         return se
 944 }
 945
 946 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
 947         if off > int64(se.length) {
 948                 return 0, io.EOF
 949         }
 950         maxlen := se.length - int(off)
 951         if len(p) > maxlen {
 952                 p = p[:maxlen]
 953                 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
 954                 if err == nil {
 955                         err = io.EOF
 956                 }
 957                 return
 958         }
 959         return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
 960 }
 961
 962 func canonicalName(name string) string {
 963         name = path.Clean("/" + name)
 964         if name == "/" || name == "./" {
 965                 name = "."
 966         } else if strings.HasPrefix(name, "/") {
 967                 name = "." + name
 968         }
 969         return name
 970 }
 971
 972 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
 973
 974 func manifestUnescapeFunc(seq string) string {
 975         if seq == `\\` {
 976                 return `\`
 977         }
 978         i, err := strconv.ParseUint(seq[1:], 8, 8)
 979         if err != nil {
 980                 // Invalid escape sequence: can't unescape.
 981                 return seq
 982         }
 983         return string([]byte{byte(i)})
 984 }
 985
 986 func manifestUnescape(s string) string {
 987         return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
 988 }
 989
 990 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
 991
 992 func manifestEscapeFunc(seq string) string {
 993         return fmt.Sprintf("\\%03o", byte(seq[0]))
 994 }
 995
 996 func manifestEscape(s string) string {
 997         return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)
 998 }