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