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