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         "context"
   9         "encoding/json"
  10         "fmt"
  11         "io"
  12         "os"
  13         "path"
  14         "regexp"
  15         "sort"
  16         "strconv"
  17         "strings"
  18         "sync"
  19         "sync/atomic"
  20         "time"
  21 )
  22
  23 var (
  24         maxBlockSize      = 1 << 26
  25         concurrentWriters = 4 // max goroutines writing to Keep in background and during flush()
  26 )
  27
  28 // A CollectionFileSystem is a FileSystem that can be serialized as a
  29 // manifest and stored as a collection.
  30 type CollectionFileSystem interface {
  31         FileSystem
  32
  33         // Flush all file data to Keep and return a snapshot of the
  34         // filesystem suitable for saving as (Collection)ManifestText.
  35         // Prefix (normally ".") is a top level directory, effectively
  36         // prepended to all paths in the returned manifest.
  37         MarshalManifest(prefix string) (string, error)
  38
  39         // Total data bytes in all files.
  40         Size() int64
  41
  42         // Memory consumed by buffered file data.
  43         memorySize() int64
  44 }
  45
  46 type collectionFileSystem struct {
  47         fileSystem
  48         uuid string
  49 }
  50
  51 // FileSystem returns a CollectionFileSystem for the collection.
  52 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
  53         modTime := c.ModifiedAt
  54         if modTime.IsZero() {
  55                 modTime = time.Now()
  56         }
  57         fs := &collectionFileSystem{
  58                 uuid: c.UUID,
  59                 fileSystem: fileSystem{
  60                         fsBackend: keepBackend{apiClient: client, keepClient: kc},
  61                         thr:       newThrottle(concurrentWriters),
  62                 },
  63         }
  64         root := &dirnode{
  65                 fs: fs,
  66                 treenode: treenode{
  67                         fileinfo: fileinfo{
  68                                 name:    ".",
  69                                 mode:    os.ModeDir | 0755,
  70                                 modTime: modTime,
  71                         },
  72                         inodes: make(map[string]inode),
  73                 },
  74         }
  75         root.SetParent(root, ".")
  76         if err := root.loadManifest(c.ManifestText); err != nil {
  77                 return nil, err
  78         }
  79         backdateTree(root, modTime)
  80         fs.root = root
  81         return fs, nil
  82 }
  83
  84 func backdateTree(n inode, modTime time.Time) {
  85         switch n := n.(type) {
  86         case *filenode:
  87                 n.fileinfo.modTime = modTime
  88         case *dirnode:
  89                 n.fileinfo.modTime = modTime
  90                 for _, n := range n.inodes {
  91                         backdateTree(n, modTime)
  92                 }
  93         }
  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         if fs.uuid == "" {
 126                 return nil
 127         }
 128         txt, err := fs.MarshalManifest(".")
 129         if err != nil {
 130                 return fmt.Errorf("sync failed: %s", err)
 131         }
 132         coll := &Collection{
 133                 UUID:         fs.uuid,
 134                 ManifestText: txt,
 135         }
 136         err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
 137                 "collection": map[string]string{
 138                         "manifest_text": coll.ManifestText,
 139                 },
 140                 "select": []string{"uuid"},
 141         })
 142         if err != nil {
 143                 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
 144         }
 145         return nil
 146 }
 147
 148 func (fs *collectionFileSystem) Flush(path string, shortBlocks bool) error {
 149         node, err := rlookup(fs.fileSystem.root, path)
 150         if err != nil {
 151                 return err
 152         }
 153         dn, ok := node.(*dirnode)
 154         if !ok {
 155                 return ErrNotADirectory
 156         }
 157         dn.Lock()
 158         defer dn.Unlock()
 159         names := dn.sortedNames()
 160         if path != "" {
 161                 // Caller only wants to flush the specified dir,
 162                 // non-recursively.  Drop subdirs from the list of
 163                 // names.
 164                 var filenames []string
 165                 for _, name := range names {
 166                         if _, ok := dn.inodes[name].(*filenode); ok {
 167                                 filenames = append(filenames, name)
 168                         }
 169                 }
 170                 names = filenames
 171         }
 172         for _, name := range names {
 173                 child := dn.inodes[name]
 174                 child.Lock()
 175                 defer child.Unlock()
 176         }
 177         return dn.flush(context.TODO(), names, flushOpts{sync: false, shortBlocks: shortBlocks})
 178 }
 179
 180 func (fs *collectionFileSystem) memorySize() int64 {
 181         fs.fileSystem.root.Lock()
 182         defer fs.fileSystem.root.Unlock()
 183         return fs.fileSystem.root.(*dirnode).memorySize()
 184 }
 185
 186 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
 187         fs.fileSystem.root.Lock()
 188         defer fs.fileSystem.root.Unlock()
 189         return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix)
 190 }
 191
 192 func (fs *collectionFileSystem) Size() int64 {
 193         return fs.fileSystem.root.(*dirnode).TreeSize()
 194 }
 195
 196 // filenodePtr is an offset into a file that is (usually) efficient to
 197 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
 198 // then
 199 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
 200 // corresponds to file offset filenodePtr.off. Otherwise, it is
 201 // necessary to reexamine len(filenode.segments[0]) etc. to find the
 202 // correct segment and offset.
 203 type filenodePtr struct {
 204         off        int64
 205         segmentIdx int
 206         segmentOff int
 207         repacked   int64
 208 }
 209
 210 // seek returns a ptr that is consistent with both startPtr.off and
 211 // the current state of fn. The caller must already hold fn.RLock() or
 212 // fn.Lock().
 213 //
 214 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
 215 // EOF.
 216 //
 217 // After seeking:
 218 //
 219 //     ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
 220 //     ||
 221 //     filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
 222 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
 223         ptr = startPtr
 224         if ptr.off < 0 {
 225                 // meaningless anyway
 226                 return
 227         } else if ptr.off >= fn.fileinfo.size {
 228                 ptr.segmentIdx = len(fn.segments)
 229                 ptr.segmentOff = 0
 230                 ptr.repacked = fn.repacked
 231                 return
 232         } else if ptr.repacked == fn.repacked {
 233                 // segmentIdx and segmentOff accurately reflect
 234                 // ptr.off, but might have fallen off the end of a
 235                 // segment
 236                 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
 237                         ptr.segmentIdx++
 238                         ptr.segmentOff = 0
 239                 }
 240                 return
 241         }
 242         defer func() {
 243                 ptr.repacked = fn.repacked
 244         }()
 245         if ptr.off >= fn.fileinfo.size {
 246                 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
 247                 return
 248         }
 249         // Recompute segmentIdx and segmentOff.  We have already
 250         // established fn.fileinfo.size > ptr.off >= 0, so we don't
 251         // have to deal with edge cases here.
 252         var off int64
 253         for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
 254                 // This would panic (index out of range) if
 255                 // fn.fileinfo.size were larger than
 256                 // sum(fn.segments[i].Len()) -- but that can't happen
 257                 // because we have ensured fn.fileinfo.size is always
 258                 // accurate.
 259                 segLen := int64(fn.segments[ptr.segmentIdx].Len())
 260                 if off+segLen > ptr.off {
 261                         ptr.segmentOff = int(ptr.off - off)
 262                         break
 263                 }
 264                 off += segLen
 265         }
 266         return
 267 }
 268
 269 // filenode implements inode.
 270 type filenode struct {
 271         parent   inode
 272         fs       FileSystem
 273         fileinfo fileinfo
 274         segments []segment
 275         // number of times `segments` has changed in a
 276         // way that might invalidate a filenodePtr
 277         repacked int64
 278         memsize  int64 // bytes in memSegments
 279         sync.RWMutex
 280         nullnode
 281 }
 282
 283 // caller must have lock
 284 func (fn *filenode) appendSegment(e segment) {
 285         fn.segments = append(fn.segments, e)
 286         fn.fileinfo.size += int64(e.Len())
 287 }
 288
 289 func (fn *filenode) SetParent(p inode, name string) {
 290         fn.Lock()
 291         defer fn.Unlock()
 292         fn.parent = p
 293         fn.fileinfo.name = name
 294 }
 295
 296 func (fn *filenode) Parent() inode {
 297         fn.RLock()
 298         defer fn.RUnlock()
 299         return fn.parent
 300 }
 301
 302 func (fn *filenode) FS() FileSystem {
 303         return fn.fs
 304 }
 305
 306 // Read reads file data from a single segment, starting at startPtr,
 307 // into p. startPtr is assumed not to be up-to-date. Caller must have
 308 // RLock or Lock.
 309 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 310         ptr = fn.seek(startPtr)
 311         if ptr.off < 0 {
 312                 err = ErrNegativeOffset
 313                 return
 314         }
 315         if ptr.segmentIdx >= len(fn.segments) {
 316                 err = io.EOF
 317                 return
 318         }
 319         n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
 320         if n > 0 {
 321                 ptr.off += int64(n)
 322                 ptr.segmentOff += n
 323                 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
 324                         ptr.segmentIdx++
 325                         ptr.segmentOff = 0
 326                         if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
 327                                 err = nil
 328                         }
 329                 }
 330         }
 331         return
 332 }
 333
 334 func (fn *filenode) Size() int64 {
 335         fn.RLock()
 336         defer fn.RUnlock()
 337         return fn.fileinfo.Size()
 338 }
 339
 340 func (fn *filenode) FileInfo() os.FileInfo {
 341         fn.RLock()
 342         defer fn.RUnlock()
 343         return fn.fileinfo
 344 }
 345
 346 func (fn *filenode) Truncate(size int64) error {
 347         fn.Lock()
 348         defer fn.Unlock()
 349         return fn.truncate(size)
 350 }
 351
 352 func (fn *filenode) truncate(size int64) error {
 353         if size == fn.fileinfo.size {
 354                 return nil
 355         }
 356         fn.repacked++
 357         if size < fn.fileinfo.size {
 358                 ptr := fn.seek(filenodePtr{off: size})
 359                 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
 360                         if seg, ok := fn.segments[i].(*memSegment); ok {
 361                                 fn.memsize -= int64(seg.Len())
 362                         }
 363                 }
 364                 if ptr.segmentOff == 0 {
 365                         fn.segments = fn.segments[:ptr.segmentIdx]
 366                 } else {
 367                         fn.segments = fn.segments[:ptr.segmentIdx+1]
 368                         switch seg := fn.segments[ptr.segmentIdx].(type) {
 369                         case *memSegment:
 370                                 seg.Truncate(ptr.segmentOff)
 371                                 fn.memsize += int64(seg.Len())
 372                         default:
 373                                 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
 374                         }
 375                 }
 376                 fn.fileinfo.size = size
 377                 return nil
 378         }
 379         for size > fn.fileinfo.size {
 380                 grow := size - fn.fileinfo.size
 381                 var seg *memSegment
 382                 var ok bool
 383                 if len(fn.segments) == 0 {
 384                         seg = &memSegment{}
 385                         fn.segments = append(fn.segments, seg)
 386                 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
 387                         seg = &memSegment{}
 388                         fn.segments = append(fn.segments, seg)
 389                 }
 390                 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
 391                         grow = maxgrow
 392                 }
 393                 seg.Truncate(seg.Len() + int(grow))
 394                 fn.fileinfo.size += grow
 395                 fn.memsize += grow
 396         }
 397         return nil
 398 }
 399
 400 // Write writes data from p to the file, starting at startPtr,
 401 // extending the file size if necessary. Caller must have Lock.
 402 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 403         if startPtr.off > fn.fileinfo.size {
 404                 if err = fn.truncate(startPtr.off); err != nil {
 405                         return 0, startPtr, err
 406                 }
 407         }
 408         ptr = fn.seek(startPtr)
 409         if ptr.off < 0 {
 410                 err = ErrNegativeOffset
 411                 return
 412         }
 413         for len(p) > 0 && err == nil {
 414                 cando := p
 415                 if len(cando) > maxBlockSize {
 416                         cando = cando[:maxBlockSize]
 417                 }
 418                 // Rearrange/grow fn.segments (and shrink cando if
 419                 // needed) such that cando can be copied to
 420                 // fn.segments[ptr.segmentIdx] at offset
 421                 // ptr.segmentOff.
 422                 cur := ptr.segmentIdx
 423                 prev := ptr.segmentIdx - 1
 424                 var curWritable bool
 425                 if cur < len(fn.segments) {
 426                         _, curWritable = fn.segments[cur].(*memSegment)
 427                 }
 428                 var prevAppendable bool
 429                 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
 430                         _, prevAppendable = fn.segments[prev].(*memSegment)
 431                 }
 432                 if ptr.segmentOff > 0 && !curWritable {
 433                         // Split a non-writable block.
 434                         if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
 435                                 // Truncate cur, and insert a new
 436                                 // segment after it.
 437                                 cando = cando[:max]
 438                                 fn.segments = append(fn.segments, nil)
 439                                 copy(fn.segments[cur+1:], fn.segments[cur:])
 440                         } else {
 441                                 // Split cur into two copies, truncate
 442                                 // the one on the left, shift the one
 443                                 // on the right, and insert a new
 444                                 // segment between them.
 445                                 fn.segments = append(fn.segments, nil, nil)
 446                                 copy(fn.segments[cur+2:], fn.segments[cur:])
 447                                 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
 448                         }
 449                         cur++
 450                         prev++
 451                         seg := &memSegment{}
 452                         seg.Truncate(len(cando))
 453                         fn.memsize += int64(len(cando))
 454                         fn.segments[cur] = seg
 455                         fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
 456                         ptr.segmentIdx++
 457                         ptr.segmentOff = 0
 458                         fn.repacked++
 459                         ptr.repacked++
 460                 } else if curWritable {
 461                         if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
 462                                 cando = cando[:fit]
 463                         }
 464                 } else {
 465                         if prevAppendable {
 466                                 // Shrink cando if needed to fit in
 467                                 // prev segment.
 468                                 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
 469                                         cando = cando[:cangrow]
 470                                 }
 471                         }
 472
 473                         if cur == len(fn.segments) {
 474                                 // ptr is at EOF, filesize is changing.
 475                                 fn.fileinfo.size += int64(len(cando))
 476                         } else if el := fn.segments[cur].Len(); el <= len(cando) {
 477                                 // cando is long enough that we won't
 478                                 // need cur any more. shrink cando to
 479                                 // be exactly as long as cur
 480                                 // (otherwise we'd accidentally shift
 481                                 // the effective position of all
 482                                 // segments after cur).
 483                                 cando = cando[:el]
 484                                 copy(fn.segments[cur:], fn.segments[cur+1:])
 485                                 fn.segments = fn.segments[:len(fn.segments)-1]
 486                         } else {
 487                                 // shrink cur by the same #bytes we're growing prev
 488                                 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
 489                         }
 490
 491                         if prevAppendable {
 492                                 // Grow prev.
 493                                 ptr.segmentIdx--
 494                                 ptr.segmentOff = fn.segments[prev].Len()
 495                                 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
 496                                 fn.memsize += int64(len(cando))
 497                                 ptr.repacked++
 498                                 fn.repacked++
 499                         } else {
 500                                 // Insert a segment between prev and
 501                                 // cur, and advance prev/cur.
 502                                 fn.segments = append(fn.segments, nil)
 503                                 if cur < len(fn.segments) {
 504                                         copy(fn.segments[cur+1:], fn.segments[cur:])
 505                                         ptr.repacked++
 506                                         fn.repacked++
 507                                 } else {
 508                                         // appending a new segment does
 509                                         // not invalidate any ptrs
 510                                 }
 511                                 seg := &memSegment{}
 512                                 seg.Truncate(len(cando))
 513                                 fn.memsize += int64(len(cando))
 514                                 fn.segments[cur] = seg
 515                                 cur++
 516                                 prev++
 517                         }
 518                 }
 519
 520                 // Finally we can copy bytes from cando to the current segment.
 521                 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
 522                 n += len(cando)
 523                 p = p[len(cando):]
 524
 525                 ptr.off += int64(len(cando))
 526                 ptr.segmentOff += len(cando)
 527                 if ptr.segmentOff >= maxBlockSize {
 528                         fn.pruneMemSegments()
 529                 }
 530                 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
 531                         ptr.segmentOff = 0
 532                         ptr.segmentIdx++
 533                 }
 534
 535                 fn.fileinfo.modTime = time.Now()
 536         }
 537         return
 538 }
 539
 540 // Write some data out to disk to reduce memory use. Caller must have
 541 // write lock.
 542 func (fn *filenode) pruneMemSegments() {
 543         // TODO: share code with (*dirnode)flush()
 544         // TODO: pack/flush small blocks too, when fragmented
 545         for idx, seg := range fn.segments {
 546                 seg, ok := seg.(*memSegment)
 547                 if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
 548                         continue
 549                 }
 550                 // Setting seg.flushing guarantees seg.buf will not be
 551                 // modified in place: WriteAt and Truncate will
 552                 // allocate a new buf instead, if necessary.
 553                 idx, buf := idx, seg.buf
 554                 done := make(chan struct{})
 555                 seg.flushing = done
 556                 // If lots of background writes are already in
 557                 // progress, block here until one finishes, rather
 558                 // than pile up an unlimited number of buffered writes
 559                 // and network flush operations.
 560                 fn.fs.throttle().Acquire()
 561                 go func() {
 562                         defer close(done)
 563                         locator, _, err := fn.FS().PutB(buf)
 564                         fn.fs.throttle().Release()
 565                         fn.Lock()
 566                         defer fn.Unlock()
 567                         if seg.flushing != done {
 568                                 // A new seg.buf has been allocated.
 569                                 return
 570                         }
 571                         seg.flushing = nil
 572                         if err != nil {
 573                                 // TODO: stall (or return errors from)
 574                                 // subsequent writes until flushing
 575                                 // starts to succeed.
 576                                 return
 577                         }
 578                         if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
 579                                 // Segment has been dropped/moved/resized.
 580                                 return
 581                         }
 582                         fn.memsize -= int64(len(buf))
 583                         fn.segments[idx] = storedSegment{
 584                                 kc:      fn.FS(),
 585                                 locator: locator,
 586                                 size:    len(buf),
 587                                 offset:  0,
 588                                 length:  len(buf),
 589                         }
 590                 }()
 591         }
 592 }
 593
 594 // Block until all pending pruneMemSegments/flush work is
 595 // finished. Caller must NOT have lock.
 596 func (fn *filenode) waitPrune() {
 597         var pending []<-chan struct{}
 598         fn.Lock()
 599         for _, seg := range fn.segments {
 600                 if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
 601                         pending = append(pending, seg.flushing)
 602                 }
 603         }
 604         fn.Unlock()
 605         for _, p := range pending {
 606                 <-p
 607         }
 608 }
 609
 610 type dirnode struct {
 611         fs *collectionFileSystem
 612         treenode
 613 }
 614
 615 func (dn *dirnode) FS() FileSystem {
 616         return dn.fs
 617 }
 618
 619 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
 620         if dn == dn.fs.rootnode() && name == ".arvados#collection" {
 621                 gn := &getternode{Getter: func() ([]byte, error) {
 622                         var coll Collection
 623                         var err error
 624                         coll.ManifestText, err = dn.fs.MarshalManifest(".")
 625                         if err != nil {
 626                                 return nil, err
 627                         }
 628                         data, err := json.Marshal(&coll)
 629                         if err == nil {
 630                                 data = append(data, '\n')
 631                         }
 632                         return data, err
 633                 }}
 634                 gn.SetParent(dn, name)
 635                 return gn, nil
 636         }
 637         return dn.treenode.Child(name, replace)
 638 }
 639
 640 type fnSegmentRef struct {
 641         fn  *filenode
 642         idx int
 643 }
 644
 645 // commitBlock concatenates the data from the given filenode segments
 646 // (which must be *memSegments), writes the data out to Keep as a
 647 // single block, and replaces the filenodes' *memSegments with
 648 // storedSegments that reference the relevant portions of the new
 649 // block.
 650 //
 651 // bufsize is the total data size in refs. It is used to preallocate
 652 // the correct amount of memory when len(refs)>1.
 653 //
 654 // If sync is false, commitBlock returns right away, after starting a
 655 // goroutine to do the writes, reacquire the filenodes' locks, and
 656 // swap out the *memSegments. Some filenodes' segments might get
 657 // modified/rearranged in the meantime, in which case commitBlock
 658 // won't replace them.
 659 //
 660 // Caller must have write lock.
 661 func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize int, sync bool) error {
 662         if len(refs) == 0 {
 663                 return nil
 664         }
 665         if err := ctx.Err(); err != nil {
 666                 return err
 667         }
 668         done := make(chan struct{})
 669         var block []byte
 670         segs := make([]*memSegment, 0, len(refs))
 671         offsets := make([]int, 0, len(refs)) // location of segment's data within block
 672         for _, ref := range refs {
 673                 seg := ref.fn.segments[ref.idx].(*memSegment)
 674                 if seg.flushing != nil && !sync {
 675                         // Let the other flushing goroutine finish. If
 676                         // it fails, we'll try again next time.
 677                         return nil
 678                 } else {
 679                         // In sync mode, we proceed regardless of
 680                         // whether another flush is in progress: It
 681                         // can't finish before we do, because we hold
 682                         // fn's lock until we finish our own writes.
 683                 }
 684                 seg.flushing = done
 685                 offsets = append(offsets, len(block))
 686                 if len(refs) == 1 {
 687                         block = seg.buf
 688                 } else if block == nil {
 689                         block = append(make([]byte, 0, bufsize), seg.buf...)
 690                 } else {
 691                         block = append(block, seg.buf...)
 692                 }
 693                 segs = append(segs, seg)
 694         }
 695         blocksize := len(block)
 696         dn.fs.throttle().Acquire()
 697         errs := make(chan error, 1)
 698         go func() {
 699                 defer close(done)
 700                 defer close(errs)
 701                 locked := map[*filenode]bool{}
 702                 locator, _, err := dn.fs.PutB(block)
 703                 dn.fs.throttle().Release()
 704                 {
 705                         if !sync {
 706                                 dn.Lock()
 707                                 defer dn.Unlock()
 708                                 for _, name := range dn.sortedNames() {
 709                                         if fn, ok := dn.inodes[name].(*filenode); ok {
 710                                                 fn.Lock()
 711                                                 defer fn.Unlock()
 712                                                 locked[fn] = true
 713                                         }
 714                                 }
 715                         }
 716                         defer func() {
 717                                 for _, seg := range segs {
 718                                         if seg.flushing == done {
 719                                                 seg.flushing = nil
 720                                         }
 721                                 }
 722                         }()
 723                 }
 724                 if err != nil {
 725                         errs <- err
 726                         return
 727                 }
 728                 for idx, ref := range refs {
 729                         if !sync {
 730                                 // In async mode, fn's lock was
 731                                 // released while we were waiting for
 732                                 // PutB(); lots of things might have
 733                                 // changed.
 734                                 if len(ref.fn.segments) <= ref.idx {
 735                                         // file segments have
 736                                         // rearranged or changed in
 737                                         // some way
 738                                         continue
 739                                 } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] {
 740                                         // segment has been replaced
 741                                         continue
 742                                 } else if seg.flushing != done {
 743                                         // seg.buf has been replaced
 744                                         continue
 745                                 } else if !locked[ref.fn] {
 746                                         // file was renamed, moved, or
 747                                         // deleted since we called
 748                                         // PutB
 749                                         continue
 750                                 }
 751                         }
 752                         data := ref.fn.segments[ref.idx].(*memSegment).buf
 753                         ref.fn.segments[ref.idx] = storedSegment{
 754                                 kc:      dn.fs,
 755                                 locator: locator,
 756                                 size:    blocksize,
 757                                 offset:  offsets[idx],
 758                                 length:  len(data),
 759                         }
 760                         // atomic is needed here despite caller having
 761                         // lock: caller might be running concurrent
 762                         // commitBlock() goroutines using the same
 763                         // lock, writing different segments from the
 764                         // same file.
 765                         atomic.AddInt64(&ref.fn.memsize, -int64(len(data)))
 766                 }
 767         }()
 768         if sync {
 769                 return <-errs
 770         } else {
 771                 return nil
 772         }
 773 }
 774
 775 type flushOpts struct {
 776         sync        bool
 777         shortBlocks bool
 778 }
 779
 780 // flush in-memory data and remote-cluster block references (for the
 781 // children with the given names, which must be children of dn) to
 782 // local-cluster persistent storage.
 783 //
 784 // Caller must have write lock on dn and the named children.
 785 //
 786 // If any children are dirs, they will be flushed recursively.
 787 func (dn *dirnode) flush(ctx context.Context, names []string, opts flushOpts) error {
 788         cg := newContextGroup(ctx)
 789         defer cg.Cancel()
 790
 791         goCommit := func(refs []fnSegmentRef, bufsize int) {
 792                 cg.Go(func() error {
 793                         return dn.commitBlock(cg.Context(), refs, bufsize, opts.sync)
 794                 })
 795         }
 796
 797         var pending []fnSegmentRef
 798         var pendingLen int = 0
 799         localLocator := map[string]string{}
 800         for _, name := range names {
 801                 switch node := dn.inodes[name].(type) {
 802                 case *dirnode:
 803                         grandchildNames := node.sortedNames()
 804                         for _, grandchildName := range grandchildNames {
 805                                 grandchild := node.inodes[grandchildName]
 806                                 grandchild.Lock()
 807                                 defer grandchild.Unlock()
 808                         }
 809                         cg.Go(func() error { return node.flush(cg.Context(), grandchildNames, opts) })
 810                 case *filenode:
 811                         for idx, seg := range node.segments {
 812                                 switch seg := seg.(type) {
 813                                 case storedSegment:
 814                                         loc, ok := localLocator[seg.locator]
 815                                         if !ok {
 816                                                 var err error
 817                                                 loc, err = dn.fs.LocalLocator(seg.locator)
 818                                                 if err != nil {
 819                                                         return err
 820                                                 }
 821                                                 localLocator[seg.locator] = loc
 822                                         }
 823                                         seg.locator = loc
 824                                         node.segments[idx] = seg
 825                                 case *memSegment:
 826                                         if seg.Len() > maxBlockSize/2 {
 827                                                 goCommit([]fnSegmentRef{{node, idx}}, seg.Len())
 828                                                 continue
 829                                         }
 830                                         if pendingLen+seg.Len() > maxBlockSize {
 831                                                 goCommit(pending, pendingLen)
 832                                                 pending = nil
 833                                                 pendingLen = 0
 834                                         }
 835                                         pending = append(pending, fnSegmentRef{node, idx})
 836                                         pendingLen += seg.Len()
 837                                 default:
 838                                         panic(fmt.Sprintf("can't sync segment type %T", seg))
 839                                 }
 840                         }
 841                 }
 842         }
 843         if opts.shortBlocks {
 844                 goCommit(pending, pendingLen)
 845         }
 846         return cg.Wait()
 847 }
 848
 849 // caller must have write lock.
 850 func (dn *dirnode) memorySize() (size int64) {
 851         for _, name := range dn.sortedNames() {
 852                 node := dn.inodes[name]
 853                 node.Lock()
 854                 defer node.Unlock()
 855                 switch node := node.(type) {
 856                 case *dirnode:
 857                         size += node.memorySize()
 858                 case *filenode:
 859                         for _, seg := range node.segments {
 860                                 switch seg := seg.(type) {
 861                                 case *memSegment:
 862                                         size += int64(seg.Len())
 863                                 }
 864                         }
 865                 }
 866         }
 867         return
 868 }
 869
 870 // caller must have write lock.
 871 func (dn *dirnode) sortedNames() []string {
 872         names := make([]string, 0, len(dn.inodes))
 873         for name := range dn.inodes {
 874                 names = append(names, name)
 875         }
 876         sort.Strings(names)
 877         return names
 878 }
 879
 880 // caller must have write lock.
 881 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string) (string, error) {
 882         cg := newContextGroup(ctx)
 883         defer cg.Cancel()
 884
 885         if len(dn.inodes) == 0 {
 886                 if prefix == "." {
 887                         return "", nil
 888                 }
 889                 // Express the existence of an empty directory by
 890                 // adding an empty file named `\056`, which (unlike
 891                 // the more obvious spelling `.`) is accepted by the
 892                 // API's manifest validator.
 893                 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
 894         }
 895
 896         names := dn.sortedNames()
 897
 898         // Wait for children to finish any pending write operations
 899         // before locking them.
 900         for _, name := range names {
 901                 node := dn.inodes[name]
 902                 if fn, ok := node.(*filenode); ok {
 903                         fn.waitPrune()
 904                 }
 905         }
 906
 907         var dirnames []string
 908         var filenames []string
 909         for _, name := range names {
 910                 node := dn.inodes[name]
 911                 node.Lock()
 912                 defer node.Unlock()
 913                 switch node := node.(type) {
 914                 case *dirnode:
 915                         dirnames = append(dirnames, name)
 916                 case *filenode:
 917                         filenames = append(filenames, name)
 918                 default:
 919                         panic(fmt.Sprintf("can't marshal inode type %T", node))
 920                 }
 921         }
 922
 923         subdirs := make([]string, len(dirnames))
 924         rootdir := ""
 925         for i, name := range dirnames {
 926                 i, name := i, name
 927                 cg.Go(func() error {
 928                         txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name)
 929                         subdirs[i] = txt
 930                         return err
 931                 })
 932         }
 933
 934         cg.Go(func() error {
 935                 var streamLen int64
 936                 type filepart struct {
 937                         name   string
 938                         offset int64
 939                         length int64
 940                 }
 941
 942                 var fileparts []filepart
 943                 var blocks []string
 944                 if err := dn.flush(cg.Context(), filenames, flushOpts{sync: true, shortBlocks: true}); err != nil {
 945                         return err
 946                 }
 947                 for _, name := range filenames {
 948                         node := dn.inodes[name].(*filenode)
 949                         if len(node.segments) == 0 {
 950                                 fileparts = append(fileparts, filepart{name: name})
 951                                 continue
 952                         }
 953                         for _, seg := range node.segments {
 954                                 switch seg := seg.(type) {
 955                                 case storedSegment:
 956                                         if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
 957                                                 streamLen -= int64(seg.size)
 958                                         } else {
 959                                                 blocks = append(blocks, seg.locator)
 960                                         }
 961                                         next := filepart{
 962                                                 name:   name,
 963                                                 offset: streamLen + int64(seg.offset),
 964                                                 length: int64(seg.length),
 965                                         }
 966                                         if prev := len(fileparts) - 1; prev >= 0 &&
 967                                                 fileparts[prev].name == name &&
 968                                                 fileparts[prev].offset+fileparts[prev].length == next.offset {
 969                                                 fileparts[prev].length += next.length
 970                                         } else {
 971                                                 fileparts = append(fileparts, next)
 972                                         }
 973                                         streamLen += int64(seg.size)
 974                                 default:
 975                                         // This can't happen: we
 976                                         // haven't unlocked since
 977                                         // calling flush(sync=true).
 978                                         panic(fmt.Sprintf("can't marshal segment type %T", seg))
 979                                 }
 980                         }
 981                 }
 982                 var filetokens []string
 983                 for _, s := range fileparts {
 984                         filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
 985                 }
 986                 if len(filetokens) == 0 {
 987                         return nil
 988                 } else if len(blocks) == 0 {
 989                         blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
 990                 }
 991                 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
 992                 return nil
 993         })
 994         err := cg.Wait()
 995         return rootdir + strings.Join(subdirs, ""), err
 996 }
 997
 998 func (dn *dirnode) loadManifest(txt string) error {
 999         var dirname string
1000         streams := strings.Split(txt, "\n")
1001         if streams[len(streams)-1] != "" {
1002                 return fmt.Errorf("line %d: no trailing newline", len(streams))
1003         }
1004         streams = streams[:len(streams)-1]
1005         segments := []storedSegment{}
1006         for i, stream := range streams {
1007                 lineno := i + 1
1008                 var anyFileTokens bool
1009                 var pos int64
1010                 var segIdx int
1011                 segments = segments[:0]
1012                 for i, token := range strings.Split(stream, " ") {
1013                         if i == 0 {
1014                                 dirname = manifestUnescape(token)
1015                                 continue
1016                         }
1017                         if !strings.Contains(token, ":") {
1018                                 if anyFileTokens {
1019                                         return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1020                                 }
1021                                 toks := strings.SplitN(token, "+", 3)
1022                                 if len(toks) < 2 {
1023                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
1024                                 }
1025                                 length, err := strconv.ParseInt(toks[1], 10, 32)
1026                                 if err != nil || length < 0 {
1027                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
1028                                 }
1029                                 segments = append(segments, storedSegment{
1030                                         locator: token,
1031                                         size:    int(length),
1032                                         offset:  0,
1033                                         length:  int(length),
1034                                 })
1035                                 continue
1036                         } else if len(segments) == 0 {
1037                                 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1038                         }
1039
1040                         toks := strings.SplitN(token, ":", 3)
1041                         if len(toks) != 3 {
1042                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1043                         }
1044                         anyFileTokens = true
1045
1046                         offset, err := strconv.ParseInt(toks[0], 10, 64)
1047                         if err != nil || offset < 0 {
1048                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1049                         }
1050                         length, err := strconv.ParseInt(toks[1], 10, 64)
1051                         if err != nil || length < 0 {
1052                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1053                         }
1054                         name := dirname + "/" + manifestUnescape(toks[2])
1055                         fnode, err := dn.createFileAndParents(name)
1056                         if fnode == nil && err == nil && length == 0 {
1057                                 // Special case: an empty file used as
1058                                 // a marker to preserve an otherwise
1059                                 // empty directory in a manifest.
1060                                 continue
1061                         }
1062                         if err != nil || (fnode == nil && length != 0) {
1063                                 return fmt.Errorf("line %d: cannot use path %q with length %d: %s", lineno, name, length, err)
1064                         }
1065                         // Map the stream offset/range coordinates to
1066                         // block/offset/range coordinates and add
1067                         // corresponding storedSegments to the filenode
1068                         if pos > offset {
1069                                 // Can't continue where we left off.
1070                                 // TODO: binary search instead of
1071                                 // rewinding all the way (but this
1072                                 // situation might be rare anyway)
1073                                 segIdx, pos = 0, 0
1074                         }
1075                         for next := int64(0); segIdx < len(segments); segIdx++ {
1076                                 seg := segments[segIdx]
1077                                 next = pos + int64(seg.Len())
1078                                 if next <= offset || seg.Len() == 0 {
1079                                         pos = next
1080                                         continue
1081                                 }
1082                                 if pos >= offset+length {
1083                                         break
1084                                 }
1085                                 var blkOff int
1086                                 if pos < offset {
1087                                         blkOff = int(offset - pos)
1088                                 }
1089                                 blkLen := seg.Len() - blkOff
1090                                 if pos+int64(blkOff+blkLen) > offset+length {
1091                                         blkLen = int(offset + length - pos - int64(blkOff))
1092                                 }
1093                                 fnode.appendSegment(storedSegment{
1094                                         kc:      dn.fs,
1095                                         locator: seg.locator,
1096                                         size:    seg.size,
1097                                         offset:  blkOff,
1098                                         length:  blkLen,
1099                                 })
1100                                 if next > offset+length {
1101                                         break
1102                                 } else {
1103                                         pos = next
1104                                 }
1105                         }
1106                         if segIdx == len(segments) && pos < offset+length {
1107                                 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1108                         }
1109                 }
1110                 if !anyFileTokens {
1111                         return fmt.Errorf("line %d: no file segments", lineno)
1112                 } else if len(segments) == 0 {
1113                         return fmt.Errorf("line %d: no locators", lineno)
1114                 } else if dirname == "" {
1115                         return fmt.Errorf("line %d: no stream name", lineno)
1116                 }
1117         }
1118         return nil
1119 }
1120
1121 // only safe to call from loadManifest -- no locking.
1122 //
1123 // If path is a "parent directory exists" marker (the last path
1124 // component is "."), the returned values are both nil.
1125 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
1126         var node inode = dn
1127         names := strings.Split(path, "/")
1128         basename := names[len(names)-1]
1129         for _, name := range names[:len(names)-1] {
1130                 switch name {
1131                 case "", ".":
1132                         continue
1133                 case "..":
1134                         if node == dn {
1135                                 // can't be sure parent will be a *dirnode
1136                                 return nil, ErrInvalidArgument
1137                         }
1138                         node = node.Parent()
1139                         continue
1140                 }
1141                 node, err = node.Child(name, func(child inode) (inode, error) {
1142                         if child == nil {
1143                                 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
1144                                 if err != nil {
1145                                         return nil, err
1146                                 }
1147                                 child.SetParent(node, name)
1148                                 return child, nil
1149                         } else if !child.IsDir() {
1150                                 return child, ErrFileExists
1151                         } else {
1152                                 return child, nil
1153                         }
1154                 })
1155                 if err != nil {
1156                         return
1157                 }
1158         }
1159         if basename == "." {
1160                 return
1161         } else if !permittedName(basename) {
1162                 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
1163                 return
1164         }
1165         _, err = node.Child(basename, func(child inode) (inode, error) {
1166                 switch child := child.(type) {
1167                 case nil:
1168                         child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
1169                         if err != nil {
1170                                 return nil, err
1171                         }
1172                         child.SetParent(node, basename)
1173                         fn = child.(*filenode)
1174                         return child, nil
1175                 case *filenode:
1176                         fn = child
1177                         return child, nil
1178                 case *dirnode:
1179                         return child, ErrIsDirectory
1180                 default:
1181                         return child, ErrInvalidArgument
1182                 }
1183         })
1184         return
1185 }
1186
1187 func (dn *dirnode) TreeSize() (bytes int64) {
1188         dn.RLock()
1189         defer dn.RUnlock()
1190         for _, i := range dn.inodes {
1191                 switch i := i.(type) {
1192                 case *filenode:
1193                         bytes += i.Size()
1194                 case *dirnode:
1195                         bytes += i.TreeSize()
1196                 }
1197         }
1198         return
1199 }
1200
1201 type segment interface {
1202         io.ReaderAt
1203         Len() int
1204         // Return a new segment with a subsection of the data from this
1205         // one. length<0 means length=Len()-off.
1206         Slice(off int, length int) segment
1207 }
1208
1209 type memSegment struct {
1210         buf []byte
1211         // If flushing is not nil, then a) buf is being shared by a
1212         // pruneMemSegments goroutine, and must be copied on write;
1213         // and b) the flushing channel will close when the goroutine
1214         // finishes, whether it succeeds or not.
1215         flushing <-chan struct{}
1216 }
1217
1218 func (me *memSegment) Len() int {
1219         return len(me.buf)
1220 }
1221
1222 func (me *memSegment) Slice(off, length int) segment {
1223         if length < 0 {
1224                 length = len(me.buf) - off
1225         }
1226         buf := make([]byte, length)
1227         copy(buf, me.buf[off:])
1228         return &memSegment{buf: buf}
1229 }
1230
1231 func (me *memSegment) Truncate(n int) {
1232         if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
1233                 newsize := 1024
1234                 for newsize < n {
1235                         newsize = newsize << 2
1236                 }
1237                 newbuf := make([]byte, n, newsize)
1238                 copy(newbuf, me.buf)
1239                 me.buf, me.flushing = newbuf, nil
1240         } else {
1241                 // reclaim existing capacity, and zero reclaimed part
1242                 oldlen := len(me.buf)
1243                 me.buf = me.buf[:n]
1244                 for i := oldlen; i < n; i++ {
1245                         me.buf[i] = 0
1246                 }
1247         }
1248 }
1249
1250 func (me *memSegment) WriteAt(p []byte, off int) {
1251         if off+len(p) > len(me.buf) {
1252                 panic("overflowed segment")
1253         }
1254         if me.flushing != nil {
1255                 me.buf, me.flushing = append([]byte(nil), me.buf...), nil
1256         }
1257         copy(me.buf[off:], p)
1258 }
1259
1260 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1261         if off > int64(me.Len()) {
1262                 err = io.EOF
1263                 return
1264         }
1265         n = copy(p, me.buf[int(off):])
1266         if n < len(p) {
1267                 err = io.EOF
1268         }
1269         return
1270 }
1271
1272 type storedSegment struct {
1273         kc      fsBackend
1274         locator string
1275         size    int // size of stored block (also encoded in locator)
1276         offset  int // position of segment within the stored block
1277         length  int // bytes in this segment (offset + length <= size)
1278 }
1279
1280 func (se storedSegment) Len() int {
1281         return se.length
1282 }
1283
1284 func (se storedSegment) Slice(n, size int) segment {
1285         se.offset += n
1286         se.length -= n
1287         if size >= 0 && se.length > size {
1288                 se.length = size
1289         }
1290         return se
1291 }
1292
1293 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1294         if off > int64(se.length) {
1295                 return 0, io.EOF
1296         }
1297         maxlen := se.length - int(off)
1298         if len(p) > maxlen {
1299                 p = p[:maxlen]
1300                 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1301                 if err == nil {
1302                         err = io.EOF
1303                 }
1304                 return
1305         }
1306         return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1307 }
1308
1309 func canonicalName(name string) string {
1310         name = path.Clean("/" + name)
1311         if name == "/" || name == "./" {
1312                 name = "."
1313         } else if strings.HasPrefix(name, "/") {
1314                 name = "." + name
1315         }
1316         return name
1317 }
1318
1319 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1320
1321 func manifestUnescapeFunc(seq string) string {
1322         if seq == `\\` {
1323                 return `\`
1324         }
1325         i, err := strconv.ParseUint(seq[1:], 8, 8)
1326         if err != nil {
1327                 // Invalid escape sequence: can't unescape.
1328                 return seq
1329         }
1330         return string([]byte{byte(i)})
1331 }
1332
1333 func manifestUnescape(s string) string {
1334         return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1335 }
1336
1337 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1338
1339 func manifestEscapeFunc(seq string) string {
1340         return fmt.Sprintf("\\%03o", byte(seq[0]))
1341 }
1342
1343 func manifestEscape(s string) string {
1344         return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)
1345 }