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         "time"
  20 )
  21
  22 var (
  23         maxBlockSize      = 1 << 26
  24         concurrentWriters = 4 // max goroutines writing to Keep in background and during flush()
  25 )
  26
  27 // A CollectionFileSystem is a FileSystem that can be serialized as a
  28 // manifest and stored as a collection.
  29 type CollectionFileSystem interface {
  30         FileSystem
  31
  32         // Flush all file data to Keep and return a snapshot of the
  33         // filesystem suitable for saving as (Collection)ManifestText.
  34         // Prefix (normally ".") is a top level directory, effectively
  35         // prepended to all paths in the returned manifest.
  36         MarshalManifest(prefix string) (string, error)
  37
  38         // Total data bytes in all files.
  39         Size() int64
  40
  41         // Memory consumed by buffered file data.
  42         memorySize() int64
  43 }
  44
  45 type collectionFileSystem struct {
  46         fileSystem
  47         uuid string
  48 }
  49
  50 // FileSystem returns a CollectionFileSystem for the collection.
  51 func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) {
  52         var modTime time.Time
  53         if c.ModifiedAt == nil {
  54                 modTime = time.Now()
  55         } else {
  56                 modTime = *c.ModifiedAt
  57         }
  58         fs := &collectionFileSystem{
  59                 uuid: c.UUID,
  60                 fileSystem: fileSystem{
  61                         fsBackend: keepBackend{apiClient: client, keepClient: kc},
  62                         thr:       newThrottle(concurrentWriters),
  63                 },
  64         }
  65         root := &dirnode{
  66                 fs: fs,
  67                 treenode: treenode{
  68                         fileinfo: fileinfo{
  69                                 name:    ".",
  70                                 mode:    os.ModeDir | 0755,
  71                                 modTime: modTime,
  72                         },
  73                         inodes: make(map[string]inode),
  74                 },
  75         }
  76         root.SetParent(root, ".")
  77         if err := root.loadManifest(c.ManifestText); err != nil {
  78                 return nil, err
  79         }
  80         backdateTree(root, modTime)
  81         fs.root = root
  82         return fs, nil
  83 }
  84
  85 func backdateTree(n inode, modTime time.Time) {
  86         switch n := n.(type) {
  87         case *filenode:
  88                 n.fileinfo.modTime = modTime
  89         case *dirnode:
  90                 n.fileinfo.modTime = modTime
  91                 for _, n := range n.inodes {
  92                         backdateTree(n, modTime)
  93                 }
  94         }
  95 }
  96
  97 func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime time.Time) (node inode, err error) {
  98         if name == "" || name == "." || name == ".." {
  99                 return nil, ErrInvalidArgument
 100         }
 101         if perm.IsDir() {
 102                 return &dirnode{
 103                         fs: fs,
 104                         treenode: treenode{
 105                                 fileinfo: fileinfo{
 106                                         name:    name,
 107                                         mode:    perm | os.ModeDir,
 108                                         modTime: modTime,
 109                                 },
 110                                 inodes: make(map[string]inode),
 111                         },
 112                 }, nil
 113         } else {
 114                 return &filenode{
 115                         fs: fs,
 116                         fileinfo: fileinfo{
 117                                 name:    name,
 118                                 mode:    perm & ^os.ModeDir,
 119                                 modTime: modTime,
 120                         },
 121                 }, nil
 122         }
 123 }
 124
 125 func (fs *collectionFileSystem) Sync() error {
 126         if fs.uuid == "" {
 127                 return nil
 128         }
 129         txt, err := fs.MarshalManifest(".")
 130         if err != nil {
 131                 return fmt.Errorf("sync failed: %s", err)
 132         }
 133         coll := &Collection{
 134                 UUID:         fs.uuid,
 135                 ManifestText: txt,
 136         }
 137         err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
 138                 "collection": map[string]string{
 139                         "manifest_text": coll.ManifestText,
 140                 },
 141                 "select": []string{"uuid"},
 142         })
 143         if err != nil {
 144                 return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
 145         }
 146         return nil
 147 }
 148
 149 func (fs *collectionFileSystem) Flush(path string, shortBlocks bool) error {
 150         node, err := rlookup(fs.fileSystem.root, path)
 151         if err != nil {
 152                 return err
 153         }
 154         dn, ok := node.(*dirnode)
 155         if !ok {
 156                 return ErrNotADirectory
 157         }
 158         dn.Lock()
 159         defer dn.Unlock()
 160         names := dn.sortedNames()
 161         if path != "" {
 162                 // Caller only wants to flush the specified dir,
 163                 // non-recursively.  Drop subdirs from the list of
 164                 // names.
 165                 var filenames []string
 166                 for _, name := range names {
 167                         if _, ok := dn.inodes[name].(*filenode); ok {
 168                                 filenames = append(filenames, name)
 169                         }
 170                 }
 171                 names = filenames
 172         }
 173         return dn.flush(context.TODO(), names, flushOpts{sync: false, shortBlocks: shortBlocks})
 174 }
 175
 176 func (fs *collectionFileSystem) memorySize() int64 {
 177         fs.fileSystem.root.Lock()
 178         defer fs.fileSystem.root.Unlock()
 179         return fs.fileSystem.root.(*dirnode).memorySize()
 180 }
 181
 182 func (fs *collectionFileSystem) MarshalManifest(prefix string) (string, error) {
 183         fs.fileSystem.root.Lock()
 184         defer fs.fileSystem.root.Unlock()
 185         return fs.fileSystem.root.(*dirnode).marshalManifest(context.TODO(), prefix)
 186 }
 187
 188 func (fs *collectionFileSystem) Size() int64 {
 189         return fs.fileSystem.root.(*dirnode).TreeSize()
 190 }
 191
 192 // filenodePtr is an offset into a file that is (usually) efficient to
 193 // seek to. Specifically, if filenode.repacked==filenodePtr.repacked
 194 // then
 195 // filenode.segments[filenodePtr.segmentIdx][filenodePtr.segmentOff]
 196 // corresponds to file offset filenodePtr.off. Otherwise, it is
 197 // necessary to reexamine len(filenode.segments[0]) etc. to find the
 198 // correct segment and offset.
 199 type filenodePtr struct {
 200         off        int64
 201         segmentIdx int
 202         segmentOff int
 203         repacked   int64
 204 }
 205
 206 // seek returns a ptr that is consistent with both startPtr.off and
 207 // the current state of fn. The caller must already hold fn.RLock() or
 208 // fn.Lock().
 209 //
 210 // If startPtr is beyond EOF, ptr.segment* will indicate precisely
 211 // EOF.
 212 //
 213 // After seeking:
 214 //
 215 //     ptr.segmentIdx == len(filenode.segments) // i.e., at EOF
 216 //     ||
 217 //     filenode.segments[ptr.segmentIdx].Len() > ptr.segmentOff
 218 func (fn *filenode) seek(startPtr filenodePtr) (ptr filenodePtr) {
 219         ptr = startPtr
 220         if ptr.off < 0 {
 221                 // meaningless anyway
 222                 return
 223         } else if ptr.off >= fn.fileinfo.size {
 224                 ptr.segmentIdx = len(fn.segments)
 225                 ptr.segmentOff = 0
 226                 ptr.repacked = fn.repacked
 227                 return
 228         } else if ptr.repacked == fn.repacked {
 229                 // segmentIdx and segmentOff accurately reflect
 230                 // ptr.off, but might have fallen off the end of a
 231                 // segment
 232                 if ptr.segmentOff >= fn.segments[ptr.segmentIdx].Len() {
 233                         ptr.segmentIdx++
 234                         ptr.segmentOff = 0
 235                 }
 236                 return
 237         }
 238         defer func() {
 239                 ptr.repacked = fn.repacked
 240         }()
 241         if ptr.off >= fn.fileinfo.size {
 242                 ptr.segmentIdx, ptr.segmentOff = len(fn.segments), 0
 243                 return
 244         }
 245         // Recompute segmentIdx and segmentOff.  We have already
 246         // established fn.fileinfo.size > ptr.off >= 0, so we don't
 247         // have to deal with edge cases here.
 248         var off int64
 249         for ptr.segmentIdx, ptr.segmentOff = 0, 0; off < ptr.off; ptr.segmentIdx++ {
 250                 // This would panic (index out of range) if
 251                 // fn.fileinfo.size were larger than
 252                 // sum(fn.segments[i].Len()) -- but that can't happen
 253                 // because we have ensured fn.fileinfo.size is always
 254                 // accurate.
 255                 segLen := int64(fn.segments[ptr.segmentIdx].Len())
 256                 if off+segLen > ptr.off {
 257                         ptr.segmentOff = int(ptr.off - off)
 258                         break
 259                 }
 260                 off += segLen
 261         }
 262         return
 263 }
 264
 265 // filenode implements inode.
 266 type filenode struct {
 267         parent   inode
 268         fs       FileSystem
 269         fileinfo fileinfo
 270         segments []segment
 271         // number of times `segments` has changed in a
 272         // way that might invalidate a filenodePtr
 273         repacked int64
 274         memsize  int64 // bytes in memSegments
 275         sync.RWMutex
 276         nullnode
 277 }
 278
 279 // caller must have lock
 280 func (fn *filenode) appendSegment(e segment) {
 281         fn.segments = append(fn.segments, e)
 282         fn.fileinfo.size += int64(e.Len())
 283 }
 284
 285 func (fn *filenode) SetParent(p inode, name string) {
 286         fn.Lock()
 287         defer fn.Unlock()
 288         fn.parent = p
 289         fn.fileinfo.name = name
 290 }
 291
 292 func (fn *filenode) Parent() inode {
 293         fn.RLock()
 294         defer fn.RUnlock()
 295         return fn.parent
 296 }
 297
 298 func (fn *filenode) FS() FileSystem {
 299         return fn.fs
 300 }
 301
 302 // Read reads file data from a single segment, starting at startPtr,
 303 // into p. startPtr is assumed not to be up-to-date. Caller must have
 304 // RLock or Lock.
 305 func (fn *filenode) Read(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 306         ptr = fn.seek(startPtr)
 307         if ptr.off < 0 {
 308                 err = ErrNegativeOffset
 309                 return
 310         }
 311         if ptr.segmentIdx >= len(fn.segments) {
 312                 err = io.EOF
 313                 return
 314         }
 315         n, err = fn.segments[ptr.segmentIdx].ReadAt(p, int64(ptr.segmentOff))
 316         if n > 0 {
 317                 ptr.off += int64(n)
 318                 ptr.segmentOff += n
 319                 if ptr.segmentOff == fn.segments[ptr.segmentIdx].Len() {
 320                         ptr.segmentIdx++
 321                         ptr.segmentOff = 0
 322                         if ptr.segmentIdx < len(fn.segments) && err == io.EOF {
 323                                 err = nil
 324                         }
 325                 }
 326         }
 327         return
 328 }
 329
 330 func (fn *filenode) Size() int64 {
 331         fn.RLock()
 332         defer fn.RUnlock()
 333         return fn.fileinfo.Size()
 334 }
 335
 336 func (fn *filenode) FileInfo() os.FileInfo {
 337         fn.RLock()
 338         defer fn.RUnlock()
 339         return fn.fileinfo
 340 }
 341
 342 func (fn *filenode) Truncate(size int64) error {
 343         fn.Lock()
 344         defer fn.Unlock()
 345         return fn.truncate(size)
 346 }
 347
 348 func (fn *filenode) truncate(size int64) error {
 349         if size == fn.fileinfo.size {
 350                 return nil
 351         }
 352         fn.repacked++
 353         if size < fn.fileinfo.size {
 354                 ptr := fn.seek(filenodePtr{off: size})
 355                 for i := ptr.segmentIdx; i < len(fn.segments); i++ {
 356                         if seg, ok := fn.segments[i].(*memSegment); ok {
 357                                 fn.memsize -= int64(seg.Len())
 358                         }
 359                 }
 360                 if ptr.segmentOff == 0 {
 361                         fn.segments = fn.segments[:ptr.segmentIdx]
 362                 } else {
 363                         fn.segments = fn.segments[:ptr.segmentIdx+1]
 364                         switch seg := fn.segments[ptr.segmentIdx].(type) {
 365                         case *memSegment:
 366                                 seg.Truncate(ptr.segmentOff)
 367                                 fn.memsize += int64(seg.Len())
 368                         default:
 369                                 fn.segments[ptr.segmentIdx] = seg.Slice(0, ptr.segmentOff)
 370                         }
 371                 }
 372                 fn.fileinfo.size = size
 373                 return nil
 374         }
 375         for size > fn.fileinfo.size {
 376                 grow := size - fn.fileinfo.size
 377                 var seg *memSegment
 378                 var ok bool
 379                 if len(fn.segments) == 0 {
 380                         seg = &memSegment{}
 381                         fn.segments = append(fn.segments, seg)
 382                 } else if seg, ok = fn.segments[len(fn.segments)-1].(*memSegment); !ok || seg.Len() >= maxBlockSize {
 383                         seg = &memSegment{}
 384                         fn.segments = append(fn.segments, seg)
 385                 }
 386                 if maxgrow := int64(maxBlockSize - seg.Len()); maxgrow < grow {
 387                         grow = maxgrow
 388                 }
 389                 seg.Truncate(seg.Len() + int(grow))
 390                 fn.fileinfo.size += grow
 391                 fn.memsize += grow
 392         }
 393         return nil
 394 }
 395
 396 // Write writes data from p to the file, starting at startPtr,
 397 // extending the file size if necessary. Caller must have Lock.
 398 func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePtr, err error) {
 399         if startPtr.off > fn.fileinfo.size {
 400                 if err = fn.truncate(startPtr.off); err != nil {
 401                         return 0, startPtr, err
 402                 }
 403         }
 404         ptr = fn.seek(startPtr)
 405         if ptr.off < 0 {
 406                 err = ErrNegativeOffset
 407                 return
 408         }
 409         for len(p) > 0 && err == nil {
 410                 cando := p
 411                 if len(cando) > maxBlockSize {
 412                         cando = cando[:maxBlockSize]
 413                 }
 414                 // Rearrange/grow fn.segments (and shrink cando if
 415                 // needed) such that cando can be copied to
 416                 // fn.segments[ptr.segmentIdx] at offset
 417                 // ptr.segmentOff.
 418                 cur := ptr.segmentIdx
 419                 prev := ptr.segmentIdx - 1
 420                 var curWritable bool
 421                 if cur < len(fn.segments) {
 422                         _, curWritable = fn.segments[cur].(*memSegment)
 423                 }
 424                 var prevAppendable bool
 425                 if prev >= 0 && fn.segments[prev].Len() < maxBlockSize {
 426                         _, prevAppendable = fn.segments[prev].(*memSegment)
 427                 }
 428                 if ptr.segmentOff > 0 && !curWritable {
 429                         // Split a non-writable block.
 430                         if max := fn.segments[cur].Len() - ptr.segmentOff; max <= len(cando) {
 431                                 // Truncate cur, and insert a new
 432                                 // segment after it.
 433                                 cando = cando[:max]
 434                                 fn.segments = append(fn.segments, nil)
 435                                 copy(fn.segments[cur+1:], fn.segments[cur:])
 436                         } else {
 437                                 // Split cur into two copies, truncate
 438                                 // the one on the left, shift the one
 439                                 // on the right, and insert a new
 440                                 // segment between them.
 441                                 fn.segments = append(fn.segments, nil, nil)
 442                                 copy(fn.segments[cur+2:], fn.segments[cur:])
 443                                 fn.segments[cur+2] = fn.segments[cur+2].Slice(ptr.segmentOff+len(cando), -1)
 444                         }
 445                         cur++
 446                         prev++
 447                         seg := &memSegment{}
 448                         seg.Truncate(len(cando))
 449                         fn.memsize += int64(len(cando))
 450                         fn.segments[cur] = seg
 451                         fn.segments[prev] = fn.segments[prev].Slice(0, ptr.segmentOff)
 452                         ptr.segmentIdx++
 453                         ptr.segmentOff = 0
 454                         fn.repacked++
 455                         ptr.repacked++
 456                 } else if curWritable {
 457                         if fit := int(fn.segments[cur].Len()) - ptr.segmentOff; fit < len(cando) {
 458                                 cando = cando[:fit]
 459                         }
 460                 } else {
 461                         if prevAppendable {
 462                                 // Shrink cando if needed to fit in
 463                                 // prev segment.
 464                                 if cangrow := maxBlockSize - fn.segments[prev].Len(); cangrow < len(cando) {
 465                                         cando = cando[:cangrow]
 466                                 }
 467                         }
 468
 469                         if cur == len(fn.segments) {
 470                                 // ptr is at EOF, filesize is changing.
 471                                 fn.fileinfo.size += int64(len(cando))
 472                         } else if el := fn.segments[cur].Len(); el <= len(cando) {
 473                                 // cando is long enough that we won't
 474                                 // need cur any more. shrink cando to
 475                                 // be exactly as long as cur
 476                                 // (otherwise we'd accidentally shift
 477                                 // the effective position of all
 478                                 // segments after cur).
 479                                 cando = cando[:el]
 480                                 copy(fn.segments[cur:], fn.segments[cur+1:])
 481                                 fn.segments = fn.segments[:len(fn.segments)-1]
 482                         } else {
 483                                 // shrink cur by the same #bytes we're growing prev
 484                                 fn.segments[cur] = fn.segments[cur].Slice(len(cando), -1)
 485                         }
 486
 487                         if prevAppendable {
 488                                 // Grow prev.
 489                                 ptr.segmentIdx--
 490                                 ptr.segmentOff = fn.segments[prev].Len()
 491                                 fn.segments[prev].(*memSegment).Truncate(ptr.segmentOff + len(cando))
 492                                 fn.memsize += int64(len(cando))
 493                                 ptr.repacked++
 494                                 fn.repacked++
 495                         } else {
 496                                 // Insert a segment between prev and
 497                                 // cur, and advance prev/cur.
 498                                 fn.segments = append(fn.segments, nil)
 499                                 if cur < len(fn.segments) {
 500                                         copy(fn.segments[cur+1:], fn.segments[cur:])
 501                                         ptr.repacked++
 502                                         fn.repacked++
 503                                 } else {
 504                                         // appending a new segment does
 505                                         // not invalidate any ptrs
 506                                 }
 507                                 seg := &memSegment{}
 508                                 seg.Truncate(len(cando))
 509                                 fn.memsize += int64(len(cando))
 510                                 fn.segments[cur] = seg
 511                                 cur++
 512                                 prev++
 513                         }
 514                 }
 515
 516                 // Finally we can copy bytes from cando to the current segment.
 517                 fn.segments[ptr.segmentIdx].(*memSegment).WriteAt(cando, ptr.segmentOff)
 518                 n += len(cando)
 519                 p = p[len(cando):]
 520
 521                 ptr.off += int64(len(cando))
 522                 ptr.segmentOff += len(cando)
 523                 if ptr.segmentOff >= maxBlockSize {
 524                         fn.pruneMemSegments()
 525                 }
 526                 if fn.segments[ptr.segmentIdx].Len() == ptr.segmentOff {
 527                         ptr.segmentOff = 0
 528                         ptr.segmentIdx++
 529                 }
 530
 531                 fn.fileinfo.modTime = time.Now()
 532         }
 533         return
 534 }
 535
 536 // Write some data out to disk to reduce memory use. Caller must have
 537 // write lock.
 538 func (fn *filenode) pruneMemSegments() {
 539         // TODO: share code with (*dirnode)flush()
 540         // TODO: pack/flush small blocks too, when fragmented
 541         for idx, seg := range fn.segments {
 542                 seg, ok := seg.(*memSegment)
 543                 if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
 544                         continue
 545                 }
 546                 // Setting seg.flushing guarantees seg.buf will not be
 547                 // modified in place: WriteAt and Truncate will
 548                 // allocate a new buf instead, if necessary.
 549                 idx, buf := idx, seg.buf
 550                 done := make(chan struct{})
 551                 seg.flushing = done
 552                 // If lots of background writes are already in
 553                 // progress, block here until one finishes, rather
 554                 // than pile up an unlimited number of buffered writes
 555                 // and network flush operations.
 556                 fn.fs.throttle().Acquire()
 557                 go func() {
 558                         defer close(done)
 559                         locator, _, err := fn.FS().PutB(buf)
 560                         fn.fs.throttle().Release()
 561                         fn.Lock()
 562                         defer fn.Unlock()
 563                         if seg.flushing != done {
 564                                 // A new seg.buf has been allocated.
 565                                 return
 566                         }
 567                         seg.flushing = nil
 568                         if err != nil {
 569                                 // TODO: stall (or return errors from)
 570                                 // subsequent writes until flushing
 571                                 // starts to succeed.
 572                                 return
 573                         }
 574                         if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
 575                                 // Segment has been dropped/moved/resized.
 576                                 return
 577                         }
 578                         fn.memsize -= int64(len(buf))
 579                         fn.segments[idx] = storedSegment{
 580                                 kc:      fn.FS(),
 581                                 locator: locator,
 582                                 size:    len(buf),
 583                                 offset:  0,
 584                                 length:  len(buf),
 585                         }
 586                 }()
 587         }
 588 }
 589
 590 // Block until all pending pruneMemSegments/flush work is
 591 // finished. Caller must NOT have lock.
 592 func (fn *filenode) waitPrune() {
 593         var pending []<-chan struct{}
 594         fn.Lock()
 595         for _, seg := range fn.segments {
 596                 if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
 597                         pending = append(pending, seg.flushing)
 598                 }
 599         }
 600         fn.Unlock()
 601         for _, p := range pending {
 602                 <-p
 603         }
 604 }
 605
 606 type dirnode struct {
 607         fs *collectionFileSystem
 608         treenode
 609 }
 610
 611 func (dn *dirnode) FS() FileSystem {
 612         return dn.fs
 613 }
 614
 615 func (dn *dirnode) Child(name string, replace func(inode) (inode, error)) (inode, error) {
 616         if dn == dn.fs.rootnode() && name == ".arvados#collection" {
 617                 gn := &getternode{Getter: func() ([]byte, error) {
 618                         var coll Collection
 619                         var err error
 620                         coll.ManifestText, err = dn.fs.MarshalManifest(".")
 621                         if err != nil {
 622                                 return nil, err
 623                         }
 624                         data, err := json.Marshal(&coll)
 625                         if err == nil {
 626                                 data = append(data, '\n')
 627                         }
 628                         return data, err
 629                 }}
 630                 gn.SetParent(dn, name)
 631                 return gn, nil
 632         }
 633         return dn.treenode.Child(name, replace)
 634 }
 635
 636 type fnSegmentRef struct {
 637         fn  *filenode
 638         idx int
 639 }
 640
 641 // commitBlock concatenates the data from the given filenode segments
 642 // (which must be *memSegments), writes the data out to Keep as a
 643 // single block, and replaces the filenodes' *memSegments with
 644 // storedSegments that reference the relevant portions of the new
 645 // block.
 646 //
 647 // If sync is false, commitBlock returns right away, after starting a
 648 // goroutine to do the writes, reacquire the filenodes' locks, and
 649 // swap out the *memSegments. Some filenodes' segments might get
 650 // modified/rearranged in the meantime, in which case commitBlock
 651 // won't replace them.
 652 //
 653 // Caller must have write lock.
 654 func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, sync bool) error {
 655         if err := ctx.Err(); err != nil {
 656                 return err
 657         }
 658         done := make(chan struct{})
 659         block := make([]byte, 0, maxBlockSize)
 660         segs := make([]*memSegment, 0, len(refs))
 661         offsets := make([]int, 0, len(refs)) // location of segment's data within block
 662         for _, ref := range refs {
 663                 seg := ref.fn.segments[ref.idx].(*memSegment)
 664                 if seg.flushing != nil && !sync {
 665                         // Let the other flushing goroutine finish. If
 666                         // it fails, we'll try again next time.
 667                         return nil
 668                 } else {
 669                         // In sync mode, we proceed regardless of
 670                         // whether another flush is in progress: It
 671                         // can't finish before we do, because we hold
 672                         // fn's lock until we finish our own writes.
 673                 }
 674                 seg.flushing = done
 675                 offsets = append(offsets, len(block))
 676                 block = append(block, seg.buf...)
 677                 segs = append(segs, seg)
 678         }
 679         dn.fs.throttle().Acquire()
 680         errs := make(chan error, 1)
 681         go func() {
 682                 defer close(done)
 683                 defer close(errs)
 684                 locked := map[*filenode]bool{}
 685                 locator, _, err := dn.fs.PutB(block)
 686                 dn.fs.throttle().Release()
 687                 {
 688                         if !sync {
 689                                 for _, name := range dn.sortedNames() {
 690                                         if fn, ok := dn.inodes[name].(*filenode); ok {
 691                                                 fn.Lock()
 692                                                 defer fn.Unlock()
 693                                                 locked[fn] = true
 694                                         }
 695                                 }
 696                         }
 697                         defer func() {
 698                                 for _, seg := range segs {
 699                                         if seg.flushing == done {
 700                                                 seg.flushing = nil
 701                                         }
 702                                 }
 703                         }()
 704                 }
 705                 if err != nil {
 706                         errs <- err
 707                         return
 708                 }
 709                 for idx, ref := range refs {
 710                         if !sync {
 711                                 // In async mode, fn's lock was
 712                                 // released while we were waiting for
 713                                 // PutB(); lots of things might have
 714                                 // changed.
 715                                 if len(ref.fn.segments) <= ref.idx {
 716                                         // file segments have
 717                                         // rearranged or changed in
 718                                         // some way
 719                                         continue
 720                                 } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] {
 721                                         // segment has been replaced
 722                                         continue
 723                                 } else if seg.flushing != done {
 724                                         // seg.buf has been replaced
 725                                         continue
 726                                 } else if !locked[ref.fn] {
 727                                         // file was renamed, moved, or
 728                                         // deleted since we called
 729                                         // PutB
 730                                         continue
 731                                 }
 732                         }
 733                         data := ref.fn.segments[ref.idx].(*memSegment).buf
 734                         ref.fn.segments[ref.idx] = storedSegment{
 735                                 kc:      dn.fs,
 736                                 locator: locator,
 737                                 size:    len(block),
 738                                 offset:  offsets[idx],
 739                                 length:  len(data),
 740                         }
 741                         ref.fn.memsize -= int64(len(data))
 742                 }
 743         }()
 744         if sync {
 745                 return <-errs
 746         } else {
 747                 return nil
 748         }
 749 }
 750
 751 type flushOpts struct {
 752         sync        bool
 753         shortBlocks bool
 754 }
 755
 756 // flush in-memory data and remote-cluster block references (for the
 757 // children with the given names, which must be children of dn) to
 758 // local-cluster persistent storage.
 759 //
 760 // Caller must have write lock on dn and the named children.
 761 //
 762 // If any children are dirs, they will be flushed recursively.
 763 func (dn *dirnode) flush(ctx context.Context, names []string, opts flushOpts) error {
 764         cg := newContextGroup(ctx)
 765         defer cg.Cancel()
 766
 767         goCommit := func(refs []fnSegmentRef) {
 768                 if len(refs) == 0 {
 769                         return
 770                 }
 771                 cg.Go(func() error {
 772                         return dn.commitBlock(cg.Context(), refs, opts.sync)
 773                 })
 774         }
 775
 776         var pending []fnSegmentRef
 777         var pendingLen int = 0
 778         localLocator := map[string]string{}
 779         for _, name := range names {
 780                 switch node := dn.inodes[name].(type) {
 781                 case *dirnode:
 782                         grandchildNames := node.sortedNames()
 783                         for _, grandchildName := range grandchildNames {
 784                                 grandchild := node.inodes[grandchildName]
 785                                 grandchild.Lock()
 786                                 defer grandchild.Unlock()
 787                         }
 788                         cg.Go(func() error { return node.flush(cg.Context(), grandchildNames, opts) })
 789                 case *filenode:
 790                         for idx, seg := range node.segments {
 791                                 switch seg := seg.(type) {
 792                                 case storedSegment:
 793                                         loc, ok := localLocator[seg.locator]
 794                                         if !ok {
 795                                                 var err error
 796                                                 loc, err = dn.fs.LocalLocator(seg.locator)
 797                                                 if err != nil {
 798                                                         return err
 799                                                 }
 800                                                 localLocator[seg.locator] = loc
 801                                         }
 802                                         seg.locator = loc
 803                                         node.segments[idx] = seg
 804                                 case *memSegment:
 805                                         if seg.Len() > maxBlockSize/2 {
 806                                                 goCommit([]fnSegmentRef{{node, idx}})
 807                                                 continue
 808                                         }
 809                                         if pendingLen+seg.Len() > maxBlockSize {
 810                                                 goCommit(pending)
 811                                                 pending = nil
 812                                                 pendingLen = 0
 813                                         }
 814                                         pending = append(pending, fnSegmentRef{node, idx})
 815                                         pendingLen += seg.Len()
 816                                 default:
 817                                         panic(fmt.Sprintf("can't sync segment type %T", seg))
 818                                 }
 819                         }
 820                 }
 821         }
 822         if opts.shortBlocks {
 823                 goCommit(pending)
 824         }
 825         return cg.Wait()
 826 }
 827
 828 // caller must have write lock.
 829 func (dn *dirnode) memorySize() (size int64) {
 830         for _, name := range dn.sortedNames() {
 831                 node := dn.inodes[name]
 832                 node.Lock()
 833                 defer node.Unlock()
 834                 switch node := node.(type) {
 835                 case *dirnode:
 836                         size += node.memorySize()
 837                 case *filenode:
 838                         for _, seg := range node.segments {
 839                                 switch seg := seg.(type) {
 840                                 case *memSegment:
 841                                         size += int64(seg.Len())
 842                                 }
 843                         }
 844                 }
 845         }
 846         return
 847 }
 848
 849 // caller must have write lock.
 850 func (dn *dirnode) sortedNames() []string {
 851         names := make([]string, 0, len(dn.inodes))
 852         for name := range dn.inodes {
 853                 names = append(names, name)
 854         }
 855         sort.Strings(names)
 856         return names
 857 }
 858
 859 // caller must have write lock.
 860 func (dn *dirnode) marshalManifest(ctx context.Context, prefix string) (string, error) {
 861         cg := newContextGroup(ctx)
 862         defer cg.Cancel()
 863
 864         if len(dn.inodes) == 0 {
 865                 if prefix == "." {
 866                         return "", nil
 867                 }
 868                 // Express the existence of an empty directory by
 869                 // adding an empty file named `\056`, which (unlike
 870                 // the more obvious spelling `.`) is accepted by the
 871                 // API's manifest validator.
 872                 return manifestEscape(prefix) + " d41d8cd98f00b204e9800998ecf8427e+0 0:0:\\056\n", nil
 873         }
 874
 875         names := dn.sortedNames()
 876
 877         // Wait for children to finish any pending write operations
 878         // before locking them.
 879         for _, name := range names {
 880                 node := dn.inodes[name]
 881                 if fn, ok := node.(*filenode); ok {
 882                         fn.waitPrune()
 883                 }
 884         }
 885
 886         var dirnames []string
 887         var filenames []string
 888         for _, name := range names {
 889                 node := dn.inodes[name]
 890                 node.Lock()
 891                 defer node.Unlock()
 892                 switch node := node.(type) {
 893                 case *dirnode:
 894                         dirnames = append(dirnames, name)
 895                 case *filenode:
 896                         filenames = append(filenames, name)
 897                 default:
 898                         panic(fmt.Sprintf("can't marshal inode type %T", node))
 899                 }
 900         }
 901
 902         subdirs := make([]string, len(dirnames))
 903         rootdir := ""
 904         for i, name := range dirnames {
 905                 i, name := i, name
 906                 cg.Go(func() error {
 907                         txt, err := dn.inodes[name].(*dirnode).marshalManifest(cg.Context(), prefix+"/"+name)
 908                         subdirs[i] = txt
 909                         return err
 910                 })
 911         }
 912
 913         cg.Go(func() error {
 914                 var streamLen int64
 915                 type filepart struct {
 916                         name   string
 917                         offset int64
 918                         length int64
 919                 }
 920
 921                 var fileparts []filepart
 922                 var blocks []string
 923                 if err := dn.flush(cg.Context(), filenames, flushOpts{sync: true, shortBlocks: true}); err != nil {
 924                         return err
 925                 }
 926                 for _, name := range filenames {
 927                         node := dn.inodes[name].(*filenode)
 928                         if len(node.segments) == 0 {
 929                                 fileparts = append(fileparts, filepart{name: name})
 930                                 continue
 931                         }
 932                         for _, seg := range node.segments {
 933                                 switch seg := seg.(type) {
 934                                 case storedSegment:
 935                                         if len(blocks) > 0 && blocks[len(blocks)-1] == seg.locator {
 936                                                 streamLen -= int64(seg.size)
 937                                         } else {
 938                                                 blocks = append(blocks, seg.locator)
 939                                         }
 940                                         next := filepart{
 941                                                 name:   name,
 942                                                 offset: streamLen + int64(seg.offset),
 943                                                 length: int64(seg.length),
 944                                         }
 945                                         if prev := len(fileparts) - 1; prev >= 0 &&
 946                                                 fileparts[prev].name == name &&
 947                                                 fileparts[prev].offset+fileparts[prev].length == next.offset {
 948                                                 fileparts[prev].length += next.length
 949                                         } else {
 950                                                 fileparts = append(fileparts, next)
 951                                         }
 952                                         streamLen += int64(seg.size)
 953                                 default:
 954                                         // This can't happen: we
 955                                         // haven't unlocked since
 956                                         // calling flush(sync=true).
 957                                         panic(fmt.Sprintf("can't marshal segment type %T", seg))
 958                                 }
 959                         }
 960                 }
 961                 var filetokens []string
 962                 for _, s := range fileparts {
 963                         filetokens = append(filetokens, fmt.Sprintf("%d:%d:%s", s.offset, s.length, manifestEscape(s.name)))
 964                 }
 965                 if len(filetokens) == 0 {
 966                         return nil
 967                 } else if len(blocks) == 0 {
 968                         blocks = []string{"d41d8cd98f00b204e9800998ecf8427e+0"}
 969                 }
 970                 rootdir = manifestEscape(prefix) + " " + strings.Join(blocks, " ") + " " + strings.Join(filetokens, " ") + "\n"
 971                 return nil
 972         })
 973         err := cg.Wait()
 974         return rootdir + strings.Join(subdirs, ""), err
 975 }
 976
 977 func (dn *dirnode) loadManifest(txt string) error {
 978         var dirname string
 979         streams := strings.Split(txt, "\n")
 980         if streams[len(streams)-1] != "" {
 981                 return fmt.Errorf("line %d: no trailing newline", len(streams))
 982         }
 983         streams = streams[:len(streams)-1]
 984         segments := []storedSegment{}
 985         for i, stream := range streams {
 986                 lineno := i + 1
 987                 var anyFileTokens bool
 988                 var pos int64
 989                 var segIdx int
 990                 segments = segments[:0]
 991                 for i, token := range strings.Split(stream, " ") {
 992                         if i == 0 {
 993                                 dirname = manifestUnescape(token)
 994                                 continue
 995                         }
 996                         if !strings.Contains(token, ":") {
 997                                 if anyFileTokens {
 998                                         return fmt.Errorf("line %d: bad file segment %q", lineno, token)
 999                                 }
1000                                 toks := strings.SplitN(token, "+", 3)
1001                                 if len(toks) < 2 {
1002                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
1003                                 }
1004                                 length, err := strconv.ParseInt(toks[1], 10, 32)
1005                                 if err != nil || length < 0 {
1006                                         return fmt.Errorf("line %d: bad locator %q", lineno, token)
1007                                 }
1008                                 segments = append(segments, storedSegment{
1009                                         locator: token,
1010                                         size:    int(length),
1011                                         offset:  0,
1012                                         length:  int(length),
1013                                 })
1014                                 continue
1015                         } else if len(segments) == 0 {
1016                                 return fmt.Errorf("line %d: bad locator %q", lineno, token)
1017                         }
1018
1019                         toks := strings.SplitN(token, ":", 3)
1020                         if len(toks) != 3 {
1021                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1022                         }
1023                         anyFileTokens = true
1024
1025                         offset, err := strconv.ParseInt(toks[0], 10, 64)
1026                         if err != nil || offset < 0 {
1027                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1028                         }
1029                         length, err := strconv.ParseInt(toks[1], 10, 64)
1030                         if err != nil || length < 0 {
1031                                 return fmt.Errorf("line %d: bad file segment %q", lineno, token)
1032                         }
1033                         name := dirname + "/" + manifestUnescape(toks[2])
1034                         fnode, err := dn.createFileAndParents(name)
1035                         if fnode == nil && err == nil && length == 0 {
1036                                 // Special case: an empty file used as
1037                                 // a marker to preserve an otherwise
1038                                 // empty directory in a manifest.
1039                                 continue
1040                         }
1041                         if err != nil || (fnode == nil && length != 0) {
1042                                 return fmt.Errorf("line %d: cannot use path %q with length %d: %s", lineno, name, length, err)
1043                         }
1044                         // Map the stream offset/range coordinates to
1045                         // block/offset/range coordinates and add
1046                         // corresponding storedSegments to the filenode
1047                         if pos > offset {
1048                                 // Can't continue where we left off.
1049                                 // TODO: binary search instead of
1050                                 // rewinding all the way (but this
1051                                 // situation might be rare anyway)
1052                                 segIdx, pos = 0, 0
1053                         }
1054                         for next := int64(0); segIdx < len(segments); segIdx++ {
1055                                 seg := segments[segIdx]
1056                                 next = pos + int64(seg.Len())
1057                                 if next <= offset || seg.Len() == 0 {
1058                                         pos = next
1059                                         continue
1060                                 }
1061                                 if pos >= offset+length {
1062                                         break
1063                                 }
1064                                 var blkOff int
1065                                 if pos < offset {
1066                                         blkOff = int(offset - pos)
1067                                 }
1068                                 blkLen := seg.Len() - blkOff
1069                                 if pos+int64(blkOff+blkLen) > offset+length {
1070                                         blkLen = int(offset + length - pos - int64(blkOff))
1071                                 }
1072                                 fnode.appendSegment(storedSegment{
1073                                         kc:      dn.fs,
1074                                         locator: seg.locator,
1075                                         size:    seg.size,
1076                                         offset:  blkOff,
1077                                         length:  blkLen,
1078                                 })
1079                                 if next > offset+length {
1080                                         break
1081                                 } else {
1082                                         pos = next
1083                                 }
1084                         }
1085                         if segIdx == len(segments) && pos < offset+length {
1086                                 return fmt.Errorf("line %d: invalid segment in %d-byte stream: %q", lineno, pos, token)
1087                         }
1088                 }
1089                 if !anyFileTokens {
1090                         return fmt.Errorf("line %d: no file segments", lineno)
1091                 } else if len(segments) == 0 {
1092                         return fmt.Errorf("line %d: no locators", lineno)
1093                 } else if dirname == "" {
1094                         return fmt.Errorf("line %d: no stream name", lineno)
1095                 }
1096         }
1097         return nil
1098 }
1099
1100 // only safe to call from loadManifest -- no locking.
1101 //
1102 // If path is a "parent directory exists" marker (the last path
1103 // component is "."), the returned values are both nil.
1104 func (dn *dirnode) createFileAndParents(path string) (fn *filenode, err error) {
1105         var node inode = dn
1106         names := strings.Split(path, "/")
1107         basename := names[len(names)-1]
1108         for _, name := range names[:len(names)-1] {
1109                 switch name {
1110                 case "", ".":
1111                         continue
1112                 case "..":
1113                         if node == dn {
1114                                 // can't be sure parent will be a *dirnode
1115                                 return nil, ErrInvalidArgument
1116                         }
1117                         node = node.Parent()
1118                         continue
1119                 }
1120                 node, err = node.Child(name, func(child inode) (inode, error) {
1121                         if child == nil {
1122                                 child, err := node.FS().newNode(name, 0755|os.ModeDir, node.Parent().FileInfo().ModTime())
1123                                 if err != nil {
1124                                         return nil, err
1125                                 }
1126                                 child.SetParent(node, name)
1127                                 return child, nil
1128                         } else if !child.IsDir() {
1129                                 return child, ErrFileExists
1130                         } else {
1131                                 return child, nil
1132                         }
1133                 })
1134                 if err != nil {
1135                         return
1136                 }
1137         }
1138         if basename == "." {
1139                 return
1140         } else if !permittedName(basename) {
1141                 err = fmt.Errorf("invalid file part %q in path %q", basename, path)
1142                 return
1143         }
1144         _, err = node.Child(basename, func(child inode) (inode, error) {
1145                 switch child := child.(type) {
1146                 case nil:
1147                         child, err = node.FS().newNode(basename, 0755, node.FileInfo().ModTime())
1148                         if err != nil {
1149                                 return nil, err
1150                         }
1151                         child.SetParent(node, basename)
1152                         fn = child.(*filenode)
1153                         return child, nil
1154                 case *filenode:
1155                         fn = child
1156                         return child, nil
1157                 case *dirnode:
1158                         return child, ErrIsDirectory
1159                 default:
1160                         return child, ErrInvalidArgument
1161                 }
1162         })
1163         return
1164 }
1165
1166 func (dn *dirnode) TreeSize() (bytes int64) {
1167         dn.RLock()
1168         defer dn.RUnlock()
1169         for _, i := range dn.inodes {
1170                 switch i := i.(type) {
1171                 case *filenode:
1172                         bytes += i.Size()
1173                 case *dirnode:
1174                         bytes += i.TreeSize()
1175                 }
1176         }
1177         return
1178 }
1179
1180 type segment interface {
1181         io.ReaderAt
1182         Len() int
1183         // Return a new segment with a subsection of the data from this
1184         // one. length<0 means length=Len()-off.
1185         Slice(off int, length int) segment
1186 }
1187
1188 type memSegment struct {
1189         buf []byte
1190         // If flushing is not nil, then a) buf is being shared by a
1191         // pruneMemSegments goroutine, and must be copied on write;
1192         // and b) the flushing channel will close when the goroutine
1193         // finishes, whether it succeeds or not.
1194         flushing <-chan struct{}
1195 }
1196
1197 func (me *memSegment) Len() int {
1198         return len(me.buf)
1199 }
1200
1201 func (me *memSegment) Slice(off, length int) segment {
1202         if length < 0 {
1203                 length = len(me.buf) - off
1204         }
1205         buf := make([]byte, length)
1206         copy(buf, me.buf[off:])
1207         return &memSegment{buf: buf}
1208 }
1209
1210 func (me *memSegment) Truncate(n int) {
1211         if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
1212                 newsize := 1024
1213                 for newsize < n {
1214                         newsize = newsize << 2
1215                 }
1216                 newbuf := make([]byte, n, newsize)
1217                 copy(newbuf, me.buf)
1218                 me.buf, me.flushing = newbuf, nil
1219         } else {
1220                 // reclaim existing capacity, and zero reclaimed part
1221                 oldlen := len(me.buf)
1222                 me.buf = me.buf[:n]
1223                 for i := oldlen; i < n; i++ {
1224                         me.buf[i] = 0
1225                 }
1226         }
1227 }
1228
1229 func (me *memSegment) WriteAt(p []byte, off int) {
1230         if off+len(p) > len(me.buf) {
1231                 panic("overflowed segment")
1232         }
1233         if me.flushing != nil {
1234                 me.buf, me.flushing = append([]byte(nil), me.buf...), nil
1235         }
1236         copy(me.buf[off:], p)
1237 }
1238
1239 func (me *memSegment) ReadAt(p []byte, off int64) (n int, err error) {
1240         if off > int64(me.Len()) {
1241                 err = io.EOF
1242                 return
1243         }
1244         n = copy(p, me.buf[int(off):])
1245         if n < len(p) {
1246                 err = io.EOF
1247         }
1248         return
1249 }
1250
1251 type storedSegment struct {
1252         kc      fsBackend
1253         locator string
1254         size    int // size of stored block (also encoded in locator)
1255         offset  int // position of segment within the stored block
1256         length  int // bytes in this segment (offset + length <= size)
1257 }
1258
1259 func (se storedSegment) Len() int {
1260         return se.length
1261 }
1262
1263 func (se storedSegment) Slice(n, size int) segment {
1264         se.offset += n
1265         se.length -= n
1266         if size >= 0 && se.length > size {
1267                 se.length = size
1268         }
1269         return se
1270 }
1271
1272 func (se storedSegment) ReadAt(p []byte, off int64) (n int, err error) {
1273         if off > int64(se.length) {
1274                 return 0, io.EOF
1275         }
1276         maxlen := se.length - int(off)
1277         if len(p) > maxlen {
1278                 p = p[:maxlen]
1279                 n, err = se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1280                 if err == nil {
1281                         err = io.EOF
1282                 }
1283                 return
1284         }
1285         return se.kc.ReadAt(se.locator, p, int(off)+se.offset)
1286 }
1287
1288 func canonicalName(name string) string {
1289         name = path.Clean("/" + name)
1290         if name == "/" || name == "./" {
1291                 name = "."
1292         } else if strings.HasPrefix(name, "/") {
1293                 name = "." + name
1294         }
1295         return name
1296 }
1297
1298 var manifestEscapeSeq = regexp.MustCompile(`\\([0-7]{3}|\\)`)
1299
1300 func manifestUnescapeFunc(seq string) string {
1301         if seq == `\\` {
1302                 return `\`
1303         }
1304         i, err := strconv.ParseUint(seq[1:], 8, 8)
1305         if err != nil {
1306                 // Invalid escape sequence: can't unescape.
1307                 return seq
1308         }
1309         return string([]byte{byte(i)})
1310 }
1311
1312 func manifestUnescape(s string) string {
1313         return manifestEscapeSeq.ReplaceAllStringFunc(s, manifestUnescapeFunc)
1314 }
1315
1316 var manifestEscapedChar = regexp.MustCompile(`[\000-\040:\s\\]`)
1317
1318 func manifestEscapeFunc(seq string) string {
1319         return fmt.Sprintf("\\%03o", byte(seq[0]))
1320 }
1321
1322 func manifestEscape(s string) string {
1323         return manifestEscapedChar.ReplaceAllStringFunc(s, manifestEscapeFunc)
1324 }