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