X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/4cadc4832ed4dd58e841628232c472e99b6f7d1b..7a7a1395490eb9b9789b79f88a77fe709a898feb:/sdk/go/arvados/fs_collection.go diff --git a/sdk/go/arvados/fs_collection.go b/sdk/go/arvados/fs_collection.go index b3e6aa96e4..060b57b493 100644 --- a/sdk/go/arvados/fs_collection.go +++ b/sdk/go/arvados/fs_collection.go @@ -16,6 +16,7 @@ import ( "strconv" "strings" "sync" + "sync/atomic" "time" ) @@ -49,11 +50,9 @@ type collectionFileSystem struct { // FileSystem returns a CollectionFileSystem for the collection. func (c *Collection) FileSystem(client apiClient, kc keepClient) (CollectionFileSystem, error) { - var modTime time.Time - if c.ModifiedAt == nil { + modTime := c.ModifiedAt + if modTime.IsZero() { modTime = time.Now() - } else { - modTime = *c.ModifiedAt } fs := &collectionFileSystem{ uuid: c.UUID, @@ -122,6 +121,62 @@ func (fs *collectionFileSystem) newNode(name string, perm os.FileMode, modTime t } } +func (fs *collectionFileSystem) Child(name string, replace func(inode) (inode, error)) (inode, error) { + return fs.rootnode().Child(name, replace) +} + +func (fs *collectionFileSystem) FS() FileSystem { + return fs +} + +func (fs *collectionFileSystem) FileInfo() os.FileInfo { + return fs.rootnode().FileInfo() +} + +func (fs *collectionFileSystem) IsDir() bool { + return true +} + +func (fs *collectionFileSystem) Lock() { + fs.rootnode().Lock() +} + +func (fs *collectionFileSystem) Unlock() { + fs.rootnode().Unlock() +} + +func (fs *collectionFileSystem) RLock() { + fs.rootnode().RLock() +} + +func (fs *collectionFileSystem) RUnlock() { + fs.rootnode().RUnlock() +} + +func (fs *collectionFileSystem) Parent() inode { + return fs.rootnode().Parent() +} + +func (fs *collectionFileSystem) Read(_ []byte, ptr filenodePtr) (int, filenodePtr, error) { + return 0, ptr, ErrInvalidOperation +} + +func (fs *collectionFileSystem) Write(_ []byte, ptr filenodePtr) (int, filenodePtr, error) { + return 0, ptr, ErrInvalidOperation +} + +func (fs *collectionFileSystem) Readdir() ([]os.FileInfo, error) { + return fs.rootnode().Readdir() +} + +func (fs *collectionFileSystem) SetParent(parent inode, name string) { + fs.rootnode().SetParent(parent, name) +} + +func (fs *collectionFileSystem) Truncate(int64) error { + return ErrInvalidOperation +} + func (fs *collectionFileSystem) Sync() error { if fs.uuid == "" { return nil @@ -513,8 +568,6 @@ func (fn *filenode) Write(p []byte, startPtr filenodePtr) (n int, ptr filenodePt seg.Truncate(len(cando)) fn.memsize += int64(len(cando)) fn.segments[cur] = seg - cur++ - prev++ } } @@ -569,7 +622,6 @@ func (fn *filenode) pruneMemSegments() { // A new seg.buf has been allocated. return } - seg.flushing = nil if err != nil { // TODO: stall (or return errors from) // subsequent writes until flushing @@ -672,9 +724,10 @@ func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize offsets := make([]int, 0, len(refs)) // location of segment's data within block for _, ref := range refs { seg := ref.fn.segments[ref.idx].(*memSegment) - if seg.flushing != nil && !sync { + if !sync && seg.flushingUnfinished() { // Let the other flushing goroutine finish. If // it fails, we'll try again next time. + close(done) return nil } else { // In sync mode, we proceed regardless of @@ -693,38 +746,21 @@ func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize } segs = append(segs, seg) } + blocksize := len(block) dn.fs.throttle().Acquire() errs := make(chan error, 1) go func() { defer close(done) defer close(errs) - locked := map[*filenode]bool{} locator, _, err := dn.fs.PutB(block) dn.fs.throttle().Release() - { - if !sync { - for _, name := range dn.sortedNames() { - if fn, ok := dn.inodes[name].(*filenode); ok { - fn.Lock() - defer fn.Unlock() - locked[fn] = true - } - } - } - defer func() { - for _, seg := range segs { - if seg.flushing == done { - seg.flushing = nil - } - } - }() - } if err != nil { errs <- err return } for idx, ref := range refs { if !sync { + ref.fn.Lock() // In async mode, fn's lock was // released while we were waiting for // PutB(); lots of things might have @@ -733,17 +769,15 @@ func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize // file segments have // rearranged or changed in // some way + ref.fn.Unlock() continue } else if seg, ok := ref.fn.segments[ref.idx].(*memSegment); !ok || seg != segs[idx] { // segment has been replaced + ref.fn.Unlock() continue } else if seg.flushing != done { // seg.buf has been replaced - continue - } else if !locked[ref.fn] { - // file was renamed, moved, or - // deleted since we called - // PutB + ref.fn.Unlock() continue } } @@ -751,11 +785,19 @@ func (dn *dirnode) commitBlock(ctx context.Context, refs []fnSegmentRef, bufsize ref.fn.segments[ref.idx] = storedSegment{ kc: dn.fs, locator: locator, - size: len(block), + size: blocksize, offset: offsets[idx], length: len(data), } - ref.fn.memsize -= int64(len(data)) + // atomic is needed here despite caller having + // lock: caller might be running concurrent + // commitBlock() goroutines using the same + // lock, writing different segments from the + // same file. + atomic.AddInt64(&ref.fn.memsize, -int64(len(data))) + if !sync { + ref.fn.Unlock() + } } }() if sync { @@ -1065,9 +1107,9 @@ func (dn *dirnode) loadManifest(txt string) error { // situation might be rare anyway) segIdx, pos = 0, 0 } - for next := int64(0); segIdx < len(segments); segIdx++ { + for ; segIdx < len(segments); segIdx++ { seg := segments[segIdx] - next = pos + int64(seg.Len()) + next := pos + int64(seg.Len()) if next <= offset || seg.Len() == 0 { pos = next continue @@ -1201,13 +1243,26 @@ type segment interface { type memSegment struct { buf []byte - // If flushing is not nil, then a) buf is being shared by a - // pruneMemSegments goroutine, and must be copied on write; - // and b) the flushing channel will close when the goroutine - // finishes, whether it succeeds or not. + // If flushing is not nil and not ready/closed, then a) buf is + // being shared by a pruneMemSegments goroutine, and must be + // copied on write; and b) the flushing channel will close + // when the goroutine finishes, whether it succeeds or not. flushing <-chan struct{} } +func (me *memSegment) flushingUnfinished() bool { + if me.flushing == nil { + return false + } + select { + case <-me.flushing: + me.flushing = nil + return false + default: + return true + } +} + func (me *memSegment) Len() int { return len(me.buf) }