"time"
)
-var maxBlockSize = 1 << 26
-
-var concurrentWriters = 4
+var (
+ maxBlockSize = 1 << 26
+ concurrentWriters = 4 // max goroutines writing to Keep during sync()
+ writeAheadBlocks = 1 // max background jobs flushing to Keep before blocking writes
+)
// A CollectionFileSystem is a FileSystem that can be serialized as a
// manifest and stored as a collection.
UUID: fs.uuid,
ManifestText: txt,
}
- err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, fs.UpdateBody(coll), map[string]interface{}{"select": []string{"uuid"}})
+ err = fs.RequestAndDecode(nil, "PUT", "arvados/v1/collections/"+fs.uuid, nil, map[string]interface{}{
+ "collection": map[string]string{
+ "manifest_text": coll.ManifestText,
+ },
+ "select": []string{"uuid"},
+ })
if err != nil {
return fmt.Errorf("sync failed: update %s: %s", fs.uuid, err)
}
memsize int64 // bytes in memSegments
sync.RWMutex
nullnode
+ throttle *throttle
}
// caller must have lock
// Write some data out to disk to reduce memory use. Caller must have
// write lock.
func (fn *filenode) pruneMemSegments() {
- // TODO: async (don't hold Lock() while waiting for Keep)
// TODO: share code with (*dirnode)sync()
// TODO: pack/flush small blocks too, when fragmented
+ if fn.throttle == nil {
+ // TODO: share a throttle with filesystem
+ fn.throttle = newThrottle(writeAheadBlocks)
+ }
for idx, seg := range fn.segments {
seg, ok := seg.(*memSegment)
- if !ok || seg.Len() < maxBlockSize {
- continue
- }
- locator, _, err := fn.FS().PutB(seg.buf)
- if err != nil {
- // TODO: stall (or return errors from)
- // subsequent writes until flushing
- // starts to succeed
+ if !ok || seg.Len() < maxBlockSize || seg.flushing != nil {
continue
}
- fn.memsize -= int64(seg.Len())
- fn.segments[idx] = storedSegment{
- kc: fn.FS(),
- locator: locator,
- size: seg.Len(),
- offset: 0,
- length: seg.Len(),
+ // Setting seg.flushing guarantees seg.buf will not be
+ // modified in place: WriteAt and Truncate will
+ // allocate a new buf instead, if necessary.
+ idx, buf := idx, seg.buf
+ done := make(chan struct{})
+ seg.flushing = done
+ // If lots of background writes are already in
+ // progress, block here until one finishes, rather
+ // than pile up an unlimited number of buffered writes
+ // and network flush operations.
+ fn.throttle.Acquire()
+ go func() {
+ defer close(done)
+ locator, _, err := fn.FS().PutB(buf)
+ fn.throttle.Release()
+ fn.Lock()
+ defer fn.Unlock()
+ if curbuf := seg.buf[:1]; &curbuf[0] != &buf[0] {
+ // A new seg.buf has been allocated.
+ return
+ }
+ seg.flushing = nil
+ if err != nil {
+ // TODO: stall (or return errors from)
+ // subsequent writes until flushing
+ // starts to succeed.
+ return
+ }
+ if len(fn.segments) <= idx || fn.segments[idx] != seg || len(seg.buf) != len(buf) {
+ // Segment has been dropped/moved/resized.
+ return
+ }
+ fn.memsize -= int64(len(buf))
+ fn.segments[idx] = storedSegment{
+ kc: fn.FS(),
+ locator: locator,
+ size: len(buf),
+ offset: 0,
+ length: len(buf),
+ }
+ }()
+ }
+}
+
+// Block until all pending pruneMemSegments work is finished. Caller
+// must NOT have lock.
+func (fn *filenode) waitPrune() {
+ var pending []<-chan struct{}
+ fn.Lock()
+ for _, seg := range fn.segments {
+ if seg, ok := seg.(*memSegment); ok && seg.flushing != nil {
+ pending = append(pending, seg.flushing)
}
}
+ fn.Unlock()
+ for _, p := range pending {
+ <-p
+ }
}
type dirnode struct {
return dn.treenode.Child(name, replace)
}
+type fnSegmentRef struct {
+ fn *filenode
+ idx int
+}
+
+// commitBlock concatenates the data from the given filenode segments
+// (which must be *memSegments), writes the data out to Keep as a
+// single block, and replaces the filenodes' *memSegments with
+// storedSegments that reference the relevant portions of the new
+// block.
+//
+// Caller must have write lock.
+func (dn *dirnode) commitBlock(ctx context.Context, throttle *throttle, refs []fnSegmentRef) error {
+ if len(refs) == 0 {
+ return nil
+ }
+ throttle.Acquire()
+ defer throttle.Release()
+ if err := ctx.Err(); err != nil {
+ return err
+ }
+ block := make([]byte, 0, maxBlockSize)
+ for _, ref := range refs {
+ block = append(block, ref.fn.segments[ref.idx].(*memSegment).buf...)
+ }
+ locator, _, err := dn.fs.PutB(block)
+ if err != nil {
+ return err
+ }
+ off := 0
+ for _, ref := range refs {
+ data := ref.fn.segments[ref.idx].(*memSegment).buf
+ ref.fn.segments[ref.idx] = storedSegment{
+ kc: dn.fs,
+ locator: locator,
+ size: len(block),
+ offset: off,
+ length: len(data),
+ }
+ off += len(data)
+ ref.fn.memsize -= int64(len(data))
+ }
+ return nil
+}
+
// sync flushes in-memory data and remote block references (for the
// children with the given names, which must be children of dn) to
// local persistent storage. Caller must have write lock on dn and the
// named children.
-func (dn *dirnode) sync(ctx context.Context, names []string, throttle *throttle) error {
+func (dn *dirnode) sync(ctx context.Context, throttle *throttle, names []string) error {
cg := newContextGroup(ctx)
defer cg.Cancel()
- type shortBlock struct {
- fn *filenode
- idx int
- }
-
- flush := func(sbs []shortBlock) error {
- if len(sbs) == 0 {
- return nil
- }
- throttle.Acquire()
- defer throttle.Release()
- if err := cg.Context().Err(); err != nil {
- return err
- }
- block := make([]byte, 0, maxBlockSize)
- for _, sb := range sbs {
- block = append(block, sb.fn.segments[sb.idx].(*memSegment).buf...)
- }
- locator, _, err := dn.fs.PutB(block)
- if err != nil {
- return err
- }
- off := 0
- for _, sb := range sbs {
- data := sb.fn.segments[sb.idx].(*memSegment).buf
- sb.fn.segments[sb.idx] = storedSegment{
- kc: dn.fs,
- locator: locator,
- size: len(block),
- offset: off,
- length: len(data),
- }
- off += len(data)
- sb.fn.memsize -= int64(len(data))
- }
- return nil
- }
-
- goFlush := func(sbs []shortBlock) {
+ goCommit := func(refs []fnSegmentRef) {
cg.Go(func() error {
- return flush(sbs)
+ return dn.commitBlock(cg.Context(), throttle, refs)
})
}
- var pending []shortBlock
- var pendingLen int
+ var pending []fnSegmentRef
+ var pendingLen int = 0
localLocator := map[string]string{}
for _, name := range names {
fn, ok := dn.inodes[name].(*filenode)
fn.segments[idx] = seg
case *memSegment:
if seg.Len() > maxBlockSize/2 {
- goFlush([]shortBlock{{fn, idx}})
+ goCommit([]fnSegmentRef{{fn, idx}})
continue
}
if pendingLen+seg.Len() > maxBlockSize {
- goFlush(pending)
+ goCommit(pending)
pending = nil
pendingLen = 0
}
- pending = append(pending, shortBlock{fn, idx})
+ pending = append(pending, fnSegmentRef{fn, idx})
pendingLen += seg.Len()
default:
panic(fmt.Sprintf("can't sync segment type %T", seg))
}
}
}
- goFlush(pending)
+ goCommit(pending)
return cg.Wait()
}
names = append(names, name)
}
sort.Strings(names)
+
+ // Wait for children to finish any pending write operations
+ // before locking them.
+ for _, name := range names {
+ node := dn.inodes[name]
+ if fn, ok := node.(*filenode); ok {
+ fn.waitPrune()
+ }
+ }
+
var dirnames []string
var filenames []string
for _, name := range names {
var fileparts []filepart
var blocks []string
- if err := dn.sync(cg.Context(), names, throttle); err != nil {
+ if err := dn.sync(cg.Context(), throttle, names); err != nil {
return err
}
for _, name := range filenames {
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.
+ flushing <-chan struct{}
}
func (me *memSegment) Len() int {
}
func (me *memSegment) Truncate(n int) {
- if n > cap(me.buf) {
+ if n > cap(me.buf) || (me.flushing != nil && n > len(me.buf)) {
newsize := 1024
for newsize < n {
newsize = newsize << 2
}
newbuf := make([]byte, n, newsize)
copy(newbuf, me.buf)
- me.buf = newbuf
+ me.buf, me.flushing = newbuf, nil
} else {
- // Zero unused part when shrinking, in case we grow
- // and start using it again later.
- for i := n; i < len(me.buf); i++ {
+ // reclaim existing capacity, and zero reclaimed part
+ oldlen := len(me.buf)
+ me.buf = me.buf[:n]
+ for i := oldlen; i < n; i++ {
me.buf[i] = 0
}
}
- me.buf = me.buf[:n]
}
func (me *memSegment) WriteAt(p []byte, off int) {
if off+len(p) > len(me.buf) {
panic("overflowed segment")
}
+ if me.flushing != nil {
+ me.buf, me.flushing = append([]byte(nil), me.buf...), nil
+ }
copy(me.buf[off:], p)
}
projects / arvados.git / blobdiff