X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/b2d825bd6059d843b1ae1917ff22e4de3cfc4f5a..41e15e626cc301d867c48ebb1f945bbec6b7d555:/services/keepstore/s3_volume.go diff --git a/services/keepstore/s3_volume.go b/services/keepstore/s3_volume.go index 3f08f6e1b8..bdab58927b 100644 --- a/services/keepstore/s3_volume.go +++ b/services/keepstore/s3_volume.go @@ -18,12 +18,12 @@ import ( "regexp" "strings" "sync" + "sync/atomic" "time" "git.curoverse.com/arvados.git/sdk/go/arvados" "github.com/AdRoll/goamz/aws" "github.com/AdRoll/goamz/s3" - log "github.com/Sirupsen/logrus" ) const ( @@ -45,6 +45,8 @@ var ( s3RaceWindow time.Duration s3ACL = s3.Private + + zeroTime time.Time ) const ( @@ -151,6 +153,7 @@ type S3Volume struct { RaceWindow arvados.Duration ReadOnly bool UnsafeDelete bool + StorageClasses []string bucket *s3bucket @@ -420,13 +423,13 @@ func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error { if err != nil { return } - err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{}) + err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) }() select { case <-ctx.Done(): theConfig.debugLogf("%s: taking PutReader's input away: %s", v, ctx.Err()) // Our pipe might be stuck in Write(), waiting for - // io.Copy() to read. If so, un-stick it. This means + // PutReader() to read. If so, un-stick it. This means // PutReader will get corrupt data, but that's OK: the // size and MD5 won't match, so the write will fail. go io.Copy(ioutil.Discard, bufr) @@ -435,6 +438,8 @@ func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error { theConfig.debugLogf("%s: abandoning PutReader goroutine", v) return ctx.Err() case <-ready: + // Unblock pipe in case PutReader did not consume it. + io.Copy(ioutil.Discard, bufr) return v.translateError(err) } } @@ -452,7 +457,7 @@ func (v *S3Volume) Touch(loc string) error { } else if err != nil { return err } - err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{}) + err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) return v.translateError(err) } @@ -466,7 +471,7 @@ func (v *S3Volume) Mtime(loc string) (time.Time, error) { err = v.translateError(err) if os.IsNotExist(err) { // The data object X exists, but recent/X is missing. - err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{}) + err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) if err != nil { log.Printf("error: creating %q: %s", "recent/"+loc, err) return zeroTime, v.translateError(err) @@ -648,7 +653,7 @@ func (v *S3Volume) Untrash(loc string) error { if err != nil { return err } - err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{}) + err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{}) return v.translateError(err) } @@ -685,6 +690,11 @@ func (v *S3Volume) Replication() int { return v.S3Replication } +// GetStorageClasses implements Volume +func (v *S3Volume) GetStorageClasses() []string { + return v.StorageClasses +} + var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`) func (v *S3Volume) isKeepBlock(s string) bool { @@ -757,26 +767,21 @@ func (v *S3Volume) translateError(err error) error { func (v *S3Volume) EmptyTrash() { var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64 - // Use a merge sort to find matching sets of trash/X and recent/X. - trashL := s3Lister{ - Bucket: v.bucket.Bucket, - Prefix: "trash/", - PageSize: v.IndexPageSize, - } // Define "ready to delete" as "...when EmptyTrash started". startT := time.Now() - for trash := trashL.First(); trash != nil; trash = trashL.Next() { + + emptyOneKey := func(trash *s3.Key) { loc := trash.Key[6:] if !v.isKeepBlock(loc) { - continue + return } - bytesInTrash += trash.Size - blocksInTrash++ + atomic.AddInt64(&bytesInTrash, trash.Size) + atomic.AddInt64(&blocksInTrash, 1) trashT, err := time.Parse(time.RFC3339, trash.LastModified) if err != nil { log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, trash.Key, trash.LastModified, err) - continue + return } recent, err := v.bucket.Head("recent/"+loc, nil) if err != nil && os.IsNotExist(v.translateError(err)) { @@ -785,15 +790,15 @@ func (v *S3Volume) EmptyTrash() { if err != nil { log.Printf("error: %s: EmptyTrash: Untrash(%q): %s", v, loc, err) } - continue + return } else if err != nil { log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err) - continue + return } recentT, err := v.lastModified(recent) if err != nil { log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, "recent/"+loc, recent.Header.Get("Last-Modified"), err) - continue + return } if trashT.Sub(recentT) < theConfig.BlobSignatureTTL.Duration() { if age := startT.Sub(recentT); age >= theConfig.BlobSignatureTTL.Duration()-time.Duration(v.RaceWindow) { @@ -808,39 +813,67 @@ func (v *S3Volume) EmptyTrash() { log.Printf("notice: %s: EmptyTrash: detected old race for %q, calling fixRace + Touch", v, loc) v.fixRace(loc) v.Touch(loc) - continue + return } _, err := v.bucket.Head(loc, nil) if os.IsNotExist(err) { log.Printf("notice: %s: EmptyTrash: detected recent race for %q, calling fixRace", v, loc) v.fixRace(loc) - continue + return } else if err != nil { log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err) - continue + return } } if startT.Sub(trashT) < theConfig.TrashLifetime.Duration() { - continue + return } err = v.bucket.Del(trash.Key) if err != nil { log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, trash.Key, err) - continue + return } - bytesDeleted += trash.Size - blocksDeleted++ + atomic.AddInt64(&bytesDeleted, trash.Size) + atomic.AddInt64(&blocksDeleted, 1) _, err = v.bucket.Head(loc, nil) - if os.IsNotExist(err) { - err = v.bucket.Del("recent/" + loc) - if err != nil { - log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err) - } - } else if err != nil { - log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err) + if err == nil { + log.Printf("warning: %s: EmptyTrash: HEAD %q succeeded immediately after deleting %q", v, loc, loc) + return + } + if !os.IsNotExist(v.translateError(err)) { + log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err) + return + } + err = v.bucket.Del("recent/" + loc) + if err != nil { + log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err) } } + + var wg sync.WaitGroup + todo := make(chan *s3.Key, theConfig.EmptyTrashWorkers) + for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ { + wg.Add(1) + go func() { + defer wg.Done() + for key := range todo { + emptyOneKey(key) + } + }() + } + + trashL := s3Lister{ + Bucket: v.bucket.Bucket, + Prefix: "trash/", + PageSize: v.IndexPageSize, + } + for trash := trashL.First(); trash != nil; trash = trashL.Next() { + todo <- trash + } + close(todo) + wg.Wait() + if err := trashL.Error(); err != nil { log.Printf("error: %s: EmptyTrash: lister: %s", v, err) } @@ -927,14 +960,18 @@ func (b *s3bucket) Head(path string, headers map[string][]string) (*http.Respons } func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error { - err := b.Bucket.PutReader(path, NewCountingReader(r, b.stats.TickOutBytes), length, contType, perm, options) - b.stats.Tick(&b.stats.Ops, &b.stats.PutOps) - b.stats.TickErr(err) - return err -} - -func (b *s3bucket) Put(path string, data []byte, contType string, perm s3.ACL, options s3.Options) error { - err := b.Bucket.PutReader(path, NewCountingReader(bytes.NewBuffer(data), b.stats.TickOutBytes), int64(len(data)), contType, perm, options) + if length == 0 { + // goamz will only send Content-Length: 0 when reader + // is nil due to net.http.Request.ContentLength + // behavior. Otherwise, Content-Length header is + // omitted which will cause some S3 services + // (including AWS and Ceph RadosGW) to fail to create + // empty objects. + r = nil + } else { + r = NewCountingReader(r, b.stats.TickOutBytes) + } + err := b.Bucket.PutReader(path, r, length, contType, perm, options) b.stats.Tick(&b.stats.Ops, &b.stats.PutOps) b.stats.TickErr(err) return err