1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
27 "git.arvados.org/arvados.git/sdk/go/arvados"
28 "github.com/AdRoll/goamz/aws"
29 "github.com/AdRoll/goamz/s3"
30 "github.com/prometheus/client_golang/prometheus"
31 "github.com/sirupsen/logrus"
35 driver["S3"] = chooseS3VolumeDriver
38 func newS3Volume(cluster *arvados.Cluster, volume arvados.Volume, logger logrus.FieldLogger, metrics *volumeMetricsVecs) (Volume, error) {
39 v := &S3Volume{cluster: cluster, volume: volume, metrics: metrics}
40 err := json.Unmarshal(volume.DriverParameters, v)
44 v.logger = logger.WithField("Volume", v.String())
48 func (v *S3Volume) check() error {
50 return errors.New("DriverParameters: Bucket must be provided")
52 if v.IndexPageSize == 0 {
53 v.IndexPageSize = 1000
56 return errors.New("DriverParameters: RaceWindow must not be negative")
60 v.region, ok = aws.Regions[v.Region]
63 return fmt.Errorf("unrecognized region %+q; try specifying endpoint instead", v.Region)
66 return fmt.Errorf("refusing to use AWS region name %+q with endpoint %+q; "+
67 "specify empty endpoint or use a different region name", v.Region, v.Endpoint)
69 v.region = aws.Region{
71 S3Endpoint: v.Endpoint,
72 S3LocationConstraint: v.LocationConstraint,
76 // Zero timeouts mean "wait forever", which is a bad
77 // default. Default to long timeouts instead.
78 if v.ConnectTimeout == 0 {
79 v.ConnectTimeout = s3DefaultConnectTimeout
81 if v.ReadTimeout == 0 {
82 v.ReadTimeout = s3DefaultReadTimeout
91 // Set up prometheus metrics
92 lbls := prometheus.Labels{"device_id": v.GetDeviceID()}
93 v.bucket.stats.opsCounters, v.bucket.stats.errCounters, v.bucket.stats.ioBytes = v.metrics.getCounterVecsFor(lbls)
95 err := v.bootstrapIAMCredentials()
97 return fmt.Errorf("error getting IAM credentials: %s", err)
104 s3DefaultReadTimeout = arvados.Duration(10 * time.Minute)
105 s3DefaultConnectTimeout = arvados.Duration(time.Minute)
109 // ErrS3TrashDisabled is returned by Trash if that operation
110 // is impossible with the current config.
111 ErrS3TrashDisabled = fmt.Errorf("trash function is disabled because Collections.BlobTrashLifetime=0 and DriverParameters.UnsafeDelete=false")
119 maxClockSkew = 600 * time.Second
120 nearlyRFC1123 = "Mon, 2 Jan 2006 15:04:05 GMT"
123 func s3regions() (okList []string) {
124 for r := range aws.Regions {
125 okList = append(okList, r)
130 // S3Volume implements Volume using an S3 bucket.
131 type S3Volume struct {
132 arvados.S3VolumeDriverParameters
133 AuthToken string // populated automatically when IAMRole is used
134 AuthExpiration time.Time // populated automatically when IAMRole is used
136 cluster *arvados.Cluster
137 volume arvados.Volume
138 logger logrus.FieldLogger
139 metrics *volumeMetricsVecs
145 // GetDeviceID returns a globally unique ID for the storage bucket.
146 func (v *S3Volume) GetDeviceID() string {
147 return "s3://" + v.Endpoint + "/" + v.Bucket
150 func (v *S3Volume) bootstrapIAMCredentials() error {
151 if v.AccessKeyID != "" || v.SecretAccessKey != "" {
153 return errors.New("invalid DriverParameters: AccessKeyID and SecretAccessKey must be blank if IAMRole is specified")
157 ttl, err := v.updateIAMCredentials()
164 ttl, err = v.updateIAMCredentials()
166 v.logger.WithError(err).Warnf("failed to update credentials for IAM role %q", v.IAMRole)
168 } else if ttl < time.Second {
169 v.logger.WithField("TTL", ttl).Warnf("received stale credentials for IAM role %q", v.IAMRole)
177 func (v *S3Volume) newS3Client() *s3.S3 {
178 auth := aws.NewAuth(v.AccessKeyID, v.SecretAccessKey, v.AuthToken, v.AuthExpiration)
179 client := s3.New(*auth, v.region)
181 client.Signature = aws.V4Signature
183 client.ConnectTimeout = time.Duration(v.ConnectTimeout)
184 client.ReadTimeout = time.Duration(v.ReadTimeout)
188 // returned by AWS metadata endpoint .../security-credentials/${rolename}
189 type iamCredentials struct {
191 LastUpdated time.Time
194 SecretAccessKey string
199 // Returns TTL of updated credentials, i.e., time to sleep until next
201 func (v *S3Volume) updateIAMCredentials() (time.Duration, error) {
202 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
205 metadataBaseURL := "http://169.254.169.254/latest/meta-data/iam/security-credentials/"
208 if strings.Contains(v.IAMRole, "://") {
209 // Configuration provides complete URL (used by tests)
211 } else if v.IAMRole != "" {
212 // Configuration provides IAM role name and we use the
213 // AWS metadata endpoint
214 url = metadataBaseURL + v.IAMRole
216 url = metadataBaseURL
217 v.logger.WithField("URL", url).Debug("looking up IAM role name")
218 req, err := http.NewRequest("GET", url, nil)
220 return 0, fmt.Errorf("error setting up request %s: %s", url, err)
222 resp, err := http.DefaultClient.Do(req.WithContext(ctx))
224 return 0, fmt.Errorf("error getting %s: %s", url, err)
226 defer resp.Body.Close()
227 if resp.StatusCode == http.StatusNotFound {
228 return 0, fmt.Errorf("this instance does not have an IAM role assigned -- either assign a role, or configure AccessKeyID and SecretAccessKey explicitly in DriverParameters (error getting %s: HTTP status %s)", url, resp.Status)
229 } else if resp.StatusCode != http.StatusOK {
230 return 0, fmt.Errorf("error getting %s: HTTP status %s", url, resp.Status)
232 body := bufio.NewReader(resp.Body)
234 _, err = fmt.Fscanf(body, "%s\n", &role)
236 return 0, fmt.Errorf("error reading response from %s: %s", url, err)
238 if n, _ := body.Read(make([]byte, 64)); n > 0 {
239 v.logger.Warnf("ignoring additional data returned by metadata endpoint %s after the single role name that we expected", url)
241 v.logger.WithField("Role", role).Debug("looked up IAM role name")
245 v.logger.WithField("URL", url).Debug("getting credentials")
246 req, err := http.NewRequest("GET", url, nil)
248 return 0, fmt.Errorf("error setting up request %s: %s", url, err)
250 resp, err := http.DefaultClient.Do(req.WithContext(ctx))
252 return 0, fmt.Errorf("error getting %s: %s", url, err)
254 defer resp.Body.Close()
255 if resp.StatusCode != http.StatusOK {
256 return 0, fmt.Errorf("error getting %s: HTTP status %s", url, resp.Status)
258 var cred iamCredentials
259 err = json.NewDecoder(resp.Body).Decode(&cred)
261 return 0, fmt.Errorf("error decoding credentials from %s: %s", url, err)
263 v.AccessKeyID, v.SecretAccessKey, v.AuthToken, v.AuthExpiration = cred.AccessKeyID, cred.SecretAccessKey, cred.Token, cred.Expiration
264 v.bucket.SetBucket(&s3.Bucket{
268 // TTL is time from now to expiration, minus 5m. "We make new
269 // credentials available at least five minutes before the
270 // expiration of the old credentials." --
271 // https://docs.aws.amazon.com/AWSEC2/latest/UserGuide/iam-roles-for-amazon-ec2.html#instance-metadata-security-credentials
272 // (If that's not true, the returned ttl might be zero or
273 // negative, which the caller can handle.)
274 ttl := cred.Expiration.Sub(time.Now()) - 5*time.Minute
275 v.logger.WithFields(logrus.Fields{
276 "AccessKeyID": cred.AccessKeyID,
277 "LastUpdated": cred.LastUpdated,
278 "Expiration": cred.Expiration,
279 "TTL": arvados.Duration(ttl),
280 }).Debug("updated credentials")
284 func (v *S3Volume) getReaderWithContext(ctx context.Context, key string) (rdr io.ReadCloser, err error) {
285 ready := make(chan bool)
287 rdr, err = v.getReader(key)
294 v.logger.Debugf("s3: abandoning getReader(%s): %s", key, ctx.Err())
301 return nil, ctx.Err()
305 // getReader wraps (Bucket)GetReader.
307 // In situations where (Bucket)GetReader would fail because the block
308 // disappeared in a Trash race, getReader calls fixRace to recover the
309 // data, and tries again.
310 func (v *S3Volume) getReader(key string) (rdr io.ReadCloser, err error) {
311 rdr, err = v.bucket.GetReader(key)
312 err = v.translateError(err)
313 if err == nil || !os.IsNotExist(err) {
317 _, err = v.bucket.Head("recent/"+key, nil)
318 err = v.translateError(err)
320 // If we can't read recent/X, there's no point in
321 // trying fixRace. Give up.
329 rdr, err = v.bucket.GetReader(key)
331 v.logger.Warnf("reading %s after successful fixRace: %s", key, err)
332 err = v.translateError(err)
337 // Get a block: copy the block data into buf, and return the number of
339 func (v *S3Volume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
341 rdr, err := v.getReaderWithContext(ctx, key)
347 ready := make(chan bool)
352 n, err = io.ReadFull(rdr, buf)
355 case nil, io.EOF, io.ErrUnexpectedEOF:
358 err = v.translateError(err)
363 v.logger.Debugf("s3: interrupting ReadFull() with Close() because %s", ctx.Err())
365 // Must wait for ReadFull to return, to ensure it
366 // doesn't write to buf after we return.
367 v.logger.Debug("s3: waiting for ReadFull() to fail")
375 // Compare the given data with the stored data.
376 func (v *S3Volume) Compare(ctx context.Context, loc string, expect []byte) error {
378 errChan := make(chan error, 1)
380 _, err := v.bucket.Head("recent/"+key, nil)
387 case err = <-errChan:
390 // Checking for "loc" itself here would interfere with
391 // future GET requests.
393 // On AWS, if X doesn't exist, a HEAD or GET request
394 // for X causes X's non-existence to be cached. Thus,
395 // if we test for X, then create X and return a
396 // signature to our client, the client might still get
397 // 404 from all keepstores when trying to read it.
399 // To avoid this, we avoid doing HEAD X or GET X until
400 // we know X has been written.
402 // Note that X might exist even though recent/X
403 // doesn't: for example, the response to HEAD recent/X
404 // might itself come from a stale cache. In such
405 // cases, we will return a false negative and
406 // PutHandler might needlessly create another replica
407 // on a different volume. That's not ideal, but it's
408 // better than passing the eventually-consistent
409 // problem on to our clients.
410 return v.translateError(err)
412 rdr, err := v.getReaderWithContext(ctx, key)
417 return v.translateError(compareReaderWithBuf(ctx, rdr, expect, loc[:32]))
420 // Put writes a block.
421 func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error {
422 if v.volume.ReadOnly {
423 return MethodDisabledError
428 md5, err := hex.DecodeString(loc)
432 opts.ContentMD5 = base64.StdEncoding.EncodeToString(md5)
433 // In AWS regions that use V4 signatures, we need to
434 // provide ContentSHA256 up front. Otherwise, the S3
435 // library reads the request body (from our buffer)
436 // into another new buffer in order to compute the
437 // SHA256 before sending the request -- which would
438 // mean consuming 128 MiB of memory for the duration
439 // of a 64 MiB write.
440 opts.ContentSHA256 = fmt.Sprintf("%x", sha256.Sum256(block))
445 // Send the block data through a pipe, so that (if we need to)
446 // we can close the pipe early and abandon our PutReader()
447 // goroutine, without worrying about PutReader() accessing our
448 // block buffer after we release it.
449 bufr, bufw := io.Pipe()
451 io.Copy(bufw, bytes.NewReader(block))
456 ready := make(chan bool)
459 if ctx.Err() != nil {
460 v.logger.Debugf("abandoned PutReader goroutine finished with err: %s", err)
464 err = v.bucket.PutReader(key, bufr, int64(size), "application/octet-stream", s3ACL, opts)
468 err = v.bucket.PutReader("recent/"+key, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
472 v.logger.Debugf("taking PutReader's input away: %s", ctx.Err())
473 // Our pipe might be stuck in Write(), waiting for
474 // PutReader() to read. If so, un-stick it. This means
475 // PutReader will get corrupt data, but that's OK: the
476 // size and MD5 won't match, so the write will fail.
477 go io.Copy(ioutil.Discard, bufr)
478 // CloseWithError() will return once pending I/O is done.
479 bufw.CloseWithError(ctx.Err())
480 v.logger.Debugf("abandoning PutReader goroutine")
483 // Unblock pipe in case PutReader did not consume it.
484 io.Copy(ioutil.Discard, bufr)
485 return v.translateError(err)
489 // Touch sets the timestamp for the given locator to the current time.
490 func (v *S3Volume) Touch(loc string) error {
491 if v.volume.ReadOnly {
492 return MethodDisabledError
495 _, err := v.bucket.Head(key, nil)
496 err = v.translateError(err)
497 if os.IsNotExist(err) && v.fixRace(key) {
498 // The data object got trashed in a race, but fixRace
500 } else if err != nil {
503 err = v.bucket.PutReader("recent/"+key, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
504 return v.translateError(err)
507 // Mtime returns the stored timestamp for the given locator.
508 func (v *S3Volume) Mtime(loc string) (time.Time, error) {
510 _, err := v.bucket.Head(key, nil)
512 return zeroTime, v.translateError(err)
514 resp, err := v.bucket.Head("recent/"+key, nil)
515 err = v.translateError(err)
516 if os.IsNotExist(err) {
517 // The data object X exists, but recent/X is missing.
518 err = v.bucket.PutReader("recent/"+key, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
520 v.logger.WithError(err).Errorf("error creating %q", "recent/"+key)
521 return zeroTime, v.translateError(err)
523 v.logger.Infof("created %q to migrate existing block to new storage scheme", "recent/"+key)
524 resp, err = v.bucket.Head("recent/"+key, nil)
526 v.logger.WithError(err).Errorf("HEAD failed after creating %q", "recent/"+key)
527 return zeroTime, v.translateError(err)
529 } else if err != nil {
530 // HEAD recent/X failed for some other reason.
533 return v.lastModified(resp)
536 // IndexTo writes a complete list of locators with the given prefix
537 // for which Get() can retrieve data.
538 func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
539 // Use a merge sort to find matching sets of X and recent/X.
542 Bucket: v.bucket.Bucket(),
543 Prefix: v.key(prefix),
544 PageSize: v.IndexPageSize,
545 Stats: &v.bucket.stats,
549 Bucket: v.bucket.Bucket(),
550 Prefix: "recent/" + v.key(prefix),
551 PageSize: v.IndexPageSize,
552 Stats: &v.bucket.stats,
554 for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() {
556 // Conveniently, "recent/*" and "trash/*" are
557 // lexically greater than all hex-encoded data
558 // hashes, so stopping here avoids iterating
559 // over all of them needlessly with dataL.
562 loc, isBlk := v.isKeepBlock(data.Key)
567 // stamp is the list entry we should use to report the
568 // last-modified time for this data block: it will be
569 // the recent/X entry if one exists, otherwise the
570 // entry for the data block itself.
573 // Advance to the corresponding recent/X marker, if any
574 for recent != nil && recentL.Error() == nil {
575 if cmp := strings.Compare(recent.Key[7:], data.Key); cmp < 0 {
576 recent = recentL.Next()
580 recent = recentL.Next()
583 // recent/X marker is missing: we'll
584 // use the timestamp on the data
589 if err := recentL.Error(); err != nil {
592 t, err := time.Parse(time.RFC3339, stamp.LastModified)
596 // We truncate sub-second precision here. Otherwise
597 // timestamps will never match the RFC1123-formatted
598 // Last-Modified values parsed by Mtime().
599 fmt.Fprintf(writer, "%s+%d %d\n", loc, data.Size, t.Unix()*1000000000)
604 // Trash a Keep block.
605 func (v *S3Volume) Trash(loc string) error {
606 if v.volume.ReadOnly {
607 return MethodDisabledError
609 if t, err := v.Mtime(loc); err != nil {
611 } else if time.Since(t) < v.cluster.Collections.BlobSigningTTL.Duration() {
615 if v.cluster.Collections.BlobTrashLifetime == 0 {
617 return ErrS3TrashDisabled
619 return v.translateError(v.bucket.Del(key))
621 err := v.checkRaceWindow(key)
625 err = v.safeCopy("trash/"+key, key)
629 return v.translateError(v.bucket.Del(key))
632 // checkRaceWindow returns a non-nil error if trash/key is, or might
633 // be, in the race window (i.e., it's not safe to trash key).
634 func (v *S3Volume) checkRaceWindow(key string) error {
635 resp, err := v.bucket.Head("trash/"+key, nil)
636 err = v.translateError(err)
637 if os.IsNotExist(err) {
638 // OK, trash/X doesn't exist so we're not in the race
641 } else if err != nil {
642 // Error looking up trash/X. We don't know whether
643 // we're in the race window
646 t, err := v.lastModified(resp)
648 // Can't parse timestamp
651 safeWindow := t.Add(v.cluster.Collections.BlobTrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow)))
653 // We can't count on "touch trash/X" to prolong
654 // trash/X's lifetime. The new timestamp might not
655 // become visible until now+raceWindow, and EmptyTrash
656 // is allowed to delete trash/X before then.
657 return fmt.Errorf("%s: same block is already in trash, and safe window ended %s ago", key, -safeWindow)
659 // trash/X exists, but it won't be eligible for deletion until
660 // after now+raceWindow, so it's safe to overwrite it.
664 // safeCopy calls PutCopy, and checks the response to make sure the
665 // copy succeeded and updated the timestamp on the destination object
666 // (PutCopy returns 200 OK if the request was received, even if the
668 func (v *S3Volume) safeCopy(dst, src string) error {
669 resp, err := v.bucket.Bucket().PutCopy(dst, s3ACL, s3.CopyOptions{
670 ContentType: "application/octet-stream",
671 MetadataDirective: "REPLACE",
672 }, v.bucket.Bucket().Name+"/"+src)
673 err = v.translateError(err)
674 if os.IsNotExist(err) {
676 } else if err != nil {
677 return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.Bucket().Name+"/"+src, err)
679 if t, err := time.Parse(time.RFC3339Nano, resp.LastModified); err != nil {
680 return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.LastModified, err)
681 } else if time.Now().Sub(t) > maxClockSkew {
682 return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.LastModified, t)
687 // Get the LastModified header from resp, and parse it as RFC1123 or
688 // -- if it isn't valid RFC1123 -- as Amazon's variant of RFC1123.
689 func (v *S3Volume) lastModified(resp *http.Response) (t time.Time, err error) {
690 s := resp.Header.Get("Last-Modified")
691 t, err = time.Parse(time.RFC1123, s)
692 if err != nil && s != "" {
693 // AWS example is "Sun, 1 Jan 2006 12:00:00 GMT",
694 // which isn't quite "Sun, 01 Jan 2006 12:00:00 GMT"
695 // as required by HTTP spec. If it's not a valid HTTP
696 // header value, it's probably AWS (or s3test) giving
697 // us a nearly-RFC1123 timestamp.
698 t, err = time.Parse(nearlyRFC1123, s)
703 // Untrash moves block from trash back into store
704 func (v *S3Volume) Untrash(loc string) error {
706 err := v.safeCopy(key, "trash/"+key)
710 err = v.bucket.PutReader("recent/"+key, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
711 return v.translateError(err)
714 // Status returns a *VolumeStatus representing the current in-use
715 // storage capacity and a fake available capacity that doesn't make
716 // the volume seem full or nearly-full.
717 func (v *S3Volume) Status() *VolumeStatus {
718 return &VolumeStatus{
720 BytesFree: BlockSize * 1000,
725 // InternalStats returns bucket I/O and API call counters.
726 func (v *S3Volume) InternalStats() interface{} {
727 return &v.bucket.stats
730 // String implements fmt.Stringer.
731 func (v *S3Volume) String() string {
732 return fmt.Sprintf("s3-bucket:%+q", v.Bucket)
735 var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
737 func (v *S3Volume) isKeepBlock(s string) (string, bool) {
738 if v.PrefixLength > 0 && len(s) == v.PrefixLength+33 && s[:v.PrefixLength] == s[v.PrefixLength+1:v.PrefixLength*2+1] {
739 s = s[v.PrefixLength+1:]
741 return s, s3KeepBlockRegexp.MatchString(s)
744 // Return the key used for a given loc. If PrefixLength==0 then
745 // key("abcdef0123") is "abcdef0123", if PrefixLength==3 then key is
746 // "abc/abcdef0123", etc.
747 func (v *S3Volume) key(loc string) string {
748 if v.PrefixLength > 0 && v.PrefixLength < len(loc)-1 {
749 return loc[:v.PrefixLength] + "/" + loc
755 // fixRace(X) is called when "recent/X" exists but "X" doesn't
756 // exist. If the timestamps on "recent/X" and "trash/X" indicate there
757 // was a race between Put and Trash, fixRace recovers from the race by
758 // Untrashing the block.
759 func (v *S3Volume) fixRace(key string) bool {
760 trash, err := v.bucket.Head("trash/"+key, nil)
762 if !os.IsNotExist(v.translateError(err)) {
763 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "trash/"+key)
767 trashTime, err := v.lastModified(trash)
769 v.logger.WithError(err).Errorf("fixRace: error parsing time %q", trash.Header.Get("Last-Modified"))
773 recent, err := v.bucket.Head("recent/"+key, nil)
775 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "recent/"+key)
778 recentTime, err := v.lastModified(recent)
780 v.logger.WithError(err).Errorf("fixRace: error parsing time %q", recent.Header.Get("Last-Modified"))
784 ageWhenTrashed := trashTime.Sub(recentTime)
785 if ageWhenTrashed >= v.cluster.Collections.BlobSigningTTL.Duration() {
786 // No evidence of a race: block hasn't been written
787 // since it became eligible for Trash. No fix needed.
791 v.logger.Infof("fixRace: %q: trashed at %s but touched at %s (age when trashed = %s < %s)", key, trashTime, recentTime, ageWhenTrashed, v.cluster.Collections.BlobSigningTTL)
792 v.logger.Infof("fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+key, key)
793 err = v.safeCopy(key, "trash/"+key)
795 v.logger.WithError(err).Error("fixRace: copy failed")
801 func (v *S3Volume) translateError(err error) error {
802 switch err := err.(type) {
804 if (err.StatusCode == http.StatusNotFound && err.Code == "NoSuchKey") ||
805 strings.Contains(err.Error(), "Not Found") {
806 return os.ErrNotExist
808 // Other 404 errors like NoSuchVersion and
809 // NoSuchBucket are different problems which should
810 // get called out downstream, so we don't convert them
811 // to os.ErrNotExist.
816 // EmptyTrash looks for trashed blocks that exceeded BlobTrashLifetime
817 // and deletes them from the volume.
818 func (v *S3Volume) EmptyTrash() {
819 if v.cluster.Collections.BlobDeleteConcurrency < 1 {
823 var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64
825 // Define "ready to delete" as "...when EmptyTrash started".
828 emptyOneKey := func(trash *s3.Key) {
830 loc, isBlk := v.isKeepBlock(key)
834 atomic.AddInt64(&bytesInTrash, trash.Size)
835 atomic.AddInt64(&blocksInTrash, 1)
837 trashT, err := time.Parse(time.RFC3339, trash.LastModified)
839 v.logger.Warnf("EmptyTrash: %q: parse %q: %s", trash.Key, trash.LastModified, err)
842 recent, err := v.bucket.Head("recent/"+key, nil)
843 if err != nil && os.IsNotExist(v.translateError(err)) {
844 v.logger.Warnf("EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", trash.Key, "recent/"+loc, err)
847 v.logger.WithError(err).Errorf("EmptyTrash: Untrash(%q) failed", loc)
850 } else if err != nil {
851 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", "recent/"+key)
854 recentT, err := v.lastModified(recent)
856 v.logger.WithError(err).Warnf("EmptyTrash: %q: error parsing %q", "recent/"+key, recent.Header.Get("Last-Modified"))
859 if trashT.Sub(recentT) < v.cluster.Collections.BlobSigningTTL.Duration() {
860 if age := startT.Sub(recentT); age >= v.cluster.Collections.BlobSigningTTL.Duration()-time.Duration(v.RaceWindow) {
861 // recent/loc is too old to protect
862 // loc from being Trashed again during
863 // the raceWindow that starts if we
864 // delete trash/X now.
866 // Note this means (TrashSweepInterval
867 // < BlobSigningTTL - raceWindow) is
868 // necessary to avoid starvation.
869 v.logger.Infof("EmptyTrash: detected old race for %q, calling fixRace + Touch", loc)
874 _, err := v.bucket.Head(key, nil)
875 if os.IsNotExist(err) {
876 v.logger.Infof("EmptyTrash: detected recent race for %q, calling fixRace", loc)
879 } else if err != nil {
880 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc)
884 if startT.Sub(trashT) < v.cluster.Collections.BlobTrashLifetime.Duration() {
887 err = v.bucket.Del(trash.Key)
889 v.logger.WithError(err).Errorf("EmptyTrash: error deleting %q", trash.Key)
892 atomic.AddInt64(&bytesDeleted, trash.Size)
893 atomic.AddInt64(&blocksDeleted, 1)
895 _, err = v.bucket.Head(key, nil)
897 v.logger.Warnf("EmptyTrash: HEAD %q succeeded immediately after deleting %q", key, key)
900 if !os.IsNotExist(v.translateError(err)) {
901 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", key)
904 err = v.bucket.Del("recent/" + key)
906 v.logger.WithError(err).Warnf("EmptyTrash: error deleting %q", "recent/"+key)
910 var wg sync.WaitGroup
911 todo := make(chan *s3.Key, v.cluster.Collections.BlobDeleteConcurrency)
912 for i := 0; i < v.cluster.Collections.BlobDeleteConcurrency; i++ {
916 for key := range todo {
924 Bucket: v.bucket.Bucket(),
926 PageSize: v.IndexPageSize,
927 Stats: &v.bucket.stats,
929 for trash := trashL.First(); trash != nil; trash = trashL.Next() {
935 if err := trashL.Error(); err != nil {
936 v.logger.WithError(err).Error("EmptyTrash: lister failed")
938 v.logger.Infof("EmptyTrash stats for %v: Deleted %v bytes in %v blocks. Remaining in trash: %v bytes in %v blocks.", v.String(), bytesDeleted, blocksDeleted, bytesInTrash-bytesDeleted, blocksInTrash-blocksDeleted)
941 type s3Lister struct {
942 Logger logrus.FieldLogger
952 // First fetches the first page and returns the first item. It returns
953 // nil if the response is the empty set or an error occurs.
954 func (lister *s3Lister) First() *s3.Key {
959 // Next returns the next item, fetching the next page if necessary. It
960 // returns nil if the last available item has already been fetched, or
962 func (lister *s3Lister) Next() *s3.Key {
963 if len(lister.buf) == 0 && lister.nextMarker != "" {
969 // Return the most recent error encountered by First or Next.
970 func (lister *s3Lister) Error() error {
974 func (lister *s3Lister) getPage() {
975 lister.Stats.TickOps("list")
976 lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps)
977 resp, err := lister.Bucket.List(lister.Prefix, "", lister.nextMarker, lister.PageSize)
978 lister.nextMarker = ""
983 if resp.IsTruncated {
984 lister.nextMarker = resp.NextMarker
986 lister.buf = make([]s3.Key, 0, len(resp.Contents))
987 for _, key := range resp.Contents {
988 if !strings.HasPrefix(key.Key, lister.Prefix) {
989 lister.Logger.Warnf("s3Lister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, key.Key)
992 lister.buf = append(lister.buf, key)
996 func (lister *s3Lister) pop() (k *s3.Key) {
997 if len(lister.buf) > 0 {
999 lister.buf = lister.buf[1:]
1004 // s3bucket wraps s3.bucket and counts I/O and API usage stats. The
1005 // wrapped bucket can be replaced atomically with SetBucket in order
1006 // to update credentials.
1007 type s3bucket struct {
1013 func (b *s3bucket) Bucket() *s3.Bucket {
1019 func (b *s3bucket) SetBucket(bucket *s3.Bucket) {
1025 func (b *s3bucket) GetReader(path string) (io.ReadCloser, error) {
1026 rdr, err := b.Bucket().GetReader(path)
1027 b.stats.TickOps("get")
1028 b.stats.Tick(&b.stats.Ops, &b.stats.GetOps)
1029 b.stats.TickErr(err)
1030 return NewCountingReader(rdr, b.stats.TickInBytes), err
1033 func (b *s3bucket) Head(path string, headers map[string][]string) (*http.Response, error) {
1034 resp, err := b.Bucket().Head(path, headers)
1035 b.stats.TickOps("head")
1036 b.stats.Tick(&b.stats.Ops, &b.stats.HeadOps)
1037 b.stats.TickErr(err)
1041 func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error {
1043 // goamz will only send Content-Length: 0 when reader
1044 // is nil due to net.http.Request.ContentLength
1045 // behavior. Otherwise, Content-Length header is
1046 // omitted which will cause some S3 services
1047 // (including AWS and Ceph RadosGW) to fail to create
1051 r = NewCountingReader(r, b.stats.TickOutBytes)
1053 err := b.Bucket().PutReader(path, r, length, contType, perm, options)
1054 b.stats.TickOps("put")
1055 b.stats.Tick(&b.stats.Ops, &b.stats.PutOps)
1056 b.stats.TickErr(err)
1060 func (b *s3bucket) Del(path string) error {
1061 err := b.Bucket().Del(path)
1062 b.stats.TickOps("delete")
1063 b.stats.Tick(&b.stats.Ops, &b.stats.DelOps)
1064 b.stats.TickErr(err)
1068 type s3bucketStats struct {
1078 func (s *s3bucketStats) TickErr(err error) {
1082 errType := fmt.Sprintf("%T", err)
1083 if err, ok := err.(*s3.Error); ok {
1084 errType = errType + fmt.Sprintf(" %d %s", err.StatusCode, err.Code)
1086 s.statsTicker.TickErr(err, errType)