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.AccessKey != "" || v.SecretKey != "" {
153 return errors.New("invalid DriverParameters: AccessKey and SecretKey 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.AccessKey, v.SecretKey, 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 AccessKey and SecretKey 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.AccessKey, v.SecretKey, 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, loc string) (rdr io.ReadCloser, err error) {
285 ready := make(chan bool)
287 rdr, err = v.getReader(loc)
294 v.logger.Debugf("s3: abandoning getReader(): %s", 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(loc string) (rdr io.ReadCloser, err error) {
311 rdr, err = v.bucket.GetReader(loc)
312 err = v.translateError(err)
313 if err == nil || !os.IsNotExist(err) {
317 _, err = v.bucket.Head("recent/"+loc, 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(loc)
331 v.logger.Warnf("reading %s after successful fixRace: %s", loc, 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) {
340 rdr, err := v.getReaderWithContext(ctx, loc)
346 ready := make(chan bool)
351 n, err = io.ReadFull(rdr, buf)
354 case nil, io.EOF, io.ErrUnexpectedEOF:
357 err = v.translateError(err)
362 v.logger.Debugf("s3: interrupting ReadFull() with Close() because %s", ctx.Err())
364 // Must wait for ReadFull to return, to ensure it
365 // doesn't write to buf after we return.
366 v.logger.Debug("s3: waiting for ReadFull() to fail")
374 // Compare the given data with the stored data.
375 func (v *S3Volume) Compare(ctx context.Context, loc string, expect []byte) error {
376 errChan := make(chan error, 1)
378 _, err := v.bucket.Head("recent/"+loc, nil)
385 case err = <-errChan:
388 // Checking for "loc" itself here would interfere with
389 // future GET requests.
391 // On AWS, if X doesn't exist, a HEAD or GET request
392 // for X causes X's non-existence to be cached. Thus,
393 // if we test for X, then create X and return a
394 // signature to our client, the client might still get
395 // 404 from all keepstores when trying to read it.
397 // To avoid this, we avoid doing HEAD X or GET X until
398 // we know X has been written.
400 // Note that X might exist even though recent/X
401 // doesn't: for example, the response to HEAD recent/X
402 // might itself come from a stale cache. In such
403 // cases, we will return a false negative and
404 // PutHandler might needlessly create another replica
405 // on a different volume. That's not ideal, but it's
406 // better than passing the eventually-consistent
407 // problem on to our clients.
408 return v.translateError(err)
410 rdr, err := v.getReaderWithContext(ctx, loc)
415 return v.translateError(compareReaderWithBuf(ctx, rdr, expect, loc[:32]))
418 // Put writes a block.
419 func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error {
420 if v.volume.ReadOnly {
421 return MethodDisabledError
426 md5, err := hex.DecodeString(loc)
430 opts.ContentMD5 = base64.StdEncoding.EncodeToString(md5)
431 // In AWS regions that use V4 signatures, we need to
432 // provide ContentSHA256 up front. Otherwise, the S3
433 // library reads the request body (from our buffer)
434 // into another new buffer in order to compute the
435 // SHA256 before sending the request -- which would
436 // mean consuming 128 MiB of memory for the duration
437 // of a 64 MiB write.
438 opts.ContentSHA256 = fmt.Sprintf("%x", sha256.Sum256(block))
441 // Send the block data through a pipe, so that (if we need to)
442 // we can close the pipe early and abandon our PutReader()
443 // goroutine, without worrying about PutReader() accessing our
444 // block buffer after we release it.
445 bufr, bufw := io.Pipe()
447 io.Copy(bufw, bytes.NewReader(block))
452 ready := make(chan bool)
455 if ctx.Err() != nil {
456 v.logger.Debugf("abandoned PutReader goroutine finished with err: %s", err)
460 err = v.bucket.PutReader(loc, bufr, int64(size), "application/octet-stream", s3ACL, opts)
464 err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
468 v.logger.Debugf("taking PutReader's input away: %s", ctx.Err())
469 // Our pipe might be stuck in Write(), waiting for
470 // PutReader() to read. If so, un-stick it. This means
471 // PutReader will get corrupt data, but that's OK: the
472 // size and MD5 won't match, so the write will fail.
473 go io.Copy(ioutil.Discard, bufr)
474 // CloseWithError() will return once pending I/O is done.
475 bufw.CloseWithError(ctx.Err())
476 v.logger.Debugf("abandoning PutReader goroutine")
479 // Unblock pipe in case PutReader did not consume it.
480 io.Copy(ioutil.Discard, bufr)
481 return v.translateError(err)
485 // Touch sets the timestamp for the given locator to the current time.
486 func (v *S3Volume) Touch(loc string) error {
487 if v.volume.ReadOnly {
488 return MethodDisabledError
490 _, err := v.bucket.Head(loc, nil)
491 err = v.translateError(err)
492 if os.IsNotExist(err) && v.fixRace(loc) {
493 // The data object got trashed in a race, but fixRace
495 } else if err != nil {
498 err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
499 return v.translateError(err)
502 // Mtime returns the stored timestamp for the given locator.
503 func (v *S3Volume) Mtime(loc string) (time.Time, error) {
504 _, err := v.bucket.Head(loc, nil)
506 return zeroTime, v.translateError(err)
508 resp, err := v.bucket.Head("recent/"+loc, nil)
509 err = v.translateError(err)
510 if os.IsNotExist(err) {
511 // The data object X exists, but recent/X is missing.
512 err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
514 v.logger.WithError(err).Errorf("error creating %q", "recent/"+loc)
515 return zeroTime, v.translateError(err)
517 v.logger.Infof("created %q to migrate existing block to new storage scheme", "recent/"+loc)
518 resp, err = v.bucket.Head("recent/"+loc, nil)
520 v.logger.WithError(err).Errorf("HEAD failed after creating %q", "recent/"+loc)
521 return zeroTime, v.translateError(err)
523 } else if err != nil {
524 // HEAD recent/X failed for some other reason.
527 return v.lastModified(resp)
530 // IndexTo writes a complete list of locators with the given prefix
531 // for which Get() can retrieve data.
532 func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
533 // Use a merge sort to find matching sets of X and recent/X.
536 Bucket: v.bucket.Bucket(),
538 PageSize: v.IndexPageSize,
539 Stats: &v.bucket.stats,
543 Bucket: v.bucket.Bucket(),
544 Prefix: "recent/" + prefix,
545 PageSize: v.IndexPageSize,
546 Stats: &v.bucket.stats,
548 for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() {
550 // Conveniently, "recent/*" and "trash/*" are
551 // lexically greater than all hex-encoded data
552 // hashes, so stopping here avoids iterating
553 // over all of them needlessly with dataL.
556 if !v.isKeepBlock(data.Key) {
560 // stamp is the list entry we should use to report the
561 // last-modified time for this data block: it will be
562 // the recent/X entry if one exists, otherwise the
563 // entry for the data block itself.
566 // Advance to the corresponding recent/X marker, if any
567 for recent != nil && recentL.Error() == nil {
568 if cmp := strings.Compare(recent.Key[7:], data.Key); cmp < 0 {
569 recent = recentL.Next()
573 recent = recentL.Next()
576 // recent/X marker is missing: we'll
577 // use the timestamp on the data
582 if err := recentL.Error(); err != nil {
585 t, err := time.Parse(time.RFC3339, stamp.LastModified)
589 // We truncate sub-second precision here. Otherwise
590 // timestamps will never match the RFC1123-formatted
591 // Last-Modified values parsed by Mtime().
592 fmt.Fprintf(writer, "%s+%d %d\n", data.Key, data.Size, t.Unix()*1000000000)
597 // Trash a Keep block.
598 func (v *S3Volume) Trash(loc string) error {
599 if v.volume.ReadOnly {
600 return MethodDisabledError
602 if t, err := v.Mtime(loc); err != nil {
604 } else if time.Since(t) < v.cluster.Collections.BlobSigningTTL.Duration() {
607 if v.cluster.Collections.BlobTrashLifetime == 0 {
609 return ErrS3TrashDisabled
611 return v.translateError(v.bucket.Del(loc))
613 err := v.checkRaceWindow(loc)
617 err = v.safeCopy("trash/"+loc, loc)
621 return v.translateError(v.bucket.Del(loc))
624 // checkRaceWindow returns a non-nil error if trash/loc is, or might
625 // be, in the race window (i.e., it's not safe to trash loc).
626 func (v *S3Volume) checkRaceWindow(loc string) error {
627 resp, err := v.bucket.Head("trash/"+loc, nil)
628 err = v.translateError(err)
629 if os.IsNotExist(err) {
630 // OK, trash/X doesn't exist so we're not in the race
633 } else if err != nil {
634 // Error looking up trash/X. We don't know whether
635 // we're in the race window
638 t, err := v.lastModified(resp)
640 // Can't parse timestamp
643 safeWindow := t.Add(v.cluster.Collections.BlobTrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow)))
645 // We can't count on "touch trash/X" to prolong
646 // trash/X's lifetime. The new timestamp might not
647 // become visible until now+raceWindow, and EmptyTrash
648 // is allowed to delete trash/X before then.
649 return fmt.Errorf("same block is already in trash, and safe window ended %s ago", -safeWindow)
651 // trash/X exists, but it won't be eligible for deletion until
652 // after now+raceWindow, so it's safe to overwrite it.
656 // safeCopy calls PutCopy, and checks the response to make sure the
657 // copy succeeded and updated the timestamp on the destination object
658 // (PutCopy returns 200 OK if the request was received, even if the
660 func (v *S3Volume) safeCopy(dst, src string) error {
661 resp, err := v.bucket.Bucket().PutCopy(dst, s3ACL, s3.CopyOptions{
662 ContentType: "application/octet-stream",
663 MetadataDirective: "REPLACE",
664 }, v.bucket.Bucket().Name+"/"+src)
665 err = v.translateError(err)
666 if os.IsNotExist(err) {
668 } else if err != nil {
669 return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.Bucket().Name+"/"+src, err)
671 if t, err := time.Parse(time.RFC3339Nano, resp.LastModified); err != nil {
672 return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.LastModified, err)
673 } else if time.Now().Sub(t) > maxClockSkew {
674 return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.LastModified, t)
679 // Get the LastModified header from resp, and parse it as RFC1123 or
680 // -- if it isn't valid RFC1123 -- as Amazon's variant of RFC1123.
681 func (v *S3Volume) lastModified(resp *http.Response) (t time.Time, err error) {
682 s := resp.Header.Get("Last-Modified")
683 t, err = time.Parse(time.RFC1123, s)
684 if err != nil && s != "" {
685 // AWS example is "Sun, 1 Jan 2006 12:00:00 GMT",
686 // which isn't quite "Sun, 01 Jan 2006 12:00:00 GMT"
687 // as required by HTTP spec. If it's not a valid HTTP
688 // header value, it's probably AWS (or s3test) giving
689 // us a nearly-RFC1123 timestamp.
690 t, err = time.Parse(nearlyRFC1123, s)
695 // Untrash moves block from trash back into store
696 func (v *S3Volume) Untrash(loc string) error {
697 err := v.safeCopy(loc, "trash/"+loc)
701 err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
702 return v.translateError(err)
705 // Status returns a *VolumeStatus representing the current in-use
706 // storage capacity and a fake available capacity that doesn't make
707 // the volume seem full or nearly-full.
708 func (v *S3Volume) Status() *VolumeStatus {
709 return &VolumeStatus{
711 BytesFree: BlockSize * 1000,
716 // InternalStats returns bucket I/O and API call counters.
717 func (v *S3Volume) InternalStats() interface{} {
718 return &v.bucket.stats
721 // String implements fmt.Stringer.
722 func (v *S3Volume) String() string {
723 return fmt.Sprintf("s3-bucket:%+q", v.Bucket)
726 var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
728 func (v *S3Volume) isKeepBlock(s string) bool {
729 return s3KeepBlockRegexp.MatchString(s)
732 // fixRace(X) is called when "recent/X" exists but "X" doesn't
733 // exist. If the timestamps on "recent/"+loc and "trash/"+loc indicate
734 // there was a race between Put and Trash, fixRace recovers from the
735 // race by Untrashing the block.
736 func (v *S3Volume) fixRace(loc string) bool {
737 trash, err := v.bucket.Head("trash/"+loc, nil)
739 if !os.IsNotExist(v.translateError(err)) {
740 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "trash/"+loc)
744 trashTime, err := v.lastModified(trash)
746 v.logger.WithError(err).Errorf("fixRace: error parsing time %q", trash.Header.Get("Last-Modified"))
750 recent, err := v.bucket.Head("recent/"+loc, nil)
752 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "recent/"+loc)
755 recentTime, err := v.lastModified(recent)
757 v.logger.WithError(err).Errorf("fixRace: error parsing time %q", recent.Header.Get("Last-Modified"))
761 ageWhenTrashed := trashTime.Sub(recentTime)
762 if ageWhenTrashed >= v.cluster.Collections.BlobSigningTTL.Duration() {
763 // No evidence of a race: block hasn't been written
764 // since it became eligible for Trash. No fix needed.
768 v.logger.Infof("fixRace: %q: trashed at %s but touched at %s (age when trashed = %s < %s)", loc, trashTime, recentTime, ageWhenTrashed, v.cluster.Collections.BlobSigningTTL)
769 v.logger.Infof("fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+loc, loc)
770 err = v.safeCopy(loc, "trash/"+loc)
772 v.logger.WithError(err).Error("fixRace: copy failed")
778 func (v *S3Volume) translateError(err error) error {
779 switch err := err.(type) {
781 if (err.StatusCode == http.StatusNotFound && err.Code == "NoSuchKey") ||
782 strings.Contains(err.Error(), "Not Found") {
783 return os.ErrNotExist
785 // Other 404 errors like NoSuchVersion and
786 // NoSuchBucket are different problems which should
787 // get called out downstream, so we don't convert them
788 // to os.ErrNotExist.
793 // EmptyTrash looks for trashed blocks that exceeded BlobTrashLifetime
794 // and deletes them from the volume.
795 func (v *S3Volume) EmptyTrash() {
796 if v.cluster.Collections.BlobDeleteConcurrency < 1 {
800 var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64
802 // Define "ready to delete" as "...when EmptyTrash started".
805 emptyOneKey := func(trash *s3.Key) {
807 if !v.isKeepBlock(loc) {
810 atomic.AddInt64(&bytesInTrash, trash.Size)
811 atomic.AddInt64(&blocksInTrash, 1)
813 trashT, err := time.Parse(time.RFC3339, trash.LastModified)
815 v.logger.Warnf("EmptyTrash: %q: parse %q: %s", trash.Key, trash.LastModified, err)
818 recent, err := v.bucket.Head("recent/"+loc, nil)
819 if err != nil && os.IsNotExist(v.translateError(err)) {
820 v.logger.Warnf("EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", trash.Key, "recent/"+loc, err)
823 v.logger.WithError(err).Errorf("EmptyTrash: Untrash(%q) failed", loc)
826 } else if err != nil {
827 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", "recent/"+loc)
830 recentT, err := v.lastModified(recent)
832 v.logger.WithError(err).Warnf("EmptyTrash: %q: error parsing %q", "recent/"+loc, recent.Header.Get("Last-Modified"))
835 if trashT.Sub(recentT) < v.cluster.Collections.BlobSigningTTL.Duration() {
836 if age := startT.Sub(recentT); age >= v.cluster.Collections.BlobSigningTTL.Duration()-time.Duration(v.RaceWindow) {
837 // recent/loc is too old to protect
838 // loc from being Trashed again during
839 // the raceWindow that starts if we
840 // delete trash/X now.
842 // Note this means (TrashSweepInterval
843 // < BlobSigningTTL - raceWindow) is
844 // necessary to avoid starvation.
845 v.logger.Infof("EmptyTrash: detected old race for %q, calling fixRace + Touch", loc)
850 _, err := v.bucket.Head(loc, nil)
851 if os.IsNotExist(err) {
852 v.logger.Infof("EmptyTrash: detected recent race for %q, calling fixRace", loc)
855 } else if err != nil {
856 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc)
860 if startT.Sub(trashT) < v.cluster.Collections.BlobTrashLifetime.Duration() {
863 err = v.bucket.Del(trash.Key)
865 v.logger.WithError(err).Errorf("EmptyTrash: error deleting %q", trash.Key)
868 atomic.AddInt64(&bytesDeleted, trash.Size)
869 atomic.AddInt64(&blocksDeleted, 1)
871 _, err = v.bucket.Head(loc, nil)
873 v.logger.Warnf("EmptyTrash: HEAD %q succeeded immediately after deleting %q", loc, loc)
876 if !os.IsNotExist(v.translateError(err)) {
877 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc)
880 err = v.bucket.Del("recent/" + loc)
882 v.logger.WithError(err).Warnf("EmptyTrash: error deleting %q", "recent/"+loc)
886 var wg sync.WaitGroup
887 todo := make(chan *s3.Key, v.cluster.Collections.BlobDeleteConcurrency)
888 for i := 0; i < v.cluster.Collections.BlobDeleteConcurrency; i++ {
892 for key := range todo {
900 Bucket: v.bucket.Bucket(),
902 PageSize: v.IndexPageSize,
903 Stats: &v.bucket.stats,
905 for trash := trashL.First(); trash != nil; trash = trashL.Next() {
911 if err := trashL.Error(); err != nil {
912 v.logger.WithError(err).Error("EmptyTrash: lister failed")
914 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)
917 type s3Lister struct {
918 Logger logrus.FieldLogger
928 // First fetches the first page and returns the first item. It returns
929 // nil if the response is the empty set or an error occurs.
930 func (lister *s3Lister) First() *s3.Key {
935 // Next returns the next item, fetching the next page if necessary. It
936 // returns nil if the last available item has already been fetched, or
938 func (lister *s3Lister) Next() *s3.Key {
939 if len(lister.buf) == 0 && lister.nextMarker != "" {
945 // Return the most recent error encountered by First or Next.
946 func (lister *s3Lister) Error() error {
950 func (lister *s3Lister) getPage() {
951 lister.Stats.TickOps("list")
952 lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps)
953 resp, err := lister.Bucket.List(lister.Prefix, "", lister.nextMarker, lister.PageSize)
954 lister.nextMarker = ""
959 if resp.IsTruncated {
960 lister.nextMarker = resp.NextMarker
962 lister.buf = make([]s3.Key, 0, len(resp.Contents))
963 for _, key := range resp.Contents {
964 if !strings.HasPrefix(key.Key, lister.Prefix) {
965 lister.Logger.Warnf("s3Lister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, key.Key)
968 lister.buf = append(lister.buf, key)
972 func (lister *s3Lister) pop() (k *s3.Key) {
973 if len(lister.buf) > 0 {
975 lister.buf = lister.buf[1:]
980 // s3bucket wraps s3.bucket and counts I/O and API usage stats. The
981 // wrapped bucket can be replaced atomically with SetBucket in order
982 // to update credentials.
983 type s3bucket struct {
989 func (b *s3bucket) Bucket() *s3.Bucket {
995 func (b *s3bucket) SetBucket(bucket *s3.Bucket) {
1001 func (b *s3bucket) GetReader(path string) (io.ReadCloser, error) {
1002 rdr, err := b.Bucket().GetReader(path)
1003 b.stats.TickOps("get")
1004 b.stats.Tick(&b.stats.Ops, &b.stats.GetOps)
1005 b.stats.TickErr(err)
1006 return NewCountingReader(rdr, b.stats.TickInBytes), err
1009 func (b *s3bucket) Head(path string, headers map[string][]string) (*http.Response, error) {
1010 resp, err := b.Bucket().Head(path, headers)
1011 b.stats.TickOps("head")
1012 b.stats.Tick(&b.stats.Ops, &b.stats.HeadOps)
1013 b.stats.TickErr(err)
1017 func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error {
1019 // goamz will only send Content-Length: 0 when reader
1020 // is nil due to net.http.Request.ContentLength
1021 // behavior. Otherwise, Content-Length header is
1022 // omitted which will cause some S3 services
1023 // (including AWS and Ceph RadosGW) to fail to create
1027 r = NewCountingReader(r, b.stats.TickOutBytes)
1029 err := b.Bucket().PutReader(path, r, length, contType, perm, options)
1030 b.stats.TickOps("put")
1031 b.stats.Tick(&b.stats.Ops, &b.stats.PutOps)
1032 b.stats.TickErr(err)
1036 func (b *s3bucket) Del(path string) error {
1037 err := b.Bucket().Del(path)
1038 b.stats.TickOps("delete")
1039 b.stats.Tick(&b.stats.Ops, &b.stats.DelOps)
1040 b.stats.TickErr(err)
1044 type s3bucketStats struct {
1054 func (s *s3bucketStats) TickErr(err error) {
1058 errType := fmt.Sprintf("%T", err)
1059 if err, ok := err.(*s3.Error); ok {
1060 errType = errType + fmt.Sprintf(" %d %s", err.StatusCode, err.Code)
1062 s.statsTicker.TickErr(err, errType)