1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
23 "git.arvados.org/arvados.git/sdk/go/arvados"
24 "github.com/aws/aws-sdk-go-v2/aws"
25 "github.com/aws/aws-sdk-go-v2/aws/awserr"
26 "github.com/aws/aws-sdk-go-v2/aws/defaults"
27 "github.com/aws/aws-sdk-go-v2/aws/ec2metadata"
28 "github.com/aws/aws-sdk-go-v2/aws/ec2rolecreds"
29 "github.com/aws/aws-sdk-go-v2/aws/endpoints"
30 "github.com/aws/aws-sdk-go-v2/service/s3"
31 "github.com/aws/aws-sdk-go-v2/service/s3/s3manager"
32 "github.com/prometheus/client_golang/prometheus"
33 "github.com/sirupsen/logrus"
36 // S3AWSVolume implements Volume using an S3 bucket.
37 type S3AWSVolume struct {
38 arvados.S3VolumeDriverParameters
39 AuthToken string // populated automatically when IAMRole is used
40 AuthExpiration time.Time // populated automatically when IAMRole is used
42 cluster *arvados.Cluster
44 logger logrus.FieldLogger
45 metrics *volumeMetricsVecs
51 // s3bucket wraps s3.bucket and counts I/O and API usage stats. The
52 // wrapped bucket can be replaced atomically with SetBucket in order
53 // to update credentials.
54 type s3AWSbucket struct {
57 stats s3awsbucketStats
61 // chooseS3VolumeDriver distinguishes between the old goamz driver and
62 // aws-sdk-go based on the UseAWSS3v2Driver feature flag
63 func chooseS3VolumeDriver(cluster *arvados.Cluster, volume arvados.Volume, logger logrus.FieldLogger, metrics *volumeMetricsVecs) (Volume, error) {
64 v := &S3Volume{cluster: cluster, volume: volume, metrics: metrics}
65 // Default value will be overriden if it happens to be defined in the config
66 v.S3VolumeDriverParameters.UseAWSS3v2Driver = true
67 err := json.Unmarshal(volume.DriverParameters, v)
71 if v.UseAWSS3v2Driver {
72 logger.Debugln("Using AWS S3 v2 driver")
73 return newS3AWSVolume(cluster, volume, logger, metrics)
75 logger.Debugln("Using goamz S3 driver")
76 return newS3Volume(cluster, volume, logger, metrics)
80 PartSize = 5 * 1024 * 1024
85 var s3AWSKeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
86 var s3AWSZeroTime time.Time
88 func (v *S3AWSVolume) isKeepBlock(s string) (string, bool) {
89 if v.PrefixLength > 0 && len(s) == v.PrefixLength+33 && s[:v.PrefixLength] == s[v.PrefixLength+1:v.PrefixLength*2+1] {
90 s = s[v.PrefixLength+1:]
92 return s, s3AWSKeepBlockRegexp.MatchString(s)
95 // Return the key used for a given loc. If PrefixLength==0 then
96 // key("abcdef0123") is "abcdef0123", if PrefixLength==3 then key is
97 // "abc/abcdef0123", etc.
98 func (v *S3AWSVolume) key(loc string) string {
99 if v.PrefixLength > 0 && v.PrefixLength < len(loc)-1 {
100 return loc[:v.PrefixLength] + "/" + loc
106 func newS3AWSVolume(cluster *arvados.Cluster, volume arvados.Volume, logger logrus.FieldLogger, metrics *volumeMetricsVecs) (Volume, error) {
107 v := &S3AWSVolume{cluster: cluster, volume: volume, metrics: metrics}
108 err := json.Unmarshal(volume.DriverParameters, v)
112 v.logger = logger.WithField("Volume", v.String())
113 return v, v.check("")
116 func (v *S3AWSVolume) translateError(err error) error {
117 if _, ok := err.(*aws.RequestCanceledError); ok {
118 return context.Canceled
119 } else if aerr, ok := err.(awserr.Error); ok {
120 if aerr.Code() == "NotFound" {
121 return os.ErrNotExist
122 } else if aerr.Code() == "NoSuchKey" {
123 return os.ErrNotExist
129 // safeCopy calls CopyObjectRequest, and checks the response to make
130 // sure the copy succeeded and updated the timestamp on the
131 // destination object
133 // (If something goes wrong during the copy, the error will be
134 // embedded in the 200 OK response)
135 func (v *S3AWSVolume) safeCopy(dst, src string) error {
136 input := &s3.CopyObjectInput{
137 Bucket: aws.String(v.bucket.bucket),
138 ContentType: aws.String("application/octet-stream"),
139 CopySource: aws.String(v.bucket.bucket + "/" + src),
140 Key: aws.String(dst),
143 req := v.bucket.svc.CopyObjectRequest(input)
144 resp, err := req.Send(context.Background())
146 err = v.translateError(err)
147 if os.IsNotExist(err) {
149 } else if err != nil {
150 return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.bucket+"/"+src, err)
153 if resp.CopyObjectResult.LastModified == nil {
154 return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.CopyObjectResult.LastModified, err)
155 } else if time.Now().Sub(*resp.CopyObjectResult.LastModified) > maxClockSkew {
156 return fmt.Errorf("PutCopy succeeded but returned an old timestamp: %q: %s", resp.CopyObjectResult.LastModified, resp.CopyObjectResult.LastModified)
161 func (v *S3AWSVolume) check(ec2metadataHostname string) error {
163 return errors.New("DriverParameters: Bucket must be provided")
165 if v.IndexPageSize == 0 {
166 v.IndexPageSize = 1000
168 if v.RaceWindow < 0 {
169 return errors.New("DriverParameters: RaceWindow must not be negative")
173 return errors.New("DriverParameters: V2Signature is not supported")
176 defaultResolver := endpoints.NewDefaultResolver()
178 cfg := defaults.Config()
180 if v.Endpoint == "" && v.Region == "" {
181 return fmt.Errorf("AWS region or endpoint must be specified")
182 } else if v.Endpoint != "" || ec2metadataHostname != "" {
183 myCustomResolver := func(service, region string) (aws.Endpoint, error) {
184 if v.Endpoint != "" && service == "s3" {
187 SigningRegion: v.Region,
189 } else if service == "ec2metadata" && ec2metadataHostname != "" {
191 URL: ec2metadataHostname,
195 return defaultResolver.ResolveEndpoint(service, region)
197 cfg.EndpointResolver = aws.EndpointResolverFunc(myCustomResolver)
200 cfg.Region = v.Region
202 // Zero timeouts mean "wait forever", which is a bad
203 // default. Default to long timeouts instead.
204 if v.ConnectTimeout == 0 {
205 v.ConnectTimeout = s3DefaultConnectTimeout
207 if v.ReadTimeout == 0 {
208 v.ReadTimeout = s3DefaultReadTimeout
211 creds := aws.NewChainProvider(
212 []aws.CredentialsProvider{
213 aws.NewStaticCredentialsProvider(v.AccessKeyID, v.SecretAccessKey, v.AuthToken),
214 ec2rolecreds.New(ec2metadata.New(cfg)),
217 cfg.Credentials = creds
219 v.bucket = &s3AWSbucket{
224 // Set up prometheus metrics
225 lbls := prometheus.Labels{"device_id": v.GetDeviceID()}
226 v.bucket.stats.opsCounters, v.bucket.stats.errCounters, v.bucket.stats.ioBytes = v.metrics.getCounterVecsFor(lbls)
231 // String implements fmt.Stringer.
232 func (v *S3AWSVolume) String() string {
233 return fmt.Sprintf("s3-bucket:%+q", v.Bucket)
236 // GetDeviceID returns a globally unique ID for the storage bucket.
237 func (v *S3AWSVolume) GetDeviceID() string {
238 return "s3://" + v.Endpoint + "/" + v.Bucket
241 // Compare the given data with the stored data.
242 func (v *S3AWSVolume) Compare(ctx context.Context, loc string, expect []byte) error {
244 errChan := make(chan error, 1)
246 _, err := v.head("recent/" + key)
253 case err = <-errChan:
256 // Checking for the key itself here would interfere
257 // with future GET requests.
259 // On AWS, if X doesn't exist, a HEAD or GET request
260 // for X causes X's non-existence to be cached. Thus,
261 // if we test for X, then create X and return a
262 // signature to our client, the client might still get
263 // 404 from all keepstores when trying to read it.
265 // To avoid this, we avoid doing HEAD X or GET X until
266 // we know X has been written.
268 // Note that X might exist even though recent/X
269 // doesn't: for example, the response to HEAD recent/X
270 // might itself come from a stale cache. In such
271 // cases, we will return a false negative and
272 // PutHandler might needlessly create another replica
273 // on a different volume. That's not ideal, but it's
274 // better than passing the eventually-consistent
275 // problem on to our clients.
276 return v.translateError(err)
279 input := &s3.GetObjectInput{
280 Bucket: aws.String(v.bucket.bucket),
281 Key: aws.String(key),
284 req := v.bucket.svc.GetObjectRequest(input)
285 result, err := req.Send(ctx)
287 return v.translateError(err)
289 return v.translateError(compareReaderWithBuf(ctx, result.Body, expect, loc[:32]))
292 // EmptyTrash looks for trashed blocks that exceeded BlobTrashLifetime
293 // and deletes them from the volume.
294 func (v *S3AWSVolume) EmptyTrash() {
295 if v.cluster.Collections.BlobDeleteConcurrency < 1 {
299 var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64
301 // Define "ready to delete" as "...when EmptyTrash started".
304 emptyOneKey := func(trash *s3.Object) {
305 key := strings.TrimPrefix(*trash.Key, "trash/")
306 loc, isblk := v.isKeepBlock(key)
310 atomic.AddInt64(&bytesInTrash, *trash.Size)
311 atomic.AddInt64(&blocksInTrash, 1)
313 trashT := *trash.LastModified
314 recent, err := v.head("recent/" + key)
315 if err != nil && os.IsNotExist(v.translateError(err)) {
316 v.logger.Warnf("EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", *trash.Key, "recent/"+key, err)
319 v.logger.WithError(err).Errorf("EmptyTrash: Untrash(%q) failed", loc)
322 } else if err != nil {
323 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", "recent/"+key)
326 if trashT.Sub(*recent.LastModified) < v.cluster.Collections.BlobSigningTTL.Duration() {
327 if age := startT.Sub(*recent.LastModified); age >= v.cluster.Collections.BlobSigningTTL.Duration()-time.Duration(v.RaceWindow) {
328 // recent/key is too old to protect
329 // loc from being Trashed again during
330 // the raceWindow that starts if we
331 // delete trash/X now.
333 // Note this means (TrashSweepInterval
334 // < BlobSigningTTL - raceWindow) is
335 // necessary to avoid starvation.
336 v.logger.Infof("EmptyTrash: detected old race for %q, calling fixRace + Touch", loc)
341 _, err := v.head(key)
342 if os.IsNotExist(err) {
343 v.logger.Infof("EmptyTrash: detected recent race for %q, calling fixRace", loc)
346 } else if err != nil {
347 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", loc)
351 if startT.Sub(trashT) < v.cluster.Collections.BlobTrashLifetime.Duration() {
354 err = v.bucket.Del(*trash.Key)
356 v.logger.WithError(err).Errorf("EmptyTrash: error deleting %q", *trash.Key)
359 atomic.AddInt64(&bytesDeleted, *trash.Size)
360 atomic.AddInt64(&blocksDeleted, 1)
362 _, err = v.head(*trash.Key)
364 v.logger.Warnf("EmptyTrash: HEAD %q succeeded immediately after deleting %q", loc, loc)
367 if !os.IsNotExist(v.translateError(err)) {
368 v.logger.WithError(err).Warnf("EmptyTrash: HEAD %q failed", key)
371 err = v.bucket.Del("recent/" + key)
373 v.logger.WithError(err).Warnf("EmptyTrash: error deleting %q", "recent/"+key)
377 var wg sync.WaitGroup
378 todo := make(chan *s3.Object, v.cluster.Collections.BlobDeleteConcurrency)
379 for i := 0; i < v.cluster.Collections.BlobDeleteConcurrency; i++ {
383 for key := range todo {
389 trashL := s3awsLister{
393 PageSize: v.IndexPageSize,
394 Stats: &v.bucket.stats,
396 for trash := trashL.First(); trash != nil; trash = trashL.Next() {
402 if err := trashL.Error(); err != nil {
403 v.logger.WithError(err).Error("EmptyTrash: lister failed")
405 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)
408 // fixRace(X) is called when "recent/X" exists but "X" doesn't
409 // exist. If the timestamps on "recent/X" and "trash/X" indicate there
410 // was a race between Put and Trash, fixRace recovers from the race by
411 // Untrashing the block.
412 func (v *S3AWSVolume) fixRace(key string) bool {
413 trash, err := v.head("trash/" + key)
415 if !os.IsNotExist(v.translateError(err)) {
416 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "trash/"+key)
421 recent, err := v.head("recent/" + key)
423 v.logger.WithError(err).Errorf("fixRace: HEAD %q failed", "recent/"+key)
427 recentTime := *recent.LastModified
428 trashTime := *trash.LastModified
429 ageWhenTrashed := trashTime.Sub(recentTime)
430 if ageWhenTrashed >= v.cluster.Collections.BlobSigningTTL.Duration() {
431 // No evidence of a race: block hasn't been written
432 // since it became eligible for Trash. No fix needed.
436 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)
437 v.logger.Infof("fixRace: copying %q to %q to recover from race between Put/Touch and Trash", "recent/"+key, key)
438 err = v.safeCopy(key, "trash/"+key)
440 v.logger.WithError(err).Error("fixRace: copy failed")
446 func (v *S3AWSVolume) head(key string) (result *s3.HeadObjectOutput, err error) {
447 input := &s3.HeadObjectInput{
448 Bucket: aws.String(v.bucket.bucket),
449 Key: aws.String(key),
452 req := v.bucket.svc.HeadObjectRequest(input)
453 res, err := req.Send(context.TODO())
455 v.bucket.stats.TickOps("head")
456 v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.HeadOps)
457 v.bucket.stats.TickErr(err)
460 return nil, v.translateError(err)
462 result = res.HeadObjectOutput
466 // Get a block: copy the block data into buf, and return the number of
468 func (v *S3AWSVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
469 // Do not use getWithPipe here: the BlockReader interface does not pass
470 // through 'buf []byte', and we don't want to allocate two buffers for each
471 // read request. Instead, use a version of ReadBlock that accepts 'buf []byte'
474 count, err := v.readWorker(ctx, key, buf)
479 err = v.translateError(err)
480 if !os.IsNotExist(err) {
484 _, err = v.head("recent/" + key)
485 err = v.translateError(err)
487 // If we can't read recent/X, there's no point in
488 // trying fixRace. Give up.
496 count, err = v.readWorker(ctx, key, buf)
498 v.logger.Warnf("reading %s after successful fixRace: %s", loc, err)
499 err = v.translateError(err)
505 func (v *S3AWSVolume) readWorker(ctx context.Context, key string, buf []byte) (int, error) {
506 awsBuf := aws.NewWriteAtBuffer(buf)
507 downloader := s3manager.NewDownloaderWithClient(v.bucket.svc, func(u *s3manager.Downloader) {
508 u.PartSize = PartSize
509 u.Concurrency = ReadConcurrency
512 v.logger.Debugf("Partsize: %d; Concurrency: %d\n", downloader.PartSize, downloader.Concurrency)
514 count, err := downloader.DownloadWithContext(ctx, awsBuf, &s3.GetObjectInput{
515 Bucket: aws.String(v.bucket.bucket),
516 Key: aws.String(key),
518 v.bucket.stats.TickOps("get")
519 v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.GetOps)
520 v.bucket.stats.TickErr(err)
521 v.bucket.stats.TickInBytes(uint64(count))
522 return int(count), v.translateError(err)
525 func (v *S3AWSVolume) writeObject(ctx context.Context, key string, r io.Reader) error {
527 // r == nil leads to a memory violation in func readFillBuf in
528 // aws-sdk-go-v2@v0.23.0/service/s3/s3manager/upload.go
529 r = bytes.NewReader(nil)
532 uploadInput := s3manager.UploadInput{
533 Bucket: aws.String(v.bucket.bucket),
534 Key: aws.String(key),
538 if loc, ok := v.isKeepBlock(key); ok {
539 var contentMD5 string
540 md5, err := hex.DecodeString(loc)
542 return v.translateError(err)
544 contentMD5 = base64.StdEncoding.EncodeToString(md5)
545 uploadInput.ContentMD5 = &contentMD5
548 // Experimentation indicated that using concurrency 5 yields the best
549 // throughput, better than higher concurrency (10 or 13) by ~5%.
550 // Defining u.BufferProvider = s3manager.NewBufferedReadSeekerWriteToPool(64 * 1024 * 1024)
551 // is detrimental to througput (minus ~15%).
552 uploader := s3manager.NewUploaderWithClient(v.bucket.svc, func(u *s3manager.Uploader) {
553 u.PartSize = PartSize
554 u.Concurrency = WriteConcurrency
557 // Unlike the goamz S3 driver, we don't need to precompute ContentSHA256:
558 // the aws-sdk-go v2 SDK uses a ReadSeeker to avoid having to copy the
559 // block, so there is no extra memory use to be concerned about. See
560 // makeSha256Reader in aws/signer/v4/v4.go. In fact, we explicitly disable
561 // calculating the Sha-256 because we don't need it; we already use md5sum
562 // hashes that match the name of the block.
563 _, err := uploader.UploadWithContext(ctx, &uploadInput, s3manager.WithUploaderRequestOptions(func(r *aws.Request) {
564 r.HTTPRequest.Header.Set("X-Amz-Content-Sha256", "UNSIGNED-PAYLOAD")
567 v.bucket.stats.TickOps("put")
568 v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.PutOps)
569 v.bucket.stats.TickErr(err)
571 return v.translateError(err)
574 // Put writes a block.
575 func (v *S3AWSVolume) Put(ctx context.Context, loc string, block []byte) error {
576 // Do not use putWithPipe here; we want to pass an io.ReadSeeker to the S3
577 // sdk to avoid memory allocation there. See #17339 for more information.
578 if v.volume.ReadOnly {
579 return MethodDisabledError
582 rdr := bytes.NewReader(block)
583 r := NewCountingReaderAtSeeker(rdr, v.bucket.stats.TickOutBytes)
585 err := v.writeObject(ctx, key, r)
589 return v.writeObject(ctx, "recent/"+key, nil)
592 type s3awsLister struct {
593 Logger logrus.FieldLogger
597 Stats *s3awsbucketStats
598 ContinuationToken string
603 // First fetches the first page and returns the first item. It returns
604 // nil if the response is the empty set or an error occurs.
605 func (lister *s3awsLister) First() *s3.Object {
610 // Next returns the next item, fetching the next page if necessary. It
611 // returns nil if the last available item has already been fetched, or
613 func (lister *s3awsLister) Next() *s3.Object {
614 if len(lister.buf) == 0 && lister.ContinuationToken != "" {
620 // Return the most recent error encountered by First or Next.
621 func (lister *s3awsLister) Error() error {
625 func (lister *s3awsLister) getPage() {
626 lister.Stats.TickOps("list")
627 lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps)
629 var input *s3.ListObjectsV2Input
630 if lister.ContinuationToken == "" {
631 input = &s3.ListObjectsV2Input{
632 Bucket: aws.String(lister.Bucket.bucket),
633 MaxKeys: aws.Int64(int64(lister.PageSize)),
634 Prefix: aws.String(lister.Prefix),
637 input = &s3.ListObjectsV2Input{
638 Bucket: aws.String(lister.Bucket.bucket),
639 MaxKeys: aws.Int64(int64(lister.PageSize)),
640 Prefix: aws.String(lister.Prefix),
641 ContinuationToken: &lister.ContinuationToken,
645 req := lister.Bucket.svc.ListObjectsV2Request(input)
646 resp, err := req.Send(context.Background())
648 if aerr, ok := err.(awserr.Error); ok {
656 if *resp.IsTruncated {
657 lister.ContinuationToken = *resp.NextContinuationToken
659 lister.ContinuationToken = ""
661 lister.buf = make([]s3.Object, 0, len(resp.Contents))
662 for _, key := range resp.Contents {
663 if !strings.HasPrefix(*key.Key, lister.Prefix) {
664 lister.Logger.Warnf("s3awsLister: S3 Bucket.List(prefix=%q) returned key %q", lister.Prefix, *key.Key)
667 lister.buf = append(lister.buf, key)
671 func (lister *s3awsLister) pop() (k *s3.Object) {
672 if len(lister.buf) > 0 {
674 lister.buf = lister.buf[1:]
679 // IndexTo writes a complete list of locators with the given prefix
680 // for which Get() can retrieve data.
681 func (v *S3AWSVolume) IndexTo(prefix string, writer io.Writer) error {
682 prefix = v.key(prefix)
683 // Use a merge sort to find matching sets of X and recent/X.
684 dataL := s3awsLister{
688 PageSize: v.IndexPageSize,
689 Stats: &v.bucket.stats,
691 recentL := s3awsLister{
694 Prefix: "recent/" + prefix,
695 PageSize: v.IndexPageSize,
696 Stats: &v.bucket.stats,
698 for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() {
699 if *data.Key >= "g" {
700 // Conveniently, "recent/*" and "trash/*" are
701 // lexically greater than all hex-encoded data
702 // hashes, so stopping here avoids iterating
703 // over all of them needlessly with dataL.
706 loc, isblk := v.isKeepBlock(*data.Key)
711 // stamp is the list entry we should use to report the
712 // last-modified time for this data block: it will be
713 // the recent/X entry if one exists, otherwise the
714 // entry for the data block itself.
717 // Advance to the corresponding recent/X marker, if any
718 for recent != nil && recentL.Error() == nil {
719 if cmp := strings.Compare((*recent.Key)[7:], *data.Key); cmp < 0 {
720 recent = recentL.Next()
724 recent = recentL.Next()
727 // recent/X marker is missing: we'll
728 // use the timestamp on the data
733 if err := recentL.Error(); err != nil {
736 // We truncate sub-second precision here. Otherwise
737 // timestamps will never match the RFC1123-formatted
738 // Last-Modified values parsed by Mtime().
739 fmt.Fprintf(writer, "%s+%d %d\n", loc, *data.Size, stamp.LastModified.Unix()*1000000000)
744 // Mtime returns the stored timestamp for the given locator.
745 func (v *S3AWSVolume) Mtime(loc string) (time.Time, error) {
747 _, err := v.head(key)
749 return s3AWSZeroTime, v.translateError(err)
751 resp, err := v.head("recent/" + key)
752 err = v.translateError(err)
753 if os.IsNotExist(err) {
754 // The data object X exists, but recent/X is missing.
755 err = v.writeObject(context.Background(), "recent/"+key, nil)
757 v.logger.WithError(err).Errorf("error creating %q", "recent/"+key)
758 return s3AWSZeroTime, v.translateError(err)
760 v.logger.Infof("Mtime: created %q to migrate existing block to new storage scheme", "recent/"+key)
761 resp, err = v.head("recent/" + key)
763 v.logger.WithError(err).Errorf("HEAD failed after creating %q", "recent/"+key)
764 return s3AWSZeroTime, v.translateError(err)
766 } else if err != nil {
767 // HEAD recent/X failed for some other reason.
768 return s3AWSZeroTime, err
770 return *resp.LastModified, err
773 // Status returns a *VolumeStatus representing the current in-use
774 // storage capacity and a fake available capacity that doesn't make
775 // the volume seem full or nearly-full.
776 func (v *S3AWSVolume) Status() *VolumeStatus {
777 return &VolumeStatus{
779 BytesFree: BlockSize * 1000,
784 // InternalStats returns bucket I/O and API call counters.
785 func (v *S3AWSVolume) InternalStats() interface{} {
786 return &v.bucket.stats
789 // Touch sets the timestamp for the given locator to the current time.
790 func (v *S3AWSVolume) Touch(loc string) error {
791 if v.volume.ReadOnly {
792 return MethodDisabledError
795 _, err := v.head(key)
796 err = v.translateError(err)
797 if os.IsNotExist(err) && v.fixRace(key) {
798 // The data object got trashed in a race, but fixRace
800 } else if err != nil {
803 err = v.writeObject(context.Background(), "recent/"+key, nil)
804 return v.translateError(err)
807 // checkRaceWindow returns a non-nil error if trash/key is, or might
808 // be, in the race window (i.e., it's not safe to trash key).
809 func (v *S3AWSVolume) checkRaceWindow(key string) error {
810 resp, err := v.head("trash/" + key)
811 err = v.translateError(err)
812 if os.IsNotExist(err) {
813 // OK, trash/X doesn't exist so we're not in the race
816 } else if err != nil {
817 // Error looking up trash/X. We don't know whether
818 // we're in the race window
821 t := resp.LastModified
822 safeWindow := t.Add(v.cluster.Collections.BlobTrashLifetime.Duration()).Sub(time.Now().Add(time.Duration(v.RaceWindow)))
824 // We can't count on "touch trash/X" to prolong
825 // trash/X's lifetime. The new timestamp might not
826 // become visible until now+raceWindow, and EmptyTrash
827 // is allowed to delete trash/X before then.
828 return fmt.Errorf("%s: same block is already in trash, and safe window ended %s ago", key, -safeWindow)
830 // trash/X exists, but it won't be eligible for deletion until
831 // after now+raceWindow, so it's safe to overwrite it.
835 func (b *s3AWSbucket) Del(path string) error {
836 input := &s3.DeleteObjectInput{
837 Bucket: aws.String(b.bucket),
838 Key: aws.String(path),
840 req := b.svc.DeleteObjectRequest(input)
841 _, err := req.Send(context.Background())
842 b.stats.TickOps("delete")
843 b.stats.Tick(&b.stats.Ops, &b.stats.DelOps)
848 // Trash a Keep block.
849 func (v *S3AWSVolume) Trash(loc string) error {
850 if v.volume.ReadOnly {
851 return MethodDisabledError
853 if t, err := v.Mtime(loc); err != nil {
855 } else if time.Since(t) < v.cluster.Collections.BlobSigningTTL.Duration() {
859 if v.cluster.Collections.BlobTrashLifetime == 0 {
861 return ErrS3TrashDisabled
863 return v.translateError(v.bucket.Del(key))
865 err := v.checkRaceWindow(key)
869 err = v.safeCopy("trash/"+key, key)
873 return v.translateError(v.bucket.Del(key))
876 // Untrash moves block from trash back into store
877 func (v *S3AWSVolume) Untrash(loc string) error {
879 err := v.safeCopy(key, "trash/"+key)
883 err = v.writeObject(context.Background(), "recent/"+key, nil)
884 return v.translateError(err)
887 type s3awsbucketStats struct {
897 func (s *s3awsbucketStats) TickErr(err error) {
901 errType := fmt.Sprintf("%T", err)
902 if aerr, ok := err.(awserr.Error); ok {
903 if reqErr, ok := err.(awserr.RequestFailure); ok {
904 // A service error occurred
905 errType = errType + fmt.Sprintf(" %d %s", reqErr.StatusCode(), aerr.Code())
907 errType = errType + fmt.Sprintf(" 000 %s", aerr.Code())
910 s.statsTicker.TickErr(err, errType)