1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
24 "git.curoverse.com/arvados.git/sdk/go/arvados"
25 "github.com/Azure/azure-sdk-for-go/storage"
26 "github.com/prometheus/client_golang/prometheus"
29 const azureDefaultRequestTimeout = arvados.Duration(10 * time.Minute)
33 azureStorageAccountName string
34 azureStorageAccountKeyFile string
35 azureStorageReplication int
36 azureWriteRaceInterval = 15 * time.Second
37 azureWriteRacePollTime = time.Second
40 func readKeyFromFile(file string) (string, error) {
41 buf, err := ioutil.ReadFile(file)
43 return "", errors.New("reading key from " + file + ": " + err.Error())
45 accountKey := strings.TrimSpace(string(buf))
47 return "", errors.New("empty account key in " + file)
49 return accountKey, nil
52 type azureVolumeAdder struct {
56 // String implements flag.Value
57 func (s *azureVolumeAdder) String() string {
61 func (s *azureVolumeAdder) Set(containerName string) error {
62 s.Config.Volumes = append(s.Config.Volumes, &AzureBlobVolume{
63 ContainerName: containerName,
64 StorageAccountName: azureStorageAccountName,
65 StorageAccountKeyFile: azureStorageAccountKeyFile,
66 AzureReplication: azureStorageReplication,
67 ReadOnly: deprecated.flagReadonly,
73 VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &AzureBlobVolume{} })
75 flag.Var(&azureVolumeAdder{theConfig},
76 "azure-storage-container-volume",
77 "Use the given container as a storage volume. Can be given multiple times.")
79 &azureStorageAccountName,
80 "azure-storage-account-name",
82 "Azure storage account name used for subsequent --azure-storage-container-volume arguments.")
84 &azureStorageAccountKeyFile,
85 "azure-storage-account-key-file",
87 "`File` containing the account key used for subsequent --azure-storage-container-volume arguments.")
89 &azureStorageReplication,
90 "azure-storage-replication",
92 "Replication level to report to clients when data is stored in an Azure container.")
95 "azure-max-get-bytes",
97 fmt.Sprintf("Maximum bytes to request in a single GET request. If smaller than %d, use multiple concurrent range requests to retrieve a block.", BlockSize))
100 // An AzureBlobVolume stores and retrieves blocks in an Azure Blob
102 type AzureBlobVolume struct {
103 StorageAccountName string
104 StorageAccountKeyFile string
105 StorageBaseURL string // "" means default, "core.windows.net"
109 RequestTimeout arvados.Duration
110 StorageClasses []string
112 azClient storage.Client
113 container *azureContainer
116 // singleSender is a single-attempt storage.Sender.
117 type singleSender struct{}
119 // Send performs req exactly once.
120 func (*singleSender) Send(c *storage.Client, req *http.Request) (resp *http.Response, err error) {
121 return c.HTTPClient.Do(req)
124 // Examples implements VolumeWithExamples.
125 func (*AzureBlobVolume) Examples() []Volume {
128 StorageAccountName: "example-account-name",
129 StorageAccountKeyFile: "/etc/azure_storage_account_key.txt",
130 ContainerName: "example-container-name",
132 RequestTimeout: azureDefaultRequestTimeout,
135 StorageAccountName: "cn-account-name",
136 StorageAccountKeyFile: "/etc/azure_cn_storage_account_key.txt",
137 StorageBaseURL: "core.chinacloudapi.cn",
138 ContainerName: "cn-container-name",
140 RequestTimeout: azureDefaultRequestTimeout,
145 // Type implements Volume.
146 func (v *AzureBlobVolume) Type() string {
150 // Start implements Volume.
151 func (v *AzureBlobVolume) Start(vm *volumeMetricsVecs) error {
152 if v.ContainerName == "" {
153 return errors.New("no container name given")
155 if v.StorageAccountName == "" || v.StorageAccountKeyFile == "" {
156 return errors.New("StorageAccountName and StorageAccountKeyFile must be given")
158 accountKey, err := readKeyFromFile(v.StorageAccountKeyFile)
162 if v.StorageBaseURL == "" {
163 v.StorageBaseURL = storage.DefaultBaseURL
165 v.azClient, err = storage.NewClient(v.StorageAccountName, accountKey, v.StorageBaseURL, storage.DefaultAPIVersion, true)
167 return fmt.Errorf("creating Azure storage client: %s", err)
169 v.azClient.Sender = &singleSender{}
171 if v.RequestTimeout == 0 {
172 v.RequestTimeout = azureDefaultRequestTimeout
174 v.azClient.HTTPClient = &http.Client{
175 Timeout: time.Duration(v.RequestTimeout),
177 bs := v.azClient.GetBlobService()
178 v.container = &azureContainer{
179 ctr: bs.GetContainerReference(v.ContainerName),
182 if ok, err := v.container.Exists(); err != nil {
185 return fmt.Errorf("Azure container %q does not exist", v.ContainerName)
187 // Set up prometheus metrics
188 lbls := prometheus.Labels{"device_id": v.DeviceID()}
189 v.container.stats.opsCounters, v.container.stats.errCounters, v.container.stats.ioBytes = vm.getCounterVecsFor(lbls)
194 // DeviceID returns a globally unique ID for the storage container.
195 func (v *AzureBlobVolume) DeviceID() string {
196 return "azure://" + v.StorageBaseURL + "/" + v.StorageAccountName + "/" + v.ContainerName
199 // Return true if expires_at metadata attribute is found on the block
200 func (v *AzureBlobVolume) checkTrashed(loc string) (bool, map[string]string, error) {
201 metadata, err := v.container.GetBlobMetadata(loc)
203 return false, metadata, v.translateError(err)
205 if metadata["expires_at"] != "" {
206 return true, metadata, nil
208 return false, metadata, nil
211 // Get reads a Keep block that has been stored as a block blob in the
214 // If the block is younger than azureWriteRaceInterval and is
215 // unexpectedly empty, assume a PutBlob operation is in progress, and
216 // wait for it to finish writing.
217 func (v *AzureBlobVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
218 trashed, _, err := v.checkTrashed(loc)
223 return 0, os.ErrNotExist
225 var deadline time.Time
226 haveDeadline := false
227 size, err := v.get(ctx, loc, buf)
228 for err == nil && size == 0 && loc != "d41d8cd98f00b204e9800998ecf8427e" {
229 // Seeing a brand new empty block probably means we're
230 // in a race with CreateBlob, which under the hood
231 // (apparently) does "CreateEmpty" and "CommitData"
232 // with no additional transaction locking.
234 t, err := v.Mtime(loc)
236 log.Print("Got empty block (possible race) but Mtime failed: ", err)
239 deadline = t.Add(azureWriteRaceInterval)
240 if time.Now().After(deadline) {
243 log.Printf("Race? Block %s is 0 bytes, %s old. Polling until %s", loc, time.Since(t), deadline)
245 } else if time.Now().After(deadline) {
251 case <-time.After(azureWriteRacePollTime):
253 size, err = v.get(ctx, loc, buf)
256 log.Printf("Race ended with size==%d", size)
261 func (v *AzureBlobVolume) get(ctx context.Context, loc string, buf []byte) (int, error) {
262 ctx, cancel := context.WithCancel(ctx)
264 expectSize := len(buf)
265 if azureMaxGetBytes < BlockSize {
266 // Unfortunately the handler doesn't tell us how long the blob
267 // is expected to be, so we have to ask Azure.
268 props, err := v.container.GetBlobProperties(loc)
270 return 0, v.translateError(err)
272 if props.ContentLength > int64(BlockSize) || props.ContentLength < 0 {
273 return 0, fmt.Errorf("block %s invalid size %d (max %d)", loc, props.ContentLength, BlockSize)
275 expectSize = int(props.ContentLength)
282 // We'll update this actualSize if/when we get the last piece.
284 pieces := (expectSize + azureMaxGetBytes - 1) / azureMaxGetBytes
285 errors := make(chan error, pieces)
286 var wg sync.WaitGroup
288 for p := 0; p < pieces; p++ {
289 // Each goroutine retrieves one piece. If we hit an
290 // error, it is sent to the errors chan so get() can
291 // return it -- but only if the error happens before
292 // ctx is done. This way, if ctx is done before we hit
293 // any other error (e.g., requesting client has hung
294 // up), we return the original ctx.Err() instead of
295 // the secondary errors from the transfers that got
296 // interrupted as a result.
299 startPos := p * azureMaxGetBytes
300 endPos := startPos + azureMaxGetBytes
301 if endPos > expectSize {
304 var rdr io.ReadCloser
306 gotRdr := make(chan struct{})
309 if startPos == 0 && endPos == expectSize {
310 rdr, err = v.container.GetBlob(loc)
312 rdr, err = v.container.GetBlobRange(loc, startPos, endPos-1, nil)
332 // Close the reader when the client
333 // hangs up or another piece fails
334 // (possibly interrupting ReadFull())
335 // or when all pieces succeed and
340 n, err := io.ReadFull(rdr, buf[startPos:endPos])
341 if pieces == 1 && (err == io.ErrUnexpectedEOF || err == io.EOF) {
342 // If we don't know the actual size,
343 // and just tried reading 64 MiB, it's
344 // normal to encounter EOF.
345 } else if err != nil {
346 if ctx.Err() == nil {
353 actualSize = startPos + n
360 return 0, v.translateError(<-errors)
362 if ctx.Err() != nil {
365 return actualSize, nil
368 // Compare the given data with existing stored data.
369 func (v *AzureBlobVolume) Compare(ctx context.Context, loc string, expect []byte) error {
370 trashed, _, err := v.checkTrashed(loc)
375 return os.ErrNotExist
377 var rdr io.ReadCloser
378 gotRdr := make(chan struct{})
381 rdr, err = v.container.GetBlob(loc)
395 return v.translateError(err)
398 return compareReaderWithBuf(ctx, rdr, expect, loc[:32])
401 // Put stores a Keep block as a block blob in the container.
402 func (v *AzureBlobVolume) Put(ctx context.Context, loc string, block []byte) error {
404 return MethodDisabledError
406 // Send the block data through a pipe, so that (if we need to)
407 // we can close the pipe early and abandon our
408 // CreateBlockBlobFromReader() goroutine, without worrying
409 // about CreateBlockBlobFromReader() accessing our block
410 // buffer after we release it.
411 bufr, bufw := io.Pipe()
413 io.Copy(bufw, bytes.NewReader(block))
416 errChan := make(chan error)
418 var body io.Reader = bufr
420 // We must send a "Content-Length: 0" header,
421 // but the http client interprets
422 // ContentLength==0 as "unknown" unless it can
423 // confirm by introspection that Body will
428 errChan <- v.container.CreateBlockBlobFromReader(loc, len(block), body, nil)
432 theConfig.debugLogf("%s: taking CreateBlockBlobFromReader's input away: %s", v, ctx.Err())
433 // Our pipe might be stuck in Write(), waiting for
434 // io.Copy() to read. If so, un-stick it. This means
435 // CreateBlockBlobFromReader will get corrupt data,
436 // but that's OK: the size won't match, so the write
438 go io.Copy(ioutil.Discard, bufr)
439 // CloseWithError() will return once pending I/O is done.
440 bufw.CloseWithError(ctx.Err())
441 theConfig.debugLogf("%s: abandoning CreateBlockBlobFromReader goroutine", v)
443 case err := <-errChan:
448 // Touch updates the last-modified property of a block blob.
449 func (v *AzureBlobVolume) Touch(loc string) error {
451 return MethodDisabledError
453 trashed, metadata, err := v.checkTrashed(loc)
458 return os.ErrNotExist
461 metadata["touch"] = fmt.Sprintf("%d", time.Now().Unix())
462 return v.container.SetBlobMetadata(loc, metadata, nil)
465 // Mtime returns the last-modified property of a block blob.
466 func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) {
467 trashed, _, err := v.checkTrashed(loc)
469 return time.Time{}, err
472 return time.Time{}, os.ErrNotExist
475 props, err := v.container.GetBlobProperties(loc)
477 return time.Time{}, err
479 return time.Time(props.LastModified), nil
482 // IndexTo writes a list of Keep blocks that are stored in the
484 func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error {
485 params := storage.ListBlobsParameters{
487 Include: &storage.IncludeBlobDataset{Metadata: true},
490 resp, err := v.container.ListBlobs(params)
494 for _, b := range resp.Blobs {
495 if !v.isKeepBlock(b.Name) {
498 modtime := time.Time(b.Properties.LastModified)
499 if b.Properties.ContentLength == 0 && modtime.Add(azureWriteRaceInterval).After(time.Now()) {
500 // A new zero-length blob is probably
501 // just a new non-empty blob that
502 // hasn't committed its data yet (see
503 // Get()), and in any case has no
507 if b.Metadata["expires_at"] != "" {
508 // Trashed blob; exclude it from response
511 fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, modtime.UnixNano())
513 if resp.NextMarker == "" {
516 params.Marker = resp.NextMarker
520 // Trash a Keep block.
521 func (v *AzureBlobVolume) Trash(loc string) error {
523 return MethodDisabledError
526 // Ideally we would use If-Unmodified-Since, but that
527 // particular condition seems to be ignored by Azure. Instead,
528 // we get the Etag before checking Mtime, and use If-Match to
529 // ensure we don't delete data if Put() or Touch() happens
530 // between our calls to Mtime() and DeleteBlob().
531 props, err := v.container.GetBlobProperties(loc)
535 if t, err := v.Mtime(loc); err != nil {
537 } else if time.Since(t) < theConfig.BlobSignatureTTL.Duration() {
541 // If TrashLifetime == 0, just delete it
542 if theConfig.TrashLifetime == 0 {
543 return v.container.DeleteBlob(loc, &storage.DeleteBlobOptions{
548 // Otherwise, mark as trash
549 return v.container.SetBlobMetadata(loc, storage.BlobMetadata{
550 "expires_at": fmt.Sprintf("%d", time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()),
551 }, &storage.SetBlobMetadataOptions{
556 // Untrash a Keep block.
557 // Delete the expires_at metadata attribute
558 func (v *AzureBlobVolume) Untrash(loc string) error {
559 // if expires_at does not exist, return NotFoundError
560 metadata, err := v.container.GetBlobMetadata(loc)
562 return v.translateError(err)
564 if metadata["expires_at"] == "" {
565 return os.ErrNotExist
568 // reset expires_at metadata attribute
569 metadata["expires_at"] = ""
570 err = v.container.SetBlobMetadata(loc, metadata, nil)
571 return v.translateError(err)
574 // Status returns a VolumeStatus struct with placeholder data.
575 func (v *AzureBlobVolume) Status() *VolumeStatus {
576 return &VolumeStatus{
578 BytesFree: BlockSize * 1000,
583 // String returns a volume label, including the container name.
584 func (v *AzureBlobVolume) String() string {
585 return fmt.Sprintf("azure-storage-container:%+q", v.ContainerName)
588 // Writable returns true, unless the -readonly flag was on when the
590 func (v *AzureBlobVolume) Writable() bool {
594 // Replication returns the replication level of the container, as
595 // specified by the -azure-storage-replication argument.
596 func (v *AzureBlobVolume) Replication() int {
597 return v.AzureReplication
600 // GetStorageClasses implements Volume
601 func (v *AzureBlobVolume) GetStorageClasses() []string {
602 return v.StorageClasses
605 // If possible, translate an Azure SDK error to a recognizable error
606 // like os.ErrNotExist.
607 func (v *AzureBlobVolume) translateError(err error) error {
611 case strings.Contains(err.Error(), "StatusCode=503"):
612 // "storage: service returned error: StatusCode=503, ErrorCode=ServerBusy, ErrorMessage=The server is busy" (See #14804)
613 return VolumeBusyError
614 case strings.Contains(err.Error(), "Not Found"):
615 // "storage: service returned without a response body (404 Not Found)"
616 return os.ErrNotExist
622 var keepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
624 func (v *AzureBlobVolume) isKeepBlock(s string) bool {
625 return keepBlockRegexp.MatchString(s)
628 // EmptyTrash looks for trashed blocks that exceeded TrashLifetime
629 // and deletes them from the volume.
630 func (v *AzureBlobVolume) EmptyTrash() {
631 var bytesDeleted, bytesInTrash int64
632 var blocksDeleted, blocksInTrash int64
634 doBlob := func(b storage.Blob) {
635 // Check whether the block is flagged as trash
636 if b.Metadata["expires_at"] == "" {
640 atomic.AddInt64(&blocksInTrash, 1)
641 atomic.AddInt64(&bytesInTrash, b.Properties.ContentLength)
643 expiresAt, err := strconv.ParseInt(b.Metadata["expires_at"], 10, 64)
645 log.Printf("EmptyTrash: ParseInt(%v): %v", b.Metadata["expires_at"], err)
649 if expiresAt > time.Now().Unix() {
653 err = v.container.DeleteBlob(b.Name, &storage.DeleteBlobOptions{
654 IfMatch: b.Properties.Etag,
657 log.Printf("EmptyTrash: DeleteBlob(%v): %v", b.Name, err)
660 atomic.AddInt64(&blocksDeleted, 1)
661 atomic.AddInt64(&bytesDeleted, b.Properties.ContentLength)
664 var wg sync.WaitGroup
665 todo := make(chan storage.Blob, theConfig.EmptyTrashWorkers)
666 for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
670 for b := range todo {
676 params := storage.ListBlobsParameters{Include: &storage.IncludeBlobDataset{Metadata: true}}
678 resp, err := v.container.ListBlobs(params)
680 log.Printf("EmptyTrash: ListBlobs: %v", err)
683 for _, b := range resp.Blobs {
686 if resp.NextMarker == "" {
689 params.Marker = resp.NextMarker
694 log.Printf("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)
697 // InternalStats returns bucket I/O and API call counters.
698 func (v *AzureBlobVolume) InternalStats() interface{} {
699 return &v.container.stats
702 type azureBlobStats struct {
707 GetMetadataOps uint64
708 GetPropertiesOps uint64
710 SetMetadataOps uint64
715 func (s *azureBlobStats) TickErr(err error) {
719 errType := fmt.Sprintf("%T", err)
720 if err, ok := err.(storage.AzureStorageServiceError); ok {
721 errType = errType + fmt.Sprintf(" %d (%s)", err.StatusCode, err.Code)
723 log.Printf("errType %T, err %s", err, err)
724 s.statsTicker.TickErr(err, errType)
727 // azureContainer wraps storage.Container in order to count I/O and
729 type azureContainer struct {
730 ctr *storage.Container
734 func (c *azureContainer) Exists() (bool, error) {
735 c.stats.TickOps("exists")
736 c.stats.Tick(&c.stats.Ops)
737 ok, err := c.ctr.Exists()
742 func (c *azureContainer) GetBlobMetadata(bname string) (storage.BlobMetadata, error) {
743 c.stats.TickOps("get_metadata")
744 c.stats.Tick(&c.stats.Ops, &c.stats.GetMetadataOps)
745 b := c.ctr.GetBlobReference(bname)
746 err := b.GetMetadata(nil)
748 return b.Metadata, err
751 func (c *azureContainer) GetBlobProperties(bname string) (*storage.BlobProperties, error) {
752 c.stats.TickOps("get_properties")
753 c.stats.Tick(&c.stats.Ops, &c.stats.GetPropertiesOps)
754 b := c.ctr.GetBlobReference(bname)
755 err := b.GetProperties(nil)
757 return &b.Properties, err
760 func (c *azureContainer) GetBlob(bname string) (io.ReadCloser, error) {
761 c.stats.TickOps("get")
762 c.stats.Tick(&c.stats.Ops, &c.stats.GetOps)
763 b := c.ctr.GetBlobReference(bname)
764 rdr, err := b.Get(nil)
766 return NewCountingReader(rdr, c.stats.TickInBytes), err
769 func (c *azureContainer) GetBlobRange(bname string, start, end int, opts *storage.GetBlobOptions) (io.ReadCloser, error) {
770 c.stats.TickOps("get_range")
771 c.stats.Tick(&c.stats.Ops, &c.stats.GetRangeOps)
772 b := c.ctr.GetBlobReference(bname)
773 rdr, err := b.GetRange(&storage.GetBlobRangeOptions{
774 Range: &storage.BlobRange{
775 Start: uint64(start),
778 GetBlobOptions: opts,
781 return NewCountingReader(rdr, c.stats.TickInBytes), err
784 // If we give it an io.Reader that doesn't also have a Len() int
785 // method, the Azure SDK determines data size by copying the data into
786 // a new buffer, which is not a good use of memory.
787 type readerWithAzureLen struct {
792 // Len satisfies the private lener interface in azure-sdk-for-go.
793 func (r *readerWithAzureLen) Len() int {
797 func (c *azureContainer) CreateBlockBlobFromReader(bname string, size int, rdr io.Reader, opts *storage.PutBlobOptions) error {
798 c.stats.TickOps("create")
799 c.stats.Tick(&c.stats.Ops, &c.stats.CreateOps)
801 rdr = &readerWithAzureLen{
802 Reader: NewCountingReader(rdr, c.stats.TickOutBytes),
806 b := c.ctr.GetBlobReference(bname)
807 err := b.CreateBlockBlobFromReader(rdr, opts)
812 func (c *azureContainer) SetBlobMetadata(bname string, m storage.BlobMetadata, opts *storage.SetBlobMetadataOptions) error {
813 c.stats.TickOps("set_metadata")
814 c.stats.Tick(&c.stats.Ops, &c.stats.SetMetadataOps)
815 b := c.ctr.GetBlobReference(bname)
817 err := b.SetMetadata(opts)
822 func (c *azureContainer) ListBlobs(params storage.ListBlobsParameters) (storage.BlobListResponse, error) {
823 c.stats.TickOps("list")
824 c.stats.Tick(&c.stats.Ops, &c.stats.ListOps)
825 resp, err := c.ctr.ListBlobs(params)
830 func (c *azureContainer) DeleteBlob(bname string, opts *storage.DeleteBlobOptions) error {
831 c.stats.TickOps("delete")
832 c.stats.Tick(&c.stats.Ops, &c.stats.DelOps)
833 b := c.ctr.GetBlobReference(bname)
834 err := b.Delete(opts)