19 "git.curoverse.com/arvados.git/sdk/go/arvados"
20 log "github.com/Sirupsen/logrus"
21 "github.com/curoverse/azure-sdk-for-go/storage"
24 const azureDefaultRequestTimeout = arvados.Duration(10 * time.Minute)
28 azureStorageAccountName string
29 azureStorageAccountKeyFile string
30 azureStorageReplication int
31 azureWriteRaceInterval = 15 * time.Second
32 azureWriteRacePollTime = time.Second
35 func readKeyFromFile(file string) (string, error) {
36 buf, err := ioutil.ReadFile(file)
38 return "", errors.New("reading key from " + file + ": " + err.Error())
40 accountKey := strings.TrimSpace(string(buf))
42 return "", errors.New("empty account key in " + file)
44 return accountKey, nil
47 type azureVolumeAdder struct {
51 // String implements flag.Value
52 func (s *azureVolumeAdder) String() string {
56 func (s *azureVolumeAdder) Set(containerName string) error {
57 s.Config.Volumes = append(s.Config.Volumes, &AzureBlobVolume{
58 ContainerName: containerName,
59 StorageAccountName: azureStorageAccountName,
60 StorageAccountKeyFile: azureStorageAccountKeyFile,
61 AzureReplication: azureStorageReplication,
62 ReadOnly: deprecated.flagReadonly,
68 VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &AzureBlobVolume{} })
70 flag.Var(&azureVolumeAdder{theConfig},
71 "azure-storage-container-volume",
72 "Use the given container as a storage volume. Can be given multiple times.")
74 &azureStorageAccountName,
75 "azure-storage-account-name",
77 "Azure storage account name used for subsequent --azure-storage-container-volume arguments.")
79 &azureStorageAccountKeyFile,
80 "azure-storage-account-key-file",
82 "`File` containing the account key used for subsequent --azure-storage-container-volume arguments.")
84 &azureStorageReplication,
85 "azure-storage-replication",
87 "Replication level to report to clients when data is stored in an Azure container.")
90 "azure-max-get-bytes",
92 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))
95 // An AzureBlobVolume stores and retrieves blocks in an Azure Blob
97 type AzureBlobVolume struct {
98 StorageAccountName string
99 StorageAccountKeyFile string
100 StorageBaseURL string // "" means default, "core.windows.net"
104 RequestTimeout arvados.Duration
106 azClient storage.Client
107 bsClient *azureBlobClient
110 // Examples implements VolumeWithExamples.
111 func (*AzureBlobVolume) Examples() []Volume {
114 StorageAccountName: "example-account-name",
115 StorageAccountKeyFile: "/etc/azure_storage_account_key.txt",
116 ContainerName: "example-container-name",
118 RequestTimeout: azureDefaultRequestTimeout,
121 StorageAccountName: "cn-account-name",
122 StorageAccountKeyFile: "/etc/azure_cn_storage_account_key.txt",
123 StorageBaseURL: "core.chinacloudapi.cn",
124 ContainerName: "cn-container-name",
126 RequestTimeout: azureDefaultRequestTimeout,
131 // Type implements Volume.
132 func (v *AzureBlobVolume) Type() string {
136 // Start implements Volume.
137 func (v *AzureBlobVolume) Start() error {
138 if v.ContainerName == "" {
139 return errors.New("no container name given")
141 if v.StorageAccountName == "" || v.StorageAccountKeyFile == "" {
142 return errors.New("StorageAccountName and StorageAccountKeyFile must be given")
144 accountKey, err := readKeyFromFile(v.StorageAccountKeyFile)
148 if v.StorageBaseURL == "" {
149 v.StorageBaseURL = storage.DefaultBaseURL
151 v.azClient, err = storage.NewClient(v.StorageAccountName, accountKey, v.StorageBaseURL, storage.DefaultAPIVersion, true)
153 return fmt.Errorf("creating Azure storage client: %s", err)
156 if v.RequestTimeout == 0 {
157 v.RequestTimeout = azureDefaultRequestTimeout
159 v.azClient.HTTPClient = &http.Client{
160 Timeout: time.Duration(v.RequestTimeout),
162 bs := v.azClient.GetBlobService()
163 v.bsClient = &azureBlobClient{
167 ok, err := v.bsClient.ContainerExists(v.ContainerName)
172 return fmt.Errorf("Azure container %q does not exist", v.ContainerName)
177 // DeviceID returns a globally unique ID for the storage container.
178 func (v *AzureBlobVolume) DeviceID() string {
179 return "azure://" + v.StorageBaseURL + "/" + v.StorageAccountName + "/" + v.ContainerName
182 // Return true if expires_at metadata attribute is found on the block
183 func (v *AzureBlobVolume) checkTrashed(loc string) (bool, map[string]string, error) {
184 metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc)
186 return false, metadata, v.translateError(err)
188 if metadata["expires_at"] != "" {
189 return true, metadata, nil
191 return false, metadata, nil
194 // Get reads a Keep block that has been stored as a block blob in the
197 // If the block is younger than azureWriteRaceInterval and is
198 // unexpectedly empty, assume a PutBlob operation is in progress, and
199 // wait for it to finish writing.
200 func (v *AzureBlobVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
201 trashed, _, err := v.checkTrashed(loc)
206 return 0, os.ErrNotExist
208 var deadline time.Time
209 haveDeadline := false
210 size, err := v.get(ctx, loc, buf)
211 for err == nil && size == 0 && loc != "d41d8cd98f00b204e9800998ecf8427e" {
212 // Seeing a brand new empty block probably means we're
213 // in a race with CreateBlob, which under the hood
214 // (apparently) does "CreateEmpty" and "CommitData"
215 // with no additional transaction locking.
217 t, err := v.Mtime(loc)
219 log.Print("Got empty block (possible race) but Mtime failed: ", err)
222 deadline = t.Add(azureWriteRaceInterval)
223 if time.Now().After(deadline) {
226 log.Printf("Race? Block %s is 0 bytes, %s old. Polling until %s", loc, time.Since(t), deadline)
228 } else if time.Now().After(deadline) {
234 case <-time.After(azureWriteRacePollTime):
236 size, err = v.get(ctx, loc, buf)
239 log.Printf("Race ended with size==%d", size)
244 func (v *AzureBlobVolume) get(ctx context.Context, loc string, buf []byte) (int, error) {
245 ctx, cancel := context.WithCancel(ctx)
247 expectSize := len(buf)
248 if azureMaxGetBytes < BlockSize {
249 // Unfortunately the handler doesn't tell us how long the blob
250 // is expected to be, so we have to ask Azure.
251 props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
253 return 0, v.translateError(err)
255 if props.ContentLength > int64(BlockSize) || props.ContentLength < 0 {
256 return 0, fmt.Errorf("block %s invalid size %d (max %d)", loc, props.ContentLength, BlockSize)
258 expectSize = int(props.ContentLength)
265 // We'll update this actualSize if/when we get the last piece.
267 pieces := (expectSize + azureMaxGetBytes - 1) / azureMaxGetBytes
268 errors := make(chan error, pieces)
269 var wg sync.WaitGroup
271 for p := 0; p < pieces; p++ {
272 // Each goroutine retrieves one piece. If we hit an
273 // error, it is sent to the errors chan so get() can
274 // return it -- but only if the error happens before
275 // ctx is done. This way, if ctx is done before we hit
276 // any other error (e.g., requesting client has hung
277 // up), we return the original ctx.Err() instead of
278 // the secondary errors from the transfers that got
279 // interrupted as a result.
282 startPos := p * azureMaxGetBytes
283 endPos := startPos + azureMaxGetBytes
284 if endPos > expectSize {
287 var rdr io.ReadCloser
289 gotRdr := make(chan struct{})
292 if startPos == 0 && endPos == expectSize {
293 rdr, err = v.bsClient.GetBlob(v.ContainerName, loc)
295 rdr, err = v.bsClient.GetBlobRange(v.ContainerName, loc, fmt.Sprintf("%d-%d", startPos, endPos-1), nil)
315 // Close the reader when the client
316 // hangs up or another piece fails
317 // (possibly interrupting ReadFull())
318 // or when all pieces succeed and
323 n, err := io.ReadFull(rdr, buf[startPos:endPos])
324 if pieces == 1 && (err == io.ErrUnexpectedEOF || err == io.EOF) {
325 // If we don't know the actual size,
326 // and just tried reading 64 MiB, it's
327 // normal to encounter EOF.
328 } else if err != nil {
329 if ctx.Err() == nil {
336 actualSize = startPos + n
343 return 0, v.translateError(<-errors)
345 if ctx.Err() != nil {
348 return actualSize, nil
351 // Compare the given data with existing stored data.
352 func (v *AzureBlobVolume) Compare(ctx context.Context, loc string, expect []byte) error {
353 trashed, _, err := v.checkTrashed(loc)
358 return os.ErrNotExist
360 var rdr io.ReadCloser
361 gotRdr := make(chan struct{})
364 rdr, err = v.bsClient.GetBlob(v.ContainerName, loc)
378 return v.translateError(err)
381 return compareReaderWithBuf(ctx, rdr, expect, loc[:32])
384 // Put stores a Keep block as a block blob in the container.
385 func (v *AzureBlobVolume) Put(ctx context.Context, loc string, block []byte) error {
387 return MethodDisabledError
389 // Send the block data through a pipe, so that (if we need to)
390 // we can close the pipe early and abandon our
391 // CreateBlockBlobFromReader() goroutine, without worrying
392 // about CreateBlockBlobFromReader() accessing our block
393 // buffer after we release it.
394 bufr, bufw := io.Pipe()
396 io.Copy(bufw, bytes.NewReader(block))
399 errChan := make(chan error)
401 errChan <- v.bsClient.CreateBlockBlobFromReader(v.ContainerName, loc, uint64(len(block)), bufr, nil)
405 theConfig.debugLogf("%s: taking CreateBlockBlobFromReader's input away: %s", v, ctx.Err())
406 // Our pipe might be stuck in Write(), waiting for
407 // io.Copy() to read. If so, un-stick it. This means
408 // CreateBlockBlobFromReader will get corrupt data,
409 // but that's OK: the size won't match, so the write
411 go io.Copy(ioutil.Discard, bufr)
412 // CloseWithError() will return once pending I/O is done.
413 bufw.CloseWithError(ctx.Err())
414 theConfig.debugLogf("%s: abandoning CreateBlockBlobFromReader goroutine", v)
416 case err := <-errChan:
421 // Touch updates the last-modified property of a block blob.
422 func (v *AzureBlobVolume) Touch(loc string) error {
424 return MethodDisabledError
426 trashed, metadata, err := v.checkTrashed(loc)
431 return os.ErrNotExist
434 metadata["touch"] = fmt.Sprintf("%d", time.Now())
435 return v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil)
438 // Mtime returns the last-modified property of a block blob.
439 func (v *AzureBlobVolume) Mtime(loc string) (time.Time, error) {
440 trashed, _, err := v.checkTrashed(loc)
442 return time.Time{}, err
445 return time.Time{}, os.ErrNotExist
448 props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
450 return time.Time{}, err
452 return time.Parse(time.RFC1123, props.LastModified)
455 // IndexTo writes a list of Keep blocks that are stored in the
457 func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error {
458 params := storage.ListBlobsParameters{
463 resp, err := v.bsClient.ListBlobs(v.ContainerName, params)
467 for _, b := range resp.Blobs {
468 t, err := time.Parse(time.RFC1123, b.Properties.LastModified)
472 if !v.isKeepBlock(b.Name) {
475 if b.Properties.ContentLength == 0 && t.Add(azureWriteRaceInterval).After(time.Now()) {
476 // A new zero-length blob is probably
477 // just a new non-empty blob that
478 // hasn't committed its data yet (see
479 // Get()), and in any case has no
483 if b.Metadata["expires_at"] != "" {
484 // Trashed blob; exclude it from response
487 fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, t.UnixNano())
489 if resp.NextMarker == "" {
492 params.Marker = resp.NextMarker
496 // Trash a Keep block.
497 func (v *AzureBlobVolume) Trash(loc string) error {
499 return MethodDisabledError
502 // Ideally we would use If-Unmodified-Since, but that
503 // particular condition seems to be ignored by Azure. Instead,
504 // we get the Etag before checking Mtime, and use If-Match to
505 // ensure we don't delete data if Put() or Touch() happens
506 // between our calls to Mtime() and DeleteBlob().
507 props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
511 if t, err := v.Mtime(loc); err != nil {
513 } else if time.Since(t) < theConfig.BlobSignatureTTL.Duration() {
517 // If TrashLifetime == 0, just delete it
518 if theConfig.TrashLifetime == 0 {
519 return v.bsClient.DeleteBlob(v.ContainerName, loc, map[string]string{
520 "If-Match": props.Etag,
524 // Otherwise, mark as trash
525 return v.bsClient.SetBlobMetadata(v.ContainerName, loc, map[string]string{
526 "expires_at": fmt.Sprintf("%d", time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()),
527 }, map[string]string{
528 "If-Match": props.Etag,
532 // Untrash a Keep block.
533 // Delete the expires_at metadata attribute
534 func (v *AzureBlobVolume) Untrash(loc string) error {
535 // if expires_at does not exist, return NotFoundError
536 metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc)
538 return v.translateError(err)
540 if metadata["expires_at"] == "" {
541 return os.ErrNotExist
544 // reset expires_at metadata attribute
545 metadata["expires_at"] = ""
546 err = v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil)
547 return v.translateError(err)
550 // Status returns a VolumeStatus struct with placeholder data.
551 func (v *AzureBlobVolume) Status() *VolumeStatus {
552 return &VolumeStatus{
554 BytesFree: BlockSize * 1000,
559 // String returns a volume label, including the container name.
560 func (v *AzureBlobVolume) String() string {
561 return fmt.Sprintf("azure-storage-container:%+q", v.ContainerName)
564 // Writable returns true, unless the -readonly flag was on when the
566 func (v *AzureBlobVolume) Writable() bool {
570 // Replication returns the replication level of the container, as
571 // specified by the -azure-storage-replication argument.
572 func (v *AzureBlobVolume) Replication() int {
573 return v.AzureReplication
576 // If possible, translate an Azure SDK error to a recognizable error
577 // like os.ErrNotExist.
578 func (v *AzureBlobVolume) translateError(err error) error {
582 case strings.Contains(err.Error(), "Not Found"):
583 // "storage: service returned without a response body (404 Not Found)"
584 return os.ErrNotExist
590 var keepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
592 func (v *AzureBlobVolume) isKeepBlock(s string) bool {
593 return keepBlockRegexp.MatchString(s)
596 // EmptyTrash looks for trashed blocks that exceeded TrashLifetime
597 // and deletes them from the volume.
598 func (v *AzureBlobVolume) EmptyTrash() {
599 var bytesDeleted, bytesInTrash int64
600 var blocksDeleted, blocksInTrash int
601 params := storage.ListBlobsParameters{Include: "metadata"}
604 resp, err := v.bsClient.ListBlobs(v.ContainerName, params)
606 log.Printf("EmptyTrash: ListBlobs: %v", err)
609 for _, b := range resp.Blobs {
610 // Check if the block is expired
611 if b.Metadata["expires_at"] == "" {
616 bytesInTrash += b.Properties.ContentLength
618 expiresAt, err := strconv.ParseInt(b.Metadata["expires_at"], 10, 64)
620 log.Printf("EmptyTrash: ParseInt(%v): %v", b.Metadata["expires_at"], err)
624 if expiresAt > time.Now().Unix() {
628 err = v.bsClient.DeleteBlob(v.ContainerName, b.Name, map[string]string{
629 "If-Match": b.Properties.Etag,
632 log.Printf("EmptyTrash: DeleteBlob(%v): %v", b.Name, err)
636 bytesDeleted += b.Properties.ContentLength
638 if resp.NextMarker == "" {
641 params.Marker = resp.NextMarker
644 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)
647 // InternalStats returns bucket I/O and API call counters.
648 func (v *AzureBlobVolume) InternalStats() interface{} {
649 return &v.bsClient.stats
652 type azureBlobStats struct {
657 GetMetadataOps uint64
658 GetPropertiesOps uint64
660 SetMetadataOps uint64
665 func (s *azureBlobStats) TickErr(err error) {
669 errType := fmt.Sprintf("%T", err)
670 if err, ok := err.(storage.AzureStorageServiceError); ok {
671 errType = errType + fmt.Sprintf(" %d (%s)", err.StatusCode, err.Code)
673 log.Printf("errType %T, err %s", err, err)
674 s.statsTicker.TickErr(err, errType)
677 // azureBlobClient wraps storage.BlobStorageClient in order to count
678 // I/O and API usage stats.
679 type azureBlobClient struct {
680 client *storage.BlobStorageClient
684 func (c *azureBlobClient) ContainerExists(cname string) (bool, error) {
685 c.stats.Tick(&c.stats.Ops)
686 ok, err := c.client.ContainerExists(cname)
691 func (c *azureBlobClient) GetBlobMetadata(cname, bname string) (map[string]string, error) {
692 c.stats.Tick(&c.stats.Ops, &c.stats.GetMetadataOps)
693 m, err := c.client.GetBlobMetadata(cname, bname)
698 func (c *azureBlobClient) GetBlobProperties(cname, bname string) (*storage.BlobProperties, error) {
699 c.stats.Tick(&c.stats.Ops, &c.stats.GetPropertiesOps)
700 p, err := c.client.GetBlobProperties(cname, bname)
705 func (c *azureBlobClient) GetBlob(cname, bname string) (io.ReadCloser, error) {
706 c.stats.Tick(&c.stats.Ops, &c.stats.GetOps)
707 rdr, err := c.client.GetBlob(cname, bname)
709 return NewCountingReader(rdr, c.stats.TickInBytes), err
712 func (c *azureBlobClient) GetBlobRange(cname, bname, byterange string, hdrs map[string]string) (io.ReadCloser, error) {
713 c.stats.Tick(&c.stats.Ops, &c.stats.GetRangeOps)
714 rdr, err := c.client.GetBlobRange(cname, bname, byterange, hdrs)
716 return NewCountingReader(rdr, c.stats.TickInBytes), err
719 func (c *azureBlobClient) CreateBlockBlobFromReader(cname, bname string, size uint64, rdr io.Reader, hdrs map[string]string) error {
720 c.stats.Tick(&c.stats.Ops, &c.stats.CreateOps)
721 rdr = NewCountingReader(rdr, c.stats.TickOutBytes)
722 err := c.client.CreateBlockBlobFromReader(cname, bname, size, rdr, hdrs)
727 func (c *azureBlobClient) SetBlobMetadata(cname, bname string, m, hdrs map[string]string) error {
728 c.stats.Tick(&c.stats.Ops, &c.stats.SetMetadataOps)
729 err := c.client.SetBlobMetadata(cname, bname, m, hdrs)
734 func (c *azureBlobClient) ListBlobs(cname string, params storage.ListBlobsParameters) (storage.BlobListResponse, error) {
735 c.stats.Tick(&c.stats.Ops, &c.stats.ListOps)
736 resp, err := c.client.ListBlobs(cname, params)
741 func (c *azureBlobClient) DeleteBlob(cname, bname string, hdrs map[string]string) error {
742 c.stats.Tick(&c.stats.Ops, &c.stats.DelOps)
743 err := c.client.DeleteBlob(cname, bname, hdrs)