+// Copyright (C) The Arvados Authors. All rights reserved.
+//
+// SPDX-License-Identifier: AGPL-3.0
+
package main
import (
"fmt"
"io"
"io/ioutil"
- "log"
"net/http"
"os"
"regexp"
"strconv"
"strings"
"sync"
+ "sync/atomic"
"time"
"git.curoverse.com/arvados.git/sdk/go/arvados"
- "github.com/curoverse/azure-sdk-for-go/storage"
+ "github.com/Azure/azure-sdk-for-go/storage"
)
const azureDefaultRequestTimeout = arvados.Duration(10 * time.Minute)
type AzureBlobVolume struct {
StorageAccountName string
StorageAccountKeyFile string
+ StorageBaseURL string // "" means default, "core.windows.net"
ContainerName string
AzureReplication int
ReadOnly bool
RequestTimeout arvados.Duration
+ StorageClasses []string
- azClient storage.Client
- bsClient storage.BlobStorageClient
+ azClient storage.Client
+ container *azureContainer
+}
+
+// singleSender is a single-attempt storage.Sender.
+type singleSender struct{}
+
+// Send performs req exactly once.
+func (*singleSender) Send(c *storage.Client, req *http.Request) (resp *http.Response, err error) {
+ return c.HTTPClient.Do(req)
}
// Examples implements VolumeWithExamples.
AzureReplication: 3,
RequestTimeout: azureDefaultRequestTimeout,
},
+ &AzureBlobVolume{
+ StorageAccountName: "cn-account-name",
+ StorageAccountKeyFile: "/etc/azure_cn_storage_account_key.txt",
+ StorageBaseURL: "core.chinacloudapi.cn",
+ ContainerName: "cn-container-name",
+ AzureReplication: 3,
+ RequestTimeout: azureDefaultRequestTimeout,
+ },
}
}
if err != nil {
return err
}
- v.azClient, err = storage.NewBasicClient(v.StorageAccountName, accountKey)
+ if v.StorageBaseURL == "" {
+ v.StorageBaseURL = storage.DefaultBaseURL
+ }
+ v.azClient, err = storage.NewClient(v.StorageAccountName, accountKey, v.StorageBaseURL, storage.DefaultAPIVersion, true)
if err != nil {
return fmt.Errorf("creating Azure storage client: %s", err)
}
+ v.azClient.Sender = &singleSender{}
if v.RequestTimeout == 0 {
v.RequestTimeout = azureDefaultRequestTimeout
v.azClient.HTTPClient = &http.Client{
Timeout: time.Duration(v.RequestTimeout),
}
- v.bsClient = v.azClient.GetBlobService()
+ bs := v.azClient.GetBlobService()
+ v.container = &azureContainer{
+ ctr: bs.GetContainerReference(v.ContainerName),
+ }
- ok, err := v.bsClient.ContainerExists(v.ContainerName)
- if err != nil {
+ if ok, err := v.container.Exists(); err != nil {
return err
- }
- if !ok {
+ } else if !ok {
return fmt.Errorf("Azure container %q does not exist", v.ContainerName)
}
return nil
}
+// DeviceID returns a globally unique ID for the storage container.
+func (v *AzureBlobVolume) DeviceID() string {
+ return "azure://" + v.StorageBaseURL + "/" + v.StorageAccountName + "/" + v.ContainerName
+}
+
// Return true if expires_at metadata attribute is found on the block
func (v *AzureBlobVolume) checkTrashed(loc string) (bool, map[string]string, error) {
- metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc)
+ metadata, err := v.container.GetBlobMetadata(loc)
if err != nil {
return false, metadata, v.translateError(err)
}
}
var deadline time.Time
haveDeadline := false
- size, err := v.get(loc, buf)
+ size, err := v.get(ctx, loc, buf)
for err == nil && size == 0 && loc != "d41d8cd98f00b204e9800998ecf8427e" {
// Seeing a brand new empty block probably means we're
// in a race with CreateBlob, which under the hood
} else if time.Now().After(deadline) {
break
}
- time.Sleep(azureWriteRacePollTime)
- size, err = v.get(loc, buf)
+ select {
+ case <-ctx.Done():
+ return 0, ctx.Err()
+ case <-time.After(azureWriteRacePollTime):
+ }
+ size, err = v.get(ctx, loc, buf)
}
if haveDeadline {
log.Printf("Race ended with size==%d", size)
return size, err
}
-func (v *AzureBlobVolume) get(loc string, buf []byte) (int, error) {
+func (v *AzureBlobVolume) get(ctx context.Context, loc string, buf []byte) (int, error) {
+ ctx, cancel := context.WithCancel(ctx)
+ defer cancel()
expectSize := len(buf)
if azureMaxGetBytes < BlockSize {
// Unfortunately the handler doesn't tell us how long the blob
// is expected to be, so we have to ask Azure.
- props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
+ props, err := v.container.GetBlobProperties(loc)
if err != nil {
return 0, v.translateError(err)
}
// We'll update this actualSize if/when we get the last piece.
actualSize := -1
pieces := (expectSize + azureMaxGetBytes - 1) / azureMaxGetBytes
- errors := make([]error, pieces)
+ errors := make(chan error, pieces)
var wg sync.WaitGroup
wg.Add(pieces)
for p := 0; p < pieces; p++ {
+ // Each goroutine retrieves one piece. If we hit an
+ // error, it is sent to the errors chan so get() can
+ // return it -- but only if the error happens before
+ // ctx is done. This way, if ctx is done before we hit
+ // any other error (e.g., requesting client has hung
+ // up), we return the original ctx.Err() instead of
+ // the secondary errors from the transfers that got
+ // interrupted as a result.
go func(p int) {
defer wg.Done()
startPos := p * azureMaxGetBytes
}
var rdr io.ReadCloser
var err error
- if startPos == 0 && endPos == expectSize {
- rdr, err = v.bsClient.GetBlob(v.ContainerName, loc)
- } else {
- rdr, err = v.bsClient.GetBlobRange(v.ContainerName, loc, fmt.Sprintf("%d-%d", startPos, endPos-1), nil)
+ gotRdr := make(chan struct{})
+ go func() {
+ defer close(gotRdr)
+ if startPos == 0 && endPos == expectSize {
+ rdr, err = v.container.GetBlob(loc)
+ } else {
+ rdr, err = v.container.GetBlobRange(loc, startPos, endPos-1, nil)
+ }
+ }()
+ select {
+ case <-ctx.Done():
+ go func() {
+ <-gotRdr
+ if err == nil {
+ rdr.Close()
+ }
+ }()
+ return
+ case <-gotRdr:
}
if err != nil {
- errors[p] = err
+ errors <- err
+ cancel()
return
}
- defer rdr.Close()
+ go func() {
+ // Close the reader when the client
+ // hangs up or another piece fails
+ // (possibly interrupting ReadFull())
+ // or when all pieces succeed and
+ // get() returns.
+ <-ctx.Done()
+ rdr.Close()
+ }()
n, err := io.ReadFull(rdr, buf[startPos:endPos])
if pieces == 1 && (err == io.ErrUnexpectedEOF || err == io.EOF) {
// If we don't know the actual size,
// and just tried reading 64 MiB, it's
// normal to encounter EOF.
} else if err != nil {
- errors[p] = err
+ if ctx.Err() == nil {
+ errors <- err
+ }
+ cancel()
+ return
}
if p == pieces-1 {
actualSize = startPos + n
}(p)
}
wg.Wait()
- for _, err := range errors {
- if err != nil {
- return 0, v.translateError(err)
- }
+ close(errors)
+ if len(errors) > 0 {
+ return 0, v.translateError(<-errors)
+ }
+ if ctx.Err() != nil {
+ return 0, ctx.Err()
}
return actualSize, nil
}
if trashed {
return os.ErrNotExist
}
- rdr, err := v.bsClient.GetBlob(v.ContainerName, loc)
+ var rdr io.ReadCloser
+ gotRdr := make(chan struct{})
+ go func() {
+ defer close(gotRdr)
+ rdr, err = v.container.GetBlob(loc)
+ }()
+ select {
+ case <-ctx.Done():
+ go func() {
+ <-gotRdr
+ if err == nil {
+ rdr.Close()
+ }
+ }()
+ return ctx.Err()
+ case <-gotRdr:
+ }
if err != nil {
return v.translateError(err)
}
if v.ReadOnly {
return MethodDisabledError
}
- return v.bsClient.CreateBlockBlobFromReader(v.ContainerName, loc, uint64(len(block)), bytes.NewReader(block), nil)
+ // Send the block data through a pipe, so that (if we need to)
+ // we can close the pipe early and abandon our
+ // CreateBlockBlobFromReader() goroutine, without worrying
+ // about CreateBlockBlobFromReader() accessing our block
+ // buffer after we release it.
+ bufr, bufw := io.Pipe()
+ go func() {
+ io.Copy(bufw, bytes.NewReader(block))
+ bufw.Close()
+ }()
+ errChan := make(chan error)
+ go func() {
+ var body io.Reader = bufr
+ if len(block) == 0 {
+ // We must send a "Content-Length: 0" header,
+ // but the http client interprets
+ // ContentLength==0 as "unknown" unless it can
+ // confirm by introspection that Body will
+ // read 0 bytes.
+ body = http.NoBody
+ bufr.Close()
+ }
+ errChan <- v.container.CreateBlockBlobFromReader(loc, len(block), body, nil)
+ }()
+ select {
+ case <-ctx.Done():
+ theConfig.debugLogf("%s: taking CreateBlockBlobFromReader's input away: %s", v, ctx.Err())
+ // Our pipe might be stuck in Write(), waiting for
+ // io.Copy() to read. If so, un-stick it. This means
+ // CreateBlockBlobFromReader will get corrupt data,
+ // but that's OK: the size won't match, so the write
+ // will fail.
+ go io.Copy(ioutil.Discard, bufr)
+ // CloseWithError() will return once pending I/O is done.
+ bufw.CloseWithError(ctx.Err())
+ theConfig.debugLogf("%s: abandoning CreateBlockBlobFromReader goroutine", v)
+ return ctx.Err()
+ case err := <-errChan:
+ return err
+ }
}
// Touch updates the last-modified property of a block blob.
return os.ErrNotExist
}
- metadata["touch"] = fmt.Sprintf("%d", time.Now())
- return v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil)
+ metadata["touch"] = fmt.Sprintf("%d", time.Now().Unix())
+ return v.container.SetBlobMetadata(loc, metadata, nil)
}
// Mtime returns the last-modified property of a block blob.
return time.Time{}, os.ErrNotExist
}
- props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
+ props, err := v.container.GetBlobProperties(loc)
if err != nil {
return time.Time{}, err
}
- return time.Parse(time.RFC1123, props.LastModified)
+ return time.Time(props.LastModified), nil
}
// IndexTo writes a list of Keep blocks that are stored in the
func (v *AzureBlobVolume) IndexTo(prefix string, writer io.Writer) error {
params := storage.ListBlobsParameters{
Prefix: prefix,
- Include: "metadata",
+ Include: &storage.IncludeBlobDataset{Metadata: true},
}
for {
- resp, err := v.bsClient.ListBlobs(v.ContainerName, params)
+ resp, err := v.container.ListBlobs(params)
if err != nil {
return err
}
for _, b := range resp.Blobs {
- t, err := time.Parse(time.RFC1123, b.Properties.LastModified)
- if err != nil {
- return err
- }
if !v.isKeepBlock(b.Name) {
continue
}
- if b.Properties.ContentLength == 0 && t.Add(azureWriteRaceInterval).After(time.Now()) {
+ modtime := time.Time(b.Properties.LastModified)
+ if b.Properties.ContentLength == 0 && modtime.Add(azureWriteRaceInterval).After(time.Now()) {
// A new zero-length blob is probably
// just a new non-empty blob that
// hasn't committed its data yet (see
// Trashed blob; exclude it from response
continue
}
- fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, t.UnixNano())
+ fmt.Fprintf(writer, "%s+%d %d\n", b.Name, b.Properties.ContentLength, modtime.UnixNano())
}
if resp.NextMarker == "" {
return nil
// we get the Etag before checking Mtime, and use If-Match to
// ensure we don't delete data if Put() or Touch() happens
// between our calls to Mtime() and DeleteBlob().
- props, err := v.bsClient.GetBlobProperties(v.ContainerName, loc)
+ props, err := v.container.GetBlobProperties(loc)
if err != nil {
return err
}
// If TrashLifetime == 0, just delete it
if theConfig.TrashLifetime == 0 {
- return v.bsClient.DeleteBlob(v.ContainerName, loc, map[string]string{
- "If-Match": props.Etag,
+ return v.container.DeleteBlob(loc, &storage.DeleteBlobOptions{
+ IfMatch: props.Etag,
})
}
// Otherwise, mark as trash
- return v.bsClient.SetBlobMetadata(v.ContainerName, loc, map[string]string{
+ return v.container.SetBlobMetadata(loc, storage.BlobMetadata{
"expires_at": fmt.Sprintf("%d", time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()),
- }, map[string]string{
- "If-Match": props.Etag,
+ }, &storage.SetBlobMetadataOptions{
+ IfMatch: props.Etag,
})
}
// Delete the expires_at metadata attribute
func (v *AzureBlobVolume) Untrash(loc string) error {
// if expires_at does not exist, return NotFoundError
- metadata, err := v.bsClient.GetBlobMetadata(v.ContainerName, loc)
+ metadata, err := v.container.GetBlobMetadata(loc)
if err != nil {
return v.translateError(err)
}
// reset expires_at metadata attribute
metadata["expires_at"] = ""
- err = v.bsClient.SetBlobMetadata(v.ContainerName, loc, metadata, nil)
+ err = v.container.SetBlobMetadata(loc, metadata, nil)
return v.translateError(err)
}
return v.AzureReplication
}
+// GetStorageClasses implements Volume
+func (v *AzureBlobVolume) GetStorageClasses() []string {
+ return v.StorageClasses
+}
+
// If possible, translate an Azure SDK error to a recognizable error
// like os.ErrNotExist.
func (v *AzureBlobVolume) translateError(err error) error {
// and deletes them from the volume.
func (v *AzureBlobVolume) EmptyTrash() {
var bytesDeleted, bytesInTrash int64
- var blocksDeleted, blocksInTrash int
- params := storage.ListBlobsParameters{Include: "metadata"}
+ var blocksDeleted, blocksInTrash int64
- for {
- resp, err := v.bsClient.ListBlobs(v.ContainerName, params)
+ doBlob := func(b storage.Blob) {
+ // Check whether the block is flagged as trash
+ if b.Metadata["expires_at"] == "" {
+ return
+ }
+
+ atomic.AddInt64(&blocksInTrash, 1)
+ atomic.AddInt64(&bytesInTrash, b.Properties.ContentLength)
+
+ expiresAt, err := strconv.ParseInt(b.Metadata["expires_at"], 10, 64)
if err != nil {
- log.Printf("EmptyTrash: ListBlobs: %v", err)
- break
+ log.Printf("EmptyTrash: ParseInt(%v): %v", b.Metadata["expires_at"], err)
+ return
}
- for _, b := range resp.Blobs {
- // Check if the block is expired
- if b.Metadata["expires_at"] == "" {
- continue
- }
- blocksInTrash++
- bytesInTrash += b.Properties.ContentLength
+ if expiresAt > time.Now().Unix() {
+ return
+ }
- expiresAt, err := strconv.ParseInt(b.Metadata["expires_at"], 10, 64)
- if err != nil {
- log.Printf("EmptyTrash: ParseInt(%v): %v", b.Metadata["expires_at"], err)
- continue
- }
+ err = v.container.DeleteBlob(b.Name, &storage.DeleteBlobOptions{
+ IfMatch: b.Properties.Etag,
+ })
+ if err != nil {
+ log.Printf("EmptyTrash: DeleteBlob(%v): %v", b.Name, err)
+ return
+ }
+ atomic.AddInt64(&blocksDeleted, 1)
+ atomic.AddInt64(&bytesDeleted, b.Properties.ContentLength)
+ }
- if expiresAt > time.Now().Unix() {
- continue
+ var wg sync.WaitGroup
+ todo := make(chan storage.Blob, theConfig.EmptyTrashWorkers)
+ for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
+ wg.Add(1)
+ go func() {
+ defer wg.Done()
+ for b := range todo {
+ doBlob(b)
}
+ }()
+ }
- err = v.bsClient.DeleteBlob(v.ContainerName, b.Name, map[string]string{
- "If-Match": b.Properties.Etag,
- })
- if err != nil {
- log.Printf("EmptyTrash: DeleteBlob(%v): %v", b.Name, err)
- continue
- }
- blocksDeleted++
- bytesDeleted += b.Properties.ContentLength
+ params := storage.ListBlobsParameters{Include: &storage.IncludeBlobDataset{Metadata: true}}
+ for {
+ resp, err := v.container.ListBlobs(params)
+ if err != nil {
+ log.Printf("EmptyTrash: ListBlobs: %v", err)
+ break
+ }
+ for _, b := range resp.Blobs {
+ todo <- b
}
if resp.NextMarker == "" {
break
}
params.Marker = resp.NextMarker
}
+ close(todo)
+ wg.Wait()
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)
}
+
+// InternalStats returns bucket I/O and API call counters.
+func (v *AzureBlobVolume) InternalStats() interface{} {
+ return &v.container.stats
+}
+
+type azureBlobStats struct {
+ statsTicker
+ Ops uint64
+ GetOps uint64
+ GetRangeOps uint64
+ GetMetadataOps uint64
+ GetPropertiesOps uint64
+ CreateOps uint64
+ SetMetadataOps uint64
+ DelOps uint64
+ ListOps uint64
+}
+
+func (s *azureBlobStats) TickErr(err error) {
+ if err == nil {
+ return
+ }
+ errType := fmt.Sprintf("%T", err)
+ if err, ok := err.(storage.AzureStorageServiceError); ok {
+ errType = errType + fmt.Sprintf(" %d (%s)", err.StatusCode, err.Code)
+ }
+ log.Printf("errType %T, err %s", err, err)
+ s.statsTicker.TickErr(err, errType)
+}
+
+// azureContainer wraps storage.Container in order to count I/O and
+// API usage stats.
+type azureContainer struct {
+ ctr *storage.Container
+ stats azureBlobStats
+}
+
+func (c *azureContainer) Exists() (bool, error) {
+ c.stats.Tick(&c.stats.Ops)
+ ok, err := c.ctr.Exists()
+ c.stats.TickErr(err)
+ return ok, err
+}
+
+func (c *azureContainer) GetBlobMetadata(bname string) (storage.BlobMetadata, error) {
+ c.stats.Tick(&c.stats.Ops, &c.stats.GetMetadataOps)
+ b := c.ctr.GetBlobReference(bname)
+ err := b.GetMetadata(nil)
+ c.stats.TickErr(err)
+ return b.Metadata, err
+}
+
+func (c *azureContainer) GetBlobProperties(bname string) (*storage.BlobProperties, error) {
+ c.stats.Tick(&c.stats.Ops, &c.stats.GetPropertiesOps)
+ b := c.ctr.GetBlobReference(bname)
+ err := b.GetProperties(nil)
+ c.stats.TickErr(err)
+ return &b.Properties, err
+}
+
+func (c *azureContainer) GetBlob(bname string) (io.ReadCloser, error) {
+ c.stats.Tick(&c.stats.Ops, &c.stats.GetOps)
+ b := c.ctr.GetBlobReference(bname)
+ rdr, err := b.Get(nil)
+ c.stats.TickErr(err)
+ return NewCountingReader(rdr, c.stats.TickInBytes), err
+}
+
+func (c *azureContainer) GetBlobRange(bname string, start, end int, opts *storage.GetBlobOptions) (io.ReadCloser, error) {
+ c.stats.Tick(&c.stats.Ops, &c.stats.GetRangeOps)
+ b := c.ctr.GetBlobReference(bname)
+ rdr, err := b.GetRange(&storage.GetBlobRangeOptions{
+ Range: &storage.BlobRange{
+ Start: uint64(start),
+ End: uint64(end),
+ },
+ GetBlobOptions: opts,
+ })
+ c.stats.TickErr(err)
+ return NewCountingReader(rdr, c.stats.TickInBytes), err
+}
+
+// If we give it an io.Reader that doesn't also have a Len() int
+// method, the Azure SDK determines data size by copying the data into
+// a new buffer, which is not a good use of memory.
+type readerWithAzureLen struct {
+ io.Reader
+ len int
+}
+
+// Len satisfies the private lener interface in azure-sdk-for-go.
+func (r *readerWithAzureLen) Len() int {
+ return r.len
+}
+
+func (c *azureContainer) CreateBlockBlobFromReader(bname string, size int, rdr io.Reader, opts *storage.PutBlobOptions) error {
+ c.stats.Tick(&c.stats.Ops, &c.stats.CreateOps)
+ if size != 0 {
+ rdr = &readerWithAzureLen{
+ Reader: NewCountingReader(rdr, c.stats.TickOutBytes),
+ len: size,
+ }
+ }
+ b := c.ctr.GetBlobReference(bname)
+ err := b.CreateBlockBlobFromReader(rdr, opts)
+ c.stats.TickErr(err)
+ return err
+}
+
+func (c *azureContainer) SetBlobMetadata(bname string, m storage.BlobMetadata, opts *storage.SetBlobMetadataOptions) error {
+ c.stats.Tick(&c.stats.Ops, &c.stats.SetMetadataOps)
+ b := c.ctr.GetBlobReference(bname)
+ b.Metadata = m
+ err := b.SetMetadata(opts)
+ c.stats.TickErr(err)
+ return err
+}
+
+func (c *azureContainer) ListBlobs(params storage.ListBlobsParameters) (storage.BlobListResponse, error) {
+ c.stats.Tick(&c.stats.Ops, &c.stats.ListOps)
+ resp, err := c.ctr.ListBlobs(params)
+ c.stats.TickErr(err)
+ return resp, err
+}
+
+func (c *azureContainer) DeleteBlob(bname string, opts *storage.DeleteBlobOptions) error {
+ c.stats.Tick(&c.stats.Ops, &c.stats.DelOps)
+ b := c.ctr.GetBlobReference(bname)
+ err := b.Delete(opts)
+ c.stats.TickErr(err)
+ return err
+}
projects / arvados.git / blobdiff