+// 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"
s3RaceWindow time.Duration
s3ACL = s3.Private
+
+ zeroTime time.Time
)
const (
&s3UnsafeDelete,
"s3-unsafe-delete",
false,
- "EXPERIMENTAL. Enable deletion (garbage collection), even though there are known race conditions that can cause data loss.")
+ "EXPERIMENTAL. Enable deletion (garbage collection) even when trash lifetime is zero, even though there are known race conditions that can cause data loss.")
}
// S3Volume implements Volume using an S3 bucket.
RaceWindow arvados.Duration
ReadOnly bool
UnsafeDelete bool
+ StorageClasses []string
- bucket *s3.Bucket
- bucketStats bucketStats
- volumeStats ioStats
+ bucket *s3bucket
startOnce sync.Once
}
-type bucketStats struct {
- Errors uint64
- Ops uint64
- GetOps uint64
- PutOps uint64
- HeadOps uint64
- DelOps uint64
- InBytes uint64
- OutBytes uint64
-}
-
// Examples implements VolumeWithExamples.
func (*S3Volume) Examples() []Volume {
return []Volume{
}
client := s3.New(auth, region)
+ if region.EC2Endpoint.Signer == aws.V4Signature {
+ // Currently affects only eu-central-1
+ client.Signature = aws.V4Signature
+ }
client.ConnectTimeout = time.Duration(v.ConnectTimeout)
client.ReadTimeout = time.Duration(v.ReadTimeout)
- v.bucket = &s3.Bucket{
- S3: client,
- Name: v.Bucket,
+ v.bucket = &s3bucket{
+ Bucket: &s3.Bucket{
+ S3: client,
+ Name: v.Bucket,
+ },
}
return nil
}
+// DeviceID returns a globally unique ID for the storage bucket.
+func (v *S3Volume) DeviceID() string {
+ return "s3://" + v.Endpoint + "/" + v.Bucket
+}
+
func (v *S3Volume) getReaderWithContext(ctx context.Context, loc string) (rdr io.ReadCloser, err error) {
ready := make(chan bool)
go func() {
// disappeared in a Trash race, getReader calls fixRace to recover the
// data, and tries again.
func (v *S3Volume) getReader(loc string) (rdr io.ReadCloser, err error) {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.GetOps)
rdr, err = v.bucket.GetReader(loc)
err = v.translateError(err)
- if err == nil {
- rdr = NewCountingReader(rdr, v.tickInBytes)
- return
- } else if !os.IsNotExist(v.tickErr(err)) {
+ if err == nil || !os.IsNotExist(err) {
return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err = v.bucket.Head("recent/"+loc, nil)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
if err != nil {
// If we can't read recent/X, there's no point in
// trying fixRace. Give up.
return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.GetOps)
rdr, err = v.bucket.GetReader(loc)
if err != nil {
log.Printf("warning: reading %s after successful fixRace: %s", loc, err)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
}
- rdr = NewCountingReader(rdr, v.tickInBytes)
return
}
// Compare the given data with the stored data.
func (v *S3Volume) Compare(ctx context.Context, loc string, expect []byte) error {
+ errChan := make(chan error, 1)
+ go func() {
+ _, err := v.bucket.Head("recent/"+loc, nil)
+ errChan <- err
+ }()
+ var err error
+ select {
+ case <-ctx.Done():
+ return ctx.Err()
+ case err = <-errChan:
+ }
+ if err != nil {
+ // Checking for "loc" itself here would interfere with
+ // future GET requests.
+ //
+ // On AWS, if X doesn't exist, a HEAD or GET request
+ // for X causes X's non-existence to be cached. Thus,
+ // if we test for X, then create X and return a
+ // signature to our client, the client might still get
+ // 404 from all keepstores when trying to read it.
+ //
+ // To avoid this, we avoid doing HEAD X or GET X until
+ // we know X has been written.
+ //
+ // Note that X might exist even though recent/X
+ // doesn't: for example, the response to HEAD recent/X
+ // might itself come from a stale cache. In such
+ // cases, we will return a false negative and
+ // PutHandler might needlessly create another replica
+ // on a different volume. That's not ideal, but it's
+ // better than passing the eventually-consistent
+ // problem on to our clients.
+ return v.translateError(err)
+ }
rdr, err := v.getReaderWithContext(ctx, loc)
if err != nil {
return err
}
}()
defer close(ready)
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
- rdr := NewCountingReader(bufr, v.tickOutBytes)
- err = v.bucket.PutReader(loc, rdr, int64(size), "application/octet-stream", s3ACL, opts)
+ err = v.bucket.PutReader(loc, bufr, int64(size), "application/octet-stream", s3ACL, opts)
if err != nil {
- v.tickErr(err)
return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- v.tickErr(err)
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
}()
select {
case <-ctx.Done():
theConfig.debugLogf("%s: taking PutReader'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
+ // PutReader() to read. If so, un-stick it. This means
// PutReader will get corrupt data, but that's OK: the
// size and MD5 won't match, so the write will fail.
go io.Copy(ioutil.Discard, bufr)
theConfig.debugLogf("%s: abandoning PutReader goroutine", v)
return ctx.Err()
case <-ready:
+ // Unblock pipe in case PutReader did not consume it.
+ io.Copy(ioutil.Discard, bufr)
return v.translateError(err)
}
}
if v.ReadOnly {
return MethodDisabledError
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err := v.bucket.Head(loc, nil)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
if os.IsNotExist(err) && v.fixRace(loc) {
// The data object got trashed in a race, but fixRace
// rescued it.
} else if err != nil {
return err
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- return v.translateError(v.tickErr(err))
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
+ return v.translateError(err)
}
// Mtime returns the stored timestamp for the given locator.
func (v *S3Volume) Mtime(loc string) (time.Time, error) {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err := v.bucket.Head(loc, nil)
if err != nil {
- return zeroTime, v.translateError(v.tickErr(err))
+ return zeroTime, v.translateError(err)
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err := v.bucket.Head("recent/"+loc, nil)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
if os.IsNotExist(err) {
// The data object X exists, but recent/X is missing.
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
if err != nil {
log.Printf("error: creating %q: %s", "recent/"+loc, err)
- return zeroTime, v.translateError(v.tickErr(err))
+ return zeroTime, v.translateError(err)
}
log.Printf("info: created %q to migrate existing block to new storage scheme", "recent/"+loc)
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err = v.bucket.Head("recent/"+loc, nil)
if err != nil {
log.Printf("error: created %q but HEAD failed: %s", "recent/"+loc, err)
- return zeroTime, v.translateError(v.tickErr(err))
+ return zeroTime, v.translateError(err)
}
} else if err != nil {
// HEAD recent/X failed for some other reason.
func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
// Use a merge sort to find matching sets of X and recent/X.
dataL := s3Lister{
- Bucket: v.bucket,
+ Bucket: v.bucket.Bucket,
Prefix: prefix,
PageSize: v.IndexPageSize,
}
recentL := s3Lister{
- Bucket: v.bucket,
+ Bucket: v.bucket.Bucket,
Prefix: "recent/" + prefix,
PageSize: v.IndexPageSize,
}
+ v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps)
+ v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps)
for data, recent := dataL.First(), recentL.First(); data != nil; data = dataL.Next() {
+ v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps)
if data.Key >= "g" {
// Conveniently, "recent/*" and "trash/*" are
// lexically greater than all hex-encoded data
for recent != nil {
if cmp := strings.Compare(recent.Key[7:], data.Key); cmp < 0 {
recent = recentL.Next()
+ v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps)
continue
} else if cmp == 0 {
stamp = recent
recent = recentL.Next()
+ v.bucket.stats.Tick(&v.bucket.stats.Ops, &v.bucket.stats.ListOps)
break
} else {
// recent/X marker is missing: we'll
if !s3UnsafeDelete {
return ErrS3TrashDisabled
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
- return v.translateError(v.tickErr(v.bucket.Del(loc)))
+ return v.translateError(v.bucket.Del(loc))
}
err := v.checkRaceWindow(loc)
if err != nil {
if err != nil {
return err
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
- return v.translateError(v.tickErr(v.bucket.Del(loc)))
+ return v.translateError(v.bucket.Del(loc))
}
// checkRaceWindow returns a non-nil error if trash/loc is, or might
// be, in the race window (i.e., it's not safe to trash loc).
func (v *S3Volume) checkRaceWindow(loc string) error {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
resp, err := v.bucket.Head("trash/"+loc, nil)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
if os.IsNotExist(err) {
// OK, trash/X doesn't exist so we're not in the race
// window
// (PutCopy returns 200 OK if the request was received, even if the
// copy failed).
func (v *S3Volume) safeCopy(dst, src string) error {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
resp, err := v.bucket.PutCopy(dst, s3ACL, s3.CopyOptions{
ContentType: "application/octet-stream",
MetadataDirective: "REPLACE",
}, v.bucket.Name+"/"+src)
- err = v.translateError(v.tickErr(err))
+ err = v.translateError(err)
if err != nil {
return err
}
if err != nil {
return err
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.PutOps)
- err = v.bucket.Put("recent/"+loc, nil, "application/octet-stream", s3ACL, s3.Options{})
- return v.translateError(v.tickErr(err))
+ err = v.bucket.PutReader("recent/"+loc, nil, 0, "application/octet-stream", s3ACL, s3.Options{})
+ return v.translateError(err)
}
// Status returns a *VolumeStatus representing the current in-use
}
}
-// IOStatus implements InternalStatser.
+// InternalStats returns bucket I/O and API call counters.
func (v *S3Volume) InternalStats() interface{} {
- return &v.bucketStats
+ return &v.bucket.stats
}
// String implements fmt.Stringer.
return v.S3Replication
}
+// GetStorageClasses implements Volume
+func (v *S3Volume) GetStorageClasses() []string {
+ return v.StorageClasses
+}
+
var s3KeepBlockRegexp = regexp.MustCompile(`^[0-9a-f]{32}$`)
func (v *S3Volume) isKeepBlock(s string) bool {
// there was a race between Put and Trash, fixRace recovers from the
// race by Untrashing the block.
func (v *S3Volume) fixRace(loc string) bool {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
trash, err := v.bucket.Head("trash/"+loc, nil)
if err != nil {
- if !os.IsNotExist(v.translateError(v.tickErr(err))) {
+ if !os.IsNotExist(v.translateError(err)) {
log.Printf("error: fixRace: HEAD %q: %s", "trash/"+loc, err)
}
return false
return false
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
recent, err := v.bucket.Head("recent/"+loc, nil)
if err != nil {
- v.tickErr(err)
log.Printf("error: fixRace: HEAD %q: %s", "recent/"+loc, err)
return false
}
func (v *S3Volume) EmptyTrash() {
var bytesInTrash, blocksInTrash, bytesDeleted, blocksDeleted int64
- // Use a merge sort to find matching sets of trash/X and recent/X.
- trashL := s3Lister{
- Bucket: v.bucket,
- Prefix: "trash/",
- PageSize: v.IndexPageSize,
- }
// Define "ready to delete" as "...when EmptyTrash started".
startT := time.Now()
- for trash := trashL.First(); trash != nil; trash = trashL.Next() {
+
+ emptyOneKey := func(trash *s3.Key) {
loc := trash.Key[6:]
if !v.isKeepBlock(loc) {
- continue
+ return
}
- bytesInTrash += trash.Size
- blocksInTrash++
+ atomic.AddInt64(&bytesInTrash, trash.Size)
+ atomic.AddInt64(&blocksInTrash, 1)
trashT, err := time.Parse(time.RFC3339, trash.LastModified)
if err != nil {
log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, trash.Key, trash.LastModified, err)
- continue
+ return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
recent, err := v.bucket.Head("recent/"+loc, nil)
- if err != nil && os.IsNotExist(v.translateError(v.tickErr(err))) {
+ if err != nil && os.IsNotExist(v.translateError(err)) {
log.Printf("warning: %s: EmptyTrash: found trash marker %q but no %q (%s); calling Untrash", v, trash.Key, "recent/"+loc, err)
err = v.Untrash(loc)
if err != nil {
log.Printf("error: %s: EmptyTrash: Untrash(%q): %s", v, loc, err)
}
- continue
+ return
} else if err != nil {
log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err)
- continue
+ return
}
recentT, err := v.lastModified(recent)
if err != nil {
log.Printf("warning: %s: EmptyTrash: %q: parse %q: %s", v, "recent/"+loc, recent.Header.Get("Last-Modified"), err)
- continue
+ return
}
if trashT.Sub(recentT) < theConfig.BlobSignatureTTL.Duration() {
if age := startT.Sub(recentT); age >= theConfig.BlobSignatureTTL.Duration()-time.Duration(v.RaceWindow) {
log.Printf("notice: %s: EmptyTrash: detected old race for %q, calling fixRace + Touch", v, loc)
v.fixRace(loc)
v.Touch(loc)
- continue
+ return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err := v.bucket.Head(loc, nil)
- if os.IsNotExist(v.tickErr(err)) {
+ if os.IsNotExist(err) {
log.Printf("notice: %s: EmptyTrash: detected recent race for %q, calling fixRace", v, loc)
v.fixRace(loc)
- continue
+ return
} else if err != nil {
log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err)
- continue
+ return
}
}
if startT.Sub(trashT) < theConfig.TrashLifetime.Duration() {
- continue
+ return
}
- v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
err = v.bucket.Del(trash.Key)
if err != nil {
- v.tickErr(err)
log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, trash.Key, err)
- continue
+ return
}
- bytesDeleted += trash.Size
- blocksDeleted++
+ atomic.AddInt64(&bytesDeleted, trash.Size)
+ atomic.AddInt64(&blocksDeleted, 1)
- v.tick(&v.bucketStats.Ops, &v.bucketStats.HeadOps)
_, err = v.bucket.Head(loc, nil)
- if os.IsNotExist(v.tickErr(err)) {
- v.tick(&v.bucketStats.Ops, &v.bucketStats.DelOps)
- err = v.bucket.Del("recent/" + loc)
- if err != nil {
- v.tickErr(err)
- log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err)
- }
- } else if err != nil {
- log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, "recent/"+loc, err)
+ if err == nil {
+ log.Printf("warning: %s: EmptyTrash: HEAD %q succeeded immediately after deleting %q", v, loc, loc)
+ return
+ }
+ if !os.IsNotExist(v.translateError(err)) {
+ log.Printf("warning: %s: EmptyTrash: HEAD %q: %s", v, loc, err)
+ return
+ }
+ err = v.bucket.Del("recent/" + loc)
+ if err != nil {
+ log.Printf("warning: %s: EmptyTrash: deleting %q: %s", v, "recent/"+loc, err)
}
}
- if err := trashL.Error(); err != nil {
- log.Printf("error: %s: EmptyTrash: lister: %s", v, err)
- }
- 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)
-}
-func (v *S3Volume) tick(counters ...*uint64) {
- for _, counter := range counters {
- atomic.AddUint64(counter, 1)
+ var wg sync.WaitGroup
+ todo := make(chan *s3.Key, theConfig.EmptyTrashWorkers)
+ for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
+ wg.Add(1)
+ go func() {
+ defer wg.Done()
+ for key := range todo {
+ emptyOneKey(key)
+ }
+ }()
}
-}
-func (v *S3Volume) tickErr(err error) error {
- if err != nil {
- atomic.AddUint64(&v.bucketStats.Errors, 1)
+ trashL := s3Lister{
+ Bucket: v.bucket.Bucket,
+ Prefix: "trash/",
+ PageSize: v.IndexPageSize,
}
- return err
-}
-
-func (v *S3Volume) tickInBytes(n uint64) {
- atomic.AddUint64(&v.bucketStats.InBytes, n)
-}
+ for trash := trashL.First(); trash != nil; trash = trashL.Next() {
+ todo <- trash
+ }
+ close(todo)
+ wg.Wait()
-func (v *S3Volume) tickOutBytes(n uint64) {
- atomic.AddUint64(&v.bucketStats.OutBytes, n)
+ if err := trashL.Error(); err != nil {
+ log.Printf("error: %s: EmptyTrash: lister: %s", v, err)
+ }
+ 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)
}
type s3Lister struct {
}
return
}
+
+// s3bucket wraps s3.bucket and counts I/O and API usage stats.
+type s3bucket struct {
+ *s3.Bucket
+ stats s3bucketStats
+}
+
+func (b *s3bucket) GetReader(path string) (io.ReadCloser, error) {
+ rdr, err := b.Bucket.GetReader(path)
+ b.stats.Tick(&b.stats.Ops, &b.stats.GetOps)
+ b.stats.TickErr(err)
+ return NewCountingReader(rdr, b.stats.TickInBytes), err
+}
+
+func (b *s3bucket) Head(path string, headers map[string][]string) (*http.Response, error) {
+ resp, err := b.Bucket.Head(path, headers)
+ b.stats.Tick(&b.stats.Ops, &b.stats.HeadOps)
+ b.stats.TickErr(err)
+ return resp, err
+}
+
+func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error {
+ if length == 0 {
+ // goamz will only send Content-Length: 0 when reader
+ // is nil due to net.http.Request.ContentLength
+ // behavior. Otherwise, Content-Length header is
+ // omitted which will cause some S3 services
+ // (including AWS and Ceph RadosGW) to fail to create
+ // empty objects.
+ r = nil
+ } else {
+ r = NewCountingReader(r, b.stats.TickOutBytes)
+ }
+ err := b.Bucket.PutReader(path, r, length, contType, perm, options)
+ b.stats.Tick(&b.stats.Ops, &b.stats.PutOps)
+ b.stats.TickErr(err)
+ return err
+}
+
+func (b *s3bucket) Del(path string) error {
+ err := b.Bucket.Del(path)
+ b.stats.Tick(&b.stats.Ops, &b.stats.DelOps)
+ b.stats.TickErr(err)
+ return err
+}
+
+type s3bucketStats struct {
+ statsTicker
+ Ops uint64
+ GetOps uint64
+ PutOps uint64
+ HeadOps uint64
+ DelOps uint64
+ ListOps uint64
+}
+
+func (s *s3bucketStats) TickErr(err error) {
+ if err == nil {
+ return
+ }
+ errType := fmt.Sprintf("%T", err)
+ if err, ok := err.(*s3.Error); ok {
+ errType = errType + fmt.Sprintf(" %d %s", err.StatusCode, err.Code)
+ }
+ s.statsTicker.TickErr(err, errType)
+}
projects / arvados.git / blobdiff