"regexp"
"strings"
"sync"
+ "sync/atomic"
"time"
"git.curoverse.com/arvados.git/sdk/go/arvados"
"github.com/AdRoll/goamz/aws"
"github.com/AdRoll/goamz/s3"
- log "github.com/Sirupsen/logrus"
)
const (
s3RaceWindow time.Duration
s3ACL = s3.Private
+
+ zeroTime time.Time
)
const (
RaceWindow arvados.Duration
ReadOnly bool
UnsafeDelete bool
+ StorageClasses []string
bucket *s3bucket
// 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
if err != nil {
return
}
- 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{})
}()
select {
case <-ctx.Done():
} else if err != nil {
return err
}
- 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{})
return v.translateError(err)
}
err = v.translateError(err)
if os.IsNotExist(err) {
// The data object X exists, but recent/X is missing.
- 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(err)
if err != nil {
return err
}
- 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{})
return v.translateError(err)
}
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 {
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.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
}
recent, err := v.bucket.Head("recent/"+loc, nil)
if err != nil && os.IsNotExist(v.translateError(err)) {
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
}
_, err := v.bucket.Head(loc, nil)
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
}
err = v.bucket.Del(trash.Key)
if err != nil {
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)
_, err = v.bucket.Head(loc, nil)
- if os.IsNotExist(err) {
- err = v.bucket.Del("recent/" + loc)
- if err != nil {
- 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)
+ }
+ }
+
+ 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)
+ }
+ }()
+ }
+
+ trashL := s3Lister{
+ Bucket: v.bucket.Bucket,
+ Prefix: "trash/",
+ PageSize: v.IndexPageSize,
+ }
+ for trash := trashL.First(); trash != nil; trash = trashL.Next() {
+ todo <- trash
}
+ close(todo)
+ wg.Wait()
+
if err := trashL.Error(); err != nil {
log.Printf("error: %s: EmptyTrash: lister: %s", v, err)
}
}
func (b *s3bucket) PutReader(path string, r io.Reader, length int64, contType string, perm s3.ACL, options s3.Options) error {
- err := b.Bucket.PutReader(path, NewCountingReader(r, b.stats.TickOutBytes), length, contType, perm, options)
- b.stats.Tick(&b.stats.Ops, &b.stats.PutOps)
- b.stats.TickErr(err)
- return err
-}
-
-func (b *s3bucket) Put(path string, data []byte, contType string, perm s3.ACL, options s3.Options) error {
- err := b.Bucket.PutReader(path, NewCountingReader(bytes.NewBuffer(data), b.stats.TickOutBytes), int64(len(data)), contType, perm, options)
+ 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
projects / arvados.git / blobdiff