X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/ba1cbe5824ddc769d8b1fdb1f51a4e02187e778c..1dfede4c84c961230954a4dad9ab348029dfa128:/services/keepstore/s3_volume.go?ds=sidebyside

diff --git a/services/keepstore/s3_volume.go b/services/keepstore/s3_volume.go
index 90e8a1b4f6..68406d5a28 100644
--- a/services/keepstore/s3_volume.go
+++ b/services/keepstore/s3_volume.go
@@ -18,6 +18,7 @@ import (
 	"regexp"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"time"
 
 	"git.curoverse.com/arvados.git/sdk/go/arvados"
@@ -152,6 +153,7 @@ type S3Volume struct {
 	RaceWindow         arvados.Duration
 	ReadOnly           bool
 	UnsafeDelete       bool
+	StorageClasses     []string
 
 	bucket *s3bucket
 
@@ -233,6 +235,10 @@ func (v *S3Volume) Start() error {
 	}
 
 	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 = &s3bucket{
@@ -427,7 +433,7 @@ func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error {
 	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)
@@ -436,6 +442,8 @@ func (v *S3Volume) Put(ctx context.Context, loc string, block []byte) error {
 		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)
 	}
 }
@@ -493,16 +501,15 @@ func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
 		Bucket:   v.bucket.Bucket,
 		Prefix:   prefix,
 		PageSize: v.IndexPageSize,
+		Stats:    &v.bucket.stats,
 	}
 	recentL := s3Lister{
 		Bucket:   v.bucket.Bucket,
 		Prefix:   "recent/" + prefix,
 		PageSize: v.IndexPageSize,
+		Stats:    &v.bucket.stats,
 	}
-	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)
+	for data, recent := dataL.First(), recentL.First(); data != nil && dataL.Error() == nil; data = dataL.Next() {
 		if data.Key >= "g" {
 			// Conveniently, "recent/*" and "trash/*" are
 			// lexically greater than all hex-encoded data
@@ -521,15 +528,13 @@ func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
 		stamp := data
 
 		// Advance to the corresponding recent/X marker, if any
-		for recent != nil {
+		for recent != nil && recentL.Error() == 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
@@ -538,13 +543,16 @@ func (v *S3Volume) IndexTo(prefix string, writer io.Writer) error {
 				break
 			}
 		}
+		if err := recentL.Error(); err != nil {
+			return err
+		}
 		t, err := time.Parse(time.RFC3339, stamp.LastModified)
 		if err != nil {
 			return err
 		}
 		fmt.Fprintf(writer, "%s+%d %d\n", data.Key, data.Size, t.UnixNano())
 	}
-	return nil
+	return dataL.Error()
 }
 
 // Trash a Keep block.
@@ -616,8 +624,10 @@ func (v *S3Volume) safeCopy(dst, src string) error {
 		MetadataDirective: "REPLACE",
 	}, v.bucket.Name+"/"+src)
 	err = v.translateError(err)
-	if err != nil {
+	if os.IsNotExist(err) {
 		return err
+	} else if err != nil {
+		return fmt.Errorf("PutCopy(%q ← %q): %s", dst, v.bucket.Name+"/"+src, err)
 	}
 	if t, err := time.Parse(time.RFC3339Nano, resp.LastModified); err != nil {
 		return fmt.Errorf("PutCopy succeeded but did not return a timestamp: %q: %s", resp.LastModified, err)
@@ -686,6 +696,11 @@ func (v *S3Volume) Replication() int {
 	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 {
@@ -758,26 +773,21 @@ func (v *S3Volume) translateError(err error) error {
 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)) {
@@ -786,15 +796,15 @@ func (v *S3Volume) EmptyTrash() {
 			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) {
@@ -809,39 +819,68 @@ func (v *S3Volume) EmptyTrash() {
 				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,
+		Stats:    &v.bucket.stats,
+	}
+	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)
 	}
@@ -852,6 +891,7 @@ type s3Lister struct {
 	Bucket     *s3.Bucket
 	Prefix     string
 	PageSize   int
+	Stats      *s3bucketStats
 	nextMarker string
 	buf        []s3.Key
 	err        error
@@ -880,6 +920,7 @@ func (lister *s3Lister) Error() error {
 }
 
 func (lister *s3Lister) getPage() {
+	lister.Stats.Tick(&lister.Stats.Ops, &lister.Stats.ListOps)
 	resp, err := lister.Bucket.List(lister.Prefix, "", lister.nextMarker, lister.PageSize)
 	lister.nextMarker = ""
 	if err != nil {