1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
5 // Package keepstore implements the keepstore service component and
6 // back-end storage drivers.
8 // It is an internal module, only intended to be imported by
9 // /cmd/arvados-server and other server-side components in this
29 "git.arvados.org/arvados.git/sdk/go/arvados"
30 "git.arvados.org/arvados.git/sdk/go/arvadosclient"
31 "git.arvados.org/arvados.git/sdk/go/auth"
32 "git.arvados.org/arvados.git/sdk/go/ctxlog"
33 "git.arvados.org/arvados.git/sdk/go/httpserver"
34 "git.arvados.org/arvados.git/sdk/go/keepclient"
35 "github.com/prometheus/client_golang/prometheus"
36 "github.com/sirupsen/logrus"
39 // Maximum size of a keep block is 64 MiB.
40 const BlockSize = 1 << 26
43 errChecksum = httpserver.ErrorWithStatus(errors.New("checksum mismatch in stored data"), http.StatusBadGateway)
44 errNoTokenProvided = httpserver.ErrorWithStatus(errors.New("no token provided in Authorization header"), http.StatusUnauthorized)
45 errMethodNotAllowed = httpserver.ErrorWithStatus(errors.New("method not allowed"), http.StatusMethodNotAllowed)
46 errVolumeUnavailable = httpserver.ErrorWithStatus(errors.New("volume unavailable"), http.StatusServiceUnavailable)
47 errCollision = httpserver.ErrorWithStatus(errors.New("hash collision"), http.StatusInternalServerError)
48 errExpiredSignature = httpserver.ErrorWithStatus(errors.New("expired signature"), http.StatusUnauthorized)
49 errInvalidSignature = httpserver.ErrorWithStatus(errors.New("invalid signature"), http.StatusBadRequest)
50 errInvalidLocator = httpserver.ErrorWithStatus(errors.New("invalid locator"), http.StatusBadRequest)
51 errFull = httpserver.ErrorWithStatus(errors.New("insufficient storage"), http.StatusInsufficientStorage)
52 errTooLarge = httpserver.ErrorWithStatus(errors.New("request entity too large"), http.StatusRequestEntityTooLarge)
53 driver = make(map[string]volumeDriver)
56 type indexOptions struct {
68 type keepstore struct {
69 cluster *arvados.Cluster
70 logger logrus.FieldLogger
71 serviceURL arvados.URL
72 mounts map[string]*mount
75 bufferPool *bufferPool
77 iostats map[volume]*ioStats
79 remoteClients map[string]*keepclient.KeepClient
80 remoteClientsMtx sync.Mutex
83 func newKeepstore(ctx context.Context, cluster *arvados.Cluster, token string, reg *prometheus.Registry, serviceURL arvados.URL) (*keepstore, error) {
84 logger := ctxlog.FromContext(ctx)
86 if cluster.API.MaxConcurrentRequests > 0 && cluster.API.MaxConcurrentRequests < cluster.API.MaxKeepBlobBuffers {
87 logger.Warnf("Possible configuration mistake: not useful to set API.MaxKeepBlobBuffers (%d) higher than API.MaxConcurrentRequests (%d)", cluster.API.MaxKeepBlobBuffers, cluster.API.MaxConcurrentRequests)
90 if cluster.Collections.BlobSigningKey != "" {
91 } else if cluster.Collections.BlobSigning {
92 return nil, errors.New("cannot enable Collections.BlobSigning with no Collections.BlobSigningKey")
94 logger.Warn("Running without a blob signing key. Block locators returned by this server will not be signed, and will be rejected by a server that enforces permissions. To fix this, configure Collections.BlobSigning and Collections.BlobSigningKey.")
97 if cluster.API.MaxKeepBlobBuffers <= 0 {
98 return nil, fmt.Errorf("API.MaxKeepBlobBuffers must be greater than zero")
100 bufferPool := newBufferPool(logger, cluster.API.MaxKeepBlobBuffers, reg)
105 serviceURL: serviceURL,
106 bufferPool: bufferPool,
107 remoteClients: make(map[string]*keepclient.KeepClient),
110 err := ks.setupMounts(newVolumeMetricsVecs(reg))
118 func (ks *keepstore) setupMounts(metrics *volumeMetricsVecs) error {
119 ks.mounts = make(map[string]*mount)
120 if len(ks.cluster.Volumes) == 0 {
121 return errors.New("no volumes configured")
123 for uuid, cfgvol := range ks.cluster.Volumes {
124 va, ok := cfgvol.AccessViaHosts[ks.serviceURL]
125 if !ok && len(cfgvol.AccessViaHosts) > 0 {
128 dri, ok := driver[cfgvol.Driver]
130 return fmt.Errorf("volume %s: invalid driver %q", uuid, cfgvol.Driver)
132 vol, err := dri(newVolumeParams{
135 ConfigVolume: cfgvol,
137 MetricsVecs: metrics,
138 BufferPool: ks.bufferPool,
141 return fmt.Errorf("error initializing volume %s: %s", uuid, err)
143 sc := cfgvol.StorageClasses
145 sc = map[string]bool{"default": true}
147 repl := cfgvol.Replication
152 for class, in := range cfgvol.StorageClasses {
153 p := ks.cluster.StorageClasses[class].Priority
161 KeepMount: arvados.KeepMount{
163 DeviceID: vol.DeviceID(),
164 AllowWrite: !va.ReadOnly && !cfgvol.ReadOnly,
165 AllowTrash: !va.ReadOnly && (!cfgvol.ReadOnly || cfgvol.AllowTrashWhenReadOnly),
170 ks.mounts[uuid] = mnt
171 ks.logger.Printf("started volume %s (%s), AllowWrite=%v, AllowTrash=%v", uuid, vol.DeviceID(), mnt.AllowWrite, mnt.AllowTrash)
173 if len(ks.mounts) == 0 {
174 return fmt.Errorf("no volumes configured for %s", ks.serviceURL)
179 for _, mnt := range ks.mounts {
180 ks.mountsR = append(ks.mountsR, mnt)
182 ks.mountsW = append(ks.mountsW, mnt)
185 // Sorting mounts by UUID makes behavior more predictable, and
186 // is convenient for testing -- for example, "index all
187 // volumes" and "trash block on all volumes" will visit
188 // volumes in predictable order.
189 sort.Slice(ks.mountsR, func(i, j int) bool { return ks.mountsR[i].UUID < ks.mountsR[j].UUID })
190 sort.Slice(ks.mountsW, func(i, j int) bool { return ks.mountsW[i].UUID < ks.mountsW[j].UUID })
194 // checkLocatorSignature checks that locator has a valid signature.
195 // If the BlobSigning config is false, it returns nil even if the
196 // signature is invalid or missing.
197 func (ks *keepstore) checkLocatorSignature(ctx context.Context, locator string) error {
198 if !ks.cluster.Collections.BlobSigning {
201 token := ctxToken(ctx)
203 return errNoTokenProvided
205 err := arvados.VerifySignature(locator, token, ks.cluster.Collections.BlobSigningTTL.Duration(), []byte(ks.cluster.Collections.BlobSigningKey))
206 if err == arvados.ErrSignatureExpired {
207 return errExpiredSignature
208 } else if err != nil {
209 return errInvalidSignature
214 // signLocator signs the locator for the given token, if possible.
215 // Note this signs if the BlobSigningKey config is available, even if
216 // the BlobSigning config is false.
217 func (ks *keepstore) signLocator(token, locator string) string {
218 if token == "" || len(ks.cluster.Collections.BlobSigningKey) == 0 {
221 ttl := ks.cluster.Collections.BlobSigningTTL.Duration()
222 return arvados.SignLocator(locator, token, time.Now().Add(ttl), ttl, []byte(ks.cluster.Collections.BlobSigningKey))
225 func (ks *keepstore) BlockRead(ctx context.Context, opts arvados.BlockReadOptions) (n int, err error) {
226 li, err := parseLocator(opts.Locator)
231 if rw, ok := out.(http.ResponseWriter); ok && li.size > 0 {
232 out = &setSizeOnWrite{ResponseWriter: rw, size: li.size}
234 if li.remote && !li.signed {
235 return ks.blockReadRemote(ctx, opts)
237 if err := ks.checkLocatorSignature(ctx, opts.Locator); err != nil {
240 hashcheck := md5.New()
242 out = newHashCheckWriter(out, hashcheck, int64(li.size), li.hash)
244 out = io.MultiWriter(out, hashcheck)
246 var errToCaller error = os.ErrNotExist
247 for _, mnt := range ks.rendezvous(li.hash, ks.mountsR) {
248 if ctx.Err() != nil {
251 n, err = mnt.BlockRead(ctx, li.hash, out)
252 if err == nil && li.size > 0 && n != li.size {
253 // If the backend read fewer bytes than
254 // expected but returns no error, we can
255 // classify this as a checksum error (even
256 // though hashCheckWriter doesn't know that
257 // yet, it's just waiting for the next
258 // write). If our caller is serving a GET
259 // request it's too late to do anything about
260 // it anyway, but if it's a HEAD request the
261 // caller can still change the response status
263 return n, errChecksum
265 if err == nil && li.size == 0 {
266 // hashCheckingWriter isn't in use because we
267 // don't know the expected size. All we can do
268 // is check after writing all the data, and
269 // trust the caller is doing a HEAD request so
270 // it's not too late to set an error code in
271 // the response header.
272 if hash := fmt.Sprintf("%x", hashcheck.Sum(nil)); hash != li.hash {
273 return n, errChecksum
276 if rw, ok := opts.WriteTo.(http.ResponseWriter); ok && li.size == 0 && err == nil {
277 // We didn't set the content-length header
278 // above because we didn't know the block size
280 rw.Header().Set("Content-Length", fmt.Sprintf("%d", n))
282 if n > 0 || err == nil {
283 // success, or there's an error but we can't
284 // retry because we've already sent some data.
287 if !os.IsNotExist(err) {
288 // If some volume returns a transient error,
289 // return it to the caller instead of "Not
290 // found" so it can retry.
294 return 0, errToCaller
297 func (ks *keepstore) blockReadRemote(ctx context.Context, opts arvados.BlockReadOptions) (int, error) {
298 ks.logger.Infof("blockReadRemote(%s)", opts.Locator)
299 token := ctxToken(ctx)
301 return 0, errNoTokenProvided
303 var remoteClient *keepclient.KeepClient
306 for i, part := range strings.Split(opts.Locator, "+") {
309 // don't try to parse hash part as hint
310 case strings.HasPrefix(part, "A"):
311 // drop local permission hint
313 case len(part) > 7 && part[0] == 'R' && part[6] == '-':
314 remoteID := part[1:6]
315 remote, ok := ks.cluster.RemoteClusters[remoteID]
317 return 0, httpserver.ErrorWithStatus(errors.New("remote cluster not configured"), http.StatusBadRequest)
319 kc, err := ks.remoteClient(remoteID, remote, token)
320 if err == auth.ErrObsoleteToken {
321 return 0, httpserver.ErrorWithStatus(err, http.StatusBadRequest)
322 } else if err != nil {
326 part = "A" + part[7:]
327 case len(part) > 0 && part[0] >= '0' && part[0] <= '9':
328 size, _ = strconv.Atoi(part)
330 parts = append(parts, part)
332 if remoteClient == nil {
333 return 0, httpserver.ErrorWithStatus(errors.New("invalid remote hint"), http.StatusBadRequest)
335 locator := strings.Join(parts, "+")
336 if opts.LocalLocator == nil {
337 // Read from remote cluster and stream response back
339 if rw, ok := opts.WriteTo.(http.ResponseWriter); ok && size > 0 {
340 rw.Header().Set("Content-Length", fmt.Sprintf("%d", size))
342 return remoteClient.BlockRead(ctx, arvados.BlockReadOptions{
344 WriteTo: opts.WriteTo,
347 // We must call LocalLocator before writing any data to
348 // opts.WriteTo, otherwise the caller can't put the local
349 // locator in a response header. So we copy into memory,
350 // generate the local signature, then copy from memory to
352 buf, err := ks.bufferPool.GetContext(ctx)
356 defer ks.bufferPool.Put(buf)
357 writebuf := bytes.NewBuffer(buf[:0])
358 ks.logger.Infof("blockReadRemote(%s): remote read(%s)", opts.Locator, locator)
359 _, err = remoteClient.BlockRead(ctx, arvados.BlockReadOptions{
366 resp, err := ks.BlockWrite(ctx, arvados.BlockWriteOptions{
368 Data: writebuf.Bytes(),
373 opts.LocalLocator(resp.Locator)
374 if rw, ok := opts.WriteTo.(http.ResponseWriter); ok {
375 rw.Header().Set("Content-Length", fmt.Sprintf("%d", writebuf.Len()))
377 n, err := io.Copy(opts.WriteTo, bytes.NewReader(writebuf.Bytes()))
381 func (ks *keepstore) remoteClient(remoteID string, remoteCluster arvados.RemoteCluster, token string) (*keepclient.KeepClient, error) {
382 ks.remoteClientsMtx.Lock()
383 kc, ok := ks.remoteClients[remoteID]
384 ks.remoteClientsMtx.Unlock()
386 c := &arvados.Client{
387 APIHost: remoteCluster.Host,
389 Insecure: remoteCluster.Insecure,
391 ac, err := arvadosclient.New(c)
395 kc, err = keepclient.MakeKeepClient(ac)
399 kc.DiskCacheSize = keepclient.DiskCacheDisabled
401 ks.remoteClientsMtx.Lock()
402 ks.remoteClients[remoteID] = kc
403 ks.remoteClientsMtx.Unlock()
405 accopy := *kc.Arvados
406 accopy.ApiToken = token
408 kccopy.Arvados = &accopy
409 token, err := auth.SaltToken(token, remoteID)
413 kccopy.Arvados.ApiToken = token
417 // BlockWrite writes a block to one or more volumes.
418 func (ks *keepstore) BlockWrite(ctx context.Context, opts arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error) {
419 var resp arvados.BlockWriteResponse
421 if opts.Data == nil {
422 buf, err := ks.bufferPool.GetContext(ctx)
426 defer ks.bufferPool.Put(buf)
427 w := bytes.NewBuffer(buf[:0])
429 limitedReader := &io.LimitedReader{R: opts.Reader, N: BlockSize}
430 n, err := io.Copy(io.MultiWriter(w, h), limitedReader)
434 if limitedReader.N == 0 {
435 // Data size is either exactly BlockSize, or too big.
436 n, err := opts.Reader.Read(make([]byte, 1))
438 return resp, httpserver.ErrorWithStatus(err, http.StatusRequestEntityTooLarge)
445 if opts.DataSize != 0 && int(n) != opts.DataSize {
446 return resp, httpserver.ErrorWithStatus(fmt.Errorf("content length %d did not match specified data size %d", n, opts.DataSize), http.StatusBadRequest)
448 hash = fmt.Sprintf("%x", h.Sum(nil))
450 hash = fmt.Sprintf("%x", md5.Sum(opts.Data))
452 if opts.Hash != "" && !strings.HasPrefix(opts.Hash, hash) {
453 return resp, httpserver.ErrorWithStatus(fmt.Errorf("content hash %s did not match specified locator %s", hash, opts.Hash), http.StatusBadRequest)
455 rvzmounts := ks.rendezvous(hash, ks.mountsW)
456 result := newPutProgress(opts.StorageClasses)
457 for _, mnt := range rvzmounts {
458 if !result.Want(mnt) {
461 cmp := &checkEqual{Expect: opts.Data}
462 if _, err := mnt.BlockRead(ctx, hash, cmp); err == nil {
464 return resp, errCollision
466 err := mnt.BlockTouch(hash)
472 var allFull atomic.Bool
474 // pending tracks what result will be if all outstanding
476 pending := result.Copy()
477 cond := sync.NewCond(new(sync.Mutex))
479 var wg sync.WaitGroup
481 for _, mnt := range rvzmounts {
483 if result.Done() || ctx.Err() != nil {
486 if !result.Want(mnt) {
489 if pending.Want(mnt) {
492 // This mount might not be needed, depending
493 // on the outcome of pending writes. Wait for
494 // a pending write to finish, then check
499 logger := ks.logger.WithField("mount", mnt.UUID)
504 logger.Debug("start write")
505 err := mnt.BlockWrite(ctx, hash, opts.Data)
507 defer cond.L.Unlock()
508 defer cond.Broadcast()
510 logger.Debug("write failed")
523 if ctx.Err() != nil {
524 return resp, ctx.Err()
526 if result.Done() || result.totalReplication > 0 {
527 resp = arvados.BlockWriteResponse{
528 Locator: ks.signLocator(ctxToken(ctx), fmt.Sprintf("%s+%d", hash, len(opts.Data))),
529 Replicas: result.totalReplication,
530 StorageClasses: result.classDone,
537 return resp, errVolumeUnavailable
540 // rendezvous sorts the given mounts by descending priority, then by
541 // rendezvous order for the given locator.
542 func (*keepstore) rendezvous(locator string, mnts []*mount) []*mount {
547 // copy the provided []*mount before doing an in-place sort
548 mnts = append([]*mount(nil), mnts...)
549 weight := make(map[*mount]string)
550 for _, mnt := range mnts {
552 if len(uuidpart) == 27 {
553 // strip zzzzz-yyyyy- prefixes
554 uuidpart = uuidpart[12:]
556 weight[mnt] = fmt.Sprintf("%x", md5.Sum([]byte(hash+uuidpart)))
558 sort.Slice(mnts, func(i, j int) bool {
559 if p := mnts[i].priority - mnts[j].priority; p != 0 {
562 return weight[mnts[i]] < weight[mnts[j]]
567 // checkEqual reports whether the data written to it (via io.Writer
568 // interface) is equal to the expected data.
570 // Expect should not be changed after the first Write.
571 type checkEqual struct {
576 func (ce *checkEqual) Equal() bool {
577 return ce.equalUntil == len(ce.Expect)
580 func (ce *checkEqual) Write(p []byte) (int, error) {
581 endpos := ce.equalUntil + len(p)
582 if ce.equalUntil >= 0 && endpos <= len(ce.Expect) && bytes.Equal(p, ce.Expect[ce.equalUntil:endpos]) {
583 ce.equalUntil = endpos
590 func (ks *keepstore) BlockUntrash(ctx context.Context, locator string) error {
591 li, err := parseLocator(locator)
595 var errToCaller error = os.ErrNotExist
596 for _, mnt := range ks.mountsW {
597 if ctx.Err() != nil {
600 err := mnt.BlockUntrash(li.hash)
603 } else if !os.IsNotExist(err) && errToCaller != nil {
610 func (ks *keepstore) BlockTouch(ctx context.Context, locator string) error {
611 li, err := parseLocator(locator)
615 var errToCaller error = os.ErrNotExist
616 for _, mnt := range ks.mountsW {
617 if ctx.Err() != nil {
620 err := mnt.BlockTouch(li.hash)
624 if !os.IsNotExist(err) {
631 func (ks *keepstore) BlockTrash(ctx context.Context, locator string) error {
632 if !ks.cluster.Collections.BlobTrash {
633 return errMethodNotAllowed
635 li, err := parseLocator(locator)
639 var errToCaller error = os.ErrNotExist
640 for _, mnt := range ks.mounts {
644 if ctx.Err() != nil {
647 t, err := mnt.Mtime(li.hash)
648 if err == nil && time.Now().Sub(t) > ks.cluster.Collections.BlobSigningTTL.Duration() {
649 err = mnt.BlockTrash(li.hash)
651 if os.IsNotExist(errToCaller) || (errToCaller == nil && !os.IsNotExist(err)) {
658 func (ks *keepstore) Mounts() []*mount {
662 func (ks *keepstore) Index(ctx context.Context, opts indexOptions) error {
664 if opts.MountUUID != "" {
665 mnt, ok := ks.mounts[opts.MountUUID]
667 return os.ErrNotExist
669 mounts = []*mount{mnt}
671 for _, mnt := range mounts {
672 err := mnt.Index(ctx, opts.Prefix, opts.WriteTo)
680 func ctxToken(ctx context.Context) string {
681 if c, ok := auth.FromContext(ctx); ok && len(c.Tokens) > 0 {
688 type locatorInfo struct {
695 func parseLocator(loc string) (locatorInfo, error) {
697 for i, part := range strings.Split(loc, "+") {
700 return li, errInvalidLocator
706 if size, err := strconv.Atoi(part); err == nil {
712 return li, errInvalidLocator
720 if part[0] >= '0' && part[0] <= '9' {
721 // size, if present at all, must come first
722 return li, errInvalidLocator