1 // Copyright (C) The Arvados Authors. All rights reserved.
3 // SPDX-License-Identifier: AGPL-3.0
25 "github.com/prometheus/client_golang/prometheus"
28 type unixVolumeAdder struct {
32 // String implements flag.Value
33 func (vs *unixVolumeAdder) String() string {
37 func (vs *unixVolumeAdder) Set(path string) error {
38 if dirs := strings.Split(path, ","); len(dirs) > 1 {
39 log.Print("DEPRECATED: using comma-separated volume list.")
40 for _, dir := range dirs {
41 if err := vs.Set(dir); err != nil {
47 vs.Config.Volumes = append(vs.Config.Volumes, &UnixVolume{
49 ReadOnly: deprecated.flagReadonly,
50 Serialize: deprecated.flagSerializeIO,
56 VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &UnixVolume{} })
58 flag.Var(&unixVolumeAdder{theConfig}, "volumes", "see Volumes configuration")
59 flag.Var(&unixVolumeAdder{theConfig}, "volume", "see Volumes configuration")
62 // Discover adds a UnixVolume for every directory named "keep" that is
63 // located at the top level of a device- or tmpfs-backed mount point
64 // other than "/". It returns the number of volumes added.
65 func (vs *unixVolumeAdder) Discover() int {
67 f, err := os.Open(ProcMounts)
69 log.Fatalf("opening %s: %s", ProcMounts, err)
71 scanner := bufio.NewScanner(f)
73 args := strings.Fields(scanner.Text())
74 if err := scanner.Err(); err != nil {
75 log.Fatalf("reading %s: %s", ProcMounts, err)
77 dev, mount := args[0], args[1]
81 if dev != "tmpfs" && !strings.HasPrefix(dev, "/dev/") {
84 keepdir := mount + "/keep"
85 if st, err := os.Stat(keepdir); err != nil || !st.IsDir() {
88 // Set the -readonly flag (but only for this volume)
89 // if the filesystem is mounted readonly.
90 flagReadonlyWas := deprecated.flagReadonly
91 for _, fsopt := range strings.Split(args[3], ",") {
93 deprecated.flagReadonly = true
100 if err := vs.Set(keepdir); err != nil {
101 log.Printf("adding %q: %s", keepdir, err)
105 deprecated.flagReadonly = flagReadonlyWas
110 // A UnixVolume stores and retrieves blocks in a local directory.
111 type UnixVolume struct {
112 Root string // path to the volume's root directory
115 DirectoryReplication int
116 StorageClasses []string
118 // something to lock during IO, typically a sync.Mutex (or nil
125 opsCounters *prometheus.CounterVec
126 errCounters *prometheus.CounterVec
127 ioBytes *prometheus.CounterVec
130 // DeviceID returns a globally unique ID for the volume's root
131 // directory, consisting of the filesystem's UUID and the path from
132 // filesystem root to storage directory, joined by "/". For example,
133 // the DeviceID for a local directory "/mnt/xvda1/keep" might be
134 // "fa0b6166-3b55-4994-bd3f-92f4e00a1bb0/keep".
135 func (v *UnixVolume) DeviceID() string {
136 giveup := func(f string, args ...interface{}) string {
137 log.Printf(f+"; using blank DeviceID for volume %s", append(args, v)...)
140 buf, err := exec.Command("findmnt", "--noheadings", "--target", v.Root).CombinedOutput()
142 return giveup("findmnt: %s (%q)", err, buf)
144 findmnt := strings.Fields(string(buf))
145 if len(findmnt) < 2 {
146 return giveup("could not parse findmnt output: %q", buf)
148 fsRoot, dev := findmnt[0], findmnt[1]
150 absRoot, err := filepath.Abs(v.Root)
152 return giveup("resolving relative path %q: %s", v.Root, err)
154 realRoot, err := filepath.EvalSymlinks(absRoot)
156 return giveup("resolving symlinks in %q: %s", absRoot, err)
159 // Find path from filesystem root to realRoot
161 if strings.HasPrefix(realRoot, fsRoot+"/") {
162 fsPath = realRoot[len(fsRoot):]
163 } else if fsRoot == "/" {
165 } else if fsRoot == realRoot {
168 return giveup("findmnt reports mount point %q which is not a prefix of volume root %q", fsRoot, realRoot)
171 if !strings.HasPrefix(dev, "/") {
172 return giveup("mount %q device %q is not a path", fsRoot, dev)
175 fi, err := os.Stat(dev)
177 return giveup("stat %q: %s\n", dev, err)
179 ino := fi.Sys().(*syscall.Stat_t).Ino
181 // Find a symlink in /dev/disk/by-uuid/ whose target is (i.e.,
182 // has the same inode as) the mounted device
183 udir := "/dev/disk/by-uuid"
184 d, err := os.Open(udir)
186 return giveup("opening %q: %s", udir, err)
188 uuids, err := d.Readdirnames(0)
190 return giveup("reading %q: %s", udir, err)
192 for _, uuid := range uuids {
193 link := filepath.Join(udir, uuid)
194 fi, err = os.Stat(link)
196 log.Printf("error: stat %q: %s", link, err)
199 if fi.Sys().(*syscall.Stat_t).Ino == ino {
203 return giveup("could not find entry in %q matching %q", udir, dev)
206 // Examples implements VolumeWithExamples.
207 func (*UnixVolume) Examples() []Volume {
210 Root: "/mnt/local-disk",
212 DirectoryReplication: 1,
215 Root: "/mnt/network-disk",
217 DirectoryReplication: 2,
222 // Type implements Volume
223 func (v *UnixVolume) Type() string {
227 // Start implements Volume
228 func (v *UnixVolume) Start(opsCounters, errCounters, ioBytes *prometheus.CounterVec) error {
230 v.locker = &sync.Mutex{}
232 if !strings.HasPrefix(v.Root, "/") {
233 return fmt.Errorf("volume root does not start with '/': %q", v.Root)
235 if v.DirectoryReplication == 0 {
236 v.DirectoryReplication = 1
238 // Set up prometheus metrics
239 lbls := prometheus.Labels{"device_id": v.DeviceID()}
240 v.opsCounters = opsCounters.MustCurryWith(lbls)
241 v.errCounters = errCounters.MustCurryWith(lbls)
242 v.ioBytes = ioBytes.MustCurryWith(lbls)
243 v.os.promSetup(v.opsCounters, v.errCounters, v.ioBytes)
245 _, err := v.os.Stat(v.Root)
250 // Touch sets the timestamp for the given locator to the current time
251 func (v *UnixVolume) Touch(loc string) error {
253 return MethodDisabledError
255 p := v.blockPath(loc)
256 f, err := v.os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
261 if err := v.lock(context.TODO()); err != nil {
265 if e := v.lockfile(f); e != nil {
268 defer v.unlockfile(f)
269 ts := syscall.NsecToTimespec(time.Now().UnixNano())
270 if v.os.opsCounters != nil {
271 v.os.opsCounters.With(prometheus.Labels{"operation": "utimes"}).Inc()
273 v.os.stats.Tick(&v.os.stats.UtimesOps)
274 err = syscall.UtimesNano(p, []syscall.Timespec{ts, ts})
276 v.os.stats.TickErr(err)
280 // Mtime returns the stored timestamp for the given locator.
281 func (v *UnixVolume) Mtime(loc string) (time.Time, error) {
282 p := v.blockPath(loc)
283 fi, err := v.os.Stat(p)
285 return time.Time{}, err
287 return fi.ModTime(), nil
290 // Lock the locker (if one is in use), open the file for reading, and
291 // call the given function if and when the file is ready to read.
292 func (v *UnixVolume) getFunc(ctx context.Context, path string, fn func(io.Reader) error) error {
293 if err := v.lock(ctx); err != nil {
297 f, err := v.os.Open(path)
302 return fn(NewCountingReader(
305 v.os.stats.TickInBytes(c)
306 if v.ioBytes != nil {
307 v.ioBytes.With(prometheus.Labels{"direction": "in"}).Add(float64(c))
312 // stat is os.Stat() with some extra sanity checks.
313 func (v *UnixVolume) stat(path string) (os.FileInfo, error) {
314 stat, err := v.os.Stat(path)
318 } else if stat.Size() > BlockSize {
325 // Get retrieves a block, copies it to the given slice, and returns
326 // the number of bytes copied.
327 func (v *UnixVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
328 return getWithPipe(ctx, loc, buf, v)
331 // ReadBlock implements BlockReader.
332 func (v *UnixVolume) ReadBlock(ctx context.Context, loc string, w io.Writer) error {
333 path := v.blockPath(loc)
334 stat, err := v.stat(path)
336 return v.translateError(err)
338 return v.getFunc(ctx, path, func(rdr io.Reader) error {
339 n, err := io.Copy(w, rdr)
340 if err == nil && n != stat.Size() {
341 err = io.ErrUnexpectedEOF
347 // Compare returns nil if Get(loc) would return the same content as
348 // expect. It is functionally equivalent to Get() followed by
349 // bytes.Compare(), but uses less memory.
350 func (v *UnixVolume) Compare(ctx context.Context, loc string, expect []byte) error {
351 path := v.blockPath(loc)
352 if _, err := v.stat(path); err != nil {
353 return v.translateError(err)
355 return v.getFunc(ctx, path, func(rdr io.Reader) error {
356 return compareReaderWithBuf(ctx, rdr, expect, loc[:32])
360 // Put stores a block of data identified by the locator string
361 // "loc". It returns nil on success. If the volume is full, it
362 // returns a FullError. If the write fails due to some other error,
363 // that error is returned.
364 func (v *UnixVolume) Put(ctx context.Context, loc string, block []byte) error {
365 return putWithPipe(ctx, loc, block, v)
368 // WriteBlock implements BlockWriter.
369 func (v *UnixVolume) WriteBlock(ctx context.Context, loc string, rdr io.Reader) error {
371 return MethodDisabledError
376 bdir := v.blockDir(loc)
377 if err := os.MkdirAll(bdir, 0755); err != nil {
378 log.Printf("%s: could not create directory %s: %s",
383 tmpfile, tmperr := v.os.TempFile(bdir, "tmp"+loc)
385 log.Printf("ioutil.TempFile(%s, tmp%s): %s", bdir, loc, tmperr)
389 bpath := v.blockPath(loc)
391 if err := v.lock(ctx); err != nil {
395 n, err := io.Copy(tmpfile, rdr)
396 if v.ioBytes != nil {
397 v.ioBytes.With(prometheus.Labels{"direction": "out"}).Add(float64(n))
399 v.os.stats.TickOutBytes(uint64(n))
401 log.Printf("%s: writing to %s: %s\n", v, bpath, err)
403 v.os.Remove(tmpfile.Name())
406 if err := tmpfile.Close(); err != nil {
407 log.Printf("closing %s: %s\n", tmpfile.Name(), err)
408 v.os.Remove(tmpfile.Name())
411 if err := v.os.Rename(tmpfile.Name(), bpath); err != nil {
412 log.Printf("rename %s %s: %s\n", tmpfile.Name(), bpath, err)
413 return v.os.Remove(tmpfile.Name())
418 // Status returns a VolumeStatus struct describing the volume's
419 // current state, or nil if an error occurs.
421 func (v *UnixVolume) Status() *VolumeStatus {
422 fi, err := v.os.Stat(v.Root)
424 log.Printf("%s: os.Stat: %s\n", v, err)
427 devnum := fi.Sys().(*syscall.Stat_t).Dev
429 var fs syscall.Statfs_t
430 if err := syscall.Statfs(v.Root, &fs); err != nil {
431 log.Printf("%s: statfs: %s\n", v, err)
434 // These calculations match the way df calculates disk usage:
435 // "free" space is measured by fs.Bavail, but "used" space
436 // uses fs.Blocks - fs.Bfree.
437 free := fs.Bavail * uint64(fs.Bsize)
438 used := (fs.Blocks - fs.Bfree) * uint64(fs.Bsize)
439 return &VolumeStatus{
447 var blockDirRe = regexp.MustCompile(`^[0-9a-f]+$`)
448 var blockFileRe = regexp.MustCompile(`^[0-9a-f]{32}$`)
450 // IndexTo writes (to the given Writer) a list of blocks found on this
451 // volume which begin with the specified prefix. If the prefix is an
452 // empty string, IndexTo writes a complete list of blocks.
454 // Each block is given in the format
456 // locator+size modification-time {newline}
460 // e4df392f86be161ca6ed3773a962b8f3+67108864 1388894303
461 // e4d41e6fd68460e0e3fc18cc746959d2+67108864 1377796043
462 // e4de7a2810f5554cd39b36d8ddb132ff+67108864 1388701136
464 func (v *UnixVolume) IndexTo(prefix string, w io.Writer) error {
466 rootdir, err := v.os.Open(v.Root)
470 defer rootdir.Close()
471 if v.opsCounters != nil {
472 v.opsCounters.With(prometheus.Labels{"operation": "readdir"}).Inc()
474 v.os.stats.Tick(&v.os.stats.ReaddirOps)
476 names, err := rootdir.Readdirnames(1)
479 } else if err != nil {
482 if !strings.HasPrefix(names[0], prefix) && !strings.HasPrefix(prefix, names[0]) {
483 // prefix excludes all blocks stored in this dir
486 if !blockDirRe.MatchString(names[0]) {
489 blockdirpath := filepath.Join(v.Root, names[0])
490 blockdir, err := v.os.Open(blockdirpath)
492 log.Print("Error reading ", blockdirpath, ": ", err)
496 if v.opsCounters != nil {
497 v.opsCounters.With(prometheus.Labels{"operation": "readdir"}).Inc()
499 v.os.stats.Tick(&v.os.stats.ReaddirOps)
501 fileInfo, err := blockdir.Readdir(1)
504 } else if err != nil {
505 log.Print("Error reading ", blockdirpath, ": ", err)
509 name := fileInfo[0].Name()
510 if !strings.HasPrefix(name, prefix) {
513 if !blockFileRe.MatchString(name) {
516 _, err = fmt.Fprint(w,
518 "+", fileInfo[0].Size(),
519 " ", fileInfo[0].ModTime().UnixNano(),
522 log.Print("Error writing : ", err)
531 // Trash trashes the block data from the unix storage
532 // If TrashLifetime == 0, the block is deleted
533 // Else, the block is renamed as path/{loc}.trash.{deadline},
534 // where deadline = now + TrashLifetime
535 func (v *UnixVolume) Trash(loc string) error {
536 // Touch() must be called before calling Write() on a block. Touch()
537 // also uses lockfile(). This avoids a race condition between Write()
538 // and Trash() because either (a) the file will be trashed and Touch()
539 // will signal to the caller that the file is not present (and needs to
540 // be re-written), or (b) Touch() will update the file's timestamp and
541 // Trash() will read the correct up-to-date timestamp and choose not to
545 return MethodDisabledError
547 if err := v.lock(context.TODO()); err != nil {
551 p := v.blockPath(loc)
552 f, err := v.os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
557 if e := v.lockfile(f); e != nil {
560 defer v.unlockfile(f)
562 // If the block has been PUT in the last blobSignatureTTL
563 // seconds, return success without removing the block. This
564 // protects data from garbage collection until it is no longer
565 // possible for clients to retrieve the unreferenced blocks
566 // anyway (because the permission signatures have expired).
567 if fi, err := v.os.Stat(p); err != nil {
569 } else if time.Since(fi.ModTime()) < time.Duration(theConfig.BlobSignatureTTL) {
573 if theConfig.TrashLifetime == 0 {
574 return v.os.Remove(p)
576 return v.os.Rename(p, fmt.Sprintf("%v.trash.%d", p, time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()))
579 // Untrash moves block from trash back into store
580 // Look for path/{loc}.trash.{deadline} in storage,
581 // and rename the first such file as path/{loc}
582 func (v *UnixVolume) Untrash(loc string) (err error) {
584 return MethodDisabledError
587 if v.opsCounters != nil {
588 v.opsCounters.With(prometheus.Labels{"operation": "readdir"}).Inc()
590 v.os.stats.Tick(&v.os.stats.ReaddirOps)
591 files, err := ioutil.ReadDir(v.blockDir(loc))
597 return os.ErrNotExist
601 prefix := fmt.Sprintf("%v.trash.", loc)
602 for _, f := range files {
603 if strings.HasPrefix(f.Name(), prefix) {
605 err = v.os.Rename(v.blockPath(f.Name()), v.blockPath(loc))
612 if foundTrash == false {
613 return os.ErrNotExist
619 // blockDir returns the fully qualified directory name for the directory
620 // where loc is (or would be) stored on this volume.
621 func (v *UnixVolume) blockDir(loc string) string {
622 return filepath.Join(v.Root, loc[0:3])
625 // blockPath returns the fully qualified pathname for the path to loc
627 func (v *UnixVolume) blockPath(loc string) string {
628 return filepath.Join(v.blockDir(loc), loc)
631 // IsFull returns true if the free space on the volume is less than
634 func (v *UnixVolume) IsFull() (isFull bool) {
635 fullSymlink := v.Root + "/full"
637 // Check if the volume has been marked as full in the last hour.
638 if link, err := os.Readlink(fullSymlink); err == nil {
639 if ts, err := strconv.Atoi(link); err == nil {
640 fulltime := time.Unix(int64(ts), 0)
641 if time.Since(fulltime).Hours() < 1.0 {
647 if avail, err := v.FreeDiskSpace(); err == nil {
648 isFull = avail < MinFreeKilobytes
650 log.Printf("%s: FreeDiskSpace: %s\n", v, err)
654 // If the volume is full, timestamp it.
656 now := fmt.Sprintf("%d", time.Now().Unix())
657 os.Symlink(now, fullSymlink)
662 // FreeDiskSpace returns the number of unused 1k blocks available on
665 func (v *UnixVolume) FreeDiskSpace() (free uint64, err error) {
666 var fs syscall.Statfs_t
667 err = syscall.Statfs(v.Root, &fs)
669 // Statfs output is not guaranteed to measure free
670 // space in terms of 1K blocks.
671 free = fs.Bavail * uint64(fs.Bsize) / 1024
676 func (v *UnixVolume) String() string {
677 return fmt.Sprintf("[UnixVolume %s]", v.Root)
680 // Writable returns false if all future Put, Mtime, and Delete calls
681 // are expected to fail.
682 func (v *UnixVolume) Writable() bool {
686 // Replication returns the number of replicas promised by the
687 // underlying device (as specified in configuration).
688 func (v *UnixVolume) Replication() int {
689 return v.DirectoryReplication
692 // GetStorageClasses implements Volume
693 func (v *UnixVolume) GetStorageClasses() []string {
694 return v.StorageClasses
697 // InternalStats returns I/O and filesystem ops counters.
698 func (v *UnixVolume) InternalStats() interface{} {
702 // lock acquires the serialize lock, if one is in use. If ctx is done
703 // before the lock is acquired, lock returns ctx.Err() instead of
704 // acquiring the lock.
705 func (v *UnixVolume) lock(ctx context.Context) error {
709 locked := make(chan struct{})
726 // unlock releases the serialize lock, if one is in use.
727 func (v *UnixVolume) unlock() {
734 // lockfile and unlockfile use flock(2) to manage kernel file locks.
735 func (v *UnixVolume) lockfile(f *os.File) error {
736 if v.opsCounters != nil {
737 v.opsCounters.With(prometheus.Labels{"operation": "flock"}).Inc()
739 v.os.stats.Tick(&v.os.stats.FlockOps)
740 err := syscall.Flock(int(f.Fd()), syscall.LOCK_EX)
742 v.os.stats.TickErr(err)
746 func (v *UnixVolume) unlockfile(f *os.File) error {
747 err := syscall.Flock(int(f.Fd()), syscall.LOCK_UN)
749 v.os.stats.TickErr(err)
753 // Where appropriate, translate a more specific filesystem error to an
754 // error recognized by handlers, like os.ErrNotExist.
755 func (v *UnixVolume) translateError(err error) error {
758 // stat() returns a PathError if the parent directory
759 // (not just the file itself) is missing
760 return os.ErrNotExist
766 var unixTrashLocRegexp = regexp.MustCompile(`/([0-9a-f]{32})\.trash\.(\d+)$`)
768 // EmptyTrash walks hierarchy looking for {hash}.trash.*
769 // and deletes those with deadline < now.
770 func (v *UnixVolume) EmptyTrash() {
771 var bytesDeleted, bytesInTrash int64
772 var blocksDeleted, blocksInTrash int64
774 doFile := func(path string, info os.FileInfo) {
775 if info.Mode().IsDir() {
778 matches := unixTrashLocRegexp.FindStringSubmatch(path)
779 if len(matches) != 3 {
782 deadline, err := strconv.ParseInt(matches[2], 10, 64)
784 log.Printf("EmptyTrash: %v: ParseInt(%v): %v", path, matches[2], err)
787 atomic.AddInt64(&bytesInTrash, info.Size())
788 atomic.AddInt64(&blocksInTrash, 1)
789 if deadline > time.Now().Unix() {
792 err = v.os.Remove(path)
794 log.Printf("EmptyTrash: Remove %v: %v", path, err)
797 atomic.AddInt64(&bytesDeleted, info.Size())
798 atomic.AddInt64(&blocksDeleted, 1)
805 var wg sync.WaitGroup
806 todo := make(chan dirent, theConfig.EmptyTrashWorkers)
807 for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
811 for e := range todo {
812 doFile(e.path, e.info)
817 err := filepath.Walk(v.Root, func(path string, info os.FileInfo, err error) error {
819 log.Printf("EmptyTrash: filepath.Walk: %v: %v", path, err)
822 todo <- dirent{path, info}
829 log.Printf("EmptyTrash error for %v: %v", v.String(), err)
832 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)
835 type unixStats struct {
847 func (s *unixStats) TickErr(err error) {
851 s.statsTicker.TickErr(err, fmt.Sprintf("%T", err))
854 type osWithStats struct {
856 opsCounters *prometheus.CounterVec
857 errCounters *prometheus.CounterVec
858 ioBytes *prometheus.CounterVec
861 func (o *osWithStats) tickErr(err error) {
862 if err == nil || o.errCounters == nil {
865 o.errCounters.With(prometheus.Labels{"error_type": fmt.Sprintf("%T", err)}).Inc()
868 func (o *osWithStats) promSetup(opsC, errC, ioB *prometheus.CounterVec) {
874 func (o *osWithStats) Open(name string) (*os.File, error) {
875 if o.opsCounters != nil {
876 o.opsCounters.With(prometheus.Labels{"operation": "open"}).Inc()
878 o.stats.Tick(&o.stats.OpenOps)
879 f, err := os.Open(name)
885 func (o *osWithStats) OpenFile(name string, flag int, perm os.FileMode) (*os.File, error) {
886 if o.opsCounters != nil {
887 o.opsCounters.With(prometheus.Labels{"operation": "open"}).Inc()
889 o.stats.Tick(&o.stats.OpenOps)
890 f, err := os.OpenFile(name, flag, perm)
896 func (o *osWithStats) Remove(path string) error {
897 if o.opsCounters != nil {
898 o.opsCounters.With(prometheus.Labels{"operation": "unlink"}).Inc()
900 o.stats.Tick(&o.stats.UnlinkOps)
901 err := os.Remove(path)
907 func (o *osWithStats) Rename(a, b string) error {
908 if o.opsCounters != nil {
909 o.opsCounters.With(prometheus.Labels{"operation": "rename"}).Inc()
911 o.stats.Tick(&o.stats.RenameOps)
912 err := os.Rename(a, b)
918 func (o *osWithStats) Stat(path string) (os.FileInfo, error) {
919 if o.opsCounters != nil {
920 o.opsCounters.With(prometheus.Labels{"operation": "stat"}).Inc()
922 o.stats.Tick(&o.stats.StatOps)
923 fi, err := os.Stat(path)
929 func (o *osWithStats) TempFile(dir, base string) (*os.File, error) {
930 if o.opsCounters != nil {
931 o.opsCounters.With(prometheus.Labels{"operation": "create"}).Inc()
933 o.stats.Tick(&o.stats.CreateOps)
934 f, err := ioutil.TempFile(dir, base)