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 // DeviceID returns a globally unique ID for the volume's root
126 // directory, consisting of the filesystem's UUID and the path from
127 // filesystem root to storage directory, joined by "/". For example,
128 // the DeviceID for a local directory "/mnt/xvda1/keep" might be
129 // "fa0b6166-3b55-4994-bd3f-92f4e00a1bb0/keep".
130 func (v *UnixVolume) DeviceID() string {
131 giveup := func(f string, args ...interface{}) string {
132 log.Printf(f+"; using blank DeviceID for volume %s", append(args, v)...)
135 buf, err := exec.Command("findmnt", "--noheadings", "--target", v.Root).CombinedOutput()
137 return giveup("findmnt: %s (%q)", err, buf)
139 findmnt := strings.Fields(string(buf))
140 if len(findmnt) < 2 {
141 return giveup("could not parse findmnt output: %q", buf)
143 fsRoot, dev := findmnt[0], findmnt[1]
145 absRoot, err := filepath.Abs(v.Root)
147 return giveup("resolving relative path %q: %s", v.Root, err)
149 realRoot, err := filepath.EvalSymlinks(absRoot)
151 return giveup("resolving symlinks in %q: %s", absRoot, err)
154 // Find path from filesystem root to realRoot
156 if strings.HasPrefix(realRoot, fsRoot+"/") {
157 fsPath = realRoot[len(fsRoot):]
158 } else if fsRoot == "/" {
160 } else if fsRoot == realRoot {
163 return giveup("findmnt reports mount point %q which is not a prefix of volume root %q", fsRoot, realRoot)
166 if !strings.HasPrefix(dev, "/") {
167 return giveup("mount %q device %q is not a path", fsRoot, dev)
170 fi, err := os.Stat(dev)
172 return giveup("stat %q: %s", dev, err)
174 ino := fi.Sys().(*syscall.Stat_t).Ino
176 // Find a symlink in /dev/disk/by-uuid/ whose target is (i.e.,
177 // has the same inode as) the mounted device
178 udir := "/dev/disk/by-uuid"
179 d, err := os.Open(udir)
181 return giveup("opening %q: %s", udir, err)
183 uuids, err := d.Readdirnames(0)
185 return giveup("reading %q: %s", udir, err)
187 for _, uuid := range uuids {
188 link := filepath.Join(udir, uuid)
189 fi, err = os.Stat(link)
191 log.Printf("error: stat %q: %s", link, err)
194 if fi.Sys().(*syscall.Stat_t).Ino == ino {
198 return giveup("could not find entry in %q matching %q", udir, dev)
201 // Examples implements VolumeWithExamples.
202 func (*UnixVolume) Examples() []Volume {
205 Root: "/mnt/local-disk",
207 DirectoryReplication: 1,
210 Root: "/mnt/network-disk",
212 DirectoryReplication: 2,
217 // Type implements Volume
218 func (v *UnixVolume) Type() string {
222 // Start implements Volume
223 func (v *UnixVolume) Start(vm *volumeMetricsVecs) error {
225 v.locker = &sync.Mutex{}
227 if !strings.HasPrefix(v.Root, "/") {
228 return fmt.Errorf("volume root does not start with '/': %q", v.Root)
230 if v.DirectoryReplication == 0 {
231 v.DirectoryReplication = 1
233 // Set up prometheus metrics
234 lbls := prometheus.Labels{"device_id": v.DeviceID()}
235 v.os.stats.opsCounters, v.os.stats.errCounters, v.os.stats.ioBytes = vm.getCounterVecsFor(lbls)
237 _, err := v.os.Stat(v.Root)
242 // Touch sets the timestamp for the given locator to the current time
243 func (v *UnixVolume) Touch(loc string) error {
245 return MethodDisabledError
247 p := v.blockPath(loc)
248 f, err := v.os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
253 if err := v.lock(context.TODO()); err != nil {
257 if e := v.lockfile(f); e != nil {
260 defer v.unlockfile(f)
261 ts := syscall.NsecToTimespec(time.Now().UnixNano())
262 v.os.stats.TickOps("utimes")
263 v.os.stats.Tick(&v.os.stats.UtimesOps)
264 err = syscall.UtimesNano(p, []syscall.Timespec{ts, ts})
265 v.os.stats.TickErr(err)
269 // Mtime returns the stored timestamp for the given locator.
270 func (v *UnixVolume) Mtime(loc string) (time.Time, error) {
271 p := v.blockPath(loc)
272 fi, err := v.os.Stat(p)
274 return time.Time{}, err
276 return fi.ModTime(), nil
279 // Lock the locker (if one is in use), open the file for reading, and
280 // call the given function if and when the file is ready to read.
281 func (v *UnixVolume) getFunc(ctx context.Context, path string, fn func(io.Reader) error) error {
282 if err := v.lock(ctx); err != nil {
286 f, err := v.os.Open(path)
291 return fn(NewCountingReader(ioutil.NopCloser(f), v.os.stats.TickInBytes))
294 // stat is os.Stat() with some extra sanity checks.
295 func (v *UnixVolume) stat(path string) (os.FileInfo, error) {
296 stat, err := v.os.Stat(path)
300 } else if stat.Size() > BlockSize {
307 // Get retrieves a block, copies it to the given slice, and returns
308 // the number of bytes copied.
309 func (v *UnixVolume) Get(ctx context.Context, loc string, buf []byte) (int, error) {
310 return getWithPipe(ctx, loc, buf, v)
313 // ReadBlock implements BlockReader.
314 func (v *UnixVolume) ReadBlock(ctx context.Context, loc string, w io.Writer) error {
315 path := v.blockPath(loc)
316 stat, err := v.stat(path)
318 return v.translateError(err)
320 return v.getFunc(ctx, path, func(rdr io.Reader) error {
321 n, err := io.Copy(w, rdr)
322 if err == nil && n != stat.Size() {
323 err = io.ErrUnexpectedEOF
329 // Compare returns nil if Get(loc) would return the same content as
330 // expect. It is functionally equivalent to Get() followed by
331 // bytes.Compare(), but uses less memory.
332 func (v *UnixVolume) Compare(ctx context.Context, loc string, expect []byte) error {
333 path := v.blockPath(loc)
334 if _, err := v.stat(path); err != nil {
335 return v.translateError(err)
337 return v.getFunc(ctx, path, func(rdr io.Reader) error {
338 return compareReaderWithBuf(ctx, rdr, expect, loc[:32])
342 // Put stores a block of data identified by the locator string
343 // "loc". It returns nil on success. If the volume is full, it
344 // returns a FullError. If the write fails due to some other error,
345 // that error is returned.
346 func (v *UnixVolume) Put(ctx context.Context, loc string, block []byte) error {
347 return putWithPipe(ctx, loc, block, v)
350 // WriteBlock implements BlockWriter.
351 func (v *UnixVolume) WriteBlock(ctx context.Context, loc string, rdr io.Reader) error {
353 return MethodDisabledError
358 bdir := v.blockDir(loc)
359 if err := os.MkdirAll(bdir, 0755); err != nil {
360 log.Printf("%s: could not create directory %s: %s",
365 tmpfile, tmperr := v.os.TempFile(bdir, "tmp"+loc)
367 log.Printf("ioutil.TempFile(%s, tmp%s): %s", bdir, loc, tmperr)
371 bpath := v.blockPath(loc)
373 if err := v.lock(ctx); err != nil {
377 n, err := io.Copy(tmpfile, rdr)
378 v.os.stats.TickOutBytes(uint64(n))
380 log.Printf("%s: writing to %s: %s", v, bpath, err)
382 v.os.Remove(tmpfile.Name())
385 if err := tmpfile.Close(); err != nil {
386 log.Printf("closing %s: %s", tmpfile.Name(), err)
387 v.os.Remove(tmpfile.Name())
390 if err := v.os.Rename(tmpfile.Name(), bpath); err != nil {
391 log.Printf("rename %s %s: %s", tmpfile.Name(), bpath, err)
392 return v.os.Remove(tmpfile.Name())
397 // Status returns a VolumeStatus struct describing the volume's
398 // current state, or nil if an error occurs.
400 func (v *UnixVolume) Status() *VolumeStatus {
401 fi, err := v.os.Stat(v.Root)
403 log.Printf("%s: os.Stat: %s", v, err)
406 devnum := fi.Sys().(*syscall.Stat_t).Dev
408 var fs syscall.Statfs_t
409 if err := syscall.Statfs(v.Root, &fs); err != nil {
410 log.Printf("%s: statfs: %s", v, err)
413 // These calculations match the way df calculates disk usage:
414 // "free" space is measured by fs.Bavail, but "used" space
415 // uses fs.Blocks - fs.Bfree.
416 free := fs.Bavail * uint64(fs.Bsize)
417 used := (fs.Blocks - fs.Bfree) * uint64(fs.Bsize)
418 return &VolumeStatus{
426 var blockDirRe = regexp.MustCompile(`^[0-9a-f]+$`)
427 var blockFileRe = regexp.MustCompile(`^[0-9a-f]{32}$`)
429 // IndexTo writes (to the given Writer) a list of blocks found on this
430 // volume which begin with the specified prefix. If the prefix is an
431 // empty string, IndexTo writes a complete list of blocks.
433 // Each block is given in the format
435 // locator+size modification-time {newline}
439 // e4df392f86be161ca6ed3773a962b8f3+67108864 1388894303
440 // e4d41e6fd68460e0e3fc18cc746959d2+67108864 1377796043
441 // e4de7a2810f5554cd39b36d8ddb132ff+67108864 1388701136
443 func (v *UnixVolume) IndexTo(prefix string, w io.Writer) error {
445 rootdir, err := v.os.Open(v.Root)
449 defer rootdir.Close()
450 v.os.stats.TickOps("readdir")
451 v.os.stats.Tick(&v.os.stats.ReaddirOps)
453 names, err := rootdir.Readdirnames(1)
456 } else if err != nil {
459 if !strings.HasPrefix(names[0], prefix) && !strings.HasPrefix(prefix, names[0]) {
460 // prefix excludes all blocks stored in this dir
463 if !blockDirRe.MatchString(names[0]) {
466 blockdirpath := filepath.Join(v.Root, names[0])
467 blockdir, err := v.os.Open(blockdirpath)
469 log.Print("Error reading ", blockdirpath, ": ", err)
473 v.os.stats.TickOps("readdir")
474 v.os.stats.Tick(&v.os.stats.ReaddirOps)
476 fileInfo, err := blockdir.Readdir(1)
479 } else if err != nil {
480 log.Print("Error reading ", blockdirpath, ": ", err)
484 name := fileInfo[0].Name()
485 if !strings.HasPrefix(name, prefix) {
488 if !blockFileRe.MatchString(name) {
491 _, err = fmt.Fprint(w,
493 "+", fileInfo[0].Size(),
494 " ", fileInfo[0].ModTime().UnixNano(),
497 log.Print("Error writing : ", err)
506 // Trash trashes the block data from the unix storage
507 // If TrashLifetime == 0, the block is deleted
508 // Else, the block is renamed as path/{loc}.trash.{deadline},
509 // where deadline = now + TrashLifetime
510 func (v *UnixVolume) Trash(loc string) error {
511 // Touch() must be called before calling Write() on a block. Touch()
512 // also uses lockfile(). This avoids a race condition between Write()
513 // and Trash() because either (a) the file will be trashed and Touch()
514 // will signal to the caller that the file is not present (and needs to
515 // be re-written), or (b) Touch() will update the file's timestamp and
516 // Trash() will read the correct up-to-date timestamp and choose not to
520 return MethodDisabledError
522 if err := v.lock(context.TODO()); err != nil {
526 p := v.blockPath(loc)
527 f, err := v.os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
532 if e := v.lockfile(f); e != nil {
535 defer v.unlockfile(f)
537 // If the block has been PUT in the last blobSignatureTTL
538 // seconds, return success without removing the block. This
539 // protects data from garbage collection until it is no longer
540 // possible for clients to retrieve the unreferenced blocks
541 // anyway (because the permission signatures have expired).
542 if fi, err := v.os.Stat(p); err != nil {
544 } else if time.Since(fi.ModTime()) < time.Duration(theConfig.BlobSignatureTTL) {
548 if theConfig.TrashLifetime == 0 {
549 return v.os.Remove(p)
551 return v.os.Rename(p, fmt.Sprintf("%v.trash.%d", p, time.Now().Add(theConfig.TrashLifetime.Duration()).Unix()))
554 // Untrash moves block from trash back into store
555 // Look for path/{loc}.trash.{deadline} in storage,
556 // and rename the first such file as path/{loc}
557 func (v *UnixVolume) Untrash(loc string) (err error) {
559 return MethodDisabledError
562 v.os.stats.TickOps("readdir")
563 v.os.stats.Tick(&v.os.stats.ReaddirOps)
564 files, err := ioutil.ReadDir(v.blockDir(loc))
570 return os.ErrNotExist
574 prefix := fmt.Sprintf("%v.trash.", loc)
575 for _, f := range files {
576 if strings.HasPrefix(f.Name(), prefix) {
578 err = v.os.Rename(v.blockPath(f.Name()), v.blockPath(loc))
585 if foundTrash == false {
586 return os.ErrNotExist
592 // blockDir returns the fully qualified directory name for the directory
593 // where loc is (or would be) stored on this volume.
594 func (v *UnixVolume) blockDir(loc string) string {
595 return filepath.Join(v.Root, loc[0:3])
598 // blockPath returns the fully qualified pathname for the path to loc
600 func (v *UnixVolume) blockPath(loc string) string {
601 return filepath.Join(v.blockDir(loc), loc)
604 // IsFull returns true if the free space on the volume is less than
607 func (v *UnixVolume) IsFull() (isFull bool) {
608 fullSymlink := v.Root + "/full"
610 // Check if the volume has been marked as full in the last hour.
611 if link, err := os.Readlink(fullSymlink); err == nil {
612 if ts, err := strconv.Atoi(link); err == nil {
613 fulltime := time.Unix(int64(ts), 0)
614 if time.Since(fulltime).Hours() < 1.0 {
620 if avail, err := v.FreeDiskSpace(); err == nil {
621 isFull = avail < MinFreeKilobytes
623 log.Printf("%s: FreeDiskSpace: %s", v, err)
627 // If the volume is full, timestamp it.
629 now := fmt.Sprintf("%d", time.Now().Unix())
630 os.Symlink(now, fullSymlink)
635 // FreeDiskSpace returns the number of unused 1k blocks available on
638 func (v *UnixVolume) FreeDiskSpace() (free uint64, err error) {
639 var fs syscall.Statfs_t
640 err = syscall.Statfs(v.Root, &fs)
642 // Statfs output is not guaranteed to measure free
643 // space in terms of 1K blocks.
644 free = fs.Bavail * uint64(fs.Bsize) / 1024
649 func (v *UnixVolume) String() string {
650 return fmt.Sprintf("[UnixVolume %s]", v.Root)
653 // Writable returns false if all future Put, Mtime, and Delete calls
654 // are expected to fail.
655 func (v *UnixVolume) Writable() bool {
659 // Replication returns the number of replicas promised by the
660 // underlying device (as specified in configuration).
661 func (v *UnixVolume) Replication() int {
662 return v.DirectoryReplication
665 // GetStorageClasses implements Volume
666 func (v *UnixVolume) GetStorageClasses() []string {
667 return v.StorageClasses
670 // InternalStats returns I/O and filesystem ops counters.
671 func (v *UnixVolume) InternalStats() interface{} {
675 // lock acquires the serialize lock, if one is in use. If ctx is done
676 // before the lock is acquired, lock returns ctx.Err() instead of
677 // acquiring the lock.
678 func (v *UnixVolume) lock(ctx context.Context) error {
683 locked := make(chan struct{})
690 log.Printf("%s: client hung up while waiting for Serialize lock (%s)", v, time.Since(t0))
701 // unlock releases the serialize lock, if one is in use.
702 func (v *UnixVolume) unlock() {
709 // lockfile and unlockfile use flock(2) to manage kernel file locks.
710 func (v *UnixVolume) lockfile(f *os.File) error {
711 v.os.stats.TickOps("flock")
712 v.os.stats.Tick(&v.os.stats.FlockOps)
713 err := syscall.Flock(int(f.Fd()), syscall.LOCK_EX)
714 v.os.stats.TickErr(err)
718 func (v *UnixVolume) unlockfile(f *os.File) error {
719 err := syscall.Flock(int(f.Fd()), syscall.LOCK_UN)
720 v.os.stats.TickErr(err)
724 // Where appropriate, translate a more specific filesystem error to an
725 // error recognized by handlers, like os.ErrNotExist.
726 func (v *UnixVolume) translateError(err error) error {
729 // stat() returns a PathError if the parent directory
730 // (not just the file itself) is missing
731 return os.ErrNotExist
737 var unixTrashLocRegexp = regexp.MustCompile(`/([0-9a-f]{32})\.trash\.(\d+)$`)
739 // EmptyTrash walks hierarchy looking for {hash}.trash.*
740 // and deletes those with deadline < now.
741 func (v *UnixVolume) EmptyTrash() {
742 var bytesDeleted, bytesInTrash int64
743 var blocksDeleted, blocksInTrash int64
745 doFile := func(path string, info os.FileInfo) {
746 if info.Mode().IsDir() {
749 matches := unixTrashLocRegexp.FindStringSubmatch(path)
750 if len(matches) != 3 {
753 deadline, err := strconv.ParseInt(matches[2], 10, 64)
755 log.Printf("EmptyTrash: %v: ParseInt(%v): %v", path, matches[2], err)
758 atomic.AddInt64(&bytesInTrash, info.Size())
759 atomic.AddInt64(&blocksInTrash, 1)
760 if deadline > time.Now().Unix() {
763 err = v.os.Remove(path)
765 log.Printf("EmptyTrash: Remove %v: %v", path, err)
768 atomic.AddInt64(&bytesDeleted, info.Size())
769 atomic.AddInt64(&blocksDeleted, 1)
776 var wg sync.WaitGroup
777 todo := make(chan dirent, theConfig.EmptyTrashWorkers)
778 for i := 0; i < 1 || i < theConfig.EmptyTrashWorkers; i++ {
782 for e := range todo {
783 doFile(e.path, e.info)
788 err := filepath.Walk(v.Root, func(path string, info os.FileInfo, err error) error {
790 log.Printf("EmptyTrash: filepath.Walk: %v: %v", path, err)
793 todo <- dirent{path, info}
800 log.Printf("EmptyTrash error for %v: %v", v.String(), err)
803 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)
806 type unixStats struct {
818 func (s *unixStats) TickErr(err error) {
822 s.statsTicker.TickErr(err, fmt.Sprintf("%T", err))
825 type osWithStats struct {
829 func (o *osWithStats) Open(name string) (*os.File, error) {
830 o.stats.TickOps("open")
831 o.stats.Tick(&o.stats.OpenOps)
832 f, err := os.Open(name)
837 func (o *osWithStats) OpenFile(name string, flag int, perm os.FileMode) (*os.File, error) {
838 o.stats.TickOps("open")
839 o.stats.Tick(&o.stats.OpenOps)
840 f, err := os.OpenFile(name, flag, perm)
845 func (o *osWithStats) Remove(path string) error {
846 o.stats.TickOps("unlink")
847 o.stats.Tick(&o.stats.UnlinkOps)
848 err := os.Remove(path)
853 func (o *osWithStats) Rename(a, b string) error {
854 o.stats.TickOps("rename")
855 o.stats.Tick(&o.stats.RenameOps)
856 err := os.Rename(a, b)
861 func (o *osWithStats) Stat(path string) (os.FileInfo, error) {
862 o.stats.TickOps("stat")
863 o.stats.Tick(&o.stats.StatOps)
864 fi, err := os.Stat(path)
869 func (o *osWithStats) TempFile(dir, base string) (*os.File, error) {
870 o.stats.TickOps("create")
871 o.stats.Tick(&o.stats.CreateOps)
872 f, err := ioutil.TempFile(dir, base)