X-Git-Url: https://git.arvados.org/arvados.git/blobdiff_plain/7930d7abaabf2fd1f3432eca10f26b821e0ef94f..f935965259f9c0476a9c5ffa79e5c27ce9da4800:/services/keepstore/volume_unix.go diff --git a/services/keepstore/volume_unix.go b/services/keepstore/volume_unix.go index 2ffa8faa39..92c897eac0 100644 --- a/services/keepstore/volume_unix.go +++ b/services/keepstore/volume_unix.go @@ -1,9 +1,8 @@ -// A UnixVolume is a Volume backed by a locally mounted disk. -// package main import ( - "bytes" + "bufio" + "flag" "fmt" "io" "io/ioutil" @@ -18,16 +17,133 @@ import ( "time" ) +type unixVolumeAdder struct { + *Config +} + +// String implements flag.Value +func (s *unixVolumeAdder) String() string { + return "-" +} + +func (vs *unixVolumeAdder) Set(path string) error { + if dirs := strings.Split(path, ","); len(dirs) > 1 { + log.Print("DEPRECATED: using comma-separated volume list.") + for _, dir := range dirs { + if err := vs.Set(dir); err != nil { + return err + } + } + return nil + } + vs.Config.Volumes = append(vs.Config.Volumes, &UnixVolume{ + Root: path, + ReadOnly: deprecated.flagReadonly, + Serialize: deprecated.flagSerializeIO, + }) + return nil +} + +func init() { + VolumeTypes = append(VolumeTypes, func() VolumeWithExamples { return &UnixVolume{} }) + + flag.Var(&unixVolumeAdder{theConfig}, "volumes", "see Volumes configuration") + flag.Var(&unixVolumeAdder{theConfig}, "volume", "see Volumes configuration") +} + +// Discover adds a UnixVolume for every directory named "keep" that is +// located at the top level of a device- or tmpfs-backed mount point +// other than "/". It returns the number of volumes added. +func (vs *unixVolumeAdder) Discover() int { + added := 0 + f, err := os.Open(ProcMounts) + if err != nil { + log.Fatalf("opening %s: %s", ProcMounts, err) + } + scanner := bufio.NewScanner(f) + for scanner.Scan() { + args := strings.Fields(scanner.Text()) + if err := scanner.Err(); err != nil { + log.Fatalf("reading %s: %s", ProcMounts, err) + } + dev, mount := args[0], args[1] + if mount == "/" { + continue + } + if dev != "tmpfs" && !strings.HasPrefix(dev, "/dev/") { + continue + } + keepdir := mount + "/keep" + if st, err := os.Stat(keepdir); err != nil || !st.IsDir() { + continue + } + // Set the -readonly flag (but only for this volume) + // if the filesystem is mounted readonly. + flagReadonlyWas := deprecated.flagReadonly + for _, fsopt := range strings.Split(args[3], ",") { + if fsopt == "ro" { + deprecated.flagReadonly = true + break + } + if fsopt == "rw" { + break + } + } + if err := vs.Set(keepdir); err != nil { + log.Printf("adding %q: %s", keepdir, err) + } else { + added++ + } + deprecated.flagReadonly = flagReadonlyWas + } + return added +} + // A UnixVolume stores and retrieves blocks in a local directory. type UnixVolume struct { - root string // path to the volume's root directory - serialize bool - readonly bool - mutex sync.Mutex + Root string // path to the volume's root directory + ReadOnly bool + Serialize bool + + // something to lock during IO, typically a sync.Mutex (or nil + // to skip locking) + locker sync.Locker +} + +// Examples implements VolumeWithExamples. +func (*UnixVolume) Examples() []Volume { + return []Volume{ + &UnixVolume{ + Root: "/mnt/local-disk", + Serialize: true, + }, + &UnixVolume{ + Root: "/mnt/network-disk", + Serialize: false, + }, + } } +// Type implements Volume +func (v *UnixVolume) Type() string { + return "Directory" +} + +// Start implements Volume +func (v *UnixVolume) Start() error { + if v.Serialize { + v.locker = &sync.Mutex{} + } + if !strings.HasPrefix(v.Root, "/") { + return fmt.Errorf("volume root does not start with '/': %q", v.Root) + } + _, err := os.Stat(v.Root) + return err +} + +// Touch sets the timestamp for the given locator to the current time func (v *UnixVolume) Touch(loc string) error { - if v.readonly { + if v.ReadOnly { return MethodDisabledError } p := v.blockPath(loc) @@ -36,40 +152,40 @@ func (v *UnixVolume) Touch(loc string) error { return err } defer f.Close() - if v.serialize { - v.mutex.Lock() - defer v.mutex.Unlock() + if v.locker != nil { + v.locker.Lock() + defer v.locker.Unlock() } if e := lockfile(f); e != nil { return e } defer unlockfile(f) - now := time.Now().Unix() - utime := syscall.Utimbuf{now, now} - return syscall.Utime(p, &utime) + ts := syscall.NsecToTimespec(time.Now().UnixNano()) + return syscall.UtimesNano(p, []syscall.Timespec{ts, ts}) } +// Mtime returns the stored timestamp for the given locator. func (v *UnixVolume) Mtime(loc string) (time.Time, error) { p := v.blockPath(loc) - if fi, err := os.Stat(p); err != nil { + fi, err := os.Stat(p) + if err != nil { return time.Time{}, err - } else { - return fi.ModTime(), nil } + return fi.ModTime(), nil } -// Open the given file, apply the serialize lock if enabled, and call -// the given function if and when the file is ready to read. +// Lock the locker (if one is in use), open the file for reading, and +// call the given function if and when the file is ready to read. func (v *UnixVolume) getFunc(path string, fn func(io.Reader) error) error { + if v.locker != nil { + v.locker.Lock() + defer v.locker.Unlock() + } f, err := os.Open(path) if err != nil { return err } defer f.Close() - if v.serialize { - v.mutex.Lock() - defer v.mutex.Unlock() - } return fn(f) } @@ -79,78 +195,43 @@ func (v *UnixVolume) stat(path string) (os.FileInfo, error) { if err == nil { if stat.Size() < 0 { err = os.ErrInvalid - } else if stat.Size() > BLOCKSIZE { + } else if stat.Size() > BlockSize { err = TooLongError } } return stat, err } -// Get retrieves a block identified by the locator string "loc", and -// returns its contents as a byte slice. -// -// Get returns a nil buffer IFF it returns a non-nil error. -func (v *UnixVolume) Get(loc string) ([]byte, error) { +// Get retrieves a block, copies it to the given slice, and returns +// the number of bytes copied. +func (v *UnixVolume) Get(loc string, buf []byte) (int, error) { path := v.blockPath(loc) stat, err := v.stat(path) if err != nil { - return nil, err + return 0, v.translateError(err) + } + if stat.Size() > int64(len(buf)) { + return 0, TooLongError } - buf := bufs.Get(int(stat.Size())) + var read int + size := int(stat.Size()) err = v.getFunc(path, func(rdr io.Reader) error { - _, err = io.ReadFull(rdr, buf) + read, err = io.ReadFull(rdr, buf[:size]) return err }) - if err != nil { - bufs.Put(buf) - return nil, err - } - return buf, nil + return read, err } // Compare returns nil if Get(loc) would return the same content as -// cmp. It is functionally equivalent to Get() followed by +// expect. It is functionally equivalent to Get() followed by // bytes.Compare(), but uses less memory. -// -// TODO(TC): Before returning CollisionError, compute the MD5 digest -// of the data on disk (i.e., known-to-be-equal data in cmp + -// remaining data on disk) and return DiskHashError instead of -// CollisionError if it doesn't equal loc[:32]. -func (v *UnixVolume) Compare(loc string, cmp []byte) error { +func (v *UnixVolume) Compare(loc string, expect []byte) error { path := v.blockPath(loc) - stat, err := v.stat(path) - if err != nil { - return err + if _, err := v.stat(path); err != nil { + return v.translateError(err) } - bufLen := 1 << 20 - if int64(bufLen) > stat.Size() { - bufLen = int(stat.Size()) - } - buf := make([]byte, bufLen) return v.getFunc(path, func(rdr io.Reader) error { - // Loop invariants: all data read so far matched what - // we expected, and the first N bytes of cmp are - // expected to equal the next N bytes read from - // reader. - for { - n, err := rdr.Read(buf) - if n > len(cmp) { - // file on disk is too long - return CollisionError - } else if n > 0 && bytes.Compare(cmp[:n], buf[:n]) != 0 { - return CollisionError - } - cmp = cmp[n:] - if err == io.EOF { - if len(cmp) != 0 { - // file on disk is too short - return CollisionError - } - return nil - } else if err != nil { - return err - } - } + return compareReaderWithBuf(rdr, expect, loc[:32]) }) } @@ -159,7 +240,7 @@ func (v *UnixVolume) Compare(loc string, cmp []byte) error { // returns a FullError. If the write fails due to some other error, // that error is returned. func (v *UnixVolume) Put(loc string, block []byte) error { - if v.readonly { + if v.ReadOnly { return MethodDisabledError } if v.IsFull() { @@ -179,9 +260,9 @@ func (v *UnixVolume) Put(loc string, block []byte) error { } bpath := v.blockPath(loc) - if v.serialize { - v.mutex.Lock() - defer v.mutex.Unlock() + if v.locker != nil { + v.locker.Lock() + defer v.locker.Unlock() } if _, err := tmpfile.Write(block); err != nil { log.Printf("%s: writing to %s: %s\n", v, bpath, err) @@ -209,14 +290,14 @@ func (v *UnixVolume) Status() *VolumeStatus { var fs syscall.Statfs_t var devnum uint64 - if fi, err := os.Stat(v.root); err == nil { + if fi, err := os.Stat(v.Root); err == nil { devnum = fi.Sys().(*syscall.Stat_t).Dev } else { log.Printf("%s: os.Stat: %s\n", v, err) return nil } - err := syscall.Statfs(v.root, &fs) + err := syscall.Statfs(v.Root, &fs) if err != nil { log.Printf("%s: statfs: %s\n", v, err) return nil @@ -226,10 +307,11 @@ func (v *UnixVolume) Status() *VolumeStatus { // uses fs.Blocks - fs.Bfree. free := fs.Bavail * uint64(fs.Bsize) used := (fs.Blocks - fs.Bfree) * uint64(fs.Bsize) - return &VolumeStatus{v.root, devnum, free, used} + return &VolumeStatus{v.Root, devnum, free, used} } var blockDirRe = regexp.MustCompile(`^[0-9a-f]+$`) +var blockFileRe = regexp.MustCompile(`^[0-9a-f]{32}$`) // IndexTo writes (to the given Writer) a list of blocks found on this // volume which begin with the specified prefix. If the prefix is an @@ -246,8 +328,8 @@ var blockDirRe = regexp.MustCompile(`^[0-9a-f]+$`) // e4de7a2810f5554cd39b36d8ddb132ff+67108864 1388701136 // func (v *UnixVolume) IndexTo(prefix string, w io.Writer) error { - var lastErr error = nil - rootdir, err := os.Open(v.root) + var lastErr error + rootdir, err := os.Open(v.Root) if err != nil { return err } @@ -266,7 +348,7 @@ func (v *UnixVolume) IndexTo(prefix string, w io.Writer) error { if !blockDirRe.MatchString(names[0]) { continue } - blockdirpath := filepath.Join(v.root, names[0]) + blockdirpath := filepath.Join(v.Root, names[0]) blockdir, err := os.Open(blockdirpath) if err != nil { log.Print("Error reading ", blockdirpath, ": ", err) @@ -286,31 +368,38 @@ func (v *UnixVolume) IndexTo(prefix string, w io.Writer) error { if !strings.HasPrefix(name, prefix) { continue } + if !blockFileRe.MatchString(name) { + continue + } _, err = fmt.Fprint(w, name, "+", fileInfo[0].Size(), - " ", fileInfo[0].ModTime().Unix(), + " ", fileInfo[0].ModTime().UnixNano(), "\n") } blockdir.Close() } } -func (v *UnixVolume) Delete(loc string) error { +// Trash trashes the block data from the unix storage +// If TrashLifetime == 0, the block is deleted +// Else, the block is renamed as path/{loc}.trash.{deadline}, +// where deadline = now + TrashLifetime +func (v *UnixVolume) Trash(loc string) error { // Touch() must be called before calling Write() on a block. Touch() // also uses lockfile(). This avoids a race condition between Write() - // and Delete() because either (a) the file will be deleted and Touch() + // and Trash() because either (a) the file will be trashed and Touch() // will signal to the caller that the file is not present (and needs to // be re-written), or (b) Touch() will update the file's timestamp and - // Delete() will read the correct up-to-date timestamp and choose not to - // delete the file. + // Trash() will read the correct up-to-date timestamp and choose not to + // trash the file. - if v.readonly { + if v.ReadOnly { return MethodDisabledError } - if v.serialize { - v.mutex.Lock() - defer v.mutex.Unlock() + if v.locker != nil { + v.locker.Lock() + defer v.locker.Unlock() } p := v.blockPath(loc) f, err := os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644) @@ -323,25 +412,63 @@ func (v *UnixVolume) Delete(loc string) error { } defer unlockfile(f) - // If the block has been PUT in the last blob_signature_ttl + // If the block has been PUT in the last blobSignatureTTL // seconds, return success without removing the block. This // protects data from garbage collection until it is no longer // possible for clients to retrieve the unreferenced blocks // anyway (because the permission signatures have expired). if fi, err := os.Stat(p); err != nil { return err - } else { - if time.Since(fi.ModTime()) < blob_signature_ttl { - return nil + } else if time.Since(fi.ModTime()) < time.Duration(theConfig.BlobSignatureTTL) { + return nil + } + + if theConfig.TrashLifetime == 0 { + return os.Remove(p) + } + return os.Rename(p, fmt.Sprintf("%v.trash.%d", p, time.Now().Add(theConfig.TrashLifetime.Duration()).Unix())) +} + +// Untrash moves block from trash back into store +// Look for path/{loc}.trash.{deadline} in storage, +// and rename the first such file as path/{loc} +func (v *UnixVolume) Untrash(loc string) (err error) { + if v.ReadOnly { + return MethodDisabledError + } + + files, err := ioutil.ReadDir(v.blockDir(loc)) + if err != nil { + return err + } + + if len(files) == 0 { + return os.ErrNotExist + } + + foundTrash := false + prefix := fmt.Sprintf("%v.trash.", loc) + for _, f := range files { + if strings.HasPrefix(f.Name(), prefix) { + foundTrash = true + err = os.Rename(v.blockPath(f.Name()), v.blockPath(loc)) + if err == nil { + break + } } } - return os.Remove(p) + + if foundTrash == false { + return os.ErrNotExist + } + + return } // blockDir returns the fully qualified directory name for the directory // where loc is (or would be) stored on this volume. func (v *UnixVolume) blockDir(loc string) string { - return filepath.Join(v.root, loc[0:3]) + return filepath.Join(v.Root, loc[0:3]) } // blockPath returns the fully qualified pathname for the path to loc @@ -351,10 +478,10 @@ func (v *UnixVolume) blockPath(loc string) string { } // IsFull returns true if the free space on the volume is less than -// MIN_FREE_KILOBYTES. +// MinFreeKilobytes. // func (v *UnixVolume) IsFull() (isFull bool) { - fullSymlink := v.root + "/full" + fullSymlink := v.Root + "/full" // Check if the volume has been marked as full in the last hour. if link, err := os.Readlink(fullSymlink); err == nil { @@ -367,7 +494,7 @@ func (v *UnixVolume) IsFull() (isFull bool) { } if avail, err := v.FreeDiskSpace(); err == nil { - isFull = avail < MIN_FREE_KILOBYTES + isFull = avail < MinFreeKilobytes } else { log.Printf("%s: FreeDiskSpace: %s\n", v, err) isFull = false @@ -386,7 +513,7 @@ func (v *UnixVolume) IsFull() (isFull bool) { // func (v *UnixVolume) FreeDiskSpace() (free uint64, err error) { var fs syscall.Statfs_t - err = syscall.Statfs(v.root, &fs) + err = syscall.Statfs(v.Root, &fs) if err == nil { // Statfs output is not guaranteed to measure free // space in terms of 1K blocks. @@ -396,11 +523,19 @@ func (v *UnixVolume) FreeDiskSpace() (free uint64, err error) { } func (v *UnixVolume) String() string { - return fmt.Sprintf("[UnixVolume %s]", v.root) + return fmt.Sprintf("[UnixVolume %s]", v.Root) } +// Writable returns false if all future Put, Mtime, and Delete calls +// are expected to fail. func (v *UnixVolume) Writable() bool { - return !v.readonly + return !v.ReadOnly +} + +// Replication returns the number of replicas promised by the +// underlying device (currently assumed to be 1). +func (v *UnixVolume) Replication() int { + return 1 } // lockfile and unlockfile use flock(2) to manage kernel file locks. @@ -411,3 +546,63 @@ func lockfile(f *os.File) error { func unlockfile(f *os.File) error { return syscall.Flock(int(f.Fd()), syscall.LOCK_UN) } + +// Where appropriate, translate a more specific filesystem error to an +// error recognized by handlers, like os.ErrNotExist. +func (v *UnixVolume) translateError(err error) error { + switch err.(type) { + case *os.PathError: + // stat() returns a PathError if the parent directory + // (not just the file itself) is missing + return os.ErrNotExist + default: + return err + } +} + +var unixTrashLocRegexp = regexp.MustCompile(`/([0-9a-f]{32})\.trash\.(\d+)$`) + +// EmptyTrash walks hierarchy looking for {hash}.trash.* +// and deletes those with deadline < now. +func (v *UnixVolume) EmptyTrash() { + var bytesDeleted, bytesInTrash int64 + var blocksDeleted, blocksInTrash int + + err := filepath.Walk(v.Root, func(path string, info os.FileInfo, err error) error { + if err != nil { + log.Printf("EmptyTrash: filepath.Walk: %v: %v", path, err) + return nil + } + if info.Mode().IsDir() { + return nil + } + matches := unixTrashLocRegexp.FindStringSubmatch(path) + if len(matches) != 3 { + return nil + } + deadline, err := strconv.ParseInt(matches[2], 10, 64) + if err != nil { + log.Printf("EmptyTrash: %v: ParseInt(%v): %v", path, matches[2], err) + return nil + } + bytesInTrash += info.Size() + blocksInTrash++ + if deadline > time.Now().Unix() { + return nil + } + err = os.Remove(path) + if err != nil { + log.Printf("EmptyTrash: Remove %v: %v", path, err) + return nil + } + bytesDeleted += info.Size() + blocksDeleted++ + return nil + }) + + if err != nil { + log.Printf("EmptyTrash error for %v: %v", v.String(), err) + } + + 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) +}