1 // A UnixVolume is a Volume backed by a locally mounted disk.
17 // IORequests are encapsulated Get or Put requests. They are used to
18 // implement serialized I/O (i.e. only one read/write operation per
19 // volume). When running in serialized mode, the Keep front end sends
20 // IORequests on a channel to an IORunner, which handles them one at a
21 // time and returns an IOResponse.
26 KeepGet IOMethod = iota
30 type IORequest struct {
34 reply chan *IOResponse
37 type IOResponse struct {
42 // A UnixVolume has the following properties:
45 // the path to the volume's root directory
47 // A channel of IORequests. If non-nil, all I/O requests for
48 // this volume should be queued on this channel; the result
49 // will be delivered on the IOResponse channel supplied in the
52 type UnixVolume struct {
53 root string // path to this volume
57 func (v *UnixVolume) IOHandler() {
58 for req := range v.queue {
62 result.data, result.err = v.Read(req.loc)
64 result.err = v.Write(req.loc, req.data)
70 func MakeUnixVolume(root string, serialize bool) (v UnixVolume) {
72 v = UnixVolume{root, make(chan *IORequest)}
75 v = UnixVolume{root, nil}
80 func (v *UnixVolume) Get(loc string) ([]byte, error) {
84 reply := make(chan *IOResponse)
85 v.queue <- &IORequest{KeepGet, loc, nil, reply}
87 return response.data, response.err
90 func (v *UnixVolume) Put(loc string, block []byte) error {
92 return v.Write(loc, block)
94 reply := make(chan *IOResponse)
95 v.queue <- &IORequest{KeepPut, loc, block, reply}
100 func (v *UnixVolume) Touch(loc string) error {
101 p := v.blockPath(loc)
102 f, err := os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
106 if e := lockfile(f); e != nil {
110 now := time.Now().Unix()
111 utime := syscall.Utimbuf{now, now}
112 return syscall.Utime(p, &utime)
115 func (v *UnixVolume) Mtime(loc string) (time.Time, error) {
116 p := v.blockPath(loc)
117 if fi, err := os.Stat(p); err != nil {
118 return time.Time{}, err
120 return fi.ModTime(), nil
124 // Read retrieves a block identified by the locator string "loc", and
125 // returns its contents as a byte slice.
127 // If the block could not be opened or read, Read returns a nil slice
128 // and the os.Error that was generated.
130 // If the block is present but its content hash does not match loc,
131 // Read returns the block and a CorruptError. It is the caller's
132 // responsibility to decide what (if anything) to do with the
133 // corrupted data block.
135 func (v *UnixVolume) Read(loc string) ([]byte, error) {
136 buf, err := ioutil.ReadFile(v.blockPath(loc))
140 // Write stores a block of data identified by the locator string
141 // "loc". It returns nil on success. If the volume is full, it
142 // returns a FullError. If the write fails due to some other error,
143 // that error is returned.
145 func (v *UnixVolume) Write(loc string, block []byte) error {
149 bdir := v.blockDir(loc)
150 if err := os.MkdirAll(bdir, 0755); err != nil {
151 log.Printf("%s: could not create directory %s: %s",
156 tmpfile, tmperr := ioutil.TempFile(bdir, "tmp"+loc)
158 log.Printf("ioutil.TempFile(%s, tmp%s): %s", bdir, loc, tmperr)
161 bpath := v.blockPath(loc)
163 if _, err := tmpfile.Write(block); err != nil {
164 log.Printf("%s: writing to %s: %s\n", v, bpath, err)
167 if err := tmpfile.Close(); err != nil {
168 log.Printf("closing %s: %s\n", tmpfile.Name(), err)
169 os.Remove(tmpfile.Name())
172 if err := os.Rename(tmpfile.Name(), bpath); err != nil {
173 log.Printf("rename %s %s: %s\n", tmpfile.Name(), bpath, err)
174 os.Remove(tmpfile.Name())
180 // Status returns a VolumeStatus struct describing the volume's
183 func (v *UnixVolume) Status() *VolumeStatus {
184 var fs syscall.Statfs_t
187 if fi, err := os.Stat(v.root); err == nil {
188 devnum = fi.Sys().(*syscall.Stat_t).Dev
190 log.Printf("%s: os.Stat: %s\n", v, err)
194 err := syscall.Statfs(v.root, &fs)
196 log.Printf("%s: statfs: %s\n", v, err)
199 // These calculations match the way df calculates disk usage:
200 // "free" space is measured by fs.Bavail, but "used" space
201 // uses fs.Blocks - fs.Bfree.
202 free := fs.Bavail * uint64(fs.Bsize)
203 used := (fs.Blocks - fs.Bfree) * uint64(fs.Bsize)
204 return &VolumeStatus{v.root, devnum, free, used}
207 // Index returns a list of blocks found on this volume which begin with
208 // the specified prefix. If the prefix is an empty string, Index returns
209 // a complete list of blocks.
211 // The return value is a multiline string (separated by
212 // newlines). Each line is in the format
214 // locator+size modification-time
218 // e4df392f86be161ca6ed3773a962b8f3+67108864 1388894303
219 // e4d41e6fd68460e0e3fc18cc746959d2+67108864 1377796043
220 // e4de7a2810f5554cd39b36d8ddb132ff+67108864 1388701136
222 func (v *UnixVolume) Index(prefix string) (output string) {
223 filepath.Walk(v.root,
224 func(path string, info os.FileInfo, err error) error {
225 // This WalkFunc inspects each path in the volume
226 // and prints an index line for all files that begin
229 log.Printf("IndexHandler: %s: walking to %s: %s",
233 locator := filepath.Base(path)
234 // Skip directories that do not match prefix.
235 // We know there is nothing interesting inside.
237 !strings.HasPrefix(locator, prefix) &&
238 !strings.HasPrefix(prefix, locator) {
239 return filepath.SkipDir
241 // Skip any file that is not apparently a locator, e.g. .meta files
242 if !IsValidLocator(locator) {
245 // Print filenames beginning with prefix
246 if !info.IsDir() && strings.HasPrefix(locator, prefix) {
247 output = output + fmt.Sprintf(
248 "%s+%d %d\n", locator, info.Size(), info.ModTime().Unix())
256 func (v *UnixVolume) Delete(loc string) error {
257 p := v.blockPath(loc)
258 f, err := os.OpenFile(p, os.O_RDWR|os.O_APPEND, 0644)
262 if e := lockfile(f); e != nil {
267 // If the block has been PUT more recently than -permission_ttl,
268 // return success without removing the block. This guards against
269 // a race condition where a block is old enough that Data Manager
270 // has added it to the trash list, but the user submitted a PUT
271 // for the block since then.
272 if fi, err := os.Stat(p); err != nil {
275 if time.Since(fi.ModTime()) < permission_ttl {
282 // blockDir returns the fully qualified directory name for the directory
283 // where loc is (or would be) stored on this volume.
284 func (v *UnixVolume) blockDir(loc string) string {
285 return filepath.Join(v.root, loc[0:3])
288 // blockPath returns the fully qualified pathname for the path to loc
290 func (v *UnixVolume) blockPath(loc string) string {
291 return filepath.Join(v.blockDir(loc), loc)
294 // IsFull returns true if the free space on the volume is less than
295 // MIN_FREE_KILOBYTES.
297 func (v *UnixVolume) IsFull() (isFull bool) {
298 fullSymlink := v.root + "/full"
300 // Check if the volume has been marked as full in the last hour.
301 if link, err := os.Readlink(fullSymlink); err == nil {
302 if ts, err := strconv.Atoi(link); err == nil {
303 fulltime := time.Unix(int64(ts), 0)
304 if time.Since(fulltime).Hours() < 1.0 {
310 if avail, err := v.FreeDiskSpace(); err == nil {
311 isFull = avail < MIN_FREE_KILOBYTES
313 log.Printf("%s: FreeDiskSpace: %s\n", v, err)
317 // If the volume is full, timestamp it.
319 now := fmt.Sprintf("%d", time.Now().Unix())
320 os.Symlink(now, fullSymlink)
325 // FreeDiskSpace returns the number of unused 1k blocks available on
328 func (v *UnixVolume) FreeDiskSpace() (free uint64, err error) {
329 var fs syscall.Statfs_t
330 err = syscall.Statfs(v.root, &fs)
332 // Statfs output is not guaranteed to measure free
333 // space in terms of 1K blocks.
334 free = fs.Bavail * uint64(fs.Bsize) / 1024
339 func (v *UnixVolume) String() string {
340 return fmt.Sprintf("[UnixVolume %s]", v.root)
343 // lockfile and unlockfile use flock(2) to manage kernel file locks.
344 func lockfile(f *os.File) error {
345 return syscall.Flock(int(f.Fd()), syscall.LOCK_EX)
348 func unlockfile(f *os.File) error {
349 return syscall.Flock(int(f.Fd()), syscall.LOCK_UN)