You can not select more than 25 topics Topics must start with a letter or number, can include dashes ('-') and can be up to 35 characters long.

238 lines
6.8 KiB

5 months ago
3 years ago
3 years ago
3 years ago
3 years ago
3 years ago
Fix dead lock (#5815) * reduce locks to avoid dead lock Flush->FlushData->uplloadPipeline.FluahAll uploaderCount>0 goroutine 1 [sync.Cond.Wait, 71 minutes]: sync.runtime_notifyListWait(0xc0007ae4d0, 0x0) /usr/local/go/src/runtime/sema.go:569 +0x159 sync.(*Cond).Wait(0xc001a59290?) /usr/local/go/src/sync/cond.go:70 +0x85 github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*UploadPipeline).waitForCurrentWritersToComplete(0xc0002ee4d0) /github/workspace/weed/mount/page_writer/upload_pipeline_lock.go:58 +0x32 github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*UploadPipeline).FlushAll(0xc0002ee4d0) /github/workspace/weed/mount/page_writer/upload_pipeline.go:151 +0x25 github.com/seaweedfs/seaweedfs/weed/mount.(*ChunkedDirtyPages).FlushData(0xc00087e840) /github/workspace/weed/mount/dirty_pages_chunked.go:54 +0x29 github.com/seaweedfs/seaweedfs/weed/mount.(*PageWriter).FlushData(...) /github/workspace/weed/mount/page_writer.go:50 github.com/seaweedfs/seaweedfs/weed/mount.(*WFS).doFlush(0xc0006ad600, 0xc00030d380, 0x0, 0x0) /github/workspace/weed/mount/weedfs_file_sync.go:101 +0x169 github.com/seaweedfs/seaweedfs/weed/mount.(*WFS).Flush(0xc0006ad600, 0xc001a594a8?, 0xc0004c1ca0) /github/workspace/weed/mount/weedfs_file_sync.go:59 +0x48 github.com/hanwen/go-fuse/v2/fuse.doFlush(0xc0000da870?, 0xc0004c1b08) SaveContent -> MemChunk.RLock -> ChunkedDirtyPages.saveChunkedFileIntervalToStorage pages.fh.AddChunks([]*filer_pb.FileChunk{chunk}) fh.entryLock.Lock() sync.(*RWMutex).Lock(0x0?) /usr/local/go/src/sync/rwmutex.go:146 +0x31 github.com/seaweedfs/seaweedfs/weed/mount.(*FileHandle).AddChunks(0xc00030d380, {0xc00028bdc8, 0x1, 0x1}) /github/workspace/weed/mount/filehandle.go:93 +0x45 github.com/seaweedfs/seaweedfs/weed/mount.(*ChunkedDirtyPages).saveChunkedFileIntervalToStorage(0xc00087e840, {0x2be7ac0, 0xc00018d9e0}, 0x0, 0x121, 0x17e3c624565ace45, 0x1?) /github/workspace/weed/mount/dirty_pages_chunked.go:80 +0x2d4 github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*MemChunk).SaveContent(0xc0008d9130, 0xc0008093e0) /github/workspace/weed/mount/page_writer/page_chunk_mem.go:115 +0x112 github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*UploadPipeline).moveToSealed.func1() /github/workspace/weed/mount/page_writer/upload_pipeline.go:187 +0x55 github.com/seaweedfs/seaweedfs/weed/util.(*LimitedConcurrentExecutor).Execute.func1() /github/workspace/weed/util/limited_executor.go:38 +0x62 created by github.com/seaweedfs/seaweedfs/weed/util.(*LimitedConcurrentExecutor).Execute in goroutine 1 /github/workspace/weed/util/limited_executor.go:33 +0x97 On metadata update fh.entryLock.Lock() fh.dirtyPages.Destroy() up.chunksLock.Lock => each sealed chunk.FreeReference => MemChunk.Lock goroutine 134 [sync.RWMutex.Lock, 71 minutes]: sync.runtime_SemacquireRWMutex(0xc0007c3558?, 0xea?, 0x3fb0800?) /usr/local/go/src/runtime/sema.go:87 +0x25 sync.(*RWMutex).Lock(0xc0007c35a8?) /usr/local/go/src/sync/rwmutex.go:151 +0x6a github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*MemChunk).FreeResource(0xc0008d9130) /github/workspace/weed/mount/page_writer/page_chunk_mem.go:38 +0x2a github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*SealedChunk).FreeReference(0xc00071cdb0, {0xc0006ba1a0, 0x20}) /github/workspace/weed/mount/page_writer/upload_pipeline.go:38 +0xb7 github.com/seaweedfs/seaweedfs/weed/mount/page_writer.(*UploadPipeline).Shutdown(0xc0002ee4d0) /github/workspace/weed/mount/page_writer/upload_pipeline.go:220 +0x185 github.com/seaweedfs/seaweedfs/weed/mount.(*ChunkedDirtyPages).Destroy(0xc0008cea40?) /github/workspace/weed/mount/dirty_pages_chunked.go:87 +0x17 github.com/seaweedfs/seaweedfs/weed/mount.(*PageWriter).Destroy(...) /github/workspace/weed/mount/page_writer.go:78 github.com/seaweedfs/seaweedfs/weed/mount.NewSeaweedFileSystem.func3({0xc00069a6c0, 0x30}, 0x6?) /github/workspace/weed/mount/weedfs.go:119 +0x17a github.com/seaweedfs/seaweedfs/weed/mount/meta_cache.NewMetaCache.func1({0xc00069a6c0?, 0xc00069a480?}, 0x4015b40?) /github/workspace/weed/mount/meta_cache/meta_cache.go:37 +0x1c github.com/seaweedfs/seaweedfs/weed/mount/meta_cache.SubscribeMetaEvents.func1(0xc000661810) /github/workspace/weed/mount/meta_cache/meta_cache_subscribe.go:43 +0x570 * use locked entry everywhere * modifiable remote entry * skip locking after getting lock from fhLockTable
5 months ago
3 years ago
  1. package mount
  2. import (
  3. "context"
  4. "github.com/hanwen/go-fuse/v2/fuse"
  5. "github.com/seaweedfs/seaweedfs/weed/filer"
  6. "github.com/seaweedfs/seaweedfs/weed/glog"
  7. "github.com/seaweedfs/seaweedfs/weed/mount/meta_cache"
  8. "github.com/seaweedfs/seaweedfs/weed/util"
  9. "math"
  10. "sync"
  11. )
  12. type DirectoryHandleId uint64
  13. const (
  14. directoryStreamBaseOffset = 2 // . & ..
  15. )
  16. type DirectoryHandle struct {
  17. isFinished bool
  18. entryStream []*filer.Entry
  19. entryStreamOffset uint64
  20. }
  21. func (dh *DirectoryHandle) reset() {
  22. *dh = DirectoryHandle{
  23. isFinished: false,
  24. entryStream: []*filer.Entry{},
  25. entryStreamOffset: directoryStreamBaseOffset,
  26. }
  27. }
  28. type DirectoryHandleToInode struct {
  29. // shares the file handle id sequencer with FileHandleToInode{nextFh}
  30. sync.Mutex
  31. dir2inode map[DirectoryHandleId]*DirectoryHandle
  32. }
  33. func NewDirectoryHandleToInode() *DirectoryHandleToInode {
  34. return &DirectoryHandleToInode{
  35. dir2inode: make(map[DirectoryHandleId]*DirectoryHandle),
  36. }
  37. }
  38. func (wfs *WFS) AcquireDirectoryHandle() (DirectoryHandleId, *DirectoryHandle) {
  39. fh := FileHandleId(util.RandomUint64())
  40. wfs.dhmap.Lock()
  41. defer wfs.dhmap.Unlock()
  42. dh := new(DirectoryHandle)
  43. dh.reset()
  44. wfs.dhmap.dir2inode[DirectoryHandleId(fh)] = dh
  45. return DirectoryHandleId(fh), dh
  46. }
  47. func (wfs *WFS) GetDirectoryHandle(dhid DirectoryHandleId) *DirectoryHandle {
  48. wfs.dhmap.Lock()
  49. defer wfs.dhmap.Unlock()
  50. if dh, found := wfs.dhmap.dir2inode[dhid]; found {
  51. return dh
  52. }
  53. dh := new(DirectoryHandle)
  54. dh.reset()
  55. wfs.dhmap.dir2inode[dhid] = dh
  56. return dh
  57. }
  58. func (wfs *WFS) ReleaseDirectoryHandle(dhid DirectoryHandleId) {
  59. wfs.dhmap.Lock()
  60. defer wfs.dhmap.Unlock()
  61. delete(wfs.dhmap.dir2inode, dhid)
  62. }
  63. // Directory handling
  64. /** Open directory
  65. *
  66. * Unless the 'default_permissions' mount option is given,
  67. * this method should check if opendir is permitted for this
  68. * directory. Optionally opendir may also return an arbitrary
  69. * filehandle in the fuse_file_info structure, which will be
  70. * passed to readdir, releasedir and fsyncdir.
  71. */
  72. func (wfs *WFS) OpenDir(cancel <-chan struct{}, input *fuse.OpenIn, out *fuse.OpenOut) (code fuse.Status) {
  73. if !wfs.inodeToPath.HasInode(input.NodeId) {
  74. return fuse.ENOENT
  75. }
  76. dhid, _ := wfs.AcquireDirectoryHandle()
  77. out.Fh = uint64(dhid)
  78. return fuse.OK
  79. }
  80. /** Release directory
  81. *
  82. * If the directory has been removed after the call to opendir, the
  83. * path parameter will be NULL.
  84. */
  85. func (wfs *WFS) ReleaseDir(input *fuse.ReleaseIn) {
  86. wfs.ReleaseDirectoryHandle(DirectoryHandleId(input.Fh))
  87. }
  88. /** Synchronize directory contents
  89. *
  90. * If the directory has been removed after the call to opendir, the
  91. * path parameter will be NULL.
  92. *
  93. * If the datasync parameter is non-zero, then only the user data
  94. * should be flushed, not the meta data
  95. */
  96. func (wfs *WFS) FsyncDir(cancel <-chan struct{}, input *fuse.FsyncIn) (code fuse.Status) {
  97. return fuse.OK
  98. }
  99. /** Read directory
  100. *
  101. * The filesystem may choose between two modes of operation:
  102. *
  103. * 1) The readdir implementation ignores the offset parameter, and
  104. * passes zero to the filler function's offset. The filler
  105. * function will not return '1' (unless an error happens), so the
  106. * whole directory is read in a single readdir operation.
  107. *
  108. * 2) The readdir implementation keeps track of the offsets of the
  109. * directory entries. It uses the offset parameter and always
  110. * passes non-zero offset to the filler function. When the buffer
  111. * is full (or an error happens) the filler function will return
  112. * '1'.
  113. */
  114. func (wfs *WFS) ReadDir(cancel <-chan struct{}, input *fuse.ReadIn, out *fuse.DirEntryList) (code fuse.Status) {
  115. return wfs.doReadDirectory(input, out, false)
  116. }
  117. func (wfs *WFS) ReadDirPlus(cancel <-chan struct{}, input *fuse.ReadIn, out *fuse.DirEntryList) (code fuse.Status) {
  118. return wfs.doReadDirectory(input, out, true)
  119. }
  120. func (wfs *WFS) doReadDirectory(input *fuse.ReadIn, out *fuse.DirEntryList, isPlusMode bool) fuse.Status {
  121. dh := wfs.GetDirectoryHandle(DirectoryHandleId(input.Fh))
  122. if input.Offset == 0 {
  123. dh.reset()
  124. } else if dh.isFinished && input.Offset >= dh.entryStreamOffset {
  125. entryCurrentIndex := input.Offset - dh.entryStreamOffset
  126. if uint64(len(dh.entryStream)) <= entryCurrentIndex {
  127. return fuse.OK
  128. }
  129. }
  130. isEarlyTerminated := false
  131. dirPath, code := wfs.inodeToPath.GetPath(input.NodeId)
  132. if code != fuse.OK {
  133. return code
  134. }
  135. var dirEntry fuse.DirEntry
  136. processEachEntryFn := func(entry *filer.Entry) bool {
  137. dirEntry.Name = entry.Name()
  138. dirEntry.Mode = toSyscallMode(entry.Mode)
  139. inode := wfs.inodeToPath.Lookup(dirPath.Child(dirEntry.Name), entry.Crtime.Unix(), entry.IsDirectory(), len(entry.HardLinkId) > 0, entry.Inode, isPlusMode)
  140. dirEntry.Ino = inode
  141. if !isPlusMode {
  142. if !out.AddDirEntry(dirEntry) {
  143. isEarlyTerminated = true
  144. return false
  145. }
  146. } else {
  147. entryOut := out.AddDirLookupEntry(dirEntry)
  148. if entryOut == nil {
  149. isEarlyTerminated = true
  150. return false
  151. }
  152. if fh, found := wfs.fhmap.FindFileHandle(inode); found {
  153. glog.V(4).Infof("readdir opened file %s", dirPath.Child(dirEntry.Name))
  154. entry = filer.FromPbEntry(string(dirPath), fh.GetEntry().GetEntry())
  155. }
  156. wfs.outputFilerEntry(entryOut, inode, entry)
  157. }
  158. return true
  159. }
  160. if input.Offset < directoryStreamBaseOffset {
  161. if !isPlusMode {
  162. if input.Offset == 0 {
  163. out.AddDirEntry(fuse.DirEntry{Mode: fuse.S_IFDIR, Name: "."})
  164. }
  165. out.AddDirEntry(fuse.DirEntry{Mode: fuse.S_IFDIR, Name: ".."})
  166. } else {
  167. if input.Offset == 0 {
  168. out.AddDirLookupEntry(fuse.DirEntry{Mode: fuse.S_IFDIR, Name: "."})
  169. }
  170. out.AddDirLookupEntry(fuse.DirEntry{Mode: fuse.S_IFDIR, Name: ".."})
  171. }
  172. input.Offset = directoryStreamBaseOffset
  173. }
  174. var lastEntryName string
  175. if input.Offset >= dh.entryStreamOffset {
  176. if input.Offset > dh.entryStreamOffset {
  177. entryPreviousIndex := (input.Offset - dh.entryStreamOffset) - 1
  178. if uint64(len(dh.entryStream)) > entryPreviousIndex {
  179. lastEntryName = dh.entryStream[entryPreviousIndex].Name()
  180. dh.entryStream = dh.entryStream[entryPreviousIndex:]
  181. dh.entryStreamOffset = input.Offset - 1
  182. }
  183. }
  184. entryCurrentIndex := input.Offset - dh.entryStreamOffset
  185. for uint64(len(dh.entryStream)) > entryCurrentIndex {
  186. entry := dh.entryStream[entryCurrentIndex]
  187. if processEachEntryFn(entry) {
  188. lastEntryName = entry.Name()
  189. entryCurrentIndex++
  190. } else {
  191. // early terminated
  192. return fuse.OK
  193. }
  194. }
  195. }
  196. var err error
  197. if err = meta_cache.EnsureVisited(wfs.metaCache, wfs, dirPath); err != nil {
  198. glog.Errorf("dir ReadDirAll %s: %v", dirPath, err)
  199. return fuse.EIO
  200. }
  201. listErr := wfs.metaCache.ListDirectoryEntries(context.Background(), dirPath, lastEntryName, false, int64(math.MaxInt32), func(entry *filer.Entry) bool {
  202. dh.entryStream = append(dh.entryStream, entry)
  203. return processEachEntryFn(entry)
  204. })
  205. if listErr != nil {
  206. glog.Errorf("list meta cache: %v", listErr)
  207. return fuse.EIO
  208. }
  209. if !isEarlyTerminated {
  210. dh.isFinished = true
  211. }
  212. return fuse.OK
  213. }