1 """FUSE driver for Arvados Keep
5 There is one `Operations` object per mount point. It is the entry point for all
6 read and write requests from the llfuse module.
8 The operations object owns an `Inodes` object. The inodes object stores the
9 mapping from numeric inode (used throughout the file system API to uniquely
10 identify files) to the Python objects that implement files and directories.
12 The `Inodes` object owns an `InodeCache` object. The inode cache records the
13 memory footprint of file system objects and when they are last used. When the
14 cache limit is exceeded, the least recently used objects are cleared.
16 File system objects inherit from `fresh.FreshBase` which manages the object lifecycle.
18 File objects inherit from `fusefile.File`. Key methods are `readfrom` and `writeto`
19 which implement actual reads and writes.
21 Directory objects inherit from `fusedir.Directory`. The directory object wraps
22 a Python dict which stores the mapping from filenames to directory entries.
23 Directory contents can be accessed through the Python operators such as `[]`
24 and `in`. These methods automatically check if the directory is fresh (up to
25 date) or stale (needs update) and will call `update` if necessary before
28 The general FUSE operation flow is as follows:
30 - The request handler is called with either an inode or file handle that is the
31 subject of the operation.
33 - Look up the inode using the Inodes table or the file handle in the
34 filehandles table to get the file system object.
36 - For methods that alter files or directories, check that the operation is
37 valid and permitted using _check_writable().
39 - Call the relevant method on the file system object.
43 The FUSE driver supports the Arvados event bus. When an event is received for
44 an object that is live in the inode cache, that object is immediately updated.
73 # Default _notify_queue has a limit of 1000 items, but it really needs to be
74 # unlimited to avoid deadlocks, see https://arvados.org/issues/3198#note-43 for
77 llfuse.capi._notify_queue = Queue.Queue()
79 from fusedir import sanitize_filename, Directory, CollectionDirectory, MagicDirectory, TagsDirectory, ProjectDirectory, SharedDirectory, CollectionDirectoryBase
80 from fusefile import StringFile, FuseArvadosFile
82 _logger = logging.getLogger('arvados.arvados_fuse')
84 # Uncomment this to enable llfuse debug logging.
85 # log_handler = logging.StreamHandler()
86 # llogger = logging.getLogger('llfuse')
87 # llogger.addHandler(log_handler)
88 # llogger.setLevel(logging.DEBUG)
91 """Connects a numeric file handle to a File or Directory object that has
92 been opened by the client."""
94 def __init__(self, fh, obj):
103 if self.obj.writable():
104 return self.obj.flush()
107 class FileHandle(Handle):
108 """Connects a numeric file handle to a File object that has
109 been opened by the client."""
113 class DirectoryHandle(Handle):
114 """Connects a numeric file handle to a Directory object that has
115 been opened by the client."""
117 def __init__(self, fh, dirobj, entries):
118 super(DirectoryHandle, self).__init__(fh, dirobj)
119 self.entries = entries
122 class InodeCache(object):
123 """Records the memory footprint of objects and when they are last used.
125 When the cache limit is exceeded, the least recently used objects are
126 cleared. Clearing the object means discarding its contents to release
127 memory. The next time the object is accessed, it must be re-fetched from
128 the server. Note that the inode cache limit is a soft limit; the cache
129 limit may be exceeded if necessary to load very large objects, it may also
130 be exceeded if open file handles prevent objects from being cleared.
134 def __init__(self, cap, min_entries=4):
135 self._entries = collections.OrderedDict()
137 self._counter = itertools.count(0)
140 self.min_entries = min_entries
145 def _remove(self, obj, clear):
146 if clear and not obj.clear():
147 _logger.debug("InodeCache could not clear %i in_use %s", obj.inode, obj.in_use())
149 self._total -= obj.cache_size
150 del self._entries[obj.cache_priority]
152 self._by_uuid[obj.cache_uuid].remove(obj)
153 if not self._by_uuid[obj.cache_uuid]:
154 del self._by_uuid[obj.cache_uuid]
155 obj.cache_uuid = None
157 _logger.debug("InodeCache cleared %i total now %i", obj.inode, self._total)
161 if self._total > self.cap:
162 for key in list(self._entries.keys()):
163 if self._total < self.cap or len(self._entries) < self.min_entries:
165 self._remove(self._entries[key], True)
167 def manage(self, obj):
169 obj.cache_priority = next(self._counter)
170 obj.cache_size = obj.objsize()
171 self._entries[obj.cache_priority] = obj
172 obj.cache_uuid = obj.uuid()
174 if obj.cache_uuid not in self._by_uuid:
175 self._by_uuid[obj.cache_uuid] = [obj]
177 if obj not in self._by_uuid[obj.cache_uuid]:
178 self._by_uuid[obj.cache_uuid].append(obj)
179 self._total += obj.objsize()
180 _logger.debug("InodeCache touched %i (size %i) (uuid %s) total now %i", obj.inode, obj.objsize(), obj.cache_uuid, self._total)
183 obj.cache_priority = None
185 def touch(self, obj):
187 if obj.cache_priority in self._entries:
188 self._remove(obj, False)
191 def unmanage(self, obj):
192 if obj.persisted() and obj.cache_priority in self._entries:
193 self._remove(obj, True)
195 def find(self, uuid):
196 return self._by_uuid.get(uuid)
199 self._entries.clear()
200 self._by_uuid.clear()
203 class Inodes(object):
204 """Manage the set of inodes. This is the mapping from a numeric id
205 to a concrete File or Directory object"""
207 def __init__(self, inode_cache, encoding="utf-8"):
209 self._counter = itertools.count(llfuse.ROOT_INODE)
210 self.inode_cache = inode_cache
211 self.encoding = encoding
212 self.deferred_invalidations = []
214 def __getitem__(self, item):
215 return self._entries[item]
217 def __setitem__(self, key, item):
218 self._entries[key] = item
221 return self._entries.iterkeys()
224 return self._entries.items()
226 def __contains__(self, k):
227 return k in self._entries
229 def touch(self, entry):
230 entry._atime = time.time()
231 self.inode_cache.touch(entry)
233 def add_entry(self, entry):
234 entry.inode = next(self._counter)
235 if entry.inode == llfuse.ROOT_INODE:
237 self._entries[entry.inode] = entry
238 self.inode_cache.manage(entry)
241 def del_entry(self, entry):
242 if entry.ref_count == 0:
243 self.inode_cache.unmanage(entry)
244 del self._entries[entry.inode]
245 with llfuse.lock_released:
247 self.invalidate_inode(entry.inode)
251 _logger.debug("del_entry on inode %i with refcount %i", entry.inode, entry.ref_count)
253 def invalidate_inode(self, inode):
254 llfuse.invalidate_inode(inode)
256 def invalidate_entry(self, inode, name):
257 llfuse.invalidate_entry(inode, name)
260 self.inode_cache.clear()
262 for k,v in self._entries.items():
265 except Exception as e:
266 _logger.exception("Error during finalize of inode %i", k)
268 self._entries.clear()
271 def catch_exceptions(orig_func):
272 """Catch uncaught exceptions and log them consistently."""
274 @functools.wraps(orig_func)
275 def catch_exceptions_wrapper(self, *args, **kwargs):
277 return orig_func(self, *args, **kwargs)
278 except llfuse.FUSEError:
280 except EnvironmentError as e:
281 raise llfuse.FUSEError(e.errno)
282 except arvados.errors.KeepWriteError as e:
283 _logger.error("Keep write error: " + str(e))
284 raise llfuse.FUSEError(errno.EIO)
285 except arvados.errors.NotFoundError as e:
286 _logger.error("Block not found error: " + str(e))
287 raise llfuse.FUSEError(errno.EIO)
289 _logger.exception("Unhandled exception during FUSE operation")
290 raise llfuse.FUSEError(errno.EIO)
292 return catch_exceptions_wrapper
295 class Operations(llfuse.Operations):
296 """This is the main interface with llfuse.
298 The methods on this object are called by llfuse threads to service FUSE
299 events to query and read from the file system.
301 llfuse has its own global lock which is acquired before calling a request handler,
302 so request handlers do not run concurrently unless the lock is explicitly released
303 using 'with llfuse.lock_released:'
307 def __init__(self, uid, gid, encoding="utf-8", inode_cache=None, num_retries=4, enable_write=False):
308 super(Operations, self).__init__()
311 inode_cache = InodeCache(cap=256*1024*1024)
312 self.inodes = Inodes(inode_cache, encoding=encoding)
315 self.enable_write = enable_write
317 # dict of inode to filehandle
318 self._filehandles = {}
319 self._filehandles_counter = itertools.count(0)
321 # Other threads that need to wait until the fuse driver
322 # is fully initialized should wait() on this event object.
323 self.initlock = threading.Event()
325 self.num_retries = num_retries
327 self.read_counter = arvados.keep.Counter()
328 self.write_counter = arvados.keep.Counter()
333 # Allow threads that are waiting for the driver to be finished
334 # initializing to continue
345 def access(self, inode, mode, ctx):
348 def listen_for_events(self, api_client):
349 self.events = arvados.events.subscribe(api_client,
350 [["event_type", "in", ["create", "update", "delete"]]],
354 def on_event(self, ev):
355 if 'event_type' in ev:
357 items = self.inodes.inode_cache.find(ev["object_uuid"])
358 if items is not None:
361 if ev["object_kind"] == "arvados#collection":
362 new_attr = ev.get("properties") and ev["properties"].get("new_attributes") and ev["properties"]["new_attributes"]
364 # new_attributes.modified_at currently lacks subsecond precision (see #6347) so use event_at which
365 # should always be the same.
366 #record_version = (new_attr["modified_at"], new_attr["portable_data_hash"]) if new_attr else None
367 record_version = (ev["event_at"], new_attr["portable_data_hash"]) if new_attr else None
369 item.update(to_record_version=record_version)
373 oldowner = ev.get("properties") and ev["properties"].get("old_attributes") and ev["properties"]["old_attributes"].get("owner_uuid")
374 olditemparent = self.inodes.inode_cache.find(oldowner)
375 if olditemparent is not None:
376 olditemparent.invalidate()
377 olditemparent.update()
379 itemparent = self.inodes.inode_cache.find(ev["object_owner_uuid"])
380 if itemparent is not None:
381 itemparent.invalidate()
386 def getattr(self, inode):
387 if inode not in self.inodes:
388 raise llfuse.FUSEError(errno.ENOENT)
390 e = self.inodes[inode]
392 entry = llfuse.EntryAttributes()
395 entry.entry_timeout = 60
396 entry.attr_timeout = 60
398 entry.st_mode = stat.S_IRUSR | stat.S_IRGRP | stat.S_IROTH
399 if isinstance(e, Directory):
400 entry.st_mode |= stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH | stat.S_IFDIR
402 entry.st_mode |= stat.S_IFREG
403 if isinstance(e, FuseArvadosFile):
404 entry.st_mode |= stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH
406 if self.enable_write and e.writable():
407 entry.st_mode |= stat.S_IWUSR | stat.S_IWGRP | stat.S_IWOTH
410 entry.st_uid = self.uid
411 entry.st_gid = self.gid
414 entry.st_size = e.size()
416 entry.st_blksize = 512
417 entry.st_blocks = (entry.st_size/512)+1
418 entry.st_atime = int(e.atime())
419 entry.st_mtime = int(e.mtime())
420 entry.st_ctime = int(e.mtime())
425 def setattr(self, inode, attr):
426 entry = self.getattr(inode)
428 e = self.inodes[inode]
430 if attr.st_size is not None and isinstance(e, FuseArvadosFile):
431 with llfuse.lock_released:
432 e.arvfile.truncate(attr.st_size)
433 entry.st_size = e.arvfile.size()
438 def lookup(self, parent_inode, name):
439 name = unicode(name, self.inodes.encoding)
445 if parent_inode in self.inodes:
446 p = self.inodes[parent_inode]
448 inode = p.parent_inode
449 elif isinstance(p, Directory) and name in p:
450 inode = p[name].inode
453 _logger.debug("arv-mount lookup: parent_inode %i name '%s' inode %i",
454 parent_inode, name, inode)
455 self.inodes[inode].inc_ref()
456 return self.getattr(inode)
458 _logger.debug("arv-mount lookup: parent_inode %i name '%s' not found",
460 raise llfuse.FUSEError(errno.ENOENT)
463 def forget(self, inodes):
464 for inode, nlookup in inodes:
465 ent = self.inodes[inode]
466 _logger.debug("arv-mount forget: inode %i nlookup %i ref_count %i", inode, nlookup, ent.ref_count)
467 if ent.dec_ref(nlookup) == 0 and ent.dead:
468 self.inodes.del_entry(ent)
471 def open(self, inode, flags):
472 if inode in self.inodes:
473 p = self.inodes[inode]
475 raise llfuse.FUSEError(errno.ENOENT)
477 if isinstance(p, Directory):
478 raise llfuse.FUSEError(errno.EISDIR)
480 if ((flags & os.O_WRONLY) or (flags & os.O_RDWR)) and not p.writable():
481 raise llfuse.FUSEError(errno.EPERM)
483 fh = next(self._filehandles_counter)
484 self._filehandles[fh] = FileHandle(fh, p)
489 def read(self, fh, off, size):
490 _logger.debug("arv-mount read %i %i %i", fh, off, size)
491 if fh in self._filehandles:
492 handle = self._filehandles[fh]
494 raise llfuse.FUSEError(errno.EBADF)
496 self.inodes.touch(handle.obj)
498 r = handle.obj.readfrom(off, size, self.num_retries)
500 self.read_counter.add(len(r))
504 def write(self, fh, off, buf):
505 _logger.debug("arv-mount write %i %i %i", fh, off, len(buf))
506 if fh in self._filehandles:
507 handle = self._filehandles[fh]
509 raise llfuse.FUSEError(errno.EBADF)
511 if not handle.obj.writable():
512 raise llfuse.FUSEError(errno.EPERM)
514 self.inodes.touch(handle.obj)
516 w = handle.obj.writeto(off, buf, self.num_retries)
518 self.write_counter.add(w)
522 def release(self, fh):
523 if fh in self._filehandles:
525 self._filehandles[fh].flush()
529 self._filehandles[fh].release()
530 del self._filehandles[fh]
531 self.inodes.inode_cache.cap_cache()
533 def releasedir(self, fh):
537 def opendir(self, inode):
538 _logger.debug("arv-mount opendir: inode %i", inode)
540 if inode in self.inodes:
541 p = self.inodes[inode]
543 raise llfuse.FUSEError(errno.ENOENT)
545 if not isinstance(p, Directory):
546 raise llfuse.FUSEError(errno.ENOTDIR)
548 fh = next(self._filehandles_counter)
549 if p.parent_inode in self.inodes:
550 parent = self.inodes[p.parent_inode]
552 raise llfuse.FUSEError(errno.EIO)
557 self._filehandles[fh] = DirectoryHandle(fh, p, [('.', p), ('..', parent)] + list(p.items()))
561 def readdir(self, fh, off):
562 _logger.debug("arv-mount readdir: fh %i off %i", fh, off)
564 if fh in self._filehandles:
565 handle = self._filehandles[fh]
567 raise llfuse.FUSEError(errno.EBADF)
569 _logger.debug("arv-mount handle.dirobj %s", handle.obj)
572 while e < len(handle.entries):
573 if handle.entries[e][1].inode in self.inodes:
574 yield (handle.entries[e][0].encode(self.inodes.encoding), self.getattr(handle.entries[e][1].inode), e+1)
579 st = llfuse.StatvfsData()
580 st.f_bsize = 128 * 1024
593 def _check_writable(self, inode_parent):
594 if not self.enable_write:
595 raise llfuse.FUSEError(errno.EROFS)
597 if inode_parent in self.inodes:
598 p = self.inodes[inode_parent]
600 raise llfuse.FUSEError(errno.ENOENT)
602 if not isinstance(p, Directory):
603 raise llfuse.FUSEError(errno.ENOTDIR)
606 raise llfuse.FUSEError(errno.EPERM)
611 def create(self, inode_parent, name, mode, flags, ctx):
612 _logger.debug("arv-mount create: %i '%s' %o", inode_parent, name, mode)
614 p = self._check_writable(inode_parent)
617 # The file entry should have been implicitly created by callback.
619 fh = next(self._filehandles_counter)
620 self._filehandles[fh] = FileHandle(fh, f)
624 return (fh, self.getattr(f.inode))
627 def mkdir(self, inode_parent, name, mode, ctx):
628 _logger.debug("arv-mount mkdir: %i '%s' %o", inode_parent, name, mode)
630 p = self._check_writable(inode_parent)
633 # The dir entry should have been implicitly created by callback.
637 return self.getattr(d.inode)
640 def unlink(self, inode_parent, name):
641 _logger.debug("arv-mount unlink: %i '%s'", inode_parent, name)
642 p = self._check_writable(inode_parent)
646 def rmdir(self, inode_parent, name):
647 _logger.debug("arv-mount rmdir: %i '%s'", inode_parent, name)
648 p = self._check_writable(inode_parent)
652 def rename(self, inode_parent_old, name_old, inode_parent_new, name_new):
653 _logger.debug("arv-mount rename: %i '%s' %i '%s'", inode_parent_old, name_old, inode_parent_new, name_new)
654 src = self._check_writable(inode_parent_old)
655 dest = self._check_writable(inode_parent_new)
656 dest.rename(name_old, name_new, src)
660 if fh in self._filehandles:
661 self._filehandles[fh].flush()
663 def fsync(self, fh, datasync):
666 def fsyncdir(self, fh, datasync):