-#
-# FUSE driver for Arvados Keep
-#
+"""FUSE driver for Arvados Keep
+
+Architecture:
+
+There is one `Operations` object per mount point. It is the entry point for all
+read and write requests from the llfuse module.
+
+The operations object owns an `Inodes` object. The inodes object stores the
+mapping from numeric inode (used throughout the file system API to uniquely
+identify files) to the Python objects that implement files and directories.
+
+The `Inodes` object owns an `InodeCache` object. The inode cache records the
+memory footprint of file system objects and when they are last used. When the
+cache limit is exceeded, the least recently used objects are cleared.
+
+File system objects inherit from `fresh.FreshBase` which manages the object lifecycle.
+
+File objects inherit from `fusefile.File`. Key methods are `readfrom` and `writeto`
+which implement actual reads and writes.
+
+Directory objects inherit from `fusedir.Directory`. The directory object wraps
+a Python dict which stores the mapping from filenames to directory entries.
+Directory contents can be accessed through the Python operators such as `[]`
+and `in`. These methods automatically check if the directory is fresh (up to
+date) or stale (needs update) and will call `update` if necessary before
+returing a result.
+
+The general FUSE operation flow is as follows:
+
+- The request handler is called with either an inode or file handle that is the
+ subject of the operation.
+
+- Look up the inode using the Inodes table or the file handle in the
+ filehandles table to get the file system object.
+
+- For methods that alter files or directories, check that the operation is
+ valid and permitted using _check_writable().
+
+- Call the relevant method on the file system object.
+
+- Return the result.
+
+The FUSE driver supports the Arvados event bus. When an event is received for
+an object that is live in the inode cache, that object is immediately updated.
+
+"""
import os
import sys
import collections
import functools
+import Queue
+
+# Default _notify_queue has a limit of 1000 items, but it really needs to be
+# unlimited to avoid deadlocks, see https://arvados.org/issues/3198#note-43 for
+# details.
+
+llfuse.capi._notify_queue = Queue.Queue()
+
from fusedir import sanitize_filename, Directory, CollectionDirectory, MagicDirectory, TagsDirectory, ProjectDirectory, SharedDirectory, CollectionDirectoryBase
from fusefile import StringFile, FuseArvadosFile
_logger = logging.getLogger('arvados.arvados_fuse')
-log_handler = logging.StreamHandler()
-llogger = logging.getLogger('llfuse')
-llogger.addHandler(log_handler)
-llogger.setLevel(logging.DEBUG)
+# Uncomment this to enable llfuse debug logging.
+# log_handler = logging.StreamHandler()
+# llogger = logging.getLogger('llfuse')
+# llogger.addHandler(log_handler)
+# llogger.setLevel(logging.DEBUG)
class Handle(object):
"""Connects a numeric file handle to a File or Directory object that has
class InodeCache(object):
+ """Records the memory footprint of objects and when they are last used.
+
+ When the cache limit is exceeded, the least recently used objects are
+ cleared. Clearing the object means discarding its contents to release
+ memory. The next time the object is accessed, it must be re-fetched from
+ the server. Note that the inode cache limit is a soft limit; the cache
+ limit may be exceeded if necessary to load very large objects, it may also
+ be exceeded if open file handles prevent objects from being cleared.
+
+ """
+
def __init__(self, cap, min_entries=4):
self._entries = collections.OrderedDict()
self._by_uuid = {}
return True
def cap_cache(self):
- #_logger.debug("InodeCache total is %i cap is %i", self._total, self.cap)
if self._total > self.cap:
for key in list(self._entries.keys()):
if self._total < self.cap or len(self._entries) < self.min_entries:
"""Manage the set of inodes. This is the mapping from a numeric id
to a concrete File or Directory object"""
- def __init__(self, inode_cache):
+ def __init__(self, inode_cache, encoding="utf-8"):
self._entries = {}
self._counter = itertools.count(llfuse.ROOT_INODE)
self.inode_cache = inode_cache
+ self.encoding = encoding
+ self.deferred_invalidations = []
def __getitem__(self, item):
return self._entries[item]
def del_entry(self, entry):
if entry.ref_count == 0:
- _logger.debug("Deleting inode %i", entry.inode)
self.inode_cache.unmanage(entry)
- llfuse.invalidate_inode(entry.inode)
del self._entries[entry.inode]
+ with llfuse.lock_released:
+ entry.finalize()
+ self.invalidate_inode(entry.inode)
entry.inode = None
else:
entry.dead = True
_logger.debug("del_entry on inode %i with refcount %i", entry.inode, entry.ref_count)
+ def invalidate_inode(self, inode):
+ llfuse.invalidate_inode(inode)
+
+ def invalidate_entry(self, inode, name):
+ llfuse.invalidate_entry(inode, name)
+
+
def catch_exceptions(orig_func):
+ """Catch uncaught exceptions and log them consistently."""
+
@functools.wraps(orig_func)
def catch_exceptions_wrapper(self, *args, **kwargs):
try:
raise
except EnvironmentError as e:
raise llfuse.FUSEError(e.errno)
+ except arvados.errors.KeepWriteError as e:
+ _logger.error("Keep write error: " + str(e))
+ raise llfuse.FUSEError(errno.EIO)
+ except arvados.errors.NotFoundError as e:
+ _logger.error("Block not found error: " + str(e))
+ raise llfuse.FUSEError(errno.EIO)
except:
_logger.exception("Unhandled exception during FUSE operation")
raise llfuse.FUSEError(errno.EIO)
"""
- def __init__(self, uid, gid, encoding="utf-8", inode_cache=None, num_retries=4):
+ def __init__(self, uid, gid, encoding="utf-8", inode_cache=None, num_retries=4, enable_write=False):
super(Operations, self).__init__()
if not inode_cache:
inode_cache = InodeCache(cap=256*1024*1024)
- self.inodes = Inodes(inode_cache)
+ self.inodes = Inodes(inode_cache, encoding=encoding)
self.uid = uid
self.gid = gid
- self.encoding = encoding
+ self.enable_write = enable_write
# dict of inode to filehandle
self._filehandles = {}
# initializing to continue
self.initlock.set()
+ @catch_exceptions
def destroy(self):
if self.events:
self.events.close()
+ self.events = None
+
+ for k,v in self.inodes.items():
+ try:
+ v.finalize()
+ except Exception as e:
+ _logger.exception("Error during finalize of inode %i", k)
+ self.inodes = None
def access(self, inode, mode, ctx):
return True
def listen_for_events(self, api_client):
- self.event = arvados.events.subscribe(api_client,
+ self.events = arvados.events.subscribe(api_client,
[["event_type", "in", ["create", "update", "delete"]]],
self.on_event)
+ @catch_exceptions
def on_event(self, ev):
if 'event_type' in ev:
with llfuse.lock:
item = self.inodes.inode_cache.find(ev["object_uuid"])
if item is not None:
item.invalidate()
- item.update()
+ if ev["object_kind"] == "arvados#collection":
+ new_attr = ev.get("properties") and ev["properties"].get("new_attributes") and ev["properties"]["new_attributes"]
+
+ # new_attributes.modified_at currently lacks subsecond precision (see #6347) so use event_at which
+ # should always be the same.
+ #record_version = (new_attr["modified_at"], new_attr["portable_data_hash"]) if new_attr else None
+ record_version = (ev["event_at"], new_attr["portable_data_hash"]) if new_attr else None
+
+ item.update(to_record_version=record_version)
+ else:
+ item.update()
+
+ oldowner = ev.get("properties") and ev["properties"].get("old_attributes") and ev["properties"]["old_attributes"].get("owner_uuid")
+ olditemparent = self.inodes.inode_cache.find(oldowner)
+ if olditemparent is not None:
+ olditemparent.invalidate()
+ olditemparent.update()
itemparent = self.inodes.inode_cache.find(ev["object_owner_uuid"])
if itemparent is not None:
itemparent.invalidate()
itemparent.update()
+
@catch_exceptions
def getattr(self, inode):
if inode not in self.inodes:
entry = llfuse.EntryAttributes()
entry.st_ino = inode
entry.generation = 0
- entry.entry_timeout = 300
- entry.attr_timeout = 300
+ entry.entry_timeout = 60
+ entry.attr_timeout = 60
entry.st_mode = stat.S_IRUSR | stat.S_IRGRP | stat.S_IROTH
if isinstance(e, Directory):
if isinstance(e, FuseArvadosFile):
entry.st_mode |= stat.S_IXUSR | stat.S_IXGRP | stat.S_IXOTH
- if e.writable():
+ if self.enable_write and e.writable():
entry.st_mode |= stat.S_IWUSR | stat.S_IWGRP | stat.S_IWOTH
entry.st_nlink = 1
entry.st_size = e.size()
entry.st_blksize = 512
- entry.st_blocks = (e.size()/512)+1
+ entry.st_blocks = (entry.st_size/512)+1
entry.st_atime = int(e.atime())
entry.st_mtime = int(e.mtime())
entry.st_ctime = int(e.mtime())
return entry
+ @catch_exceptions
+ def setattr(self, inode, attr):
+ entry = self.getattr(inode)
+
+ e = self.inodes[inode]
+
+ if attr.st_size is not None and isinstance(e, FuseArvadosFile):
+ with llfuse.lock_released:
+ e.arvfile.truncate(attr.st_size)
+ entry.st_size = e.arvfile.size()
+
+ return entry
+
@catch_exceptions
def lookup(self, parent_inode, name):
- name = unicode(name, self.encoding)
+ name = unicode(name, self.inodes.encoding)
inode = None
if name == '.':
self.inodes.touch(handle.obj)
- try:
- with llfuse.lock_released:
- return handle.obj.readfrom(off, size, self.num_retries)
- except arvados.errors.NotFoundError as e:
- _logger.warning("Block not found: " + str(e))
- raise llfuse.FUSEError(errno.EIO)
+ return handle.obj.readfrom(off, size, self.num_retries)
@catch_exceptions
def write(self, fh, off, buf):
self.inodes.touch(handle.obj)
- with llfuse.lock_released:
- return handle.obj.writeto(off, buf, self.num_retries)
+ return handle.obj.writeto(off, buf, self.num_retries)
@catch_exceptions
def release(self, fh):
if fh in self._filehandles:
try:
self._filehandles[fh].flush()
- except EnvironmentError as e:
- raise llfuse.FUSEError(e.errno)
except Exception:
- _logger.exception("Flush error")
- self._filehandles[fh].release()
- del self._filehandles[fh]
+ raise
+ finally:
+ self._filehandles[fh].release()
+ del self._filehandles[fh]
self.inodes.inode_cache.cap_cache()
def releasedir(self, fh):
e = off
while e < len(handle.entries):
if handle.entries[e][1].inode in self.inodes:
- try:
- yield (handle.entries[e][0].encode(self.encoding), self.getattr(handle.entries[e][1].inode), e+1)
- except UnicodeEncodeError:
- pass
+ yield (handle.entries[e][0].encode(self.inodes.encoding), self.getattr(handle.entries[e][1].inode), e+1)
e += 1
@catch_exceptions
def statfs(self):
st = llfuse.StatvfsData()
- st.f_bsize = 64 * 1024
+ st.f_bsize = 128 * 1024
st.f_blocks = 0
st.f_files = 0
return st
def _check_writable(self, inode_parent):
+ if not self.enable_write:
+ raise llfuse.FUSEError(errno.EROFS)
+
if inode_parent in self.inodes:
p = self.inodes[inode_parent]
else:
@catch_exceptions
def create(self, inode_parent, name, mode, flags, ctx):
+ _logger.debug("arv-mount create: %i '%s' %o", inode_parent, name, mode)
+
p = self._check_writable(inode_parent)
p.create(name)
projects / arvados.git / blobdiff