4 # Copyright (C) 2006, 2007, 2008, 2009, 2010 Google Inc.
6 # This program is free software; you can redistribute it and/or modify
7 # it under the terms of the GNU General Public License as published by
8 # the Free Software Foundation; either version 2 of the License, or
9 # (at your option) any later version.
11 # This program is distributed in the hope that it will be useful, but
12 # WITHOUT ANY WARRANTY; without even the implied warranty of
13 # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
14 # General Public License for more details.
16 # You should have received a copy of the GNU General Public License
17 # along with this program; if not, write to the Free Software
18 # Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA
21 """Module implementing the Ganeti locking code."""
23 # pylint: disable-msg=W0212
25 # W0212 since e.g. LockSet methods use (a lot) the internals of
38 from ganeti import errors
39 from ganeti import utils
40 from ganeti import compat
41 from ganeti import query
44 _EXCLUSIVE_TEXT = "exclusive"
45 _SHARED_TEXT = "shared"
46 _DELETED_TEXT = "deleted"
51 def ssynchronized(mylock, shared=0):
52 """Shared Synchronization decorator.
54 Calls the function holding the given lock, either in exclusive or shared
55 mode. It requires the passed lock to be a SharedLock (or support its
58 @type mylock: lockable object or string
59 @param mylock: lock to acquire or class member name of the lock to acquire
63 def sync_function(*args, **kwargs):
64 if isinstance(mylock, basestring):
65 assert args, "cannot ssynchronize on non-class method: self not found"
67 lock = getattr(args[0], mylock)
70 lock.acquire(shared=shared)
72 return fn(*args, **kwargs)
79 class _SingleNotifyPipeConditionWaiter(object):
80 """Helper class for SingleNotifyPipeCondition
88 def __init__(self, poller, fd):
89 """Constructor for _SingleNotifyPipeConditionWaiter
91 @type poller: select.poll
92 @param poller: Poller object
94 @param fd: File descriptor to wait for
101 def __call__(self, timeout):
102 """Wait for something to happen on the pipe.
104 @type timeout: float or None
105 @param timeout: Timeout for waiting (can be None)
108 running_timeout = utils.RunningTimeout(timeout, True)
111 remaining_time = running_timeout.Remaining()
113 if remaining_time is not None:
114 if remaining_time < 0.0:
117 # Our calculation uses seconds, poll() wants milliseconds
118 remaining_time *= 1000
121 result = self._poller.poll(remaining_time)
122 except EnvironmentError, err:
123 if err.errno != errno.EINTR:
127 # Check whether we were notified
128 if result and result[0][0] == self._fd:
132 class _BaseCondition(object):
133 """Base class containing common code for conditions.
135 Some of this code is taken from python's threading module.
147 def __init__(self, lock):
148 """Constructor for _BaseCondition.
150 @type lock: threading.Lock
151 @param lock: condition base lock
154 object.__init__(self)
157 self._release_save = lock._release_save
158 except AttributeError:
159 self._release_save = self._base_release_save
161 self._acquire_restore = lock._acquire_restore
162 except AttributeError:
163 self._acquire_restore = self._base_acquire_restore
165 self._is_owned = lock._is_owned
166 except AttributeError:
167 self._is_owned = self._base_is_owned
171 # Export the lock's acquire() and release() methods
172 self.acquire = lock.acquire
173 self.release = lock.release
175 def _base_is_owned(self):
176 """Check whether lock is owned by current thread.
179 if self._lock.acquire(0):
184 def _base_release_save(self):
187 def _base_acquire_restore(self, _):
190 def _check_owned(self):
191 """Raise an exception if the current thread doesn't own the lock.
194 if not self._is_owned():
195 raise RuntimeError("cannot work with un-aquired lock")
198 class SingleNotifyPipeCondition(_BaseCondition):
199 """Condition which can only be notified once.
201 This condition class uses pipes and poll, internally, to be able to wait for
202 notification with a timeout, without resorting to polling. It is almost
203 compatible with Python's threading.Condition, with the following differences:
204 - notifyAll can only be called once, and no wait can happen after that
205 - notify is not supported, only notifyAll
217 _waiter_class = _SingleNotifyPipeConditionWaiter
219 def __init__(self, lock):
220 """Constructor for SingleNotifyPipeCondition
223 _BaseCondition.__init__(self, lock)
225 self._notified = False
227 self._write_fd = None
230 def _check_unnotified(self):
231 """Throws an exception if already notified.
235 raise RuntimeError("cannot use already notified condition")
238 """Cleanup open file descriptors, if any.
241 if self._read_fd is not None:
242 os.close(self._read_fd)
245 if self._write_fd is not None:
246 os.close(self._write_fd)
247 self._write_fd = None
250 def wait(self, timeout):
251 """Wait for a notification.
253 @type timeout: float or None
254 @param timeout: Waiting timeout (can be None)
258 self._check_unnotified()
262 if self._poller is None:
263 (self._read_fd, self._write_fd) = os.pipe()
264 self._poller = select.poll()
265 self._poller.register(self._read_fd, select.POLLHUP)
267 wait_fn = self._waiter_class(self._poller, self._read_fd)
268 state = self._release_save()
270 # Wait for notification
274 self._acquire_restore(state)
277 if self._nwaiters == 0:
280 def notifyAll(self): # pylint: disable-msg=C0103
281 """Close the writing side of the pipe to notify all waiters.
285 self._check_unnotified()
286 self._notified = True
287 if self._write_fd is not None:
288 os.close(self._write_fd)
289 self._write_fd = None
292 class PipeCondition(_BaseCondition):
293 """Group-only non-polling condition with counters.
295 This condition class uses pipes and poll, internally, to be able to wait for
296 notification with a timeout, without resorting to polling. It is almost
297 compatible with Python's threading.Condition, but only supports notifyAll and
298 non-recursive locks. As an additional features it's able to report whether
299 there are any waiting threads.
307 _single_condition_class = SingleNotifyPipeCondition
309 def __init__(self, lock):
310 """Initializes this class.
313 _BaseCondition.__init__(self, lock)
314 self._waiters = set()
315 self._single_condition = self._single_condition_class(self._lock)
317 def wait(self, timeout):
318 """Wait for a notification.
320 @type timeout: float or None
321 @param timeout: Waiting timeout (can be None)
326 # Keep local reference to the pipe. It could be replaced by another thread
327 # notifying while we're waiting.
328 cond = self._single_condition
330 self._waiters.add(threading.currentThread())
335 self._waiters.remove(threading.currentThread())
337 def notifyAll(self): # pylint: disable-msg=C0103
338 """Notify all currently waiting threads.
342 self._single_condition.notifyAll()
343 self._single_condition = self._single_condition_class(self._lock)
345 def get_waiting(self):
346 """Returns a list of all waiting threads.
353 def has_waiting(self):
354 """Returns whether there are active waiters.
359 return bool(self._waiters)
362 class _PipeConditionWithMode(PipeCondition):
367 def __init__(self, lock, shared):
368 """Initializes this class.
372 PipeCondition.__init__(self, lock)
375 class SharedLock(object):
376 """Implements a shared lock.
378 Multiple threads can acquire the lock in a shared way by calling
379 C{acquire(shared=1)}. In order to acquire the lock in an exclusive way
380 threads can call C{acquire(shared=0)}.
382 Notes on data structures: C{__pending} contains a priority queue (heapq) of
383 all pending acquires: C{[(priority1: prioqueue1), (priority2: prioqueue2),
384 ...]}. Each per-priority queue contains a normal in-order list of conditions
385 to be notified when the lock can be acquired. Shared locks are grouped
386 together by priority and the condition for them is stored in
387 C{__pending_shared} if it already exists. C{__pending_by_prio} keeps
388 references for the per-priority queues indexed by priority for faster access.
391 @ivar name: the name of the lock
406 __condition_class = _PipeConditionWithMode
408 def __init__(self, name, monitor=None):
409 """Construct a new SharedLock.
411 @param name: the name of the lock
412 @type monitor: L{LockMonitor}
413 @param monitor: Lock monitor with which to register
416 object.__init__(self)
421 self.__lock = threading.Lock()
423 # Queue containing waiting acquires
425 self.__pending_by_prio = {}
426 self.__pending_shared = {}
428 # Current lock holders
432 # is this lock in the deleted state?
433 self.__deleted = False
435 # Register with lock monitor
437 monitor.RegisterLock(self)
439 def GetInfo(self, requested):
440 """Retrieves information for querying locks.
443 @param requested: Requested information, see C{query.LQ_*}
446 self.__lock.acquire()
448 # Note: to avoid unintentional race conditions, no references to
449 # modifiable objects should be returned unless they were created in this
454 if query.LQ_MODE in requested:
457 assert not (self.__exc or self.__shr)
459 mode = _EXCLUSIVE_TEXT
463 # Current owner(s) are wanted
464 if query.LQ_OWNER in requested:
471 assert not self.__deleted
472 owner_names = [i.getName() for i in owner]
474 # Pending acquires are wanted
475 if query.LQ_PENDING in requested:
478 # Sorting instead of copying and using heaq functions for simplicity
479 for (_, prioqueue) in sorted(self.__pending):
480 for cond in prioqueue:
482 pendmode = _SHARED_TEXT
484 pendmode = _EXCLUSIVE_TEXT
486 # List of names will be sorted in L{query._GetLockPending}
487 pending.append((pendmode, [i.getName()
488 for i in cond.get_waiting()]))
492 return (self.name, mode, owner_names, pending)
494 self.__lock.release()
496 def __check_deleted(self):
497 """Raises an exception if the lock has been deleted.
501 raise errors.LockError("Deleted lock %s" % self.name)
503 def __is_sharer(self):
504 """Is the current thread sharing the lock at this time?
507 return threading.currentThread() in self.__shr
509 def __is_exclusive(self):
510 """Is the current thread holding the lock exclusively at this time?
513 return threading.currentThread() == self.__exc
515 def __is_owned(self, shared=-1):
516 """Is the current thread somehow owning the lock at this time?
518 This is a private version of the function, which presumes you're holding
523 return self.__is_sharer() or self.__is_exclusive()
525 return self.__is_sharer()
527 return self.__is_exclusive()
529 def _is_owned(self, shared=-1):
530 """Is the current thread somehow owning the lock at this time?
533 - < 0: check for any type of ownership (default)
534 - 0: check for exclusive ownership
535 - > 0: check for shared ownership
538 self.__lock.acquire()
540 return self.__is_owned(shared=shared)
542 self.__lock.release()
544 def _count_pending(self):
545 """Returns the number of pending acquires.
550 self.__lock.acquire()
552 return sum(len(prioqueue) for (_, prioqueue) in self.__pending)
554 self.__lock.release()
556 def _check_empty(self):
557 """Checks whether there are any pending acquires.
562 self.__lock.acquire()
564 # Order is important: __find_first_pending_queue modifies __pending
565 (_, prioqueue) = self.__find_first_pending_queue()
567 return not (prioqueue or
569 self.__pending_by_prio or
570 self.__pending_shared)
572 self.__lock.release()
574 def __do_acquire(self, shared):
575 """Actually acquire the lock.
579 self.__shr.add(threading.currentThread())
581 self.__exc = threading.currentThread()
583 def __can_acquire(self, shared):
584 """Determine whether lock can be acquired.
588 return self.__exc is None
590 return len(self.__shr) == 0 and self.__exc is None
592 def __find_first_pending_queue(self):
593 """Tries to find the topmost queued entry with pending acquires.
595 Removes empty entries while going through the list.
598 while self.__pending:
599 (priority, prioqueue) = self.__pending[0]
602 return (priority, prioqueue)
605 heapq.heappop(self.__pending)
606 del self.__pending_by_prio[priority]
607 assert priority not in self.__pending_shared
611 def __is_on_top(self, cond):
612 """Checks whether the passed condition is on top of the queue.
614 The caller must make sure the queue isn't empty.
617 (_, prioqueue) = self.__find_first_pending_queue()
619 return cond == prioqueue[0]
621 def __acquire_unlocked(self, shared, timeout, priority):
622 """Acquire a shared lock.
624 @param shared: whether to acquire in shared mode; by default an
625 exclusive lock will be acquired
626 @param timeout: maximum waiting time before giving up
627 @type priority: integer
628 @param priority: Priority for acquiring lock
631 self.__check_deleted()
633 # We cannot acquire the lock if we already have it
634 assert not self.__is_owned(), ("double acquire() on a non-recursive lock"
637 # Remove empty entries from queue
638 self.__find_first_pending_queue()
640 # Check whether someone else holds the lock or there are pending acquires.
641 if not self.__pending and self.__can_acquire(shared):
642 # Apparently not, can acquire lock directly.
643 self.__do_acquire(shared)
646 prioqueue = self.__pending_by_prio.get(priority, None)
649 # Try to re-use condition for shared acquire
650 wait_condition = self.__pending_shared.get(priority, None)
651 assert (wait_condition is None or
652 (wait_condition.shared and wait_condition in prioqueue))
654 wait_condition = None
656 if wait_condition is None:
657 if prioqueue is None:
658 assert priority not in self.__pending_by_prio
661 heapq.heappush(self.__pending, (priority, prioqueue))
662 self.__pending_by_prio[priority] = prioqueue
664 wait_condition = self.__condition_class(self.__lock, shared)
665 prioqueue.append(wait_condition)
668 # Keep reference for further shared acquires on same priority. This is
669 # better than trying to find it in the list of pending acquires.
670 assert priority not in self.__pending_shared
671 self.__pending_shared[priority] = wait_condition
674 # Wait until we become the topmost acquire in the queue or the timeout
676 # TODO: Decrease timeout with spurious notifications
677 while not (self.__is_on_top(wait_condition) and
678 self.__can_acquire(shared)):
679 # Wait for notification
680 wait_condition.wait(timeout)
681 self.__check_deleted()
683 # A lot of code assumes blocking acquires always succeed. Loop
684 # internally for that case.
685 if timeout is not None:
688 if self.__is_on_top(wait_condition) and self.__can_acquire(shared):
689 self.__do_acquire(shared)
692 # Remove condition from queue if there are no more waiters
693 if not wait_condition.has_waiting():
694 prioqueue.remove(wait_condition)
695 if wait_condition.shared:
696 # Remove from list of shared acquires if it wasn't while releasing
697 # (e.g. on lock deletion)
698 self.__pending_shared.pop(priority, None)
702 def acquire(self, shared=0, timeout=None, priority=None,
704 """Acquire a shared lock.
706 @type shared: integer (0/1) used as a boolean
707 @param shared: whether to acquire in shared mode; by default an
708 exclusive lock will be acquired
710 @param timeout: maximum waiting time before giving up
711 @type priority: integer
712 @param priority: Priority for acquiring lock
713 @type test_notify: callable or None
714 @param test_notify: Special callback function for unittesting
718 priority = _DEFAULT_PRIORITY
720 self.__lock.acquire()
722 # We already got the lock, notify now
723 if __debug__ and callable(test_notify):
726 return self.__acquire_unlocked(shared, timeout, priority)
728 self.__lock.release()
731 """Changes the lock mode from exclusive to shared.
733 Pending acquires in shared mode on the same priority will go ahead.
736 self.__lock.acquire()
738 assert self.__is_owned(), "Lock must be owned"
740 if self.__is_exclusive():
741 # Do nothing if the lock is already acquired in shared mode
745 # Important: pending shared acquires should only jump ahead if there
746 # was a transition from exclusive to shared, otherwise an owner of a
747 # shared lock can keep calling this function to push incoming shared
749 (priority, prioqueue) = self.__find_first_pending_queue()
751 # Is there a pending shared acquire on this priority?
752 cond = self.__pending_shared.pop(priority, None)
755 assert cond in prioqueue
757 # Ensure shared acquire is on top of queue
758 if len(prioqueue) > 1:
759 prioqueue.remove(cond)
760 prioqueue.insert(0, cond)
765 assert not self.__is_exclusive()
766 assert self.__is_sharer()
770 self.__lock.release()
773 """Release a Shared Lock.
775 You must have acquired the lock, either in shared or in exclusive mode,
776 before calling this function.
779 self.__lock.acquire()
781 assert self.__is_exclusive() or self.__is_sharer(), \
782 "Cannot release non-owned lock"
784 # Autodetect release type
785 if self.__is_exclusive():
788 self.__shr.remove(threading.currentThread())
790 # Notify topmost condition in queue
791 (priority, prioqueue) = self.__find_first_pending_queue()
796 # Prevent further shared acquires from sneaking in while waiters are
798 self.__pending_shared.pop(priority, None)
801 self.__lock.release()
803 def delete(self, timeout=None, priority=None):
804 """Delete a Shared Lock.
806 This operation will declare the lock for removal. First the lock will be
807 acquired in exclusive mode if you don't already own it, then the lock
808 will be put in a state where any future and pending acquire() fail.
811 @param timeout: maximum waiting time before giving up
812 @type priority: integer
813 @param priority: Priority for acquiring lock
817 priority = _DEFAULT_PRIORITY
819 self.__lock.acquire()
821 assert not self.__is_sharer(), "Cannot delete() a lock while sharing it"
823 self.__check_deleted()
825 # The caller is allowed to hold the lock exclusively already.
826 acquired = self.__is_exclusive()
829 acquired = self.__acquire_unlocked(0, timeout, priority)
831 assert self.__is_exclusive() and not self.__is_sharer(), \
832 "Lock wasn't acquired in exclusive mode"
835 self.__deleted = True
838 assert not (self.__exc or self.__shr), "Found owner during deletion"
840 # Notify all acquires. They'll throw an error.
841 for (_, prioqueue) in self.__pending:
842 for cond in prioqueue:
845 assert self.__deleted
849 self.__lock.release()
851 def _release_save(self):
852 shared = self.__is_sharer()
856 def _acquire_restore(self, shared):
857 self.acquire(shared=shared)
860 # Whenever we want to acquire a full LockSet we pass None as the value
861 # to acquire. Hide this behind this nicely named constant.
865 class _AcquireTimeout(Exception):
866 """Internal exception to abort an acquire on a timeout.
872 """Implements a set of locks.
874 This abstraction implements a set of shared locks for the same resource type,
875 distinguished by name. The user can lock a subset of the resources and the
876 LockSet will take care of acquiring the locks always in the same order, thus
879 All the locks needed in the same set must be acquired together, though.
882 @ivar name: the name of the lockset
885 def __init__(self, members, name, monitor=None):
886 """Constructs a new LockSet.
888 @type members: list of strings
889 @param members: initial members of the set
890 @type monitor: L{LockMonitor}
891 @param monitor: Lock monitor with which to register member locks
894 assert members is not None, "members parameter is not a list"
898 self.__monitor = monitor
900 # Used internally to guarantee coherency
901 self.__lock = SharedLock(self._GetLockName("[lockset]"), monitor=monitor)
903 # The lockdict indexes the relationship name -> lock
904 # The order-of-locking is implied by the alphabetical order of names
907 for mname in members:
908 self.__lockdict[mname] = SharedLock(self._GetLockName(mname),
911 # The owner dict contains the set of locks each thread owns. For
912 # performance each thread can access its own key without a global lock on
913 # this structure. It is paramount though that *no* other type of access is
914 # done to this structure (eg. no looping over its keys). *_owner helper
915 # function are defined to guarantee access is correct, but in general never
916 # do anything different than __owners[threading.currentThread()], or there
920 def _GetLockName(self, mname):
921 """Returns the name for a member lock.
924 return "%s/%s" % (self.name, mname)
927 """Returns the lockset-internal lock.
932 def _get_lockdict(self):
933 """Returns the lockset-internal lock dictionary.
935 Accessing this structure is only safe in single-thread usage or when the
936 lockset-internal lock is held.
939 return self.__lockdict
942 """Is the current thread a current level owner?"""
943 return threading.currentThread() in self.__owners
945 def _add_owned(self, name=None):
946 """Note the current thread owns the given lock"""
948 if not self._is_owned():
949 self.__owners[threading.currentThread()] = set()
952 self.__owners[threading.currentThread()].add(name)
954 self.__owners[threading.currentThread()] = set([name])
956 def _del_owned(self, name=None):
957 """Note the current thread owns the given lock"""
959 assert not (name is None and self.__lock._is_owned()), \
960 "Cannot hold internal lock when deleting owner status"
963 self.__owners[threading.currentThread()].remove(name)
965 # Only remove the key if we don't hold the set-lock as well
966 if (not self.__lock._is_owned() and
967 not self.__owners[threading.currentThread()]):
968 del self.__owners[threading.currentThread()]
970 def _list_owned(self):
971 """Get the set of resource names owned by the current thread"""
973 return self.__owners[threading.currentThread()].copy()
977 def _release_and_delete_owned(self):
978 """Release and delete all resources owned by the current thread"""
979 for lname in self._list_owned():
980 lock = self.__lockdict[lname]
983 self._del_owned(name=lname)
986 """Return the current set of names.
988 Only call this function while holding __lock and don't iterate on the
989 result after releasing the lock.
992 return self.__lockdict.keys()
995 """Return a copy of the current set of elements.
997 Used only for debugging purposes.
1000 # If we don't already own the set-level lock acquired
1001 # we'll get it and note we need to release it later.
1002 release_lock = False
1003 if not self.__lock._is_owned():
1005 self.__lock.acquire(shared=1)
1007 result = self.__names()
1010 self.__lock.release()
1013 def acquire(self, names, timeout=None, shared=0, priority=None,
1015 """Acquire a set of resource locks.
1017 @type names: list of strings (or string)
1018 @param names: the names of the locks which shall be acquired
1019 (special lock names, or instance/node names)
1020 @type shared: integer (0/1) used as a boolean
1021 @param shared: whether to acquire in shared mode; by default an
1022 exclusive lock will be acquired
1023 @type timeout: float or None
1024 @param timeout: Maximum time to acquire all locks
1025 @type priority: integer
1026 @param priority: Priority for acquiring locks
1027 @type test_notify: callable or None
1028 @param test_notify: Special callback function for unittesting
1030 @return: Set of all locks successfully acquired or None in case of timeout
1032 @raise errors.LockError: when any lock we try to acquire has
1033 been deleted before we succeed. In this case none of the
1034 locks requested will be acquired.
1037 assert timeout is None or timeout >= 0.0
1039 # Check we don't already own locks at this level
1040 assert not self._is_owned(), ("Cannot acquire locks in the same set twice"
1041 " (lockset %s)" % self.name)
1043 if priority is None:
1044 priority = _DEFAULT_PRIORITY
1046 # We need to keep track of how long we spent waiting for a lock. The
1047 # timeout passed to this function is over all lock acquires.
1048 running_timeout = utils.RunningTimeout(timeout, False)
1051 if names is not None:
1052 # Support passing in a single resource to acquire rather than many
1053 if isinstance(names, basestring):
1056 return self.__acquire_inner(names, False, shared, priority,
1057 running_timeout.Remaining, test_notify)
1060 # If no names are given acquire the whole set by not letting new names
1061 # being added before we release, and getting the current list of names.
1062 # Some of them may then be deleted later, but we'll cope with this.
1064 # We'd like to acquire this lock in a shared way, as it's nice if
1065 # everybody else can use the instances at the same time. If we are
1066 # acquiring them exclusively though they won't be able to do this
1067 # anyway, though, so we'll get the list lock exclusively as well in
1068 # order to be able to do add() on the set while owning it.
1069 if not self.__lock.acquire(shared=shared, priority=priority,
1070 timeout=running_timeout.Remaining()):
1071 raise _AcquireTimeout()
1073 # note we own the set-lock
1076 return self.__acquire_inner(self.__names(), True, shared, priority,
1077 running_timeout.Remaining, test_notify)
1079 # We shouldn't have problems adding the lock to the owners list, but
1080 # if we did we'll try to release this lock and re-raise exception.
1081 # Of course something is going to be really wrong, after this.
1082 self.__lock.release()
1086 except _AcquireTimeout:
1089 def __acquire_inner(self, names, want_all, shared, priority,
1090 timeout_fn, test_notify):
1091 """Inner logic for acquiring a number of locks.
1093 @param names: Names of the locks to be acquired
1094 @param want_all: Whether all locks in the set should be acquired
1095 @param shared: Whether to acquire in shared mode
1096 @param timeout_fn: Function returning remaining timeout
1097 @param priority: Priority for acquiring locks
1098 @param test_notify: Special callback function for unittesting
1103 # First we look the locks up on __lockdict. We have no way of being sure
1104 # they will still be there after, but this makes it a lot faster should
1105 # just one of them be the already wrong. Using a sorted sequence to prevent
1107 for lname in sorted(utils.UniqueSequence(names)):
1109 lock = self.__lockdict[lname] # raises KeyError if lock is not there
1112 # We are acquiring all the set, it doesn't matter if this particular
1113 # element is not there anymore.
1116 raise errors.LockError("Non-existing lock %s in set %s (it may have"
1117 " been removed)" % (lname, self.name))
1119 acquire_list.append((lname, lock))
1121 # This will hold the locknames we effectively acquired.
1125 # Now acquire_list contains a sorted list of resources and locks we
1126 # want. In order to get them we loop on this (private) list and
1127 # acquire() them. We gave no real guarantee they will still exist till
1128 # this is done but .acquire() itself is safe and will alert us if the
1129 # lock gets deleted.
1130 for (lname, lock) in acquire_list:
1131 if __debug__ and callable(test_notify):
1132 test_notify_fn = lambda: test_notify(lname)
1134 test_notify_fn = None
1136 timeout = timeout_fn()
1139 # raises LockError if the lock was deleted
1140 acq_success = lock.acquire(shared=shared, timeout=timeout,
1142 test_notify=test_notify_fn)
1143 except errors.LockError:
1145 # We are acquiring all the set, it doesn't matter if this
1146 # particular element is not there anymore.
1149 raise errors.LockError("Non-existing lock %s in set %s (it may"
1150 " have been removed)" % (lname, self.name))
1153 # Couldn't get lock or timeout occurred
1155 # This shouldn't happen as SharedLock.acquire(timeout=None) is
1157 raise errors.LockError("Failed to get lock %s (set %s)" %
1160 raise _AcquireTimeout()
1163 # now the lock cannot be deleted, we have it!
1164 self._add_owned(name=lname)
1168 # We shouldn't have problems adding the lock to the owners list, but
1169 # if we did we'll try to release this lock and re-raise exception.
1170 # Of course something is going to be really wrong after this.
1171 if lock._is_owned():
1176 # Release all owned locks
1177 self._release_and_delete_owned()
1182 def downgrade(self, names=None):
1183 """Downgrade a set of resource locks from exclusive to shared mode.
1185 The locks must have been acquired in exclusive mode.
1188 assert self._is_owned(), ("downgrade on lockset %s while not owning any"
1189 " lock" % self.name)
1191 # Support passing in a single resource to downgrade rather than many
1192 if isinstance(names, basestring):
1195 owned = self._list_owned()
1201 assert owned.issuperset(names), \
1202 ("downgrade() on unheld resources %s (set %s)" %
1203 (names.difference(owned), self.name))
1205 for lockname in names:
1206 self.__lockdict[lockname].downgrade()
1208 # Do we own the lockset in exclusive mode?
1209 if self.__lock._is_owned(shared=0):
1210 # Have all locks been downgraded?
1211 if not compat.any(lock._is_owned(shared=0)
1212 for lock in self.__lockdict.values()):
1213 self.__lock.downgrade()
1214 assert self.__lock._is_owned(shared=1)
1218 def release(self, names=None):
1219 """Release a set of resource locks, at the same level.
1221 You must have acquired the locks, either in shared or in exclusive mode,
1222 before releasing them.
1224 @type names: list of strings, or None
1225 @param names: the names of the locks which shall be released
1226 (defaults to all the locks acquired at that level).
1229 assert self._is_owned(), ("release() on lock set %s while not owner" %
1232 # Support passing in a single resource to release rather than many
1233 if isinstance(names, basestring):
1237 names = self._list_owned()
1240 assert self._list_owned().issuperset(names), (
1241 "release() on unheld resources %s (set %s)" %
1242 (names.difference(self._list_owned()), self.name))
1244 # First of all let's release the "all elements" lock, if set.
1245 # After this 'add' can work again
1246 if self.__lock._is_owned():
1247 self.__lock.release()
1250 for lockname in names:
1251 # If we are sure the lock doesn't leave __lockdict without being
1252 # exclusively held we can do this...
1253 self.__lockdict[lockname].release()
1254 self._del_owned(name=lockname)
1256 def add(self, names, acquired=0, shared=0):
1257 """Add a new set of elements to the set
1259 @type names: list of strings
1260 @param names: names of the new elements to add
1261 @type acquired: integer (0/1) used as a boolean
1262 @param acquired: pre-acquire the new resource?
1263 @type shared: integer (0/1) used as a boolean
1264 @param shared: is the pre-acquisition shared?
1267 # Check we don't already own locks at this level
1268 assert not self._is_owned() or self.__lock._is_owned(shared=0), \
1269 ("Cannot add locks if the set %s is only partially owned, or shared" %
1272 # Support passing in a single resource to add rather than many
1273 if isinstance(names, basestring):
1276 # If we don't already own the set-level lock acquired in an exclusive way
1277 # we'll get it and note we need to release it later.
1278 release_lock = False
1279 if not self.__lock._is_owned():
1281 self.__lock.acquire()
1284 invalid_names = set(self.__names()).intersection(names)
1286 # This must be an explicit raise, not an assert, because assert is
1287 # turned off when using optimization, and this can happen because of
1288 # concurrency even if the user doesn't want it.
1289 raise errors.LockError("duplicate add(%s) on lockset %s" %
1290 (invalid_names, self.name))
1292 for lockname in names:
1293 lock = SharedLock(self._GetLockName(lockname), monitor=self.__monitor)
1296 # No need for priority or timeout here as this lock has just been
1298 lock.acquire(shared=shared)
1299 # now the lock cannot be deleted, we have it!
1301 self._add_owned(name=lockname)
1303 # We shouldn't have problems adding the lock to the owners list,
1304 # but if we did we'll try to release this lock and re-raise
1305 # exception. Of course something is going to be really wrong,
1306 # after this. On the other hand the lock hasn't been added to the
1307 # __lockdict yet so no other threads should be pending on it. This
1308 # release is just a safety measure.
1312 self.__lockdict[lockname] = lock
1315 # Only release __lock if we were not holding it previously.
1317 self.__lock.release()
1321 def remove(self, names):
1322 """Remove elements from the lock set.
1324 You can either not hold anything in the lockset or already hold a superset
1325 of the elements you want to delete, exclusively.
1327 @type names: list of strings
1328 @param names: names of the resource to remove.
1330 @return: a list of locks which we removed; the list is always
1331 equal to the names list if we were holding all the locks
1335 # Support passing in a single resource to remove rather than many
1336 if isinstance(names, basestring):
1339 # If we own any subset of this lock it must be a superset of what we want
1340 # to delete. The ownership must also be exclusive, but that will be checked
1341 # by the lock itself.
1342 assert not self._is_owned() or self._list_owned().issuperset(names), (
1343 "remove() on acquired lockset %s while not owning all elements" %
1349 # Calling delete() acquires the lock exclusively if we don't already own
1350 # it, and causes all pending and subsequent lock acquires to fail. It's
1351 # fine to call it out of order because delete() also implies release(),
1352 # and the assertion above guarantees that if we either already hold
1353 # everything we want to delete, or we hold none.
1355 self.__lockdict[lname].delete()
1356 removed.append(lname)
1357 except (KeyError, errors.LockError):
1358 # This cannot happen if we were already holding it, verify:
1359 assert not self._is_owned(), ("remove failed while holding lockset %s"
1362 # If no LockError was raised we are the ones who deleted the lock.
1363 # This means we can safely remove it from lockdict, as any further or
1364 # pending delete() or acquire() will fail (and nobody can have the lock
1365 # since before our call to delete()).
1367 # This is done in an else clause because if the exception was thrown
1368 # it's the job of the one who actually deleted it.
1369 del self.__lockdict[lname]
1370 # And let's remove it from our private list if we owned it.
1371 if self._is_owned():
1372 self._del_owned(name=lname)
1377 # Locking levels, must be acquired in increasing order.
1378 # Current rules are:
1379 # - at level LEVEL_CLUSTER resides the Big Ganeti Lock (BGL) which must be
1380 # acquired before performing any operation, either in shared or in exclusive
1381 # mode. acquiring the BGL in exclusive mode is discouraged and should be
1383 # - at levels LEVEL_NODE and LEVEL_INSTANCE reside node and instance locks.
1384 # If you need more than one node, or more than one instance, acquire them at
1391 LEVELS = [LEVEL_CLUSTER,
1396 # Lock levels which are modifiable
1397 LEVELS_MOD = [LEVEL_NODE, LEVEL_NODEGROUP, LEVEL_INSTANCE]
1400 LEVEL_CLUSTER: "cluster",
1401 LEVEL_INSTANCE: "instance",
1402 LEVEL_NODEGROUP: "nodegroup",
1406 # Constant for the big ganeti lock
1410 class GanetiLockManager:
1411 """The Ganeti Locking Library
1413 The purpose of this small library is to manage locking for ganeti clusters
1414 in a central place, while at the same time doing dynamic checks against
1415 possible deadlocks. It will also make it easier to transition to a different
1416 lock type should we migrate away from python threads.
1421 def __init__(self, nodes, nodegroups, instances):
1422 """Constructs a new GanetiLockManager object.
1424 There should be only a GanetiLockManager object at any time, so this
1425 function raises an error if this is not the case.
1427 @param nodes: list of node names
1428 @param nodegroups: list of nodegroup uuids
1429 @param instances: list of instance names
1432 assert self.__class__._instance is None, \
1433 "double GanetiLockManager instance"
1435 self.__class__._instance = self
1437 self._monitor = LockMonitor()
1439 # The keyring contains all the locks, at their level and in the correct
1442 LEVEL_CLUSTER: LockSet([BGL], "BGL", monitor=self._monitor),
1443 LEVEL_NODE: LockSet(nodes, "nodes", monitor=self._monitor),
1444 LEVEL_NODEGROUP: LockSet(nodegroups, "nodegroups", monitor=self._monitor),
1445 LEVEL_INSTANCE: LockSet(instances, "instances",
1446 monitor=self._monitor),
1449 def QueryLocks(self, fields):
1450 """Queries information from all locks.
1452 See L{LockMonitor.QueryLocks}.
1455 return self._monitor.QueryLocks(fields)
1457 def OldStyleQueryLocks(self, fields):
1458 """Queries information from all locks, returning old-style data.
1460 See L{LockMonitor.OldStyleQueryLocks}.
1463 return self._monitor.OldStyleQueryLocks(fields)
1465 def _names(self, level):
1466 """List the lock names at the given level.
1468 This can be used for debugging/testing purposes.
1470 @param level: the level whose list of locks to get
1473 assert level in LEVELS, "Invalid locking level %s" % level
1474 return self.__keyring[level]._names()
1476 def _is_owned(self, level):
1477 """Check whether we are owning locks at the given level
1480 return self.__keyring[level]._is_owned()
1482 is_owned = _is_owned
1484 def _list_owned(self, level):
1485 """Get the set of owned locks at the given level
1488 return self.__keyring[level]._list_owned()
1490 list_owned = _list_owned
1492 def _upper_owned(self, level):
1493 """Check that we don't own any lock at a level greater than the given one.
1496 # This way of checking only works if LEVELS[i] = i, which we check for in
1498 return compat.any((self._is_owned(l) for l in LEVELS[level + 1:]))
1500 def _BGL_owned(self): # pylint: disable-msg=C0103
1501 """Check if the current thread owns the BGL.
1503 Both an exclusive or a shared acquisition work.
1506 return BGL in self.__keyring[LEVEL_CLUSTER]._list_owned()
1509 def _contains_BGL(level, names): # pylint: disable-msg=C0103
1510 """Check if the level contains the BGL.
1512 Check if acting on the given level and set of names will change
1513 the status of the Big Ganeti Lock.
1516 return level == LEVEL_CLUSTER and (names is None or BGL in names)
1518 def acquire(self, level, names, timeout=None, shared=0, priority=None):
1519 """Acquire a set of resource locks, at the same level.
1521 @type level: member of locking.LEVELS
1522 @param level: the level at which the locks shall be acquired
1523 @type names: list of strings (or string)
1524 @param names: the names of the locks which shall be acquired
1525 (special lock names, or instance/node names)
1526 @type shared: integer (0/1) used as a boolean
1527 @param shared: whether to acquire in shared mode; by default
1528 an exclusive lock will be acquired
1529 @type timeout: float
1530 @param timeout: Maximum time to acquire all locks
1531 @type priority: integer
1532 @param priority: Priority for acquiring lock
1535 assert level in LEVELS, "Invalid locking level %s" % level
1537 # Check that we are either acquiring the Big Ganeti Lock or we already own
1538 # it. Some "legacy" opcodes need to be sure they are run non-concurrently
1539 # so even if we've migrated we need to at least share the BGL to be
1540 # compatible with them. Of course if we own the BGL exclusively there's no
1541 # point in acquiring any other lock, unless perhaps we are half way through
1542 # the migration of the current opcode.
1543 assert (self._contains_BGL(level, names) or self._BGL_owned()), (
1544 "You must own the Big Ganeti Lock before acquiring any other")
1546 # Check we don't own locks at the same or upper levels.
1547 assert not self._upper_owned(level), ("Cannot acquire locks at a level"
1548 " while owning some at a greater one")
1550 # Acquire the locks in the set.
1551 return self.__keyring[level].acquire(names, shared=shared, timeout=timeout,
1554 def downgrade(self, level, names=None):
1555 """Downgrade a set of resource locks from exclusive to shared mode.
1557 You must have acquired the locks in exclusive mode.
1559 @type level: member of locking.LEVELS
1560 @param level: the level at which the locks shall be downgraded
1561 @type names: list of strings, or None
1562 @param names: the names of the locks which shall be downgraded
1563 (defaults to all the locks acquired at the level)
1566 assert level in LEVELS, "Invalid locking level %s" % level
1568 return self.__keyring[level].downgrade(names=names)
1570 def release(self, level, names=None):
1571 """Release a set of resource locks, at the same level.
1573 You must have acquired the locks, either in shared or in exclusive
1574 mode, before releasing them.
1576 @type level: member of locking.LEVELS
1577 @param level: the level at which the locks shall be released
1578 @type names: list of strings, or None
1579 @param names: the names of the locks which shall be released
1580 (defaults to all the locks acquired at that level)
1583 assert level in LEVELS, "Invalid locking level %s" % level
1584 assert (not self._contains_BGL(level, names) or
1585 not self._upper_owned(LEVEL_CLUSTER)), (
1586 "Cannot release the Big Ganeti Lock while holding something"
1587 " at upper levels (%r)" %
1588 (utils.CommaJoin(["%s=%r" % (LEVEL_NAMES[i], self._list_owned(i))
1589 for i in self.__keyring.keys()]), ))
1591 # Release will complain if we don't own the locks already
1592 return self.__keyring[level].release(names)
1594 def add(self, level, names, acquired=0, shared=0):
1595 """Add locks at the specified level.
1597 @type level: member of locking.LEVELS_MOD
1598 @param level: the level at which the locks shall be added
1599 @type names: list of strings
1600 @param names: names of the locks to acquire
1601 @type acquired: integer (0/1) used as a boolean
1602 @param acquired: whether to acquire the newly added locks
1603 @type shared: integer (0/1) used as a boolean
1604 @param shared: whether the acquisition will be shared
1607 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
1608 assert self._BGL_owned(), ("You must own the BGL before performing other"
1610 assert not self._upper_owned(level), ("Cannot add locks at a level"
1611 " while owning some at a greater one")
1612 return self.__keyring[level].add(names, acquired=acquired, shared=shared)
1614 def remove(self, level, names):
1615 """Remove locks from the specified level.
1617 You must either already own the locks you are trying to remove
1618 exclusively or not own any lock at an upper level.
1620 @type level: member of locking.LEVELS_MOD
1621 @param level: the level at which the locks shall be removed
1622 @type names: list of strings
1623 @param names: the names of the locks which shall be removed
1624 (special lock names, or instance/node names)
1627 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
1628 assert self._BGL_owned(), ("You must own the BGL before performing other"
1630 # Check we either own the level or don't own anything from here
1631 # up. LockSet.remove() will check the case in which we don't own
1632 # all the needed resources, or we have a shared ownership.
1633 assert self._is_owned(level) or not self._upper_owned(level), (
1634 "Cannot remove locks at a level while not owning it or"
1635 " owning some at a greater one")
1636 return self.__keyring[level].remove(names)
1639 def _MonitorSortKey((num, item)):
1640 """Sorting key function.
1642 Sort by name, then by incoming order.
1645 (name, _, _, _) = item
1647 return (utils.NiceSortKey(name), num)
1650 class LockMonitor(object):
1651 _LOCK_ATTR = "_lock"
1654 """Initializes this class.
1657 self._lock = SharedLock("LockMonitor")
1659 # Counter for stable sorting
1660 self._counter = itertools.count(0)
1662 # Tracked locks. Weak references are used to avoid issues with circular
1663 # references and deletion.
1664 self._locks = weakref.WeakKeyDictionary()
1666 @ssynchronized(_LOCK_ATTR)
1667 def RegisterLock(self, lock):
1668 """Registers a new lock.
1671 logging.debug("Registering lock %s", lock.name)
1672 assert lock not in self._locks, "Duplicate lock registration"
1674 # There used to be a check for duplicate names here. As it turned out, when
1675 # a lock is re-created with the same name in a very short timeframe, the
1676 # previous instance might not yet be removed from the weakref dictionary.
1677 # By keeping track of the order of incoming registrations, a stable sort
1678 # ordering can still be guaranteed.
1680 self._locks[lock] = self._counter.next()
1682 @ssynchronized(_LOCK_ATTR)
1683 def _GetLockInfo(self, requested):
1684 """Get information from all locks while the monitor lock is held.
1687 return [(num, lock.GetInfo(requested)) for lock, num in self._locks.items()]
1689 def _Query(self, fields):
1690 """Queries information from all locks.
1692 @type fields: list of strings
1693 @param fields: List of fields to return
1696 qobj = query.Query(query.LOCK_FIELDS, fields)
1698 # Get all data with internal lock held and then sort by name and incoming
1700 lockinfo = sorted(self._GetLockInfo(qobj.RequestedData()),
1701 key=_MonitorSortKey)
1703 # Extract lock information and build query data
1704 return (qobj, query.LockQueryData(map(operator.itemgetter(1), lockinfo)))
1706 def QueryLocks(self, fields):
1707 """Queries information from all locks.
1709 @type fields: list of strings
1710 @param fields: List of fields to return
1713 (qobj, ctx) = self._Query(fields)
1715 # Prepare query response
1716 return query.GetQueryResponse(qobj, ctx)
1718 def OldStyleQueryLocks(self, fields):
1719 """Queries information from all locks, returning old-style data.
1721 @type fields: list of strings
1722 @param fields: List of fields to return
1725 (qobj, ctx) = self._Query(fields)
1727 return qobj.OldStyleQuery(ctx)