4 # Copyright (C) 2006, 2007, 2008, 2009, 2010, 2011 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=W0212
25 # W0212 since e.g. LockSet methods use (a lot) the internals of
37 from ganeti import errors
38 from ganeti import utils
39 from ganeti import compat
40 from ganeti import query
43 _EXCLUSIVE_TEXT = "exclusive"
44 _SHARED_TEXT = "shared"
45 _DELETED_TEXT = "deleted"
50 def ssynchronized(mylock, shared=0):
51 """Shared Synchronization decorator.
53 Calls the function holding the given lock, either in exclusive or shared
54 mode. It requires the passed lock to be a SharedLock (or support its
57 @type mylock: lockable object or string
58 @param mylock: lock to acquire or class member name of the lock to acquire
62 def sync_function(*args, **kwargs):
63 if isinstance(mylock, basestring):
64 assert args, "cannot ssynchronize on non-class method: self not found"
66 lock = getattr(args[0], mylock)
69 lock.acquire(shared=shared)
71 return fn(*args, **kwargs)
78 class _SingleNotifyPipeConditionWaiter(object):
79 """Helper class for SingleNotifyPipeCondition
87 def __init__(self, poller, fd):
88 """Constructor for _SingleNotifyPipeConditionWaiter
90 @type poller: select.poll
91 @param poller: Poller object
93 @param fd: File descriptor to wait for
100 def __call__(self, timeout):
101 """Wait for something to happen on the pipe.
103 @type timeout: float or None
104 @param timeout: Timeout for waiting (can be None)
107 running_timeout = utils.RunningTimeout(timeout, True)
110 remaining_time = running_timeout.Remaining()
112 if remaining_time is not None:
113 if remaining_time < 0.0:
116 # Our calculation uses seconds, poll() wants milliseconds
117 remaining_time *= 1000
120 result = self._poller.poll(remaining_time)
121 except EnvironmentError, err:
122 if err.errno != errno.EINTR:
126 # Check whether we were notified
127 if result and result[0][0] == self._fd:
131 class _BaseCondition(object):
132 """Base class containing common code for conditions.
134 Some of this code is taken from python's threading module.
146 def __init__(self, lock):
147 """Constructor for _BaseCondition.
149 @type lock: threading.Lock
150 @param lock: condition base lock
153 object.__init__(self)
156 self._release_save = lock._release_save
157 except AttributeError:
158 self._release_save = self._base_release_save
160 self._acquire_restore = lock._acquire_restore
161 except AttributeError:
162 self._acquire_restore = self._base_acquire_restore
164 self._is_owned = lock.is_owned
165 except AttributeError:
166 self._is_owned = self._base_is_owned
170 # Export the lock's acquire() and release() methods
171 self.acquire = lock.acquire
172 self.release = lock.release
174 def _base_is_owned(self):
175 """Check whether lock is owned by current thread.
178 if self._lock.acquire(0):
183 def _base_release_save(self):
186 def _base_acquire_restore(self, _):
189 def _check_owned(self):
190 """Raise an exception if the current thread doesn't own the lock.
193 if not self._is_owned():
194 raise RuntimeError("cannot work with un-aquired lock")
197 class SingleNotifyPipeCondition(_BaseCondition):
198 """Condition which can only be notified once.
200 This condition class uses pipes and poll, internally, to be able to wait for
201 notification with a timeout, without resorting to polling. It is almost
202 compatible with Python's threading.Condition, with the following differences:
203 - notifyAll can only be called once, and no wait can happen after that
204 - notify is not supported, only notifyAll
216 _waiter_class = _SingleNotifyPipeConditionWaiter
218 def __init__(self, lock):
219 """Constructor for SingleNotifyPipeCondition
222 _BaseCondition.__init__(self, lock)
224 self._notified = False
226 self._write_fd = None
229 def _check_unnotified(self):
230 """Throws an exception if already notified.
234 raise RuntimeError("cannot use already notified condition")
237 """Cleanup open file descriptors, if any.
240 if self._read_fd is not None:
241 os.close(self._read_fd)
244 if self._write_fd is not None:
245 os.close(self._write_fd)
246 self._write_fd = None
249 def wait(self, timeout):
250 """Wait for a notification.
252 @type timeout: float or None
253 @param timeout: Waiting timeout (can be None)
257 self._check_unnotified()
261 if self._poller is None:
262 (self._read_fd, self._write_fd) = os.pipe()
263 self._poller = select.poll()
264 self._poller.register(self._read_fd, select.POLLHUP)
266 wait_fn = self._waiter_class(self._poller, self._read_fd)
267 state = self._release_save()
269 # Wait for notification
273 self._acquire_restore(state)
276 if self._nwaiters == 0:
279 def notifyAll(self): # pylint: disable=C0103
280 """Close the writing side of the pipe to notify all waiters.
284 self._check_unnotified()
285 self._notified = True
286 if self._write_fd is not None:
287 os.close(self._write_fd)
288 self._write_fd = None
291 class PipeCondition(_BaseCondition):
292 """Group-only non-polling condition with counters.
294 This condition class uses pipes and poll, internally, to be able to wait for
295 notification with a timeout, without resorting to polling. It is almost
296 compatible with Python's threading.Condition, but only supports notifyAll and
297 non-recursive locks. As an additional features it's able to report whether
298 there are any waiting threads.
306 _single_condition_class = SingleNotifyPipeCondition
308 def __init__(self, lock):
309 """Initializes this class.
312 _BaseCondition.__init__(self, lock)
313 self._waiters = set()
314 self._single_condition = self._single_condition_class(self._lock)
316 def wait(self, timeout):
317 """Wait for a notification.
319 @type timeout: float or None
320 @param timeout: Waiting timeout (can be None)
325 # Keep local reference to the pipe. It could be replaced by another thread
326 # notifying while we're waiting.
327 cond = self._single_condition
329 self._waiters.add(threading.currentThread())
334 self._waiters.remove(threading.currentThread())
336 def notifyAll(self): # pylint: disable=C0103
337 """Notify all currently waiting threads.
341 self._single_condition.notifyAll()
342 self._single_condition = self._single_condition_class(self._lock)
344 def get_waiting(self):
345 """Returns a list of all waiting threads.
352 def has_waiting(self):
353 """Returns whether there are active waiters.
358 return bool(self._waiters)
361 class _PipeConditionWithMode(PipeCondition):
366 def __init__(self, lock, shared):
367 """Initializes this class.
371 PipeCondition.__init__(self, lock)
374 class SharedLock(object):
375 """Implements a shared lock.
377 Multiple threads can acquire the lock in a shared way by calling
378 C{acquire(shared=1)}. In order to acquire the lock in an exclusive way
379 threads can call C{acquire(shared=0)}.
381 Notes on data structures: C{__pending} contains a priority queue (heapq) of
382 all pending acquires: C{[(priority1: prioqueue1), (priority2: prioqueue2),
383 ...]}. Each per-priority queue contains a normal in-order list of conditions
384 to be notified when the lock can be acquired. Shared locks are grouped
385 together by priority and the condition for them is stored in
386 C{__pending_shared} if it already exists. C{__pending_by_prio} keeps
387 references for the per-priority queues indexed by priority for faster access.
390 @ivar name: the name of the lock
405 __condition_class = _PipeConditionWithMode
407 def __init__(self, name, monitor=None):
408 """Construct a new SharedLock.
410 @param name: the name of the lock
411 @type monitor: L{LockMonitor}
412 @param monitor: Lock monitor with which to register
415 object.__init__(self)
420 self.__lock = threading.Lock()
422 # Queue containing waiting acquires
424 self.__pending_by_prio = {}
425 self.__pending_shared = {}
427 # Current lock holders
431 # is this lock in the deleted state?
432 self.__deleted = False
434 # Register with lock monitor
436 logging.debug("Adding lock %s to monitor", name)
437 monitor.RegisterLock(self)
439 def GetLockInfo(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 #: Necessary to remain compatible with threading.Condition, which tries to
545 #: retrieve a locks' "_is_owned" attribute
548 def _count_pending(self):
549 """Returns the number of pending acquires.
554 self.__lock.acquire()
556 return sum(len(prioqueue) for (_, prioqueue) in self.__pending)
558 self.__lock.release()
560 def _check_empty(self):
561 """Checks whether there are any pending acquires.
566 self.__lock.acquire()
568 # Order is important: __find_first_pending_queue modifies __pending
569 (_, prioqueue) = self.__find_first_pending_queue()
571 return not (prioqueue or
573 self.__pending_by_prio or
574 self.__pending_shared)
576 self.__lock.release()
578 def __do_acquire(self, shared):
579 """Actually acquire the lock.
583 self.__shr.add(threading.currentThread())
585 self.__exc = threading.currentThread()
587 def __can_acquire(self, shared):
588 """Determine whether lock can be acquired.
592 return self.__exc is None
594 return len(self.__shr) == 0 and self.__exc is None
596 def __find_first_pending_queue(self):
597 """Tries to find the topmost queued entry with pending acquires.
599 Removes empty entries while going through the list.
602 while self.__pending:
603 (priority, prioqueue) = self.__pending[0]
606 return (priority, prioqueue)
609 heapq.heappop(self.__pending)
610 del self.__pending_by_prio[priority]
611 assert priority not in self.__pending_shared
615 def __is_on_top(self, cond):
616 """Checks whether the passed condition is on top of the queue.
618 The caller must make sure the queue isn't empty.
621 (_, prioqueue) = self.__find_first_pending_queue()
623 return cond == prioqueue[0]
625 def __acquire_unlocked(self, shared, timeout, priority):
626 """Acquire a shared lock.
628 @param shared: whether to acquire in shared mode; by default an
629 exclusive lock will be acquired
630 @param timeout: maximum waiting time before giving up
631 @type priority: integer
632 @param priority: Priority for acquiring lock
635 self.__check_deleted()
637 # We cannot acquire the lock if we already have it
638 assert not self.__is_owned(), ("double acquire() on a non-recursive lock"
641 # Remove empty entries from queue
642 self.__find_first_pending_queue()
644 # Check whether someone else holds the lock or there are pending acquires.
645 if not self.__pending and self.__can_acquire(shared):
646 # Apparently not, can acquire lock directly.
647 self.__do_acquire(shared)
650 prioqueue = self.__pending_by_prio.get(priority, None)
653 # Try to re-use condition for shared acquire
654 wait_condition = self.__pending_shared.get(priority, None)
655 assert (wait_condition is None or
656 (wait_condition.shared and wait_condition in prioqueue))
658 wait_condition = None
660 if wait_condition is None:
661 if prioqueue is None:
662 assert priority not in self.__pending_by_prio
665 heapq.heappush(self.__pending, (priority, prioqueue))
666 self.__pending_by_prio[priority] = prioqueue
668 wait_condition = self.__condition_class(self.__lock, shared)
669 prioqueue.append(wait_condition)
672 # Keep reference for further shared acquires on same priority. This is
673 # better than trying to find it in the list of pending acquires.
674 assert priority not in self.__pending_shared
675 self.__pending_shared[priority] = wait_condition
678 # Wait until we become the topmost acquire in the queue or the timeout
680 # TODO: Decrease timeout with spurious notifications
681 while not (self.__is_on_top(wait_condition) and
682 self.__can_acquire(shared)):
683 # Wait for notification
684 wait_condition.wait(timeout)
685 self.__check_deleted()
687 # A lot of code assumes blocking acquires always succeed. Loop
688 # internally for that case.
689 if timeout is not None:
692 if self.__is_on_top(wait_condition) and self.__can_acquire(shared):
693 self.__do_acquire(shared)
696 # Remove condition from queue if there are no more waiters
697 if not wait_condition.has_waiting():
698 prioqueue.remove(wait_condition)
699 if wait_condition.shared:
700 # Remove from list of shared acquires if it wasn't while releasing
701 # (e.g. on lock deletion)
702 self.__pending_shared.pop(priority, None)
706 def acquire(self, shared=0, timeout=None, priority=None,
708 """Acquire a shared lock.
710 @type shared: integer (0/1) used as a boolean
711 @param shared: whether to acquire in shared mode; by default an
712 exclusive lock will be acquired
714 @param timeout: maximum waiting time before giving up
715 @type priority: integer
716 @param priority: Priority for acquiring lock
717 @type test_notify: callable or None
718 @param test_notify: Special callback function for unittesting
722 priority = _DEFAULT_PRIORITY
724 self.__lock.acquire()
726 # We already got the lock, notify now
727 if __debug__ and callable(test_notify):
730 return self.__acquire_unlocked(shared, timeout, priority)
732 self.__lock.release()
735 """Changes the lock mode from exclusive to shared.
737 Pending acquires in shared mode on the same priority will go ahead.
740 self.__lock.acquire()
742 assert self.__is_owned(), "Lock must be owned"
744 if self.__is_exclusive():
745 # Do nothing if the lock is already acquired in shared mode
749 # Important: pending shared acquires should only jump ahead if there
750 # was a transition from exclusive to shared, otherwise an owner of a
751 # shared lock can keep calling this function to push incoming shared
753 (priority, prioqueue) = self.__find_first_pending_queue()
755 # Is there a pending shared acquire on this priority?
756 cond = self.__pending_shared.pop(priority, None)
759 assert cond in prioqueue
761 # Ensure shared acquire is on top of queue
762 if len(prioqueue) > 1:
763 prioqueue.remove(cond)
764 prioqueue.insert(0, cond)
769 assert not self.__is_exclusive()
770 assert self.__is_sharer()
774 self.__lock.release()
777 """Release a Shared Lock.
779 You must have acquired the lock, either in shared or in exclusive mode,
780 before calling this function.
783 self.__lock.acquire()
785 assert self.__is_exclusive() or self.__is_sharer(), \
786 "Cannot release non-owned lock"
788 # Autodetect release type
789 if self.__is_exclusive():
792 self.__shr.remove(threading.currentThread())
794 # Notify topmost condition in queue
795 (priority, prioqueue) = self.__find_first_pending_queue()
800 # Prevent further shared acquires from sneaking in while waiters are
802 self.__pending_shared.pop(priority, None)
805 self.__lock.release()
807 def delete(self, timeout=None, priority=None):
808 """Delete a Shared Lock.
810 This operation will declare the lock for removal. First the lock will be
811 acquired in exclusive mode if you don't already own it, then the lock
812 will be put in a state where any future and pending acquire() fail.
815 @param timeout: maximum waiting time before giving up
816 @type priority: integer
817 @param priority: Priority for acquiring lock
821 priority = _DEFAULT_PRIORITY
823 self.__lock.acquire()
825 assert not self.__is_sharer(), "Cannot delete() a lock while sharing it"
827 self.__check_deleted()
829 # The caller is allowed to hold the lock exclusively already.
830 acquired = self.__is_exclusive()
833 acquired = self.__acquire_unlocked(0, timeout, priority)
835 assert self.__is_exclusive() and not self.__is_sharer(), \
836 "Lock wasn't acquired in exclusive mode"
839 self.__deleted = True
842 assert not (self.__exc or self.__shr), "Found owner during deletion"
844 # Notify all acquires. They'll throw an error.
845 for (_, prioqueue) in self.__pending:
846 for cond in prioqueue:
849 assert self.__deleted
853 self.__lock.release()
855 def _release_save(self):
856 shared = self.__is_sharer()
860 def _acquire_restore(self, shared):
861 self.acquire(shared=shared)
864 # Whenever we want to acquire a full LockSet we pass None as the value
865 # to acquire. Hide this behind this nicely named constant.
869 class _AcquireTimeout(Exception):
870 """Internal exception to abort an acquire on a timeout.
876 """Implements a set of locks.
878 This abstraction implements a set of shared locks for the same resource type,
879 distinguished by name. The user can lock a subset of the resources and the
880 LockSet will take care of acquiring the locks always in the same order, thus
883 All the locks needed in the same set must be acquired together, though.
886 @ivar name: the name of the lockset
889 def __init__(self, members, name, monitor=None):
890 """Constructs a new LockSet.
892 @type members: list of strings
893 @param members: initial members of the set
894 @type monitor: L{LockMonitor}
895 @param monitor: Lock monitor with which to register member locks
898 assert members is not None, "members parameter is not a list"
902 self.__monitor = monitor
904 # Used internally to guarantee coherency
905 self.__lock = SharedLock(self._GetLockName("[lockset]"), monitor=monitor)
907 # The lockdict indexes the relationship name -> lock
908 # The order-of-locking is implied by the alphabetical order of names
911 for mname in members:
912 self.__lockdict[mname] = SharedLock(self._GetLockName(mname),
915 # The owner dict contains the set of locks each thread owns. For
916 # performance each thread can access its own key without a global lock on
917 # this structure. It is paramount though that *no* other type of access is
918 # done to this structure (eg. no looping over its keys). *_owner helper
919 # function are defined to guarantee access is correct, but in general never
920 # do anything different than __owners[threading.currentThread()], or there
924 def _GetLockName(self, mname):
925 """Returns the name for a member lock.
928 return "%s/%s" % (self.name, mname)
931 """Returns the lockset-internal lock.
936 def _get_lockdict(self):
937 """Returns the lockset-internal lock dictionary.
939 Accessing this structure is only safe in single-thread usage or when the
940 lockset-internal lock is held.
943 return self.__lockdict
946 """Is the current thread a current level owner?"""
947 return threading.currentThread() in self.__owners
949 def _add_owned(self, name=None):
950 """Note the current thread owns the given lock"""
952 if not self.is_owned():
953 self.__owners[threading.currentThread()] = set()
956 self.__owners[threading.currentThread()].add(name)
958 self.__owners[threading.currentThread()] = set([name])
960 def _del_owned(self, name=None):
961 """Note the current thread owns the given lock"""
963 assert not (name is None and self.__lock.is_owned()), \
964 "Cannot hold internal lock when deleting owner status"
967 self.__owners[threading.currentThread()].remove(name)
969 # Only remove the key if we don't hold the set-lock as well
970 if (not self.__lock.is_owned() and
971 not self.__owners[threading.currentThread()]):
972 del self.__owners[threading.currentThread()]
974 def list_owned(self):
975 """Get the set of resource names owned by the current thread"""
977 return self.__owners[threading.currentThread()].copy()
981 def _release_and_delete_owned(self):
982 """Release and delete all resources owned by the current thread"""
983 for lname in self.list_owned():
984 lock = self.__lockdict[lname]
987 self._del_owned(name=lname)
990 """Return the current set of names.
992 Only call this function while holding __lock and don't iterate on the
993 result after releasing the lock.
996 return self.__lockdict.keys()
999 """Return a copy of the current set of elements.
1001 Used only for debugging purposes.
1004 # If we don't already own the set-level lock acquired
1005 # we'll get it and note we need to release it later.
1006 release_lock = False
1007 if not self.__lock.is_owned():
1009 self.__lock.acquire(shared=1)
1011 result = self.__names()
1014 self.__lock.release()
1017 def acquire(self, names, timeout=None, shared=0, priority=None,
1019 """Acquire a set of resource locks.
1021 @type names: list of strings (or string)
1022 @param names: the names of the locks which shall be acquired
1023 (special lock names, or instance/node names)
1024 @type shared: integer (0/1) used as a boolean
1025 @param shared: whether to acquire in shared mode; by default an
1026 exclusive lock will be acquired
1027 @type timeout: float or None
1028 @param timeout: Maximum time to acquire all locks
1029 @type priority: integer
1030 @param priority: Priority for acquiring locks
1031 @type test_notify: callable or None
1032 @param test_notify: Special callback function for unittesting
1034 @return: Set of all locks successfully acquired or None in case of timeout
1036 @raise errors.LockError: when any lock we try to acquire has
1037 been deleted before we succeed. In this case none of the
1038 locks requested will be acquired.
1041 assert timeout is None or timeout >= 0.0
1043 # Check we don't already own locks at this level
1044 assert not self.is_owned(), ("Cannot acquire locks in the same set twice"
1045 " (lockset %s)" % self.name)
1047 if priority is None:
1048 priority = _DEFAULT_PRIORITY
1050 # We need to keep track of how long we spent waiting for a lock. The
1051 # timeout passed to this function is over all lock acquires.
1052 running_timeout = utils.RunningTimeout(timeout, False)
1055 if names is not None:
1056 # Support passing in a single resource to acquire rather than many
1057 if isinstance(names, basestring):
1060 return self.__acquire_inner(names, False, shared, priority,
1061 running_timeout.Remaining, test_notify)
1064 # If no names are given acquire the whole set by not letting new names
1065 # being added before we release, and getting the current list of names.
1066 # Some of them may then be deleted later, but we'll cope with this.
1068 # We'd like to acquire this lock in a shared way, as it's nice if
1069 # everybody else can use the instances at the same time. If we are
1070 # acquiring them exclusively though they won't be able to do this
1071 # anyway, though, so we'll get the list lock exclusively as well in
1072 # order to be able to do add() on the set while owning it.
1073 if not self.__lock.acquire(shared=shared, priority=priority,
1074 timeout=running_timeout.Remaining()):
1075 raise _AcquireTimeout()
1077 # note we own the set-lock
1080 return self.__acquire_inner(self.__names(), True, shared, priority,
1081 running_timeout.Remaining, test_notify)
1083 # We shouldn't have problems adding the lock to the owners list, but
1084 # if we did we'll try to release this lock and re-raise exception.
1085 # Of course something is going to be really wrong, after this.
1086 self.__lock.release()
1090 except _AcquireTimeout:
1093 def __acquire_inner(self, names, want_all, shared, priority,
1094 timeout_fn, test_notify):
1095 """Inner logic for acquiring a number of locks.
1097 @param names: Names of the locks to be acquired
1098 @param want_all: Whether all locks in the set should be acquired
1099 @param shared: Whether to acquire in shared mode
1100 @param timeout_fn: Function returning remaining timeout
1101 @param priority: Priority for acquiring locks
1102 @param test_notify: Special callback function for unittesting
1107 # First we look the locks up on __lockdict. We have no way of being sure
1108 # they will still be there after, but this makes it a lot faster should
1109 # just one of them be the already wrong. Using a sorted sequence to prevent
1111 for lname in sorted(utils.UniqueSequence(names)):
1113 lock = self.__lockdict[lname] # raises KeyError if lock is not there
1116 # We are acquiring all the set, it doesn't matter if this particular
1117 # element is not there anymore.
1120 raise errors.LockError("Non-existing lock %s in set %s (it may have"
1121 " been removed)" % (lname, self.name))
1123 acquire_list.append((lname, lock))
1125 # This will hold the locknames we effectively acquired.
1129 # Now acquire_list contains a sorted list of resources and locks we
1130 # want. In order to get them we loop on this (private) list and
1131 # acquire() them. We gave no real guarantee they will still exist till
1132 # this is done but .acquire() itself is safe and will alert us if the
1133 # lock gets deleted.
1134 for (lname, lock) in acquire_list:
1135 if __debug__ and callable(test_notify):
1136 test_notify_fn = lambda: test_notify(lname)
1138 test_notify_fn = None
1140 timeout = timeout_fn()
1143 # raises LockError if the lock was deleted
1144 acq_success = lock.acquire(shared=shared, timeout=timeout,
1146 test_notify=test_notify_fn)
1147 except errors.LockError:
1149 # We are acquiring all the set, it doesn't matter if this
1150 # particular element is not there anymore.
1153 raise errors.LockError("Non-existing lock %s in set %s (it may"
1154 " have been removed)" % (lname, self.name))
1157 # Couldn't get lock or timeout occurred
1159 # This shouldn't happen as SharedLock.acquire(timeout=None) is
1161 raise errors.LockError("Failed to get lock %s (set %s)" %
1164 raise _AcquireTimeout()
1167 # now the lock cannot be deleted, we have it!
1168 self._add_owned(name=lname)
1172 # We shouldn't have problems adding the lock to the owners list, but
1173 # if we did we'll try to release this lock and re-raise exception.
1174 # Of course something is going to be really wrong after this.
1180 # Release all owned locks
1181 self._release_and_delete_owned()
1186 def downgrade(self, names=None):
1187 """Downgrade a set of resource locks from exclusive to shared mode.
1189 The locks must have been acquired in exclusive mode.
1192 assert self.is_owned(), ("downgrade on lockset %s while not owning any"
1193 " lock" % self.name)
1195 # Support passing in a single resource to downgrade rather than many
1196 if isinstance(names, basestring):
1199 owned = self.list_owned()
1205 assert owned.issuperset(names), \
1206 ("downgrade() on unheld resources %s (set %s)" %
1207 (names.difference(owned), self.name))
1209 for lockname in names:
1210 self.__lockdict[lockname].downgrade()
1212 # Do we own the lockset in exclusive mode?
1213 if self.__lock.is_owned(shared=0):
1214 # Have all locks been downgraded?
1215 if not compat.any(lock.is_owned(shared=0)
1216 for lock in self.__lockdict.values()):
1217 self.__lock.downgrade()
1218 assert self.__lock.is_owned(shared=1)
1222 def release(self, names=None):
1223 """Release a set of resource locks, at the same level.
1225 You must have acquired the locks, either in shared or in exclusive mode,
1226 before releasing them.
1228 @type names: list of strings, or None
1229 @param names: the names of the locks which shall be released
1230 (defaults to all the locks acquired at that level).
1233 assert self.is_owned(), ("release() on lock set %s while not owner" %
1236 # Support passing in a single resource to release rather than many
1237 if isinstance(names, basestring):
1241 names = self.list_owned()
1244 assert self.list_owned().issuperset(names), (
1245 "release() on unheld resources %s (set %s)" %
1246 (names.difference(self.list_owned()), self.name))
1248 # First of all let's release the "all elements" lock, if set.
1249 # After this 'add' can work again
1250 if self.__lock.is_owned():
1251 self.__lock.release()
1254 for lockname in names:
1255 # If we are sure the lock doesn't leave __lockdict without being
1256 # exclusively held we can do this...
1257 self.__lockdict[lockname].release()
1258 self._del_owned(name=lockname)
1260 def add(self, names, acquired=0, shared=0):
1261 """Add a new set of elements to the set
1263 @type names: list of strings
1264 @param names: names of the new elements to add
1265 @type acquired: integer (0/1) used as a boolean
1266 @param acquired: pre-acquire the new resource?
1267 @type shared: integer (0/1) used as a boolean
1268 @param shared: is the pre-acquisition shared?
1271 # Check we don't already own locks at this level
1272 assert not self.is_owned() or self.__lock.is_owned(shared=0), \
1273 ("Cannot add locks if the set %s is only partially owned, or shared" %
1276 # Support passing in a single resource to add rather than many
1277 if isinstance(names, basestring):
1280 # If we don't already own the set-level lock acquired in an exclusive way
1281 # we'll get it and note we need to release it later.
1282 release_lock = False
1283 if not self.__lock.is_owned():
1285 self.__lock.acquire()
1288 invalid_names = set(self.__names()).intersection(names)
1290 # This must be an explicit raise, not an assert, because assert is
1291 # turned off when using optimization, and this can happen because of
1292 # concurrency even if the user doesn't want it.
1293 raise errors.LockError("duplicate add(%s) on lockset %s" %
1294 (invalid_names, self.name))
1296 for lockname in names:
1297 lock = SharedLock(self._GetLockName(lockname), monitor=self.__monitor)
1300 # No need for priority or timeout here as this lock has just been
1302 lock.acquire(shared=shared)
1303 # now the lock cannot be deleted, we have it!
1305 self._add_owned(name=lockname)
1307 # We shouldn't have problems adding the lock to the owners list,
1308 # but if we did we'll try to release this lock and re-raise
1309 # exception. Of course something is going to be really wrong,
1310 # after this. On the other hand the lock hasn't been added to the
1311 # __lockdict yet so no other threads should be pending on it. This
1312 # release is just a safety measure.
1316 self.__lockdict[lockname] = lock
1319 # Only release __lock if we were not holding it previously.
1321 self.__lock.release()
1325 def remove(self, names):
1326 """Remove elements from the lock set.
1328 You can either not hold anything in the lockset or already hold a superset
1329 of the elements you want to delete, exclusively.
1331 @type names: list of strings
1332 @param names: names of the resource to remove.
1334 @return: a list of locks which we removed; the list is always
1335 equal to the names list if we were holding all the locks
1339 # Support passing in a single resource to remove rather than many
1340 if isinstance(names, basestring):
1343 # If we own any subset of this lock it must be a superset of what we want
1344 # to delete. The ownership must also be exclusive, but that will be checked
1345 # by the lock itself.
1346 assert not self.is_owned() or self.list_owned().issuperset(names), (
1347 "remove() on acquired lockset %s while not owning all elements" %
1353 # Calling delete() acquires the lock exclusively if we don't already own
1354 # it, and causes all pending and subsequent lock acquires to fail. It's
1355 # fine to call it out of order because delete() also implies release(),
1356 # and the assertion above guarantees that if we either already hold
1357 # everything we want to delete, or we hold none.
1359 self.__lockdict[lname].delete()
1360 removed.append(lname)
1361 except (KeyError, errors.LockError):
1362 # This cannot happen if we were already holding it, verify:
1363 assert not self.is_owned(), ("remove failed while holding lockset %s" %
1366 # If no LockError was raised we are the ones who deleted the lock.
1367 # This means we can safely remove it from lockdict, as any further or
1368 # pending delete() or acquire() will fail (and nobody can have the lock
1369 # since before our call to delete()).
1371 # This is done in an else clause because if the exception was thrown
1372 # it's the job of the one who actually deleted it.
1373 del self.__lockdict[lname]
1374 # And let's remove it from our private list if we owned it.
1376 self._del_owned(name=lname)
1381 # Locking levels, must be acquired in increasing order.
1382 # Current rules are:
1383 # - at level LEVEL_CLUSTER resides the Big Ganeti Lock (BGL) which must be
1384 # acquired before performing any operation, either in shared or in exclusive
1385 # mode. acquiring the BGL in exclusive mode is discouraged and should be
1387 # - at levels LEVEL_NODE and LEVEL_INSTANCE reside node and instance locks.
1388 # If you need more than one node, or more than one instance, acquire them at
1404 # Lock levels which are modifiable
1405 LEVELS_MOD = frozenset([
1412 #: Lock level names (make sure to use singular form)
1414 LEVEL_CLUSTER: "cluster",
1415 LEVEL_INSTANCE: "instance",
1416 LEVEL_NODEGROUP: "nodegroup",
1418 LEVEL_NODE_RES: "node-res",
1421 # Constant for the big ganeti lock
1425 class GanetiLockManager:
1426 """The Ganeti Locking Library
1428 The purpose of this small library is to manage locking for ganeti clusters
1429 in a central place, while at the same time doing dynamic checks against
1430 possible deadlocks. It will also make it easier to transition to a different
1431 lock type should we migrate away from python threads.
1436 def __init__(self, nodes, nodegroups, instances):
1437 """Constructs a new GanetiLockManager object.
1439 There should be only a GanetiLockManager object at any time, so this
1440 function raises an error if this is not the case.
1442 @param nodes: list of node names
1443 @param nodegroups: list of nodegroup uuids
1444 @param instances: list of instance names
1447 assert self.__class__._instance is None, \
1448 "double GanetiLockManager instance"
1450 self.__class__._instance = self
1452 self._monitor = LockMonitor()
1454 # The keyring contains all the locks, at their level and in the correct
1457 LEVEL_CLUSTER: LockSet([BGL], "cluster", monitor=self._monitor),
1458 LEVEL_NODE: LockSet(nodes, "node", monitor=self._monitor),
1459 LEVEL_NODE_RES: LockSet(nodes, "node-res", monitor=self._monitor),
1460 LEVEL_NODEGROUP: LockSet(nodegroups, "nodegroup", monitor=self._monitor),
1461 LEVEL_INSTANCE: LockSet(instances, "instance",
1462 monitor=self._monitor),
1465 assert compat.all(ls.name == LEVEL_NAMES[level]
1466 for (level, ls) in self.__keyring.items())
1468 def AddToLockMonitor(self, provider):
1469 """Registers a new lock with the monitor.
1471 See L{LockMonitor.RegisterLock}.
1474 return self._monitor.RegisterLock(provider)
1476 def QueryLocks(self, fields):
1477 """Queries information from all locks.
1479 See L{LockMonitor.QueryLocks}.
1482 return self._monitor.QueryLocks(fields)
1484 def OldStyleQueryLocks(self, fields):
1485 """Queries information from all locks, returning old-style data.
1487 See L{LockMonitor.OldStyleQueryLocks}.
1490 return self._monitor.OldStyleQueryLocks(fields)
1492 def _names(self, level):
1493 """List the lock names at the given level.
1495 This can be used for debugging/testing purposes.
1497 @param level: the level whose list of locks to get
1500 assert level in LEVELS, "Invalid locking level %s" % level
1501 return self.__keyring[level]._names()
1503 def is_owned(self, level):
1504 """Check whether we are owning locks at the given level
1507 return self.__keyring[level].is_owned()
1509 def list_owned(self, level):
1510 """Get the set of owned locks at the given level
1513 return self.__keyring[level].list_owned()
1515 def _upper_owned(self, level):
1516 """Check that we don't own any lock at a level greater than the given one.
1519 # This way of checking only works if LEVELS[i] = i, which we check for in
1521 return compat.any((self.is_owned(l) for l in LEVELS[level + 1:]))
1523 def _BGL_owned(self): # pylint: disable=C0103
1524 """Check if the current thread owns the BGL.
1526 Both an exclusive or a shared acquisition work.
1529 return BGL in self.__keyring[LEVEL_CLUSTER].list_owned()
1532 def _contains_BGL(level, names): # pylint: disable=C0103
1533 """Check if the level contains the BGL.
1535 Check if acting on the given level and set of names will change
1536 the status of the Big Ganeti Lock.
1539 return level == LEVEL_CLUSTER and (names is None or BGL in names)
1541 def acquire(self, level, names, timeout=None, shared=0, priority=None):
1542 """Acquire a set of resource locks, at the same level.
1544 @type level: member of locking.LEVELS
1545 @param level: the level at which the locks shall be acquired
1546 @type names: list of strings (or string)
1547 @param names: the names of the locks which shall be acquired
1548 (special lock names, or instance/node names)
1549 @type shared: integer (0/1) used as a boolean
1550 @param shared: whether to acquire in shared mode; by default
1551 an exclusive lock will be acquired
1552 @type timeout: float
1553 @param timeout: Maximum time to acquire all locks
1554 @type priority: integer
1555 @param priority: Priority for acquiring lock
1558 assert level in LEVELS, "Invalid locking level %s" % level
1560 # Check that we are either acquiring the Big Ganeti Lock or we already own
1561 # it. Some "legacy" opcodes need to be sure they are run non-concurrently
1562 # so even if we've migrated we need to at least share the BGL to be
1563 # compatible with them. Of course if we own the BGL exclusively there's no
1564 # point in acquiring any other lock, unless perhaps we are half way through
1565 # the migration of the current opcode.
1566 assert (self._contains_BGL(level, names) or self._BGL_owned()), (
1567 "You must own the Big Ganeti Lock before acquiring any other")
1569 # Check we don't own locks at the same or upper levels.
1570 assert not self._upper_owned(level), ("Cannot acquire locks at a level"
1571 " while owning some at a greater one")
1573 # Acquire the locks in the set.
1574 return self.__keyring[level].acquire(names, shared=shared, timeout=timeout,
1577 def downgrade(self, level, names=None):
1578 """Downgrade a set of resource locks from exclusive to shared mode.
1580 You must have acquired the locks in exclusive mode.
1582 @type level: member of locking.LEVELS
1583 @param level: the level at which the locks shall be downgraded
1584 @type names: list of strings, or None
1585 @param names: the names of the locks which shall be downgraded
1586 (defaults to all the locks acquired at the level)
1589 assert level in LEVELS, "Invalid locking level %s" % level
1591 return self.__keyring[level].downgrade(names=names)
1593 def release(self, level, names=None):
1594 """Release a set of resource locks, at the same level.
1596 You must have acquired the locks, either in shared or in exclusive
1597 mode, before releasing them.
1599 @type level: member of locking.LEVELS
1600 @param level: the level at which the locks shall be released
1601 @type names: list of strings, or None
1602 @param names: the names of the locks which shall be released
1603 (defaults to all the locks acquired at that level)
1606 assert level in LEVELS, "Invalid locking level %s" % level
1607 assert (not self._contains_BGL(level, names) or
1608 not self._upper_owned(LEVEL_CLUSTER)), (
1609 "Cannot release the Big Ganeti Lock while holding something"
1610 " at upper levels (%r)" %
1611 (utils.CommaJoin(["%s=%r" % (LEVEL_NAMES[i], self.list_owned(i))
1612 for i in self.__keyring.keys()]), ))
1614 # Release will complain if we don't own the locks already
1615 return self.__keyring[level].release(names)
1617 def add(self, level, names, acquired=0, shared=0):
1618 """Add locks at the specified level.
1620 @type level: member of locking.LEVELS_MOD
1621 @param level: the level at which the locks shall be added
1622 @type names: list of strings
1623 @param names: names of the locks to acquire
1624 @type acquired: integer (0/1) used as a boolean
1625 @param acquired: whether to acquire the newly added locks
1626 @type shared: integer (0/1) used as a boolean
1627 @param shared: whether the acquisition will be shared
1630 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
1631 assert self._BGL_owned(), ("You must own the BGL before performing other"
1633 assert not self._upper_owned(level), ("Cannot add locks at a level"
1634 " while owning some at a greater one")
1635 return self.__keyring[level].add(names, acquired=acquired, shared=shared)
1637 def remove(self, level, names):
1638 """Remove locks from the specified level.
1640 You must either already own the locks you are trying to remove
1641 exclusively or not own any lock at an upper level.
1643 @type level: member of locking.LEVELS_MOD
1644 @param level: the level at which the locks shall be removed
1645 @type names: list of strings
1646 @param names: the names of the locks which shall be removed
1647 (special lock names, or instance/node names)
1650 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
1651 assert self._BGL_owned(), ("You must own the BGL before performing other"
1653 # Check we either own the level or don't own anything from here
1654 # up. LockSet.remove() will check the case in which we don't own
1655 # all the needed resources, or we have a shared ownership.
1656 assert self.is_owned(level) or not self._upper_owned(level), (
1657 "Cannot remove locks at a level while not owning it or"
1658 " owning some at a greater one")
1659 return self.__keyring[level].remove(names)
1662 def _MonitorSortKey((item, idx, num)):
1663 """Sorting key function.
1665 Sort by name, registration order and then order of information. This provides
1666 a stable sort order over different providers, even if they return the same
1670 (name, _, _, _) = item
1672 return (utils.NiceSortKey(name), num, idx)
1675 class LockMonitor(object):
1676 _LOCK_ATTR = "_lock"
1679 """Initializes this class.
1682 self._lock = SharedLock("LockMonitor")
1684 # Counter for stable sorting
1685 self._counter = itertools.count(0)
1687 # Tracked locks. Weak references are used to avoid issues with circular
1688 # references and deletion.
1689 self._locks = weakref.WeakKeyDictionary()
1691 @ssynchronized(_LOCK_ATTR)
1692 def RegisterLock(self, provider):
1693 """Registers a new lock.
1695 @param provider: Object with a callable method named C{GetLockInfo}, taking
1696 a single C{set} containing the requested information items
1697 @note: It would be nicer to only receive the function generating the
1698 requested information but, as it turns out, weak references to bound
1699 methods (e.g. C{self.GetLockInfo}) are tricky; there are several
1700 workarounds, but none of the ones I found works properly in combination
1701 with a standard C{WeakKeyDictionary}
1704 assert provider not in self._locks, "Duplicate registration"
1706 # There used to be a check for duplicate names here. As it turned out, when
1707 # a lock is re-created with the same name in a very short timeframe, the
1708 # previous instance might not yet be removed from the weakref dictionary.
1709 # By keeping track of the order of incoming registrations, a stable sort
1710 # ordering can still be guaranteed.
1712 self._locks[provider] = self._counter.next()
1714 def _GetLockInfo(self, requested):
1715 """Get information from all locks.
1718 # Must hold lock while getting consistent list of tracked items
1719 self._lock.acquire(shared=1)
1721 items = self._locks.items()
1723 self._lock.release()
1725 return [(info, idx, num)
1726 for (provider, num) in items
1727 for (idx, info) in enumerate(provider.GetLockInfo(requested))]
1729 def _Query(self, fields):
1730 """Queries information from all locks.
1732 @type fields: list of strings
1733 @param fields: List of fields to return
1736 qobj = query.Query(query.LOCK_FIELDS, fields)
1738 # Get all data with internal lock held and then sort by name and incoming
1740 lockinfo = sorted(self._GetLockInfo(qobj.RequestedData()),
1741 key=_MonitorSortKey)
1743 # Extract lock information and build query data
1744 return (qobj, query.LockQueryData(map(compat.fst, lockinfo)))
1746 def QueryLocks(self, fields):
1747 """Queries information from all locks.
1749 @type fields: list of strings
1750 @param fields: List of fields to return
1753 (qobj, ctx) = self._Query(fields)
1755 # Prepare query response
1756 return query.GetQueryResponse(qobj, ctx)
1758 def OldStyleQueryLocks(self, fields):
1759 """Queries information from all locks, returning old-style data.
1761 @type fields: list of strings
1762 @param fields: List of fields to return
1765 (qobj, ctx) = self._Query(fields)
1767 return qobj.OldStyleQuery(ctx)