4 # Copyright (C) 2006, 2007 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=W0613,W0201
26 # Wouldn't it be better to define LockingError in the locking module?
27 # Well, for now that's how the rest of the code does it...
28 from ganeti import errors
29 from ganeti import utils
32 def ssynchronized(lock, shared=0):
33 """Shared Synchronization decorator.
35 Calls the function holding the given lock, either in exclusive or shared
36 mode. It requires the passed lock to be a SharedLock (or support its
41 def sync_function(*args, **kwargs):
42 lock.acquire(shared=shared)
44 return fn(*args, **kwargs)
52 """Implements a shared lock.
54 Multiple threads can acquire the lock in a shared way, calling
55 acquire_shared(). In order to acquire the lock in an exclusive way threads
56 can call acquire_exclusive().
58 The lock prevents starvation but does not guarantee that threads will acquire
59 the shared lock in the order they queued for it, just that they will
64 """Construct a new SharedLock"""
65 # we have two conditions, c_shr and c_exc, sharing the same lock.
66 self.__lock = threading.Lock()
67 self.__turn_shr = threading.Condition(self.__lock)
68 self.__turn_exc = threading.Condition(self.__lock)
70 # current lock holders
79 # is this lock in the deleted state?
80 self.__deleted = False
82 def __is_sharer(self):
83 """Is the current thread sharing the lock at this time?"""
84 return threading.currentThread() in self.__shr
86 def __is_exclusive(self):
87 """Is the current thread holding the lock exclusively at this time?"""
88 return threading.currentThread() == self.__exc
90 def __is_owned(self, shared=-1):
91 """Is the current thread somehow owning the lock at this time?
93 This is a private version of the function, which presumes you're holding
98 return self.__is_sharer() or self.__is_exclusive()
100 return self.__is_sharer()
102 return self.__is_exclusive()
104 def _is_owned(self, shared=-1):
105 """Is the current thread somehow owning the lock at this time?
108 - < 0: check for any type of ownership (default)
109 - 0: check for exclusive ownership
110 - > 0: check for shared ownership
113 self.__lock.acquire()
115 result = self.__is_owned(shared=shared)
117 self.__lock.release()
122 """Wait on the given condition, and raise an exception if the current lock
123 is declared deleted in the meantime.
125 @param c: the condition to wait on
130 raise errors.LockError('deleted lock')
132 def __exclusive_acquire(self):
133 """Acquire the lock exclusively.
135 This is a private function that presumes you are already holding the
136 internal lock. It's defined separately to avoid code duplication between
137 acquire() and delete()
140 self.__nwait_exc += 1
142 # This is to save ourselves from a nasty race condition that could
143 # theoretically make the sharers starve.
144 if self.__nwait_shr > 0 or self.__nwait_exc > 1:
145 self.__wait(self.__turn_exc)
147 while len(self.__shr) > 0 or self.__exc is not None:
148 self.__wait(self.__turn_exc)
150 self.__exc = threading.currentThread()
152 self.__nwait_exc -= 1
154 assert self.__npass_shr == 0, "SharedLock: internal fairness violation"
156 def __shared_acquire(self):
157 """Acquire the lock in shared mode
159 This is a private function that presumes you are already holding the
163 self.__nwait_shr += 1
166 # If there is an exclusive holder waiting we have to wait.
167 # We'll only do this once, though, when we start waiting for
168 # the lock. Then we'll just wait while there are no
170 if self.__nwait_exc > 0:
171 # TODO: if !blocking...
173 self.__wait(self.__turn_shr)
175 while self.__exc is not None:
177 # TODO: if !blocking...
178 self.__wait(self.__turn_shr)
180 self.__shr.add(threading.currentThread())
182 # If we were waiting note that we passed
184 self.__npass_shr -= 1
187 self.__nwait_shr -= 1
189 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
191 def acquire(self, blocking=1, shared=0):
192 """Acquire a shared lock.
194 @param shared: whether to acquire in shared mode; by default an
195 exclusive lock will be acquired
196 @param blocking: whether to block while trying to acquire or to
197 operate in try-lock mode (this locking mode is not supported yet)
201 # We don't have non-blocking mode for now
202 raise NotImplementedError
204 self.__lock.acquire()
207 raise errors.LockError('deleted lock')
209 # We cannot acquire the lock if we already have it
210 assert not self.__is_owned(), "double acquire() on a non-recursive lock"
211 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
214 self.__shared_acquire()
216 # TODO: if !blocking...
217 # (or modify __exclusive_acquire for non-blocking mode)
218 self.__exclusive_acquire()
221 self.__lock.release()
226 """Release a Shared Lock.
228 You must have acquired the lock, either in shared or in exclusive mode,
229 before calling this function.
232 self.__lock.acquire()
234 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
235 # Autodetect release type
236 if self.__is_exclusive():
239 # An exclusive holder has just had the lock, time to put it in shared
240 # mode if there are shared holders waiting. Otherwise wake up the next
242 if self.__nwait_shr > 0:
243 # Make sure at least the ones which were blocked pass.
244 self.__npass_shr = self.__nwait_shr
245 self.__turn_shr.notifyAll()
246 elif self.__nwait_exc > 0:
247 self.__turn_exc.notify()
249 elif self.__is_sharer():
250 self.__shr.remove(threading.currentThread())
252 # If there are shared holders waiting (and not just scheduled to pass)
253 # there *must* be an exclusive holder waiting as well; otherwise what
254 # were they waiting for?
255 assert (self.__nwait_exc > 0 or
256 self.__npass_shr == self.__nwait_shr), \
257 "Lock sharers waiting while no exclusive is queueing"
259 # If there are no more shared holders either in or scheduled to pass,
260 # and some exclusive holders are waiting let's wake one up.
261 if (len(self.__shr) == 0 and
262 self.__nwait_exc > 0 and
263 not self.__npass_shr > 0):
264 self.__turn_exc.notify()
267 assert False, "Cannot release non-owned lock"
270 self.__lock.release()
272 def delete(self, blocking=1):
273 """Delete a Shared Lock.
275 This operation will declare the lock for removal. First the lock will be
276 acquired in exclusive mode if you don't already own it, then the lock
277 will be put in a state where any future and pending acquire() fail.
279 @param blocking: whether to block while trying to acquire or to
280 operate in try-lock mode. this locking mode is not supported
281 yet unless you are already holding exclusively the lock.
284 self.__lock.acquire()
286 assert not self.__is_sharer(), "cannot delete() a lock while sharing it"
289 raise errors.LockError('deleted lock')
291 if not self.__is_exclusive():
293 # We don't have non-blocking mode for now
294 raise NotImplementedError
295 self.__exclusive_acquire()
297 self.__deleted = True
299 # Wake up everybody, they will fail acquiring the lock and
300 # raise an exception instead.
301 self.__turn_exc.notifyAll()
302 self.__turn_shr.notifyAll()
305 self.__lock.release()
308 # Whenever we want to acquire a full LockSet we pass None as the value
309 # to acquire. Hide this behind this nicely named constant.
314 """Implements a set of locks.
316 This abstraction implements a set of shared locks for the same resource type,
317 distinguished by name. The user can lock a subset of the resources and the
318 LockSet will take care of acquiring the locks always in the same order, thus
321 All the locks needed in the same set must be acquired together, though.
324 def __init__(self, members=None):
325 """Constructs a new LockSet.
327 @param members: initial members of the set
330 # Used internally to guarantee coherency.
331 self.__lock = SharedLock()
333 # The lockdict indexes the relationship name -> lock
334 # The order-of-locking is implied by the alphabetical order of names
337 if members is not None:
339 self.__lockdict[name] = SharedLock()
341 # The owner dict contains the set of locks each thread owns. For
342 # performance each thread can access its own key without a global lock on
343 # this structure. It is paramount though that *no* other type of access is
344 # done to this structure (eg. no looping over its keys). *_owner helper
345 # function are defined to guarantee access is correct, but in general never
346 # do anything different than __owners[threading.currentThread()], or there
351 """Is the current thread a current level owner?"""
352 return threading.currentThread() in self.__owners
354 def _add_owned(self, name=None):
355 """Note the current thread owns the given lock"""
357 if not self._is_owned():
358 self.__owners[threading.currentThread()] = set()
361 self.__owners[threading.currentThread()].add(name)
363 self.__owners[threading.currentThread()] = set([name])
365 def _del_owned(self, name=None):
366 """Note the current thread owns the given lock"""
369 self.__owners[threading.currentThread()].remove(name)
371 # Only remove the key if we don't hold the set-lock as well
372 if (not self.__lock._is_owned() and
373 not self.__owners[threading.currentThread()]):
374 del self.__owners[threading.currentThread()]
376 def _list_owned(self):
377 """Get the set of resource names owned by the current thread"""
379 return self.__owners[threading.currentThread()].copy()
384 """Return the current set of names.
386 Only call this function while holding __lock and don't iterate on the
387 result after releasing the lock.
390 return self.__lockdict.keys()
393 """Return a copy of the current set of elements.
395 Used only for debugging purposes.
398 # If we don't already own the set-level lock acquired
399 # we'll get it and note we need to release it later.
401 if not self.__lock._is_owned():
403 self.__lock.acquire(shared=1)
405 result = self.__names()
408 self.__lock.release()
411 def acquire(self, names, blocking=1, shared=0):
412 """Acquire a set of resource locks.
414 @param names: the names of the locks which shall be acquired
415 (special lock names, or instance/node names)
416 @param shared: whether to acquire in shared mode; by default an
417 exclusive lock will be acquired
418 @param blocking: whether to block while trying to acquire or to
419 operate in try-lock mode (this locking mode is not supported yet)
421 @return: True when all the locks are successfully acquired
423 @raise errors.LockError: when any lock we try to acquire has
424 been deleted before we succeed. In this case none of the
425 locks requested will be acquired.
429 # We don't have non-blocking mode for now
430 raise NotImplementedError
432 # Check we don't already own locks at this level
433 assert not self._is_owned(), "Cannot acquire locks in the same set twice"
436 # If no names are given acquire the whole set by not letting new names
437 # being added before we release, and getting the current list of names.
438 # Some of them may then be deleted later, but we'll cope with this.
440 # We'd like to acquire this lock in a shared way, as it's nice if
441 # everybody else can use the instances at the same time. If are acquiring
442 # them exclusively though they won't be able to do this anyway, though,
443 # so we'll get the list lock exclusively as well in order to be able to
444 # do add() on the set while owning it.
445 self.__lock.acquire(shared=shared)
447 # note we own the set-lock
449 names = self.__names()
451 # We shouldn't have problems adding the lock to the owners list, but
452 # if we did we'll try to release this lock and re-raise exception.
453 # Of course something is going to be really wrong, after this.
454 self.__lock.release()
458 # Support passing in a single resource to acquire rather than many
459 if isinstance(names, basestring):
462 names = sorted(names)
465 # First we look the locks up on __lockdict. We have no way of being sure
466 # they will still be there after, but this makes it a lot faster should
467 # just one of them be the already wrong
468 for lname in utils.UniqueSequence(names):
470 lock = self.__lockdict[lname] # raises KeyError if lock is not there
471 acquire_list.append((lname, lock))
473 if self.__lock._is_owned():
474 # We are acquiring all the set, it doesn't matter if this
475 # particular element is not there anymore.
478 raise errors.LockError('non-existing lock in set (%s)' % lname)
480 # This will hold the locknames we effectively acquired.
482 # Now acquire_list contains a sorted list of resources and locks we want.
483 # In order to get them we loop on this (private) list and acquire() them.
484 # We gave no real guarantee they will still exist till this is done but
485 # .acquire() itself is safe and will alert us if the lock gets deleted.
486 for (lname, lock) in acquire_list:
488 lock.acquire(shared=shared) # raises LockError if the lock is deleted
489 # now the lock cannot be deleted, we have it!
490 self._add_owned(name=lname)
492 except (errors.LockError):
493 if self.__lock._is_owned():
494 # We are acquiring all the set, it doesn't matter if this
495 # particular element is not there anymore.
499 for lname in self._list_owned():
500 self.__lockdict[lname].release()
501 self._del_owned(name=lname)
502 raise errors.LockError('non-existing lock in set (%s)' % name_fail)
504 # We shouldn't have problems adding the lock to the owners list, but
505 # if we did we'll try to release this lock and re-raise exception.
506 # Of course something is going to be really wrong, after this.
512 # If something went wrong and we had the set-lock let's release it...
513 if self.__lock._is_owned():
514 self.__lock.release()
519 def release(self, names=None):
520 """Release a set of resource locks, at the same level.
522 You must have acquired the locks, either in shared or in exclusive mode,
523 before releasing them.
525 @param names: the names of the locks which shall be released
526 (defaults to all the locks acquired at that level).
529 assert self._is_owned(), "release() on lock set while not owner"
531 # Support passing in a single resource to release rather than many
532 if isinstance(names, basestring):
536 names = self._list_owned()
539 assert self._list_owned().issuperset(names), (
540 "release() on unheld resources %s" %
541 names.difference(self._list_owned()))
543 # First of all let's release the "all elements" lock, if set.
544 # After this 'add' can work again
545 if self.__lock._is_owned():
546 self.__lock.release()
549 for lockname in names:
550 # If we are sure the lock doesn't leave __lockdict without being
551 # exclusively held we can do this...
552 self.__lockdict[lockname].release()
553 self._del_owned(name=lockname)
555 def add(self, names, acquired=0, shared=0):
556 """Add a new set of elements to the set
558 @param names: names of the new elements to add
559 @param acquired: pre-acquire the new resource?
560 @param shared: is the pre-acquisition shared?
563 # Check we don't already own locks at this level
564 assert not self._is_owned() or self.__lock._is_owned(shared=0), \
565 "Cannot add locks if the set is only partially owned, or shared"
567 # Support passing in a single resource to add rather than many
568 if isinstance(names, basestring):
571 # If we don't already own the set-level lock acquired in an exclusive way
572 # we'll get it and note we need to release it later.
574 if not self.__lock._is_owned():
576 self.__lock.acquire()
579 invalid_names = set(self.__names()).intersection(names)
581 # This must be an explicit raise, not an assert, because assert is
582 # turned off when using optimization, and this can happen because of
583 # concurrency even if the user doesn't want it.
584 raise errors.LockError("duplicate add() (%s)" % invalid_names)
586 for lockname in names:
590 lock.acquire(shared=shared)
591 # now the lock cannot be deleted, we have it!
593 self._add_owned(name=lockname)
595 # We shouldn't have problems adding the lock to the owners list,
596 # but if we did we'll try to release this lock and re-raise
597 # exception. Of course something is going to be really wrong,
598 # after this. On the other hand the lock hasn't been added to the
599 # __lockdict yet so no other threads should be pending on it. This
600 # release is just a safety measure.
604 self.__lockdict[lockname] = lock
607 # Only release __lock if we were not holding it previously.
609 self.__lock.release()
613 def remove(self, names, blocking=1):
614 """Remove elements from the lock set.
616 You can either not hold anything in the lockset or already hold a superset
617 of the elements you want to delete, exclusively.
619 @param names: names of the resource to remove.
620 @param blocking: whether to block while trying to acquire or to
621 operate in try-lock mode (this locking mode is not supported
622 yet unless you are already holding exclusively the locks)
624 @return:: a list of locks which we removed; the list is always
625 equal to the names list if we were holding all the locks
629 if not blocking and not self._is_owned():
630 # We don't have non-blocking mode for now
631 raise NotImplementedError
633 # Support passing in a single resource to remove rather than many
634 if isinstance(names, basestring):
637 # If we own any subset of this lock it must be a superset of what we want
638 # to delete. The ownership must also be exclusive, but that will be checked
639 # by the lock itself.
640 assert not self._is_owned() or self._list_owned().issuperset(names), (
641 "remove() on acquired lockset while not owning all elements")
646 # Calling delete() acquires the lock exclusively if we don't already own
647 # it, and causes all pending and subsequent lock acquires to fail. It's
648 # fine to call it out of order because delete() also implies release(),
649 # and the assertion above guarantees that if we either already hold
650 # everything we want to delete, or we hold none.
652 self.__lockdict[lname].delete()
653 removed.append(lname)
654 except (KeyError, errors.LockError):
655 # This cannot happen if we were already holding it, verify:
656 assert not self._is_owned(), "remove failed while holding lockset"
658 # If no LockError was raised we are the ones who deleted the lock.
659 # This means we can safely remove it from lockdict, as any further or
660 # pending delete() or acquire() will fail (and nobody can have the lock
661 # since before our call to delete()).
663 # This is done in an else clause because if the exception was thrown
664 # it's the job of the one who actually deleted it.
665 del self.__lockdict[lname]
666 # And let's remove it from our private list if we owned it.
668 self._del_owned(name=lname)
673 # Locking levels, must be acquired in increasing order.
675 # - at level LEVEL_CLUSTER resides the Big Ganeti Lock (BGL) which must be
676 # acquired before performing any operation, either in shared or in exclusive
677 # mode. acquiring the BGL in exclusive mode is discouraged and should be
679 # - at levels LEVEL_NODE and LEVEL_INSTANCE reside node and instance locks.
680 # If you need more than one node, or more than one instance, acquire them at
686 LEVELS = [LEVEL_CLUSTER,
690 # Lock levels which are modifiable
691 LEVELS_MOD = [LEVEL_NODE, LEVEL_INSTANCE]
694 LEVEL_CLUSTER: "cluster",
695 LEVEL_INSTANCE: "instance",
699 # Constant for the big ganeti lock
703 class GanetiLockManager:
704 """The Ganeti Locking Library
706 The purpose of this small library is to manage locking for ganeti clusters
707 in a central place, while at the same time doing dynamic checks against
708 possible deadlocks. It will also make it easier to transition to a different
709 lock type should we migrate away from python threads.
714 def __init__(self, nodes=None, instances=None):
715 """Constructs a new GanetiLockManager object.
717 There should be only a GanetiLockManager object at any time, so this
718 function raises an error if this is not the case.
720 @param nodes: list of node names
721 @param instances: list of instance names
724 assert self.__class__._instance is None, \
725 "double GanetiLockManager instance"
727 self.__class__._instance = self
729 # The keyring contains all the locks, at their level and in the correct
732 LEVEL_CLUSTER: LockSet([BGL]),
733 LEVEL_NODE: LockSet(nodes),
734 LEVEL_INSTANCE: LockSet(instances),
737 def _names(self, level):
738 """List the lock names at the given level.
740 This can be used for debugging/testing purposes.
742 @param level: the level whose list of locks to get
745 assert level in LEVELS, "Invalid locking level %s" % level
746 return self.__keyring[level]._names()
748 def _is_owned(self, level):
749 """Check whether we are owning locks at the given level
752 return self.__keyring[level]._is_owned()
756 def _list_owned(self, level):
757 """Get the set of owned locks at the given level
760 return self.__keyring[level]._list_owned()
762 def _upper_owned(self, level):
763 """Check that we don't own any lock at a level greater than the given one.
766 # This way of checking only works if LEVELS[i] = i, which we check for in
768 return utils.any((self._is_owned(l) for l in LEVELS[level + 1:]))
770 def _BGL_owned(self):
771 """Check if the current thread owns the BGL.
773 Both an exclusive or a shared acquisition work.
776 return BGL in self.__keyring[LEVEL_CLUSTER]._list_owned()
778 def _contains_BGL(self, level, names):
779 """Check if the level contains the BGL.
781 Check if acting on the given level and set of names will change
782 the status of the Big Ganeti Lock.
785 return level == LEVEL_CLUSTER and (names is None or BGL in names)
787 def acquire(self, level, names, blocking=1, shared=0):
788 """Acquire a set of resource locks, at the same level.
790 @param level: the level at which the locks shall be acquired;
791 it must be a member of LEVELS.
792 @param names: the names of the locks which shall be acquired
793 (special lock names, or instance/node names)
794 @param shared: whether to acquire in shared mode; by default
795 an exclusive lock will be acquired
796 @param blocking: whether to block while trying to acquire or to
797 operate in try-lock mode (this locking mode is not supported yet)
800 assert level in LEVELS, "Invalid locking level %s" % level
802 # Check that we are either acquiring the Big Ganeti Lock or we already own
803 # it. Some "legacy" opcodes need to be sure they are run non-concurrently
804 # so even if we've migrated we need to at least share the BGL to be
805 # compatible with them. Of course if we own the BGL exclusively there's no
806 # point in acquiring any other lock, unless perhaps we are half way through
807 # the migration of the current opcode.
808 assert (self._contains_BGL(level, names) or self._BGL_owned()), (
809 "You must own the Big Ganeti Lock before acquiring any other")
811 # Check we don't own locks at the same or upper levels.
812 assert not self._upper_owned(level), ("Cannot acquire locks at a level"
813 " while owning some at a greater one")
815 # Acquire the locks in the set.
816 return self.__keyring[level].acquire(names, shared=shared,
819 def release(self, level, names=None):
820 """Release a set of resource locks, at the same level.
822 You must have acquired the locks, either in shared or in exclusive
823 mode, before releasing them.
825 @param level: the level at which the locks shall be released;
826 it must be a member of LEVELS
827 @param names: the names of the locks which shall be released
828 (defaults to all the locks acquired at that level)
831 assert level in LEVELS, "Invalid locking level %s" % level
832 assert (not self._contains_BGL(level, names) or
833 not self._upper_owned(LEVEL_CLUSTER)), (
834 "Cannot release the Big Ganeti Lock while holding something"
837 # Release will complain if we don't own the locks already
838 return self.__keyring[level].release(names)
840 def add(self, level, names, acquired=0, shared=0):
841 """Add locks at the specified level.
843 @param level: the level at which the locks shall be added;
844 it must be a member of LEVELS_MOD.
845 @param names: names of the locks to acquire
846 @param acquired: whether to acquire the newly added locks
847 @param shared: whether the acquisition will be shared
850 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
851 assert self._BGL_owned(), ("You must own the BGL before performing other"
853 assert not self._upper_owned(level), ("Cannot add locks at a level"
854 " while owning some at a greater one")
855 return self.__keyring[level].add(names, acquired=acquired, shared=shared)
857 def remove(self, level, names, blocking=1):
858 """Remove locks from the specified level.
860 You must either already own the locks you are trying to remove
861 exclusively or not own any lock at an upper level.
863 @param level: the level at which the locks shall be removed;
864 it must be a member of LEVELS_MOD
865 @param names: the names of the locks which shall be removed
866 (special lock names, or instance/node names)
867 @param blocking: whether to block while trying to operate in
868 try-lock mode (this locking mode is not supported yet)
871 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
872 assert self._BGL_owned(), ("You must own the BGL before performing other"
874 # Check we either own the level or don't own anything from here
875 # up. LockSet.remove() will check the case in which we don't own
876 # all the needed resources, or we have a shared ownership.
877 assert self._is_owned(level) or not self._upper_owned(level), (
878 "Cannot remove locks at a level while not owning it or"
879 " owning some at a greater one")
880 return self.__keyring[level].remove(names, blocking=blocking)