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=W0212
25 # W0212 since e.g. LockSet methods use (a lot) the internals of
29 # Wouldn't it be better to define LockingError in the locking module?
30 # Well, for now that's how the rest of the code does it...
31 from ganeti import errors
32 from ganeti import utils
35 def ssynchronized(lock, shared=0):
36 """Shared Synchronization decorator.
38 Calls the function holding the given lock, either in exclusive or shared
39 mode. It requires the passed lock to be a SharedLock (or support its
44 def sync_function(*args, **kwargs):
45 lock.acquire(shared=shared)
47 return fn(*args, **kwargs)
55 """Implements a shared lock.
57 Multiple threads can acquire the lock in a shared way, calling
58 acquire_shared(). In order to acquire the lock in an exclusive way threads
59 can call acquire_exclusive().
61 The lock prevents starvation but does not guarantee that threads will acquire
62 the shared lock in the order they queued for it, just that they will
67 """Construct a new SharedLock"""
68 # we have two conditions, c_shr and c_exc, sharing the same lock.
69 self.__lock = threading.Lock()
70 self.__turn_shr = threading.Condition(self.__lock)
71 self.__turn_exc = threading.Condition(self.__lock)
73 # current lock holders
82 # is this lock in the deleted state?
83 self.__deleted = False
85 def __is_sharer(self):
86 """Is the current thread sharing the lock at this time?"""
87 return threading.currentThread() in self.__shr
89 def __is_exclusive(self):
90 """Is the current thread holding the lock exclusively at this time?"""
91 return threading.currentThread() == self.__exc
93 def __is_owned(self, shared=-1):
94 """Is the current thread somehow owning the lock at this time?
96 This is a private version of the function, which presumes you're holding
101 return self.__is_sharer() or self.__is_exclusive()
103 return self.__is_sharer()
105 return self.__is_exclusive()
107 def _is_owned(self, shared=-1):
108 """Is the current thread somehow owning the lock at this time?
111 - < 0: check for any type of ownership (default)
112 - 0: check for exclusive ownership
113 - > 0: check for shared ownership
116 self.__lock.acquire()
118 result = self.__is_owned(shared=shared)
120 self.__lock.release()
125 """Wait on the given condition, and raise an exception if the current lock
126 is declared deleted in the meantime.
128 @param c: the condition to wait on
133 raise errors.LockError('deleted lock')
135 def __exclusive_acquire(self):
136 """Acquire the lock exclusively.
138 This is a private function that presumes you are already holding the
139 internal lock. It's defined separately to avoid code duplication between
140 acquire() and delete()
143 self.__nwait_exc += 1
145 # This is to save ourselves from a nasty race condition that could
146 # theoretically make the sharers starve.
147 if self.__nwait_shr > 0 or self.__nwait_exc > 1:
148 self.__wait(self.__turn_exc)
150 while len(self.__shr) > 0 or self.__exc is not None:
151 self.__wait(self.__turn_exc)
153 self.__exc = threading.currentThread()
155 self.__nwait_exc -= 1
157 assert self.__npass_shr == 0, "SharedLock: internal fairness violation"
159 def acquire(self, blocking=1, shared=0):
160 """Acquire a shared lock.
162 @param shared: whether to acquire in shared mode; by default an
163 exclusive lock will be acquired
164 @param blocking: whether to block while trying to acquire or to
165 operate in try-lock mode (this locking mode is not supported yet)
169 # We don't have non-blocking mode for now
170 raise NotImplementedError
172 self.__lock.acquire()
175 raise errors.LockError('deleted lock')
177 # We cannot acquire the lock if we already have it
178 assert not self.__is_owned(), "double acquire() on a non-recursive lock"
179 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
182 self.__nwait_shr += 1
185 # If there is an exclusive holder waiting we have to wait.
186 # We'll only do this once, though, when we start waiting for
187 # the lock. Then we'll just wait while there are no
189 if self.__nwait_exc > 0:
190 # TODO: if !blocking...
192 self.__wait(self.__turn_shr)
194 while self.__exc is not None:
196 # TODO: if !blocking...
197 self.__wait(self.__turn_shr)
199 self.__shr.add(threading.currentThread())
201 # If we were waiting note that we passed
203 self.__npass_shr -= 1
206 self.__nwait_shr -= 1
208 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
210 # TODO: if !blocking...
211 # (or modify __exclusive_acquire for non-blocking mode)
212 self.__exclusive_acquire()
215 self.__lock.release()
220 """Release a Shared Lock.
222 You must have acquired the lock, either in shared or in exclusive mode,
223 before calling this function.
226 self.__lock.acquire()
228 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
229 # Autodetect release type
230 if self.__is_exclusive():
233 # An exclusive holder has just had the lock, time to put it in shared
234 # mode if there are shared holders waiting. Otherwise wake up the next
236 if self.__nwait_shr > 0:
237 # Make sure at least the ones which were blocked pass.
238 self.__npass_shr = self.__nwait_shr
239 self.__turn_shr.notifyAll()
240 elif self.__nwait_exc > 0:
241 self.__turn_exc.notify()
243 elif self.__is_sharer():
244 self.__shr.remove(threading.currentThread())
246 # If there are shared holders waiting (and not just scheduled to pass)
247 # there *must* be an exclusive holder waiting as well; otherwise what
248 # were they waiting for?
249 assert (self.__nwait_exc > 0 or
250 self.__npass_shr == self.__nwait_shr), \
251 "Lock sharers waiting while no exclusive is queueing"
253 # If there are no more shared holders either in or scheduled to pass,
254 # and some exclusive holders are waiting let's wake one up.
255 if (len(self.__shr) == 0 and
256 self.__nwait_exc > 0 and
257 not self.__npass_shr > 0):
258 self.__turn_exc.notify()
261 assert False, "Cannot release non-owned lock"
264 self.__lock.release()
266 def delete(self, blocking=1):
267 """Delete a Shared Lock.
269 This operation will declare the lock for removal. First the lock will be
270 acquired in exclusive mode if you don't already own it, then the lock
271 will be put in a state where any future and pending acquire() fail.
273 @param blocking: whether to block while trying to acquire or to
274 operate in try-lock mode. this locking mode is not supported
275 yet unless you are already holding exclusively the lock.
278 self.__lock.acquire()
280 assert not self.__is_sharer(), "cannot delete() a lock while sharing it"
283 raise errors.LockError('deleted lock')
285 if not self.__is_exclusive():
287 # We don't have non-blocking mode for now
288 raise NotImplementedError
289 self.__exclusive_acquire()
291 self.__deleted = True
293 # Wake up everybody, they will fail acquiring the lock and
294 # raise an exception instead.
295 self.__turn_exc.notifyAll()
296 self.__turn_shr.notifyAll()
299 self.__lock.release()
302 # Whenever we want to acquire a full LockSet we pass None as the value
303 # to acquire. Hide this behind this nicely named constant.
308 """Implements a set of locks.
310 This abstraction implements a set of shared locks for the same resource type,
311 distinguished by name. The user can lock a subset of the resources and the
312 LockSet will take care of acquiring the locks always in the same order, thus
315 All the locks needed in the same set must be acquired together, though.
318 def __init__(self, members=None):
319 """Constructs a new LockSet.
321 @param members: initial members of the set
324 # Used internally to guarantee coherency.
325 self.__lock = SharedLock()
327 # The lockdict indexes the relationship name -> lock
328 # The order-of-locking is implied by the alphabetical order of names
331 if members is not None:
333 self.__lockdict[name] = SharedLock()
335 # The owner dict contains the set of locks each thread owns. For
336 # performance each thread can access its own key without a global lock on
337 # this structure. It is paramount though that *no* other type of access is
338 # done to this structure (eg. no looping over its keys). *_owner helper
339 # function are defined to guarantee access is correct, but in general never
340 # do anything different than __owners[threading.currentThread()], or there
345 """Is the current thread a current level owner?"""
346 return threading.currentThread() in self.__owners
348 def _add_owned(self, name=None):
349 """Note the current thread owns the given lock"""
351 if not self._is_owned():
352 self.__owners[threading.currentThread()] = set()
355 self.__owners[threading.currentThread()].add(name)
357 self.__owners[threading.currentThread()] = set([name])
360 def _del_owned(self, name=None):
361 """Note the current thread owns the given lock"""
364 self.__owners[threading.currentThread()].remove(name)
366 # Only remove the key if we don't hold the set-lock as well
367 if (not self.__lock._is_owned() and
368 not self.__owners[threading.currentThread()]):
369 del self.__owners[threading.currentThread()]
371 def _list_owned(self):
372 """Get the set of resource names owned by the current thread"""
374 return self.__owners[threading.currentThread()].copy()
379 """Return the current set of names.
381 Only call this function while holding __lock and don't iterate on the
382 result after releasing the lock.
385 return self.__lockdict.keys()
388 """Return a copy of the current set of elements.
390 Used only for debugging purposes.
393 # If we don't already own the set-level lock acquired
394 # we'll get it and note we need to release it later.
396 if not self.__lock._is_owned():
398 self.__lock.acquire(shared=1)
400 result = self.__names()
403 self.__lock.release()
406 def acquire(self, names, blocking=1, shared=0):
407 """Acquire a set of resource locks.
409 @param names: the names of the locks which shall be acquired
410 (special lock names, or instance/node names)
411 @param shared: whether to acquire in shared mode; by default an
412 exclusive lock will be acquired
413 @param blocking: whether to block while trying to acquire or to
414 operate in try-lock mode (this locking mode is not supported yet)
416 @return: True when all the locks are successfully acquired
418 @raise errors.LockError: when any lock we try to acquire has
419 been deleted before we succeed. In this case none of the
420 locks requested will be acquired.
424 # We don't have non-blocking mode for now
425 raise NotImplementedError
427 # Check we don't already own locks at this level
428 assert not self._is_owned(), "Cannot acquire locks in the same set twice"
431 # If no names are given acquire the whole set by not letting new names
432 # being added before we release, and getting the current list of names.
433 # Some of them may then be deleted later, but we'll cope with this.
435 # We'd like to acquire this lock in a shared way, as it's nice if
436 # everybody else can use the instances at the same time. If are acquiring
437 # them exclusively though they won't be able to do this anyway, though,
438 # so we'll get the list lock exclusively as well in order to be able to
439 # do add() on the set while owning it.
440 self.__lock.acquire(shared=shared)
442 # note we own the set-lock
444 names = self.__names()
446 # We shouldn't have problems adding the lock to the owners list, but
447 # if we did we'll try to release this lock and re-raise exception.
448 # Of course something is going to be really wrong, after this.
449 self.__lock.release()
453 # Support passing in a single resource to acquire rather than many
454 if isinstance(names, basestring):
457 names = sorted(names)
460 # First we look the locks up on __lockdict. We have no way of being sure
461 # they will still be there after, but this makes it a lot faster should
462 # just one of them be the already wrong
463 for lname in utils.UniqueSequence(names):
465 lock = self.__lockdict[lname] # raises KeyError if lock is not there
466 acquire_list.append((lname, lock))
468 if self.__lock._is_owned():
469 # We are acquiring all the set, it doesn't matter if this
470 # particular element is not there anymore.
473 raise errors.LockError('non-existing lock in set (%s)' % lname)
475 # This will hold the locknames we effectively acquired.
477 # Now acquire_list contains a sorted list of resources and locks we want.
478 # In order to get them we loop on this (private) list and acquire() them.
479 # We gave no real guarantee they will still exist till this is done but
480 # .acquire() itself is safe and will alert us if the lock gets deleted.
481 for (lname, lock) in acquire_list:
483 lock.acquire(shared=shared) # raises LockError if the lock is deleted
484 # now the lock cannot be deleted, we have it!
485 self._add_owned(name=lname)
487 except (errors.LockError):
488 if self.__lock._is_owned():
489 # We are acquiring all the set, it doesn't matter if this
490 # particular element is not there anymore.
494 for lname in self._list_owned():
495 self.__lockdict[lname].release()
496 self._del_owned(name=lname)
497 raise errors.LockError('non-existing lock in set (%s)' % name_fail)
499 # We shouldn't have problems adding the lock to the owners list, but
500 # if we did we'll try to release this lock and re-raise exception.
501 # Of course something is going to be really wrong, after this.
507 # If something went wrong and we had the set-lock let's release it...
508 if self.__lock._is_owned():
509 self.__lock.release()
514 def release(self, names=None):
515 """Release a set of resource locks, at the same level.
517 You must have acquired the locks, either in shared or in exclusive mode,
518 before releasing them.
520 @param names: the names of the locks which shall be released
521 (defaults to all the locks acquired at that level).
524 assert self._is_owned(), "release() on lock set while not owner"
526 # Support passing in a single resource to release rather than many
527 if isinstance(names, basestring):
531 names = self._list_owned()
534 assert self._list_owned().issuperset(names), (
535 "release() on unheld resources %s" %
536 names.difference(self._list_owned()))
538 # First of all let's release the "all elements" lock, if set.
539 # After this 'add' can work again
540 if self.__lock._is_owned():
541 self.__lock.release()
544 for lockname in names:
545 # If we are sure the lock doesn't leave __lockdict without being
546 # exclusively held we can do this...
547 self.__lockdict[lockname].release()
548 self._del_owned(name=lockname)
550 def add(self, names, acquired=0, shared=0):
551 """Add a new set of elements to the set
553 @param names: names of the new elements to add
554 @param acquired: pre-acquire the new resource?
555 @param shared: is the pre-acquisition shared?
558 # Check we don't already own locks at this level
559 assert not self._is_owned() or self.__lock._is_owned(shared=0), \
560 "Cannot add locks if the set is only partially owned, or shared"
562 # Support passing in a single resource to add rather than many
563 if isinstance(names, basestring):
566 # If we don't already own the set-level lock acquired in an exclusive way
567 # we'll get it and note we need to release it later.
569 if not self.__lock._is_owned():
571 self.__lock.acquire()
574 invalid_names = set(self.__names()).intersection(names)
576 # This must be an explicit raise, not an assert, because assert is
577 # turned off when using optimization, and this can happen because of
578 # concurrency even if the user doesn't want it.
579 raise errors.LockError("duplicate add() (%s)" % invalid_names)
581 for lockname in names:
585 lock.acquire(shared=shared)
586 # now the lock cannot be deleted, we have it!
588 self._add_owned(name=lockname)
590 # We shouldn't have problems adding the lock to the owners list,
591 # but if we did we'll try to release this lock and re-raise
592 # exception. Of course something is going to be really wrong,
593 # after this. On the other hand the lock hasn't been added to the
594 # __lockdict yet so no other threads should be pending on it. This
595 # release is just a safety measure.
599 self.__lockdict[lockname] = lock
602 # Only release __lock if we were not holding it previously.
604 self.__lock.release()
608 def remove(self, names, blocking=1):
609 """Remove elements from the lock set.
611 You can either not hold anything in the lockset or already hold a superset
612 of the elements you want to delete, exclusively.
614 @param names: names of the resource to remove.
615 @param blocking: whether to block while trying to acquire or to
616 operate in try-lock mode (this locking mode is not supported
617 yet unless you are already holding exclusively the locks)
619 @return:: a list of locks which we removed; the list is always
620 equal to the names list if we were holding all the locks
624 if not blocking and not self._is_owned():
625 # We don't have non-blocking mode for now
626 raise NotImplementedError
628 # Support passing in a single resource to remove rather than many
629 if isinstance(names, basestring):
632 # If we own any subset of this lock it must be a superset of what we want
633 # to delete. The ownership must also be exclusive, but that will be checked
634 # by the lock itself.
635 assert not self._is_owned() or self._list_owned().issuperset(names), (
636 "remove() on acquired lockset while not owning all elements")
641 # Calling delete() acquires the lock exclusively if we don't already own
642 # it, and causes all pending and subsequent lock acquires to fail. It's
643 # fine to call it out of order because delete() also implies release(),
644 # and the assertion above guarantees that if we either already hold
645 # everything we want to delete, or we hold none.
647 self.__lockdict[lname].delete()
648 removed.append(lname)
649 except (KeyError, errors.LockError):
650 # This cannot happen if we were already holding it, verify:
651 assert not self._is_owned(), "remove failed while holding lockset"
653 # If no LockError was raised we are the ones who deleted the lock.
654 # This means we can safely remove it from lockdict, as any further or
655 # pending delete() or acquire() will fail (and nobody can have the lock
656 # since before our call to delete()).
658 # This is done in an else clause because if the exception was thrown
659 # it's the job of the one who actually deleted it.
660 del self.__lockdict[lname]
661 # And let's remove it from our private list if we owned it.
663 self._del_owned(name=lname)
668 # Locking levels, must be acquired in increasing order.
670 # - at level LEVEL_CLUSTER resides the Big Ganeti Lock (BGL) which must be
671 # acquired before performing any operation, either in shared or in exclusive
672 # mode. acquiring the BGL in exclusive mode is discouraged and should be
674 # - at levels LEVEL_NODE and LEVEL_INSTANCE reside node and instance locks.
675 # If you need more than one node, or more than one instance, acquire them at
681 LEVELS = [LEVEL_CLUSTER,
685 # Lock levels which are modifiable
686 LEVELS_MOD = [LEVEL_NODE, LEVEL_INSTANCE]
688 # Constant for the big ganeti lock
692 class GanetiLockManager:
693 """The Ganeti Locking Library
695 The purpose of this small library is to manage locking for ganeti clusters
696 in a central place, while at the same time doing dynamic checks against
697 possible deadlocks. It will also make it easier to transition to a different
698 lock type should we migrate away from python threads.
703 def __init__(self, nodes=None, instances=None):
704 """Constructs a new GanetiLockManager object.
706 There should be only a GanetiLockManager object at any time, so this
707 function raises an error if this is not the case.
709 @param nodes: list of node names
710 @param instances: list of instance names
713 assert self.__class__._instance is None, \
714 "double GanetiLockManager instance"
716 self.__class__._instance = self
718 # The keyring contains all the locks, at their level and in the correct
721 LEVEL_CLUSTER: LockSet([BGL]),
722 LEVEL_NODE: LockSet(nodes),
723 LEVEL_INSTANCE: LockSet(instances),
726 def _names(self, level):
727 """List the lock names at the given level.
729 This can be used for debugging/testing purposes.
731 @param level: the level whose list of locks to get
734 assert level in LEVELS, "Invalid locking level %s" % level
735 return self.__keyring[level]._names()
737 def _is_owned(self, level):
738 """Check whether we are owning locks at the given level
741 return self.__keyring[level]._is_owned()
745 def _list_owned(self, level):
746 """Get the set of owned locks at the given level
749 return self.__keyring[level]._list_owned()
751 def _upper_owned(self, level):
752 """Check that we don't own any lock at a level greater than the given one.
755 # This way of checking only works if LEVELS[i] = i, which we check for in
757 return utils.any((self._is_owned(l) for l in LEVELS[level + 1:]))
759 def _BGL_owned(self): # pylint: disable-msg=C0103
760 """Check if the current thread owns the BGL.
762 Both an exclusive or a shared acquisition work.
765 return BGL in self.__keyring[LEVEL_CLUSTER]._list_owned()
768 def _contains_BGL(level, names): # pylint: disable-msg=C0103
769 """Check if the level contains the BGL.
771 Check if acting on the given level and set of names will change
772 the status of the Big Ganeti Lock.
775 return level == LEVEL_CLUSTER and (names is None or BGL in names)
777 def acquire(self, level, names, blocking=1, shared=0):
778 """Acquire a set of resource locks, at the same level.
780 @param level: the level at which the locks shall be acquired;
781 it must be a member of LEVELS.
782 @param names: the names of the locks which shall be acquired
783 (special lock names, or instance/node names)
784 @param shared: whether to acquire in shared mode; by default
785 an exclusive lock will be acquired
786 @param blocking: whether to block while trying to acquire or to
787 operate in try-lock mode (this locking mode is not supported yet)
790 assert level in LEVELS, "Invalid locking level %s" % level
792 # Check that we are either acquiring the Big Ganeti Lock or we already own
793 # it. Some "legacy" opcodes need to be sure they are run non-concurrently
794 # so even if we've migrated we need to at least share the BGL to be
795 # compatible with them. Of course if we own the BGL exclusively there's no
796 # point in acquiring any other lock, unless perhaps we are half way through
797 # the migration of the current opcode.
798 assert (self._contains_BGL(level, names) or self._BGL_owned()), (
799 "You must own the Big Ganeti Lock before acquiring any other")
801 # Check we don't own locks at the same or upper levels.
802 assert not self._upper_owned(level), ("Cannot acquire locks at a level"
803 " while owning some at a greater one")
805 # Acquire the locks in the set.
806 return self.__keyring[level].acquire(names, shared=shared,
809 def release(self, level, names=None):
810 """Release a set of resource locks, at the same level.
812 You must have acquired the locks, either in shared or in exclusive
813 mode, before releasing them.
815 @param level: the level at which the locks shall be released;
816 it must be a member of LEVELS
817 @param names: the names of the locks which shall be released
818 (defaults to all the locks acquired at that level)
821 assert level in LEVELS, "Invalid locking level %s" % level
822 assert (not self._contains_BGL(level, names) or
823 not self._upper_owned(LEVEL_CLUSTER)), (
824 "Cannot release the Big Ganeti Lock while holding something"
827 # Release will complain if we don't own the locks already
828 return self.__keyring[level].release(names)
830 def add(self, level, names, acquired=0, shared=0):
831 """Add locks at the specified level.
833 @param level: the level at which the locks shall be added;
834 it must be a member of LEVELS_MOD.
835 @param names: names of the locks to acquire
836 @param acquired: whether to acquire the newly added locks
837 @param shared: whether the acquisition will be shared
840 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
841 assert self._BGL_owned(), ("You must own the BGL before performing other"
843 assert not self._upper_owned(level), ("Cannot add locks at a level"
844 " while owning some at a greater one")
845 return self.__keyring[level].add(names, acquired=acquired, shared=shared)
847 def remove(self, level, names, blocking=1):
848 """Remove locks from the specified level.
850 You must either already own the locks you are trying to remove
851 exclusively or not own any lock at an upper level.
853 @param level: the level at which the locks shall be removed;
854 it must be a member of LEVELS_MOD
855 @param names: the names of the locks which shall be removed
856 (special lock names, or instance/node names)
857 @param blocking: whether to block while trying to operate in
858 try-lock mode (this locking mode is not supported yet)
861 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
862 assert self._BGL_owned(), ("You must own the BGL before performing other"
864 # Check we either own the level or don't own anything from here
865 # up. LockSet.remove() will check the case in which we don't own
866 # all the needed resources, or we have a shared ownership.
867 assert self._is_owned(level) or not self._upper_owned(level), (
868 "Cannot remove locks at a level while not owning it or"
869 " owning some at a greater one")
870 return self.__keyring[level].remove(names, blocking=blocking)