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 acquire(self, blocking=1, shared=0):
157 """Acquire a shared lock.
159 @param shared: whether to acquire in shared mode; by default an
160 exclusive lock will be acquired
161 @param blocking: whether to block while trying to acquire or to
162 operate in try-lock mode (this locking mode is not supported yet)
166 # We don't have non-blocking mode for now
167 raise NotImplementedError
169 self.__lock.acquire()
172 raise errors.LockError('deleted lock')
174 # We cannot acquire the lock if we already have it
175 assert not self.__is_owned(), "double acquire() on a non-recursive lock"
176 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
179 self.__nwait_shr += 1
182 # If there is an exclusive holder waiting we have to wait. We'll
183 # only do this once, though, when we start waiting for the lock. Then
184 # we'll just wait while there are no exclusive holders.
185 if self.__nwait_exc > 0:
186 # TODO: if !blocking...
188 self.__wait(self.__turn_shr)
190 while self.__exc is not None:
192 # TODO: if !blocking...
193 self.__wait(self.__turn_shr)
195 self.__shr.add(threading.currentThread())
197 # If we were waiting note that we passed
199 self.__npass_shr -= 1
202 self.__nwait_shr -= 1
204 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
206 # TODO: if !blocking...
207 # (or modify __exclusive_acquire for non-blocking mode)
208 self.__exclusive_acquire()
211 self.__lock.release()
216 """Release a Shared Lock.
218 You must have acquired the lock, either in shared or in exclusive mode,
219 before calling this function.
222 self.__lock.acquire()
224 assert self.__npass_shr >= 0, "Internal fairness condition weirdness"
225 # Autodetect release type
226 if self.__is_exclusive():
229 # An exclusive holder has just had the lock, time to put it in shared
230 # mode if there are shared holders waiting. Otherwise wake up the next
232 if self.__nwait_shr > 0:
233 # Make sure at least the ones which were blocked pass.
234 self.__npass_shr = self.__nwait_shr
235 self.__turn_shr.notifyAll()
236 elif self.__nwait_exc > 0:
237 self.__turn_exc.notify()
239 elif self.__is_sharer():
240 self.__shr.remove(threading.currentThread())
242 # If there are shared holders waiting (and not just scheduled to pass)
243 # there *must* be an exclusive holder waiting as well; otherwise what
244 # were they waiting for?
245 assert (self.__nwait_exc > 0 or self.__npass_shr == self.__nwait_shr), \
246 "Lock sharers waiting while no exclusive is queueing"
248 # If there are no more shared holders either in or scheduled to pass,
249 # and some exclusive holders are waiting let's wake one up.
250 if (len(self.__shr) == 0 and
251 self.__nwait_exc > 0 and
252 not self.__npass_shr > 0):
253 self.__turn_exc.notify()
256 assert False, "Cannot release non-owned lock"
259 self.__lock.release()
261 def delete(self, blocking=1):
262 """Delete a Shared Lock.
264 This operation will declare the lock for removal. First the lock will be
265 acquired in exclusive mode if you don't already own it, then the lock
266 will be put in a state where any future and pending acquire() fail.
268 @param blocking: whether to block while trying to acquire or to
269 operate in try-lock mode. this locking mode is not supported
270 yet unless you are already holding exclusively the lock.
273 self.__lock.acquire()
275 assert not self.__is_sharer(), "cannot delete() a lock while sharing it"
278 raise errors.LockError('deleted lock')
280 if not self.__is_exclusive():
282 # We don't have non-blocking mode for now
283 raise NotImplementedError
284 self.__exclusive_acquire()
286 self.__deleted = True
288 # Wake up everybody, they will fail acquiring the lock and
289 # raise an exception instead.
290 self.__turn_exc.notifyAll()
291 self.__turn_shr.notifyAll()
294 self.__lock.release()
297 # Whenever we want to acquire a full LockSet we pass None as the value to acquire.
298 # Hide this behing this nicely named constant.
303 """Implements a set of locks.
305 This abstraction implements a set of shared locks for the same resource type,
306 distinguished by name. The user can lock a subset of the resources and the
307 LockSet will take care of acquiring the locks always in the same order, thus
310 All the locks needed in the same set must be acquired together, though.
313 def __init__(self, members=None):
314 """Constructs a new LockSet.
316 @param members: initial members of the set
319 # Used internally to guarantee coherency.
320 self.__lock = SharedLock()
322 # The lockdict indexes the relationship name -> lock
323 # The order-of-locking is implied by the alphabetical order of names
326 if members is not None:
328 self.__lockdict[name] = SharedLock()
330 # The owner dict contains the set of locks each thread owns. For
331 # performance each thread can access its own key without a global lock on
332 # this structure. It is paramount though that *no* other type of access is
333 # done to this structure (eg. no looping over its keys). *_owner helper
334 # function are defined to guarantee access is correct, but in general never
335 # do anything different than __owners[threading.currentThread()], or there
340 """Is the current thread a current level owner?"""
341 return threading.currentThread() in self.__owners
343 def _add_owned(self, name=None):
344 """Note the current thread owns the given lock"""
346 if not self._is_owned():
347 self.__owners[threading.currentThread()] = set()
350 self.__owners[threading.currentThread()].add(name)
352 self.__owners[threading.currentThread()] = set([name])
355 def _del_owned(self, name=None):
356 """Note the current thread owns the given lock"""
359 self.__owners[threading.currentThread()].remove(name)
361 # Only remove the key if we don't hold the set-lock as well
362 if (not self.__lock._is_owned() and
363 not self.__owners[threading.currentThread()]):
364 del self.__owners[threading.currentThread()]
366 def _list_owned(self):
367 """Get the set of resource names owned by the current thread"""
369 return self.__owners[threading.currentThread()].copy()
374 """Return the current set of names.
376 Only call this function while holding __lock and don't iterate on the
377 result after releasing the lock.
380 return self.__lockdict.keys()
383 """Return a copy of the current set of elements.
385 Used only for debugging purposes.
388 # If we don't already own the set-level lock acquired
389 # we'll get it and note we need to release it later.
391 if not self.__lock._is_owned():
393 self.__lock.acquire(shared=1)
395 result = self.__names()
398 self.__lock.release()
401 def acquire(self, names, blocking=1, shared=0):
402 """Acquire a set of resource locks.
404 @param names: the names of the locks which shall be acquired
405 (special lock names, or instance/node names)
406 @param shared: whether to acquire in shared mode; by default an
407 exclusive lock will be acquired
408 @param blocking: whether to block while trying to acquire or to
409 operate in try-lock mode (this locking mode is not supported yet)
411 @return: True when all the locks are successfully acquired
413 @raise errors.LockError: when any lock we try to acquire has
414 been deleted before we succeed. In this case none of the
415 locks requested will be acquired.
419 # We don't have non-blocking mode for now
420 raise NotImplementedError
422 # Check we don't already own locks at this level
423 assert not self._is_owned(), "Cannot acquire locks in the same set twice"
426 # If no names are given acquire the whole set by not letting new names
427 # being added before we release, and getting the current list of names.
428 # Some of them may then be deleted later, but we'll cope with this.
430 # We'd like to acquire this lock in a shared way, as it's nice if
431 # everybody else can use the instances at the same time. If are acquiring
432 # them exclusively though they won't be able to do this anyway, though,
433 # so we'll get the list lock exclusively as well in order to be able to
434 # do add() on the set while owning it.
435 self.__lock.acquire(shared=shared)
437 # note we own the set-lock
439 names = self.__names()
441 # We shouldn't have problems adding the lock to the owners list, but
442 # if we did we'll try to release this lock and re-raise exception.
443 # Of course something is going to be really wrong, after this.
444 self.__lock.release()
448 # Support passing in a single resource to acquire rather than many
449 if isinstance(names, basestring):
455 # First we look the locks up on __lockdict. We have no way of being sure
456 # they will still be there after, but this makes it a lot faster should
457 # just one of them be the already wrong
458 for lname in utils.UniqueSequence(names):
460 lock = self.__lockdict[lname] # raises KeyError if lock is not there
461 acquire_list.append((lname, lock))
463 if self.__lock._is_owned():
464 # We are acquiring all the set, it doesn't matter if this particular
465 # element is not there anymore.
468 raise errors.LockError('non-existing lock in set (%s)' % lname)
470 # This will hold the locknames we effectively acquired.
472 # Now acquire_list contains a sorted list of resources and locks we want.
473 # In order to get them we loop on this (private) list and acquire() them.
474 # We gave no real guarantee they will still exist till this is done but
475 # .acquire() itself is safe and will alert us if the lock gets deleted.
476 for (lname, lock) in acquire_list:
478 lock.acquire(shared=shared) # raises LockError if the lock is deleted
479 # now the lock cannot be deleted, we have it!
480 self._add_owned(name=lname)
482 except (errors.LockError):
483 if self.__lock._is_owned():
484 # We are acquiring all the set, it doesn't matter if this particular
485 # element is not there anymore.
489 for lname in self._list_owned():
490 self.__lockdict[lname].release()
491 self._del_owned(name=lname)
492 raise errors.LockError('non-existing lock in set (%s)' % name_fail)
494 # We shouldn't have problems adding the lock to the owners list, but
495 # if we did we'll try to release this lock and re-raise exception.
496 # Of course something is going to be really wrong, after this.
502 # If something went wrong and we had the set-lock let's release it...
503 if self.__lock._is_owned():
504 self.__lock.release()
509 def release(self, names=None):
510 """Release a set of resource locks, at the same level.
512 You must have acquired the locks, either in shared or in exclusive mode,
513 before releasing them.
515 @param names: the names of the locks which shall be released
516 (defaults to all the locks acquired at that level).
519 assert self._is_owned(), "release() on lock set while not owner"
521 # Support passing in a single resource to release rather than many
522 if isinstance(names, basestring):
526 names = self._list_owned()
529 assert self._list_owned().issuperset(names), (
530 "release() on unheld resources %s" %
531 names.difference(self._list_owned()))
533 # First of all let's release the "all elements" lock, if set.
534 # After this 'add' can work again
535 if self.__lock._is_owned():
536 self.__lock.release()
539 for lockname in names:
540 # If we are sure the lock doesn't leave __lockdict without being
541 # exclusively held we can do this...
542 self.__lockdict[lockname].release()
543 self._del_owned(name=lockname)
545 def add(self, names, acquired=0, shared=0):
546 """Add a new set of elements to the set
548 @param names: names of the new elements to add
549 @param acquired: pre-acquire the new resource?
550 @param shared: is the pre-acquisition shared?
553 # Check we don't already own locks at this level
554 assert not self._is_owned() or self.__lock._is_owned(shared=0), \
555 "Cannot add locks if the set is only partially owned, or shared"
557 # Support passing in a single resource to add rather than many
558 if isinstance(names, basestring):
561 # If we don't already own the set-level lock acquired in an exclusive way
562 # we'll get it and note we need to release it later.
564 if not self.__lock._is_owned():
566 self.__lock.acquire()
569 invalid_names = set(self.__names()).intersection(names)
571 # This must be an explicit raise, not an assert, because assert is
572 # turned off when using optimization, and this can happen because of
573 # concurrency even if the user doesn't want it.
574 raise errors.LockError("duplicate add() (%s)" % invalid_names)
576 for lockname in names:
580 lock.acquire(shared=shared)
581 # now the lock cannot be deleted, we have it!
583 self._add_owned(name=lockname)
585 # We shouldn't have problems adding the lock to the owners list,
586 # but if we did we'll try to release this lock and re-raise
587 # exception. Of course something is going to be really wrong,
588 # after this. On the other hand the lock hasn't been added to the
589 # __lockdict yet so no other threads should be pending on it. This
590 # release is just a safety measure.
594 self.__lockdict[lockname] = lock
597 # Only release __lock if we were not holding it previously.
599 self.__lock.release()
603 def remove(self, names, blocking=1):
604 """Remove elements from the lock set.
606 You can either not hold anything in the lockset or already hold a superset
607 of the elements you want to delete, exclusively.
609 @param names: names of the resource to remove.
610 @param blocking: whether to block while trying to acquire or to
611 operate in try-lock mode (this locking mode is not supported
612 yet unless you are already holding exclusively the locks)
614 @return:: a list of locks which we removed; the list is always
615 equal to the names list if we were holding all the locks
619 if not blocking and not self._is_owned():
620 # We don't have non-blocking mode for now
621 raise NotImplementedError
623 # Support passing in a single resource to remove rather than many
624 if isinstance(names, basestring):
627 # If we own any subset of this lock it must be a superset of what we want
628 # to delete. The ownership must also be exclusive, but that will be checked
629 # by the lock itself.
630 assert not self._is_owned() or self._list_owned().issuperset(names), (
631 "remove() on acquired lockset while not owning all elements")
636 # Calling delete() acquires the lock exclusively if we don't already own
637 # it, and causes all pending and subsequent lock acquires to fail. It's
638 # fine to call it out of order because delete() also implies release(),
639 # and the assertion above guarantees that if we either already hold
640 # everything we want to delete, or we hold none.
642 self.__lockdict[lname].delete()
643 removed.append(lname)
644 except (KeyError, errors.LockError):
645 # This cannot happen if we were already holding it, verify:
646 assert not self._is_owned(), "remove failed while holding lockset"
648 # If no LockError was raised we are the ones who deleted the lock.
649 # This means we can safely remove it from lockdict, as any further or
650 # pending delete() or acquire() will fail (and nobody can have the lock
651 # since before our call to delete()).
653 # This is done in an else clause because if the exception was thrown
654 # it's the job of the one who actually deleted it.
655 del self.__lockdict[lname]
656 # And let's remove it from our private list if we owned it.
658 self._del_owned(name=lname)
663 # Locking levels, must be acquired in increasing order.
665 # - at level LEVEL_CLUSTER resides the Big Ganeti Lock (BGL) which must be
666 # acquired before performing any operation, either in shared or in exclusive
667 # mode. acquiring the BGL in exclusive mode is discouraged and should be
669 # - at levels LEVEL_NODE and LEVEL_INSTANCE reside node and instance locks.
670 # If you need more than one node, or more than one instance, acquire them at
676 LEVELS = [LEVEL_CLUSTER,
680 # Lock levels which are modifiable
681 LEVELS_MOD = [LEVEL_NODE, LEVEL_INSTANCE]
683 # Constant for the big ganeti lock
687 class GanetiLockManager:
688 """The Ganeti Locking Library
690 The purpouse of this small library is to manage locking for ganeti clusters
691 in a central place, while at the same time doing dynamic checks against
692 possible deadlocks. It will also make it easier to transition to a different
693 lock type should we migrate away from python threads.
698 def __init__(self, nodes=None, instances=None):
699 """Constructs a new GanetiLockManager object.
701 There should be only a GanetiLockManager object at any time, so this
702 function raises an error if this is not the case.
704 @param nodes: list of node names
705 @param instances: list of instance names
708 assert self.__class__._instance is None, \
709 "double GanetiLockManager instance"
711 self.__class__._instance = self
713 # The keyring contains all the locks, at their level and in the correct
716 LEVEL_CLUSTER: LockSet([BGL]),
717 LEVEL_NODE: LockSet(nodes),
718 LEVEL_INSTANCE: LockSet(instances),
721 def _names(self, level):
722 """List the lock names at the given level.
724 This can be used for debugging/testing purposes.
726 @param level: the level whose list of locks to get
729 assert level in LEVELS, "Invalid locking level %s" % level
730 return self.__keyring[level]._names()
732 def _is_owned(self, level):
733 """Check whether we are owning locks at the given level
736 return self.__keyring[level]._is_owned()
740 def _list_owned(self, level):
741 """Get the set of owned locks at the given level
744 return self.__keyring[level]._list_owned()
746 def _upper_owned(self, level):
747 """Check that we don't own any lock at a level greater than the given one.
750 # This way of checking only works if LEVELS[i] = i, which we check for in
752 return utils.any((self._is_owned(l) for l in LEVELS[level + 1:]))
754 def _BGL_owned(self):
755 """Check if the current thread owns the BGL.
757 Both an exclusive or a shared acquisition work.
760 return BGL in self.__keyring[LEVEL_CLUSTER]._list_owned()
762 def _contains_BGL(self, level, names):
763 """Check if the level contains the BGL.
765 Check if acting on the given level and set of names will change
766 the status of the Big Ganeti Lock.
769 return level == LEVEL_CLUSTER and (names is None or BGL in names)
771 def acquire(self, level, names, blocking=1, shared=0):
772 """Acquire a set of resource locks, at the same level.
774 @param level: the level at which the locks shall be acquired;
775 it must be a memmber of LEVELS.
776 @param names: the names of the locks which shall be acquired
777 (special lock names, or instance/node names)
778 @param shared: whether to acquire in shared mode; by default
779 an exclusive lock will be acquired
780 @param blocking: whether to block while trying to acquire or to
781 operate in try-lock mode (this locking mode is not supported yet)
784 assert level in LEVELS, "Invalid locking level %s" % level
786 # Check that we are either acquiring the Big Ganeti Lock or we already own
787 # it. Some "legacy" opcodes need to be sure they are run non-concurrently
788 # so even if we've migrated we need to at least share the BGL to be
789 # compatible with them. Of course if we own the BGL exclusively there's no
790 # point in acquiring any other lock, unless perhaps we are half way through
791 # the migration of the current opcode.
792 assert (self._contains_BGL(level, names) or self._BGL_owned()), (
793 "You must own the Big Ganeti Lock before acquiring any other")
795 # Check we don't own locks at the same or upper levels.
796 assert not self._upper_owned(level), ("Cannot acquire locks at a level"
797 " while owning some at a greater one")
799 # Acquire the locks in the set.
800 return self.__keyring[level].acquire(names, shared=shared,
803 def release(self, level, names=None):
804 """Release a set of resource locks, at the same level.
806 You must have acquired the locks, either in shared or in exclusive
807 mode, before releasing them.
809 @param level: the level at which the locks shall be released;
810 it must be a memmber of LEVELS
811 @param names: the names of the locks which shall be released
812 (defaults to all the locks acquired at that level)
815 assert level in LEVELS, "Invalid locking level %s" % level
816 assert (not self._contains_BGL(level, names) or
817 not self._upper_owned(LEVEL_CLUSTER)), (
818 "Cannot release the Big Ganeti Lock while holding something"
821 # Release will complain if we don't own the locks already
822 return self.__keyring[level].release(names)
824 def add(self, level, names, acquired=0, shared=0):
825 """Add locks at the specified level.
827 @param level: the level at which the locks shall be added;
828 it must be a memmber of LEVELS_MOD.
829 @param names: names of the locks to acquire
830 @param acquired: whether to acquire the newly added locks
831 @param shared: whether the acquisition will be shared
834 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
835 assert self._BGL_owned(), ("You must own the BGL before performing other"
837 assert not self._upper_owned(level), ("Cannot add locks at a level"
838 " while owning some at a greater one")
839 return self.__keyring[level].add(names, acquired=acquired, shared=shared)
841 def remove(self, level, names, blocking=1):
842 """Remove locks from the specified level.
844 You must either already own the locks you are trying to remove
845 exclusively or not own any lock at an upper level.
847 @param level: the level at which the locks shall be removed;
848 it must be a member of LEVELS_MOD
849 @param names: the names of the locks which shall be removed
850 (special lock names, or instance/node names)
851 @param blocking: whether to block while trying to operate in
852 try-lock mode (this locking mode is not supported yet)
855 assert level in LEVELS_MOD, "Invalid or immutable level %s" % level
856 assert self._BGL_owned(), ("You must own the BGL before performing other"
858 # Check we either own the level or don't own anything from here up.
859 # LockSet.remove() will check the case in which we don't own all the needed
860 # resources, or we have a shared ownership.
861 assert self._is_owned(level) or not self._upper_owned(level), (
862 "Cannot remove locks at a level while not owning it or"
863 " owning some at a greater one")
864 return self.__keyring[level].remove(names, blocking=blocking)