X-Git-Url: https://code.grnet.gr/git/ganeti-local/blobdiff_plain/08a6c581169019d14e8438de21762ccccb1843e0..d24bc00093906f86c8f054557b4fe6869cb75885:/lib/locking.py diff --git a/lib/locking.py b/lib/locking.py index 733fa84..3575085 100644 --- a/lib/locking.py +++ b/lib/locking.py @@ -1,7 +1,7 @@ # # -# Copyright (C) 2006, 2007 Google Inc. +# Copyright (C) 2006, 2007, 2008, 2009, 2010 Google Inc. # # This program is free software; you can redistribute it and/or modify # it under the terms of the GNU General Public License as published by @@ -20,25 +20,52 @@ """Module implementing the Ganeti locking code.""" -# pylint: disable-msg=W0613,W0201 +# pylint: disable-msg=W0212 +# W0212 since e.g. LockSet methods use (a lot) the internals of +# SharedLock + +import os +import select import threading -# Wouldn't it be better to define LockingError in the locking module? -# Well, for now that's how the rest of the code does it... +import errno +import weakref +import logging +import heapq +import operator + from ganeti import errors from ganeti import utils +from ganeti import compat +from ganeti import query + + +_EXCLUSIVE_TEXT = "exclusive" +_SHARED_TEXT = "shared" +_DELETED_TEXT = "deleted" + +_DEFAULT_PRIORITY = 0 -def ssynchronized(lock, shared=0): +def ssynchronized(mylock, shared=0): """Shared Synchronization decorator. Calls the function holding the given lock, either in exclusive or shared mode. It requires the passed lock to be a SharedLock (or support its semantics). + @type mylock: lockable object or string + @param mylock: lock to acquire or class member name of the lock to acquire + """ def wrap(fn): def sync_function(*args, **kwargs): + if isinstance(mylock, basestring): + assert args, "cannot ssynchronize on non-class method: self not found" + # args[0] is "self" + lock = getattr(args[0], mylock) + else: + lock = mylock lock.acquire(shared=shared) try: return fn(*args, **kwargs) @@ -48,43 +75,440 @@ def ssynchronized(lock, shared=0): return wrap -class SharedLock: +class _SingleNotifyPipeConditionWaiter(object): + """Helper class for SingleNotifyPipeCondition + + """ + __slots__ = [ + "_fd", + "_poller", + ] + + def __init__(self, poller, fd): + """Constructor for _SingleNotifyPipeConditionWaiter + + @type poller: select.poll + @param poller: Poller object + @type fd: int + @param fd: File descriptor to wait for + + """ + object.__init__(self) + self._poller = poller + self._fd = fd + + def __call__(self, timeout): + """Wait for something to happen on the pipe. + + @type timeout: float or None + @param timeout: Timeout for waiting (can be None) + + """ + running_timeout = utils.RunningTimeout(timeout, True) + + while True: + remaining_time = running_timeout.Remaining() + + if remaining_time is not None: + if remaining_time < 0.0: + break + + # Our calculation uses seconds, poll() wants milliseconds + remaining_time *= 1000 + + try: + result = self._poller.poll(remaining_time) + except EnvironmentError, err: + if err.errno != errno.EINTR: + raise + result = None + + # Check whether we were notified + if result and result[0][0] == self._fd: + break + + +class _BaseCondition(object): + """Base class containing common code for conditions. + + Some of this code is taken from python's threading module. + + """ + __slots__ = [ + "_lock", + "acquire", + "release", + "_is_owned", + "_acquire_restore", + "_release_save", + ] + + def __init__(self, lock): + """Constructor for _BaseCondition. + + @type lock: threading.Lock + @param lock: condition base lock + + """ + object.__init__(self) + + try: + self._release_save = lock._release_save + except AttributeError: + self._release_save = self._base_release_save + try: + self._acquire_restore = lock._acquire_restore + except AttributeError: + self._acquire_restore = self._base_acquire_restore + try: + self._is_owned = lock._is_owned + except AttributeError: + self._is_owned = self._base_is_owned + + self._lock = lock + + # Export the lock's acquire() and release() methods + self.acquire = lock.acquire + self.release = lock.release + + def _base_is_owned(self): + """Check whether lock is owned by current thread. + + """ + if self._lock.acquire(0): + self._lock.release() + return False + return True + + def _base_release_save(self): + self._lock.release() + + def _base_acquire_restore(self, _): + self._lock.acquire() + + def _check_owned(self): + """Raise an exception if the current thread doesn't own the lock. + + """ + if not self._is_owned(): + raise RuntimeError("cannot work with un-aquired lock") + + +class SingleNotifyPipeCondition(_BaseCondition): + """Condition which can only be notified once. + + This condition class uses pipes and poll, internally, to be able to wait for + notification with a timeout, without resorting to polling. It is almost + compatible with Python's threading.Condition, with the following differences: + - notifyAll can only be called once, and no wait can happen after that + - notify is not supported, only notifyAll + + """ + + __slots__ = [ + "_poller", + "_read_fd", + "_write_fd", + "_nwaiters", + "_notified", + ] + + _waiter_class = _SingleNotifyPipeConditionWaiter + + def __init__(self, lock): + """Constructor for SingleNotifyPipeCondition + + """ + _BaseCondition.__init__(self, lock) + self._nwaiters = 0 + self._notified = False + self._read_fd = None + self._write_fd = None + self._poller = None + + def _check_unnotified(self): + """Throws an exception if already notified. + + """ + if self._notified: + raise RuntimeError("cannot use already notified condition") + + def _Cleanup(self): + """Cleanup open file descriptors, if any. + + """ + if self._read_fd is not None: + os.close(self._read_fd) + self._read_fd = None + + if self._write_fd is not None: + os.close(self._write_fd) + self._write_fd = None + self._poller = None + + def wait(self, timeout=None): + """Wait for a notification. + + @type timeout: float or None + @param timeout: Waiting timeout (can be None) + + """ + self._check_owned() + self._check_unnotified() + + self._nwaiters += 1 + try: + if self._poller is None: + (self._read_fd, self._write_fd) = os.pipe() + self._poller = select.poll() + self._poller.register(self._read_fd, select.POLLHUP) + + wait_fn = self._waiter_class(self._poller, self._read_fd) + state = self._release_save() + try: + # Wait for notification + wait_fn(timeout) + finally: + # Re-acquire lock + self._acquire_restore(state) + finally: + self._nwaiters -= 1 + if self._nwaiters == 0: + self._Cleanup() + + def notifyAll(self): # pylint: disable-msg=C0103 + """Close the writing side of the pipe to notify all waiters. + + """ + self._check_owned() + self._check_unnotified() + self._notified = True + if self._write_fd is not None: + os.close(self._write_fd) + self._write_fd = None + + +class PipeCondition(_BaseCondition): + """Group-only non-polling condition with counters. + + This condition class uses pipes and poll, internally, to be able to wait for + notification with a timeout, without resorting to polling. It is almost + compatible with Python's threading.Condition, but only supports notifyAll and + non-recursive locks. As an additional features it's able to report whether + there are any waiting threads. + + """ + __slots__ = [ + "_waiters", + "_single_condition", + ] + + _single_condition_class = SingleNotifyPipeCondition + + def __init__(self, lock): + """Initializes this class. + + """ + _BaseCondition.__init__(self, lock) + self._waiters = set() + self._single_condition = self._single_condition_class(self._lock) + + def wait(self, timeout=None): + """Wait for a notification. + + @type timeout: float or None + @param timeout: Waiting timeout (can be None) + + """ + self._check_owned() + + # Keep local reference to the pipe. It could be replaced by another thread + # notifying while we're waiting. + cond = self._single_condition + + self._waiters.add(threading.currentThread()) + try: + cond.wait(timeout) + finally: + self._check_owned() + self._waiters.remove(threading.currentThread()) + + def notifyAll(self): # pylint: disable-msg=C0103 + """Notify all currently waiting threads. + + """ + self._check_owned() + self._single_condition.notifyAll() + self._single_condition = self._single_condition_class(self._lock) + + def get_waiting(self): + """Returns a list of all waiting threads. + + """ + self._check_owned() + + return self._waiters + + def has_waiting(self): + """Returns whether there are active waiters. + + """ + self._check_owned() + + return bool(self._waiters) + + +class _PipeConditionWithMode(PipeCondition): + __slots__ = [ + "shared", + ] + + def __init__(self, lock, shared): + """Initializes this class. + + """ + self.shared = shared + PipeCondition.__init__(self, lock) + + +class SharedLock(object): """Implements a shared lock. - Multiple threads can acquire the lock in a shared way, calling - acquire_shared(). In order to acquire the lock in an exclusive way threads - can call acquire_exclusive(). + Multiple threads can acquire the lock in a shared way by calling + C{acquire(shared=1)}. In order to acquire the lock in an exclusive way + threads can call C{acquire(shared=0)}. - The lock prevents starvation but does not guarantee that threads will acquire - the shared lock in the order they queued for it, just that they will - eventually do so. + Notes on data structures: C{__pending} contains a priority queue (heapq) of + all pending acquires: C{[(priority1: prioqueue1), (priority2: prioqueue2), + ...]}. Each per-priority queue contains a normal in-order list of conditions + to be notified when the lock can be acquired. Shared locks are grouped + together by priority and the condition for them is stored in + C{__pending_shared} if it already exists. C{__pending_by_prio} keeps + references for the per-priority queues indexed by priority for faster access. + + @type name: string + @ivar name: the name of the lock """ - def __init__(self): - """Construct a new SharedLock""" - # we have two conditions, c_shr and c_exc, sharing the same lock. + __slots__ = [ + "__weakref__", + "__deleted", + "__exc", + "__lock", + "__pending", + "__pending_by_prio", + "__pending_shared", + "__shr", + "name", + ] + + __condition_class = _PipeConditionWithMode + + def __init__(self, name, monitor=None): + """Construct a new SharedLock. + + @param name: the name of the lock + @type monitor: L{LockMonitor} + @param monitor: Lock monitor with which to register + + """ + object.__init__(self) + + self.name = name + + # Internal lock self.__lock = threading.Lock() - self.__turn_shr = threading.Condition(self.__lock) - self.__turn_exc = threading.Condition(self.__lock) - # current lock holders + # Queue containing waiting acquires + self.__pending = [] + self.__pending_by_prio = {} + self.__pending_shared = {} + + # Current lock holders self.__shr = set() self.__exc = None - # lock waiters - self.__nwait_exc = 0 - self.__nwait_shr = 0 - self.__npass_shr = 0 - # is this lock in the deleted state? self.__deleted = False + # Register with lock monitor + if monitor: + monitor.RegisterLock(self) + + def GetInfo(self, requested): + """Retrieves information for querying locks. + + @type requested: set + @param requested: Requested information, see C{query.LQ_*} + + """ + self.__lock.acquire() + try: + # Note: to avoid unintentional race conditions, no references to + # modifiable objects should be returned unless they were created in this + # function. + mode = None + owner_names = None + + if query.LQ_MODE in requested: + if self.__deleted: + mode = _DELETED_TEXT + assert not (self.__exc or self.__shr) + elif self.__exc: + mode = _EXCLUSIVE_TEXT + elif self.__shr: + mode = _SHARED_TEXT + + # Current owner(s) are wanted + if query.LQ_OWNER in requested: + if self.__exc: + owner = [self.__exc] + else: + owner = self.__shr + + if owner: + assert not self.__deleted + owner_names = [i.getName() for i in owner] + + # Pending acquires are wanted + if query.LQ_PENDING in requested: + pending = [] + + # Sorting instead of copying and using heaq functions for simplicity + for (_, prioqueue) in sorted(self.__pending): + for cond in prioqueue: + if cond.shared: + pendmode = _SHARED_TEXT + else: + pendmode = _EXCLUSIVE_TEXT + + # List of names will be sorted in L{query._GetLockPending} + pending.append((pendmode, [i.getName() + for i in cond.get_waiting()])) + else: + pending = None + + return (self.name, mode, owner_names, pending) + finally: + self.__lock.release() + + def __check_deleted(self): + """Raises an exception if the lock has been deleted. + + """ + if self.__deleted: + raise errors.LockError("Deleted lock %s" % self.name) + def __is_sharer(self): - """Is the current thread sharing the lock at this time?""" + """Is the current thread sharing the lock at this time? + + """ return threading.currentThread() in self.__shr def __is_exclusive(self): - """Is the current thread holding the lock exclusively at this time?""" + """Is the current thread holding the lock exclusively at this time? + + """ return threading.currentThread() == self.__exc def __is_owned(self, shared=-1): @@ -104,116 +528,198 @@ class SharedLock: def _is_owned(self, shared=-1): """Is the current thread somehow owning the lock at this time? - Args: - shared: - < 0: check for any type of ownership (default) - 0: check for exclusive ownership - > 0: check for shared ownership + @param shared: + - < 0: check for any type of ownership (default) + - 0: check for exclusive ownership + - > 0: check for shared ownership """ self.__lock.acquire() try: - result = self.__is_owned(shared=shared) + return self.__is_owned(shared=shared) finally: self.__lock.release() - return result + def _count_pending(self): + """Returns the number of pending acquires. - def __wait(self, c): - """Wait on the given condition, and raise an exception if the current lock - is declared deleted in the meantime. - - Args: - c: condition to wait on + @rtype: int """ - c.wait() - if self.__deleted: - raise errors.LockError('deleted lock') + self.__lock.acquire() + try: + return sum(len(prioqueue) for (_, prioqueue) in self.__pending) + finally: + self.__lock.release() - def __exclusive_acquire(self): - """Acquire the lock exclusively. + def _check_empty(self): + """Checks whether there are any pending acquires. - This is a private function that presumes you are already holding the - internal lock. It's defined separately to avoid code duplication between - acquire() and delete() + @rtype: bool """ - self.__nwait_exc += 1 + self.__lock.acquire() try: - # This is to save ourselves from a nasty race condition that could - # theoretically make the sharers starve. - if self.__nwait_shr > 0 or self.__nwait_exc > 1: - self.__wait(self.__turn_exc) + # Order is important: __find_first_pending_queue modifies __pending + return not (self.__find_first_pending_queue() or + self.__pending or + self.__pending_by_prio or + self.__pending_shared) + finally: + self.__lock.release() - while len(self.__shr) > 0 or self.__exc is not None: - self.__wait(self.__turn_exc) + def __do_acquire(self, shared): + """Actually acquire the lock. + """ + if shared: + self.__shr.add(threading.currentThread()) + else: self.__exc = threading.currentThread() - finally: - self.__nwait_exc -= 1 - assert self.__npass_shr == 0, "SharedLock: internal fairness violation" + def __can_acquire(self, shared): + """Determine whether lock can be acquired. + + """ + if shared: + return self.__exc is None + else: + return len(self.__shr) == 0 and self.__exc is None + + def __find_first_pending_queue(self): + """Tries to find the topmost queued entry with pending acquires. + + Removes empty entries while going through the list. + + """ + while self.__pending: + (priority, prioqueue) = self.__pending[0] + + if not prioqueue: + heapq.heappop(self.__pending) + del self.__pending_by_prio[priority] + assert priority not in self.__pending_shared + continue + + if prioqueue: + return prioqueue + + return None + + def __is_on_top(self, cond): + """Checks whether the passed condition is on top of the queue. + + The caller must make sure the queue isn't empty. - def acquire(self, blocking=1, shared=0): + """ + return cond == self.__find_first_pending_queue()[0] + + def __acquire_unlocked(self, shared, timeout, priority): """Acquire a shared lock. - Args: - shared: whether to acquire in shared mode. By default an exclusive lock - will be acquired. - blocking: whether to block while trying to acquire or to operate in - try-lock mode. this locking mode is not supported yet. + @param shared: whether to acquire in shared mode; by default an + exclusive lock will be acquired + @param timeout: maximum waiting time before giving up + @type priority: integer + @param priority: Priority for acquiring lock """ - if not blocking: - # We don't have non-blocking mode for now - raise NotImplementedError + self.__check_deleted() - self.__lock.acquire() - try: - if self.__deleted: - raise errors.LockError('deleted lock') + # We cannot acquire the lock if we already have it + assert not self.__is_owned(), ("double acquire() on a non-recursive lock" + " %s" % self.name) + + # Remove empty entries from queue + self.__find_first_pending_queue() - # We cannot acquire the lock if we already have it - assert not self.__is_owned(), "double acquire() on a non-recursive lock" - assert self.__npass_shr >= 0, "Internal fairness condition weirdness" + # Check whether someone else holds the lock or there are pending acquires. + if not self.__pending and self.__can_acquire(shared): + # Apparently not, can acquire lock directly. + self.__do_acquire(shared) + return True + + prioqueue = self.__pending_by_prio.get(priority, None) + + if shared: + # Try to re-use condition for shared acquire + wait_condition = self.__pending_shared.get(priority, None) + assert (wait_condition is None or + (wait_condition.shared and wait_condition in prioqueue)) + else: + wait_condition = None + + if wait_condition is None: + if prioqueue is None: + assert priority not in self.__pending_by_prio + + prioqueue = [] + heapq.heappush(self.__pending, (priority, prioqueue)) + self.__pending_by_prio[priority] = prioqueue + + wait_condition = self.__condition_class(self.__lock, shared) + prioqueue.append(wait_condition) if shared: - self.__nwait_shr += 1 - try: - wait = False - # If there is an exclusive holder waiting we have to wait. We'll - # only do this once, though, when we start waiting for the lock. Then - # we'll just wait while there are no exclusive holders. - if self.__nwait_exc > 0: - # TODO: if !blocking... - wait = True - self.__wait(self.__turn_shr) - - while self.__exc is not None: - wait = True - # TODO: if !blocking... - self.__wait(self.__turn_shr) - - self.__shr.add(threading.currentThread()) - - # If we were waiting note that we passed - if wait: - self.__npass_shr -= 1 - - finally: - self.__nwait_shr -= 1 - - assert self.__npass_shr >= 0, "Internal fairness condition weirdness" - else: - # TODO: if !blocking... - # (or modify __exclusive_acquire for non-blocking mode) - self.__exclusive_acquire() + # Keep reference for further shared acquires on same priority. This is + # better than trying to find it in the list of pending acquires. + assert priority not in self.__pending_shared + self.__pending_shared[priority] = wait_condition + try: + # Wait until we become the topmost acquire in the queue or the timeout + # expires. + # TODO: Decrease timeout with spurious notifications + while not (self.__is_on_top(wait_condition) and + self.__can_acquire(shared)): + # Wait for notification + wait_condition.wait(timeout) + self.__check_deleted() + + # A lot of code assumes blocking acquires always succeed. Loop + # internally for that case. + if timeout is not None: + break + + if self.__is_on_top(wait_condition) and self.__can_acquire(shared): + self.__do_acquire(shared) + return True finally: - self.__lock.release() + # Remove condition from queue if there are no more waiters + if not wait_condition.has_waiting(): + prioqueue.remove(wait_condition) + if wait_condition.shared: + del self.__pending_shared[priority] - return True + return False + + def acquire(self, shared=0, timeout=None, priority=None, + test_notify=None): + """Acquire a shared lock. + + @type shared: integer (0/1) used as a boolean + @param shared: whether to acquire in shared mode; by default an + exclusive lock will be acquired + @type timeout: float + @param timeout: maximum waiting time before giving up + @type priority: integer + @param priority: Priority for acquiring lock + @type test_notify: callable or None + @param test_notify: Special callback function for unittesting + + """ + if priority is None: + priority = _DEFAULT_PRIORITY + + self.__lock.acquire() + try: + # We already got the lock, notify now + if __debug__ and callable(test_notify): + test_notify() + + return self.__acquire_unlocked(shared, timeout, priority) + finally: + self.__lock.release() def release(self): """Release a Shared Lock. @@ -224,79 +730,90 @@ class SharedLock: """ self.__lock.acquire() try: - assert self.__npass_shr >= 0, "Internal fairness condition weirdness" + assert self.__is_exclusive() or self.__is_sharer(), \ + "Cannot release non-owned lock" + # Autodetect release type if self.__is_exclusive(): self.__exc = None - - # An exclusive holder has just had the lock, time to put it in shared - # mode if there are shared holders waiting. Otherwise wake up the next - # exclusive holder. - if self.__nwait_shr > 0: - # Make sure at least the ones which were blocked pass. - self.__npass_shr = self.__nwait_shr - self.__turn_shr.notifyAll() - elif self.__nwait_exc > 0: - self.__turn_exc.notify() - - elif self.__is_sharer(): + else: self.__shr.remove(threading.currentThread()) - # If there are shared holders waiting (and not just scheduled to pass) - # there *must* be an exclusive holder waiting as well; otherwise what - # were they waiting for? - assert (self.__nwait_exc > 0 or self.__npass_shr == self.__nwait_shr), \ - "Lock sharers waiting while no exclusive is queueing" - - # If there are no more shared holders either in or scheduled to pass, - # and some exclusive holders are waiting let's wake one up. - if (len(self.__shr) == 0 and - self.__nwait_exc > 0 and - not self.__npass_shr > 0): - self.__turn_exc.notify() - - else: - assert False, "Cannot release non-owned lock" + # Notify topmost condition in queue + prioqueue = self.__find_first_pending_queue() + if prioqueue: + prioqueue[0].notifyAll() finally: self.__lock.release() - def delete(self, blocking=1): + def delete(self, timeout=None, priority=None): """Delete a Shared Lock. This operation will declare the lock for removal. First the lock will be acquired in exclusive mode if you don't already own it, then the lock will be put in a state where any future and pending acquire() fail. - Args: - blocking: whether to block while trying to acquire or to operate in - try-lock mode. this locking mode is not supported yet unless - you are already holding exclusively the lock. + @type timeout: float + @param timeout: maximum waiting time before giving up + @type priority: integer + @param priority: Priority for acquiring lock """ + if priority is None: + priority = _DEFAULT_PRIORITY + self.__lock.acquire() try: - assert not self.__is_sharer(), "cannot delete() a lock while sharing it" + assert not self.__is_sharer(), "Cannot delete() a lock while sharing it" + + self.__check_deleted() - if self.__deleted: - raise errors.LockError('deleted lock') + # The caller is allowed to hold the lock exclusively already. + acquired = self.__is_exclusive() - if not self.__is_exclusive(): - if not blocking: - # We don't have non-blocking mode for now - raise NotImplementedError - self.__exclusive_acquire() + if not acquired: + acquired = self.__acquire_unlocked(0, timeout, priority) - self.__deleted = True - self.__exc = None - # Wake up everybody, they will fail acquiring the lock and - # raise an exception instead. - self.__turn_exc.notifyAll() - self.__turn_shr.notifyAll() + assert self.__is_exclusive() and not self.__is_sharer(), \ + "Lock wasn't acquired in exclusive mode" + if acquired: + self.__deleted = True + self.__exc = None + + assert not (self.__exc or self.__shr), "Found owner during deletion" + + # Notify all acquires. They'll throw an error. + for (_, prioqueue) in self.__pending: + for cond in prioqueue: + cond.notifyAll() + + assert self.__deleted + + return acquired finally: self.__lock.release() + def _release_save(self): + shared = self.__is_sharer() + self.release() + return shared + + def _acquire_restore(self, shared): + self.acquire(shared=shared) + + +# Whenever we want to acquire a full LockSet we pass None as the value +# to acquire. Hide this behind this nicely named constant. +ALL_SET = None + + +class _AcquireTimeout(Exception): + """Internal exception to abort an acquire on a timeout. + + """ + class LockSet: """Implements a set of locks. @@ -308,24 +825,35 @@ class LockSet: All the locks needed in the same set must be acquired together, though. + @type name: string + @ivar name: the name of the lockset + """ - def __init__(self, members=None): + def __init__(self, members, name, monitor=None): """Constructs a new LockSet. - Args: - members: initial members of the set + @type members: list of strings + @param members: initial members of the set + @type monitor: L{LockMonitor} + @param monitor: Lock monitor with which to register member locks """ + assert members is not None, "members parameter is not a list" + self.name = name + + # Lock monitor + self.__monitor = monitor + # Used internally to guarantee coherency. - self.__lock = SharedLock() + self.__lock = SharedLock(name) # The lockdict indexes the relationship name -> lock # The order-of-locking is implied by the alphabetical order of names self.__lockdict = {} - if members is not None: - for name in members: - self.__lockdict[name] = SharedLock() + for mname in members: + self.__lockdict[mname] = SharedLock(self._GetLockName(mname), + monitor=monitor) # The owner dict contains the set of locks each thread owns. For # performance each thread can access its own key without a global lock on @@ -336,6 +864,12 @@ class LockSet: # will be trouble. self.__owners = {} + def _GetLockName(self, mname): + """Returns the name for a member lock. + + """ + return "%s/%s" % (self.name, mname) + def _is_owned(self): """Is the current thread a current level owner?""" return threading.currentThread() in self.__owners @@ -351,10 +885,12 @@ class LockSet: else: self.__owners[threading.currentThread()] = set([name]) - def _del_owned(self, name=None): """Note the current thread owns the given lock""" + assert not (name is None and self.__lock._is_owned()), \ + "Cannot hold internal lock when deleting owner status" + if name is not None: self.__owners[threading.currentThread()].remove(name) @@ -370,6 +906,14 @@ class LockSet: else: return set() + def _release_and_delete_owned(self): + """Release and delete all resources owned by the current thread""" + for lname in self._list_owned(): + lock = self.__lockdict[lname] + if lock._is_owned(): + lock.release() + self._del_owned(name=lname) + def __names(self): """Return the current set of names. @@ -385,120 +929,184 @@ class LockSet: Used only for debugging purposes. """ - self.__lock.acquire(shared=1) + # If we don't already own the set-level lock acquired + # we'll get it and note we need to release it later. + release_lock = False + if not self.__lock._is_owned(): + release_lock = True + self.__lock.acquire(shared=1) try: result = self.__names() finally: - self.__lock.release() + if release_lock: + self.__lock.release() return set(result) - def acquire(self, names, blocking=1, shared=0): + def acquire(self, names, timeout=None, shared=0, priority=None, + test_notify=None): """Acquire a set of resource locks. - Args: - names: the names of the locks which shall be acquired. - (special lock names, or instance/node names) - shared: whether to acquire in shared mode. By default an exclusive lock - will be acquired. - blocking: whether to block while trying to acquire or to operate in - try-lock mode. this locking mode is not supported yet. - - Returns: - True: when all the locks are successfully acquired - - Raises: - errors.LockError: when any lock we try to acquire has been deleted - before we succeed. In this case none of the locks requested will be - acquired. + @type names: list of strings (or string) + @param names: the names of the locks which shall be acquired + (special lock names, or instance/node names) + @type shared: integer (0/1) used as a boolean + @param shared: whether to acquire in shared mode; by default an + exclusive lock will be acquired + @type timeout: float or None + @param timeout: Maximum time to acquire all locks + @type priority: integer + @param priority: Priority for acquiring locks + @type test_notify: callable or None + @param test_notify: Special callback function for unittesting + + @return: Set of all locks successfully acquired or None in case of timeout + + @raise errors.LockError: when any lock we try to acquire has + been deleted before we succeed. In this case none of the + locks requested will be acquired. """ - if not blocking: - # We don't have non-blocking mode for now - raise NotImplementedError + assert timeout is None or timeout >= 0.0 # Check we don't already own locks at this level - assert not self._is_owned(), "Cannot acquire locks in the same set twice" + assert not self._is_owned(), ("Cannot acquire locks in the same set twice" + " (lockset %s)" % self.name) - if names is None: - # If no names are given acquire the whole set by not letting new names - # being added before we release, and getting the current list of names. - # Some of them may then be deleted later, but we'll cope with this. - # - # We'd like to acquire this lock in a shared way, as it's nice if - # everybody else can use the instances at the same time. If are acquiring - # them exclusively though they won't be able to do this anyway, though, - # so we'll get the list lock exclusively as well in order to be able to - # do add() on the set while owning it. - self.__lock.acquire(shared=shared) - try: - # note we own the set-lock - self._add_owned() - names = self.__names() - except: - # We shouldn't have problems adding the lock to the owners list, but - # if we did we'll try to release this lock and re-raise exception. - # Of course something is going to be really wrong, after this. - self.__lock.release() - raise + if priority is None: + priority = _DEFAULT_PRIORITY + + # We need to keep track of how long we spent waiting for a lock. The + # timeout passed to this function is over all lock acquires. + running_timeout = utils.RunningTimeout(timeout, False) try: - # Support passing in a single resource to acquire rather than many - if isinstance(names, basestring): - names = [names] + if names is not None: + # Support passing in a single resource to acquire rather than many + if isinstance(names, basestring): + names = [names] + + return self.__acquire_inner(names, False, shared, priority, + running_timeout.Remaining, test_notify) + else: - names.sort() + # If no names are given acquire the whole set by not letting new names + # being added before we release, and getting the current list of names. + # Some of them may then be deleted later, but we'll cope with this. + # + # We'd like to acquire this lock in a shared way, as it's nice if + # everybody else can use the instances at the same time. If we are + # acquiring them exclusively though they won't be able to do this + # anyway, though, so we'll get the list lock exclusively as well in + # order to be able to do add() on the set while owning it. + if not self.__lock.acquire(shared=shared, priority=priority, + timeout=running_timeout.Remaining()): + raise _AcquireTimeout() + try: + # note we own the set-lock + self._add_owned() + + return self.__acquire_inner(self.__names(), True, shared, priority, + running_timeout.Remaining, test_notify) + except: + # We shouldn't have problems adding the lock to the owners list, but + # if we did we'll try to release this lock and re-raise exception. + # Of course something is going to be really wrong, after this. + self.__lock.release() + self._del_owned() + raise + + except _AcquireTimeout: + return None + + def __acquire_inner(self, names, want_all, shared, priority, + timeout_fn, test_notify): + """Inner logic for acquiring a number of locks. + + @param names: Names of the locks to be acquired + @param want_all: Whether all locks in the set should be acquired + @param shared: Whether to acquire in shared mode + @param timeout_fn: Function returning remaining timeout + @param priority: Priority for acquiring locks + @param test_notify: Special callback function for unittesting + + """ + acquire_list = [] + + # First we look the locks up on __lockdict. We have no way of being sure + # they will still be there after, but this makes it a lot faster should + # just one of them be the already wrong. Using a sorted sequence to prevent + # deadlocks. + for lname in sorted(utils.UniqueSequence(names)): + try: + lock = self.__lockdict[lname] # raises KeyError if lock is not there + except KeyError: + if want_all: + # We are acquiring all the set, it doesn't matter if this particular + # element is not there anymore. + continue + + raise errors.LockError("Non-existing lock %s in set %s (it may have" + " been removed)" % (lname, self.name)) + + acquire_list.append((lname, lock)) + + # This will hold the locknames we effectively acquired. + acquired = set() + + try: + # Now acquire_list contains a sorted list of resources and locks we + # want. In order to get them we loop on this (private) list and + # acquire() them. We gave no real guarantee they will still exist till + # this is done but .acquire() itself is safe and will alert us if the + # lock gets deleted. + for (lname, lock) in acquire_list: + if __debug__ and callable(test_notify): + test_notify_fn = lambda: test_notify(lname) + else: + test_notify_fn = None + + timeout = timeout_fn() - acquire_list = [] - # First we look the locks up on __lockdict. We have no way of being sure - # they will still be there after, but this makes it a lot faster should - # just one of them be the already wrong - for lname in names: try: - lock = self.__lockdict[lname] # raises KeyError if lock is not there - acquire_list.append((lname, lock)) - except (KeyError): - if self.__lock._is_owned(): - # We are acquiring all the set, it doesn't matter if this particular - # element is not there anymore. + # raises LockError if the lock was deleted + acq_success = lock.acquire(shared=shared, timeout=timeout, + priority=priority, + test_notify=test_notify_fn) + except errors.LockError: + if want_all: + # We are acquiring all the set, it doesn't matter if this + # particular element is not there anymore. continue - else: - raise errors.LockError('non-existing lock in set (%s)' % lname) - - # This will hold the locknames we effectively acquired. - acquired = set() - # Now acquire_list contains a sorted list of resources and locks we want. - # In order to get them we loop on this (private) list and acquire() them. - # We gave no real guarantee they will still exist till this is done but - # .acquire() itself is safe and will alert us if the lock gets deleted. - for (lname, lock) in acquire_list: + + raise errors.LockError("Non-existing lock %s in set %s (it may" + " have been removed)" % (lname, self.name)) + + if not acq_success: + # Couldn't get lock or timeout occurred + if timeout is None: + # This shouldn't happen as SharedLock.acquire(timeout=None) is + # blocking. + raise errors.LockError("Failed to get lock %s (set %s)" % + (lname, self.name)) + + raise _AcquireTimeout() + try: - lock.acquire(shared=shared) # raises LockError if the lock is deleted # now the lock cannot be deleted, we have it! self._add_owned(name=lname) acquired.add(lname) - except (errors.LockError): - if self.__lock._is_owned(): - # We are acquiring all the set, it doesn't matter if this particular - # element is not there anymore. - continue - else: - name_fail = lname - for lname in self._list_owned(): - self.__lockdict[lname].release() - self._del_owned(name=lname) - raise errors.LockError('non-existing lock in set (%s)' % name_fail) + except: # We shouldn't have problems adding the lock to the owners list, but # if we did we'll try to release this lock and re-raise exception. - # Of course something is going to be really wrong, after this. + # Of course something is going to be really wrong after this. if lock._is_owned(): lock.release() - raise + raise except: - # If something went wrong and we had the set-lock let's release it... - if self.__lock._is_owned(): - self.__lock.release() + # Release all owned locks + self._release_and_delete_owned() raise return acquired @@ -509,12 +1117,13 @@ class LockSet: You must have acquired the locks, either in shared or in exclusive mode, before releasing them. - Args: - names: the names of the locks which shall be released. - (defaults to all the locks acquired at that level). + @type names: list of strings, or None + @param names: the names of the locks which shall be released + (defaults to all the locks acquired at that level). """ - assert self._is_owned(), "release() on lock set while not owner" + assert self._is_owned(), ("release() on lock set %s while not owner" % + self.name) # Support passing in a single resource to release rather than many if isinstance(names, basestring): @@ -525,8 +1134,8 @@ class LockSet: else: names = set(names) assert self._list_owned().issuperset(names), ( - "release() on unheld resources %s" % - names.difference(self._list_owned())) + "release() on unheld resources %s (set %s)" % + (names.difference(self._list_owned()), self.name)) # First of all let's release the "all elements" lock, if set. # After this 'add' can work again @@ -543,21 +1152,24 @@ class LockSet: def add(self, names, acquired=0, shared=0): """Add a new set of elements to the set - Args: - names: names of the new elements to add - acquired: pre-acquire the new resource? - shared: is the pre-acquisition shared? + @type names: list of strings + @param names: names of the new elements to add + @type acquired: integer (0/1) used as a boolean + @param acquired: pre-acquire the new resource? + @type shared: integer (0/1) used as a boolean + @param shared: is the pre-acquisition shared? """ - - assert not self.__lock._is_owned(shared=1), ( - "Cannot add new elements while sharing the set-lock") + # Check we don't already own locks at this level + assert not self._is_owned() or self.__lock._is_owned(shared=0), \ + ("Cannot add locks if the set %s is only partially owned, or shared" % + self.name) # Support passing in a single resource to add rather than many if isinstance(names, basestring): names = [names] - # If we don't already own the set-level lock acquire it in an exclusive way + # If we don't already own the set-level lock acquired in an exclusive way # we'll get it and note we need to release it later. release_lock = False if not self.__lock._is_owned(): @@ -570,12 +1182,15 @@ class LockSet: # This must be an explicit raise, not an assert, because assert is # turned off when using optimization, and this can happen because of # concurrency even if the user doesn't want it. - raise errors.LockError("duplicate add() (%s)" % invalid_names) + raise errors.LockError("duplicate add(%s) on lockset %s" % + (invalid_names, self.name)) for lockname in names: - lock = SharedLock() + lock = SharedLock(self._GetLockName(lockname), monitor=self.__monitor) if acquired: + # No need for priority or timeout here as this lock has just been + # created lock.acquire(shared=shared) # now the lock cannot be deleted, we have it! try: @@ -599,27 +1214,20 @@ class LockSet: return True - def remove(self, names, blocking=1): + def remove(self, names): """Remove elements from the lock set. You can either not hold anything in the lockset or already hold a superset of the elements you want to delete, exclusively. - Args: - names: names of the resource to remove. - blocking: whether to block while trying to acquire or to operate in - try-lock mode. this locking mode is not supported yet unless - you are already holding exclusively the locks. + @type names: list of strings + @param names: names of the resource to remove. - Returns: - A list of lock which we removed. The list is always equal to the names - list if we were holding all the locks exclusively. + @return: a list of locks which we removed; the list is always + equal to the names list if we were holding all the locks + exclusively """ - if not blocking and not self._is_owned(): - # We don't have non-blocking mode for now - raise NotImplementedError - # Support passing in a single resource to remove rather than many if isinstance(names, basestring): names = [names] @@ -628,7 +1236,8 @@ class LockSet: # to delete. The ownership must also be exclusive, but that will be checked # by the lock itself. assert not self._is_owned() or self._list_owned().issuperset(names), ( - "remove() on acquired lockset while not owning all elements") + "remove() on acquired lockset %s while not owning all elements" % + self.name) removed = [] @@ -643,7 +1252,8 @@ class LockSet: removed.append(lname) except (KeyError, errors.LockError): # This cannot happen if we were already holding it, verify: - assert not self._is_owned(), "remove failed while holding lockset" + assert not self._is_owned(), ("remove failed while holding lockset %s" + % self.name) else: # If no LockError was raised we are the ones who deleted the lock. # This means we can safely remove it from lockdict, as any further or @@ -670,15 +1280,24 @@ class LockSet: # If you need more than one node, or more than one instance, acquire them at # the same time. LEVEL_CLUSTER = 0 -LEVEL_NODE = 1 -LEVEL_INSTANCE = 2 +LEVEL_INSTANCE = 1 +LEVEL_NODEGROUP = 2 +LEVEL_NODE = 3 LEVELS = [LEVEL_CLUSTER, - LEVEL_NODE, - LEVEL_INSTANCE] + LEVEL_INSTANCE, + LEVEL_NODEGROUP, + LEVEL_NODE] # Lock levels which are modifiable -LEVELS_MOD = [LEVEL_NODE, LEVEL_INSTANCE] +LEVELS_MOD = [LEVEL_NODE, LEVEL_NODEGROUP, LEVEL_INSTANCE] + +LEVEL_NAMES = { + LEVEL_CLUSTER: "cluster", + LEVEL_INSTANCE: "instance", + LEVEL_NODEGROUP: "nodegroup", + LEVEL_NODE: "node", + } # Constant for the big ganeti lock BGL = 'BGL' @@ -687,7 +1306,7 @@ BGL = 'BGL' class GanetiLockManager: """The Ganeti Locking Library - The purpouse of this small library is to manage locking for ganeti clusters + The purpose of this small library is to manage locking for ganeti clusters in a central place, while at the same time doing dynamic checks against possible deadlocks. It will also make it easier to transition to a different lock type should we migrate away from python threads. @@ -695,34 +1314,56 @@ class GanetiLockManager: """ _instance = None - def __init__(self, nodes=None, instances=None): + def __init__(self, nodes, nodegroups, instances): """Constructs a new GanetiLockManager object. There should be only a GanetiLockManager object at any time, so this function raises an error if this is not the case. - Args: - nodes: list of node names - instances: list of instance names + @param nodes: list of node names + @param nodegroups: list of nodegroup uuids + @param instances: list of instance names """ - assert self.__class__._instance is None, "double GanetiLockManager instance" + assert self.__class__._instance is None, \ + "double GanetiLockManager instance" + self.__class__._instance = self + self._monitor = LockMonitor() + # The keyring contains all the locks, at their level and in the correct # locking order. self.__keyring = { - LEVEL_CLUSTER: LockSet([BGL]), - LEVEL_NODE: LockSet(nodes), - LEVEL_INSTANCE: LockSet(instances), - } + LEVEL_CLUSTER: LockSet([BGL], "BGL", monitor=self._monitor), + LEVEL_NODE: LockSet(nodes, "nodes", monitor=self._monitor), + LEVEL_NODEGROUP: LockSet(nodegroups, "nodegroups", monitor=self._monitor), + LEVEL_INSTANCE: LockSet(instances, "instances", + monitor=self._monitor), + } + + def QueryLocks(self, fields): + """Queries information from all locks. + + See L{LockMonitor.QueryLocks}. + + """ + return self._monitor.QueryLocks(fields) + + def OldStyleQueryLocks(self, fields): + """Queries information from all locks, returning old-style data. + + See L{LockMonitor.OldStyleQueryLocks}. + + """ + return self._monitor.OldStyleQueryLocks(fields) def _names(self, level): """List the lock names at the given level. - Used for debugging/testing purposes. - Args: - level: the level whose list of locks to get + This can be used for debugging/testing purposes. + + @param level: the level whose list of locks to get """ assert level in LEVELS, "Invalid locking level %s" % level @@ -734,6 +1375,8 @@ class GanetiLockManager: """ return self.__keyring[level]._is_owned() + is_owned = _is_owned + def _list_owned(self, level): """Get the set of owned locks at the given level @@ -746,9 +1389,9 @@ class GanetiLockManager: """ # This way of checking only works if LEVELS[i] = i, which we check for in # the test cases. - return utils.any((self._is_owned(l) for l in LEVELS[level + 1:])) + return compat.any((self._is_owned(l) for l in LEVELS[level + 1:])) - def _BGL_owned(self): + def _BGL_owned(self): # pylint: disable-msg=C0103 """Check if the current thread owns the BGL. Both an exclusive or a shared acquisition work. @@ -756,25 +1399,31 @@ class GanetiLockManager: """ return BGL in self.__keyring[LEVEL_CLUSTER]._list_owned() - def _contains_BGL(self, level, names): - """Check if acting on the given level and set of names will change the - status of the Big Ganeti Lock. + @staticmethod + def _contains_BGL(level, names): # pylint: disable-msg=C0103 + """Check if the level contains the BGL. + + Check if acting on the given level and set of names will change + the status of the Big Ganeti Lock. """ return level == LEVEL_CLUSTER and (names is None or BGL in names) - def acquire(self, level, names, blocking=1, shared=0): + def acquire(self, level, names, timeout=None, shared=0, priority=None): """Acquire a set of resource locks, at the same level. - Args: - level: the level at which the locks shall be acquired. - It must be a memmber of LEVELS. - names: the names of the locks which shall be acquired. - (special lock names, or instance/node names) - shared: whether to acquire in shared mode. By default an exclusive lock - will be acquired. - blocking: whether to block while trying to acquire or to operate in - try-lock mode. this locking mode is not supported yet. + @type level: member of locking.LEVELS + @param level: the level at which the locks shall be acquired + @type names: list of strings (or string) + @param names: the names of the locks which shall be acquired + (special lock names, or instance/node names) + @type shared: integer (0/1) used as a boolean + @param shared: whether to acquire in shared mode; by default + an exclusive lock will be acquired + @type timeout: float + @param timeout: Maximum time to acquire all locks + @type priority: integer + @param priority: Priority for acquiring lock """ assert level in LEVELS, "Invalid locking level %s" % level @@ -793,27 +1442,29 @@ class GanetiLockManager: " while owning some at a greater one") # Acquire the locks in the set. - return self.__keyring[level].acquire(names, shared=shared, - blocking=blocking) + return self.__keyring[level].acquire(names, shared=shared, timeout=timeout, + priority=priority) def release(self, level, names=None): """Release a set of resource locks, at the same level. - You must have acquired the locks, either in shared or in exclusive mode, - before releasing them. + You must have acquired the locks, either in shared or in exclusive + mode, before releasing them. - Args: - level: the level at which the locks shall be released. - It must be a memmber of LEVELS. - names: the names of the locks which shall be released. - (defaults to all the locks acquired at that level). + @type level: member of locking.LEVELS + @param level: the level at which the locks shall be released + @type names: list of strings, or None + @param names: the names of the locks which shall be released + (defaults to all the locks acquired at that level) """ assert level in LEVELS, "Invalid locking level %s" % level assert (not self._contains_BGL(level, names) or not self._upper_owned(LEVEL_CLUSTER)), ( "Cannot release the Big Ganeti Lock while holding something" - " at upper levels") + " at upper levels (%r)" % + (utils.CommaJoin(["%s=%r" % (LEVEL_NAMES[i], self._list_owned(i)) + for i in self.__keyring.keys()]), )) # Release will complain if we don't own the locks already return self.__keyring[level].release(names) @@ -821,12 +1472,15 @@ class GanetiLockManager: def add(self, level, names, acquired=0, shared=0): """Add locks at the specified level. - Args: - level: the level at which the locks shall be added. - It must be a memmber of LEVELS_MOD. - names: names of the locks to acquire - acquired: whether to acquire the newly added locks - shared: whether the acquisition will be shared + @type level: member of locking.LEVELS_MOD + @param level: the level at which the locks shall be added + @type names: list of strings + @param names: names of the locks to acquire + @type acquired: integer (0/1) used as a boolean + @param acquired: whether to acquire the newly added locks + @type shared: integer (0/1) used as a boolean + @param shared: whether the acquisition will be shared + """ assert level in LEVELS_MOD, "Invalid or immutable level %s" % level assert self._BGL_owned(), ("You must own the BGL before performing other" @@ -835,28 +1489,96 @@ class GanetiLockManager: " while owning some at a greater one") return self.__keyring[level].add(names, acquired=acquired, shared=shared) - def remove(self, level, names, blocking=1): + def remove(self, level, names): """Remove locks from the specified level. - You must either already own the locks you are trying to remove exclusively - or not own any lock at an upper level. + You must either already own the locks you are trying to remove + exclusively or not own any lock at an upper level. - Args: - level: the level at which the locks shall be removed. - It must be a memmber of LEVELS_MOD. - names: the names of the locks which shall be removed. - (special lock names, or instance/node names) - blocking: whether to block while trying to operate in try-lock mode. - this locking mode is not supported yet. + @type level: member of locking.LEVELS_MOD + @param level: the level at which the locks shall be removed + @type names: list of strings + @param names: the names of the locks which shall be removed + (special lock names, or instance/node names) """ assert level in LEVELS_MOD, "Invalid or immutable level %s" % level assert self._BGL_owned(), ("You must own the BGL before performing other" " operations") - # Check we either own the level or don't own anything from here up. - # LockSet.remove() will check the case in which we don't own all the needed - # resources, or we have a shared ownership. + # Check we either own the level or don't own anything from here + # up. LockSet.remove() will check the case in which we don't own + # all the needed resources, or we have a shared ownership. assert self._is_owned(level) or not self._upper_owned(level), ( "Cannot remove locks at a level while not owning it or" " owning some at a greater one") - return self.__keyring[level].remove(names, blocking=blocking) + return self.__keyring[level].remove(names) + + +class LockMonitor(object): + _LOCK_ATTR = "_lock" + + def __init__(self): + """Initializes this class. + + """ + self._lock = SharedLock("LockMonitor") + + # Tracked locks. Weak references are used to avoid issues with circular + # references and deletion. + self._locks = weakref.WeakKeyDictionary() + + @ssynchronized(_LOCK_ATTR) + def RegisterLock(self, lock): + """Registers a new lock. + + """ + logging.debug("Registering lock %s", lock.name) + assert lock not in self._locks, "Duplicate lock registration" + assert not compat.any(lock.name == i.name for i in self._locks.keys()), \ + "Found duplicate lock name" + self._locks[lock] = None + + @ssynchronized(_LOCK_ATTR) + def _GetLockInfo(self, requested): + """Get information from all locks while the monitor lock is held. + + """ + return [lock.GetInfo(requested) for lock in self._locks.keys()] + + def _Query(self, fields): + """Queries information from all locks. + + @type fields: list of strings + @param fields: List of fields to return + + """ + qobj = query.Query(query.LOCK_FIELDS, fields) + + # Get all data and sort by name + lockinfo = utils.NiceSort(self._GetLockInfo(qobj.RequestedData()), + key=operator.itemgetter(0)) + + return (qobj, query.LockQueryData(lockinfo)) + + def QueryLocks(self, fields): + """Queries information from all locks. + + @type fields: list of strings + @param fields: List of fields to return + + """ + (qobj, ctx) = self._Query(fields) + + # Prepare query response + return query.GetQueryResponse(qobj, ctx) + + def OldStyleQueryLocks(self, fields): + """Queries information from all locks, returning old-style data. + + @type fields: list of strings + @param fields: List of fields to return + + """ + (qobj, ctx) = self._Query(fields) + + return qobj.OldStyleQuery(ctx)