• Home
  • Line#
  • Scopes#
  • Navigate#
  • Raw
  • Download
1 // Copyright (c) 2011 The Chromium Authors. All rights reserved.
2 // Use of this source code is governed by a BSD-style license that can be
3 // found in the LICENSE file.
4 
5 #include "base/synchronization/condition_variable.h"
6 
7 #include <stack>
8 
9 #include "base/logging.h"
10 #include "base/synchronization/lock.h"
11 #include "base/time.h"
12 
13 namespace base {
14 
ConditionVariable(Lock * user_lock)15 ConditionVariable::ConditionVariable(Lock* user_lock)
16     : user_lock_(*user_lock),
17       run_state_(RUNNING),
18       allocation_counter_(0),
19       recycling_list_size_(0) {
20   DCHECK(user_lock);
21 }
22 
~ConditionVariable()23 ConditionVariable::~ConditionVariable() {
24   AutoLock auto_lock(internal_lock_);
25   run_state_ = SHUTDOWN;  // Prevent any more waiting.
26 
27   DCHECK_EQ(recycling_list_size_, allocation_counter_);
28   if (recycling_list_size_ != allocation_counter_) {  // Rare shutdown problem.
29     // There are threads of execution still in this->TimedWait() and yet the
30     // caller has instigated the destruction of this instance :-/.
31     // A common reason for such "overly hasty" destruction is that the caller
32     // was not willing to wait for all the threads to terminate.  Such hasty
33     // actions are a violation of our usage contract, but we'll give the
34     // waiting thread(s) one last chance to exit gracefully (prior to our
35     // destruction).
36     // Note: waiting_list_ *might* be empty, but recycling is still pending.
37     AutoUnlock auto_unlock(internal_lock_);
38     Broadcast();  // Make sure all waiting threads have been signaled.
39     Sleep(10);  // Give threads a chance to grab internal_lock_.
40     // All contained threads should be blocked on user_lock_ by now :-).
41   }  // Reacquire internal_lock_.
42 
43   DCHECK_EQ(recycling_list_size_, allocation_counter_);
44 }
45 
Wait()46 void ConditionVariable::Wait() {
47   // Default to "wait forever" timing, which means have to get a Signal()
48   // or Broadcast() to come out of this wait state.
49   TimedWait(TimeDelta::FromMilliseconds(INFINITE));
50 }
51 
TimedWait(const TimeDelta & max_time)52 void ConditionVariable::TimedWait(const TimeDelta& max_time) {
53   Event* waiting_event;
54   HANDLE handle;
55   {
56     AutoLock auto_lock(internal_lock_);
57     if (RUNNING != run_state_) return;  // Destruction in progress.
58     waiting_event = GetEventForWaiting();
59     handle = waiting_event->handle();
60     DCHECK(handle);
61   }  // Release internal_lock.
62 
63   {
64     AutoUnlock unlock(user_lock_);  // Release caller's lock
65     WaitForSingleObject(handle, static_cast<DWORD>(max_time.InMilliseconds()));
66     // Minimize spurious signal creation window by recycling asap.
67     AutoLock auto_lock(internal_lock_);
68     RecycleEvent(waiting_event);
69     // Release internal_lock_
70   }  // Reacquire callers lock to depth at entry.
71 }
72 
73 // Broadcast() is guaranteed to signal all threads that were waiting (i.e., had
74 // a cv_event internally allocated for them) before Broadcast() was called.
Broadcast()75 void ConditionVariable::Broadcast() {
76   std::stack<HANDLE> handles;  // See FAQ-question-10.
77   {
78     AutoLock auto_lock(internal_lock_);
79     if (waiting_list_.IsEmpty())
80       return;
81     while (!waiting_list_.IsEmpty())
82       // This is not a leak from waiting_list_.  See FAQ-question 12.
83       handles.push(waiting_list_.PopBack()->handle());
84   }  // Release internal_lock_.
85   while (!handles.empty()) {
86     SetEvent(handles.top());
87     handles.pop();
88   }
89 }
90 
91 // Signal() will select one of the waiting threads, and signal it (signal its
92 // cv_event).  For better performance we signal the thread that went to sleep
93 // most recently (LIFO).  If we want fairness, then we wake the thread that has
94 // been sleeping the longest (FIFO).
Signal()95 void ConditionVariable::Signal() {
96   HANDLE handle;
97   {
98     AutoLock auto_lock(internal_lock_);
99     if (waiting_list_.IsEmpty())
100       return;  // No one to signal.
101     // Only performance option should be used.
102     // This is not a leak from waiting_list.  See FAQ-question 12.
103      handle = waiting_list_.PopBack()->handle();  // LIFO.
104   }  // Release internal_lock_.
105   SetEvent(handle);
106 }
107 
108 // GetEventForWaiting() provides a unique cv_event for any caller that needs to
109 // wait.  This means that (worst case) we may over time create as many cv_event
110 // objects as there are threads simultaneously using this instance's Wait()
111 // functionality.
GetEventForWaiting()112 ConditionVariable::Event* ConditionVariable::GetEventForWaiting() {
113   // We hold internal_lock, courtesy of Wait().
114   Event* cv_event;
115   if (0 == recycling_list_size_) {
116     DCHECK(recycling_list_.IsEmpty());
117     cv_event = new Event();
118     cv_event->InitListElement();
119     allocation_counter_++;
120     CHECK(cv_event->handle());
121   } else {
122     cv_event = recycling_list_.PopFront();
123     recycling_list_size_--;
124   }
125   waiting_list_.PushBack(cv_event);
126   return cv_event;
127 }
128 
129 // RecycleEvent() takes a cv_event that was previously used for Wait()ing, and
130 // recycles it for use in future Wait() calls for this or other threads.
131 // Note that there is a tiny chance that the cv_event is still signaled when we
132 // obtain it, and that can cause spurious signals (if/when we re-use the
133 // cv_event), but such is quite rare (see FAQ-question-5).
RecycleEvent(Event * used_event)134 void ConditionVariable::RecycleEvent(Event* used_event) {
135   // We hold internal_lock, courtesy of Wait().
136   // If the cv_event timed out, then it is necessary to remove it from
137   // waiting_list_.  If it was selected by Broadcast() or Signal(), then it is
138   // already gone.
139   used_event->Extract();  // Possibly redundant
140   recycling_list_.PushBack(used_event);
141   recycling_list_size_++;
142 }
143 //------------------------------------------------------------------------------
144 // The next section provides the implementation for the private Event class.
145 //------------------------------------------------------------------------------
146 
147 // Event provides a doubly-linked-list of events for use exclusively by the
148 // ConditionVariable class.
149 
150 // This custom container was crafted because no simple combination of STL
151 // classes appeared to support the functionality required.  The specific
152 // unusual requirement for a linked-list-class is support for the Extract()
153 // method, which can remove an element from a list, potentially for insertion
154 // into a second list.  Most critically, the Extract() method is idempotent,
155 // turning the indicated element into an extracted singleton whether it was
156 // contained in a list or not.  This functionality allows one (or more) of
157 // threads to do the extraction.  The iterator that identifies this extractable
158 // element (in this case, a pointer to the list element) can be used after
159 // arbitrary manipulation of the (possibly) enclosing list container.  In
160 // general, STL containers do not provide iterators that can be used across
161 // modifications (insertions/extractions) of the enclosing containers, and
162 // certainly don't provide iterators that can be used if the identified
163 // element is *deleted* (removed) from the container.
164 
165 // It is possible to use multiple redundant containers, such as an STL list,
166 // and an STL map, to achieve similar container semantics.  This container has
167 // only O(1) methods, while the corresponding (multiple) STL container approach
168 // would have more complex O(log(N)) methods (yeah... N isn't that large).
169 // Multiple containers also makes correctness more difficult to assert, as
170 // data is redundantly stored and maintained, which is generally evil.
171 
Event()172 ConditionVariable::Event::Event() : handle_(0) {
173   next_ = prev_ = this;  // Self referencing circular.
174 }
175 
~Event()176 ConditionVariable::Event::~Event() {
177   if (0 == handle_) {
178     // This is the list holder
179     while (!IsEmpty()) {
180       Event* cv_event = PopFront();
181       DCHECK(cv_event->ValidateAsItem());
182       delete cv_event;
183     }
184   }
185   DCHECK(IsSingleton());
186   if (0 != handle_) {
187     int ret_val = CloseHandle(handle_);
188     DCHECK(ret_val);
189   }
190 }
191 
192 // Change a container instance permanently into an element of a list.
InitListElement()193 void ConditionVariable::Event::InitListElement() {
194   DCHECK(!handle_);
195   handle_ = CreateEvent(NULL, false, false, NULL);
196   CHECK(handle_);
197 }
198 
199 // Methods for use on lists.
IsEmpty() const200 bool ConditionVariable::Event::IsEmpty() const {
201   DCHECK(ValidateAsList());
202   return IsSingleton();
203 }
204 
PushBack(Event * other)205 void ConditionVariable::Event::PushBack(Event* other) {
206   DCHECK(ValidateAsList());
207   DCHECK(other->ValidateAsItem());
208   DCHECK(other->IsSingleton());
209   // Prepare other for insertion.
210   other->prev_ = prev_;
211   other->next_ = this;
212   // Cut into list.
213   prev_->next_ = other;
214   prev_ = other;
215   DCHECK(ValidateAsDistinct(other));
216 }
217 
PopFront()218 ConditionVariable::Event* ConditionVariable::Event::PopFront() {
219   DCHECK(ValidateAsList());
220   DCHECK(!IsSingleton());
221   return next_->Extract();
222 }
223 
PopBack()224 ConditionVariable::Event* ConditionVariable::Event::PopBack() {
225   DCHECK(ValidateAsList());
226   DCHECK(!IsSingleton());
227   return prev_->Extract();
228 }
229 
230 // Methods for use on list elements.
231 // Accessor method.
handle() const232 HANDLE ConditionVariable::Event::handle() const {
233   DCHECK(ValidateAsItem());
234   return handle_;
235 }
236 
237 // Pull an element from a list (if it's in one).
Extract()238 ConditionVariable::Event* ConditionVariable::Event::Extract() {
239   DCHECK(ValidateAsItem());
240   if (!IsSingleton()) {
241     // Stitch neighbors together.
242     next_->prev_ = prev_;
243     prev_->next_ = next_;
244     // Make extractee into a singleton.
245     prev_ = next_ = this;
246   }
247   DCHECK(IsSingleton());
248   return this;
249 }
250 
251 // Method for use on a list element or on a list.
IsSingleton() const252 bool ConditionVariable::Event::IsSingleton() const {
253   DCHECK(ValidateLinks());
254   return next_ == this;
255 }
256 
257 // Provide pre/post conditions to validate correct manipulations.
ValidateAsDistinct(Event * other) const258 bool ConditionVariable::Event::ValidateAsDistinct(Event* other) const {
259   return ValidateLinks() && other->ValidateLinks() && (this != other);
260 }
261 
ValidateAsItem() const262 bool ConditionVariable::Event::ValidateAsItem() const {
263   return (0 != handle_) && ValidateLinks();
264 }
265 
ValidateAsList() const266 bool ConditionVariable::Event::ValidateAsList() const {
267   return (0 == handle_) && ValidateLinks();
268 }
269 
ValidateLinks() const270 bool ConditionVariable::Event::ValidateLinks() const {
271   // Make sure both of our neighbors have links that point back to us.
272   // We don't do the O(n) check and traverse the whole loop, and instead only
273   // do a local check to (and returning from) our immediate neighbors.
274   return (next_->prev_ == this) && (prev_->next_ == this);
275 }
276 
277 
278 /*
279 FAQ On subtle implementation details:
280 
281 1) What makes this problem subtle?  Please take a look at "Strategies
282 for Implementing POSIX Condition Variables on Win32" by Douglas
283 C. Schmidt and Irfan Pyarali.
284 http://www.cs.wustl.edu/~schmidt/win32-cv-1.html It includes
285 discussions of numerous flawed strategies for implementing this
286 functionality.  I'm not convinced that even the final proposed
287 implementation has semantics that are as nice as this implementation
288 (especially with regard to Broadcast() and the impact on threads that
289 try to Wait() after a Broadcast() has been called, but before all the
290 original waiting threads have been signaled).
291 
292 2) Why can't you use a single wait_event for all threads that call
293 Wait()?  See FAQ-question-1, or consider the following: If a single
294 event were used, then numerous threads calling Wait() could release
295 their cs locks, and be preempted just before calling
296 WaitForSingleObject().  If a call to Broadcast() was then presented on
297 a second thread, it would be impossible to actually signal all
298 waiting(?) threads.  Some number of SetEvent() calls *could* be made,
299 but there could be no guarantee that those led to to more than one
300 signaled thread (SetEvent()'s may be discarded after the first!), and
301 there could be no guarantee that the SetEvent() calls didn't just
302 awaken "other" threads that hadn't even started waiting yet (oops).
303 Without any limit on the number of requisite SetEvent() calls, the
304 system would be forced to do many such calls, allowing many new waits
305 to receive spurious signals.
306 
307 3) How does this implementation cause spurious signal events?  The
308 cause in this implementation involves a race between a signal via
309 time-out and a signal via Signal() or Broadcast().  The series of
310 actions leading to this are:
311 
312 a) Timer fires, and a waiting thread exits the line of code:
313 
314     WaitForSingleObject(waiting_event, max_time.InMilliseconds());
315 
316 b) That thread (in (a)) is randomly pre-empted after the above line,
317 leaving the waiting_event reset (unsignaled) and still in the
318 waiting_list_.
319 
320 c) A call to Signal() (or Broadcast()) on a second thread proceeds, and
321 selects the waiting cv_event (identified in step (b)) as the event to revive
322 via a call to SetEvent().
323 
324 d) The Signal() method (step c) calls SetEvent() on waiting_event (step b).
325 
326 e) The waiting cv_event (step b) is now signaled, but no thread is
327 waiting on it.
328 
329 f) When that waiting_event (step b) is reused, it will immediately
330 be signaled (spuriously).
331 
332 
333 4) Why do you recycle events, and cause spurious signals?  First off,
334 the spurious events are very rare.  They can only (I think) appear
335 when the race described in FAQ-question-3 takes place.  This should be
336 very rare.  Most(?)  uses will involve only timer expiration, or only
337 Signal/Broadcast() actions.  When both are used, it will be rare that
338 the race will appear, and it would require MANY Wait() and signaling
339 activities.  If this implementation did not recycle events, then it
340 would have to create and destroy events for every call to Wait().
341 That allocation/deallocation and associated construction/destruction
342 would be costly (per wait), and would only be a rare benefit (when the
343 race was "lost" and a spurious signal took place). That would be bad
344 (IMO) optimization trade-off.  Finally, such spurious events are
345 allowed by the specification of condition variables (such as
346 implemented in Vista), and hence it is better if any user accommodates
347 such spurious events (see usage note in condition_variable.h).
348 
349 5) Why don't you reset events when you are about to recycle them, or
350 about to reuse them, so that the spurious signals don't take place?
351 The thread described in FAQ-question-3 step c may be pre-empted for an
352 arbitrary length of time before proceeding to step d.  As a result,
353 the wait_event may actually be re-used *before* step (e) is reached.
354 As a result, calling reset would not help significantly.
355 
356 6) How is it that the callers lock is released atomically with the
357 entry into a wait state?  We commit to the wait activity when we
358 allocate the wait_event for use in a given call to Wait().  This
359 allocation takes place before the caller's lock is released (and
360 actually before our internal_lock_ is released).  That allocation is
361 the defining moment when "the wait state has been entered," as that
362 thread *can* now be signaled by a call to Broadcast() or Signal().
363 Hence we actually "commit to wait" before releasing the lock, making
364 the pair effectively atomic.
365 
366 8) Why do you need to lock your data structures during waiting, as the
367 caller is already in possession of a lock?  We need to Acquire() and
368 Release() our internal lock during Signal() and Broadcast().  If we tried
369 to use a callers lock for this purpose, we might conflict with their
370 external use of the lock.  For example, the caller may use to consistently
371 hold a lock on one thread while calling Signal() on another, and that would
372 block Signal().
373 
374 9) Couldn't a more efficient implementation be provided if you
375 preclude using more than one external lock in conjunction with a
376 single ConditionVariable instance?  Yes, at least it could be viewed
377 as a simpler API (since you don't have to reiterate the lock argument
378 in each Wait() call).  One of the constructors now takes a specific
379 lock as an argument, and a there are corresponding Wait() calls that
380 don't specify a lock now.  It turns that the resulting implmentation
381 can't be made more efficient, as the internal lock needs to be used by
382 Signal() and Broadcast(), to access internal data structures.  As a
383 result, I was not able to utilize the user supplied lock (which is
384 being used by the user elsewhere presumably) to protect the private
385 member access.
386 
387 9) Since you have a second lock, how can be be sure that there is no
388 possible deadlock scenario?  Our internal_lock_ is always the last
389 lock acquired, and the first one released, and hence a deadlock (due
390 to critical section problems) is impossible as a consequence of our
391 lock.
392 
393 10) When doing a Broadcast(), why did you copy all the events into
394 an STL queue, rather than making a linked-loop, and iterating over it?
395 The iterating during Broadcast() is done so outside the protection
396 of the internal lock. As a result, other threads, such as the thread
397 wherein a related event is waiting, could asynchronously manipulate
398 the links around a cv_event.  As a result, the link structure cannot
399 be used outside a lock.  Broadcast() could iterate over waiting
400 events by cycling in-and-out of the protection of the internal_lock,
401 but that appears more expensive than copying the list into an STL
402 stack.
403 
404 11) Why did the lock.h file need to be modified so much for this
405 change?  Central to a Condition Variable is the atomic release of a
406 lock during a Wait().  This places Wait() functionality exactly
407 mid-way between the two classes, Lock and Condition Variable.  Given
408 that there can be nested Acquire()'s of locks, and Wait() had to
409 Release() completely a held lock, it was necessary to augment the Lock
410 class with a recursion counter. Even more subtle is the fact that the
411 recursion counter (in a Lock) must be protected, as many threads can
412 access it asynchronously.  As a positive fallout of this, there are
413 now some DCHECKS to be sure no one Release()s a Lock more than they
414 Acquire()ed it, and there is ifdef'ed functionality that can detect
415 nested locks (legal under windows, but not under Posix).
416 
417 12) Why is it that the cv_events removed from list in Broadcast() and Signal()
418 are not leaked?  How are they recovered??  The cv_events that appear to leak are
419 taken from the waiting_list_.  For each element in that list, there is currently
420 a thread in or around the WaitForSingleObject() call of Wait(), and those
421 threads have references to these otherwise leaked events. They are passed as
422 arguments to be recycled just aftre returning from WaitForSingleObject().
423 
424 13) Why did you use a custom container class (the linked list), when STL has
425 perfectly good containers, such as an STL list?  The STL list, as with any
426 container, does not guarantee the utility of an iterator across manipulation
427 (such as insertions and deletions) of the underlying container.  The custom
428 double-linked-list container provided that assurance.  I don't believe any
429 combination of STL containers provided the services that were needed at the same
430 O(1) efficiency as the custom linked list.  The unusual requirement
431 for the container class is that a reference to an item within a container (an
432 iterator) needed to be maintained across an arbitrary manipulation of the
433 container.  This requirement exposes itself in the Wait() method, where a
434 waiting_event must be selected prior to the WaitForSingleObject(), and then it
435 must be used as part of recycling to remove the related instance from the
436 waiting_list.  A hash table (STL map) could be used, but I was embarrased to
437 use a complex and relatively low efficiency container when a doubly linked list
438 provided O(1) performance in all required operations.  Since other operations
439 to provide performance-and/or-fairness required queue (FIFO) and list (LIFO)
440 containers, I would also have needed to use an STL list/queue as well as an STL
441 map.  In the end I decided it would be "fun" to just do it right, and I
442 put so many assertions (DCHECKs) into the container class that it is trivial to
443 code review and validate its correctness.
444 
445 */
446 
447 }  // namespace base
448