Why does pthread_cond_wait have spurious wakeups?

To quote the man page:

> When using condition variables there is always a Boolean predicate involving shared variables associated with each condition wait that is true if the thread should proceed. Spurious wakeups from the pthread_cond_timedwait() or pthread_cond_wait() functions may occur. Since the return from pthread_cond_timedwait() or pthread_cond_wait() does not imply anything about the value of this predicate, the predicate should be re-evaluated upon such return.

So, pthread_cond_wait can return even if you haven't signaled it. At first glance at least, that seems pretty atrocious. It would be like a function which randomly returned the wrong value or randomly returned before it actually reached a proper return statement. It seems like a major bug. But the fact that they chose to document this in the man page rather than fix it would seem to indicate that there is a legitimate reason why pthread_cond_wait ends up waking up spuriously. Presumably, there's something intrinsic about how it works that makes it so that that can't be helped. The question is what.

Why does pthread_cond_wait return spuriously? Why can't it guarantee that it's only going to wake up when it's been properly signaled? Can anyone explain the reason for its spurious behavior?

There are at least two things 'spurious wakeup' could mean:

• A thread blocked in pthread_cond_wait can return from the call even though no call to pthread_call_signal or pthread_cond_broadcast on the condition occurred.
• A thread blocked in pthread_cond_wait returns because of a call to pthread_cond_signal or pthread_cond_broadcast, however after reacquiring the mutex the underlying predicate is found to no longer be true.

But the latter case can occur even if the condition variable implementation does not allow the former case. Consider a producer consumer queue, and three threads.

• Thread 1 has just dequeued an element and released the mutex, and the queue is now empty. The thread is doing whatever it does with the element it acquired on some CPU.
• Thread 2 attempts to dequeue an element, but finds the queue to be empty when checked under the mutex, calls pthread_cond_wait, and blocks in the call awaiting signal/broadcast.
• Thread 3 obtains the mutex, inserts a new element into the queue, notifies the condition variable, and releases the lock.
• In response to the notification from thread 3, thread 2, which was waiting on the condition, is scheduled to run.
• However before thread 2 manages to get on the CPU and grab the queue lock, thread 1 completes its current task, and returns to the queue for more work. It obtains the queue lock, checks the predicate, and finds that there is work in the queue. It proceeds to dequeue the item that thread 3 inserted, releases the lock, and does whatever it does with the item that thread 3 enqueued.
• Thread 2 now gets on a CPU and obtains the lock, but when it checks the predicate, it finds that the queue is empty. Thread 1 'stole' the item, so the wakeup appears to be spurious. Thread 2 needs to wait on the condition again.

So since you already always need to check the predicate under a loop, it makes no difference if the underlying condition variables can have other sorts of spurious wakeups.

The following explanation is given by David R. Butenhof in "Programming with POSIX Threads" (p. 80):

> Spurious wakeups may sound strange, but on some multiprocessor systems, making condition wakeup completely predictable might substantially slow all condition variable operations.

In the following comp.programming.threads discussion, he expands on the thinking behind the design:

Patrick Doyle wrote:
> In article <9couq0$[email protected]>, Tom Payne  <[email protected]> wrote:
> >Kaz Kylheku <[email protected]> wrote:
> >: It is so because implementations can sometimes not avoid inserting
> >: these spurious wakeups; it might be costly to prevent them.

> >But why?  Why is this so difficult?  For example, are we talking about
> >situations where a wait times out just as a signal arrives?
> You know, I wonder if the designers of pthreads used logic like this:
> users of condition variables have to check the condition on exit anyway,
> so we will not be placing any additional burden on them if we allow
> spurious wakeups; and since it is conceivable that allowing spurious
> wakeups could make an implementation faster, it can only help if we
> allow them.
> They may not have had any particular implementation in mind.
You're actually not far off at all, except you didn't push it far enough.
The intent was to force correct/robust code by requiring predicate loops. This was
driven by the provably correct academic contingent among the "core threadies" in
the working group, though I don't think anyone really disagreed with the intent
once they understood what it meant.
We followed that intent with several levels of justification. The first was that
"religiously" using a loop protects the application against its own imperfect
coding practices. The second was that it wasn't difficult to abstractly imagine
machines and implementation code that could exploit this requirement to improve
the performance of average condition wait operations through optimizing the
synchronization mechanisms.
/------------------[ [email protected] ]------------------\
| Compaq Computer Corporation              POSIX Thread Architect |
|     My book: http://www.awl.com/cseng/titles/0-201-63392-2/     |
-----[ http://home.earthlink.net/~anneart/family/dave.html ]-----/

Section "Multiple Awakenings by Condition Signal" in pthread_cond_signal has an example implementation of pthread_cond_wait and pthread_cond_signal which involves spurious wakekups.

While I don't think it was considered at the time of design, here is an actual technical reason: In combination with thread cancellation, there are conditions under which taking the option to wake "spuriously" may be absolutely necessary, at least unless you're willing to impose very very strong constraints on what sort of implementation strategies are possible.

The key problem is that, if a thread acts on cancellation while blocked in pthread_cond_wait, the side effects must be as if it did not consume any signal on the condition variable. However, it's difficult (and highly constraining) to ensure that you have not already consumed a signal when you begin acting on cancellation, and at this stage it may be impossible to "re-post" the signal to the condition variable, since you may be in a situation where the caller of pthread_cond_signal is already justified to have destroyed the condvar and freed the memory in which it resided.

The allowance for spurious wake gives you an easy out. Instead of continuing to act on cancellation when it arrives while blocked on a condition variable, if you may have already consumed a signal (or if you want to be lazy, no matter what), you can declare a spurious wake to have happened instead, and return with success. This does not interfere at all with the operation of cancellation, because a correct caller will simply act on the pending cancellation the next time it loops and calls pthread_cond_wait again.