There are many text on the subject of condition variables and their usage, so I’ll not bore you with a ton of ugly details. The reason they exist at all is to allow you to notify change in a predicate state. The following are critical in understanding proper use of condition variables and their mutex association:
-
pthread_cond_wait()simultaneously unlocks the mutex and begins waiting for the condition variable to be signalled. thus you must always have ownership of the mutex before invoking it. -
pthread_cond_wait()returns with the mutex locked, thus you must unlock the mutex to allow its use somewhere else when finished with it. Whether the return happened because the condition variable was signalled or not isn’t relevant. You still need to check your predicate regardless to account for potential spurious wakeups. -
The purpose of the mutex is not to protect the condition variable; it is to protect the predicate on which the condition variable is being used as a signaling mechanism. This is hands-down the most often misunderstood idiom of pthread condition variables and their mutexes. The condition variable doesn’t need mutual exclusion protection; the predicate data does. Think of the predicate as an outside-state which is being monitored by the users of the condition-variable/mutex pair.
For example, a trivial yet obviously wrong piece of code to wait for a boolean flag fSet:
bool fSet = false;
int WaitForTrue()
{
while (!fSet)
{
sleep(n);
}
}
I should be obvious the main problem is the predicate, fSet, is not protected at all. Many things can go wrong here. Ex: From the time you evaluate the while-conditon until the time you begin waiting (or spinning, or whatever) the value may have changed. If that change notification is somehow missed, you’re needlessly waiting.
We can change this a little so at least the predicate is protected somehow. Mutual exclusion in both modifying and evaluating the predicate is easily provided with (what else) a mutex.
pthread_mutex_t mtx = PTHREAD_MUTEX_INITIALIZER;
bool fSet = false;
int WaitForTrue()
{
pthread_mutex_lock(&mtx);
while (!fSet)
sleep(n);
pthread_mutex_unlock(&mtx);
}
Well, that seems simple enough.. Now we never evaluate the predicate without first getting exclusive access to it (by latching the mutex). But this is still a major problem. We latched the mutex, but we never release it until our loop is finished. If everyone else plays by the rules and waits for the mutex lock before evaluation or modification of fSet, they’re never be able to do so until we give up the mutex. The only “someone” that can do that in this case is us.
So what about adding still more layers to this. Will this work?
pthread_mutex_t mtx = PTHREAD_MUTEX_INITIALIZER;
bool fSet = false;
int WaitForTrue()
{
pthread_mutex_lock(&mtx);
while (!fSet)
{
pthread_mutex_unlock(&mtx);
// XXXXX
sleep(n);
// YYYYY
pthread_mutex_lock(&mtx);
}
pthread_mutex_unlock(&mtx);
}
Well, yes it will “work”, but still is not much better. The period between XXXXX and YYYYY we don’t own the mutex (which is ok, since we’re not checking or modifying fSet anyway). But anytime during that period some other thread can (a) obtain the mutex, (b) modify fSet, and (c) release the mutex, and we won’t know a thing about it until we finish our sleep(), once-again obtain the mutex lock, and loop around for another check.
There has to be a better way. Somehow there should be a way that we can release the mutex and begin waiting for some sort of signal that tells us a change in the predicate may have happened. Equally important, when we receive that signal and return to our code, we should already own the lock that grants us access to check the predicate data. This is exactly what a condition-variable is designed to provide.
The Condition Variable In Action
Enter the condition-variable + mutex pair. The mutex protects access to changing or checking the predicate, while the condition variable sets up a system of monitoring a change, and more importantly, doing so atomically (as far as you’re concerned, anyway) with the predicate mutual exclusion:
int WaitForPredicate()
{
// lock mutex (means:lock access to the predicate)
pthread_mutex_lock(&mtx);
// we can safely check this, since no one else should be
// changing it unless they have the mutex, which they don't
// because we just locked it.
while (!predicate)
{
// predicate not met, so begin waiting for notification
// it has been changed *and* release access to change it
// to anyone wanting to by unlatching the mutex, doing
// both (start waiting and unlatching) atomically
pthread_cond_wait(&cv,&mtx);
// upon arriving here, the above returns with the mutex
// latched (we own it). The predicate *may* be true, and
// we'll be looping around to see if it is, but we can
// safely do so because we own the mutex coming out of
// the cv-wait call.
}
// we still own the mutex here. further, we have assessed the
// predicate is true (thus how we broke the loop).
// take whatever action needed.
// You *must* release the mutex before we leave. Remember, we
// still own it even after the code above.
pthread_mutex_unlock(&mtx);
}
For some other thread to signal the loop above, there are several ways to do it, the two most popular below:
pthread_mutex_lock(&mtx);
TODO: change predicate state here as needed.
pthread_mutex_unlock(&mtx);
pthread_cond_signal(&cv);
Another way…
pthread_mutex_lock(&mtx);
TODO: change predicate state here as needed.
pthread_cond_signal(&cv);
pthread_mutex_unlock(&mtx);
Each has different intrinsic behavior and I invite you to do some homework on those differences and determine which is more appropriate for specific circumstances. The former provides better program flow at the expense of introducing potentially unwarranted wake-ups. The latter reduces those wake-ups but at the price of less context synergy. Either will work in our sample, and you can experiment with how each affects your waiting loops. Regardless, one thing paramount, and both methods fulfill this mandate:
Never change, nor check, the predicate condition unless the mutex is locked. Ever.
Simple Monitor Thread
This type of operation is common in a monitor thread that acts on a specific predicate condition, which (sans’ error checking) typically looks something like this:
void* monitor_proc(void *pv)
{
// acquire mutex ownership
// (which means we own change-control to the predicate)
pthread_mutex_lock(&mtx);
// heading into monitor loop, we own the predicate mutex
while (true)
{
// safe to check; we own the mutex
while (!predicate)
pthread_cond_wait(&cv, &mtx);
// TODO: the cv has been signalled. our predicate data should include
// data to signal a break-state to exit this loop and finish the proc,
// as well as data that we may check for other processing.
}
// we still own the mutex. remember to release it on exit
pthread_mutex_unlock(&mtx);
return pv;
}
A More Complex Monitor Thread
Modifying this basic form to account for a notification system that doesn’t require you to keep the mutex latched once you’ve picked up the notification becomes a little more involved, but not by very much. Below is a monitor proc that does not keep the mutex latched during regular processing once we’ve established we’ve been served (so to speak).
void* monitor_proc(void *pv)
{
// acquire mutex ownership
// (which means we own change-control to the predicate)
pthread_mutex_lock(&mtx);
// heading into monitor loop, we own the predicate mutex
while (true)
{
// check predicate
while (!predicate)
pthread_cond_wait(&cv, &mtx);
// some state that is part of the predicate to
// inform us we're finished
if (break-state)
break;
// TODO: perform latch-required work here.
// unlatch the mutex to do our predicate-independant work.
pthread_mutex_unlock(&mtx);
// TODO: perform no-latch-required work here.
// re-latch mutex prior to heading into wait
pthread_mutex_lock(&mtx);
}
// we still own the mutex. remember to release it on exit
pthread_mutex_unlock(&mtx);
return pv;
}
Where would someone use something like that ? Well, suppose your “predicate” is the “state” of a work queue as well as some flag to tell you to stop looping and exit. Upon receiving the notification that something is “different”, you check to see if you should continue executing your loop, and deciding you should continue, pop some data off the queue. Modifying the queue requires the mutex be latched (remember, its “state” is part of our predicate). Once we have popped our data, we have it locally and can process it independent of the queue state, so we release the mutex, do our thing, then require the mutex for the next go-around. There are many ways to code the above concept, including judicious use of pthread_cond_broadcast, etc. But the basic form is hopefully understandable.
This turned out to be considerably longer than I had hoped, but this is a major hurdle for people learning pthread-programming, and I feel it is worth the extra time/effort. I hope you got something out of it.