Question

I was reading through my STL implementation (standard-issue g++ 4.6.2) and came across this bit of race condition inside of condition_variable:

template<typename _Rep, typename _Period>
cv_status
wait_for(unique_lock<mutex>& __lock,
         const chrono::duration<_Rep, _Period>& __rtime)
{
    return wait_until(__lock, __clock_t::now() + __rtime);
}

Because __clock_t is an std::chrono::system_clock, we are tied to the whims of things like NTP (if the clock is moved back by a day after __clock_t::now() + __rtime, then we’ll wait for a day).

The C++ standard (30.5.1) appears to get it right:

26

Effects: as if

return wait_until(lock, chrono::steady_clock::now() + rel_time);

Boost’s condition_variable implementation has the same problem:

template<typename duration_type>
bool timed_wait(unique_lock<mutex>& m,duration_type const& wait_duration)
{
    return timed_wait(m,get_system_time()+wait_duration);
}

In fact, the underlying pthreads implementation seems to be the problem:

int pthread_cond_timedwait(pthread_cond_t *restrict cond,
   pthread_mutex_t *restrict mutex,
   const struct timespec *restrict abstime);

because abstime is specified as “system time,” not a monotonic clock.

So my question is: How would one implement something like std::condition_variable::wait_for correctly? Is there an existing implementation that gets this right? Or am I missing something?

Answer 1

The trick is to use a pthread_condattr_setclock to tell the pthread_condattr_t to use CLOCK_MONOTONIC. The C code for doing this is pretty simple:

#include <time.h>
#include <pthread.h>

#include <errno.h>
#include <stdio.h>

int main()
{
    // Set the clock to be CLOCK_MONOTONIC
    pthread_condattr_t attr;
    pthread_condattr_init(&attr);
    if (int err = pthread_condattr_setclock(&attr, CLOCK_MONOTONIC))
    {
        printf("Error setting clock: %d\n", err);
    }

    // Now we can initialize the pthreads objects with that condattr
    pthread_mutex_t mutex = PTHREAD_MUTEX_INITIALIZER;
    pthread_cond_t  cond;
    pthread_cond_init(&cond, &attr);

    // when getting the time, we must poll from CLOCK_MONOTONIC
    struct timespec timeout;
    struct timespec now;
    clock_gettime(CLOCK_MONOTONIC, &now);
    timeout.tv_sec = now.tv_sec + 5;
    timeout.tv_nsec = now.tv_nsec;

    // business as usual...
    pthread_mutex_lock(&mutex);
    int rc = pthread_cond_timedwait(&cond, &mutex, &timeout);
    if (rc == ETIMEDOUT)
        printf("Success!\n");
    else
        printf("Got return that wasn't timeout: %d\n", rc);
    pthread_mutex_unlock(&mutex);

    return 0;
}

I’m going to leave this open for a while because somebody might have an easier answer. The thing I’m not happy about here is that it means a wait_until is rather difficult to implement with a real-time clock (my best solution to that is to convert the provided Clock in the time_point into the steady_clock‘s time and go from there…it is still subject to time-change race conditions, but if you’re specifying a timeout in real time, you’re already making a terrible mistake).