Re: [PATCH v2] perf bench: Use condition variables in numa.

[Ian Rogers]

On Wed, Oct 14, 2020 at 4:45 AM Jiri Olsa <jolsa@xxxxxxxxxx> wrote:
>
> On Mon, Oct 12, 2020 at 09:16:11AM -0700, Ian Rogers wrote:
>
> SNIP
>
> > @@ -483,6 +484,18 @@ static void init_global_mutex(pthread_mutex_t *mutex)
> >       pthread_mutex_init(mutex, &attr);
> >  }
> >
> > +/*
> > + * Return a process-shared (global) condition variable:
> > + */
> > +static void init_global_cond(pthread_cond_t *cond)
> > +{
> > +     pthread_condattr_t attr;
> > +
> > +     pthread_condattr_init(&attr);
> > +     pthread_condattr_setpshared(&attr, PTHREAD_PROCESS_SHARED);
> > +     pthread_cond_init(cond, &attr);
> > +}
> > +
> >  static int parse_cpu_list(const char *arg)
> >  {
> >       p0.cpu_list_str = strdup(arg);
> > @@ -1136,15 +1149,18 @@ static void *worker_thread(void *__tdata)
> >       if (g->p.serialize_startup) {
> >               pthread_mutex_lock(&g->startup_mutex);
> >               g->nr_tasks_started++;
> > +             /* The last thread wakes the main process. */
> > +             if (g->nr_tasks_started == g->p.nr_tasks)
> > +                     pthread_cond_signal(&g->startup_cond);
>
> should you remove the condition? it's not necessary
> and making this racy, no?
>
> just single pthread_cond_signal should be enough,
> because the wait code is checking the number of tasks

The pthread_mutex_lock avoids any race on g->nr_tasks_started and
g->p.nr_tasks is set up in init() along with all the global state. I
don't think there's any race on g->nr_tasks_started and doing a signal
for every thread starting will just cause unnecessary wake-ups for the
main thread. I think it is better to keep it. I added loops on all the
pthread_cond_waits so the code is robust against spurious wake ups.

Thanks,
Ian

> jirka
>