[igt-dev] [Intel-gfx] [PATH i-g-t 04/13] gem_wsim: Check sleep times

[Ville Syrjälä]

On Fri, Sep 07, 2018 at 09:45:14AM +0100, Chris Wilson wrote:
> Quoting Tvrtko Ursulin (2018-09-07 09:37:00)
> > 
> > On 05/09/2018 15:09, Ville Syrjälä wrote:
> > > On Wed, Sep 05, 2018 at 02:49:30PM +0100, Tvrtko Ursulin wrote:
> > >> From: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > >>
> > >> Notice in more places if we are running behind.
> > >>
> > >> Signed-off-by: Tvrtko Ursulin <tvrtko.ursulin at intel.com>
> > >> ---
> > >>   benchmarks/gem_wsim.c | 52 ++++++++++++++++++++++++++++++++++++++-----
> > >>   1 file changed, 46 insertions(+), 6 deletions(-)
> > >>
> > >> diff --git a/benchmarks/gem_wsim.c b/benchmarks/gem_wsim.c
> > >> index 25af4d678ba4..b05e9760f419 100644
> > >> --- a/benchmarks/gem_wsim.c
> > >> +++ b/benchmarks/gem_wsim.c
> > >> @@ -1718,6 +1718,21 @@ static bool sync_deps(struct workload *wrk, struct w_step *w)
> > >>      return synced;
> > >>   }
> > >>   
> > >> +static unsigned int measured_usleep(unsigned int usec)
> > >> +{
> > >> +    struct timespec ts = { };
> > >> +    unsigned int slept;
> > >> +
> > >> +    slept = igt_nsec_elapsed(&ts);
> > >> +    igt_assert(slept == 0);
> > >> +    do {
> > >> +            usleep(usec - slept);
> > >> +            slept = igt_nsec_elapsed(&ts) / 1000;
> > >> +    } while (slept < usec);
> > > 
> > > clock_nanosleep(ABS)?
> > 
> > Hm I think I see what you mean. Rather than a relative sleep trying to 
> > hit the loop period, ask from the kernel (or glibc, I don't know who 
> > implements it) to sleep until an absolute target. This totally makes 
> > sense and would simplify the code from one angle, I am just not sure if 
> > absolute sleep can be relied upon any better to not oversleep. Well, 
> > actually for scheduling delays not to affect the caller. However maybe 
> > it doesn't matter since AFAIR my main problem were dropped period due 
> > GPU activity (the first pair of warning messages in the patch), and 
> > again AFAIR, it was quite hard to hit the second ones.
> 
> Right, it removes the loop but we still want to keep the measurement.

I guess we still want a loop if we're worried about signals? Not sure
why else we'd need a loop anyway. But the loop could be just something
like 'while (clock_nanosleep()) ;'

-- 
Ville Syrjälä
Intel