Re: [PATCH] drm/i915: use hrtimer in wait for vblank

[Daniel Vetter <daniel@ffwll.ch>]

On Mon, Mar 24, 2014 at 09:48:09AM +0000, Chris Wilson wrote:
> On Mon, Mar 24, 2014 at 09:34:35AM +0000, Murthy, Arun R wrote:
> > >> diff --git a/drivers/gpu/drm/i915/intel_drv.h 
> > >> b/drivers/gpu/drm/i915/intel_drv.h
> > >> index 44067bc..079280a 100644
> > >> --- a/drivers/gpu/drm/i915/intel_drv.h
> > >> +++ b/drivers/gpu/drm/i915/intel_drv.h
> > >> @@ -52,7 +52,7 @@
> > >>  			break;						\
> > >>  		}							\
> > >>  		if (W && drm_can_sleep())  {				\
> > >> -			msleep(W);					\
> > >> +			usleep_range(W * 1000, W * 2 * 1000);		\
> > >>  		} else {						\
> > >>  			cpu_relax();					\
> > >>  		}							\
> > >
> > > Ok. But W is still just a random value we picked for being the mininum 
> > > legal value for msleep(). So just usleep_range(500, 2000) or somesuch 
> > > will be fine. We can rename W to CAN_SLEEP it that helps.
> > 
> > We do use _wait_for directly from intel_dp.c with W == 10 to not retry so many times on what's expected to be a long wait.
> 
> Hmm, missed that. We can fallback to the fallback plan of switching
> based on W between msleep and usleep_range. Using -1, would be best.
> 
>  if (W && drm_can_sleep()) {
>   if (__builtin_constant_p(W) && W < 0)
>      usleep_range(500, 2000);
>   else
>      msleep(W);
>  }
>  
> > Its not expected to be too long, we tend to get vblank every 16ms.  With using usleep_range
> > with min and max as 1-2ms, we have observed that this function consuming 4ms to 17ms.
> 
> vs msleep()? That would be nice information to capture in the changelog.
> We expect the average wait here to be 8ms, ofc.
> 
> > Having 10ms sleep timer, we might tend to be blocking for 6ms in some cases.(from the above data)
> > Moreover 10ms usleep doesn't guarantee that we get a chance to execute after 10ms, it might
> > get increased due to scheduling. So ideal solution would be to use useep_range which uses hrtimers.
> 
> Ideal for what? We only use the sleeping path where we do not care too
> great about the latency, otherwise we use the busy-wait path. Improving
> the latency of the sleep without adversely affecting overall system
> performance/power is ideal.

We've used it in a few places where we'd care a bit more about latency
(like edid reads) but on most platforms those switched to interrupts now.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
+41 (0) 79 365 57 48 - http://blog.ffwll.ch