[Intel-gfx] [PATCH] drm/i915: Do not trigger the pageflip stall check too early

[Chris Wilson]

On Wed, Dec 17, 2014 at 08:19:00PM +0200, Ville Syrjälä wrote:
> On Wed, Dec 17, 2014 at 04:48:19PM +0000, Chris Wilson wrote:
> > On Wed, Dec 17, 2014 at 04:41:42PM +0000, Chris Wilson wrote:
> > > On gen2-4, we have a separate pageflip prepare/finish phase. The intent
> > > of the stall check is to detect when we have incurred a delay,
> > > potentially indefinite, after the pageflip is submitted and before the
> > > flip is processed by the hardware. However, our notion of
> > > INTEL_FLIP_PENDING/INTEL_FLIP_COMPLETE do not tally with how we set the
> > > values during prepare/finish and the current stall check erroneously
> > > assumes that when the pending value >= COMPLETE, the driver has seen the
> > > hardware completion flag. But what the driver actually means, is that it
> > > has seen the acknowlegement that the flip is queued and is now pending
> > > the completion event.
> > 
> > Bah, otoh, as we don't mark the flip as pending before completion on
> > gen5+, we know don't detect when the flip is wedged there after applying
> > this patch.
> > 
> > It is quite possible that the work->pending check is entirely bogus.
> > Back to the drawing board.
> 
> We just apply my fix. And then we can even get rid of the whole
> prepare/finish mess.

I concur. I was trying to see if there was any value in trusting IIR but
not ISR for gen2/3, but that's silly. So with the current flip interrupt
handling we really do not have the two phases anymore.

https://bugs.freedesktop.org/attachment.cgi?id=110912
Reviewed-by: Chris Wilson <chris at chris-wilson.co.uk>
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre