Re: [PATCH 3/4] drm/i915: Drop the excessive vblank waits from modeset codepaths

["Ville Syrjälä" <ville.syrjala@linux.intel.com>]

On Wed, Mar 12, 2014 at 09:25:42AM +0000, Chris Wilson wrote:
> On Wed, Mar 12, 2014 at 11:16:59AM +0200, Ville Syrjälä wrote:
> > On Wed, Mar 12, 2014 at 08:35:39AM +0000, Chris Wilson wrote:
> > > On Tue, Mar 11, 2014 at 07:37:35PM +0200, ville.syrjala@linux.intel.com wrote:
> > > > From: Ville Syrjälä <ville.syrjala@linux.intel.com>
> > > > 
> > > > Now that we've plugged the mmio vs. ring flip race, we shouldn't need
> > > > these vblank waits in the modeset codepaths anymore. So get rid of
> > > > them.
> > > 
> > > Hmm, could we not add an assert(DSPSURFLIVE ==
> > > intel_crtc->dspsurf)?
> > 
> > Where would you want the assert?
> 
> assert_plane_is_bound(new_fb) in crtc_enable() and
> assert_plane_is_bound(old_fb) in crt_disable(). i.e. add them to our set
> of plane/pipe checks through modeset.

We can't really add live base address checks to such places w/o adding
extra vblank waits. For example if we first do a set_base w/o waiting for
vblank and then disable the crtc. By the time we get to crtc_disable()
there's no guarantee that the mmio flip from set_base has completed, and
hence the live surface address may still have the old value.

For the atomic page flip stuff the live address can serve as a decent
debug tool to make sure the flip has really completed or not completed
when we expect. IIRC I had such a debug patch in my atomic branch.

-- 
Ville Syrjälä
Intel OTC