[Intel-gfx] [PATCH 49/89] drm/i915/skl: Program the DDB allocation

[Damien Lespiau]

On Fri, Sep 19, 2014 at 01:03:15PM +0300, Ville Syrjälä wrote:
> > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > index 0ddcbad..756ff16 100644
> > --- a/drivers/gpu/drm/i915/intel_pm.c
> > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > @@ -3393,6 +3393,15 @@ static void skl_write_wm_values(struct drm_i915_private *dev_priv,
> >  				I915_WRITE(PLANE_WM_TRANS(pipe, i),
> >  						new->plane_trans[pipe][i]);
> >  			I915_WRITE(CUR_WM_TRANS(pipe), new->cursor_trans[pipe]);
> > +
> > +			for (i = 0; i < intel_num_planes(crtc); i++)
> > +				I915_WRITE(PLANE_BUF_CFG(pipe, i),
> > +					   new->ddb.plane[pipe][i].end << 16 |
> 
> The spec doesn't say if end is inclusive or exclusive. Someone actually
> confirmed that from someone or tested it?

I remember testing it and having the end of a plane match the start of
another produces artefacts.

> > +					   new->ddb.plane[pipe][i].start);
> > +
> > +			I915_WRITE(CUR_BUF_CFG(pipe),
> > +				   new->ddb.cursor[pipe].end << 16 |
> > +				   new->ddb.cursor[pipe].start);
> >  		}
> >  	}
> 
> And here we land in deep doodoo wrt. changing the DDB allocation
> for active pipes. Someone needs to implement the correct logic to
> sequence the DDB allocation changes so that the allocations don't
> overlap at any point in time.
> 
> So this needs a big FIXME at the very least.

As seen on IRC, the DDB regs are double buffered and the "flush" patch
is supposed to take care of that.

-- 
Damien