[Intel-gfx] [PATCH 08/24] drm/i915: Do not add planes from intel_atomic_setup_scalers.

[Konduru, Chandra]

> >>>>>> @@ -396,6 +377,24 @@ int intel_atomic_setup_scalers(struct
> drm_device
> >>>>> *dev,
> >>>>>>  		}
> >>>>>>  	}
> >>>>>>
> >>>>>> +	/* plane not part of mask must leave hq mode? */
> >>>>>> +	if (num_scalers_need > 1 && scaler_state->scalers[0].in_use &&
> >>>>>> +	    scaler_state->scalers[0].mode == PS_SCALER_MODE_HQ) {
> >>>>>> +		scaler_state->scalers[0].mode =
> PS_SCALER_MODE_DYN;
> >>>>>> +
> >>>>>> +		intel_crtc->atomic.skl_update_scaler0 =
> >>>>>> +			PS_SCALER_EN | PS_SCALER_MODE_DYN;
> >>>>>> +	}
> >>>>>> +
> >>>>>> +	/* plane not part of mask can enter hq mode? */
> >>>>>> +	if (num_scalers_need == 1 && scaler_state->scalers[0].in_use
> &&
> >>>>>> +	    intel_crtc->pipe != PIPE_C && scaler_state->scalers[0].mode
> !=
> >>>>> PS_SCALER_MODE_HQ) {
> >>>>>> +		scaler_state->scalers[0].mode =
> PS_SCALER_MODE_HQ;
> >>>>>> +
> >>>>>> +		intel_crtc->atomic.skl_update_scaler0 =
> >>>>>> +			PS_SCALER_EN | PS_SCALER_MODE_HQ;
> >>>>>> +	}
> >>>>>> +
> >>>>> I don't have access to the hw spec at the moment; is scaler #0 the only
> >>>>> one that can ever go into HQ mode?
> >>>> Yes
> >>>>
> >>>>> If there isn't a hardware
> >>>>> requirement about this, then it seems like we're missing the case where
> >>>>> planes A and B get scalers 0 and 1.  Then plane A (and thus scaler 0) is
> >>>>> disabled, which should allow scaler 1 to go into HQ mode.
> >>>> In this case, scaler 0 to be allocated to plane B to operate in HQ mode.
> >>> Is it really bad to keep it on scaler 1 for a while until the next time the plane
> is
> >>> added?
> >>>
> >>>>> I guess it's not immediately clear to me why we need to not pull the
> >>>>> other planes into the transaction.  Is this just to avoid doing some
> >>>>> extra work for a plane that hasn't changed, or does it cause a problem
> >>>>> somehow?
> >>>> Per atomic design, unchanged planes can be added to transaction.
> >>>> And scaler implementation is using this design feature.
> >>>> Not sure what the issue here, but we need this feature continue
> >>>> to available.
> >>>>
> >>> Unchanged planes can be added, but this could pull in a primary plane,
> which
> >>> would need
> >>> to set atomic.wait_for_flips then. I can do that as special case when adding
> a
> >>> plane if
> >>> that's preferred.
> >> Here primary plane can get added if that is the only plane using scaler which
> >> isn't part of the transaction. But here addition of primary plane isn't adding
> >> or changing its FB. So why it needs to set atomic.wait_for_flips?
> > In the atomic case, intel_check_primary_plane() sets
> > atomic.wait_for_flips if the primary plane is part of the original
> > transaction (no matter what about the primary plane is changing).
> > However if we pull in the primary plane via
> > intel_atomic_setup_scalers(), that gets called after we've finished
> > checking all of the planes (that were originally part of the
> > transaction), so I don't think wait_for_flips will get set in that case.
> > So I think pulling in the plane as Chandra was will still avoid an
> > unnecessary wait.  Even though it's in the transaction, I believe the
> > 'check' step is bypassed (which means we need to be careful about doing
> > stuff like this if it could affect derived state...but I think in this
> > case it's safe).
> >
> > N.B.  The wait we're talking about here will only be triggered when a
> > SKL or BXT platform disables a "sprite" plane (thus triggering the
> > primary's scaler to switch to HQ mode) while there's a pending legacy
> > pageflip on the primary plane.  I don't think we have a userspace today
> > that can trigger this...Weston won't use sprites at all in a non-atomic
> > manner and although SNA has an Xv sprite adapter, it won't ever be
> > performing scaling of the primary plane afaik.
> Correct. The wait will be a nop if there are no legacy page flips. But for
> completeness sake I included it. :-)

So you will be keeping current code to add plane as-is and let respective 
update_plane function take care of updating scaler quality instead 
of the above update_scaler0?

> 
> ~Maarten