[Intel-gfx] [PATCH 4/4] drm/i915: Don't allocate extra ddb during async flip for DG2

[Ville Syrjälä]

On Sun, Jan 23, 2022 at 10:34:17PM +0200, Lisovskiy, Stanislav wrote:
> On Fri, Jan 21, 2022 at 02:06:12PM +0200, Ville Syrjälä wrote:
> > On Fri, Jan 21, 2022 at 10:06:15AM +0200, Stanislav Lisovskiy wrote:
> > > In terms of async flip optimization we don't to allocate
> > > extra ddb space, so lets skip it.
> > > 
> > > v2: - Extracted min ddb async flip check to separate function
> > >       (Ville Syrjälä)
> > >     - Used this function to prevent false positive WARN
> > >       to be triggered(Ville Syrjälä)
> > > 
> > > v3: - Renamed dg2_need_min_ddb to need_min_ddb thus making
> > >       it more universal.
> > >     - Also used DISPLAY_VER instead of IS_DG2(Ville Syrjälä)
> > >     - Use rate = 0 instead of just setting extra = 0, thus
> > >       letting other planes to use extra ddb and avoiding WARN
> > >       (Ville Syrjälä)
> > > 
> > > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy at intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/intel_pm.c | 17 +++++++++++++++++
> > >  1 file changed, 17 insertions(+)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/intel_pm.c b/drivers/gpu/drm/i915/intel_pm.c
> > > index 5fb022a2a4d7..18fb35c480ef 100644
> > > --- a/drivers/gpu/drm/i915/intel_pm.c
> > > +++ b/drivers/gpu/drm/i915/intel_pm.c
> > > @@ -5118,6 +5118,12 @@ static bool icl_need_wm1_wa(struct drm_i915_private *i915,
> > >  	       (IS_DISPLAY_VER(i915, 12, 13) && plane_id == PLANE_CURSOR);
> > >  }
> > >  
> > > +static bool needs_min_ddb(struct drm_i915_private *i915,
> > > +			  struct intel_crtc_state *crtc_state)
> > 
> > s/needs/use/ to match the wm0 counterpart?
> > 
> > Could use a comment as well perhaps, or maybe just put this right
> > next to the wm0 counterpart so the reader can see both together and
> > make the connection.
> > 
> > Hmm. Actually I think this would also need the plane->async_flip
> > check here too or else we'll drop all the planes to min ddb
> > instead of just the plane doing async flips.
> > 
> > Oh, and I think we need this same thing when calculating the
> > total_data_rate or else the numbers won't match.
> 
> Yes, there seems to be a problem with that approach, we use ratio
> from data plane_data_rate/total_data_rate to determine how we split
> extra ddb blocks, however if plane data rate can be just set as 0
> here localle, total_data_rate is obtained from crtc_state->plane_data_rate,
> which is being calculated first.
> So if we trick icl_get_total_relative_data_rate function to calculate
> total_data_rate corresponding to rate = 0, we will then have 
> crtc_state->plane_data_rate[plane_id] set to 0, which is probably
> not what we want.

These are just the relative data rates so they're not actually used for
anything else. So I guess we could even set them to 0. Though I don't
even recall if the current code really works or not. I think there
might have been some problem with calculating these that I perhaps fixed
with my latest ddb series (or maybe I already fixed it with some earlier
series, can't remember anymore).

> 
> Or should I just edit icl_get_total_relative_data_rate so that it
> still calculates crtc_state->plane_data_rate properly however, the
> doesn't add those to total_data_rate, if use_min_ddb(plane) is set?

This should work too. Can't immediately think why one approach would
be strictly better than the other.

-- 
Ville Syrjälä
Intel