[PATCH 3/3] drm/i915: Disable SAGV on bw init, to force QGV point recalculation

[Ville Syrjälä]

On Thu, Jan 18, 2024 at 10:50:30AM +0200, Lisovskiy, Stanislav wrote:
> On Thu, Jan 18, 2024 at 10:35:56AM +0200, Ville Syrjälä wrote:
> > On Wed, Jan 17, 2024 at 05:57:18PM +0200, Stanislav Lisovskiy wrote:
> > > Problem is that on some platforms, we do get QGV point mask in wrong
> > > state on boot. However driver assumes it is set to 0
> > > (i.e all points allowed), however in reality we might get them all restricted,
> > > causing issues.
> > > Lets disable SAGV initially to force proper QGV point state.
> > > If more QGV points are available, driver will recalculate and update
> > > those then after next commit.
> > > 
> > > v2: - Added trace to see which QGV/PSF GV point is used when SAGV is
> > >       disabled.
> > > v3: - Move force disable function to intel_bw_init in order to initialize
> > >       bw state as well, so that hw/sw are immediately in sync after init.
> > > v4: - Don't try sending PCode request, seems like it is not possible at
> > >       intel_bw_init, however assigning bw->state to be restricted as if
> > >       SAGV is off, still forces driveer to send PCode request anyway on
> > >       next modeset, so the solution still works.
> > >       However we still need to address the case, when no display is connected,
> > >       which anyway requires much more changes.
> > > 
> > > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy at intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/display/intel_bw.c | 24 ++++++++++++++++++++++++
> > >  drivers/gpu/drm/i915/display/intel_bw.h |  2 ++
> > >  2 files changed, 26 insertions(+)
> > > 
> > > diff --git a/drivers/gpu/drm/i915/display/intel_bw.c b/drivers/gpu/drm/i915/display/intel_bw.c
> > > index 7baa1c13eccd..36a6304207ba 100644
> > > --- a/drivers/gpu/drm/i915/display/intel_bw.c
> > > +++ b/drivers/gpu/drm/i915/display/intel_bw.c
> > > @@ -852,6 +852,27 @@ static unsigned int icl_max_bw_qgv_point(struct drm_i915_private *i915,
> > >  	return max_bw_point;
> > >  }
> > >  
> > > +void icl_force_disable_sagv(struct drm_i915_private *i915, struct intel_bw_state *bw_state)
> > > +{
> > > +	unsigned int max_bw_qgv_point = icl_max_bw_qgv_point(i915, 0);
> > > +	unsigned int qgv_points;
> > > +	unsigned int psf_points;
> > > +
> > > +	qgv_points = BIT(max_bw_qgv_point);
> > > +
> > > +	/*
> > > +	 * We don't restrict PSF GV points, when disabling SAGV
> > > +	 */
> > > +	psf_points = 0;
> > 
> > Using 0 looks very wrong here. Since we have no idea how much
> > bandwidth the display is consuming at this time we should
> > restrict this to the max psf gv point as well.
> 
> Didn't we just agree that we are not restricting to max PSF GV
> point, in the last revision?..

That was about restricting PSF GV points during normal SAGV disable,
which is all about the SAGV latency and nothing to do with bandwidth.

> 
> "
> > Yep, but I really suspect we should. BSpec states that we should restrict all the GV points
> > except highest one, also that some PSF GV points aren't same or usable, depending on BW reqs.
> > So I would restrict that as well, in case if SAGV is off, just to be on safe side.
> 
> Pretty sure it's explicitly noted that PSF doesn't cause issues with
> latency and hence doesn't need this.
> 
> In any case, a change like this has no business being in a patch
> that's just supposed to refactor code.
> "
> 
> 
> 
> > 
> > > +
> > > +	bw_state->qgv_points_mask = ~(ICL_PCODE_REQ_QGV_PT(qgv_points) |
> > > +				      ADLS_PCODE_REQ_PSF_PT(psf_points)) &
> > > +				      icl_qgv_points_mask(i915);
> > > +
> > > +	drm_dbg_kms(&i915->drm, "Forcing SAGV disable: leaving QGV point %d\n",
> > > +				max_bw_qgv_point);
> > 
> > You didn't actually poke the hardware to disable anything.
> 
> I know, problem is that PCode request doesn't work at this stage.

Probably because you passed in garbage in the PSF mask. Please
check what pcode is reporting back to you.

> Need to figure out why, but apparently it seems a bit too early.
> PCode just rejects that request.
> 
> However that still works, because if more QGV points are available, driver
> will send a new request anyway on next modeset.

And what happens if the first commit needs to disable SAGV?

> 
> Stan
> 
> > 
> > > +}
> > > +
> > >  static int mtl_find_qgv_points(struct drm_i915_private *i915,
> > >  			       unsigned int data_rate,
> > >  			       unsigned int num_active_planes,
> > > @@ -1351,5 +1372,8 @@ int intel_bw_init(struct drm_i915_private *dev_priv)
> > >  	intel_atomic_global_obj_init(dev_priv, &dev_priv->display.bw.obj,
> > >  				     &state->base, &intel_bw_funcs);
> > >  
> > > +	if (DISPLAY_VER(dev_priv) < 14)
> > 
> > Should be some kind of range check to avoid putting garbage in there on
> > old platforms that don't support QGV.
> > 
> > > +		icl_force_disable_sagv(dev_priv, state);
> > > +
> > >  	return 0;
> > >  }
> > > diff --git a/drivers/gpu/drm/i915/display/intel_bw.h b/drivers/gpu/drm/i915/display/intel_bw.h
> > > index 59cb4fc5db76..243192fd4cae 100644
> > > --- a/drivers/gpu/drm/i915/display/intel_bw.h
> > > +++ b/drivers/gpu/drm/i915/display/intel_bw.h
> > > @@ -74,5 +74,7 @@ int intel_bw_calc_min_cdclk(struct intel_atomic_state *state,
> > >  			    bool *need_cdclk_calc);
> > >  int intel_bw_min_cdclk(struct drm_i915_private *i915,
> > >  		       const struct intel_bw_state *bw_state);
> > > +void icl_force_disable_sagv(struct drm_i915_private *dev_priv,
> > > +			    struct intel_bw_state *bw_state);
> > 
> > Why?
> > 
> > >  
> > >  #endif /* __INTEL_BW_H__ */
> > > -- 
> > > 2.37.3
> > 
> > -- 
> > Ville Syrjälä
> > Intel

-- 
Ville Syrjälä
Intel