Re: [PATCH 2/3] drm/i915/bigjoiner: Refactor bigjoiner state readout

["Lisovskiy, Stanislav" <stanislav.lisovskiy@intel.com>]

On Fri, Jan 12, 2024 at 06:42:15PM +0200, Ville Syrjälä wrote:
> On Mon, Jan 08, 2024 at 02:07:24PM +0200, Stanislav Lisovskiy wrote:
> > Don't call enabled_bigjoiner_pipes twice, lets just move
> > intel_get_bigjoiner_config earlier, because it is anyway
> > calling same function.
> > Also cleanup hsw_enabled_transcoders from irrelevant bigjoiner code.
> 
> It's not irrelevant. The function is supposed to return the set of
> enabled transcoders associated with the pipe. With this change the
> function no longer does what it says on the tin.

Yes, but I guess it is just a matter what we define to be higher in a
logical hierarchy: a pipe or a bigjoiner?
I thought it won't harm hsw_enabled_transcoders won't have any excess
logic and will return only transcoders naturally associated with a
physical pipe, while for higher complexity level constructs like bigjoiner
we would have some logic on top. 
In fact my main motivation was to avoid calling enabled_bigjoiner_pipe as 
it is quite heavy and call intel_crtc_is_bigjoiner_slave here instead. 

enabled_bigjoiner_pipes reads too much information, which 
we don't need in that function(here we just need to know if we are slave or master)

The absolute need for calling enabled_bigjoiner_pipes happens in 
intel_bigjoiner_get_config, which we moved earlier, which seems to be
logical since hsw_get_transcoder_state needs to have bigjoiner info and
now we can use its results to call more lightweight intel_crtc_is_bigjoiner_slave there
because pipe_config->bigjoiner_pipes is now initialized, so we avoid calling enabled_bigjoiner_pipes
second time..

> 
> > 
> > Signed-off-by: Stanislav Lisovskiy <stanislav.lisovskiy@intel.com>
> > ---
> >  drivers/gpu/drm/i915/display/intel_display.c | 22 ++++++++++----------
> >  1 file changed, 11 insertions(+), 11 deletions(-)
> > 
> > diff --git a/drivers/gpu/drm/i915/display/intel_display.c b/drivers/gpu/drm/i915/display/intel_display.c
> > index 927d124457b61..eec76ec167069 100644
> > --- a/drivers/gpu/drm/i915/display/intel_display.c
> > +++ b/drivers/gpu/drm/i915/display/intel_display.c
> > @@ -3525,7 +3525,6 @@ static u8 hsw_enabled_transcoders(struct intel_crtc *crtc)
> >  	struct drm_i915_private *dev_priv = to_i915(dev);
> >  	u8 panel_transcoder_mask = hsw_panel_transcoders(dev_priv);
> >  	enum transcoder cpu_transcoder;
> > -	u8 master_pipes, slave_pipes;
> >  	u8 enabled_transcoders = 0;
> >  
> >  	/*
> > @@ -3576,15 +3575,6 @@ static u8 hsw_enabled_transcoders(struct intel_crtc *crtc)
> >  	if (transcoder_ddi_func_is_enabled(dev_priv, cpu_transcoder))
> >  		enabled_transcoders |= BIT(cpu_transcoder);
> >  
> > -	/* bigjoiner slave -> consider the master pipe's transcoder as well */
> > -	enabled_bigjoiner_pipes(dev_priv, &master_pipes, &slave_pipes);
> > -	if (slave_pipes & BIT(crtc->pipe)) {
> > -		cpu_transcoder = (enum transcoder)
> > -			get_bigjoiner_master_pipe(crtc->pipe, master_pipes, slave_pipes);
> > -		if (transcoder_ddi_func_is_enabled(dev_priv, cpu_transcoder))
> > -			enabled_transcoders |= BIT(cpu_transcoder);
> > -	}
> > -
> >  	return enabled_transcoders;
> >  }
> >  
> > @@ -3631,6 +3621,15 @@ static bool hsw_get_transcoder_state(struct intel_crtc *crtc,
> >  	u32 tmp;
> >  
> >  	enabled_transcoders = hsw_enabled_transcoders(crtc);
> > +
> > +	/* bigjoiner slave -> consider the master pipe's transcoder as well */
> > +	if (intel_crtc_is_bigjoiner_slave(pipe_config)) {
> > +		unsigned long cpu_transcoder = (enum transcoder)
> > +			bigjoiner_master_pipe(pipe_config);
> > +		if (transcoder_ddi_func_is_enabled(dev_priv, cpu_transcoder))
> > +			enabled_transcoders |= BIT(cpu_transcoder);
> > +	}
> > +
> >  	if (!enabled_transcoders)
> >  		return false;
> >  
> > @@ -3735,6 +3734,8 @@ static bool hsw_get_pipe_config(struct intel_crtc *crtc,
> >  
> >  	pipe_config->shared_dpll = NULL;
> >  
> > +	intel_bigjoiner_get_config(pipe_config);
> 
> So this is what avoids the "call it twice" part, but it also makes the
> state potentially inconsistent as in all other cases we leave everything
> zeroed if the transcoder is not enabled. So I'm not sure this is
> entirely safe or whether we could end up with some weird state
> mismatches due to the inconsistency.

Isn't it vice versa? intel_bigjoiner_get_config is now called way earlier,
before hsw_get_transcoder_state is called(previously it was called later),
the only difference is just that we now have pipe_config->bigjoiner_pipes
filled and enabled_bigjoiner_pipes was called there, so we can now
use that info to call intel_crtc_is_bigjoiner_slave in hsw_get_transcoder_state,
as I mentioned above.

Also if none of the transcoders are enabled, we now in fact have more information
filled than before this change(before we had only enabled_bigjoiner_pipes called
in hsw_get_transcoder_state, but now we have also pipe_config->bigjoiner_pipes
initialized), otherwise if none of the transcoders are active everything should
be pretty much the same.

Stan


> 
> Why do you think calling it twice is problematic? It's supposed to be 
> idempotent (ignoring the actual register reads/etc.).
> 
> > +
> >  	active = hsw_get_transcoder_state(crtc, pipe_config, &crtc->hw_readout_power_domains);
> >  
> >  	if ((IS_GEMINILAKE(dev_priv) || IS_BROXTON(dev_priv)) &&
> > @@ -3746,7 +3747,6 @@ static bool hsw_get_pipe_config(struct intel_crtc *crtc,
> >  	if (!active)
> >  		goto out;
> >  
> > -	intel_bigjoiner_get_config(pipe_config);
> >  	intel_dsc_get_config(pipe_config);
> >  
> >  	if (!transcoder_is_dsi(pipe_config->cpu_transcoder) ||
> > -- 
> > 2.37.3
> 
> -- 
> Ville Syrjälä
> Intel