[Intel-gfx] [PATCH] drm/i915: Lock gmbus/aux mutexes while changing cdclk

[Ville Syrjälä]

On Fri, Feb 28, 2020 at 01:10:45PM +0200, Ville Syrjälä wrote:
> On Fri, Feb 28, 2020 at 11:06:41AM +0200, Jani Nikula wrote:
> > On Thu, 27 Feb 2020, Ville Syrjala <ville.syrjala at linux.intel.com> wrote:
> > > From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > >
> > > gmbus/aux may be clocked by cdclk, thus we should make sure no
> > > transfers are ongoing while the cdclk frequency is being changed.
> > > We do that by simply grabbing all the gmbus/aux mutexes. No one
> > > else should be holding any more than one of those at a time so
> > > the lock ordering here shouldn't matter.
> > >
> > > Signed-off-by: Ville Syrjälä <ville.syrjala at linux.intel.com>
> > > ---
> > >  drivers/gpu/drm/i915/display/intel_cdclk.c | 23 ++++++++++++++++++++++
> > >  1 file changed, 23 insertions(+)
> > >
> > > diff --git a/drivers/gpu/drm/i915/display/intel_cdclk.c b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > index 0741d643455b..f69bf4a4eb1c 100644
> > > --- a/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > +++ b/drivers/gpu/drm/i915/display/intel_cdclk.c
> > > @@ -1868,6 +1868,9 @@ static void intel_set_cdclk(struct drm_i915_private *dev_priv,
> > >  			    const struct intel_cdclk_config *cdclk_config,
> > >  			    enum pipe pipe)
> > >  {
> > > +	struct intel_encoder *encoder;
> > > +	unsigned int aux_mutex_lockclass = 0;
> > > +
> > >  	if (!intel_cdclk_changed(&dev_priv->cdclk.hw, cdclk_config))
> > >  		return;
> > >  
> > > @@ -1876,8 +1879,28 @@ static void intel_set_cdclk(struct drm_i915_private *dev_priv,
> > >  
> > >  	intel_dump_cdclk_config(cdclk_config, "Changing CDCLK to");
> > >  
> > > +	/*
> > > +	 * Lock aux/gmbus while we change cdclk in case those
> > > +	 * functions use cdclk. Not all platforms/ports do,
> > > +	 * but we'll lock them all for simplicity.
> > > +	 */
> > > +	mutex_lock(&dev_priv->gmbus_mutex);
> > > +	for_each_intel_dp(&dev_priv->drm, encoder) {
> > > +		struct intel_dp *intel_dp = enc_to_intel_dp(encoder);
> > > +
> > > +		mutex_lock_nested(&intel_dp->aux.hw_mutex,
> > > +				  aux_mutex_lockclass++);
> > > +	}
> > > +
> > >  	dev_priv->display.set_cdclk(dev_priv, cdclk_config, pipe);
> > >  
> > > +	for_each_intel_dp(&dev_priv->drm, encoder) {
> > > +		struct intel_dp *intel_dp = enc_to_intel_dp(encoder);
> > > +
> > > +		mutex_unlock(&intel_dp->aux.hw_mutex);
> > > +	}
> > > +	mutex_unlock(&dev_priv->gmbus_mutex);
> > > +
> > 
> > I'm becoming increasingly sensitive to directly touching the private
> > parts of other modules... gmbus_mutex is really for intel_gmbus.c and
> > aux.hw_mutex for drm_dp_helper.c.
> > 
> > One could also argue that the cdclk is a lower level function used by
> > higher level functions aux/gmbus, and it seems like the higher level
> > function should lock the cdclk while it depends on it, not the other way
> > around.
> 
> That would require a rwsem. Otherwise it all gets serialized needlessly.
> Not sure what's the state of rwsems these days, but IIRC at some point
> the rt patches converted them all to normal mutexes.

Some googling suggests that my infromation may be out of date. So we
could introduce an rwsem for this I guess. Would add a bit more cost
to the common codepaths though (not that they're really performance
critical) whereas this version only adds extra cost to the much more
rare .set_cdclk() path.

-- 
Ville Syrjälä
Intel