[PATCH 2/3] drm/i915/dp_mst: Fix side-band message timeouts due to long PPS delays

[Jani Nikula]

On Fri, 21 Mar 2025, Imre Deak <imre.deak at intel.com> wrote:
> On Fri, Mar 21, 2025 at 08:44:22PM +0200, Ville Syrjälä wrote:
>> On Fri, Mar 21, 2025 at 08:38:45PM +0200, Imre Deak wrote:
>> > On Fri, Mar 21, 2025 at 08:00:29PM +0200, Ville Syrjälä wrote:
>> > > On Fri, Mar 21, 2025 at 04:56:25PM +0200, Imre Deak wrote:
>> > > > The Panel Power Sequencer lock held on an eDP port (a) blocks a DP AUX
>> > > > transfer on another port (b), since the PPS lock is device global, thus
>> > > > shared by all ports. The PPS lock can be held on port (a) for a longer
>> > > > period due to the various PPS delays (panel/backlight on/off,
>> > > > power-cycle delays). This in turn can cause an MST down-message request
>> > > > on port (b) time out, if the above PPS delay defers the handling of the
>> > > > reply to the request by more than 100ms: the MST branch device sending
>> > > > the reply (signaling this via the DP_DOWN_REP_MSG_RDY flag in the
>> > > > DP_DEVICE_SERVICE_IRQ_VECTOR DPCD register) may cancel the reply
>> > > > (clearing DP_DOWN_REP_MSG_RDY and the reply message buffer) after 110
>> > > > ms, if the reply is not processed by that time.
>> > > > 
>> > > > Avoid MST down-message timeouts described above, by locking the PPS
>> > > > state for AUX transfers only if this is actually required: on eDP ports,
>> > > > where the VDD power depends on the PPS state and on all DP and eDP ports
>> > > > on VLV/CHV, where the PPS is a pipe instance and hence a modeset on any
>> > > > port possibly affecting the PPS state.
>> > > > 
>> > > > Signed-off-by: Imre Deak <imre.deak at intel.com>
>> > > > ---
>> > > >  drivers/gpu/drm/i915/display/intel_pps.c | 34 ++++++++++++++++++++++++
>> > > >  1 file changed, 34 insertions(+)
>> > > > 
>> > > > diff --git a/drivers/gpu/drm/i915/display/intel_pps.c b/drivers/gpu/drm/i915/display/intel_pps.c
>> > > > index 3c078fd53fbfa..7d7157983f25e 100644
>> > > > --- a/drivers/gpu/drm/i915/display/intel_pps.c
>> > > > +++ b/drivers/gpu/drm/i915/display/intel_pps.c
>> > > > @@ -26,6 +26,11 @@ static void vlv_steal_power_sequencer(struct intel_display *display,
>> > > >  static void pps_init_delays(struct intel_dp *intel_dp);
>> > > >  static void pps_init_registers(struct intel_dp *intel_dp, bool force_disable_vdd);
>> > > >  
>> > > > +static bool intel_pps_is_pipe_instance(struct intel_display *display)
>> > > > +{
>> > > > +	return display->platform.valleyview || display->platform.cherryview;
>> > > > +}
>> > > > +
>> > > >  static const char *pps_name(struct intel_dp *intel_dp)
>> > > >  {
>> > > >  	struct intel_display *display = to_intel_display(intel_dp);
>> > > > @@ -955,10 +960,32 @@ void intel_pps_vdd_off(struct intel_dp *intel_dp)
>> > > >  		intel_pps_vdd_off_unlocked(intel_dp, false);
>> > > >  }
>> > > >  
>> > > > +static bool aux_needs_pps_lock(struct intel_dp *intel_dp)
>> > > > +{
>> > > > +	struct intel_display *display = to_intel_display(intel_dp);
>> > > > +
>> > > > +	/*
>> > > > +	 * The PPS state needs to be locked for:
>> > > > +	 * - eDP on all platforms, since AUX transfers on eDP need VDD power
>> > > > +	 *   (either forced or via panel power) which depends on the PPS
>> > > > +	 *   state.
>> > > > +	 * - non-eDP on platforms where the PPS is a pipe instance (VLV/CHV),
>> > > > +	 *   since changing the PPS state (via a parallel modeset for
>> > > > +	 *   instance) may interfere with the AUX transfers on a non-eDP
>> > > > +	 *   output as well.
>> > > > +	 */
>> > > > +	return intel_dp_is_edp(intel_dp) || intel_pps_is_pipe_instance(display);
>> > > > +}
>> > > > +
>> > > >  intel_wakeref_t intel_pps_lock_for_aux(struct intel_dp *intel_dp, bool *vdd_ref)
>> > > >  {
>> > > >  	intel_wakeref_t wakeref;
>> > > >  
>> > > > +	if (!aux_needs_pps_lock(intel_dp)) {
>> > > > +		*vdd_ref = false;
>> > > > +		return NULL;
>> > > 
>> > > I was pondering if we need a define for this since intel_wakeref_t
>> > > doesn't look like a pointer, but apparently we use NULLs elsewhere
>> > > as well for this stuff.
>> > 
>> > Ok, makes sense. It is a bigger a change though, so is it ok to do that
>> > as a follow up?
>> 
>> I'm not sure what we even should do about it. Should all the
>> naked NULLs be hidden, or should we make the thing look like the
>> pointer it actually is?
>
> The latter, i.e.
>
> #define INTEL_WAKEREF_NONE ((intel_wakeref_t)0)

I've been leaning towards making it the pointer it actually is,
i.e. struct ref_tracker *. See the new intel_display_rpm.[ch].

But I have much stronger objections to patches 1 and 2 than this [1].


BR,
Jani.


[1] https://lore.kernel.org/r/874izibtvx.fsf@intel.com




>
> ?
>
>> > > > +	}
>> > > > +
>> > > >  	wakeref = intel_pps_lock(intel_dp);
>> > > >  
>> > > >  	/*
>> > > > @@ -976,6 +1003,13 @@ intel_wakeref_t intel_pps_lock_for_aux(struct intel_dp *intel_dp, bool *vdd_ref)
>> > > >  
>> > > >  void intel_pps_unlock_for_aux(struct intel_dp *intel_dp, intel_wakeref_t wakeref, bool vdd_ref)
>> > > >  {
>> > > > +	struct intel_display *display = to_intel_display(intel_dp);
>> > > > +
>> > > > +	if (!wakeref) {
>> > > > +		drm_WARN_ON(display->drm, vdd_ref || aux_needs_pps_lock(intel_dp));
>> > > > +		return;
>> > > > +	}
>> > > > +
>> > > >  	if (vdd_ref)
>> > > >  		intel_pps_vdd_off_unlocked(intel_dp, false);
>> > > >  
>> > > > -- 
>> > > > 2.44.2
>> > > 
>> > > -- 
>> > > Ville Syrjälä
>> > > Intel
>> 
>> -- 
>> Ville Syrjälä
>> Intel

-- 
Jani Nikula, Intel