| Message ID | 20210107112500.16216-1-anshuman.gupta@intel.com (mailing list archive) |
|---|---|
| State | New, archived |
| Headers | show |
| Series | drm/i915/pps: Reuse POWER_DOMAIN_DISPLAY_CORE in pps_{lock, unlock} | expand |
On Thu, 07 Jan 2021, Anshuman Gupta <anshuman.gupta@intel.com> wrote: > We need a power_domain wakeref in pps_{lock,unlock} to prevent > a race while resetting pps state in intel_power_sequencer_reset(). > > intel_power_sequencer_reset() need a pps_mutex to access pps_pipe > but it can't grab pps_mutex due to deadlock with power_well > functions are called while holding pps_mutex. > intel_power_sequencer_reset() is called by power_well function > associated with legacy platforms like vlv and chv therefore re-use > the POWER_DOMAIN_DISPLAY_CORE power domain, which only used > by vlv and chv display power domain. > > This will avoids the unnecessary noise of unrelated power wells > in pps_{lock,unlock}. > > Cc: Jani Nikula <jani.nikula@intel.com> > Cc: Imre Deak <imre.deak@intel.com> > Signed-off-by: Anshuman Gupta <anshuman.gupta@intel.com> Imre convinced me yesterday on irc that this should work. Reviewed-by: Jani Nikula <jani.nikula@intel.com> On the surface, this reduces the need to enable/disable the aux power so much. It's unnecessary, so it stands to reason to optimize it. We should only grab the domain references we actually need. However, this *also* papers over an issue we've been seeing [1]. We need to be aware the root cause for that remains unknown, and needs to be figured out. I presume simply doing aux transfers won't reproduce the problem, because that disables the power asynchronously since commit f39194a7a8b9 ("drm/i915: Disable power asynchronously during DP AUX transfers"). Perhaps we wouldn't have seen this at all if pps_unlock() also did that as suggested in the commit message. Anyway, I'd like to get acks or rb's from Imre and Ville before merging this. BR, Jani. [1] http://lore.kernel.org/r/20201204081845.26528-1-anshuman.gupta@intel.com > --- > drivers/gpu/drm/i915/display/intel_dp.c | 8 ++------ > 1 file changed, 2 insertions(+), 6 deletions(-) > > diff --git a/drivers/gpu/drm/i915/display/intel_dp.c b/drivers/gpu/drm/i915/display/intel_dp.c > index 8a00e609085f..4f190a82d4ad 100644 > --- a/drivers/gpu/drm/i915/display/intel_dp.c > +++ b/drivers/gpu/drm/i915/display/intel_dp.c > @@ -895,9 +895,7 @@ pps_lock(struct intel_dp *intel_dp) > * See intel_power_sequencer_reset() why we need > * a power domain reference here. > */ > - wakeref = intel_display_power_get(dev_priv, > - intel_aux_power_domain(dp_to_dig_port(intel_dp))); > - > + wakeref = intel_display_power_get(dev_priv, POWER_DOMAIN_DISPLAY_CORE); > mutex_lock(&dev_priv->pps_mutex); > > return wakeref; > @@ -909,9 +907,7 @@ pps_unlock(struct intel_dp *intel_dp, intel_wakeref_t wakeref) > struct drm_i915_private *dev_priv = dp_to_i915(intel_dp); > > mutex_unlock(&dev_priv->pps_mutex); > - intel_display_power_put(dev_priv, > - intel_aux_power_domain(dp_to_dig_port(intel_dp)), > - wakeref); > + intel_display_power_put(dev_priv, POWER_DOMAIN_DISPLAY_CORE, wakeref); > return 0; > }
On Fri, Jan 08, 2021 at 11:38:04AM +0200, Jani Nikula wrote: > On Thu, 07 Jan 2021, Anshuman Gupta <anshuman.gupta@intel.com> wrote: > > We need a power_domain wakeref in pps_{lock,unlock} to prevent > > a race while resetting pps state in intel_power_sequencer_reset(). > > > > intel_power_sequencer_reset() need a pps_mutex to access pps_pipe > > but it can't grab pps_mutex due to deadlock with power_well > > functions are called while holding pps_mutex. > > intel_power_sequencer_reset() is called by power_well function > > associated with legacy platforms like vlv and chv therefore re-use > > the POWER_DOMAIN_DISPLAY_CORE power domain, which only used > > by vlv and chv display power domain. > > > > This will avoids the unnecessary noise of unrelated power wells > > in pps_{lock,unlock}. > > > > Cc: Jani Nikula <jani.nikula@intel.com> > > Cc: Imre Deak <imre.deak@intel.com> > > Signed-off-by: Anshuman Gupta <anshuman.gupta@intel.com> > > Imre convinced me yesterday on irc that this should work. > > Reviewed-by: Jani Nikula <jani.nikula@intel.com> > > On the surface, this reduces the need to enable/disable the aux power so > much. It's unnecessary, so it stands to reason to optimize it. We should > only grab the domain references we actually need. > > However, this *also* papers over an issue we've been seeing [1]. We need > to be aware the root cause for that remains unknown, and needs to be > figured out. > > I presume simply doing aux transfers won't reproduce the problem, > because that disables the power asynchronously since commit f39194a7a8b9 > ("drm/i915: Disable power asynchronously during DP AUX > transfers"). Perhaps we wouldn't have seen this at all if pps_unlock() > also did that as suggested in the commit message. > > Anyway, I'd like to get acks or rb's from Imre and Ville before merging > this. Looks ok to me: Acked-by: Imre Deak <imre.deak@intel.com> > > > BR, > Jani. > > > [1] http://lore.kernel.org/r/20201204081845.26528-1-anshuman.gupta@intel.com > > > > --- > > drivers/gpu/drm/i915/display/intel_dp.c | 8 ++------ > > 1 file changed, 2 insertions(+), 6 deletions(-) > > > > diff --git a/drivers/gpu/drm/i915/display/intel_dp.c b/drivers/gpu/drm/i915/display/intel_dp.c > > index 8a00e609085f..4f190a82d4ad 100644 > > --- a/drivers/gpu/drm/i915/display/intel_dp.c > > +++ b/drivers/gpu/drm/i915/display/intel_dp.c > > @@ -895,9 +895,7 @@ pps_lock(struct intel_dp *intel_dp) > > * See intel_power_sequencer_reset() why we need > > * a power domain reference here. > > */ > > - wakeref = intel_display_power_get(dev_priv, > > - intel_aux_power_domain(dp_to_dig_port(intel_dp))); > > - > > + wakeref = intel_display_power_get(dev_priv, POWER_DOMAIN_DISPLAY_CORE); > > mutex_lock(&dev_priv->pps_mutex); > > > > return wakeref; > > @@ -909,9 +907,7 @@ pps_unlock(struct intel_dp *intel_dp, intel_wakeref_t wakeref) > > struct drm_i915_private *dev_priv = dp_to_i915(intel_dp); > > > > mutex_unlock(&dev_priv->pps_mutex); > > - intel_display_power_put(dev_priv, > > - intel_aux_power_domain(dp_to_dig_port(intel_dp)), > > - wakeref); > > + intel_display_power_put(dev_priv, POWER_DOMAIN_DISPLAY_CORE, wakeref); > > return 0; > > } > > -- > Jani Nikula, Intel Open Source Graphics Center
On Fri, 08 Jan 2021, Imre Deak <imre.deak@intel.com> wrote: > On Fri, Jan 08, 2021 at 11:38:04AM +0200, Jani Nikula wrote: >> On Thu, 07 Jan 2021, Anshuman Gupta <anshuman.gupta@intel.com> wrote: >> > We need a power_domain wakeref in pps_{lock,unlock} to prevent >> > a race while resetting pps state in intel_power_sequencer_reset(). >> > >> > intel_power_sequencer_reset() need a pps_mutex to access pps_pipe >> > but it can't grab pps_mutex due to deadlock with power_well >> > functions are called while holding pps_mutex. >> > intel_power_sequencer_reset() is called by power_well function >> > associated with legacy platforms like vlv and chv therefore re-use >> > the POWER_DOMAIN_DISPLAY_CORE power domain, which only used >> > by vlv and chv display power domain. >> > >> > This will avoids the unnecessary noise of unrelated power wells >> > in pps_{lock,unlock}. >> > >> > Cc: Jani Nikula <jani.nikula@intel.com> >> > Cc: Imre Deak <imre.deak@intel.com> >> > Signed-off-by: Anshuman Gupta <anshuman.gupta@intel.com> >> >> Imre convinced me yesterday on irc that this should work. >> >> Reviewed-by: Jani Nikula <jani.nikula@intel.com> >> >> On the surface, this reduces the need to enable/disable the aux power so >> much. It's unnecessary, so it stands to reason to optimize it. We should >> only grab the domain references we actually need. >> >> However, this *also* papers over an issue we've been seeing [1]. We need >> to be aware the root cause for that remains unknown, and needs to be >> figured out. >> >> I presume simply doing aux transfers won't reproduce the problem, >> because that disables the power asynchronously since commit f39194a7a8b9 >> ("drm/i915: Disable power asynchronously during DP AUX >> transfers"). Perhaps we wouldn't have seen this at all if pps_unlock() >> also did that as suggested in the commit message. >> >> Anyway, I'd like to get acks or rb's from Imre and Ville before merging >> this. > > Looks ok to me: > Acked-by: Imre Deak <imre.deak@intel.com> Thanks, pushed to din. BR, Jani. > >> >> >> BR, >> Jani. >> >> >> [1] http://lore.kernel.org/r/20201204081845.26528-1-anshuman.gupta@intel.com >> >> >> > --- >> > drivers/gpu/drm/i915/display/intel_dp.c | 8 ++------ >> > 1 file changed, 2 insertions(+), 6 deletions(-) >> > >> > diff --git a/drivers/gpu/drm/i915/display/intel_dp.c b/drivers/gpu/drm/i915/display/intel_dp.c >> > index 8a00e609085f..4f190a82d4ad 100644 >> > --- a/drivers/gpu/drm/i915/display/intel_dp.c >> > +++ b/drivers/gpu/drm/i915/display/intel_dp.c >> > @@ -895,9 +895,7 @@ pps_lock(struct intel_dp *intel_dp) >> > * See intel_power_sequencer_reset() why we need >> > * a power domain reference here. >> > */ >> > - wakeref = intel_display_power_get(dev_priv, >> > - intel_aux_power_domain(dp_to_dig_port(intel_dp))); >> > - >> > + wakeref = intel_display_power_get(dev_priv, POWER_DOMAIN_DISPLAY_CORE); >> > mutex_lock(&dev_priv->pps_mutex); >> > >> > return wakeref; >> > @@ -909,9 +907,7 @@ pps_unlock(struct intel_dp *intel_dp, intel_wakeref_t wakeref) >> > struct drm_i915_private *dev_priv = dp_to_i915(intel_dp); >> > >> > mutex_unlock(&dev_priv->pps_mutex); >> > - intel_display_power_put(dev_priv, >> > - intel_aux_power_domain(dp_to_dig_port(intel_dp)), >> > - wakeref); >> > + intel_display_power_put(dev_priv, POWER_DOMAIN_DISPLAY_CORE, wakeref); >> > return 0; >> > } >> >> -- >> Jani Nikula, Intel Open Source Graphics Center > _______________________________________________ > Intel-gfx mailing list > Intel-gfx@lists.freedesktop.org > https://lists.freedesktop.org/mailman/listinfo/intel-gfx
diff --git a/drivers/gpu/drm/i915/display/intel_dp.c b/drivers/gpu/drm/i915/display/intel_dp.c index 8a00e609085f..4f190a82d4ad 100644 --- a/drivers/gpu/drm/i915/display/intel_dp.c +++ b/drivers/gpu/drm/i915/display/intel_dp.c @@ -895,9 +895,7 @@ pps_lock(struct intel_dp *intel_dp) * See intel_power_sequencer_reset() why we need * a power domain reference here. */ - wakeref = intel_display_power_get(dev_priv, - intel_aux_power_domain(dp_to_dig_port(intel_dp))); - + wakeref = intel_display_power_get(dev_priv, POWER_DOMAIN_DISPLAY_CORE); mutex_lock(&dev_priv->pps_mutex); return wakeref; @@ -909,9 +907,7 @@ pps_unlock(struct intel_dp *intel_dp, intel_wakeref_t wakeref) struct drm_i915_private *dev_priv = dp_to_i915(intel_dp); mutex_unlock(&dev_priv->pps_mutex); - intel_display_power_put(dev_priv, - intel_aux_power_domain(dp_to_dig_port(intel_dp)), - wakeref); + intel_display_power_put(dev_priv, POWER_DOMAIN_DISPLAY_CORE, wakeref); return 0; }
We need a power_domain wakeref in pps_{lock,unlock} to prevent a race while resetting pps state in intel_power_sequencer_reset(). intel_power_sequencer_reset() need a pps_mutex to access pps_pipe but it can't grab pps_mutex due to deadlock with power_well functions are called while holding pps_mutex. intel_power_sequencer_reset() is called by power_well function associated with legacy platforms like vlv and chv therefore re-use the POWER_DOMAIN_DISPLAY_CORE power domain, which only used by vlv and chv display power domain. This will avoids the unnecessary noise of unrelated power wells in pps_{lock,unlock}. Cc: Jani Nikula <jani.nikula@intel.com> Cc: Imre Deak <imre.deak@intel.com> Signed-off-by: Anshuman Gupta <anshuman.gupta@intel.com> --- drivers/gpu/drm/i915/display/intel_dp.c | 8 ++------ 1 file changed, 2 insertions(+), 6 deletions(-)