Re: [PATCH] drm/xe/pm: Change HPD to polling on runtime suspend

[Jani Nikula <jani.nikula@linux.intel.com>]

On Wed, 07 Aug 2024, Imre Deak <imre.deak@intel.com> wrote:
> On Wed, Aug 07, 2024 at 12:15:33PM +0300, Murthy, Arun R wrote:
>> 
>> > -----Original Message-----
>> > From: Deak, Imre <imre.deak@intel.com>
>> > Sent: Tuesday, August 6, 2024 10:11 PM
>> > To: Murthy, Arun R <arun.r.murthy@intel.com>
>> > Cc: intel-xe@lists.freedesktop.org; intel-gfx@lists.freedesktop.org
>> > Subject: Re: [PATCH] drm/xe/pm: Change HPD to polling on runtime suspend
>> >
>> > On Tue, Aug 06, 2024 at 02:25:13PM +0530, Arun R Murthy wrote:
>> > > HPD is interrupt based and on runtime suspend change it to polling as
>> > > HPD is not a wakeup event. A worker thread is scheduled for doing this
>> > > polling and it keeps polling for HPD live status on an internval of 10s.
>> > > On runtime resume disable polling and fallback to interrupt mechanism.
>> > >
>> > > Signed-off-by: Arun R Murthy <arun.r.murthy@intel.com>
>> > > ---
>> > >  drivers/gpu/drm/xe/xe_pm.c | 6 ++++++
>> > >  1 file changed, 6 insertions(+)
>> > >
>> > > diff --git a/drivers/gpu/drm/xe/xe_pm.c b/drivers/gpu/drm/xe/xe_pm.c
>> > > index 9f3c14fd9f33..d952b06ebfb4 100644
>> > > --- a/drivers/gpu/drm/xe/xe_pm.c
>> > > +++ b/drivers/gpu/drm/xe/xe_pm.c
>> > > @@ -22,6 +22,7 @@
>> > >  #include "xe_pcode.h"
>> > >  #include "xe_trace.h"
>> > >  #include "xe_wa.h"
>> > > +#include "intel_hotplug.h"
>> > >
>> > >  /**
>> > >   * DOC: Xe Power Management
>> > > @@ -378,6 +379,8 @@ int xe_pm_runtime_suspend(struct xe_device *xe)
>> > >             if (err)
>> > >                     goto out;
>> > >     }
>> > > +   /* Enable hpd polling on runtime suspend */
>> > > +   intel_hpd_poll_enable(xe);
>> > >
>> > >     xe_irq_suspend(xe);
>> > >
>> > > @@ -427,6 +430,9 @@ int xe_pm_runtime_resume(struct xe_device *xe)
>> > >
>> > >     xe_irq_resume(xe);
>> > >
>> > > +   /* Disable hpd polling on runtime resume */
>> > > +   intel_hpd_poll_disable(xe);
>> >
>> > This should be handled in xe_display_pm_resume() (likewise enabling polling in
>> > xe_display_pm_suspend()), in fact it's already called there.
>>
>> In xe_display_pm_suspend we have drm_kms_helper_poll_disable and it disables
>> output polling.
>
> What I meant is that intel_hpd_poll_disable() is already called in
> xe_display_pm_resume(), which you would call now twice during runtime
> resume if d3cold is allowed.
>
> drm_kms_helper_poll_disable() in xe_display_pm_suspend() should be
> called only during system suspend, which is another thing to fix there.
>
>> Moreover we don't have the functions defined for enable/disable_hpd
>> defined in connector_ops which is referenced by
>> drm_kms_helper_poll_enable/disable.
>> 
>> What we need is on runtime_suspend the poll work should be
>> enabled so that hotplug can be detected and can come out of runtime_suspend.
>> The same poll work is not required on system suspend.
>
> All the display related suspend/resume handling is done already in
> xe_display_pm_suspend()/resume() or in xe_display.c in general, both for
> runtime and system suspend/resume. The reason for that is to keep the
> display specific parts separate from the generic/higher-level device
> suspend/resume handling in xe_pm.c. New display specific things added
> should keep this separation.

I'll keep repeating here too: We should aim to move all of this under
i915/display, and not duplicate it in i915 and xe specific code.

BR,
Jani.



>
>> > Also this is missing the xe->info.enable_display check and polling shouldn't be
>> > enabled when system (vs. runtime) suspending.
>> Probably HAS_DISPLAY can be added before calling this intel_hpd_poll_enable.
>> 
>> Thanks and Regards,
>> Arun R Murthy
>> --------------------
>> >
>> > Not sure why the display is suspended/resumed only if d3cold is allowed, for
>> > many of the display s/r steps at least it doesn't make sense to me.
>> > For now moving the d3cold check within
>> > xe_display_pm_suspend()/resume() and enabling/disabling polling regardless of
>> > that flag would be the simplest imo.
>> >
>> > > +
>> > >     for_each_gt(gt, xe, id)
>> > >             xe_gt_resume(gt);
>> > >
>> > > --
>> > > 2.25.1
>> > >

-- 
Jani Nikula, Intel