Re: [PATCH v2] drm/xe/pm: Handle GT resume failure

[Raag Jadav <raag.jadav@intel.com>]

On Fri, Jan 09, 2026 at 08:20:12PM +0530, Vivi, Rodrigo wrote:
> On Fri, 2026-01-09 at 14:43 +0100, Raag Jadav wrote:
> > On Fri, Jan 09, 2026 at 04:37:39PM +0530, Nilawar, Badal wrote:
> > > On 09-01-2026 11:28, Raag Jadav wrote:
> > > > On Fri, Jan 09, 2026 at 08:14:05AM +0530, Nilawar, Badal wrote:
> > > > > On 20-12-2025 13:06, Raag Jadav wrote:
> > > > > > We've been historically ignoring GT resume failure. Since the
> > > > > > function
> > > > > > can return error, handle it properly.
> > > > > > 
> > > > > > v2: Bring up display before bailing (Matt Roper, Rodrigo)
> > > > > > 
> > > > > > Signed-off-by: Raag Jadav <raag.jadav@intel.com>
> > > > > > ---
> > > > > >    drivers/gpu/drm/xe/xe_pm.c | 26 ++++++++++++++++++++++----
> > > > > >    1 file changed, 22 insertions(+), 4 deletions(-)
> > > > > > 
> > > > > > diff --git a/drivers/gpu/drm/xe/xe_pm.c
> > > > > > b/drivers/gpu/drm/xe/xe_pm.c
> > > > > > index 4390ba69610d..559cf5490ac0 100644
> > > > > > --- a/drivers/gpu/drm/xe/xe_pm.c
> > > > > > +++ b/drivers/gpu/drm/xe/xe_pm.c
> > > > > > @@ -260,10 +260,19 @@ int xe_pm_resume(struct xe_device *xe)
> > > > > >    	xe_irq_resume(xe);
> > > > > > -	for_each_gt(gt, xe, id)
> > > > > > -		xe_gt_resume(gt);
> > > > > > +	for_each_gt(gt, xe, id) {
> > > > > > +		err = xe_gt_resume(gt);
> > > > > > +		if (err)
> > > > > > +			break;
> > > > > GT and SAMedia are different entities (even if both are treated
> > > > > as GTs in
> > > > > software), should we not continue attempting to resume the
> > > > > remaining GT even
> > > > > if resuming first one fails.
> > > > My limited understanding is that GUI needs render engine, which
> > > > the user
> > > > won't be getting back either way.
> > > 
> > > Ok, but how about multi-tile platform like PVC which doesn't have
> > > Render
> > > engine and Media?
> > 
> > I'm assuming we don't need display to be able to debug them?
> > 
> > Regardless, a semi-working hardware would cause even more problems
> > than
> > solve IMHO but I'll leave to the experts to decide.
> 
> What I liked in this Raag's version is that this is the most generic
> best-effort attempt we can get to still bring some display for some
> debug or information instead of only a blank screen or only keep
> moving as if nothing had happened.
> 
> A more complete version would be checks per platform and per type of
> GT etc. We can still attempt to try this as a follow up.
> 
> For server platforms such as PVC it is a bogus discussion. System
> Suspend is not a valid case there. Period.
> 
> For some discrete like BMG where render is there in the main GT,
> but display is fused off, we can simply return the error.

Display helpers already handle fused off cases.

	if (!xe->info.probe_display)
		return;

> For the case where main GT failed, we still can try to bring display
> up and get some output like Raag's attempt.
> 
> For the case where media GT failed we could proceed with the resume,
> but kind of a partial wedge of the media gt...

We can try to introduce some kind of per-gt wedging but that's pretty
much a driver wide redesign.

Raag

> > > > > > +	}
> > > > > > +	/*
> > > > > > +	 * Try to bring up display before bailing from GT
> > > > > > resume failure,
> > > > > > +	 * so we don't leave the user clueless with a blank
> > > > > > screen.
> > > > > > +	 */
> > > > > >    	xe_display_pm_resume(xe);
> > > > > > +	if (err)
> > > > > > +		goto err;
> > > > > >    	err = xe_bo_restore_late(xe);
> > > > > >    	if (err)
> > > > > > @@ -656,10 +665,19 @@ int xe_pm_runtime_resume(struct
> > > > > > xe_device *xe)
> > > > > >    	xe_irq_resume(xe);
> > > > > > -	for_each_gt(gt, xe, id)
> > > > > > -		xe->d3cold.allowed ? xe_gt_resume(gt) :
> > > > > > xe_gt_runtime_resume(gt);
> > > > > > +	for_each_gt(gt, xe, id) {
> > > > > > +		err = xe->d3cold.allowed ? xe_gt_resume(gt)
> > > > > > : xe_gt_runtime_resume(gt);
> > > > > > +		if (err)
> > > > > > +			break;
> > > > > > +	}
> > > > > > +	/*
> > > > > > +	 * Try to bring up display before bailing from GT
> > > > > > resume failure,
> > > > > > +	 * so we don't leave the user clueless with a blank
> > > > > > screen.
> > > > > > +	 */
> > > > > >    	xe_display_pm_runtime_resume(xe);
> > > > > > +	if (err)
> > > > > > +		goto out;
> > > > > >    	if (xe->d3cold.allowed) {
> > > > > >    		err = xe_bo_restore_late(xe);