Re: [PATCH 2/2] drm/xe/shrinker: hold a stable RPM ref across scan

[Matthew Auld <matthew.auld@intel.com>]

On 26/09/2026 17:56, Karl Mehltretter wrote:
> On Fri, Sep 25, 2026 at 11:55:23AM +0000, Matthew Auld wrote:
>> @@ -102,7 +84,7 @@ static int __xe_shrinker_walk(struct xe_shrinker *shrinker,
>>   
>>   		if (mem_type != XE_PL_SYSTEM && !rpm &&
>>   		    xe_device_is_l2_flush_optimized(xe)) {
>> -			if (!__xe_shrinker_runtime_pm_get(shrinker))
>> +			if (!xe_pm_runtime_get_if_active(xe))
>>   				break;
>>   			rpm = true;
>>   		}
> 
> Could this skip TT reclaim with CONFIG_PM=n on integrated graphics
> version 35 or newer?
> 
> The !CONFIG_PM stub of pm_runtime_get_if_active() returns -EINVAL, so
> xe_pm_runtime_get_if_active() always returns false here. Taking the
> outer synchronous reference cannot change that stub's result.
> 
> Previously, kswapd could take the synchronous fallback in
> __xe_shrinker_runtime_pm_get() when D3cold was unavailable and continue
> into the TT manager (e.g. with CONFIG_ACPI=n).

Interesting. If this is true then I think this is a widespread issue and 
driver was already busted here, since various places use if_active() and 
if_in_use(). In fact just below this call if we trigger xe_bo_shrink(), 
we have the same pattern, I think. If that returns "false" it will skip 
the operation there also.

Assuming CONFIG_PM=n is valid, then maybe something like:

  bool xe_pm_runtime_get_if_active(struct xe_device *xe)
  {
+       if (!IS_ENABLED(CONFIG_PM))
+               return true;
+
         return pm_runtime_get_if_active(xe->drm.dev) > 0;
  }

And then same thing for if_in_use().

> 
> Sashiko dismissed the skipped-TT concern assuming the outer resume
> makes the active check succeed. Might that dismissal have overlooked
> CONFIG_PM=n?
> https://sashiko.dev/#/patchset/80538
> 
> Thanks,
> Karl