[PATCH] drm/xe/xe_pmu: Validate gt in event supported

[Matt Roper]

On Tue, Jul 01, 2025 at 04:09:44PM -0500, Lucas De Marchi wrote:
> On Tue, Jul 01, 2025 at 01:31:28PM -0700, Matt Roper wrote:
> > On Mon, Jun 30, 2025 at 05:28:53PM -0500, Lucas De Marchi wrote:
> > > On Mon, Jun 30, 2025 at 03:07:41PM +0530, Riana Tauro wrote:
> > > > Validate gt instead of checking gt_id is lesser
> > > > than max gts per tile
> > > 
> > > I prefer the series from Matt Roper that is refactoring this part well.
> > 
> > I think we want this change regardless.  My series is about refactoring
> > to future-proof the behavior in case we wind up with multi-tile+multi-GT
> > platforms in the future.  However Riana's patch here is fixing a real
> > bug in the code that exists for existing platforms like PVC (i.e., the
> > code currently lets "GT1" through as valid even on single-tile variants
> > of PVC).
> 
> I misread this patch and thought it was an impossible condition that was
> already protected during event init and it was added more as an "assert
> we always have this". Now I see this is the call that actually protects
> the event creation. So, yes, this looks the correct thing to do.
> 
> Riana, could you expand the commit message to emphasize this is a real
> issue for platforms like PVC?
> 
> Reviewed-by: Lucas De Marchi <lucas.demarchi at intel.com>
> 
> thanks
> Lucas De Marchi

I did just notice (as Karthik also pointed out on another part of this
thread) that if we apply this patch _without_ my series, then
xe_device_get_gt() could throw a WARN and return something non-NULL if
the GT ID is too high, which doesn't actually help.  But if we apply
this fix after my refactoring, then xe_device_get_gt() will return NULL
on any invalid GT ID and the logic here works as expected.  So I guess
rather than applying this on its own, we could wait and apply it after
my series (and I'd drop the patch #1 that I had added to v4 of my
series).


Matt

-- 
Matt Roper
Graphics Software Engineer
Linux GPU Platform Enablement
Intel Corporation