[Intel-gfx] [PATCH v2 05/21] drm/i915/gt: Skip TLB invalidations once wedged

Tvrtko Ursulin tvrtko.ursulin at linux.intel.com
Tue Jul 19 07:19:54 UTC 2022


On 18/07/2022 17:06, Mauro Carvalho Chehab wrote:
> On Mon, 18 Jul 2022 14:45:22 +0100
> Tvrtko Ursulin <tvrtko.ursulin at linux.intel.com> wrote:
> 
>> On 14/07/2022 13:06, Mauro Carvalho Chehab wrote:
>>> From: Chris Wilson <chris.p.wilson at intel.com>
>>>
>>> Skip all further TLB invalidations once the device is wedged and
>>> had been reset, as, on such cases, it can no longer process instructions
>>> on the GPU and the user no longer has access to the TLB's in each engine.
>>>
>>> That helps to reduce the performance regression introduced by TLB
>>> invalidate logic.
>>>
>>> Cc: stable at vger.kernel.org
>>> Fixes: 7938d61591d3 ("drm/i915: Flush TLBs before releasing backing store")
>>
>> Is the claim of a performance regression this solved based on a wedged
>> GPU which does not work any more to the extend where mmio tlb
>> invalidation requests keep timing out? If so please clarify in the
>> commit text and then it looks good to me. Even if it is IMO a very
>> borderline situation to declare something a fix.
> 
> Indeed this helps on a borderline situation: if GT is wedged, TLB
> invalidation will timeout, so it makes sense to keep the patch with a
> comment like:
> 
>      drm/i915/gt: Skip TLB invalidations once wedged
>      
>      Skip all further TLB invalidations once the device is wedged and
>      had been reset, as, on such cases, it can no longer process instructions
>      on the GPU and the user no longer has access to the TLB's in each engine.
>      
>      So, an attempt to do a TLB cache invalidation will produce a timeout.
>      
>      That helps to reduce the performance regression introduced by TLB
>      invalidate logic.

Yeah that is better but whether bothering stable with it is the 
question. Wedged GPU means constant endless -EIO to userspace so very 
hard to imagine that after a TLB invalidation timeout or two there would 
be further ones. But okay, it's tiny so fine I guess.

Regards,

Tvrtko


More information about the Intel-gfx mailing list