[Intel-gfx] [RFC] tentative fix for drm/i915/gt regression on preempt-rt

[Sebastian Andrzej Siewior]

On 2023-07-04 09:02:07 [+0100], Tvrtko Ursulin wrote:

> Lets make it a two patch series and then see. First patch to see if we can
> really get away without the top level preempt_disable, and then second patch
> to see if we can get away with preemptible short sleeps too.
oki.

> I guess on RT the top priority is consistent scheduling latency and not so
> much potential UI latency in some edge cases? Saying that because if waiting
I would says, yes. If you do RT and you provide some kind of GUI then
you prefer to meet your deadlines for your RT load over some UI latency.

> on the hw reset is made preemptible, _in theory_ it can prolong the reset
> completion (as observed by i915), and so result in more UI glitching than it
> normally would. Could be a theoretical point only because it requires both
> CPU over-subscribe and GPU hangs. It could also easily be that the reset
> path is only one path, and not so interesting one even, which can cause this
> on RT.

I see.

> Regards,
> 
> Tvrtko

Sebastian