[Intel-gfx] [PATCH 08/11] drm/i915/execlists: Force preemption via reset on timeout

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-03-27 09:51:20)
> 
> On 26/03/2018 12:50, Chris Wilson wrote:
> > Install a timer when trying to preempt on behalf of an important
> > context such that if the active context does not honour the preemption
> > request within the desired timeout, then we reset the GPU to allow the
> > important context to run.
> 
> I suggest renaming patch title to "Implement optional preemption delay 
> timeout", or "upper bound", or something, as long as it is not "force 
> preemption". :)
> 
> > (Open: should not the timer be started from receiving the high priority
> > request...)
> 
> If you think receiving as in execbuf I think not - that would be 
> something else and not preempt timeout.

I'm thinking of submit_request -> [insert huge delay] -> ksoftirqd.

Actually I think it's better just to solve the ksoftirqd adding random
delays since it affects us all over the shop.

The tricky part is that I expect any request to improve (i.e. avoid)
ksoftirqd for our use case is likely to end up in "use threaded-irq"
with which we always incur a scheduler delay as we then do not have our
irq_exit hook. I currently hack kernel/softirq.c to always run at least
one loop for HI_SOFTIRQ before deferring to ksoftirqd.
-Chris