[Intel-gfx] [PATCH 3/7] drm/i915/selftests: Relax timeout for error-interrupt reset processing

[Chris Wilson]

Quoting Mika Kuoppala (2020-02-11 15:23:24)
> Chris Wilson <chris at chris-wilson.co.uk> writes:
> > +                             /* Kick the tasklet to process the error */
> > +                             intel_engine_flush_submission(engine);
> 
> This pattern of usage in selftests makes me want to add mb(); to
> intel_engine_flush_submission(), as it does not seem to do it
> implicitly.
> 
> We want to flush submission and observe the effects, thus
> it seems like a good place.

Hmm. From the cpu perspective the memory barrier would be on the other
side in clear_bit_unlock(), and flush_submission does unlock_wait.

But, then, we have to factor in that we have to communicate with an
external client the GPU, so perhaps an explicit memory barrier...

We certainly do perform the memory barrier in order to set the GPU in
motion, but have not relied on them for observing effects (aside from
the CSB ringbuf).

I don't see a strong argument that adding a 'mb/rmb' here will make any
difference, I don't see what we are pairing with from the GPU
perspective. But maybe there is?
-Chris