[igt-dev] [PATCH igt] lib: Avoid rcu_barrier() for gem_quiescent_gpu()

[Chris Wilson]

Quoting Antonio Argenziano (2018-02-01 17:23:45)
> 
> 
> On 31/01/18 01:35, Chris Wilson wrote:
> > Using DROP_FREED invokes an rcu_barrier() to ensure that all RCU
> > protected objects are freed before it returns. rcu_barrier() is a very
> > high latency operation (~50ms at best) which is only required under
> > specific circumstances such as when we need a stable object count. By
> 
> Do we expect those specific circumstances to be of interest in our 
> testing? Or better, do we need a gem_quiescent_gpu that waits on all RCU 
> protected objects?

Yes/no. It's really nice to be able to say at the start of the test
there is no leftover state whatsoever from any previous test. To be
honest, I think that is worth the cost, so just read this as shock that
I had forgotten how expensive DROP_FREED was.
-Chris