[Intel-gfx] Prevent trivial oom from gem_exec_nop/sequential

[Chris Wilson]

Quoting Tvrtko Ursulin (2018-01-16 09:52:10)
> 
> On 15/01/2018 21:24, Chris Wilson wrote:
> > About the third resend of this series that tries to curtail the prolonged
> > reference lifetimes of fences via the signaler thread when the CPUs are
> > saturated by interrupts.
> 
> Happens on CI, regularly, occasionally? In the real-world it shouldnt 
> right? Trying to prioritize, spotted the last patch in the series 
> rewriting breadcrumbs so grumble. :)

There's a very minor performance improvement overall, plus closing an
easy to trigger DoS. That style of DoS pops up everywhere we have
references to requests (with so many fence contexts floating around,
it's very easy to manipulate), so getting a grip on it is very useful.

The gem_exec_nop/sequential one was a little more surprising because the
lifetime there wasn't ostensibly extended, it was just the result of a
huge backlog for the RT thread!
-Chris