[Intel-gfx] [PATCH] drm/i915: Mark hardware context support optional

[Chris Wilson]

On Tue, 25 Sep 2012 13:35:31 -0700, Ben Widawsky <ben at bwidawsk.net> wrote:
> On Tue, 25 Sep 2012 20:41:45 +0100
> Chris Wilson <chris at chris-wilson.co.uk> wrote:
> 
> > On Tue, 25 Sep 2012 12:04:14 -0700, Ben Widawsky <ben at bwidawsk.net>
> > wrote:
> > > On Tue, 25 Sep 2012 14:53:37 +0100
> > > Chris Wilson <chris at chris-wilson.co.uk> wrote:
> > > 
> > > > "Enable hardware context support for userspace (default:
> > > > disabled))");
> > > 
> > > You've found one platform this doesn't work on, and a bunch of
> > > features rely on this, and yet we default to disabled? That seems a
> > > bit harsh to me.
> > 
> > Exactly, it's meant to be harsh. In the limited testing that hw
> > contexts have been exposed to, we have a number of hangs for which
> > they are implicated. Therefore they are not safe to enable yet...
> > Unless you can prove otherwise. :-p
> > -Chris
> > 
> 
> The reason I've always resisted a module parameter is that rc6 and
> contexts are tied so very closely together. We've had a number of
> issues around rc6 already, I do not want contexts to be conflated with
> those issues. It's interesting if rc6=0 and failures still occur. I've
> yet to hear of such an issue, but I'd like to know if that's the case
> here.

Disabling rc6 but keeping hw contexts works but regresses performance in
old school games.

rc6=0 hw_contexts=0 -> works
rc6=0 hw_contexts=1 -> works
rc6=1 hw_contexts=0 -> works (current default up to 3.6)
rc6=1 hw_contexts=1 -> hard hang with gl (new default in 3.6)
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre