[Intel-gfx] [PATCH] drm/i915: Suppress spurious EIO when moving away from the GPU domain

Chris Wilson chris at chris-wilson.co.uk
Fri May 24 11:36:29 CEST 2013


On Fri, May 24, 2013 at 11:03:06AM +0200, Daniel Vetter wrote:
> On Wed, May 1, 2013 at 6:23 PM, Daniel Vetter <daniel at ffwll.ch> wrote:
> > The atomic_set(WEDGED) is imo very dangerous, since it'll wreak havoc
> > with our accounting. But in general I really prefer an angry user with
> > a stuck-process backtrace than trying to paper over our own
> > shortcomings with hacks. And at least for the gpu hang vs.
> > pageflip/set_base we should now be fairly well-covered with tests. And
> > I haven't yet seen a report indicating that there's another issue
> > looming ...
> >
> > I guess an obnoxious WARN with a return 0; would also work for the
> > time out case. But I prefer to just ditch the timeout.
> 
> Poke ... I still think setting wedged here is a bit risky.

The alternative is that every ioctl then takes 10s. But if we do set
wedged and it recovers, it is reset to 0. If the reset fails, it is also
set to wedged. It's a fugly thing to do, but at that level of paranoia,
I think it is the right thing to do.

> And maybe
> add a comment why wait_for_error eats the -EIO (and why
> intel_ring_begin does not eat the -EIO)?

Sure. The patch today is slightly different anyway...
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre



More information about the Intel-gfx mailing list