Re: [PATCH 03/15] drm/i915: Upgrade execbuffer fail after resume failure to EIO

[Chris Wilson <chris@chris-wilson.co.uk>]

On Fri, Aug 08, 2014 at 11:46:07AM +0200, Daniel Vetter wrote:
> On Fri, Aug 08, 2014 at 10:17:10AM +0100, Chris Wilson wrote:
> > On Wed, Aug 06, 2014 at 10:39:16AM +0200, Daniel Vetter wrote:
> > > On Wed, Aug 06, 2014 at 09:12:32AM +0100, Chris Wilson wrote:
> > > > On Wed, Aug 06, 2014 at 09:56:45AM +0200, Daniel Vetter wrote:
> > > > > On Tue, Aug 05, 2014 at 07:51:14AM -0700, Rodrigo Vivi wrote:
> > > > > > From: Chris Wilson <chris@chris-wilson.co.uk>
> > > > > > 
> > > > > > If we try to execute on a known ring, but it has failed to be
> > > > > > initialised correctly, report that the GPU is hung rather than the
> > > > > > command invalid. This leaves us reporting EINVAL only if the user
> > > > > > requests execution on a ring that is not supported by the device.
> > > > > > 
> > > > > > This should prevent UXA from getting stuck in a null render loop after a
> > > > > > failed resume.
> > > > > > 
> > > > > > v2 (Rodrigo): Fix conflict and add VCS2 ring and
> > > > > >    	      s/intel_ring_buffer/intel_engine_cs.
> > > > > > 
> > > > > > Reported-by: Jiri Kosina <jikos@jikos.cz>
> > > > > > References: https://bugs.freedesktop.org/show_bug.cgi?id=76554
> > > > > > Signed-off-by: Chris Wilson <chris@chris-wilson.co.uk>
> > > > > > Signed-off-by: Rodrigo Vivi <rodrigo.vivi@intel.com>
> > > > > 
> > > > > This isn't required any more, see
> > > > > 
> > > > > commit 074c6adaf4e7d1423d373bd5d1afc20b683cb4d0
> > > > > Author: Chris Wilson <chris@chris-wilson.co.uk>
> > > > > Date:   Wed Apr 9 09:19:43 2014 +0100
> > > > > 
> > > > >     drm/i915: Mark device as wedged if we fail to resume
> > > > > 
> > > > > for the alternate merged patch.
> > > > 
> > > > Hmm, there is still a path that ends here, but the example above is
> > > > already fixed as you say.
> > > 
> > > We have the EIO check both in the resume and driver load paths. Which
> > > other path are we missing?
> > 
> > The GPU may be set to wedged, but this check in execbuffer occurs before
> > we check for a wedged GPU.
> 
> But we no longer free the ring structures over suspedn/resume, so at least
> the commit message is outdated.

Went off on an incorrect tangent.
 
> I wonder whether the easier fix wouldn't be to continue ring init if we
> get an -EIO.

Right, that's the problem I remember. But I am not sure if it is really
worth it. It's only to hide log spew in a corner case of a corner case in
UXA.
-Chris

-- 
Chris Wilson, Intel Open Source Technology Centre