[PATCH] drm/i915: Mark forced GPU wedging unrecoverable on driver unregister

[Janusz Krzysztofik]

(off-list)

On Tue, 2020-10-20 at 16:27 +0100, Chris Wilson wrote:
> Quoting Janusz Krzysztofik (2020-10-20 16:21:09)
> > On Tue, 2020-10-20 at 13:55 +0100, Chris Wilson wrote:
> > > Quoting Janusz Krzysztofik (2020-10-20 13:49:35)
> > > > Signed-off-by: Janusz Krzysztofik <janusz.krzysztofik at linux.intel.com>
> > > > ---
> > > >  drivers/gpu/drm/i915/gt/intel_gt.c | 2 +-
> > > >  1 file changed, 1 insertion(+), 1 deletion(-)
> > > > 
> > > > diff --git a/drivers/gpu/drm/i915/gt/intel_gt.c b/drivers/gpu/drm/i915/gt/intel_gt.c
> > > > index 44f1d51e5ae52..6644d5485a1b6 100644
> > > > --- a/drivers/gpu/drm/i915/gt/intel_gt.c
> > > > +++ b/drivers/gpu/drm/i915/gt/intel_gt.c
> > > > @@ -623,7 +623,7 @@ void intel_gt_driver_unregister(struct intel_gt *gt)
> > > >          * all in-flight requests so that we can quickly unbind the active
> > > >          * resources.
> > > >          */
> > > > -       intel_gt_set_wedged(gt);
> > > > +       intel_gt_set_wedged_on_fini(gt);
> > > 
> > > Currently a few moments later at intel_gt_driver_remove.
> > 
> > Hi Chris,
> > 
> > I know it is there.  Any reason for not moving that from
> > intel_gt_driver_unregister() to intel_gt_driver_unregister()?
> 
> None, but also no reason to move it either. So if you can write a
> compelling case for one or the other, I'm easily convinced. I was
> uncertain what you were looking for here.

I'm looking for a chance for pre-Si to unbind without a taint for CI. 
That's what is now happening apparently in i915_gem_driver_remove()
calling i915_gem_suspend_late()->intel_gt_suspend_late()->gt_sanitize-
>intel_gt_unset_wedged(), still before intel_gt_driver_remove().  That
trybot submission was just a dirty check for potential regressions,  my
intent was to submit the change internally.

Thanks,
Janusz

>  Shutting down iommu remains
> the biggest fail, although that intel_iommu isn't squawking over misuse
> is concerning as well. core_hotplug lateclose also demonstrates a nice
> uaf we have with our exported fences.
> -Chris