[Intel-gfx] [PATCH] drm/i915: Disable outputs during unregister

[Chris Wilson]

Quoting Ville Syrjälä (2020-12-01 16:05:17)
> On Fri, Nov 27, 2020 at 10:05:48PM +0000, Chris Wilson wrote:
> > Switch off the scanout during driver unregister, so we can shutdown the
> > HW immediately for unbind.
> > 
> > Signed-off-by: Chris Wilson <chris at chris-wilson.co.uk>
> > ---
> >  drivers/gpu/drm/i915/i915_drv.c | 1 +
> >  1 file changed, 1 insertion(+)
> > 
> > diff --git a/drivers/gpu/drm/i915/i915_drv.c b/drivers/gpu/drm/i915/i915_drv.c
> > index 320856b665a1..62d188e5cb8d 100644
> > --- a/drivers/gpu/drm/i915/i915_drv.c
> > +++ b/drivers/gpu/drm/i915/i915_drv.c
> > @@ -738,6 +738,7 @@ static void i915_driver_unregister(struct drm_i915_private *dev_priv)
> >        * events.
> >        */
> >       drm_kms_helper_poll_fini(&dev_priv->drm);
> > +     drm_atomic_helper_shutdown(&dev_priv->drm);
> 
> Looks like we already have this in remove(). Is that too late?

For the operations we do during unbind, yes.

For the core_hotplug/rebind dance, we have to reset the GPU while we
still have runtime-pm operational and have pushed the reset to
unregister (from experimentation that's as late as we can put it where
the GPU works after rebinding and we don't corrupt the system on unbind,
with the current hooks). You can guess how well gen3 likes that.

But I don't think the right answer is to skip the reset for gen3.
Suppose we enable context support for gen3, then the reset would be
required as well, and so we would still need the whole display
shenanigans to turn it off. Moving the modeset to turn the display off
to the end of userspace seems reasonable.
-Chris