[Intel-gfx] [PATCH v2] drm/i915: Do a better job at disabling primary plane in the noatomic case.

[Daniel Vetter]

On Mon, Nov 23, 2015 at 10:25:28AM +0100, Maarten Lankhorst wrote:
> Op 13-11-15 om 14:28 schreef Ander Conselvan De Oliveira:
> > On Thu, 2015-11-12 at 14:58 +0100, Maarten Lankhorst wrote:
> >> Op 12-11-15 om 14:37 schreef Ander Conselvan De Oliveira:
> >>> On Wed, 2015-11-11 at 15:36 +0100, Maarten Lankhorst wrote:
> >>>> When disable_noatomic is called plane_mask is not reliable yet,
> >>>> and plane_state->visible = true even after disabling the primary plane.
> >>> So the stale value of plane_state->visible causes a subsequent modeset to
> >>> enable
> >>> the primary again?
> >> Probably not because it would get recalculated in calc_changes, but it should
> >> really be set to false afterwards.
> > So basically I didn't understand how the wrong value of plane_state->visible
> > causes the bug that was mentioned. I think a brief explanation in the commit
> > message would be good.
> Well, apply with git am --scissors. Same patch but mentioning this.
> 
> Is this better?
> ---->8-----
> 
> When disable_noatomic is called plane_mask is not correct yet,
> and plane_state->visible = true is left as true after disabling
> the primary plane.
> 
> Other planes are already disabled as part of crtc sanitization, only the
> primary is left active. But the plane_mask is not updated here. It gets updated
> during fb takeover in modeset_gem_init, or set to the new value on resume.
> 
> This means that to disable the primary plane 1 << drm_plane_index(primary)
> needs to be used.
> 
> Afterwards because the crtc is no longer active it's forbidden to keep
> plane_state->visible set, or a WARN_ON in intel_plane_atomic_calc_changes
> triggers. There are other code points that rely on accurate plane_state->visible
> too, so make sure the bool is cleared.

Oh dear is this a mess :(

As a bugfix this is Reviewed-by: Daniel Vetter <daniel.vetter at ffwll.ch>

But we need to bring a _lot_ more sanity to this in -next. Would probably
be better if we keep the primary plane state all consistent, and then
clear both visible and plane_mask when fb reconstruction fails.

Also, some WARN_ON to bitch about inconsistency between plane_mask and
->visible in the disable code paths would be good I think.
-Daniel

> 
> 
> The other planes are already disabled in intel_sanitize_crtc, so
> they don't have to be handled here.
> 
> Cc: stable at vger.kernel.org #v4.3, v4.2?
> Bugzilla: https://bugs.freedesktop.org/show_bug.cgi?id=92655
> Tested-by: Tomas Mezzadra <tmezzadra at gmail.com>
> Signed-off-by: Maarten Lankhorst <maarten.lankhorst at linux.intel.com>
> ---
>  drivers/gpu/drm/i915/intel_display.c | 4 +++-
>  1 file changed, 3 insertions(+), 1 deletion(-)
> 
> diff --git a/drivers/gpu/drm/i915/intel_display.c b/drivers/gpu/drm/i915/intel_display.c
> index b5f7493213b7..bc3282ab5ed2 100644
> --- a/drivers/gpu/drm/i915/intel_display.c
> +++ b/drivers/gpu/drm/i915/intel_display.c
> @@ -6267,9 +6267,11 @@ static void intel_crtc_disable_noatomic(struct drm_crtc *crtc)
>  		WARN_ON(intel_crtc->unpin_work);
>  
>  		intel_pre_disable_primary(crtc);
> +
> +		intel_crtc_disable_planes(crtc, 1 << drm_plane_index(crtc->primary));
> +		to_intel_plane_state(crtc->primary->state)->visible = false;
>  	}
>  
> -	intel_crtc_disable_planes(crtc, crtc->state->plane_mask);
>  	dev_priv->display.crtc_disable(crtc);
>  	intel_crtc->active = false;
>  	intel_update_watermarks(crtc);
> -- 2.1.0
> 
> _______________________________________________
> Intel-gfx mailing list
> Intel-gfx at lists.freedesktop.org
> http://lists.freedesktop.org/mailman/listinfo/intel-gfx

-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch