[Intel-gfx] [PATCH 3/5] drm/i915: Update CRTC state if connector link status property changed

[Daniel Vetter]

On Fri, Nov 18, 2016 at 09:44:49AM -0800, Manasi Navare wrote:
> On Fri, Nov 18, 2016 at 06:21:21PM +0200, Ville Syrjälä wrote:
> > On Fri, Nov 18, 2016 at 04:35:25PM +0100, Daniel Vetter wrote:
> > > On Fri, Nov 18, 2016 at 05:28:54PM +0200, Ville Syrjälä wrote:
> > > > On Fri, Nov 18, 2016 at 03:18:06PM +0100, Maarten Lankhorst wrote:
> > > > > Op 18-11-16 om 15:11 schreef Ville Syrjälä:
> > > > > > On Fri, Nov 18, 2016 at 02:50:52PM +0100, Maarten Lankhorst wrote:
> > > > > >> Op 18-11-16 om 08:13 schreef Manasi Navare:
> > > > > >>> CRTC state connector_changed needs to be set to true
> > > > > >>> if connector link status property has changed. This will tell the
> > > > > >>> driver to do a complete modeset due to change in connector property.
> > > > > >>>
> > > > > >>> Acked-by: Harry Wentland <harry.wentland at amd.com>
> > > > > >>> Acked-by: Tony Cheng <tony.cheng at amd.com>
> > > > > >>> Cc: dri-devel at lists.freedesktop.org
> > > > > >>> Cc: Jani Nikula <jani.nikula at linux.intel.com>
> > > > > >>> Cc: Daniel Vetter <daniel.vetter at intel.com>
> > > > > >>> Cc: Ville Syrjala <ville.syrjala at linux.intel.com>
> > > > > >>> Signed-off-by: Manasi Navare <manasi.d.navare at intel.com>
> > > > > >>> ---
> > > > > >>>  drivers/gpu/drm/drm_atomic_helper.c | 7 +++++++
> > > > > >>>  1 file changed, 7 insertions(+)
> > > > > >>>
> > > > > >>> diff --git a/drivers/gpu/drm/drm_atomic_helper.c b/drivers/gpu/drm/drm_atomic_helper.c
> > > > > >>> index 0b16587..2125fd1 100644
> > > > > >>> --- a/drivers/gpu/drm/drm_atomic_helper.c
> > > > > >>> +++ b/drivers/gpu/drm/drm_atomic_helper.c
> > > > > >>> @@ -519,6 +519,13 @@ static int handle_conflicting_encoders(struct drm_atomic_state *state,
> > > > > >>>  					       connector_state);
> > > > > >>>  		if (ret)
> > > > > >>>  			return ret;
> > > > > >>> +
> > > > > >>> +		if (connector->state->crtc) {
> > > > > >>> +			crtc_state = drm_atomic_get_existing_crtc_state(state,
> > > > > >>> +									connector->state->crtc);
> > > > > >>> +			if (connector->link_status == DRM_MODE_LINK_STATUS_BAD)
> > > > > >>> +				crtc_state->connectors_changed = true;
> > > > > >>> +		}
> > > > > >>>  	}
> > > > > >>>  
> > > > > >>>  	/*
> > > > > >> This will cause ordinary atomic commits that happen to change connector flags to potentially fail with -EINVAL if ALLOW_MODESET is not set.
> > > > > >> For this reason I'm not sure this flag should be set automatically by the kernel. Could we add add a retrain link property instead, that
> > > > > >> always return 0 when queried, but writing a 1 causing connectors_changed to be set on bad link status?
> > > > > > Or just check for allow_modeset before setting connectors_changed=true here?
> > > > > 
> > > > > I don't think modesets should be done automatically like that, even if ALLOW_MODESET is set a modeset may not be expected by userspace.
> > > > 
> > > > Presumably userspace would want a picture on the screen using any means
> > > > if it said ALLOW_MODESET. So if it can't tolerate the modeset it should
> > > > probably say as much?
> > > 
> > > Yeah, agreed. Also, if the link is bad then we pretty much have to do a
> > > modeset to recover it, otherwise you'll be forever stuck with a bad
> > > screen.
> > > 
> > > What we could try is to gate this of whether userspace touches the mode
> > > property on the corresponding CRTC. I.e. if that's touched (even if it's
> > > the same mode), and a link is bad in one of the connectors in the state
> > > then we do a full modeset to recover.
> > > 
> > > Another option would be to make the link status writeable. Trying to
> > > change it from bad->good would force the modeset. That would be 100% clear
> > > to userspace, not special hacks needed with checking for allow_modeset,
> > > no magic property that auto-changes its value. And 100% backwards compat
> > > because existing userspace should never touch properties it doesn't
> > > understand (except when restoring a mode, and then it should allow a full
> > > modeset). And if someone does try a good->bad transition, we just silently
> > > keep it at good.
> > > 
> > > Definitely need to document this properly in the property docs, no matter
> > > what we decide.
> > 
> > Hmm. I think I kinda like this idea of userspace clear the state back
> > to good explicitly, if it happens with the same prop. So it's like
> > Maarten's retrain_link prop idea, but without having to add the second
> > prop to the mix.
> > 
> > It would also save me from pointing out (for the nth time) that the
> > link status should really be cleared to good during the commit state
> > swap and not at some random point during the commit ;)
> >
> 
> Okay, so change 1 is to make the userspace clear the state back to Good for the property..
> Then Change 2 is to set connector_changed flag in crtc_state to true if this property changed
> from BAD to GOOD. I am not quite how and where to change this to state connector_changed to true.
> Without this the full modeset will not happen and the whole design of retrianing is lost.

To make this work we need a few more bits:

- link-status needs to become a full-blown atomic property. This means we
  need to move link_status into drm_connector_state, mark the property as
  type ATOMIC and wire up the get/set stuff.

- once that's done it's also pretty easy to set crtc->connectors_changed
  from the atomic helpers. You can just compare old and new link_status in
  drm_connector_state, similar to how we compare old/new CRTC in
  drm_connector_state->crtc already.

- Another fallout is that legacy clients will no longer see the
  link-status property. And they won't be able to set it through the
  SETCRTC ioctl, which would kinda defaut the point. I think the best
  solution would be to check for link_status == BAD in
  drm_atomic_helper_set_config, and reset it to good automatically for
  legacy clients.

This way there's no magic auto-switching properties from an atomic pov
(which is a pet peeve of Ville, and I agree with it), legacy clients still
work, we have a clear model of when link-retrain happens (no magic
behavioural changes depending upon ALLOW_MODESET, making me happy) and DP
compliance should also still work.

So ... everyone happy with the above?
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch