[PATCH 3/3 v3] drm: bridge/dw-hdmi: Move edid reading to .detect() callback

[Daniel Vetter]

On Fri, Aug 05, 2016 at 12:01:25AM +0100, Russell King - ARM Linux wrote:
> On Thu, Aug 04, 2016 at 06:13:18PM +0100, Jose Abreu wrote:
> > Hi Russell,
> > 
> > So, I didn't use framebuffer console but used X instead and it is
> > working as it should. I think we can drop this patch. I am now
> > making interoperability with DVI and I am facing the following
> > scenario:
> >     - I start the driver
> >     - An EDID is sent which tells the driver that HDMI is NOT
> > supported;
> >     - The driver configures itself to a DVI mode;
> > 
> > Until this point everything is working as it should. But:
> > 
> >     - Now I send an EDID which tells the driver that HDMI is
> > supported;
> >     - As the EDID has the same preferred mode the user will not
> > reconfigure the mode and there will be no change to HDMI mode.
> 
> The EDID should still be read, but as you say, userspace doesn't take
> any action because it sees that the mode parameters are still the same,
> as you have identified.
> 
> > The missing change to HDMI mode will cause the test to fail. The
> > workaround that I am using is to reconfigure to another video
> > mode and then configure to the preferred one but I think this
> > could be fixed in the driver, right?
> 
> This one is extremely awkward - to fix it, we would need to check to
> see whether we reconfigure the hardware each time we read the EDID,
> and I don't think that's a particularly nice thing to do.
> 
> I'd like to hear what other DRM developers think about this issue.

I pondored whether we should have a counter on each connector for probe
state, which helpers would increment whenever something changes, like:
- connector_status (as tracked by probe helpers)
- anything in the edid changes (when setting it
  drm_mode_connector_update_edid_property)
- other changes (like sink state changes in dpcd or whatever)

That way userspace would be able to reliably spot such changes and do a
new modeset.

The other side of this is making sure that the kernel doesn't optimize
away a modeset in this case. The problem is that current helpers (both
legacy and atomic) only look at the output routing and the display mode
and nothing else to decide whether a full modeset is required. At least
with atomic that's easy to fix:
- When the connector state changes we simply need to set the
  connectors_changed boolean in the corresponding CRTC state.
- We can use the probe lifecycle counter from above to do that: If it's
  changed in atomic_check, then we set connectors_changed to force a full
  modeset.

I guess with just the kernel changes (and not even upgrading userspace)
we'd probably fix many cases already.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch