[Intel-gfx] [PATCH 2/2] drm/hdmi: Allow HDMI infoframe without VIC or S3D

[Ville Syrjälä]

On Tue, Jul 04, 2017 at 05:09:36PM +0200, Andrzej Hajda wrote:
> On 04.07.2017 16:25, Ville Syrjälä wrote:
> > On Tue, Jul 04, 2017 at 03:58:02PM +0200, Andrzej Hajda wrote:
> >> On 04.07.2017 14:44, Ville Syrjälä wrote:
> >>> On Tue, Jul 04, 2017 at 01:56:07PM +0200, Andrzej Hajda wrote:
> >>>> On 03.07.2017 21:19, ville.syrjala at linux.intel.com wrote:
> >>>>> From: Ville Syrjälä <ville.syrjala at linux.intel.com>
> >>>>>
> >>>>> Appedix F of HDMI 2.0 says that some HDMI sink may fail to switch from
> >>>>> 3D to 2D mode in a timely fashion if the source simply stops sending the
> >>>>> HDMI infoframe. The suggested workaround is to keep sending the
> >>>>> infoframe even when strictly not necessary (ie. no VIC and no S3D).
> >>>>> HDMI 1.4 does allow for this behaviour, stating that sending the
> >>>>> infoframe is optional in this case.
> >>>> My impression from the specs is that it should be done only after
> >>>> switching from 3d to 2d mode.
> >>>> In such case we just need to remember previous mode, if it was 3d, empty
> >>>> VSIF infoframe should be still generated for 2seconds.
> >>>> No need to do guesses from EDID.
> >>>> Am I right, or just missing something?
> >>> This code has no idea about any 3D->2D transitions, trying to make it
> >>> do that would just result in a lot of complexity. Much easier to just
> >>> always send the infoframe.
> >> With such approach I see following 'issues':
> >> 0. It does not follow advices from specs.
> > Sure it does. The spec says you should keep sending it for at least two
> > seconds, but it doesn't say that you can't keep sending it for longer,
> > or even when there are no 3D->2D transitions.
> >
> >> 1. It changes behavior of old drivers, probably in harmless way but
> >> still there can be some sinks which will stop working.
> > I think this is a justified risk. If we start to worry too much about
> > every little change we stop making progress altogether. I did minimize
> > the danger by making sure we don't send it for pre-1.4 sinks, and
> > I was almost leaning towards not checking for that. But then I saw some
> > language in the spec which might be interpreted to mean that the source
> > isn't allowed to send unknown infoframe types to the sink, and I figured
> > that maybe it's better to play it safe.
> >
> >> 2. What if EDID does not advertises 3d/4k modes but the sink supports
> >> it, in such case userspace can set 3d mode, but after switch 3d->2d
> >> empty infoframe will not be sent.
> > IMO no point in worrying about broken EDIDs until one is proven to
> > exist.
> 
> I though the whole 'fake edid' interface, possibility to provide custom
> mode, and avoiding connector:mode_valid in such case is enough proof of
> existence of such EDIDs.

It's proof that there are buggy EDIDs in general. It isn't proof of
displays that can do 3D + don't advertize 3D/4k + require the 2D infoframe
to switch out of 3D mode.

> Internet is also full of advices how to force certain mode (also 3d
> mode), even if EDID does not advertise it.

If people insist on shooting themselves in the foot they can.

> 
> >
> >> 3. With this patch connector is required to generate infoframe, but
> >> there are pipelines where infoframe can be generated in non-connector
> >> driver, for example:
> >>     crtc -> hdmi_encoder -> mhl_encoder -> connector
> >> In such case encoder has no access to the connector, of course it can
> >> violate abstraction layers and localize one, but it shows that something
> >> here is probably wrong.
> > Just pass the connector down if needed. We'll need the connector to
> > decide whether to send the new CEA-864-F VICs or not as well. And you
> > could need it (well really the connector state) to figure out the value
> > if some connector properties and whanot.
> >
> > Also I didn't actually run into any cases where the connector is
> > unavailable in the tree.
> 
> Not unavailable, just requires digging across layers.
> 
> >
> >> Maybe another helper drm_hdmi_vendor_infoframe_from_connector will be
> >> enough to solve it.
> >> 4. Video mode provided by user has nothing to do with EDID, why
> >> infoframe generated for that mode should depend on EDID.
> >>
> >> All above 'issues' are not serious ones but suggests that the solution
> >> is not the ideal one.
> >>
> >> On the other side I do not see much complication in 3D->2D transition,
> >> it requires just recording if 3d mode was set before and again
> >> drm_hdmi_vendor_infoframe_from_connector can perform this check.
> > Well, then we'd have to pass in the old and new crtc/connector states
> > etc. We don't want to start maintaining some custom state in the
> > infoframe helpers that bypasses the normal atomic state handling
> > altogether.
> 
> I understand, on the other side the issue the patch tries to solve has
> such nature.
> 
> >
> > And what causes us to stop sending it after the 2 seconds? Is two
> > second the correct amount of time or should we do it longer? etc.
> 
> I though here about simpler approach: to send frame always after 3d mode
> was ever used.

That would just result in inconsistent behaviour depending on whether
you start with the 2D mode from the get go, or if you start with 3D and
later switch to 2D. That kind of inconsistent behaviour just leads to
harder to analyze bugs in my experience.

> Approach with 2s timeout would require more work on drivers side.
> 
> >
> > So this line of thinking leads to a lot of new problems we can avoid
> > by keeping it simple.
> >
> I have pointed out some weakness IMO, but as I said earlier they are
> non-blockers, so if you still think this approach is better it is OK.
> 
> 
> Regards
> 
> Andrzej
> 

-- 
Ville Syrjälä
Intel OTC