[PATCH] drm/imx: parallel-display: Adjust bus_flags and bus_format handling

[Laurent Pinchart]

Hi Boris,

On Mon, Mar 09, 2020 at 09:22:18PM +0100, Boris Brezillon wrote:
> On Mon, 9 Mar 2020 21:59:26 +0200 Laurent Pinchart wrote:
> > On Mon, Mar 09, 2020 at 08:55:59PM +0100, Boris Brezillon wrote:
> > > On Mon, 9 Mar 2020 21:23:06 +0200 Laurent Pinchart wrote:  
> > > > On Mon, Mar 09, 2020 at 11:50:59AM +0100, Philipp Zabel wrote:  
> > > > > On Thu, 2019-11-14 at 14:17 +0100, Marek Vasut wrote:    
> > > > > > The bus_flags and bus_format handling logic does not seem to cover
> > > > > > all potential usecases. Specifically, this seems to fail with an
> > > > > > "edt,etm0700g0edh6" display attached to an 24bit display interface,
> > > > > > with interface-pix-fmt = "rgb24" set in DT.    
> > > > > 
> > > > > interface-pix-fmt is a legacy property that was never intended to be
> > > > > used as an override for the panel bus format. The bus flags were
> > > > > supposed to be set from the display-timings node, back when there was no
> > > > > of-graph connected panel at all.
> > > > > 
> > > > > That being said, there isn't really a proper alternative that allows to
> > > > > override the bus format requested by the panel driver in the device tree
> > > > > to account for weird wiring. We could reuse the bus-width endpoint
> > > > > property documented in [1], but that wouldn't completely specify how the
> > > > > RGB components are to be mapped onto the parallel bus.
> > > > > 
> > > > > [1] Documentation/devicetree/bindings/media/video-interfaces.txt    
> > > > 
> > > > Things are funny sometimes, I've run into the exact same problem with a
> > > > different display controller today.
> > > > 
> > > > Shouldn't we use the data-shift property from [1] to specify this ?
> > > > Combined with Boris' bus format negotiation for bridges, I think we
> > > > would have all the components in place to solve this problem properly.  
> > > 
> > > I wonder if we shouldn't take more complex pin mappings into account
> > > now and go directly for a data-mapping property describing those
> > > mappings using a string. This way we'd have a single property that
> > > would work for both fully parallel buses (DPI/RGB) and serial (or
> > > partially parallel) ones (LVDS).  
> > 
> > I'm all for standardization, but I'm not sure data-mapping is the right
> > property, at least with its current definition. It's really meant to
> > describe how individual bits are mapped to the LVDS time slots. I'm fine
> > extending it, but we need to define it clearly. How would you envision
> > it being used in this case ?
> 
> Well, clearly the data-width/data-shift approach does not solve all
> problems: what do you do if the source R pins are connected to the sink
> B pins? Well, the first answer would probably be 'have a serious
> discussion with the HW designer responsible for this insanity' :-), but
> once you've passed this 'WTF' stage, you'll have to find a way to tell
> the source component it should use RGBxyx while the sink should use
> BGRxyx (or vice-versa). This is something you can't extract that from
> those width/shift props though. My suggestion would be to have one
> string per MEDIA_BUS_FMT definition, so we can force things at the DT
> level if we really have to. That's basically what the interface-pix-fmt
> property was doing, except we would standardize the prop and values and
> probably provide helpers so bridge elements don't have to parse this
> prop manually.

I don't think that would work in the general case though. We may want to
use different formats and pick one of them at runtime based on external
information (for instance when the sink can accept both RGB and YUV),
hardcoding formats in DT isn't a good option. We instead need to add
information to DT to specify how lines are connected, and deduce formats
based on that.

-- 
Regards,

Laurent Pinchart