[PATCH 0/4 v5] Support bridge timings

[Laurent Pinchart]

Hi Daniel,

On Friday, 15 December 2017 17:54:15 EET Daniel Vetter wrote:
> On Fri, Dec 15, 2017 at 01:30:24PM +0100, Linus Walleij wrote:
> > On Fri, Dec 15, 2017 at 1:10 PM, Linus Walleij <linus.walleij at linaro.org> 
wrote:
> >> - The connector is apparently not the right abstraction to carry
> >>   the detailed timings specification between DRI drivers and bridge
> >>   drivers.
> >> 
> >> - Instead put detailed timing data into the bridge itself as an
> >>   optional information pointer.
> > 
> > Notice that this is just my fumbling attempts to deal with the situation.
> > 
> > Laurent made me understand what the actual technical problem was,
> > how come my pixels were flickering.
> > 
> > Both Laurent and DVetter mentioned that we may need to convey
> > information between the bridge and the display engine in some
> > way.
> > 
> > Alternatively I could go and hack on adding this to e.g. drm_display_info
> > which was used in the previous patch sets by setting the negede flag
> > in bus_formats.
> > 
> > I don't know. struct drm_display_info is getting a bit heavy as
> > container of misc settings related to "some kind of display".
> > The bridge isn't even a display itself, that is on the other side
> > of it. So using the connector and treating a bridge as "some kind
> > of display" seems wrong too.
> > 
> > Is there a third way?
> 
> If you don't plan to nest bridges too deeply, there is. Atm we have 2
> modes in drm_crtc_state:
> 
> - mode, which is what userspace requested, and what it expects logically
>   to be the actual real thing. I.e. timing, resolution and all that that
>   userspace can observe (through plane positioning and vblank timestamps)
>   should match this mode. For external screens this should also match
>   what's physically going over the cable.
> 
> - adjusted_mode, which is something entirely undefined and to be used by
>   drivers internally. Most drivers use it as the thing that's actually
>   transported between the CRTC and the encoder.
> 
> There's a few common reasons for adjusted mode to be different:
> - integrated panel, and your CRTC has a scaler. In that case the
>   userspace-requested mode is what you feed into into the scaler, and the
>   adjusted mode is what comes out of your scaler and then goes down the
>   wire to the panel.
> 
> - your encoder is funky, and e.g. transcodes to the output mode itself,
>   but expects that you program the input mode always the same. Usual
>   reasons for this are transcoders that always want non-interlaced mode
>   (and do the interlacing themselves), if the transcoder has some scaler
>   itself (some TV-out transcoders had that), or if it has a strict
>   expectation about signalling edges and stuff (and then transcodes the
>   signal again). DACs are common doing that.
> 
> Anyway, sounds like your bridge is of the 2nd kind, so all you have to do
> is
> - in your bridge->mode_fixup function, adjust the adjusted_mode as needed
> - in your pl111 driver, program the adjusted mode, not the originally
>   requested mode
> 
> adjusted mode is set to be a copy of the requested mode by atomic helpers,
> so this should keep working as-is on any other bridge driver.

I don't think that's the right fix.

The problem here is that the display engine has to output data in a way that 
doesn't violate the DAC setup and hold times. Depending on the display engine, 
you can just select the output clock edge, or adjust the phase of the data 
compared to the pixel clock by a fraction of a clock cycle (1/4 is common, 
I've seen smaller steps too). Note that selecting the opposite clock edge is 
simple a 1/2 clock cycle delay.

The delay has to be computed based on the receiver's setup and hold times, but 
also take into account other components on the board (such as buffer or 
voltage shifters, or even inverters) and the PCB delay itself. This 
computation doesn't belong to the bridge driver, especially given than its 
goal is to compute a delay that depends on the display engine's capabilities 
(inverting the clock vs. smaller step delays for instance). For this reason I 
think the bridge driver should expose its timing parameters, and the display 
engine should then decide how to output its data accordingly.

> No idea why I didn't tell you this right away, or maybe I'm missing
> something this time around.
> 
> > I'm just a bit lost.
> 
> Once your un-lost, pls review the docs for drm_crtc_state and the various
> mode_fixup functions, to make sure they're clear on how this is supposed
> to work. Might need a new overview DOC: comment that ties it all together.

-- 
Regards,

Laurent Pinchart