Re: [PATCH RESEND v2 08/12] dt-bindings: add binding for generic eDP panel

[Thierry Reding <thierry.reding-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>]

[-- Attachment #1: Type: text/plain, Size: 5586 bytes --]

On Mon, Feb 04, 2019 at 10:40:12AM +0100, Daniel Vetter wrote:
> On Mon, Feb 04, 2019 at 09:23:59AM +0100, Thierry Reding wrote:
> > On Mon, Feb 04, 2019 at 12:13:55AM -0800, Vasily Khoruzhick wrote:
> > > On Sun, Feb 3, 2019 at 11:43 PM Thierry Reding <thierry.reding-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org> wrote:
> > > >
> > > > On Sun, Feb 03, 2019 at 10:54:57AM -0800, Vasily Khoruzhick wrote:
> > > > > eDP panels usually have EDID EEPROM, so there's no need to define panel
> > > > > width/height or any modes/timings in dts. But this panel still may have
> > > > > regulator and/or backlight.
> > > > >
> > > > > Signed-off-by: Vasily Khoruzhick <anarsoul-Re5JQEeQqe8AvxtiuMwx3w@public.gmane.org>
> > > > > ---
> > > > >  .../devicetree/bindings/display/panel/panel-edp.txt        | 7 +++++++
> > > > >  1 file changed, 7 insertions(+)
> > > > >  create mode 100644 Documentation/devicetree/bindings/display/panel/panel-edp.txt
> > > >
> > > > Please don't try to make panels look more generic than they really are.
> > > > You're going to have to provide a compatible string for your device that
> > > > is more specific than "panel-edp". You claim that you don't need any
> > > > extra information that is panel specific, but you don't know that now.
> > > > We have in the past thought that we didn't need things like prepare
> > > > delay, but then we ran into situations where we did need them.
> > > >
> > > > Just do what everybody else does. Provide a specific compatible string
> > > > and match on that in the panel-simple driver. Even if you can read all
> > > > the video timings from an EDID EEPROM, you can still provide a mode in
> > > > the panel descriptor to serve as a fallback if for example the EEPROM
> > > > is faulty on some device.
> > > 
> > > Pinebook used several 768p panels that have slightly different timings
> > > and recent batch uses 1080p panel.
> > > 
> > > What panel descriptor should I use as fallback?
> > 
> > You don't use panel descriptors as fallback. The simple-panel driver
> > will bind to a panel device and use the corresponding descriptor. If
> > your device tree contains the correct information, the descriptor is
> > correct for the panel you have.
> > 
> > In other words you need to ensure that you have the correct panel in
> > device tree for the board that you're using. This is exactly the same
> > thing as for other devices.
> > 
> > One way to to this is to have separate device trees for each variant
> > of the board that you want to support. Another variant may be to have
> > a common device tree and then have some early firmware update the DTB
> > with the correct panel information.
> 
> This would defeat the point of edp, which is to standardize the mess of
> panels (at least somewhat) and avoid having to change the DT/ACPI
> tables/firmware for every board you ship. Also, we do have DP quirking
> infrastructure already (using the OUI), I think if there's something that
> doesn't work then we should quirk it there.

The problem is that while the attempt may have been to standardize, it
failed. It doesn't take into account any of the details such as timing
between things like powering up the display and enabling the backlight
or similar. I don't know how you'd want to "quirk" those kinds of
requirements because they are highly panel specific.

> What does make sense though imo is if we try not to stuff the edp panel
> into panel-simple, because it's anything like a simple dumb panel. There's
> also some integration awkwardness since with this panel you need to do dp
> aux/i2c transactions to get at the information (edid alone isn't good
> enough for edp), and I'm not sure how exactly that's supposed to be
> instantiated. Maybe a special function to instantiate an edp panel, which
> takes both a DT node and the dp_aux controller would be much better,
> instead of trying to auto-match against a DT compatible string and load a
> panel driver which is almost all fake.
> 
> Or we teach dp_aux to register itself and somehow teach panel-edp how it
> can get hold of the dp_aux channel it needs.

We already do that. drm_dp_aux registers as an I2C adapter that can be
used to read EDID EEPROMs using I2C-over-AUX transactions. We already
use that on some platforms.

Also note that simple-panel already supports getting video timings from
EDID. If a DDC link is present in DT, the driver will load the modes
from EDID and use them.

> But fake mode in panel-simple sounds like the wrong approach. At least
> from our experience with i915 (and I think other drivers supporting edp),
> the standardization of panels for basic stuff at least worked.
> Self-refresh seems an entirely different story unfortunately ... but again
> quirking that for DP stuff should be done using the OUI I think.

I can imagine that on x86 platforms OEMs can also easily hide some of
the necessary quirks in ACPI. None of that exists on ARM or other
architectures that solely rely on device tree to describe the device.
Since there's no way of specifying power sequences in DT (there have
been attempts in the past to make that work, but they failed miserably)
all we can do is match on a compatible string and encode the necessary
logic in the driver. Luckily we haven't run into anything too fancy up
to now and we've been able to make things work mostly with just a couple
of fairly generic parameters.

Also note that initially people thought this wasn't going to be
necessary and we only added the various "delay" parameters later on
because it did turn out that not all panels are created equal.

Thierry