[RFC PATCH] drm/panel: Make backlight attachment lazy

[Daniel Vetter]

On Wed, Dec 09, 2020 at 02:28:18PM -0600, Bjorn Andersson wrote:
> On Tue 08 Dec 17:52 CST 2020, Daniel Vetter wrote:
> 
> > On Tue, Dec 08, 2020 at 04:02:16PM -0600, Bjorn Andersson wrote:
> > > On Tue 08 Dec 06:47 CST 2020, Thierry Reding wrote:
> > > 
> > > > On Mon, Dec 07, 2020 at 10:44:46PM -0600, Bjorn Andersson wrote:
> > > > > Some bridge chips, such as the TI SN65DSI86 DSI/eDP bridge, provides
> > > > > means of generating a PWM signal for backlight control of the attached
> > > > > panel. The provided PWM chip is typically controlled by the
> > > > > pwm-backlight driver, which if tied to the panel will provide DPMS.
> > > > > 
> > > > > But with the current implementation the panel will refuse to probe
> > > > > because the bridge driver has yet to probe and register the PWM chip,
> > > > > and the bridge driver will refuse to probe because it's unable to find
> > > > > the panel.
> > > > 
> > > > What you're describing is basically a circular dependency. Can't we get
> > > > rid of that in some other way? Why exactly does the bridge driver refuse
> > > > to probe if the panel can't be found?
> > > > 
> > > > In other words, I see how the bridge would /use/ the panel in that it
> > > > forward a video stream to it. But how does the panel /use/ the bridge?
> > > > 
> > > 
> > > Yes, this is indeed a circular dependency between the components.
> > > 
> > > The involved parts are:
> > > * the bridge driver that implements the PWM chip probe defers on
> > >   drm_of_find_panel_or_bridge() failing to find the panel.
> > > * the pwm-backlight driver that consumes the PWM channel probe defer
> > >   because the pwm_chip was not registered by the bridge.
> > > * the panel that uses the backlight for DPMS purposes probe defer
> > >   because drm_panel_of_backlight() fails to find the pwm-backlight.
> > > 
> > > I looked at means of postponing drm_of_find_panel_or_bridge() to
> > > drm_bridge_funcs->attach(), but at that time "deferral" would be fatal.
> > > I looked at registering the pwm_chip earlier, but that would depend on a
> > > guarantee of the pwm-backlight and panel driver to probe concurrently.
> > > And the current solution of not tying the backlight to the panel means
> > > that when userspace decides to DPMS the display the backlight stays on.
> > > 
> > > 
> > > The proposed solution (hack?) means that DPMS operations happening
> > > before the pwm-backlight has probed will be missed, so it's not perfect.
> > > It does however allow the backlight on my laptop to turn off, which is a
> > > big improvement.
> > > 
> > > But I'm certainly welcome to suggestions.
> > 
> > Entirely hand-waving, why doesn't the following work:
> > 
> > 1. driver for the platform device which is the bridge loads
> > 2. that platform driver registers the pwm
> > 3. it registers some magic for later on (more below)
> > 4. panel driver has deferred loading until step 2 happened
> > 5. panel driver registers drm_panel
> > 6. the magic from step 3 picks up (after having been deferred for a few
> > times probably) grabs the panel, and sets up the actual drm_bridge driver
> > 
> > Everyone happy, or not? From the description it looks like the problem
> > that the pwm that we need for the backlight is tied to the same driver as
> > the drm_bridge, and always torn down too if the drm_bridge setup fails
> > somehow for a reason. And that reason is the circular dependency this
> > creates.
> > 
> > Now for the magic in step 3, there's options:
> > - change DT to split out that pwm as a separate platform_device, that way
> >   bridge and panel can load indepedently (hopefully)
> > 
> 
> This is an i2c device, so describing it multiple times would mean we
> have multiple devices with the same address...
> 
> > - convert bridge to a multi-function device (mfd), essentially a way to
> >   instantiate more devices with their drivers at runtime. Then the actual
> >   pwm and drm_bridge parts of your bridge driver bind against those
> >   sub-functions, and can defer indepedently
> > 
> 
> But, this sounds reasonable and would rely on the existing probe
> deferral logic and if there's ever any improvements in this area we
> would directly benefit from it.
> 
> > - we could create a callback/wait function for "pls wait for any panel to
> >   show up". Then your bridge driver could launch a work_struct with that
> >   wait function, which will do the bridge setup once the panel has shown
> >   up. The pwm will be registered right away. It's essentially hand-rolling
> >   EPROBE_DEFERRED for work_struct in drm/panel. Maybe we might even have
> >   that exported from the driver core, e.g.
> > 
> > register_bridge_fn(struct work *)
> > {
> > 	do_wait_probe_defer();
> > 	panel = drm_of_find_panel_or_bridge();
> > 	if (!panel) {
> > 		schedule_work(); /* want to restart the work so it can be stopped on driver unload */
> > 		return;
> > 	}
> > 
> > 	/* we have the panel now, register drm_bridge */
> > }
> > 
> > - cobble something together with component.c, but that's more for
> >   collecting unrelated struct device into a logical one than splitting it
> >   up more.
> > 
> > tldr; I think you can split this loop here at the bridge by untangling the
> > pwm from the drm_bridge part sufficiently.
> 
> Yes, it seems like a reasonable path forward. But I wanted some input
> before refactoring the whole thing.

Yeah it's unfortunately a bit of work. But I think it's the proper
approach since EPROBE_DEFERRED is fundamentally linked to struct device
and bound drivers. So we do need a struct device for every part in our
dependency graph to make sure we can resolve the dependencies all
correctly with reprobing.
-Daniel
-- 
Daniel Vetter
Software Engineer, Intel Corporation
http://blog.ffwll.ch