Re: [PATCH 30/37] drm/bridge: add drm_bridge_is_tail() to know whether a bridge completes the pipeline

["Luca Ceresoli" <luca.ceresoli@bootlin.com>]

On Wed Jun 24, 2026 at 3:04 PM CEST, Maxime Ripard wrote:
> On Tue, Jun 09, 2026 at 10:23:24AM +0200, Luca Ceresoli wrote:
>> Hi Maxime,
>>
>> thanks for the review of this long series!
>>
>> On Mon Jun 8, 2026 at 2:34 PM CEST, Maxime Ripard wrote:
>> ...
>> >> --- a/include/drm/drm_bridge.h
>> >> +++ b/include/drm/drm_bridge.h
>> >> @@ -78,6 +78,19 @@ struct drm_bridge_funcs {
>> >>  	int (*attach)(struct drm_bridge *bridge, struct drm_encoder *encoder,
>> >>  		      enum drm_bridge_attach_flags flags);
>> >>
>> >> +	/**
>> >> +	 * @is_tail:
>> >> +	 *
>> >> +	 * Returns true if this is a tail bridge, i.e. it does not need a
>> >> +	 * next bridge to work. E.g. a panel_bridge is a tail bridge, a
>> >> +	 * DSI-to-LVDS bridge is not a tail bridge (no matter whether the
>> >> +	 * next bridge is present or not).
>> >
>> > Why a DSI-to-LDVS bridge isn't a tail bridge? It only needs a panel
>> > next, right?
>>
>> Uhm, good point, but I'd say it can be a tail bridge or not: in the
>> ATTACH_NO_CONNECTOR case it will need a bridge (a panel_bridge most
>> likely), no?
>
> Yeah, I think it cannot (always) be a blanket statement from the driver.
> For drivers that do not support ATTACH_NO_CONNECTOR, then it's always
> going to be a tail, but if the driver supports it, we should use a
> helper because it's going to depend on the DT, basically.
>
>> However this is possibly irrelevant as the whole .is_tail is meant to
>> disappear in v2, see below.
>>
>> >> +	 * The @is_tail callback is optional but it is required if the
>> >> +	 * bridge is part of a pipeline with hot-pluggable components.
>> >> +	 */
>> >> +	bool (*is_tail)(struct drm_bridge *bridge);
>> >> +
>> >
>> > I don't think that's the right way to think about it, if only because
>> > you never really know at the driver level if you're supposed to be last
>> > or not. A DSI-to-LVDS bridge might just as well be chained with an
>> > LVDS-to-eDP bridge, or feed the panel directly without any additional
>> > bridge.
>> >
>> > I *think* that what you're trying to find out here is whether the chain
>> > is complete or not.
>>
>> You nailed it.
>>
>> That was the main point discussed during the Display Next Hackfest 2026
>> (see the report [0]) and we all agreed the .is_tail along with the
>> -EPROBE_DEFER changes (patches 20+35) are not a good approach.
>>
>> This is actually the most crucial aspect of the whole work still not having
>> no well-defined solution.
>
> Ok
>
>> > I think you can get the same information by checking
>> > whether you have a connector for that bridge chain. If you don't, you
>> > know the chain isn't complete, and if you do, it's supposed to be.
>>
>> There's a checken-egg problem here. Knowing whether the pipeline is
>> complete or not is needed to know whether we have to create a
>> connector. See patch 36:
>>
>>   +  if (drm_bridge_connector_pipeline_is_complete(bridge_connector))
>>   +          drm_bridge_connector_add_connector(bridge_connector);
>>
>> So the question is still open, what I need the most right now is comments
>> on the overall architecture of this aspecs. Maybe replying to [0] can be
>> more effective than here.
>>
>> In a nutshell what we need is:
>>
>>  * when a bridge needs a following element (a bridge, or a panel which
>>    however might/should have a panel_bridge), but that element is not
>>    present, instead of deferring the whole card probe the card should be
>>    allowed to probe but without a connector
>>
>>  * when something is added to an incomplete pipeline we need to know
>>    whether the pipeline has become complete (in order to create the
>>    connector)
>>
>> How to implement this is still an open point. I'll welcome proposals in
>> addition to the ones in [0].
>
> Thanks for the notes, I think I largely agree with the discussion.

Good! :)

> Reading through it, I thought it would be nice for a new callback called
> get_next_bridge or something that would return either an error, NULL or a
> (refcounted) pointer to the next bridge in the chain. And have some kind
> of special error (ENODEV?) when the bridge should be there but isn't
> (and thus the chain isn't complete).

I initially preferred exposing the fwnode of the next bridge as discussed
with Dmitry during the Display Next Hackfest, which looks simpler for
driver writers. But then I realized it would be tricyk in cases such as
dual-LVDS output bridges (ti-sn65dsi83.c) which have two output ports in
DT, none of which is mandatory. So I've come to thinking a callback is
better as it should be flexible enough.

Picking the best errno is probably the only aspect needing special
attention now.

Luca

--
Luca Ceresoli, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com