Re: [PATCH v6 03/12] pinctrl: Add MAX7360 pinctrl driver

["Mathieu Dubois-Briand" <mathieu.dubois-briand@bootlin.com>]

On Wed Apr 9, 2025 at 6:32 PM CEST, Andy Shevchenko wrote:
> On Wed, Apr 09, 2025 at 05:03:02PM +0200, Mathieu Dubois-Briand wrote:
>> On Wed Apr 9, 2025 at 4:55 PM CEST, Mathieu Dubois-Briand wrote:
>> > Add driver for Maxim Integrated MAX7360 pinctrl on the PORT pins. Pins
>> > can be used either for GPIO, PWM or rotary encoder functionalities.
>
> ...
>
> The all the rest of the driver LGTM, but the below.
>
>> > +	device_set_of_node_from_dev(dev, dev->parent);
>> 
>> Ok, so this goes a bit against what I said I was going to do on my
>> previous series, let me explain why. Same reasoning applies for both
>> uses, in PWM and pinctrl drivers.
>> 
>> With my previous experiments, I came to the conclusion that:
>> - Either we should use device_set_of_node_from_dev() as I do here.
>> - Or we should add more subnodes in the device tree binding.
>
>> - Also, copying the fwnode with device_set_node() was not possible, as
>>   the kernel would then try to apply pinctrl on both the parent and
>>   child device.
>
> Hmm... I need to refresh my memory with the old discussions. Can you point out
> to the problem statement with that approach?
>

I mentioned here briefly in my previous series: https://lore.kernel.org/lkml/D8R4B2PKIWSU.2LWTN50YP7SMX@bootlin.com/

So the issue is, if I copy the parent fwnode using device_set_node(),
the kernel is trying to apply any pinctrl defined on the node with
pinctrl- properties on both the parent and the child node. Of course,
only the first one will succeed, as two devices cannot request the same
pins at the same time.

>> I previously said the second solution was probably the way to go, but I
>> changed my mind for two reasons.
>> 
>> First having more subnodes in the device tree was already rejected in
>> the past in the reviews of the dt-bindings patch. This do makes sense as
>> it would be describing device internals (which should not be made in
>> DT), just to ease one specific software implementation (which should
>> also be avoided). So I believe this change would again be rejected.
>> https://lore.kernel.org/lkml/58c80c2a-2532-4bc5-9c9f-52480b3af52a@kernel.org/
>> 
>> But the the second reason is, doing
>> 'git grep "device_set_of_node_from_dev.*parent"', I found several
>> drivers using device_set_of_node_from_dev() for a similar need. Some of
>> these uses are also for MFD child devices:
>> - gpio-adp5585.c / pwm-adp5585.c,
>> - pwm-ntxec.c,
>> - max77620-regulator.c / max77620_thermal.c.
>> 
>> So, based on this, I believe using device_set_of_node_from_dev() in
>> these two drivers is the way to go.
>
> The problem with this solution is that, It's OF-centric. Which shouldn't be
> done in a new code (and I don't see impediments to avoid it). Yes, it does
> the right thing for the case, but only on OF systems. Note, fwnode is a list
> of maximum of two entries (yeah, designed like that right now), can you utilise
> that somehow?

Looking at MFD code, I believe ACPI MFD child devices already get the
parent fwnode, except if a fwnode exists for them.

https://elixir.bootlin.com/linux/v6.13.7/source/drivers/mfd/mfd-core.c#L90

Thanks for your review.

-- 
Mathieu Dubois-Briand, Bootlin
Embedded Linux and Kernel engineering
https://bootlin.com