Re: [PATCH 05/13] dt-bindings: mfd: add Documentation for Airoha EN7581 SCU

[Krzysztof Kozlowski <krzk@kernel.org>]

On 11/03/2025 20:09, Christian Marangi wrote:
>>
>> This does not explain that binding said "This node defines the System
>> Control Unit".
>>
>> So what are you adding here if not SCU?
>>
> 
> With "This node defines the System Control Unit" I mean, the entire
> register space of the IP block is defined and each child specifically
> define each part of the IP block.

And the binding said - "this is the SCU". Now you claim that the binding
which said "this is the SCU" is not the SCU but a part of it.


> 
>>>
>>> The driver solution for this is syscon and the simple-mfd node
>>> structure.
>>
>> Let's keep driver entirely separate, we don't talk about them and mixing
>> arguments won't make it easier.
>>
> 
> Ok.
> 
>>>
>>> Now the main problem is how to modle this in DT. There are lots of case
>>> where the simple-mfd model is used (like the one proposed) but probably
>>> this is not accepted anymore. But again this should be clearly stated or
>>> we have a chicken-egg problem when other devs implement similar thing and
>>> have to implement simple MFD driver to handle this. (and driver
>>> maintainers say "Use the simple-mfd model like it was already done)
>>
>> simple-mfd has nothing to do here. Describe the hardware - what is the SCU?
>>
>>
> 
> As I said below, SCU is just the name used in Airoha Documentation for
> this IP block. In this register space there are multiple things so it's
> not strictly a clock-controller (as it's currently defined)
> 
> It was proposed as clock-controller previously as we weren't aware this
> IP block was used also for other usage that a strict clock controller.


I am pretty sure everyone was, just drivers were developed for current
needs and then binding was adjusted to match drivers.

> 
>>>
>>> For this specific case (and to give an answer to the clock patch after
>>> this) the problem is that this register space was originally used only
>>> to control the clock and I wasn't aware that it was also used to control
>>> USB. Now that I'm implementing support for it, the disaster happened.
>>>
>>> So In short SCU is lots of thing, both a system-controller, a
>>> clock-controller and even a reset-controller.
>>
>> And you have bindings for that already. Done.
>>
> 
> It's currently defined in DTS as clock-controller, should we change it
> to system-controller to make it more clear?

It does not matter for this discussion about bindings, it is just a
name. You can rename it, sure, although also a bit of a churn since this
device has a "clock controller" as one of its main features. And "system
controller" is also way generic.

BTW, it is frequent (e.g. all Qcom platforms, many Allwiner and
Mediatek) that clock controller is also reset or power domain controllers.


> 
>>>
>>> To make it short, 2 different solution:
>>> 1. We can keep the current node structure of the node-controller and add a
>>> child node for the SSR part (with a dedicated compatible).
>>
>> No, you do not add child nodes just because you want some drivers.
>>
>> What is SSR? How is it a device?
> 
> SSR is the name used in Documentation for the register used to configure
> the Serdes and PCIe port.

Register(s) for configuring something is not a device.

> 
>>
>>> 2. Those property need to be be defined in the clock-controller node?
>>
>> In the SCU node. Do you have only one SCU or more?
> 
> Strictly speaking it's one register space. One clock-controller, one
> reset-controller and one set of SSR registers, and from what I can
> understand it's ALWAYS "One device/compatible for Register space"

No, it's not the reason. You define children if they are distinctive. By
distinction I mean some sort of controllable resources which they need
independently of parent. No resources here, except address space, which
seems superficial choice.



> 
> The simple-mfd pattern can't really work for case like this where in one
> register space there are multiple stuff.

We agreed not to talk about simple-mfd, but answering this: yes, but
because your split is neither necessary nor really accurate.

> 
> Is everything clear now?
> 
> To summarize:
> - no child nodes
> - add additional property for SSR in the SCU .yaml

Yes.

Unless I made mistake here in the judgement, but that will be easier to
spot after seeing entire binding/DTS for this device.

Best regards,
Krzysztof