Re: [PATCH v2 1/2] dt-bindings: pinctrl: airoha: Add EN7581 pinctrl controller

On Fri, Aug 23, 2024 at 05:14:30PM +0100, Conor Dooley wrote:

quoted

On Thu, Aug 22, 2024 at 10:50:52PM +0200, Benjamin Larsson wrote:

quoted

On 22/08/2024 18:06, Conor Dooley wrote:


Hi.

quoted

before looking at v1:
I would really like to see an explanation for why this is a correct
model of the hardware as part of the commit message. To me this screams
syscon/MFD and instead of describing this as a child of a syscon and
using regmap to access it you're doing whatever this is...

Can you post a link to a good example dts that uses syscon/MFD ?

It is not only pinctrl, pwm and gpio that are entangled in each other. A
good example would help with developing a proper implementation.

Off the top of my head, no unfortunately. Maybe Rob or Krzk have a good
example. I would suggest to start by looking at drivers within gpio or
pinctrl that use syscon_to_regmap() where the argument is sourced from
either of_node->parent or dev.parent->of_node (which you use depends on
whether or not you have a child node or not).

I recently had some questions myself for Rob about child nodes for mfd
devices and when they were suitable to use:
https://lore.kernel.org/all/20240815200003.GA2956351-robh@kernel.org/ (local)

Following Rob's line of thought, I'd kinda expect an mfd driver to create
the devices for gpio and pwm using devm_mfd_add_devices() and the
pinctrl to have a child node.

Just to not get confused and staring to focus on the wrong kind of
API/too complex solution, I would suggest to check the example from
Lorenzo.

The pinctrl/gpio is an entire separate block and is mapped separately.
What is problematic is that chip SCU is a mix and address are not in
order and is required by many devices. (clock, pinctrl, gpio...)

IMHO a mfd is overkill and wouldn't suite the task. MDF still support a
single big region and in our case we need to map 2 different one (gpio
AND chip SCU) (or for clock SCU and chip SCU)

Similar problem is present in many other place and syscon is just for
the task.

Simple proposed solution is:
- chip SCU entirely mapped and we use syscon
- pinctrl mapped and reference chip SCU by phandle
- pwm a child of pinctrl as it's scrambled in the pinctrl mapped regs

Hope this can clear any confusion.

-- 
	Ansuel

To clarify the hw architecture we are discussing about 3 memory regions:
- chip_scu: <0x1fa20000 0x384>
- scu: <0x1fb00020 0x94c>
- gpio: <0x1fbf0200 0xbc>

The memory regions above are used by the following IC blocks:
- clock: chip_scu and scu
- pinctrl (io-muxing/gpio_chip/irq_chip): chip_scu and gpio
- pwm: gpio

clock and pinctrl devices share the chip_scu memory region but they need to
access even other separated memory areas (scu and gpio respectively).
pwm needs to just read/write few gpio registers.
As pointed out in my previous email, we can define the chip_scu block as
syscon node that can be accessed via phandle by clock and pinctrl drivers.
clock driver will map scu area while pinctrl one will map gpio memory block.
pwm can be just a child of pinctrl node.
What do you think about this approach? Can we address the requirements above
via classic mfd driver?

Regards,
Lorenzo