Re: [PATCH v1 1/1] i2c: cht-wc: Use fwnode for the controller and IRQ domain

[Hans de Goede]

Hi,

On 2/25/21 4:44 PM, Andy Shevchenko wrote:
> On Thu, Feb 25, 2021 at 5:11 AM Hans de Goede <hdegoede@xxxxxxxxxx> wrote:
>> On 2/24/21 1:51 PM, Andy Shevchenko wrote:
>>> On Tue, Feb 23, 2021 at 08:25:35PM +0100, Hans de Goede wrote:
>>>> On 2/23/21 6:22 PM, Andy Shevchenko wrote:
>>>>> It's better to describe the I²C controller and associated IRQ domain with
>>>>> fwnode, so they will find their place in the hierarchy in sysfs and also
>>>>> make easier to debug.
> 
> ...
> 
>>>>> +   set_primary_fwnode(&adap->adapter.dev, fwnode);
>>>>
>>>> So now we have the main PMIC device i2c-client, the platform-device instantiated
>>>> for the MFD-cell for the PMIC's builtin I2C-controller; and the device instantiated
>>>> for the adapter-device all 3 share the same ACPI-companion fwnode.
>>>
>>> Okay, this step in this patch maybe not needed (or should be a separate change,
>>> but I don't see clearly what would be the benefit out of it).
> 
> Shall I leave this or should be removed in v2?
> 
> ...
> 
>>>>> -   adap->irq_domain = irq_domain_add_linear(pdev->dev.of_node, 1,
>>>>> -                                            &irq_domain_simple_ops, NULL);
>>>>> +   adap->irq_domain = irq_domain_create_linear(fwnode, 1,
>>>>> +                                               &irq_domain_simple_ops, NULL);
>>>>
>>>> Hmm, not sure this is right, admittedly the old code looks weird too, but now we
>>>> are creating a second irq_domain at the same level as the irq_domain created for
>>>> the IRQ-chip part of the PMIC. But this is really more of a child-domain of just
>>>> the I2C-controller MFD-cell. The IRQ-CHIP part of the PMIC has a single IRQ for the
>>>> I2C controller which gets raised both on i2c-transfer completions and when the
>>>> pin on the PMIC which is reserved as input for the IRQ coming out of the charger-chip
>>>> gets triggered.
>>>>
>>>> IOW we have this:
>>>>
>>>>
>>>>                PMIC
>>>>                  |
>>>>     ------------------------------
>>>>     |       |        |           |
>>>>    IRQ1   IRQ2      IRQ3       I2C-IRQ
>>>>                                  |
>>>>                    ----------------------------------
>>>>                    |        |         |             |
>>>>                  READIRQ   WRIRQ    NACKIRQ     CLIENT-IRQ
>>>>
>>>> Where READIRQ, WRIRQ and NACKIRQ are directly consumed
>>>> and the CLIENT-IRQ is being represented as a single IRQ on
>>>> a new irqchip so that we can pass it along to the i2c-driver
>>>> for the charger-chip which is connected to the Whiskey Cove's
>>>> builtin I2C controller.
>>>>
>>>> But doing as you suggest would model the IRQs as:
>>>>
>>>>                PMIC
>>>>                  |
>>>>     --------------------------------------------------
>>>>     |       |        |           |                    |
>>>>    IRQ1   IRQ2      IRQ3       I2C-IRQ           CLIENT-IRQ
>>>>
>>>> Which is not the same really. I guess it is better then what we
>>>> have though ?
>>>
>>> Hmm... There should not be difference in the hierarchy. add_linear ==
>>> create_linear. The propagation of *device* (not an IRQ) fwnode is just
>>> convenient way to have IRQ domain be named (instead of 'unknown-N' or so).
>>> Maybe I have read __irq_domain_add() code wrongly.
>>
>> Sorry, this is probably my bad. The first ASCII-art which I posted is
>> how things actually work in HW. The second one is how I assumed that
>> things would look like in some nested representation of the IRQ-domains
>> given that all the IRQs mentioned in the ASCII-art now use the same fwnode
>> as parent for their domain. But poking around in sysfs I don't see any
>> hierarchical representation of the domains at all. Actually I cannot
>> find any representation of the IRQ domains inside sysfs (I've never
>> looked at / into this before) ?
> 
> I have enabled  GENERIC_IRQ_DEBUGFS to see some information.
> 
>> If what you say is right and the fwnode is only used to set a name (where can
>> I see those names ?) then your patch is probably correct.
> 
> I have checked again and I don't see anything except it uses it as a
> domain name and takes reference count.
> 
>>> Nevertheless, thinking more about it, why we don't add an IRQ chip via regmap
>>> IRQ API?
>>
>> There already is a regmap IRQ chip associated with the MFD device and the
>> IRQ handling required here is somewhat tricky (see the comments in the driver)
>> so I would prefer to keep this as is.
> 
> Ah, that makes things complicated a bit.
> 
>>>> Note I can test any changes made here, but I'm not 100% convinced that
>>>> the current version of this patch is correct.
>>>
>>> If we settle on the idea first. I'm (slowly) looking forward to check another
>>> CherryTrail device we have at the lab, but we lack of some (power) equipment
>>> right now to setup it properly. I hope it may have the Whiskey Cove PMIC there.
>>
>> More testing is always welcome :)   With that said, testing these changes really
>> is not a lot of work for me.
> 
> I would expect that we will have a clash with IRQ domain names and
> thus we would need our own fwnode here.
> 
> I will think about it, but it sounds like we need to create a
> hierarchy of the IRQ domains and take the device's fwnode as a parent
> here.
> 
> Overall, I stumbled over of_node use in pure ACPI case (simplest "fix"
> is to provide a NULL pointer there). If you think we can get rid of
> of_node as intermediate step, I will send v2 with that.

Sorry for being slow to respond.

I agree that the of_node use is weird, so a patch which simply replaces the
pdev->dev.of_node with NULL would be good. Otherwise I would just leave the
code as is.

Regards,

Hans