Re: [PATCH v2 10/16] gpio: add a reusable generic gpio_chip using regmap

[Michael Walle <michael@walle.cc>]

Hi Linus,

Am 2020-04-17 08:34, schrieb Michael Walle:
> Hi Linus,
> 
> Am 2020-04-16 11:27, schrieb Linus Walleij:
>> On Thu, Apr 2, 2020 at 10:37 PM Michael Walle <michael@walle.cc> 
>> wrote:
>> 
>>> There are quite a lot simple GPIO controller which are using regmap 
>>> to
>>> access the hardware. This driver tries to be a base to unify existing
>>> code into one place. This won't cover everything but it should be a 
>>> good
>>> starting point.
>>> 
>>> It does not implement its own irq_chip because there is already a
>>> generic one for regmap based devices. Instead, the irq_chip will be
>>> instanciated in the parent driver and its irq domain will be 
>>> associate
>>> to this driver.
>>> 
>>> For now it consists of the usual registers, like set (and an optional
>>> clear) data register, an input register and direction registers.
>>> Out-of-the-box, it supports consecutive register mappings and 
>>> mappings
>>> where the registers have gaps between them with a linear mapping 
>>> between
>>> GPIO offset and bit position. For weirder mappings the user can 
>>> register
>>> its own .xlate().
>>> 
>>> Signed-off-by: Michael Walle <michael@walle.cc>
>> 
>> Overall I really like this driver and I think we should merge is as 
>> soon
>> as it is in reasonable shape and then improve on top so we can start
>> migrating drivers to it.
>> 
>>> +static int gpio_regmap_to_irq(struct gpio_chip *chip, unsigned int 
>>> offset)
>>> +{
>>> +       struct gpio_regmap_data *data = gpiochip_get_data(chip);
>>> +       struct gpio_regmap *gpio = data->gpio;
>>> +
>>> +       /* the user might have its own .to_irq callback */
>>> +       if (gpio->to_irq)
>>> +               return gpio->to_irq(gpio, offset);
>>> +
>>> +       return irq_create_mapping(gpio->irq_domain, offset);
>> 
>> I think that should at least be irq_find_mapping(), the mapping should
>> definately not be created by the .to_irq() callback since that is just
>> a convenience function.
> 
> what do you mean by conenience function? are there other ways? if you 
> use
> irq_find_mapping() who will create the mappings? most gpio drivers use 
> a
> similar function like gpio_regmap_to_irq().
> 
>> 
>>> +       if (gpio->irq_domain)
>>> +               chip->to_irq = gpio_regmap_to_irq;
>> 
>> I don't know about this.
>> (...)
>>> + * @irq_domain:                (Optional) IRQ domain if the 
>>> controller is
>>> + *                     interrupt-capable
>> (...)
>>> +       struct irq_domain *irq_domain;
>> 
>> I don't think this is a good storage place for the irqdomain, we 
>> already have
>> gpio_irq_chip inside gpio_chip and that has an irqdomain, we should
>> strive to reuse that infrastructure also for regmap GPIO I think, for 
>> now
>> I would just leave .to_irq() out of this and let the driver deal with 
>> any
>> irqs.
> 
> How would a driver attach the to_irq callback then? At the moment, the
> gpio_regmap doesn't expose the gpio_chip. So either we have to do that 
> or
> the config still have to have a .to_irq property.

Also, if I move the interrupt hanling completely out of the gpio-regmap, 
the
driver would have to deal with "struct gpio_chip" which I would like to 
avoid
if possible and keep it private to gpio-regmap.

Unfortunately, I don't have much experience how a good API for the 
interrupt
handling and the gpio-regmap might look like. And there seems to be some
overlap between regmap-irq and the interrupt stuff in gpiolib. For 
example,
both provide and set the irq_domain_ops. Thus handing the domain over to
gpio-regmap looked like a good idea to me. I get you point, that there 
is
already a irqdomain in gpiolib and also a _to_irq() which is the same as
the current implementation in gpio-regmap. Maybe it makes sense to just
have a new function

int gpiolib_add_irqdomain(struct gpio_chip *gc, struct irq_domain 
domain)
{
   gc->irq.domain = domain;
   gc->to_irq = gpiochip_to_irq;
}

which is called by gpio_regmap_register() if a config->irq_domain is 
given.

-michael