Re: [PATCH v4 05/10] mfd: max77650: new core mfd driver

[Bartosz Golaszewski <bgolaszewski@xxxxxxxxxxxx>]

śr., 13 lut 2019 o 10:53 Lee Jones <lee.jones@xxxxxxxxxx> napisał(a):
>
> On Wed, 13 Feb 2019, Bartosz Golaszewski wrote:
>
> > śr., 13 lut 2019 o 10:25 Lee Jones <lee.jones@xxxxxxxxxx> napisał(a):
> > >
> > > On Tue, 12 Feb 2019, Lee Jones wrote:
> > >
> > > > On Tue, 12 Feb 2019, Bartosz Golaszewski wrote:
> > > >
> > > > > wt., 12 lut 2019 o 12:14 Lee Jones <lee.jones@xxxxxxxxxx> napisał(a):
> > > > > >
> > > > > > On Tue, 12 Feb 2019, Bartosz Golaszewski wrote:
> > > > > >
> > > > > > > wt., 12 lut 2019 o 11:18 Lee Jones <lee.jones@xxxxxxxxxx> napisał(a):
> > > > > > > >
> > > > > > > > On Tue, 12 Feb 2019, Bartosz Golaszewski wrote:
> > > > > > > >
> > > > > > > > > wt., 12 lut 2019 o 10:55 Lee Jones <lee.jones@xxxxxxxxxx> napisał(a):
> > > > > > > > > >
> > > > > > > > > >  * The declaration of a superfluous struct
> > > > > > > > > >  * 100 lines of additional/avoidable code
> > > > > > > > > >  * Hacky hoop jumping trying to fudge VIRQs into resources
> > > > > > > > > >  * Resources were designed for HWIRQs (unless a domain is present)
> > > > > > > > > >  * Loads of additional/avoidable CPU cycles setting all this up
> > > > > > > > >
> > > > > > > > > While the above may be right, this one is negligible and you know it. :)
> > > > > > > >
> > > > > > > > You have nested for() loops.  You *are* wasting lots of cycles.
> > > > > > > >
> > > > > > > > > > Need I go on? :)
> > > > > > > > > >
> > > > > > > > > > Surely the fact that you are using both sides of an API
> > > > > > > > > > (devm_regmap_init_i2c and regmap_irq_get_*) in the same driver, must
> > > > > > > > > > set some alarm bells ringing?
> > > > > > > > > >
> > > > > > > > > > This whole HWIRQ setting, VIRQ getting, resource hacking is a mess.
> > > > > > > > > >
> > > > > > > > > > And for what?  To avoid passing IRQ data to a child driver?
> > > > > > > > >
> > > > > > > > > What do you propose? Should I go back to the approach in v1 and pass
> > > > > > > > > the regmap_irq_chip_data to child drivers?
> > > > > > > >
> > > > > > > > I'm saying you should remove all of this hackery and pass IRQs as they
> > > > > > > > are supposed to be passed (like everyone else does).
> > > > > > >
> > > > > > > I'm not sure what you mean by "like everyone else does" - different
> > > > > > > mfd drivers seem to be doing different things. Is a simple struct
> > > > > > > containing virtual irq numbers passed to sub-drivers fine?
> > > > > >
> > > > > > How do you plan on deriving the VIRQs to place into the struct?
> > > > >
> > > > > Exampe:
> > > > >
> > > > > struct max77650_gpio_pdata {
> > > > >     int gpi_irq;
> > > > > };
> > > > >
> > > > > In MFD driver:
> > > > >
> > > > > struct max77650_gpio_pdata *gpio_data = devm_kmalloc(dev, sizeof(*gpio_data));
> > > > >
> > > > > gpio_data->gpi_irq = regmap_irq_get_virq(irqchip_data, GPI_NUM);
> > > > >
> > > > > gpio_cell.platform_data = gpio_data;
> > > > >
> > > > > In GPIO driver:
> > > > >
> > > > > struct max77650_gpio_pdata *gpio_data = pdev->dev.platform_data;
> > > > >
> > > > > int irq = gpio_data->gpi_irq;
> > > >
> > > > Definitely not.  What you're trying to do is a hack.
> > > >
> > > > If you're using Regmap to handle your IRQs, then you should use Regmap
> > > > in the client to pull them out.  Setting them via Regmap, then pulling
> > > > them out again in the *same driver*, only to store them in platform
> > > > data to be passed to a child device is bonkers.
> > > >
> > > > *Either* use the MFD provided platform-data helpers *or* pass and
> > > > handle them via the Regmap APIs, *not* both.
> > >
> > > Right, a plan has been formed.
> > >
> > > Hopefully this works and you can avoid all this dancing around.
> > >
> > > Firstly, you need to make a small change to:
> > >
> > >   drivers/base/regmap/regmap-irq.c
> > >
> > > Add the following function:
> > >
> > >   struct irq_domain *regmap_irq_get_domain(struct regmap *map)
> >
> > We already do have such function and a lot of mfd drivers actually use it.
>
> Even better.
>
> > > As you can see, it will return the IRQ Domain for the chip.
> > >
> > > You can then pass this IRQ domain to mfd_add_devices() and it will do
> > > the HWIRQ => VIRQ mapping for you on the fly.  Meaning that you can
> > > remove all the nastiness in max77650_setup_irqs() and have the Input
> > > device use the standard (e.g. platform_get_irq()) APIs.
> > >
> > > How does that Sound?
> >
> > This does sound better! Why didn't you lead with that in the first place?
>
> I'm not even going to dignify that stupid question with a response.
>

It's not a stupid question and you're being unnecessarily rude. As an
expert in the subsystem you maintain you could have answered simply
with a "this is wrong, retrieve the irq domain from the regmap
irq_chip and pass it over to mfd_add_devices, the mfd core will create
appropriate mappings".

> > It's a pity it's not documented, I had to look at the code to find out
> > irq resources would get translated in mfd_add_devices() if a domain is
> > present.
>
> Where is it likely to be documented?  MFD is pretty simple and seldom
> needs explanation.  A 3 second look at the API you're trying to use
> (instead of blind copy & paste) would have told you that it's possible
> to take an IRQ domain as an argument.
>
> It's only the craziness in this patch which forced me to look into how
> Regmap handles IRQs and come up with a subsequent solution which fits
> your use-case.
>
> > In that case - I really don't see a reason to create a superfluous
> > structure to only hold the main regmap - we can very well get it from
> > the parent device in sub-drivers as I do now.
>
> The whole point of this solution is that you don't need to pass
> anything non-standard (i.e. not provided by the current APIs) to the
> child device.
>

I don't understand what you're saying here.

Bartosz