Re: [PATCH v1 4/4] gpiolib: Reuse device's fwnode to create IRQ domain

[Andy Shevchenko]

On Thu, Mar 04, 2021 at 09:06:08AM +0100, Linus Walleij wrote:
> On Wed, Mar 3, 2021 at 10:35 AM Andy Shevchenko
> <andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> > On Wed, Mar 03, 2021 at 10:22:02AM +0100, Linus Walleij wrote:
> 
> > > But this:
> > >
> > > > @@ -1504,15 +1497,14 @@ static int gpiochip_add_irqchip(struct gpio_chip *gc,
> > > >                         return ret;
> > > >         } else {
> > > >                 /* Some drivers provide custom irqdomain ops */
> > > > -               if (gc->irq.domain_ops)
> > > > -                       ops = gc->irq.domain_ops;
> > > > -
> > > > -               if (!ops)
> > > > -                       ops = &gpiochip_domain_ops;
> > > > -               gc->irq.domain = irq_domain_add_simple(np,
> > > > -                       gc->ngpio,
> > > > -                       gc->irq.first,
> > > > -                       ops, gc);
> > > > +               ops = gc->irq.domain_ops ?: &gpiochip_domain_ops;
> > > > +               if (gc->irq.first)
> > > > +                       gc->irq.domain = irq_domain_create_legacy(fwnode, gc->ngpio,
> > > > +                                                                 gc->irq.first, 0,
> > > > +                                                                 ops, gc);
> > > > +               else
> > > > +                       gc->irq.domain = irq_domain_create_linear(fwnode, gc->ngpio,
> > > > +                                                                 ops, gc);
> > >
> > > This looks like a refactoring and reimplementation of irq_domain_add_simple()?
> >
> > If you named it as irq_domain_create_simple(), then yes, but the problem is
> > that we don't have irq_domain_create_simple() API right now.
> >
> > > Why, and should it rather be a separate patch?
> >
> > Nope.
> 
> OK I looked closer at irq_domain_add_simple(), and what it does different
> is to call irq_alloc_descs() for all lines if using sparse IRQs and then
> associate them. irq_domain_create_linear|legacy() does not allocate IRQ
> descriptors because it assumes something like DT or ACPI will do that
> on-demand when drivers request IRQs.
> 
> This may be dangerous because some old platforms do not resolve IRQs
> at runtime and you will get NULL pointer exceptions.
> 
> We then need to make sure all callers do what is done in e.g.
> drivers/gpio/gpio-omap.c in the #ifdef CONFIG_ARCH_OMAP1 clause:
> they need to be augmented to call irq_alloc_descs() explicitly,
> and I don't think all of them do it as nicely for us as OMAP1.
> 
> I might be overly cautious though, however that is why this code
> uses irq_domain_add_simple(), came in commit
> commit 2854d167cc545d0642277bf8b77f972a91146fc6

Ah, thanks! I was puzzled how and why the approach above had been extended like
now. This explains it. Okay, I will introduce irq_domain_create_simple().

-- 
With Best Regards,
Andy Shevchenko