Re: [PATCH v3 5/5] pinctrl: renesas: pinctrl-rzg2l: Add IRQ domain to handle GPIO interrupt

[Geert Uytterhoeven <geert@linux-m68k.org>]

Hi Prabhakar,

On Fri, May 13, 2022 at 3:56 PM Lad, Prabhakar
<prabhakar.csengg@gmail.com> wrote:
> On Fri, May 13, 2022 at 7:53 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> > On Thu, May 12, 2022 at 7:36 PM Lad, Prabhakar
> > <prabhakar.csengg@gmail.com> wrote:
> > > On Thu, May 12, 2022 at 8:39 AM Geert Uytterhoeven <geert@linux-m68k.org> wrote:
> > > > On Wed, May 11, 2022 at 8:32 PM Lad Prabhakar
> > > > <prabhakar.mahadev-lad.rj@bp.renesas.com> wrote:
> > > > > Add IRQ domian to RZ/G2L pinctrl driver to handle GPIO interrupt.
> > > > > GPIO0-GPIO122 pins can be used as IRQ lines but only 32 pins can be
> > > > > used as IRQ lines at given time. Selection of pins as IRQ lines
> > > > > is handled by IA55 (which is the IRQC block) which sits in between the
> > > > > GPIO and GIC.
> > > > >
> > > > > Signed-off-by: Lad Prabhakar <prabhakar.mahadev-lad.rj@bp.renesas.com>
> > > >
> > > > Thanks for your patch!
> > > >
> > > > > --- a/drivers/pinctrl/renesas/pinctrl-rzg2l.c
> > > > > +++ b/drivers/pinctrl/renesas/pinctrl-rzg2l.c
> > > >
> > > > >  static int rzg2l_gpio_register(struct rzg2l_pinctrl *pctrl)
> > > > >  {
> > > > >         struct device_node *np = pctrl->dev->of_node;
> > > > >         struct gpio_chip *chip = &pctrl->gpio_chip;
> > > > >         const char *name = dev_name(pctrl->dev);
> > > > > +       struct irq_domain *parent_domain;
> > > > >         struct of_phandle_args of_args;
> > > > > +       struct device_node *parent_np;
> > > > > +       struct gpio_irq_chip *girq;
> > > > >         int ret;
> > > > >
> > > > > +       parent_np = of_irq_find_parent(np);
> > > > > +       if (!parent_np)
> > > > > +               return -ENXIO;
> > > > > +
> > > > > +       parent_domain = irq_find_host(parent_np);
> > > > > +       of_node_put(parent_np);
> > > > > +       if (!parent_domain)
> > > > > +               return -EPROBE_DEFER;
> > > > > +
> > > > >         ret = of_parse_phandle_with_fixed_args(np, "gpio-ranges", 3, 0, &of_args);
> > > > >         if (ret) {
> > > > >                 dev_err(pctrl->dev, "Unable to parse gpio-ranges\n");
> > > > > @@ -1138,6 +1330,15 @@ static int rzg2l_gpio_register(struct rzg2l_pinctrl *pctrl)
> > > > >         chip->base = -1;
> > > > >         chip->ngpio = of_args.args[2];
> > > > >
> > > > > +       girq = &chip->irq;
> > > > > +       girq->chip = &rzg2l_gpio_irqchip;
> > > > > +       girq->fwnode = of_node_to_fwnode(np);
> > > > > +       girq->parent_domain = parent_domain;
> > > > > +       girq->child_to_parent_hwirq = rzg2l_gpio_child_to_parent_hwirq;
> > > > > +       girq->populate_parent_alloc_arg = rzg2l_gpio_populate_parent_fwspec;
> > > > > +       girq->child_irq_domain_ops.free = rzg2l_gpio_irq_domain_free;
> > > > > +       girq->ngirq = RZG2L_TINT_MAX_INTERRUPT;
> > > > > +
> > > >
> > > > I think you need to provide a .init_valid_mask() callback, as
> > > > gpiochip_irqchip_remove() relies on that for destroying interrupts.
> > > Are you suggesting  the callback to avoid looping through all the GPIO pins?
> >
> > gpiochip_irqchip_remove() does:
> >
> >         /* Remove all IRQ mappings and delete the domain */
> >         if (gc->irq.domain) {
> >                 unsigned int irq;
> >
> >                 for (offset = 0; offset < gc->ngpio; offset++) {
> >                        if (!gpiochip_irqchip_irq_valid(gc, offset))
> >                                 continue;
> >
> >                         irq = irq_find_mapping(gc->irq.domain, offset);
> >                         irq_dispose_mapping(irq);
> >                 }
> >
> >                 irq_domain_remove(gc->irq.domain);
> >
> >         }
> >
> > The main thing is not about avoiding to loop through all GPIO pins,
> > but to avoid irq_{find,dispose}_mapping() doing the wrong thing.
> So in our case if we don't implement valid masks, that would mean all
> the pins are valid. irq_find_mapping() would return 0 if no mapping is
> found to the corresponding offset and irq_dispose_mapping() would
> simply return back without doing anything if virq == 0.(In this patch
> rzg2l_gpio_free() does call irq_{find,dispose}_mapping())

But "offset" is a number from the GPIO offset space (0-122), while
irq_find_mapping() expects a number from the domain's IRQ space,
which is only 0-31?

> > The loop is over all GPIO offsets, while not all of them are mapped
> > to valid interrupts. Does the above work correctly?
> >
> I haven't tested unloading the pinctrl driver which should call
> gpiochip_irqchip_remove() (we don't have remove call back for pinctrl
> driver)
>
> > > > However, the mask will need to be dynamic, as GPIO interrupts can be
> > > > mapped and unmapped to one of the 32 available interrupts dynamically,
> > > > right?
> > > Yep that's correct.
> > >
> > > > I'm not sure if that can be done easily: if gpiochip_irqchip_irq_valid()
> > > > is ever called too early, before the mapping is done, it would fail.
> > > >
> > > The mask initialization is a one time process and that is during
> > > adding the GPIO chip. At this stage we won't be knowing what will be
> > > the valid GPIO pins used as interrupts. Maybe the core needs to
> > > implement a callback which lands in the GPIO controller driver to tell
> > > if the gpio irq line is valid. This way we can handle dynamic
> > > interrupts.
> >
> > Upon closer look, I think the mask is a red herring, and we don't
> > need it.
> Agreed.
>
> > But we do need to handle the (possible) mismatch between GPIO
> > offset (index) and IRQ offset in the above code.
> >
> Agreed, do you see any possibility of the mismatch I have missed?

gpiochip_to_irq():

        if (irq_domain_is_hierarchy(domain)) {
                struct irq_fwspec spec;

                spec.fwnode = domain->fwnode;
                spec.param_count = 2;
                spec.param[0] = gc->irq.child_offset_to_irq(gc, offset);
                spec.param[1] = IRQ_TYPE_NONE;

                return irq_create_fwspec_mapping(&spec);
        }

Same here: in the absence of a child_offset_to_irq() callback,
the default gpiochip_child_offset_to_irq_noop() will be used,
assuming an identity mapping between GPIO numbers and IRQ
numbers.

So perhaps
  1. you need to provide a child_offset_to_irq() callback,
  2. gpiochip_irqchip_remove() needs to apply the child_offset_to_irq()
    mapping too?
  3. you do need the mask, or let child_offset_to_irq() an error code,
     to avoid irq_{find,dispose}_mapping() handling non-existent irqs?

Or am I missing something?

I guess this is easy to verify by adding some debug prints to the code.

Gr{oetje,eeting}s,

                        Geert

--
Geert Uytterhoeven -- There's lots of Linux beyond ia32 -- geert@linux-m68k.org

In personal conversations with technical people, I call myself a hacker. But
when I'm talking to journalists I just say "programmer" or something like that.
                                -- Linus Torvalds