Re: [PATCH] pinctrl: samsung: add irq_set_affinity() for non wake up external gpio interrupt

[Sam Protsenko <semen.protsenko@xxxxxxxxxx>]

On Thu, Nov 23, 2023 at 10:00 PM Youngmin Nam <youngmin.nam@xxxxxxxxxxx> wrote:
>
> On Tue, Nov 21, 2023 at 12:33:51PM -0600, Sam Protsenko wrote:
> > On Sun, Nov 19, 2023 at 2:54 AM Youngmin Nam <youngmin.nam@xxxxxxxxxxx> wrote:
> > >
> > > To support affinity setting for non wake up external gpio interrupt,
> > > we add a new irq_set_affinity callback using irq number which is in pinctrl
> > > driver data.
> > >
> > > Before applying this patch, we couldn't change irq affinity of gpio interrupt.
> > > * before
> > > erd9945:/proc/irq/418 # cat smp_affinity
> > > 3ff
> > > erd9945:/proc/irq/418 # echo 00f > smp_affinity
> > > erd9945:/proc/irq/418 # cat smp_affinity
> > > 3ff
> > > erd9945:/proc/irq/418 # cat /proc/interrupts
> > >            CPU0       CPU1       CPU2       CPU3       CPU4       CPU5       CPU6       CPU7       CPU8       CPU9
> > > 418:       3631          0          0          0          0          0          0          0          0          0      gpg2   0 Edge      19100000.drmdecon
> > >
> > > After applying this patch, we can change irq affinity of gpio interrupt as below.
> > > * after
> > > erd9945:/proc/irq/418 # cat smp_affinity
> > > 3ff
> > > erd9945:/proc/irq/418 # echo 00f > smp_affinity
> > > erd9945:/proc/irq/418 # cat smp_affinity
> > > 00f
> > > erd9945:/proc/irq/418 # cat /proc/interrupts
> > >            CPU0       CPU1       CPU2       CPU3       CPU4       CPU5       CPU6       CPU7       CPU8       CPU9
> > > 418:       3893        201        181        188          0          0          0          0          0          0      gpg2   0 Edge      19100000.drmdecon
> > >
> >
> > Suggest formatting the commit message as follows, to make it more readable:
> >
> > 8<-------------------------------------------------------------------------->8
> > To support affinity setting for non wake up external gpio interrupt,
> > add irq_set_affinity callback using irq number from pinctrl driver
> > data.
> >
> > Before this patch, changing the irq affinity of gpio interrupt is not
> > possible:
> >
> >     # cat /proc/irq/418/smp_affinity
> >     3ff
> >     # echo 00f > /proc/irq/418/smp_affinity
> >     # cat /proc/irq/418/smp_affinity
> >     3ff
> >     # cat /proc/interrupts
> >                CPU0       CPU1       CPU2       CPU3    ...
> >     418:       3631          0          0          0    ...
> >
> > With this patch applied, it's possible to change irq affinity of gpio
> > interrupt:
> >
> >     # cat /proc/irq/418/smp_affinity
> >     3ff
> >     # echo 00f > /proc/irq/418/smp_affinity
> >     # cat /proc/irq/418/smp_affinity
> >     00f
> >     # cat /proc/interrupts
> >                CPU0       CPU1       CPU2       CPU3      ...
> >     418:       3893        201        181        188      ...
> > 8<-------------------------------------------------------------------------->8
> >
>
> Thanks for your suggestion. I'll modify it.
>
> > > Signed-off-by: Youngmin Nam <youngmin.nam@xxxxxxxxxxx>
> > > ---
> > >  drivers/pinctrl/samsung/pinctrl-exynos.c | 14 ++++++++++++++
> > >  1 file changed, 14 insertions(+)
> > >
> > > diff --git a/drivers/pinctrl/samsung/pinctrl-exynos.c b/drivers/pinctrl/samsung/pinctrl-exynos.c
> > > index 6b58ec84e34b..5d7b788282e9 100644
> > > --- a/drivers/pinctrl/samsung/pinctrl-exynos.c
> > > +++ b/drivers/pinctrl/samsung/pinctrl-exynos.c
> > > @@ -147,6 +147,19 @@ static int exynos_irq_set_type(struct irq_data *irqd, unsigned int type)
> > >         return 0;
> > >  }
> > >
> > > +static int exynos_irq_set_affinity(struct irq_data *irqd,
> > > +                                  const struct cpumask *dest, bool force)
> > > +{
> > > +       struct samsung_pin_bank *bank = irq_data_get_irq_chip_data(irqd);
> > > +       struct samsung_pinctrl_drv_data *d = bank->drvdata;
> > > +       struct irq_data *parent = irq_get_irq_data(d->irq);
> >
> > I'm probably missing something, but: why not just use "irqd" parameter
> > and avoid declaring "bank" and "d"? Is "d->irq" somehow different from
> > "irqd"?
> >
>
> Yes, irqd->irq is different from d->irq as below.
>
> [  188.230707] irqd->irq is 417
> [  188.230837] d->irq is 133
>
> We have to use d->irq(133) instead of irqd->irq(417) because d->irq has GICv3 as a IRQ chip.
> To use set_affinity() call back of GICv3, d->irq is needed.
>
> IRQ  HWIRQ  Type   Affinity    IRQ_DESC             CPU0    CPU1    CPU2    CPU3 ... Chip   Name
> 133    603  Level  0x3ff       0xffffff883b25d800  52260       0       0       0 ... GICv3  11030000.pinctrl
> 417      0  Edge   0xffffffff  0xffffff883b68a800  52259       0       0       0 ... gpg2   19100000.drmdecon
>
> erd9945: # cat /proc/interrupts | grep gpg2
> 417:       9250         48         45         45 ...  gpg2   0 Edge      19100000.drmdecon
>
> erd9945: # cat /proc/interrupts | grep 11030000
> 133:       9250         48         45         45 ...  GICv3 603 Level     11030000.pinctrl
>

Thanks for the explanation! Apart from my suggestion for the commit message:

Reviewed-by: Sam Protsenko <semen.protsenko@xxxxxxxxxx>

> > > +
> > > +       if (parent)
> > > +               return parent->chip->irq_set_affinity(parent, dest, force);
> > > +
> >
> > Why not use irq_chip_set_affinity_parent() API?
> >
> > > +       return -EINVAL;
> >
> > Maybe use something like this instead:
> >
> >     if (!irqd->parent_data)
> >             return -EINVAL;
> >
> >     return irq_chip_set_affinity_parent(irqd, dest, force);
> >
> > Can you please test if this code works?
> >
>
> I tested as you suggested as below.
>
> diff --git a/drivers/pinctrl/samsung/pinctrl-exynos.c b/drivers/pinctrl/samsung/pinctrl-exynos.c
> index bf8dd5e3c3d2..593320b408ce 100644
> --- a/drivers/pinctrl/samsung/pinctrl-exynos.c
> +++ b/drivers/pinctrl/samsung/pinctrl-exynos.c
> @@ -153,14 +153,12 @@ static int exynos_irq_set_type(struct irq_data *irqd, unsigned int type)
>  static int exynos_irq_set_affinity(struct irq_data *irqd,
>                                    const struct cpumask *dest, bool force)
>  {
> -       struct samsung_pin_bank *bank = irq_data_get_irq_chip_data(irqd);
> -       struct samsung_pinctrl_drv_data *d = bank->drvdata;
> -       struct irq_data *parent = irq_get_irq_data(d->irq);
> -
> -       if (parent)
> -               return parent->chip->irq_set_affinity(parent, dest, force);
> +       if (!irqd->parent_data) {
> +               pr_err("irqd->parent_data is null!!\n");
> +               return -EINVAL;
> +       }
>
> -       return -EINVAL;
> +       return irq_chip_set_affinity_parent(irqd, dest, force);
>  }
>
> [  149.658395] irqd->parent_data is null!!
>
> Currently, irqd->paranet_data is null.
>
> > > +}
> > > +
> > >  static int exynos_irq_request_resources(struct irq_data *irqd)
> > >  {
> > >         struct samsung_pin_bank *bank = irq_data_get_irq_chip_data(irqd);
> > > @@ -212,6 +225,7 @@ static const struct exynos_irq_chip exynos_gpio_irq_chip __initconst = {
> > >                 .irq_mask = exynos_irq_mask,
> > >                 .irq_ack = exynos_irq_ack,
> > >                 .irq_set_type = exynos_irq_set_type,
> > > +               .irq_set_affinity = exynos_irq_set_affinity,
> >
> > What happens if we just assign irq_chip_set_affinity_parent() here?
> > Would it work, or Exynos case is more complicated than this?
> >
>
> Yes, I couldn't find how to use irq_chip_set_affinity_parent() directly yet.
>
> > >                 .irq_request_resources = exynos_irq_request_resources,
> > >                 .irq_release_resources = exynos_irq_release_resources,
> > >         },
> > > --
> > > 2.39.2
> > >
> >