Re: [PATCH v3 4/7] gpio: gpiolib: fix allocation order in hierarchical IRQ domains

[Bartosz Golaszewski]

On Fri, Mar 13, 2026 at 3:05 PM Tommaso Merciai
<tommaso.merciai.xr@xxxxxxxxxxxxxx> wrote:
>
> Hi Oleksij,
>
> On 3/13/26 14:08, Oleksij Rempel wrote:
> > Hi Tommaso,
> >
> > On Fri, Mar 13, 2026 at 11:42:34AM +0100, Tommaso Merciai wrote:
> >> Hi Oleksij,
> >> Thanks for your patch.
> >>
> >> I'm working on DSI support for RZ/G3E
> >>
> >> from this morning rebasing on top of next-20260312 I'm seeing
> >> the following:
> >> I found out the the issue is related to the interrupt of the adv7535
> >> bridge:
> >>
> >>          adv7535: hdmi1@3d {
> >>                  compatible = "adi,adv7535";
> >>                  ...
> >>                  ...
> >>                  interrupts-extended = <&pinctrl RZG3E_GPIO(L, 4) IRQ_TYPE_EDGE_FALLING>;
> >>
> >> RZ/G3E is using:
> >>   - drivers/pinctrl/renesas/pinctrl-rzg2l.c
> >>
> >> Reverting this patch fix the issue.
> >> (git revert a23463beb3d5)
> >
> > Thank you for the feedback! If I understand the problem correctly, the
> > adv7535 is asserting its IRQ line early during probe, which creates an
> > irq storm due to a missing handler.
> >
> > My patch moved irq_domain_set_info() after the parent allocation. When
> > the parent allocates the IRQ, the pending hardware interrupt fires
> > immediately. Because the child descriptor isn't fully configured yet, it
> > routes to handle_bad_irq. This fails to acknowledge the hardware
> > interrupt, locking up the CPU and causing the RCU stall.
> >
> > I hope splitting the irq_domain_set_info() should fix the regression.
> > Can you please test if this change resolve the RCU stalls on your setup:
>
> Thanks for sharing.
>
> >
> > diff --git a/drivers/gpio/gpiolib.c b/drivers/gpio/gpiolib.c
> > index 13dd97344b26..376daeddbbbb 100644
> > --- a/drivers/gpio/gpiolib.c
> > +++ b/drivers/gpio/gpiolib.c
> > @@ -1628,6 +1628,9 @@ static int gpiochip_hierarchy_irq_domain_alloc(struct irq_domain *d,
> >       }
> >       gpiochip_dbg(gc, "found parent hwirq %u\n", parent_hwirq);
> >
> > +     irq_set_handler(irq, girq->handler);
> > +     irq_set_handler_data(irq, gc);
> > +
> >       /* This parent only handles asserted level IRQs */
> >       ret = girq->populate_parent_alloc_arg(gc, &gpio_parent_fwspec,
> >                                             parent_hwirq, parent_type);
> > @@ -1655,13 +1658,7 @@ static int gpiochip_hierarchy_irq_domain_alloc(struct irq_domain *d,
> >        * We set handle_bad_irq because the .set_type() should
> >        * always be invoked and set the right type of handler.
> >        */
> > -     irq_domain_set_info(d,
> > -                         irq,
> > -                         hwirq,
> > -                         gc->irq.chip,
> > -                         gc,
> > -                         girq->handler,
> > -                         NULL, NULL);
> > +     irq_domain_set_hwirq_and_chip(d, irq, hwirq, gc->irq.chip, gc);
> >       irq_set_probe(irq);
> >
> >       return 0;
>
> Tested on RZ/G3E + adv7535.
>
> With this fix all is working fine on my side.
> I'm not more seeing the seeing the RCU stall.
>

Ah, I sent this patch upstream for v7.0. I will tell Linus to not pull
it. How do we want to handle it then? Should this patch go together
with the rest of the series?

Bartosz