Re: [PATCH v3 2/2] pinctrl: stm32: Add level interrupt support to gpio irq chip

From: Marc Zyngier
Date: Mon Mar 23 2020 - 15:49:52 EST


On Mon, 23 Mar 2020 20:37:54 +0100
Marek Vasut <marex@xxxxxxx> wrote:

> On 3/23/20 8:31 PM, Marc Zyngier wrote:
> > On Mon, 23 Mar 2020 20:19:39 +0100
> > Marek Vasut <marex@xxxxxxx> wrote:
> >
> >> On 3/23/20 8:04 PM, Marek Vasut wrote:
> >>> On 2/20/20 10:17 AM, Marc Zyngier wrote:
> >>>> On 2020-02-20 09:04, Linus Walleij wrote:
> >>>>> On Wed, Feb 19, 2020 at 3:32 PM Alexandre Torgue
> >>>>> <alexandre.torgue@xxxxxx> wrote:
> >>>>>
> >>>>>> GPIO hardware block is directly linked to EXTI block but EXTI handles
> >>>>>> external interrupts only on edge. To be able to handle GPIO interrupt on
> >>>>>> level a "hack" is done in gpio irq chip: parent interrupt (exti irq
> >>>>>> chip)
> >>>>>> is retriggered following interrupt type and gpio line value.
> >>>>>>
> >>>>>> Signed-off-by: Alexandre Torgue <alexandre.torgue@xxxxxx>
> >>>>>> Tested-by: Marek Vasut <marex@xxxxxxx>
> >>>>>
> >>>>> Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>
> >>>>>
> >>>>> If Marc want to merge it with patch 1/2 go ahead!
> >>>>
> >>>> I'll queue the whole thing for 5.7.
> >>>
> >>> I have a feeling this doesn't work with threaded interrupts.
> >>>
> >>> If the interrupt handler runs in a thread context, the EOI will happen
> >>> almost right away (while the IRQ handler runs) and so will the code
> >>> handling the IRQ retriggering. But since the IRQ handler still runs and
> >>> didn't return yet, the retriggering doesn't cause the IRQ handler to be
> >>> called again once it finishes, even if the IRQ line is still asserted.
> >>> And that could result in some of the retriggers now happening I think.
> >>> Or am I doing something wrong ?
> >>
> >> The patch below makes my usecase work, but I don't know whether it's
> >> correct. Basically once the threaded IRQ handler finishes and unmasks
> >> the IRQ, check whether the line is asserted and retrigger if so.
> >>
> >> diff --git a/drivers/pinctrl/stm32/pinctrl-stm32.c
> >> b/drivers/pinctrl/stm32/pinctrl-stm32.c
> >> index 9ac9ecfc2f34..060dbcb7ae72 100644
> >> --- a/drivers/pinctrl/stm32/pinctrl-stm32.c
> >> +++ b/drivers/pinctrl/stm32/pinctrl-stm32.c
> >> @@ -371,12 +371,26 @@ static void
> >> stm32_gpio_irq_release_resources(struct irq_data *irq_data)
> >> gpiochip_unlock_as_irq(&bank->gpio_chip, irq_data->hwirq);
> >> }
> >>
> >> +static void stm32_gpio_irq_unmask(struct irq_data *d)
> >> +{
> >> + struct stm32_gpio_bank *bank = d->domain->host_data;
> >> + int level;
> >> +
> >> + irq_chip_unmask_parent(d);
> >> +
> >> + /* If level interrupt type then retrig */
> >> + level = stm32_gpio_get(&bank->gpio_chip, d->hwirq);
> >> + if ((level == 0 && bank->irq_type[d->hwirq] ==
> >> IRQ_TYPE_LEVEL_LOW) ||
> >> + (level == 1 && bank->irq_type[d->hwirq] == IRQ_TYPE_LEVEL_HIGH))
> >> + irq_chip_retrigger_hierarchy(d);
> >> +}
> >> +
> >> static struct irq_chip stm32_gpio_irq_chip = {
> >> .name = "stm32gpio",
> >> .irq_eoi = stm32_gpio_irq_eoi,
> >> .irq_ack = irq_chip_ack_parent,
> >> .irq_mask = irq_chip_mask_parent,
> >> - .irq_unmask = irq_chip_unmask_parent,
> >> + .irq_unmask = stm32_gpio_irq_unmask,
> >> .irq_set_type = stm32_gpio_set_type,
> >> .irq_set_wake = irq_chip_set_wake_parent,
> >> .irq_request_resources = stm32_gpio_irq_request_resources,
> >>
> >
> > OK, I see your problem now.
> >
> > The usual flow is along the line of Ack+Eoi, and that's what the
> > current code guarantees.
> >
> > Threaded interrupts do Ack+Mask+Eoi, followed by an Unmask once the
> > thread finishes. This unmask needs to do the retrigger as well, as you
> > found out.
> >
> > Can you please refactor the above so that we have the common code
> > between unmask and eoi in a separate function, send a proper patch, and
> > I'll apply it on top of the current irq/irqchip-5.7 branch.
>
> Sure, I can. Do we still need this retriggering in the irq_eoi too ?

Absolutely, because that's what matters for the non-threaded case
(there is no mask/unmask on that path). It is also never wrong to
over-resample (it just slows things down).

> Also, are there any other hidden details I might've missed ?

Probably. But let's fix one bug at a time, shall we? ;-) And let's hope
that ST doesn't take this as a excuse not to clean up their act in
their next SoC!

Thanks,

M.
--
Jazz is not dead. It just smells funny...