On Tue, Nov 01, 2016 at 02:24:38PM +0000, Jon Hunter wrote:
> Hi Mika,
>
> On 01/11/16 13:02, Mika Westerberg wrote:
> > Hi,
> >
> > I started seeing following messages on Intel Broxton when the
> > pinctrl/GPIO driver [1] loads:
> >
> > [ 0.645786] genirq: irq 14 uses trigger mode 8; requested 0
> >
> > The driver shares interrupt with other GPIO "communities" or banks so it
> > uses request_irq() instead of irq_set_chained_handler_and_data(). The
> > driver does not specify IRQ flags as those come from ACPI resources.
> >
> > This started happen after commit 4b357daed698 ("genirq: Look-up trigger
> > type if not specified by caller").
> >
> > I think this is what happens:
> >
> > 1. ACPI platform sets up the interrupt according what is in the _CRS
> > of the GPIO device. This ends up setting trigger type for irq_data of
> > the irq.
> >
> > 2. First GPIO device is found and the driver calls request_irq() which
> > calls __setup_irq() where shared == 0.
> >
> > 3. Since new->flags is read back from irq_data we call __irq_set_trigger()
> > passing the flags.
> >
> > 4. The parent IRQ chip, IO-APIC, does not have ->irq_set_type callback
> > so __irq_set_trigger() never calls irq_settings_set_trigger_mask() for
> > the desciptor.
> >
> > 5. The second GPIO device is found and this time shared == 1 so we
> > end up comparing nmsk with omsk where nmsk was read from irq_data
> > and omsk is read using irq_settings_get_trigger_mask().
> >
> > 6. Because we never called irq_settings_set_trigger_mask() for the
> > descriptor, omsk is 0 and we print out a warning:
> >
> > [ 0.645786] genirq: irq 14 uses trigger mode 8; requested 0
> >
> > If I revert commit 4b357daed698 the warning goes away.
> >
> > Do you have any ideas how to get rid of the warning properly?
>
> May be I am misunderstanding something here, but if the parent does not have
> a ->irq_set_type callback, then it would seem that the type for the
> interrupt should be not specified/set in the ACPI _CRS for the GPIO device,
> right?
Not sure.
Why the parent driver (IO-APIC) does not have ->irq_set_type callback is
beyond me. I guess it might have something to do with the IRQ hierarchy
domains it is part of.
When the ACPI core parses _CRS for the GPIO device it calls
acpi_register_gsi() with the triggering flags from _CRS and that ends up
calling acpi_register_gsi_ioapic() that programs the hardware
accordingly. So we definitely need to have the type in _CRS.
