Re: [PATCH 2/2] pinctrl: ocelot: add support for interrupt controller

[Quentin Schulz]

Hi Linus,

On Mon, Aug 06, 2018 at 01:06:23PM +0200, Linus Walleij wrote:
> Hi Quentin, sorry for delays!
> 

No worries :)

> On Wed, Jul 25, 2018 at 2:27 PM Quentin Schulz
> <quentin.schulz@xxxxxxxxxxx> wrote:
> 
> > This GPIO controller can serve as an interrupt controller as well on the
> > GPIOs it handles.
> >
> > An interrupt is generated whenever a GPIO line changes and the
> > interrupt for this GPIO line is enabled. This means that both the
> > changes from low to high and high to low generate an interrupt.
> >
> > For some use cases, it makes sense to ignore the high to low change and
> > not generate an interrupt. Such a use case is a line that is hold in a
> > level high/low manner until the event holding the line gets acked.
> > This can be achieved by making sure the interrupt on the GPIO controller
> > side gets acked and masked only after the line gets hold in its default
> > state, this is what's done with the fasteoi functions.
> >
> > Only IRQ_TYPE_EDGE_BOTH and IRQ_TYPE_LEVEL_HIGH are supported for now.
> >
> > Signed-off-by: Quentin Schulz <quentin.schulz@xxxxxxxxxxx>
> 
> Patch applied, it's such a pretty and straight-forward patch.
> Also IRQ is probably very nice to have, so let's get this in and
> supported.
> 
> Please consider addressing the following in follow-up patch(es):
> 
> > +static int ocelot_irq_set_type(struct irq_data *data, unsigned int type);
> 
> Can't you just move the function above so you don't have to forward-declare
> this?
> 

No I can't, not in the current implementation at least.

In this function, I set the irq chip handler which needs one of the
below irq_chip. As you can see, the set_type function is also defined in
those two structures.

> > +static struct irq_chip ocelot_eoi_irqchip = {
> > +       .name           = "gpio",
> > +       .irq_mask       = ocelot_irq_mask,
> > +       .irq_eoi        = ocelot_irq_ack,
> > +       .irq_unmask     = ocelot_irq_unmask,
> > +       .flags          = IRQCHIP_EOI_THREADED | IRQCHIP_EOI_IF_HANDLED,
> 
> As you see the latter part of the define is "IF_HANDLED".
> 
> > +       .irq_set_type   = ocelot_irq_set_type,
> > +};
> > +
> > +static struct irq_chip ocelot_irqchip = {
> > +       .name           = "gpio",
> > +       .irq_mask       = ocelot_irq_mask,
> > +       .irq_ack        = ocelot_irq_ack,
> > +       .irq_unmask     = ocelot_irq_unmask,
> > +       .irq_set_type   = ocelot_irq_set_type,
> > +};
> 
> Is it really neccessary to have two irqchips?
> 
> Is this to separate ACK and EOI because the EOI version
> doesn't survive an ACK?
> 

Let me give you a real world use case, the reason why I used EOI.

I've a PHY interrupt line that connects to a GPIO of this controller.
The PHY holds the line until the PHY acknowledges the reason for
generating an interrupt. So we can say that it's a LEVEL_HIGH
"interrupt".

The GPIO interrupt controller generates an interruption whenever the
GPIO line has changed (0->1 or 1->0).

In a "normal" use case, I'd have the PHY holding high the line, the GPIO
controller generating an interrupt because the line has changed. We
acknowledge the interrupt in the GPIO controller. Later, we acknowledge
in the PHY the reason why we had to generate an "interrupt" on the PHY
side, resulting in the line being now held low. This generates another
interrupt on the GPIO controller side. However, only one actual
interrupt should have been generated by the GPIO controller as there's
only one event that resulted in an "interrupt" generation from the PHY.

So, we need EOI so that the GPIO controller interrupt gets acked only
when the PHY event is acked so that we don't receive an unexpected
interrupt.

However, we also need the common irq handler behaviour as we might want
to have the interrupt generated for 0->1 AND 1->0 transition depending
on the IP connected to the GPIO.

I could make my use case work with EOI but I find it very specific to my
use case and thus added the support for a more commonly found (IMHO) use
case.

I'd say that, yes, the two irqchips are needed but I may be missing
another way to handle all that.

Thanks,
Quentin

Attachment: signature.asc
Description: PGP signature