Re: [PATCH] irqchip/gic: Convert to handle_strict_flow_irq()

[Marc Zyngier <maz@kernel.org>]

On Tue, 17 Aug 2021 01:30:43 +0100,
Valentin Schneider <valentin.schneider@arm.com> wrote:
> 
> On 15/08/21 07:54, Marc Zyngier wrote:
> > This is going and-up in a wack-a-mole game. There is probably a bunch
> > of these all over the place. I'd rather squash it at the root,
> > i.e. with something like this (untested):
> >
> > diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c
> > index 099bc7e13d1b..601ad3fc47cd 100644
> > --- a/kernel/irq/chip.c
> > +++ b/kernel/irq/chip.c
> > @@ -410,7 +410,12 @@ void irq_percpu_disable(struct irq_desc *desc, unsigned int cpu)
> >
> >  void ack_irq(struct irq_desc *desc)
> >  {
> > -	desc->irq_data.chip->irq_ack(&desc->irq_data);
> > +	struct irq_data *data = &desc->irq_data;
> > +
> > +	while (!data->chip->irq_ack)
> > +		data = data->parent_data;
> > +
> > +	data->chip->irq_ack(&desc->irq_data);
> >
> >       if (desc->irq_data.chip->flags & IRQCHIP_AUTOMASKS_FLOW)
> >               irq_state_set_flow_masked(desc);
> >
> > We probably need something similar for irq_eoi().
> >
> > This however shows a more fundamental problem, I'm afraid. We set
> > IRQCHIP_AUTOMASKS_FLOW in the GIC drivers (i.e. at the root), but test
> > for it at the top of the hierarchy. As soon as we have more than a
> > single layer of irqchip, this will do the wrong thing (or at least
> > miss the masking optimisation).
> >
> 
> Yup.
> 
> > This probably advocates for moving the flag into the descriptor. This
> > really makes sense, as the flow is global to the whole stack, not just
> > to the localised irqchip.
> >
> 
> Are we guaranteed to have
> 
>   .irq_ack \in {NULL, irq_chip_ack_parent}
> 
> for all intermediate (!root) irqchips? I don't see why that wouldn't
> be the case, and with that in mind what you described makes sense to
> me.

An intermediate layer is allowed to implement its own irq_ack that is
not irq_chip_ack_parent, but it then has to call irq_chip_ack_parent
itself.

There is the bizarre case of drivers/gpio/gpio-thunderx.c that changes
the irqchip flow to use either handle_fasteoi_ack_irq or
handle_fasteoi_mask_irq, which won't play very nicely with this.
Someone said Cavium?

> 
> > In order to restore -next into a working state, I'm temporarily
> > dropping this series. Hopefully, we can sort this out before the merge
> > window and reinstate it.
> >
> 
> I'm away from any keyboard for most of this week, but I'll get to it by the
> weekend.

No worries, enjoy your break!

	M.

-- 
Without deviation from the norm, progress is not possible.