On Fri, Oct 25, 2024 at 10:15 AM Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On Fri, 25 Oct 2024 06:07:45 +0100,
> Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> >
> > Hi Marc,
> >
> > On Tue, Oct 22, 2024 at 9:03 AM Marc Zyngier <maz@xxxxxxxxxx> wrote:
> > >
> > > On Mon, 21 Oct 2024 05:22:15 +0100,
> > > Yu Zhao <yuzhao@xxxxxxxxxx> wrote:
> > > >
> > > > @@ -1407,6 +1418,13 @@ static void gic_ipi_send_mask(struct irq_data *d, const struct cpumask *mask)
> > > > */
> > > > dsb(ishst);
> > > >
> > > > + cpumask_copy(&broadcast, cpu_present_mask);
> > >
> > > Why cpu_present_mask? I'd expect that cpu_online_mask should be the
> > > correct mask to use -- we don't IPI offline CPUs, in general.
> >
> > This is exactly because "we don't IPI offline CPUs, in general",
> > assuming "we" means the kernel, not GIC.
> >
> > My interpretation of what the GIC spec says ("0b1: Interrupts routed
> > to all PEs in the system, excluding self") is that it broadcasts IPIs to
> > "cpu_present_mask" (minus the local one). So if the kernel uses
> > "cpu_online_mask" here, GIC would send IPIs to offline CPUs
> > (cpu_present_mask ^ cpu_online_mask), which I don't know whether it's
> > a defined behavior.
Thanks for clarifying.
> Offline CPUs are not known to the kernel.
I assume it wouldn't matter to firmware either, correct? IOW, we
wouldn't cause firmware any trouble by letting GIC send IPIs to
(cpu_present_mask ^ cpu_online_mask), assuming those two masks can be
different on arm64 when hotplug is enabled?
> Most likely, they are either
> powered off, or spending quality time in Secure or Realm mode. Either
> way, this is none of our business.
>
> Your approach would make also the kernel perform pretty inconsistently
> depending on whether CPUs are offline and not.
>
> >
> > But if you actually meant GIC doesn't IPI offline CPUs, then yes, here
> > the kernel should use "cpu_online_mask".
> >
> > > > + cpumask_clear_cpu(smp_processor_id(), &broadcast);
> > > > + if (cpumask_equal(&broadcast, mask)) {
> > > > + gic_broadcast_sgi(d->hwirq);
> > > > + goto done;
> > > > + }
> > >
> > > So the (valid) case where you would IPI *everyone* is not handled as a
> > > fast path? That seems a missed opportunity.
> >
> > You are right: it should handle that case.
> >
> > > This also seem an like expensive way to do it. How about something
> > > like:
> > >
> > > int mcnt = cpumask_weight(mask);
> > > int ocnt = cpumask_weight(cpu_online_mask);
> > > if (mcnt == ocnt) {
> > > /* Broadcast to all CPUs including self */
> >
> > Does the comment mean the following two steps?
> > 1. Broadcasting to everyone else.
> > 2. Sending to self.
>
> Correct.
>
> > My understanding of the "Interrupt Routing Mode" is that it can't
> > broadcast to all CPUs including self, and therefore we need the above
> > two steps, which still can be a lot faster. Is my understanding
> > correct?
>
> Yes.
>
> Thanks,
>
> M.
>
> --
> Without deviation from the norm, progress is not possible.
