Re: [PATCH v5 1/5] arm64: Add framework to turn IPI as NMI

[Sumit Garg]

On Wed, 21 Oct 2020 at 15:57, Marc Zyngier <maz@xxxxxxxxxx> wrote:
>
> On 2020-10-20 12:22, Sumit Garg wrote:
> > On Tue, 20 Oct 2020 at 15:38, Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >>
> >> On 2020-10-20 07:43, Sumit Garg wrote:
> >> > On Mon, 19 Oct 2020 at 17:07, Marc Zyngier <maz@xxxxxxxxxx> wrote:
> >>
> >> [...]
> >>
> >> >> > +{
> >> >> > +     if (!ipi_desc)
> >> >> > +             return;
> >> >> > +
> >> >> > +     if (is_nmi) {
> >> >> > +             if (!prepare_percpu_nmi(ipi_id))
> >> >> > +                     enable_percpu_nmi(ipi_id, IRQ_TYPE_NONE);
> >> >> > +     } else {
> >> >> > +             enable_percpu_irq(ipi_id, IRQ_TYPE_NONE);
> >> >>
> >> >> I'm not keen on this. Normal IRQs can't reliably work, so why do you
> >> >> even bother with this?
> >> >
> >> > Yeah I agree but we need to support existing functionality for kgdb
> >> > roundup and sysrq backtrace using normal IRQs as well.
> >>
> >> When has this become a requirement? I don't really see the point in
> >> implementing something that is known not to work.
> >>
> >
> > For kgdb:
> >
> > Default implementation [1] uses smp_call_function_single_async() which
> > in turn will invoke IPI as a normal IRQ to roundup CPUs.
> >
> > [1]
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/debug/debug_core.c#n244
> >
> > For sysrq backtrace:
> >
> > Default implementation [2] fallbacks to smp_call_function() (IPI as a
> > normal IRQ) to print backtrace in case architecture doesn't provide
> > arch_trigger_cpumask_backtrace() hook.
> >
> > [2]
> > https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/drivers/tty/sysrq.c#n250
> >
> > So in general, IPI as a normal IRQ is still useful for debugging but
> > it can't debug a core which is stuck in deadlock with interrupts
> > disabled.
>
> And that's not something we implement today for good reasons:
> it *cannot* work reliably. What changed that we all of a sudden need it?
>
> > And since we choose override default implementations for pseudo NMI
> > support, we need to be backwards compatible for platforms which don't
> > possess pseudo NMI support.
>
> No. There is nothing to be "backward compatible" with, because
> - this isn't a userspace visible feature
> - *it doesn't work*
>
> So please drop this non-feature from this series.
>

Okay, fair enough. I will drop support for new IPI being normal IRQ
and instead update sysrq backtrace and kgdb roundup frameworks to use
existing cross-calls stuff in case a platform doesn't possess pseudo
NMI support.

-Sumit

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