Re: [PATCH v2 17/17] irq: remove handle_domain_{irq,nmi}()

[Thomas Gleixner <tglx@xxxxxxxxxxxxx>]

On Tue, May 10 2022 at 15:15, Mark Rutland wrote:
> On Tue, May 10, 2022 at 02:13:20PM +0200, Lukas Wunner wrote:
>> Actually, since you're mentioning the in_nmi() check, I suspect
>> there's another problem here:
>> 
>> generic_handle_domain_nmi() warns if !in_nmi(), then calls down 
>> to handle_irq_desc() which warns if !in_hardirq().  Doesn't this
>> cause a false-positive !in_hardirq() warning for a NMI on GIC/GICv3?
>
> I agree that doesn't look right.
>
>> The only driver calling request_nmi() or request_percpu_nmi() is
>> drivers/perf/arm_pmu.c.  So that's the only one affected.
>> You may want to test if that driver indeed exhibits such a
>> false-positive warning since c16816acd086.
>
> In testing with v5.18-rc5, I can't see that going wrong.
>
> I also hacked the following in:
>
> -------->8--------
> diff --git a/kernel/irq/irqdesc.c b/kernel/irq/irqdesc.c
> index 939d21cd55c38..3c85608a8779f 100644
> --- a/kernel/irq/irqdesc.c
> +++ b/kernel/irq/irqdesc.c
> @@ -718,6 +718,7 @@ EXPORT_SYMBOL_GPL(generic_handle_domain_irq);
>  int generic_handle_domain_nmi(struct irq_domain *domain, unsigned int hwirq)
>  {
>         WARN_ON_ONCE(!in_nmi());
> +       WARN_ON_ONCE(!in_hardirq());
>         return handle_irq_desc(irq_resolve_mapping(domain, hwirq));

which is pointless because NMI entry code has to invoke [__]nmi_enter()
before invoking this function. [__]nmi_enter() does:

    __preempt_count_add(NMI_OFFSET + HARDIRQ_OFFSET);

So it's more than bloody obvious why there is no warning triggered for a
regular hardware induced NMI invocation.

For a software invocation from the wrong context it does not matter how
many redundant WARN_ONs you add. The existing ones are covering it
nicely already.

Thanks,

        tglx