Re: [RFC PATCH v2 39/45] arm64: smp: Fall back to IRQ when IPI NMI request fails

[Jinjie Ruan <ruanjinjie@huawei.com>]

在 2026/8/21 18:16, Vladimir Murzin 写道:
> On 8/13/26 10:23, Jinjie Ruan wrote:
>>
>> 在 2026/7/28 0:34, Vladimir Murzin 写道:
>>> Existing IPI setup logic relies on ipi_should_be_nmi() to decide
>>> whether to request an NMI (if supported) or an ordinary IRQ. If NMI is
>>> not supported, using an ordinary IRQ is acceptable.
>>>
>>> FEAT_NMI requires complementary support in the interrupt
>>> controller. If FEAT_NMI is present but interrupt controller support is
>>> absent, request_percpu_nmi() can fail.
>>>
>>> One way to prevent this would be to extend ipi_should_be_nmi() to also
>>> check for interrupt controller support. However, since falling back to
>>> an ordinary IRQ is acceptable, treat ipi_should_be_nmi() as a hint
>>> instead. Whether an IPI is actually registered as an NMI is determined
>>> by the success or failure of request_percpu_nmi(). If the request
>>> fails, fall back to request_irq().
>>>
>>> Track which IPIs were actually registered as NMIs and use that state
>>> in the per-CPU enable and disable paths. This avoids calling the
>>> percpu NMI helpers for an IPI that fell back to regular IRQ handling.
>>>
>>> Signed-off-by: Vladimir Murzin <vladimir.murzin@arm.com>
>>
>> Should we replace ipi_should_be_nmi() with test_bit(ipi,
>> get_ipi_nmi_bitmap(cpu) in smp_send_stop()?
>>
> 
> Right. We can do something like
> 
> @@ -1312,7 +1328,7 @@ void smp_send_stop(void)
>          * If CPUs are still online, try an NMI. There's no excuse for this to
>          * be slow, so we only give them an extra 10 ms to respond.
>          */
> -       if (num_other_online_cpus() && ipi_should_be_nmi(IPI_CPU_STOP_NMI)) {
> +       if (num_other_online_cpus() && ipi_is_nmi(smp_processor_id(), IPI_CPU_STOP_NMI)) {
>                 smp_rmb();
>                 cpumask_copy(&mask, cpu_online_mask);
>                 cpumask_clear_cpu(smp_processor_id(), &mask);
> 
> where 
> 
> +#define ipi_is_nmi(__cpu, __ipi) test_bit(__ipi, get_ipi_nmi_bitmap(__cpu))
> 
> can be used in other places to simpify the code 

Agree.

> 
>> irq_desc_get_irq(get_ipi_desc(cpu, ipi)) has been used more than 3
>> times, can we define a new macro for use?
>>
>> git grep "irq_desc_get_irq(get_ipi_desc(cpu, ipi))"
>> arch/arm64/kernel/smp.c:
>> enable_nmi(irq_desc_get_irq(get_ipi_desc(cpu, ipi)));
>> arch/arm64/kernel/smp.c:
>> enable_irq(irq_desc_get_irq(get_ipi_desc(cpu, ipi)));
>> arch/arm64/kernel/smp.c:
>> disable_nmi_nosync(irq_desc_get_irq(get_ipi_desc(cpu, ipi)));
>> arch/arm64/kernel/smp.c:
>> disable_irq(irq_desc_get_irq(get_ipi_desc(cpu, ipi)));
>>
>>
> 
> I assume you did grep after full series applied, since
> irq_desc_get_irq(get_ipi_desc(cpu, ipi)) comes from LPI flow.
> 
> Anyway, we can easily avoid some duplication and get usage
> of irq_desc_get_irq(get_ipi_desc(cpu, ipi)) back to two

Anyway, it is always good to avoid repetition.

> 
> $ git grep "irq_desc_get_irq(get_ipi_desc(cpu, ipi))"
> arch/arm64/kernel/smp.c:        int irq = irq_desc_get_irq(get_ipi_desc(cpu, ipi));
> arch/arm64/kernel/smp.c:        int irq = irq_desc_get_irq(get_ipi_desc(cpu, ipi));
> 
> Cheers
> Vladimir
> 
>