Hi Shijith,
On 04/07/2019 05:22, Shijith Thotton wrote:
> The NMI handlers handle_percpu_devid_fasteoi_nmi() and
> handle_fasteoi_nmi() added by commit 2dcf1fbcad35 ("genirq: Provide NMI
> handlers") do not update the interrupt counts. Due to that the NMI
> interrupt count does not show up correctly in /proc/interrupts.
>
> Update the functions to fix this. With this change, we can see stats of
> the perf NMI interrupts on arm64.
>
> Fixes: 2dcf1fbcad35 ("genirq: Provide NMI handlers")
>
> Signed-off-by: Shijith Thotton <sthot...@marvell.com>
> ---
> kernel/irq/chip.c | 4 ++++
> 1 file changed, 4 insertions(+)
>
> diff --git a/kernel/irq/chip.c b/kernel/irq/chip.c
> index 29d6c7d070b4..88d1e054c6ea 100644
> --- a/kernel/irq/chip.c
> +++ b/kernel/irq/chip.c
> @@ -748,6 +748,8 @@ void handle_fasteoi_nmi(struct irq_desc *desc)
> unsigned int irq = irq_desc_get_irq(desc);
> irqreturn_t res;
>
> + kstat_incr_irqs_this_cpu(desc);
> +
This needs to be called with the desc lock taken, otherwise we're likely
to corrupt desc->tot_count.
But taking the desc lock is something we can't do in NMI context (
*spin_lock_irq*() won't prevent an NMI from happening).
> trace_irq_handler_entry(irq, action);
> /*
> * NMIs cannot be shared, there is only one action.
> @@ -962,6 +964,8 @@ void handle_percpu_devid_fasteoi_nmi(struct irq_desc
> *desc)
> unsigned int irq = irq_desc_get_irq(desc);
> irqreturn_t res;
>
> + __kstat_incr_irqs_this_cpu(desc);
> +
Looking at handle_percpu_irq(), I think this might be acceptable. But
does it make sense to only have kstats for percpu NMIs?
Cheers,
--
Julien Thierry
