[PATCH RFC] irq: Set CPU affinity as soon as the irq thread is created

[PATCH RFC] irq: Set CPU affinity as soon as the irq thread is created

[Sankara Muthukrishnan]

Hi there,

As irq_thread_check_affinity is called inside the while loop in the
IRQ thread, the core affinity is set only when an interrupt happens.
It does not seem to be a bad idea to set the core affinity soon after
the thread is created also (as shown in the patch).
irq_thread_check_affinity calls kmalloc_node that can sleep. In an
embedded target where IRQ affinities are pre-configured during
boot-time (and not modified through /proc/irq/../smp_affinity) during
run-time, the extra latency due to irq_thread_check_affinity can be
avoided during processing of the first interrupt with this patch.
Thoughts?

diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
index 87dc053..bf2cdc2 100644
--- a/kernel/irq/manage.c
+++ b/kernel/irq/manage.c
@@ -791,6 +791,7 @@ static int irq_thread(void *data)

        sched_setscheduler(current, SCHED_FIFO, &param);
        current->irqaction = action;
+       irq_thread_check_affinity(desc, action);

        while (!irq_wait_for_interrupt(action)) {

- Sankara

Re: [PATCH RFC] irq: Set CPU affinity as soon as the irq thread is created

[Steven Rostedt]

On Wed, 2012-09-05 at 18:07 -0500, Sankara Muthukrishnan wrote:
> Hi there,

This also affects mainline. Please Cc LKML (as I did).

> 
> As irq_thread_check_affinity is called inside the while loop in the
> IRQ thread, the core affinity is set only when an interrupt happens.
> It does not seem to be a bad idea to set the core affinity soon after
> the thread is created also (as shown in the patch).
> irq_thread_check_affinity calls kmalloc_node that can sleep.

Note, it only calls kmalloc if you enable CONFIG_CPUMASK_OFFSTACK, which
you probably don't on an embedded device. But you might. Your embedded
device may have more than 32 or 64 CPUs.

I'd be more worried about calling set_cpus_allowed_ptr() than kmalloc().

>  In an
> embedded target where IRQ affinities are pre-configured during
> boot-time (and not modified through /proc/irq/../smp_affinity) during
> run-time, the extra latency due to irq_thread_check_affinity can be
> avoided during processing of the first interrupt with this patch.
> Thoughts?
> 
> diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
> index 87dc053..bf2cdc2 100644
> --- a/kernel/irq/manage.c
> +++ b/kernel/irq/manage.c
> @@ -791,6 +791,7 @@ static int irq_thread(void *data)
> 
>         sched_setscheduler(current, SCHED_FIFO, &param);
>         current->irqaction = action;
> +       irq_thread_check_affinity(desc, action);
> 
>         while (!irq_wait_for_interrupt(action)) {

I see no problem with this patch.

Acked-by: Steven Rostedt <rostedt@goodmis.org>

-- Steve

Re: [PATCH RFC] irq: Set CPU affinity as soon as the irq thread is created

[Thomas Gleixner]

On Wed, 5 Sep 2012, Steven Rostedt wrote:

> On Wed, 2012-09-05 at 18:07 -0500, Sankara Muthukrishnan wrote:
> > Hi there,
> 
> This also affects mainline. Please Cc LKML (as I did).
> 
> > 
> > As irq_thread_check_affinity is called inside the while loop in the
> > IRQ thread, the core affinity is set only when an interrupt happens.
> > It does not seem to be a bad idea to set the core affinity soon after
> > the thread is created also (as shown in the patch).
> > irq_thread_check_affinity calls kmalloc_node that can sleep.
> 
> Note, it only calls kmalloc if you enable CONFIG_CPUMASK_OFFSTACK, which
> you probably don't on an embedded device. But you might. Your embedded
> device may have more than 32 or 64 CPUs.
> 
> I'd be more worried about calling set_cpus_allowed_ptr() than kmalloc().
> 
> >  In an
> > embedded target where IRQ affinities are pre-configured during
> > boot-time (and not modified through /proc/irq/../smp_affinity) during
> > run-time, the extra latency due to irq_thread_check_affinity can be
> > avoided during processing of the first interrupt with this patch.
> > Thoughts?

Yes, this needs a proper Signed-off-by and an update against current
mainline.

Thanks,

	tglx

> > diff --git a/kernel/irq/manage.c b/kernel/irq/manage.c
> > index 87dc053..bf2cdc2 100644
> > --- a/kernel/irq/manage.c
> > +++ b/kernel/irq/manage.c
> > @@ -791,6 +791,7 @@ static int irq_thread(void *data)
> > 
> >         sched_setscheduler(current, SCHED_FIFO, &param);
> >         current->irqaction = action;
> > +       irq_thread_check_affinity(desc, action);
> > 
> >         while (!irq_wait_for_interrupt(action)) {
> 
> I see no problem with this patch.
> 
> Acked-by: Steven Rostedt <rostedt@goodmis.org>
> 
> -- Steve
> 
> 
> --
> To unsubscribe from this list: send the line "unsubscribe linux-rt-users" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
>