On Sat, Jul 19, 2014 at 02:44:12AM +0200, Frederic Weisbecker wrote:
> Being able to trigger an empty IPI appears to be useless in the first
> place. Yet it is expected to be very useful for callers who just need
> to execute irq_enter() or irq_exit() to a remote target.
>
> More precisely this is going to be useful for the nohz subsystem which
> often needs a remote CPU to reschedule its tick when some state changed
> and require periodicity for any reason. Similar cases have been handled
> with random IPIs until now. But they surely bring unwanted overhead
> all along since they are all dedicated for specific tasks.
>
> Triggering an irq work/smp_call_function IPI should be enough to solve
> that. If competing and spurious IPIs become a problem, we can still
> optimize that later.
>
> Cc: Ingo Molnar <[email protected]>
> Cc: Paul E. McKenney <[email protected]>
> Cc: Peter Zijlstra <[email protected]>
> Cc: Steven Rostedt <[email protected]>
> Cc: Thomas Gleixner <[email protected]>
> Cc: Viresh Kumar <[email protected]>
> Signed-off-by: Frederic Weisbecker <[email protected]>

Reviewed-by: Paul E. McKenney <[email protected]>

> ---
> include/linux/irq_work.h | 1 +
> kernel/irq_work.c | 21 +++++++++++++++++++++
> 2 files changed, 22 insertions(+)
>
> diff --git a/include/linux/irq_work.h b/include/linux/irq_work.h
> index bf9422c..b2ad065 100644
> --- a/include/linux/irq_work.h
> +++ b/include/linux/irq_work.h
> @@ -36,6 +36,7 @@ bool irq_work_queue(struct irq_work *work);
>
> #ifdef CONFIG_SMP
> bool irq_work_queue_on(struct irq_work *work, int cpu);
> +void irq_work_void_on(int cpu);
> #endif
>
> void irq_work_run(void);
> diff --git a/kernel/irq_work.c b/kernel/irq_work.c
> index 4b0a890..36b7fb2 100644
> --- a/kernel/irq_work.c
> +++ b/kernel/irq_work.c
> @@ -81,6 +81,27 @@ bool irq_work_queue_on(struct irq_work *work, int cpu)
> return true;
> }
> EXPORT_SYMBOL_GPL(irq_work_queue_on);
> +
> +/**
> + * irq_work_void_on(): Run a void IRQ on the target
> + * @cpu: The cpu to run the IRQ on
> + *
> + * Run a void IRQ for its own sake on the target. It's generally
> + * useful for callers which want to run irq_enter() or irq_exit()
> + * on a remote CPU.
> + */
> +void irq_work_void_on(int cpu)
> +{
> + /*
> + * NOTE: we could optimize that and spare some IPIs
> + * after checking that raised_list isn't empty before raising.
> + * This can't be done properly without cmpxchg() though so
> + * it may make things worse after all. But lets leave that
> + * possibility open in case people report such issue in the
> + * future.
> + */
> + arch_send_call_function_single_ipi(cpu);
> +}
> #endif
>
> /* Enqueue the irq work @work on the current CPU */
> --
> 1.8.3.1
>

2014-07-25 21:27:30

by Frederic Weisbecker

[permalink] [raw]

Subject: Re: [PATCH 06/10] nohz: Define meaningful symbol for nohz full timekeeper

On Mon, Jul 21, 2014 at 11:12:55AM -0700, Paul E. McKenney wrote:
> On Sat, Jul 19, 2014 at 02:44:17AM +0200, Frederic Weisbecker wrote:
> > The nohz full timekeeper is always CPU 0. Lets add it to the list of
> > special tick_do_timer_cpu symbols for more self explanatory code.
> >
> > Cc: Ingo Molnar <[email protected]>
> > Cc: Paul E. McKenney <[email protected]>
> > Cc: Peter Zijlstra <[email protected]>
> > Cc: Steven Rostedt <[email protected]>
> > Cc: Thomas Gleixner <[email protected]>
> > Cc: Viresh Kumar <[email protected]>
> > Signed-off-by: Frederic Weisbecker <[email protected]>
>
> Sounds like Nicolas Pitre would like to be able to specify the default,
> perhaps at build time, but aside from that requested enhancement:

Ok. That's outside the scope of this patchset though as it doesn't change much things
in this regard: CPU 0 has been the timekeeper since the early days of nohz full.
And we still need a default timekeeper in any case, whether we can override it or not
in the future. So we still need that patch.

Then if big.LITTLE people are really interested in nohz full, we can still enhance
that in the future.

Thanks!