Re: [patch 7/8] hrtimer: Avoid unnecessary SMP function calls in clock_was_set()

[Peter Zijlstra]

On Tue, Apr 27, 2021 at 10:25:44AM +0200, Thomas Gleixner wrote:
> --- a/kernel/time/hrtimer.c
> +++ b/kernel/time/hrtimer.c
> @@ -896,11 +896,42 @@ static void hrtimer_reprogram(struct hrt
>   */
>  void clock_was_set(unsigned int bases)
>  {
> +	cpumask_var_t mask;
> +	int cpu;
> +
>  	if (!hrtimer_hres_active() && !tick_nohz_active)
>  		goto out_timerfd;
>  
> -	/* Retrigger the CPU local events everywhere */
> -	on_each_cpu(retrigger_next_event, NULL, 1);
> +	if (!zalloc_cpumask_var(&mask, GFP_KERNEL)) {
> +		on_each_cpu(retrigger_next_event, NULL, 1);

This will violate NOHZ_FULL;

> +		goto out_timerfd;
> +	}
> +
> +	/* Avoid interrupting CPUs if possible */
> +	cpus_read_lock();
> +	for_each_online_cpu(cpu) {
> +		struct hrtimer_cpu_base *cpu_base = &per_cpu(hrtimer_bases, cpu);
> +		unsigned long flags;
> +
> +		raw_spin_lock_irqsave(&cpu_base->lock, flags);
> +		/*
> +		 * Only send the IPI when there are timers queued in one of
> +		 * the affected clock bases. Otherwise update the base
> +		 * remote to ensure that the next enqueue of a timer on
> +		 * such a clock base will see the correct offsets.
> +		 */
> +		if (cpu_base->active_bases & bases)
> +			cpumask_set_cpu(cpu, mask);
> +		else
> +			hrtimer_update_base(cpu_base);
> +		raw_spin_unlock_irqrestore(&cpu_base->lock, flags);
> +	}
> +
> +	preempt_disable();
> +	smp_call_function_many(mask, retrigger_next_event, NULL, 1);

The sane option is:

	smp_call_function_many_cond(cpu_online_mask, retrigger_next_event,
				    NULL, SCF_WAIT, update_needs_ipi);

Which does all of the above, but better.


> +	preempt_enable();
> +	cpus_read_unlock();
> +	free_cpumask_var(mask);
>  
>  out_timerfd:
>  	timerfd_clock_was_set();
>