Re: [PATCH 2/3] sched: Correctly handle nohz ticks cpu load accounting

[Peter Zijlstra]

On Fri, Apr 08, 2016 at 03:07:12AM +0200, Frederic Weisbecker wrote:
> +void cpu_load_update_nohz_start(void)
>  {
>  	struct rq *this_rq = this_rq();
> +
> +	/*
> +	 * This is all lockless but should be fine. If weighted_cpuload changes
> +	 * concurrently we'll exit nohz. And cpu_load write can race with
> +	 * cpu_load_update_idle() but both updater would be writing the same.
> +	 */
> +	this_rq->cpu_load[0] = weighted_cpuload(cpu_of(this_rq));
> +}

There is more to this; this also updates ->cpu_load[0] at possibly much
higher frequency than we've done before, while not updating the other
->cpu_load[] members.

Now, I'm not sure we care, but it is a bit odd.

> +/*
> + * Account the tickless load in the end of a nohz frame.
> + */
> +void cpu_load_update_nohz_stop(void)
> +{
>  	unsigned long curr_jiffies = READ_ONCE(jiffies);
> +	struct rq *this_rq = this_rq();
> +	unsigned long load;
>  
>  	if (curr_jiffies == this_rq->last_load_update_tick)
>  		return;
>  
> +	load = weighted_cpuload(cpu_of(this_rq));
>  	raw_spin_lock(&this_rq->lock);
> +	cpu_load_update_nohz(this_rq, curr_jiffies, load);
>  	raw_spin_unlock(&this_rq->lock);
>  }

And this makes us take rq->lock when waking from nohz; a bit
unfortunate. Do we really need this though? Will not a tick be
forthcoming real-soon-now?