Re: [PATCH v2] hrtimers: add fast path to hrtimer_get_next_event

[Thomas Gleixner <tglx@linutronix.de>]

On Mon, 9 Jun 2014, Stanislav Fomichev wrote:

> In hrtimer_get_next_event we unconditionally lock/unlock spinlock, even if it's
> not required (hrtimer_hres_active() != 0). This patch adds fast path
> when highres is active so we don't execute unnecessary operations.
> 
> We can safely do lockless check because:
> - hrtimer_get_next_event is always called with interrupts disabled;
> - we may switch to hres only from softirq handler with disabled interrupts.
> 
> Because we only care about hres_active which may be changed only from
> local CPU, we can use interrupt context for synchronization.
> 
> run_timer_softirq
>   hrtimer_run_pending
>     hrtimer_switch_to_hres
>       local_irq_save <-
>       base->hres_active = 0
> 
> tick_nohz_idle_enter
>   local_irq_disable <-
>   __tick_nohz_idle_enter
>     tick_nohz_stop_sched_tick
>       get_next_timer_interrupt
>         cmp_next_hrtimer_event
>           hrtimer_get_next_event
>           check base->hres_active
> 
> irq_exit <- irq context
>   tick_irq_exit
>     tick_nohz_irq_exit
>       tick_nohz_full_stop_tick
>         tick_nohz_stop_sched_tick
>             ... <see above>
>       __tick_nohz_idle_enter
>         ... <see above>

Sigh. Is anybody actually reading and trying to understand what I
write in reviews?

> Also the lockless check wants a comment why it is correct.

Is it that hard? A comment is NOT a uberlenghty explanation in the
changelog. A comment starts with /* and ends with */ and is in the
code.

	/*
	 * Called with interrupts disabled and therefor
	 * protected against a switch to high resolution mode.
	 */
 
That's a comment, right?

Thanks,

	tglx