On Thu, Apr 10, 2014 at 06:13:54PM +0900, Hidetoshi Seto wrote:
> [WHAT THIS PATCH PROPOSED]:
>
> To fix problem 1, this patch adds seqcount for NO_HZ idle
> accounting to avoid possible races between reader/writer.
>
> And to cope with problem 2, I introduced delayed iowait
> accounting to get approximate value without making observers
> to writers. Refer comment in patch for the detail.
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -407,15 +407,42 @@ static void tick_nohz_stop_idle(struct tick_sched *ts, ktime_t now)
> {
> ktime_t delta;
>
> + write_seqcount_begin(&ts->idle_sleeptime_seq);
> +
> /* Updates the per cpu time idle statistics counters */
> delta = ktime_sub(now, ts->idle_entrytime);
> +
> + /*
> + * Perform delayed iowait accounting:
> + *
> + * We account sleep time as iowait when nr_iowait of cpu indicates
> + * there are taskes blocked by io, at the end of idle (=here).
> + * It means we can not determine whether the sleep time will be idle
> + * or iowait on the fly.
> + * Therefore introduce a new rule:
> + * - basically observers assign delta to idle
> + * - if cpu find nr_iowait>0 at idle exit, accumulate delta as missed
> + * iowait, and account it in next turn of sleep instead.
> + * - if observer find accumulated iowait while cpu is in sleep, it
> + * can calculate proper value to be accounted.
> + */
> + if (ktime_compare(ts->iowait_pending, delta) > 0) {
> ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime, delta);
> + ts->iowait_pending = ktime_sub(ts->iowait_pending, delta);
> + } else {
> + ts->idle_sleeptime = ktime_add(ts->idle_sleeptime,
> + ktime_sub(delta, ts->iowait_pending));
> + ts->iowait_sleeptime = ktime_add(ts->iowait_sleeptime,
> + ts->iowait_pending);
> + ts->iowait_pending = ktime_set(0, 0);
> + }
> + if (nr_iowait_cpu(smp_processor_id()) > 0)
> + ts->iowait_pending = ktime_add(ts->iowait_pending, delta);
> +
> ts->idle_active = 0;
>
> + write_seqcount_end(&ts->idle_sleeptime_seq);
> +
> sched_clock_idle_wakeup_event(0);
> }
Why!? Both changelog and comment are silent on this. This doesn't appear
to make any sense nor really solve anything.
--
To unsubscribe from this list: send the line "unsubscribe stable" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html