Re: [BUG nohz]: wrong user and system time accounting

[Frederic Weisbecker]

On Thu, Mar 30, 2017 at 02:47:11PM +0800, Wanpeng Li wrote:
> Cc Peterz, Thomas,
> 2017-03-30 12:27 GMT+08:00 Mike Galbraith <efault@xxxxxx>:
> > On Wed, 2017-03-29 at 16:08 -0400, Rik van Riel wrote:
> >
> >> In other words, the tick on cpu0 is aligned
> >> with the tick on the nohz_full cpus, and
> >> jiffies is advanced while the nohz_full cpus
> >> with an active tick happen to be in kernel
> >> mode?
> >
> > You really want skew_tick=1, especially on big boxen.
> >
> >> Frederic, can you think of any reason why
> >> the tick on nohz_full CPUs would end up aligned
> >> with the tick on cpu0, instead of running at some
> >> random offset?
> >
> > (I or low rq->clock bits as crude NOHZ collision avoidance)
> >
> >> A random offset, or better yet a somewhat randomized
> >> tick length to make sure that simultaneous ticks are
> >> fairly rare and the vtime sampling does not end up
> >> "in phase" with the jiffies incrementing, could make
> >> the accounting work right again.
> >
> > That improves jitter, especially on big boxen.  I have an 8 socket box
> > that thinks it's an extra large PC, there, collision avoidance matters
> > hugely.  I couldn't reproduce bean counting woes, no idea if collision
> > avoidance will help that.
> 
> So I implement two methods, one is from Rik's random offset proposal
> through skew tick, the other one is from Frederic's proposal and it is
> the same as my original idea through use nanosecond granularity to
> check deltas but only perform an actual cputime update when that delta
> >= TICK_NSEC. Both methods can solve the bug which Luiz reported.
> Peterz, Thomas, any ideas?
> 
> --------------------------->8-------------------------------------------------------------
> 
> skew tick:
> 
> diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> index 7fe53be..9981437 100644
> --- a/kernel/time/tick-sched.c
> +++ b/kernel/time/tick-sched.c
> @@ -1198,7 +1198,11 @@ void tick_setup_sched_timer(void)
>      hrtimer_set_expires(&ts->sched_timer, tick_init_jiffy_update());
> 
>      /* Offset the tick to avert jiffies_lock contention. */
> +#ifdef CONFIG_NO_HZ_FULL
> +    if (sched_skew_tick || tick_nohz_full_running) {
> +#else
>      if (sched_skew_tick) {
> +#endif

Please rather use tick_nohz_full_enabled() to avoid ifdeffery.

>          u64 offset = ktime_to_ns(tick_period) >> 1;
>          do_div(offset, num_possible_cpus());
>          offset *= smp_processor_id();

If it works, we may want to take that solution, likely less performance sensitive
than using sched_clock(). In fact sched_clock() is fast, especially as we require it to
be stable for nohz_full, but using it involves costly conversion back and forth to jiffies.

> 
> -------------------------------------->8-----------------------------------------------------
> 
> use nanosecond granularity to check deltas but only perform an actual
> cputime update when that delta >= TICK_NSEC.
> 
> diff --git a/kernel/sched/cputime.c b/kernel/sched/cputime.c
> index f3778e2b..f1ee393 100644
> --- a/kernel/sched/cputime.c
> +++ b/kernel/sched/cputime.c
> @@ -676,18 +676,21 @@ void thread_group_cputime_adjusted(struct
> task_struct *p, u64 *ut, u64 *st)
>  #ifdef CONFIG_VIRT_CPU_ACCOUNTING_GEN
>  static u64 vtime_delta(struct task_struct *tsk)
>  {
> -    unsigned long now = READ_ONCE(jiffies);
> +    u64 now = local_clock();

I fear we need a global clock, because the reader (task_cputime()) needs
to compute the delta and therefore use the same clock from any CPU.

Or we can use the local_clock() but the reader must access the same.

So there would be vtime_delta_writer() which uses local_clock and stores
the current CPU to tsk->vtime_cpu (under the vtime_seqcount). And then
vtime_delta_reader() which calls sched_clock_cpu(tsk->vtime_cpu) which
is protected by vtime_seqcount as well.

Although those sched_clock_cpu() things seem to only matter when the
sched_clock() is unstable. And that stability is a condition for nohz_full
to work anyway. So probably sched_clock() alone would be enough.

Thanks.