Re: [PATCH] tick/nohz: Fix wrong user and system time accouting against vtime sampling

[Frederic Weisbecker]

On Mon, Apr 10, 2017 at 05:45:56PM +0200, Thomas Gleixner wrote:
> On Wed, 5 Apr 2017, Wanpeng Li wrote:
> > +	/*
> > +	 * Offset the tick to avert jiffies_lock contention, and all ticks
> > +	 * alignment in order that the vtime sampling does not end up "in
> > +	 * phase" with the jiffies incrementing.
> > +	 */
> > +	if (sched_skew_tick || tick_nohz_full_enabled()) {
> >  		u64 offset = ktime_to_ns(tick_period) >> 1;
> >  		do_div(offset, num_possible_cpus());
> >  		offset *= smp_processor_id();
> 
> That's not a fix, that's just papering over the problem.
> 
>        offset = 1ms / 2 = 500us = 500000ns;
>        offset /= 144 = 3472ns
> 
> So CPU0 and CPU1 ticks are ~3 microseconds apart. That merily reduces the
> probability of the issue, but does not prevent it.

I worried about it but didn't realize it could be that tight.

So the alternative is the solution involving sched_clock() as the source for
cputime. Wanpeng Li could you please resubmit your patch that does that?

Thanks.