Re: [PATCH] timers/nohz: Update nohz load even if tick already stopped

[Peter Zijlstra]

On Tue, Nov 05, 2019 at 01:43:51PM +0100, Peter Zijlstra wrote:
> On Tue, Nov 05, 2019 at 01:30:58AM -0600, Scott Wood wrote:
> > The warning is due to kernel/sched/idle.c not updating curr->se.exec_start.
> 
> Ah, indeed so.
> 
> > While debugging I noticed an issue with a particular load pattern.  The CPU
> > goes non-nohz for a brief time at an interval very close to twice 
> > tick_period.  When the tick is started, the timer expiration is more than
> > tick_period in the past, so hrtimer_forward() tries to catch up by adding
> > 2*tick_period to the expiration.  Then the tick is stopped before that new
> > expiration, and when the tick is woken up the expiry is again advanced by
> > 2*tick_period with the timer never actually running.  sched_tick_remote()
> > does fire every second, but there are streaks of several seconds where it
> > keeps catching the CPU in a non-nohz state, so neither the normal nor remote
> > ticks are calling calc_load_nohz_remote().
> > 
> > Is there a reason to not just remove the hrtimer_forward() from
> > tick_nohz_restart(), letting the timer fire if it's in the past, which will
> > take care of doing hrtimer_forward()?
> 
> I'll have to look into that. I always get confused by all that nohz code
> :/
> 
> > As for the warning in sched_tick_remote(), it seems like a test for time
> > since the last tick on this cpu (remote or otherwise) would be better than
> > relying on curr->se.exec_start, in order to detect things like this.
> 
> I don't think we have a timestamp that is shared between the remote and
> local tick. Also, there is a reason this warning uses the task time
> accounting, there used to be (as in, I can't find it in a hurry) code
> that could not deal with >u32 (~4s) clock updates.
> 
> The below should have idle keep the timestamp up-to-date. Keeping
> accurate idle->se.sum_exec_runtime doesn't seem too interesting, the
> idle code already keeps track of total idle times.
> 
> ---

The obvious alternative is something like:

	if (rq->curr != rq->idle) {
		s64 delta = rq_clock_task(rq) - curr->se.exec_start;
		WARN_ON_ONCE(delta > 3ULL * NSEC_PER_SEC);
	}

Which would avoid polluting the idle path with that extra assignment.