On Wed, 6 May 2020 at 18:03, Valentin Schneider <valentin.schnei...@arm.com> wrote: > > > On 06/05/20 14:45, Vincent Guittot wrote: > >> But then we may skip an update if we goto abort, no? Imagine we have just > >> NOHZ_STATS_KICK, so we don't call any rebalance_domains(), and then as we > >> go through the last NOHZ CPU in the loop we hit need_resched(). We would > >> end in the abort part without any update to nohz.next_balance, despite > >> having accumulated relevant data in the local next_balance variable. > > > > Yes but on the other end, the last CPU has not been able to run the > > rebalance_domain so we must not move nohz.next_balance otherwise it > > will have to wait for at least another full period > > In fact, I think that we have a problem with current implementation > > because if we abort because local cpu because busy we might end up > > skipping idle load balance for a lot of idle CPUs > > > > As an example, imagine that we have 10 idle CPUs with the same > > rq->next_balance which equal nohz.next_balance. _nohz_idle_balance > > starts on CPU0, it processes idle lb for CPU1 but then has to abort > > because of need_resched. If we update nohz.next_balance like > > currently, the next idle load balance will happen after a full > > balance interval whereas we still have 8 CPUs waiting for running an > > idle load balance. > > > > My proposal also fixes this problem > > > > That's a very good point; so with NOHZ_BALANCE_KICK we can reduce > nohz.next_balance via rebalance_domains(), and otherwise we would only > increase it if we go through a complete for_each_cpu() loop in > _nohz_idle_balance(). > > That said, if for some reason we keep bailing out of the loop, we won't > push nohz.next_balance forward and thus may repeatedly nohz-balance only > the first few CPUs in the NOHZ mask. I think that can happen if we have > say 2 tasks pinned to a single rq, in that case nohz_balancer_kick() will > kick a NOHZ balance whenever now >= nohz.next_balance.
If we take my example above and we have CPU0 which is idle at every tick and selected as ilb_cpu but unluckily CPU0 has to abort before running ilb for CPU1 everytime, I agree that we can end up trying to run ilb on CPU0 at every tick without any success. We might consider to kick_ilb in _nohz_idle_balance if we have to abort to let another CPU handle the ilb