Re: [PATCH 6/7 v3] sched/fair: trigger the update of blocked load on newly idle cpu

[Vincent Guittot]

On Wed, 17 Feb 2021 at 12:51, Valentin Schneider
<valentin.schneider@xxxxxxx> wrote:
>
> On 15/02/21 16:02, Vincent Guittot wrote:
> > On Fri, 12 Feb 2021 at 20:19, Valentin Schneider
> > <valentin.schneider@xxxxxxx> wrote:
> >> I don't think there is anything inherently wrong with it - the
> >> nohz_idle_balance() call resulting from the kick_ilb() IPI will just bail
> >> out due to the flags being cleared here. This wasn't immediately clear to
> >> me however.
> >
> > In fact, I forgot to replace the WARN_ON in nohz_csd_func() by a
> > simple return as reported by kernel test robot / oliver.sang@xxxxxxxxx
> >
>
> Can't that actually be a problem? kick_ilb() says:
>
>          * Access to rq::nohz_csd is serialized by NOHZ_KICK_MASK; he who sets
>          * the first flag owns it; cleared by nohz_csd_func().
>
> So if you have:
>
>   kick_ilb() -> kicks CPU42
>
> And then said CPU42 goes through, before nohz_csd_func(),:
>
>   do_idle() -> nohz_run_idle_balance()
>
> you could have yet another CPU do:
>
>   kick_ilb() -> kicks CPU42
>
> which would break rq->nohz_csd serialization.

Yeah there are ever further problems and I get some rcu_sched log on
my large server with one benchmark with one specific parameter which I
can't reproduce on my smaller system. Right now, I'm working on making
both exclusive which should be mainly about testing if this_cpu is set
in nohz.idle_cpus_mask

>
> >>
> >> > +}
> >> > +