Re: [PATCH RFC] sched/fair: Avoid unnecessary IPIs for ILB

[Vineeth Pillai]

On Sun, Oct 8, 2023 at 1:35 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
>
[...snip...]
> > The patch does make _nohz_idle_balance() run more parallel, as previously
> > it would be generally run by the first-idle CPU in nohz.idle_cpus_mask (at
> > least for next_balance updates), but I think it's still SMP-safe, as all
> > key data structure updates are already rq-locked AFAICS.
>
> One thing I am confused about in the original code is:
>
> tick_nohz_idle_stop_tick() is what sets the nohz.idle_cpus_mask.
> However, nohz_run_idle_balance() is called before that can happen, in
> do_idle(). So it is possible that NOHZ_NEWILB_KICK is set for a CPU but it is
> not yet in the mask.
>
> So will this code in _nohz_idle_balance() really run in such a scenario?
>
>                 if (flags & NOHZ_STATS_KICK)
>                         has_blocked_load |= update_nohz_stats(rq);
>
> AFAICS, this loop may not select the CPU due to its absence from the mask:
>         for_each_cpu_wrap(balance_cpu,  nohz.idle_cpus_mask, this_cpu+1)
>
I have traced this a bit further. As Joel mentioned, the
nohz.idle_cpus_mask shouldn't contain this cpu when
nohz_run_idle_balance () is called from do_idle(), but on tracing I
have seen that it does have it mostly with HIGHRES. And I feel this is
a bug. We call nohz_balance_enter_idle() when we turn off the tick,
but we don't always call nohz_balance_exit_idle() when we turn the
tick back on. We call it only on the next tick on this cpu in
nohz_balancer_kick. If a wakeup happens on this cpu while the tick is
off, we re-enable the tick, but do not remove ourselves from the
nohz.idle_cpus_mask. So, ILB will consider this cpu to be a valid pick
until the next tick on this cpu where it gets removed. I am not sure
if this is intentional.

If this is a bug and we fix it by calling nohz_balance_exit_idle
during restart_tick, then we might not probably need NOHZ_NEWIDLE_KICK
flag and could use NOHZ_STATS_KICK as there will not be any overlap
between nohz_run_idle_balance and nohz_idle_balance.

Thanks,
Vineeth