On Wed, Jul 16, 2014 at 02:15:50PM -0400, Pranith Kumar wrote:
>
> On 07/16/2014 01:24 PM, Paul E. McKenney wrote:
> >>> And of course if you don't actually specify a nohz_full= mask, then
> >>> tick_nohz_full_mask can be NULL at RCU initialization time, and if it
> >>> is also true that CONFIG_NO_HZ_FULL_ALL=n, this condition can persist
> >>> forever.
> >>>
> >> Hi Paul,
> >>
> >> The other location where tick_nohz_full_mask is being allocated is in
> >> tick_nohz_init_all(), called from tick_nohz_init(). rcu_init() is called
> >> before
> >> tick_nohz_init() in init/main.c. CONFIG_NO_HZ_FULL_ALL for allocation of
> >> the
> >> mask does not take effect when rcu_init() runs.
> >>
> >> So if nohz_full command line arg is not specified, tick_nohz_full_mask will
> >> always be NULL when rcu_init() runs. So just checking for
> >> tick_nohz_full_mask
> >> is NULL should be enough I guess.
> > Yep, that is indeed one of the conditions called out in the commit log
> > below.
> >
> > Thanx, Paul
> >
>
> Sorry, I was not clear.
>
> nohz_full= boot time parameter overrides the CONFIG_NO_HZ_FULL_ALL build time
> parameter.
> The check for !CONFIG_NO_HZ_FULL_ALL will disable masking out the cpus on
> which NOHZ is
> not enabled using the nohz_full boot time parameter, and hence you will
> enable callback
> offloading on all CPUs as was previously being done.
>
> So, if I have CONFIG_NO_HZ_FULL_ALL enabled at build time but pass nohz_cpu=1
> at boot
> time, all the CPUs will have callback offloading enabled.
>
> I think the following looks like what is needed in our case:
>
> diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
> index 00dc411..1123ed4 100644
> --- a/kernel/rcu/tree_plugin.h
> +++ b/kernel/rcu/tree_plugin.h
> @@ -2479,10 +2479,10 @@ static void __init rcu_spawn_nocb_kthreads(struct
> rcu_state *rsp)
>
> if (rcu_nocb_mask == NULL)
> return;
> -#if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL)
> +
> if (tick_nohz_full_running)
> cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
> -#endif /* #if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL)
> */
> +
> if (ls == -1) {
> ls = int_sqrt(nr_cpu_ids);
> rcu_nocb_leader_stride = ls;
Agreed, and I queued something very similar to this. Testing is in progress.
If it works, I will send it out. If you want to see it earlier, it is
at branch rcu/urgent on the -rcu tree.
Thanx, Paul
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majord...@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/
