Re: [PATCH v5] nohz: set isolcpus when nohz_full is set

[Chris Metcalf]

On 04/09/2015 09:05 PM, Mike Galbraith wrote:
> On Thu, 2015-04-09 at 19:12 +0200, Peter Zijlstra wrote:
> > On Thu, Apr 09, 2015 at 12:59:39PM -0400, Chris Metcalf wrote:
> > > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > > index 6d77432e14ff..18a961b9beba 100644
> > > --- a/include/linux/sched.h
> > > +++ b/include/linux/sched.h
> > > @@ -323,6 +323,7 @@ struct task_struct;
> > >   extern int lockdep_tasklist_lock_is_held(void);
> > >   #endif /* #ifdef CONFIG_PROVE_RCU */
> > >   
> > > +extern void sched_isolated_map_add(const struct cpumask *);
> > >   extern void sched_init(void);
> > >   extern void sched_init_smp(void);
> > >   extern asmlinkage void schedule_tail(struct task_struct *prev);
> > > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > > index f0f831e8a345..b055c5e0e65c 100644
> > > --- a/kernel/sched/core.c
> > > +++ b/kernel/sched/core.c
> > > @@ -5824,6 +5824,11 @@ static int __init isolated_cpu_setup(char
> > > *str)
> > >   
> > >   __setup("isolcpus=", isolated_cpu_setup);
> > >   
> > > +void sched_isolated_map_add(const struct cpumask *cpumask)
> > > +{
> > > +   cpumask_or(cpu_isolated_map, cpu_isolated_map, cpumask);
> > > +}
> > > +
> > >   struct s_data {
> > >      struct sched_domain ** __percpu sd;
> > >      struct root_domain      *rd;
> > > diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> > > index a4c4edac4528..b0092d02ca3f 100644
> > > --- a/kernel/time/tick-sched.c
> > > +++ b/kernel/time/tick-sched.c
> > > @@ -385,6 +385,9 @@ void __init tick_nohz_init(void)
> > >      for_each_cpu(cpu, tick_nohz_full_mask)
> > >              context_tracking_cpu_set(cpu);
> > >   
> > > +   /* It's not meaningful to be nohz without disabling the
> > > scheduler. */
> > > +   sched_isolated_map_add(tick_nohz_full_mask);
> > > +
> > >      cpu_notifier(tick_nohz_cpu_down_callback, 0);
> > >      pr_info("NO_HZ: Full dynticks CPUs: %*pbl.\n",
> > >              cpumask_pr_args(tick_nohz_full_mask));
> > Right, this could work. Although I would suggest adding a comment
> > somewhere that we should be careful with init order. I checked, this
> > appears to be ordered right, but...
> I'd embed it in domain construction, that way it'd be ready for the
> day nohz_full becomes dynamic, and people can start using cpusets to
> set up /tear down isolated sets on the fly.

So, move it to the top of build_sched_domains(), and then for
every "const struct cpumask *cpu_map" argument, create a
temporary cpu_map so we can mask out the nohz_full cores?

The problem is, we already allow partition_sched_domains()
to override "isolcpus=", so it seems appropriate that you should
be able to override "nohz_full=" in the same way, which my
current patch (v6) does.

So I think the proposed solution is certainly no worse than what
we have now in terms of a future migration to cpusets.

--
Chris Metcalf, EZChip Semiconductor
http://www.ezchip.com

--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/