Re: [PATCH 4/5] sched/isolation: Residual 1Hz scheduler tick offload

[Frederic Weisbecker <frederic@kernel.org>]

On Fri, Jan 12, 2018 at 02:22:58PM -0500, Luiz Capitulino wrote:
> On Thu,  4 Jan 2018 05:25:36 +0100
> Frederic Weisbecker <frederic@kernel.org> wrote:
> 
> > When a CPU runs in full dynticks mode, a 1Hz tick remains in order to
> > keep the scheduler stats alive. However this residual tick is a burden
> > for bare metal tasks that can't stand any interruption at all, or want
> > to minimize them.
> > 
> > Adding the boot parameter "isolcpus=nohz_offload" will now outsource
> > these scheduler ticks to the global workqueue so that a housekeeping CPU
> > handles that tick remotely.
> > 
> > Note it's still up to the user to affine the global workqueues to the
> > housekeeping CPUs through /sys/devices/virtual/workqueue/cpumask or
> > domains isolation.
> > 
> > Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
> > Cc: Chris Metcalf <cmetcalf@mellanox.com>
> > Cc: Christoph Lameter <cl@linux.com>
> > Cc: Luiz Capitulino <lcapitulino@redhat.com>
> > Cc: Mike Galbraith <efault@gmx.de>
> > Cc: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
> > Cc: Peter Zijlstra <peterz@infradead.org>
> > Cc: Rik van Riel <riel@redhat.com>
> > Cc: Thomas Gleixner <tglx@linutronix.de>
> > Cc: Wanpeng Li <kernellwp@gmail.com>
> > Cc: Ingo Molnar <mingo@kernel.org>
> > ---
> >  kernel/sched/core.c      | 88 ++++++++++++++++++++++++++++++++++++++++++++++--
> >  kernel/sched/isolation.c |  4 +++
> >  kernel/sched/sched.h     |  2 ++
> >  3 files changed, 91 insertions(+), 3 deletions(-)
> > 
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index d72d0e9..b964890 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -3052,9 +3052,14 @@ void scheduler_tick(void)
> >   */
> >  u64 scheduler_tick_max_deferment(void)
> >  {
> > -	struct rq *rq = this_rq();
> > -	unsigned long next, now = READ_ONCE(jiffies);
> > +	struct rq *rq;
> > +	unsigned long next, now;
> >  
> > +	if (!housekeeping_cpu(smp_processor_id(), HK_FLAG_TICK_SCHED))
> > +		return ktime_to_ns(KTIME_MAX);
> > +
> > +	rq = this_rq();
> > +	now = READ_ONCE(jiffies);
> >  	next = rq->last_sched_tick + HZ;
> >  
> >  	if (time_before_eq(next, now))
> > @@ -3062,7 +3067,82 @@ u64 scheduler_tick_max_deferment(void)
> >  
> >  	return jiffies_to_nsecs(next - now);
> >  }
> > -#endif
> > +
> > +struct tick_work {
> > +	int			cpu;
> > +	struct delayed_work	work;
> > +};
> > +
> > +static struct tick_work __percpu *tick_work_cpu;
> > +
> > +static void sched_tick_remote(struct work_struct *work)
> > +{
> > +	struct delayed_work *dwork = to_delayed_work(work);
> > +	struct tick_work *twork = container_of(dwork, struct tick_work, work);
> > +	int cpu = twork->cpu;
> > +	struct rq *rq = cpu_rq(cpu);
> > +	struct rq_flags rf;
> > +
> > +	/*
> > +	 * Handle the tick only if it appears the remote CPU is running
> > +	 * in full dynticks mode. The check is racy by nature, but
> > +	 * missing a tick or having one too much is no big deal.
> > +	 */
> > +	if (!idle_cpu(cpu) && tick_nohz_tick_stopped_cpu(cpu)) {
> > +		rq_lock_irq(rq, &rf);
> > +		update_rq_clock(rq);
> > +		rq->curr->sched_class->task_tick(rq, rq->curr, 0);
> > +		rq_unlock_irq(rq, &rf);
> > +	}
> 
> OK, so this executes task_tick() remotely. What about account_process_tick()?
> Don't we need it as well?

Nope, tasks in nohz_full mode have their special accounting that doesn't
rely on the tick.

> 
> In particular, when I run a hog application on a nohz_full core configured
> with tick offload, I can see in top that the CPU usage goes from 100%
> to idle for a few seconds every couple of seconds. Could this be related?
> 
> Also, in my testing I'm sometimes seeing the tick. Sometimes at 10 or
> 20 seconds interval. Is this expected? I'll dig deeper next week.

That's expected, see the changelog: the offload is not affine by default.
You need to either also isolate the domains:

    isolcpus=nohz_offload,domain

or tweak the workqueue cpumask through:

    /sys/devices/virtual/workqueue/cpumask

Thanks.