Re: [PATCH v3] sched/idle: disable tick in idle=poll idle entry

[Marcelo Tosatti <mtosatti@redhat.com>]

On Mon, Nov 03, 2025 at 01:30:23PM +0100, Peter Zijlstra wrote:
> On Mon, Nov 03, 2025 at 08:48:14AM -0300, Marcelo Tosatti wrote:
> > 
> > Commit a5183862e76fdc25f36b39c2489b816a5c66e2e5
> > ("tick/nohz: Conditionally restart tick on idle exit") allows
> 
> Quoting a commit usually shortens the hash to 12 charters, no?
> 
> > a nohz_full CPU to enter idle and return from it with the
> > scheduler tick disabled (since the tick might be undesired noise).
> > 
> > The idle=poll case still unconditionally restarts the tick when entering
> > idle.
> > 
> > To reduce the noise for that case as well, stop the tick when entering
> > idle, for the idle=poll case.
> > 
> > Change tick_nohz_full_kick_cpu to set NEED_RESCHED bit, to handle the
> > case where a new timer is added from an interrupt. This breaks out of
> > cpu_idle_poll and rearms the timer if necessary.
> > 
> > Signed-off-by: Marcelo Tosatti <mtosatti@redhat.com>
> > 
> > ---
> > 
> > v3: Add comment with proper explanation (Frederic Weisbecker)
> >     Add signed-off-by			(Thomas Gleixner)
> > v2: Handle the case where a new timer is added from an interrupt (Frederic Weisbecker)
> > 
> >  include/linux/sched.h    |    2 ++
> >  kernel/sched/core.c      |   10 ++++++++++
> >  kernel/sched/idle.c      |    2 +-
> >  kernel/time/tick-sched.c |    7 +++++++
> >  4 files changed, 20 insertions(+), 1 deletion(-)
> > 
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index cbb7340c5866..1f6938dc20cd 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -2428,4 +2428,6 @@ extern void migrate_enable(void);
> >  
> >  DEFINE_LOCK_GUARD_0(migrate, migrate_disable(), migrate_enable())
> >  
> > +void set_tif_resched_if_polling(int cpu);
> > +
> >  #endif
> > diff --git a/kernel/sched/core.c b/kernel/sched/core.c
> > index f1ebf67b48e2..f0b84600084b 100644
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -988,6 +988,11 @@ static bool set_nr_if_polling(struct task_struct *p)
> >  	return true;
> >  }
> >  
> > +void set_tif_resched_if_polling(int cpu)
> > +{
> > +	set_nr_if_polling(cpu_rq(cpu)->idle);
> > +}
> > +
> >  #else
> >  static inline bool set_nr_and_not_polling(struct thread_info *ti, int tif)
> >  {
> > @@ -999,6 +1004,11 @@ static inline bool set_nr_if_polling(struct task_struct *p)
> >  {
> >  	return false;
> >  }
> > +
> > +void set_tif_resched_if_polling(int cpu)
> > +{
> > +	set_tsk_need_resched(cpu_rq(cpu)->idle);
> > +}
> >  #endif
> >  
> >  static bool __wake_q_add(struct wake_q_head *head, struct task_struct *task)
> > diff --git a/kernel/sched/idle.c b/kernel/sched/idle.c
> > index c39b089d4f09..428c2d1cbd1b 100644
> > --- a/kernel/sched/idle.c
> > +++ b/kernel/sched/idle.c
> > @@ -324,7 +324,7 @@ static void do_idle(void)
> >  		 * idle as we know that the IPI is going to arrive right away.
> >  		 */
> >  		if (cpu_idle_force_poll || tick_check_broadcast_expired()) {
> > -			tick_nohz_idle_restart_tick();
> > +			tick_nohz_idle_stop_tick();
> >  			cpu_idle_poll();
> >  		} else {
> >  			cpuidle_idle_call();
> > diff --git a/kernel/time/tick-sched.c b/kernel/time/tick-sched.c
> > index c527b421c865..9ec51da49591 100644
> > --- a/kernel/time/tick-sched.c
> > +++ b/kernel/time/tick-sched.c
> > @@ -408,6 +408,13 @@ void tick_nohz_full_kick_cpu(int cpu)
> >  	if (!tick_nohz_full_cpu(cpu))
> >  		return;
> >  
> > +	/*
> > +	 * When idle=poll, with the tick disabled (therefore idle CPU looping
> > +	 * at cpu_idle_poll), if a new timer is added from an interrupt,
> > +	 * the cpu_idle_poll only exits when TIF_NEED_RESCHED gets set.
> > +	 */
> > +	set_tif_resched_if_polling(cpu);
> > +
> >  	irq_work_queue_on(&per_cpu(nohz_full_kick_work, cpu), cpu);
> >  }
> 
> I'm confused. Why is this here and not in tick_nohz_start_idle() or
> something?
> 
> 

Hi Peter,

The codepath being followed is:

enqueue_timer -> trigger_dyntick_cpu -> wake_up_nohz_cpu ->
wake_up_full_nohz_cpu -> tick_nohz_full_kick_cpu ->
set_tif_resched_if_polling. 

So we only set the PF_RESCHED bit if there is a pending timer
on the CPU.

Calling unconditionally from tick_nohz_start_idle seems strange:

/**
 * tick_nohz_idle_enter - prepare for entering idle on the current CPU
 *
 * Called when we start the idle loop.
 */
void tick_nohz_idle_enter(void)
{
        struct tick_sched *ts;
        
        lockdep_assert_irqs_enabled();

        local_irq_disable();

        ts = this_cpu_ptr(&tick_cpu_sched);

        WARN_ON_ONCE(ts->timer_expires_base);

        tick_sched_flag_set(ts, TS_FLAG_INIDLE);
        tick_nohz_start_idle(ts);

        local_irq_enable();
}

Can test for TS_FLAG_INIDLE before calling set_tif_resched_if_polling 
(but seems not necessary since tick_nohz_full_kick_cpu will wake up the 
CPU anyway and is a slow path (timer addition)).

What do you think?