Re: [PATCH 4/4] sched: Exclude CPU boot code from PF_IDLE area

[Frederic Weisbecker]

Le Wed, Oct 25, 2023 at 10:48:33AM +0200, Peter Zijlstra a écrit :
> On Tue, Oct 24, 2023 at 11:46:25PM +0200, Frederic Weisbecker wrote:
> 
> > diff --git a/include/linux/sched.h b/include/linux/sched.h
> > index 8885be2c143e..ad18962b921d 100644
> > --- a/include/linux/sched.h
> > +++ b/include/linux/sched.h
> > @@ -1945,7 +1945,7 @@ extern struct task_struct *idle_task(int cpu);
> >   */
> >  static __always_inline bool is_idle_task(const struct task_struct *p)
> >  {
> > -	return !!(p->flags & PF_IDLE);
> > +	return !!(READ_ONCE(p->flags) & PF_IDLE);
> >  }
> >  
> >  extern struct task_struct *curr_task(int cpu);
> > diff --git a/kernel/cpu.c b/kernel/cpu.c
> > index 3b9d5c7eb4a2..3a1991010f4e 100644
> > --- a/kernel/cpu.c
> > +++ b/kernel/cpu.c
> > @@ -1394,7 +1394,9 @@ void cpuhp_report_idle_dead(void)
> >  {
> >  	struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
> >  
> > +	WRITE_ONCE(current->flags, current->flags & ~PF_IDLE);
> >  	BUG_ON(st->state != CPUHP_AP_OFFLINE);
> > +
> >  	rcutree_report_cpu_dead();
> >  	st->state = CPUHP_AP_IDLE_DEAD;
> >  	/*
> > @@ -1642,6 +1644,8 @@ void cpuhp_online_idle(enum cpuhp_state state)
> >  {
> >  	struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
> >  
> > +	WRITE_ONCE(current->flags, current->flags | PF_IDLE);
> > +
> >  	/* Happens for the boot cpu */
> >  	if (state != CPUHP_AP_ONLINE_IDLE)
> >  		return;
> 
> Without changing *ALL* ->flags stores to WRITE_ONCE() I don't see the
> point of this. Also, since we only care about a single bit, how does
> store tearing affect things?
> 
> Not to mention if we're really paranoid, what are the SMP ordering
> considerations :-)
> 
> [ also, PF_ is used for Protocol Family, Page Flag and Process Flag,
>   grepping is a pain in the arse :-( ]

Indeed. Also cpuhp_online_idle() is called with preemption disabled
and cpuhp_report_idle_dead() with interrupts disabled. As for idle
injection in play_idle_precise(), the flag is set and cleared with
preemption disabled.

This means that all writes are in an RCU read side critical section
that RCU-tasks pre-gp's synchronize_rcu() waits for. So I don't think
we need those WRITE_ONCE/READ_ONCE.

Paul are you ok with that?

Thanks.