Re: [PATCH v3 09/57] sched: Simplify ttwu()

[Peter Zijlstra]

On Mon, Jun 12, 2023 at 04:51:17PM +0300, Dan Carpenter wrote:
> On Mon, Jun 12, 2023 at 11:07:22AM +0200, Peter Zijlstra wrote:
> > --- a/kernel/sched/core.c
> > +++ b/kernel/sched/core.c
> > @@ -3664,16 +3664,15 @@ ttwu_stat(struct task_struct *p, int cpu
> >  		__schedstat_inc(p->stats.nr_wakeups_local);
> >  	} else {
> >  		struct sched_domain *sd;
> > +		guard(rcu)();
> >  
> >  		__schedstat_inc(p->stats.nr_wakeups_remote);
> > -		rcu_read_lock();
> 
> We can't put the guard(rcu)(); here?  I have unpublished static analysis
> which assumes that the first and last statements guarded by a lock are
> important.  But if we always put it at the top of the scope then we
> lose that information.

we can definitely put it there. that one schedstat doesn't matter either
way around.

> >  		for_each_domain(rq->cpu, sd) {
> >  			if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
> >  				__schedstat_inc(sd->ttwu_wake_remote);
> >  				break;
> >  			}
> >  		}
> > -		rcu_read_unlock();
> >  	}