Re: [PATCH v3] sched/fair: Skip sched_balance_running cmpxchg when balance is not due

[Tim Chen]

On Fri, 2025-11-07 at 08:27 +0530, K Prateek Nayak wrote:
> Hello Tim,
> 
> On 11/7/2025 4:57 AM, Tim Chen wrote:
> > @@ -11757,6 +11772,7 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
> >  		.fbq_type	= all,
> >  		.tasks		= LIST_HEAD_INIT(env.tasks),
> >  	};
> > +	int need_unlock = false;
> >  
> >  	cpumask_and(cpus, sched_domain_span(sd), cpu_active_mask);
> >  
> > @@ -11768,6 +11784,13 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
> >  		goto out_balanced;
> >  	}
> >  
> > +	if (idle != CPU_NEWLY_IDLE && (sd->flags & SD_SERIALIZE)) {
> > +		if (atomic_cmpxchg_acquire(&sched_balance_running, 0, 1)) {
> > +			goto out_balanced;
> > +		}
> > +		need_unlock = true;
> > +	}
> > +
> >  	group = sched_balance_find_src_group(&env);
> >  	if (!group) {
> >  		schedstat_inc(sd->lb_nobusyg[idle]);
> > @@ -11892,6 +11915,9 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
> >  			if (!cpumask_subset(cpus, env.dst_grpmask)) {
> >  				env.loop = 0;
> >  				env.loop_break = SCHED_NR_MIGRATE_BREAK;
> > +				if (need_unlock)
> > +					atomic_set_release(&sched_balance_running, 0);
> 
> I believe we should reset "need_unlock" to false here since "redo" can
> fail the atomic_cmpxchg_acquire() while still having "need_unlock" set
> to "true" and the "out_balanced" path will then perform the
> atomic_set_release() when another CPU is in middle of a busy / idle
> balance on a SD_SERIALIZE domain.

Makes sense.


> 
> We can also initialize the "need_unlock" to false just after
> the redo label too - whichever you prefer.
> 
> nit. "need_unlock" can just be a bool instead of an int.

Sure.

Tim
> 
> Apart from that, feel free to include:
> 
> Reviewed-by: K Prateek Nayak <kprateek.nayak@xxxxxxx>
> 
> > +
> >  				goto redo;
> >  			}
> >  			goto out_all_pinned;
> > @@ -12008,6 +12034,9 @@ static int sched_balance_rq(int this_cpu, struct rq *this_rq,
> >  	    sd->balance_interval < sd->max_interval)
> >  		sd->balance_interval *= 2;
> >  out:
> > +	if (need_unlock)
> > +		atomic_set_release(&sched_balance_running, 0);
> > +
> >  	return ld_moved;
> >  }
> >