sched: restore the sliding search window for select_idle_cpu()

Commit d8ff690 ("sched/fair: Optimize select_idle_cpu")
dropped a per_cpu assignment and replaced "target_tmp" with "target" as
the starting cpu in for_each_cpu_wrap(),
both of which were originally introduced in
commit 7a0590a ("sched: rotate the cpu search window for better spread").
Adding them back allows the per-cpu search window to move
based on the last search result not the target CPU. Target CPUs may not be
uniformly spreading out due to, for example, certain IRQ bindings.

Fixes: d8ff690 ("sched/fair: Optimize select_idle_cpu")

Orabug: 33965297

Signed-off-by: Libo Chen <[email protected]>
Reviewed-by: Chris Hyser <[email protected]>
Signed-off-by: John Donnelly <[email protected]>

Author: Libo Chen

kernel/sched/fair.c

@@ -5866,7 +5866,8 @@ static int select_idle_cpu(struct task_struct *p, struct sched_domain *sd, int t
 
 	cpumask_and(cpus, sched_domain_span(sd), &p->cpus_allowed);
 
-	for_each_cpu_wrap(cpu, cpus, target) {
+	for_each_cpu_wrap(cpu, cpus, target_tmp) {
+		per_cpu(next_cpu, target) = cpu;
 		if (!--nr)
 			return -1;
 		if (idle_cpu(cpu))