sched: fix load average accounting vs. cpu hotplug
authorThomas Gleixner <tglx@linutronix.de>
Fri, 17 Jul 2009 12:15:46 +0000 (14:15 +0200)
committerThomas Gleixner <tglx@linutronix.de>
Sat, 18 Jul 2009 12:19:52 +0000 (14:19 +0200)
The new load average code clears rq->calc_load_active on
CPU_ONLINE. That's wrong as the new onlined CPU might have got a
scheduler tick already and accounted the delta to the stale value of
the time we offlined the CPU.

Clear the value when we cleanup the dead CPU instead.

Also move the update of the calc_load_update time for the newly online
CPU to CPU_UP_PREPARE to avoid that the CPU plays catch up with the
stale update time value.

Signed-off-by: Thomas Gleixner <tglx@linutronix.de>
kernel/sched.c

index 98972d366fdc5bd8c2bf6232ee212e90d20be5cb..1b59e265273b032d6aac2baec9b059646ca563bf 100644 (file)
@@ -7289,6 +7289,7 @@ static void migrate_dead_tasks(unsigned int dead_cpu)
 static void calc_global_load_remove(struct rq *rq)
 {
        atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
+       rq->calc_load_active = 0;
 }
 #endif /* CONFIG_HOTPLUG_CPU */
 
@@ -7515,6 +7516,7 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
                task_rq_unlock(rq, &flags);
                get_task_struct(p);
                cpu_rq(cpu)->migration_thread = p;
+               rq->calc_load_update = calc_load_update;
                break;
 
        case CPU_ONLINE:
@@ -7525,8 +7527,6 @@ migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
                /* Update our root-domain */
                rq = cpu_rq(cpu);
                spin_lock_irqsave(&rq->lock, flags);
-               rq->calc_load_update = calc_load_update;
-               rq->calc_load_active = 0;
                if (rq->rd) {
                        BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));