rcu: put per-CPU kthread at non-RT priority during CPU hotplug operations
Paul E. McKenney [Mon, 18 Apr 2011 22:31:26 +0000 (15:31 -0700)]
If you are doing CPU hotplug operations, it is best not to have
CPU-bound realtime tasks running CPU-bound on the outgoing CPU.
So this commit makes per-CPU kthreads run at non-realtime priority
during that time.

Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>

kernel/rcutree.c

index 51eef41..198e4df 100644 (file)
@@ -1464,6 +1464,30 @@ static void invoke_rcu_node_kthread(struct rcu_node *rnp)
 }
 
 /*
+ * Set the specified CPU's kthread to run RT or not, as specified by
+ * the to_rt argument.  The CPU-hotplug locks are held, so the task
+ * is not going away.
+ */
+static void rcu_cpu_kthread_setrt(int cpu, int to_rt)
+{
+       int policy;
+       struct sched_param sp;
+       struct task_struct *t;
+
+       t = per_cpu(rcu_cpu_kthread_task, cpu);
+       if (t == NULL)
+               return;
+       if (to_rt) {
+               policy = SCHED_FIFO;
+               sp.sched_priority = RCU_KTHREAD_PRIO;
+       } else {
+               policy = SCHED_NORMAL;
+               sp.sched_priority = 0;
+       }
+       sched_setscheduler_nocheck(t, policy, &sp);
+}
+
+/*
  * Timer handler to initiate the waking up of per-CPU kthreads that
  * have yielded the CPU due to excess numbers of RCU callbacks.
  * We wake up the per-rcu_node kthread, which in turn will wake up
@@ -2166,9 +2190,11 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
        case CPU_ONLINE:
        case CPU_DOWN_FAILED:
                rcu_node_kthread_setaffinity(rnp, -1);
+               rcu_cpu_kthread_setrt(cpu, 1);
                break;
        case CPU_DOWN_PREPARE:
                rcu_node_kthread_setaffinity(rnp, cpu);
+               rcu_cpu_kthread_setrt(cpu, 0);
                break;
        case CPU_DYING:
        case CPU_DYING_FROZEN: