rcu: Allow RCU grace-period initialization to be preempted
authorPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Wed, 20 Jun 2012 00:18:20 +0000 (17:18 -0700)
committerPaul E. McKenney <paulmck@linux.vnet.ibm.com>
Sun, 23 Sep 2012 14:41:52 +0000 (07:41 -0700)
RCU grace-period initialization is currently carried out with interrupts
disabled, which can result in 200-microsecond latency spikes on systems
on which RCU has been configured for 4096 CPUs.  This patch therefore
makes the RCU grace-period initialization be preemptible, which should
eliminate those latency spikes.  Similar spikes from grace-period cleanup
and the forcing of quiescent states will be dealt with similarly by later
patches.

Reported-by: Mike Galbraith <mgalbraith@suse.de>
Reported-by: Dimitri Sivanich <sivanich@sgi.com>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
Reviewed-by: Josh Triplett <josh@joshtriplett.org>
kernel/rcutree.c

index e7a534498aa0a3ba87f4a20dbceb2557b9f100b0..781e5f0b7b1754bc6f559cade0ba23fab7b2d465 100644 (file)
@@ -1030,7 +1030,7 @@ rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
 /*
  * Body of kthread that handles grace periods.
  */
-static int rcu_gp_kthread(void *arg)
+static int __noreturn rcu_gp_kthread(void *arg)
 {
        struct rcu_data *rdp;
        struct rcu_node *rnp;
@@ -1056,6 +1056,7 @@ static int rcu_gp_kthread(void *arg)
                         * don't start another one.
                         */
                        raw_spin_unlock_irq(&rnp->lock);
+                       cond_resched();
                        continue;
                }
 
@@ -1066,6 +1067,7 @@ static int rcu_gp_kthread(void *arg)
                         */
                        rsp->fqs_need_gp = 1;
                        raw_spin_unlock_irq(&rnp->lock);
+                       cond_resched();
                        continue;
                }
 
@@ -1076,10 +1078,10 @@ static int rcu_gp_kthread(void *arg)
                rsp->fqs_state = RCU_GP_INIT; /* Stop force_quiescent_state. */
                rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
                record_gp_stall_check_time(rsp);
-               raw_spin_unlock(&rnp->lock);  /* leave irqs disabled. */
+               raw_spin_unlock_irq(&rnp->lock);
 
                /* Exclude any concurrent CPU-hotplug operations. */
-               raw_spin_lock(&rsp->onofflock);  /* irqs already disabled. */
+               get_online_cpus();
 
                /*
                 * Set the quiescent-state-needed bits in all the rcu_node
@@ -1091,15 +1093,9 @@ static int rcu_gp_kthread(void *arg)
                 * indicate that no grace period is in progress, at least
                 * until the corresponding leaf node has been initialized.
                 * In addition, we have excluded CPU-hotplug operations.
-                *
-                * Note that the grace period cannot complete until
-                * we finish the initialization process, as there will
-                * be at least one qsmask bit set in the root node until
-                * that time, namely the one corresponding to this CPU,
-                * due to the fact that we have irqs disabled.
                 */
                rcu_for_each_node_breadth_first(rsp, rnp) {
-                       raw_spin_lock(&rnp->lock); /* irqs already disabled. */
+                       raw_spin_lock_irq(&rnp->lock);
                        rcu_preempt_check_blocked_tasks(rnp);
                        rnp->qsmask = rnp->qsmaskinit;
                        rnp->gpnum = rsp->gpnum;
@@ -1110,17 +1106,17 @@ static int rcu_gp_kthread(void *arg)
                        trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
                                                    rnp->level, rnp->grplo,
                                                    rnp->grphi, rnp->qsmask);
-                       raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
+                       raw_spin_unlock_irq(&rnp->lock);
+                       cond_resched();
                }
 
                rnp = rcu_get_root(rsp);
-               raw_spin_lock(&rnp->lock); /* irqs already disabled. */
+               raw_spin_lock_irq(&rnp->lock);
                /* force_quiescent_state() now OK. */
                rsp->fqs_state = RCU_SIGNAL_INIT;
-               raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
-               raw_spin_unlock_irq(&rsp->onofflock);
+               raw_spin_unlock_irq(&rnp->lock);
+               put_online_cpus();
        }
-       return 0;
 }
 
 /*