rcu: Move RCU grace-period initialization into a kthread

As the first step towards allowing grace-period initialization to be
preemptible, this commit moves the RCU grace-period initialization
into its own kthread.  This is needed to keep large-system scheduling
latency at reasonable levels.

Also change raw_spin_lock_irqsave() to raw_spin_lock_irq() as suggested
by Peter Zijlstra in review comments.

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

@@ -1042,49 +1042,52 @@ rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_dat
 }
 
 /*
- * Start a new RCU grace period if warranted, re-initializing the hierarchy
+ * Body of kthread that handles grace periods.
- * in preparation for detecting the next grace period.  The caller must hold
- * the root node's ->lock, which is released before return.  Hard irqs must
- * be disabled.
- *
- * Note that it is legal for a dying CPU (which is marked as offline) to
- * invoke this function.  This can happen when the dying CPU reports its
- * quiescent state.
  */
-static void
+static int rcu_gp_kthread(void *arg)
-rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
-	__releases(rcu_get_root(rsp)->lock)
 {
-	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
+	struct rcu_data *rdp;
-	struct rcu_node *rnp = rcu_get_root(rsp);
+	struct rcu_node *rnp;
+	struct rcu_state *rsp = arg;
 
-	if (!rcu_scheduler_fully_active ||
+	for (;;) {
-	    !cpu_needs_another_gp(rsp, rdp)) {
+
+		/* Handle grace-period start. */
+		rnp = rcu_get_root(rsp);
+		for (;;) {
+			wait_event_interruptible(rsp->gp_wq, rsp->gp_flags);
+			if (rsp->gp_flags)
+				break;
+			flush_signals(current);
+		}
+		raw_spin_lock_irq(&rnp->lock);
+		rsp->gp_flags = 0;
+		rdp = this_cpu_ptr(rsp->rda);
+
+		if (rcu_gp_in_progress(rsp)) {
 			/*
-		 * Either the scheduler hasn't yet spawned the first
+			 * A grace period is already in progress, so
-		 * non-idle task or this CPU does not need another
+			 * don't start another one.
-		 * grace period.  Either way, don't start a new grace
-		 * period.
 			 */
-		raw_spin_unlock_irqrestore(&rnp->lock, flags);
+			raw_spin_unlock_irq(&rnp->lock);
-		return;
+			continue;
 		}
 
 		if (rsp->fqs_active) {
 			/*
-		 * This CPU needs a grace period, but force_quiescent_state()
+			 * We need a grace period, but force_quiescent_state()
-		 * is running.  Tell it to start one on this CPU's behalf.
+			 * is running.  Tell it to start one on our behalf.
 			 */
 			rsp->fqs_need_gp = 1;
-		raw_spin_unlock_irqrestore(&rnp->lock, flags);
+			raw_spin_unlock_irq(&rnp->lock);
-		return;
+			continue;
 		}
 
 		/* Advance to a new grace period and initialize state. */
 		rsp->gpnum++;
 		trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
 		WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
-	rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
+		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. */
@@ -1095,19 +1098,19 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 		/*
 		 * Set the quiescent-state-needed bits in all the rcu_node
 		 * structures for all currently online CPUs in breadth-first
-	 * order, starting from the root rcu_node structure.  This
+		 * order, starting from the root rcu_node structure.
-	 * operation relies on the layout of the hierarchy within the
+		 * This operation relies on the layout of the hierarchy
-	 * rsp->node[] array.  Note that other CPUs will access only
+		 * within the rsp->node[] array.  Note that other CPUs will
-	 * the leaves of the hierarchy, which still indicate that no
+		 * access only the leaves of the hierarchy, which still
-	 * grace period is in progress, at least until the corresponding
+		 * indicate that no grace period is in progress, at least
-	 * leaf node has been initialized.  In addition, we have excluded
+		 * until the corresponding leaf node has been initialized.
-	 * CPU-hotplug operations.
+		 * In addition, we have excluded CPU-hotplug operations.
 		 *
-	 * Note that the grace period cannot complete until we finish
+		 * Note that the grace period cannot complete until
-	 * the initialization process, as there will be at least one
+		 * we finish the initialization process, as there will
-	 * qsmask bit set in the root node until that time, namely the
+		 * be at least one qsmask bit set in the root node until
-	 * one corresponding to this CPU, due to the fact that we have
+		 * that time, namely the one corresponding to this CPU,
-	 * irqs disabled.
+		 * 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. */
@@ -1126,9 +1129,45 @@ rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
 
 		rnp = rcu_get_root(rsp);
 		raw_spin_lock(&rnp->lock); /* irqs already disabled. */
-	rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
+		/* force_quiescent_state() now OK. */
+		rsp->fqs_state = RCU_SIGNAL_INIT;
 		raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
-	raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
+		raw_spin_unlock_irq(&rsp->onofflock);
+	}
+	return 0;
+}
+
+/*
+ * Start a new RCU grace period if warranted, re-initializing the hierarchy
+ * in preparation for detecting the next grace period.  The caller must hold
+ * the root node's ->lock, which is released before return.  Hard irqs must
+ * be disabled.
+ *
+ * Note that it is legal for a dying CPU (which is marked as offline) to
+ * invoke this function.  This can happen when the dying CPU reports its
+ * quiescent state.
+ */
+static void
+rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
+	__releases(rcu_get_root(rsp)->lock)
+{
+	struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
+	struct rcu_node *rnp = rcu_get_root(rsp);
+
+	if (!rsp->gp_kthread ||
+	    !cpu_needs_another_gp(rsp, rdp)) {
+		/*
+		 * Either we have not yet spawned the grace-period
+		 * task or this CPU does not need another grace period.
+		 * Either way, don't start a new grace period.
+		 */
+		raw_spin_unlock_irqrestore(&rnp->lock, flags);
+		return;
+	}
+
+	rsp->gp_flags = 1;
+	raw_spin_unlock_irqrestore(&rnp->lock, flags);
+	wake_up(&rsp->gp_wq);
 }
 
 /*
@@ -2628,6 +2667,28 @@ static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
 	return NOTIFY_OK;
 }
 
+/*
+ * Spawn the kthread that handles this RCU flavor's grace periods.
+ */
+static int __init rcu_spawn_gp_kthread(void)
+{
+	unsigned long flags;
+	struct rcu_node *rnp;
+	struct rcu_state *rsp;
+	struct task_struct *t;
+
+	for_each_rcu_flavor(rsp) {
+		t = kthread_run(rcu_gp_kthread, rsp, rsp->name);
+		BUG_ON(IS_ERR(t));
+		rnp = rcu_get_root(rsp);
+		raw_spin_lock_irqsave(&rnp->lock, flags);
+		rsp->gp_kthread = t;
+		raw_spin_unlock_irqrestore(&rnp->lock, flags);
+	}
+	return 0;
+}
+early_initcall(rcu_spawn_gp_kthread);
+
 /*
  * This function is invoked towards the end of the scheduler's initialization
  * process.  Before this is called, the idle task might contain
@@ -2729,6 +2790,7 @@ static void __init rcu_init_one(struct rcu_state *rsp,
 	}
 
 	rsp->rda = rda;
+	init_waitqueue_head(&rsp->gp_wq);
 	rnp = rsp->level[rcu_num_lvls - 1];
 	for_each_possible_cpu(i) {
 		while (i > rnp->grphi)

kernel/rcutree.h

@@ -385,6 +385,9 @@ struct rcu_state {
 	u8	boost;				/* Subject to priority boost. */
 	unsigned long gpnum;			/* Current gp number. */
 	unsigned long completed;		/* # of last completed gp. */
+	struct task_struct *gp_kthread;		/* Task for grace periods. */
+	wait_queue_head_t gp_wq;		/* Where GP task waits. */
+	int gp_flags;				/* Commands for GP task. */
 
 	/* End of fields guarded by root rcu_node's lock. */