rcu: Avoid tick_dep_set_cpu() misordering

In the current code, rcu_nmi_enter_common() might decide to turn on
the tick using tick_dep_set_cpu(), but be delayed just before doing so.
Then the grace-period kthread might notice that the CPU in question had
in fact gone through a quiescent state, thus turning off the tick using
tick_dep_clear_cpu().  The later invocation of tick_dep_set_cpu() would
then incorrectly leave the tick on.

This commit therefore enlists the aid of the leaf rcu_node structure's
->lock to ensure that decisions to enable or disable the tick are
carried out before they can be reversed.

Signed-off-by: Paul E. McKenney <paulmck@kernel.org>

kernel/rcu/tree.c

@@ -800,8 +800,8 @@ void rcu_user_exit(void)
  */
 static __always_inline void rcu_nmi_enter_common(bool irq)
 {
-	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 	long incby = 2;
+	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
 
 	/* Complain about underflow. */
 	WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
@@ -828,8 +828,13 @@ static __always_inline void rcu_nmi_enter_common(bool irq)
 	} else if (tick_nohz_full_cpu(rdp->cpu) &&
 		   rdp->dynticks_nmi_nesting == DYNTICK_IRQ_NONIDLE &&
 		   READ_ONCE(rdp->rcu_urgent_qs) && !rdp->rcu_forced_tick) {
-		rdp->rcu_forced_tick = true;
-		tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
+		raw_spin_lock_rcu_node(rdp->mynode);
+		// Recheck under lock.
+		if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
+			rdp->rcu_forced_tick = true;
+			tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
+		}
+		raw_spin_unlock_rcu_node(rdp->mynode);
 	}
 	trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
 			  rdp->dynticks_nmi_nesting,
@@ -898,6 +903,7 @@ void rcu_irq_enter_irqson(void)
  */
 static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
 {
+	raw_lockdep_assert_held_rcu_node(rdp->mynode);
 	WRITE_ONCE(rdp->rcu_urgent_qs, false);
 	WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
 	if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {