rcu: Make expedited GP CPU stoppage asynchronous

Sequentially stopping the CPUs slows down expedited grace periods by
at least a factor of two, based on rcutorture's grace-period-per-second
rate.  This is a conservative measure because rcutorture uses unusually
long RCU read-side critical sections and because rcutorture periodically
quiesces the system in order to test RCU's ability to ramp down to and
up from the idle state.  This commit therefore replaces the stop_one_cpu()
with stop_one_cpu_nowait(), using an atomic-counter scheme to determine
when all CPUs have passed through the stopped state.

Signed-off-by: Peter Zijlstra <peterz@infradead.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

1 files changed, 17 insertions, 14 deletions

diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
index b310b40a49a2..c5c8509054ef 100644
--- a/kernel/rcu/tree.c
+++ b/kernel/rcu/tree.c
@@ -3257,18 +3257,11 @@ EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
 
 static int synchronize_sched_expedited_cpu_stop(void *data)
 {
-	/*
-	 * There must be a full memory barrier on each affected CPU
-	 * between the time that try_stop_cpus() is called and the
-	 * time that it returns.
-	 *
-	 * In the current initial implementation of cpu_stop, the
-	 * above condition is already met when the control reaches
-	 * this point and the following smp_mb() is not strictly
-	 * necessary.  Do smp_mb() anyway for documentation and
-	 * robustness against future implementation changes.
-	 */
-	smp_mb(); /* See above comment block. */
+	struct rcu_state *rsp = data;
+
+	/* We are here: If we are last, do the wakeup. */
+	if (atomic_dec_and_test(&rsp->expedited_need_qs))
+		wake_up(&rsp->expedited_wq);
 	return 0;
 }
 
@@ -3308,9 +3301,9 @@ void synchronize_sched_expedited(void)
 {
 	int cpu;
 	long s;
-	struct rcu_state *rsp = &rcu_sched_state;
 	struct rcu_node *rnp0;
 	struct rcu_node *rnp1 = NULL;
+	struct rcu_state *rsp = &rcu_sched_state;
 
 	/* Take a snapshot of the sequence number.  */
 	smp_mb(); /* Caller's modifications seen first by other CPUs. */
@@ -3351,16 +3344,26 @@ void synchronize_sched_expedited(void)
 	WARN_ON_ONCE(!(rsp->expedited_sequence & 0x1));
 
 	/* Stop each CPU that is online, non-idle, and not us. */
+	init_waitqueue_head(&rsp->expedited_wq);
+	atomic_set(&rsp->expedited_need_qs, 1); /* Extra count avoids race. */
 	for_each_online_cpu(cpu) {
+		struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
 		struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
 
 		/* Skip our CPU and any idle CPUs. */
 		if (raw_smp_processor_id() == cpu ||
 		    !(atomic_add_return(0, &rdtp->dynticks) & 0x1))
 			continue;
-		stop_one_cpu(cpu, synchronize_sched_expedited_cpu_stop, NULL);
+		atomic_inc(&rsp->expedited_need_qs);
+		stop_one_cpu_nowait(cpu, synchronize_sched_expedited_cpu_stop,
+				    rsp, &rdp->exp_stop_work);
 	}
 
+	/* Remove extra count and, if necessary, wait for CPUs to stop. */
+	if (!atomic_dec_and_test(&rsp->expedited_need_qs))
+		wait_event(rsp->expedited_wq,
+			   !atomic_read(&rsp->expedited_need_qs));
+
 	smp_mb(); /* Ensure expedited GP seen before counter increment. */
 	WRITE_ONCE(rsp->expedited_sequence, rsp->expedited_sequence + 1);
 	WARN_ON_ONCE(rsp->expedited_sequence & 0x1);