rcu: Reduce expedited GP memory contention via per-CPU variables

Currently, the piggybacked-work checks carried out by sync_exp_work_done()
atomically increment a small set of variables (the ->expedited_workdone0,
->expedited_workdone1, ->expedited_workdone2, ->expedited_workdone3
fields in the rcu_state structure), which will form a memory-contention
bottleneck given a sufficiently large number of CPUs concurrently invoking
either synchronize_rcu_expedited() or synchronize_sched_expedited().

This commit therefore moves these for fields to the per-CPU rcu_data
structure, eliminating the memory contention.  The show_rcuexp() function
also changes to sum up each field in the rcu_data structures.
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

kernel/rcu/tree.c

@@ -3585,7 +3585,7 @@ static bool sync_exp_work_done(struct rcu_state *rsp, struct rcu_node *rnp,
  */
 static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 {
-	struct rcu_data *rdp;
+	struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
 	struct rcu_node *rnp0;
 	struct rcu_node *rnp1 = NULL;
 
@@ -3599,7 +3599,7 @@ static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 	if (!mutex_is_locked(&rnp0->exp_funnel_mutex)) {
 		if (mutex_trylock(&rnp0->exp_funnel_mutex)) {
 			if (sync_exp_work_done(rsp, rnp0, NULL,
-					       &rsp->expedited_workdone0, s))
+					       &rdp->expedited_workdone0, s))
 				return NULL;
 			return rnp0;
 		}
@@ -3613,14 +3613,13 @@ static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 	 * can be inexact, as it is just promoting locality and is not
 	 * strictly needed for correctness.
 	 */
-	rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id());
-	if (sync_exp_work_done(rsp, NULL, NULL, &rsp->expedited_workdone1, s))
+	if (sync_exp_work_done(rsp, NULL, NULL, &rdp->expedited_workdone1, s))
 		return NULL;
 	mutex_lock(&rdp->exp_funnel_mutex);
 	rnp0 = rdp->mynode;
 	for (; rnp0 != NULL; rnp0 = rnp0->parent) {
 		if (sync_exp_work_done(rsp, rnp1, rdp,
-				       &rsp->expedited_workdone2, s))
+				       &rdp->expedited_workdone2, s))
 			return NULL;
 		mutex_lock(&rnp0->exp_funnel_mutex);
 		if (rnp1)
@@ -3630,7 +3629,7 @@ static struct rcu_node *exp_funnel_lock(struct rcu_state *rsp, unsigned long s)
 		rnp1 = rnp0;
 	}
 	if (sync_exp_work_done(rsp, rnp1, rdp,
-			       &rsp->expedited_workdone3, s))
+			       &rdp->expedited_workdone3, s))
 		return NULL;
 	return rnp1;
 }

kernel/rcu/tree.h

@@ -386,6 +386,10 @@ struct rcu_data {
 	struct rcu_head oom_head;
 #endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
 	struct mutex exp_funnel_mutex;
+	atomic_long_t expedited_workdone0;	/* # done by others #0. */
+	atomic_long_t expedited_workdone1;	/* # done by others #1. */
+	atomic_long_t expedited_workdone2;	/* # done by others #2. */
+	atomic_long_t expedited_workdone3;	/* # done by others #3. */
 
 	/* 7) Callback offloading. */
 #ifdef CONFIG_RCU_NOCB_CPU
@@ -500,10 +504,6 @@ struct rcu_state {
 	/* End of fields guarded by barrier_mutex. */
 
 	unsigned long expedited_sequence;	/* Take a ticket. */
-	atomic_long_t expedited_workdone0;	/* # done by others #0. */
-	atomic_long_t expedited_workdone1;	/* # done by others #1. */
-	atomic_long_t expedited_workdone2;	/* # done by others #2. */
-	atomic_long_t expedited_workdone3;	/* # done by others #3. */
 	atomic_long_t expedited_normal;		/* # fallbacks to normal. */
 	atomic_t expedited_need_qs;		/* # CPUs left to check in. */
 	wait_queue_head_t expedited_wq;		/* Wait for check-ins. */

kernel/rcu/tree_trace.c

@@ -183,14 +183,20 @@ static const struct file_operations rcudata_fops = {
 
 static int show_rcuexp(struct seq_file *m, void *v)
 {
+	int cpu;
 	struct rcu_state *rsp = (struct rcu_state *)m->private;
-
+	struct rcu_data *rdp;
+	unsigned long s0 = 0, s1 = 0, s2 = 0, s3 = 0;
+
+	for_each_possible_cpu(cpu) {
+		rdp = per_cpu_ptr(rsp->rda, cpu);
+		s0 += atomic_long_read(&rdp->expedited_workdone0);
+		s1 += atomic_long_read(&rdp->expedited_workdone1);
+		s2 += atomic_long_read(&rdp->expedited_workdone2);
+		s3 += atomic_long_read(&rdp->expedited_workdone3);
+	}
 	seq_printf(m, "s=%lu wd0=%lu wd1=%lu wd2=%lu wd3=%lu n=%lu enq=%d sc=%lu\n",
-		   rsp->expedited_sequence,
-		   atomic_long_read(&rsp->expedited_workdone0),
-		   atomic_long_read(&rsp->expedited_workdone1),
-		   atomic_long_read(&rsp->expedited_workdone2),
-		   atomic_long_read(&rsp->expedited_workdone3),
+		   rsp->expedited_sequence, s0, s1, s2, s3,
 		   atomic_long_read(&rsp->expedited_normal),
 		   atomic_read(&rsp->expedited_need_qs),
 		   rsp->expedited_sequence / 2);