sched/debug: Optimize sched_domain sysctl generation

Currently we unconditionally destroy all sysctl bits and regenerate
them after we've rebuild the domains (even if that rebuild is a
no-op).

And since we unconditionally (re)build the sysctl for all possible
CPUs, onlining all CPUs gets us O(n^2) time. Instead change this to
only rebuild the bits for CPUs we've actually installed new domains
on.

Reported-by: Ofer Levi(SW) <oferle@mellanox.com>
Signed-off-by: Peter Zijlstra (Intel) <peterz@infradead.org>
Cc: Linus Torvalds <torvalds@linux-foundation.org>
Cc: Peter Zijlstra <peterz@infradead.org>
Cc: Thomas Gleixner <tglx@linutronix.de>
Signed-off-by: Ingo Molnar <mingo@kernel.org>

kernel/sched/debug.c

@@ -327,38 +327,78 @@ static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu)
 	return table;
 }
 
+static cpumask_var_t sd_sysctl_cpus;
 static struct ctl_table_header *sd_sysctl_header;
+
 void register_sched_domain_sysctl(void)
 {
-	int i, cpu_num = num_possible_cpus();
+	static struct ctl_table *cpu_entries;
-	struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
+	static struct ctl_table **cpu_idx;
 	char buf[32];
+	int i;
 
-	WARN_ON(sd_ctl_dir[0].child);
+	if (!cpu_entries) {
-	sd_ctl_dir[0].child = entry;
+		cpu_entries = sd_alloc_ctl_entry(num_possible_cpus() + 1);
-
+		if (!cpu_entries)
-	if (entry == NULL)
 			return;
 
+		WARN_ON(sd_ctl_dir[0].child);
+		sd_ctl_dir[0].child = cpu_entries;
+	}
+
+	if (!cpu_idx) {
+		struct ctl_table *e = cpu_entries;
+
+		cpu_idx = kcalloc(nr_cpu_ids, sizeof(struct ctl_table*), GFP_KERNEL);
+		if (!cpu_idx)
+			return;
+
+		/* deal with sparse possible map */
 		for_each_possible_cpu(i) {
+			cpu_idx[i] = e;
+			e++;
+		}
+	}
+
+	if (!cpumask_available(sd_sysctl_cpus)) {
+		if (!alloc_cpumask_var(&sd_sysctl_cpus, GFP_KERNEL))
+			return;
+
+		/* init to possible to not have holes in @cpu_entries */
+		cpumask_copy(sd_sysctl_cpus, cpu_possible_mask);
+	}
+
+	for_each_cpu(i, sd_sysctl_cpus) {
+		struct ctl_table *e = cpu_idx[i];
+
+		if (e->child)
+			sd_free_ctl_entry(&e->child);
+
+		if (!e->procname) {
 			snprintf(buf, 32, "cpu%d", i);
-		entry->procname = kstrdup(buf, GFP_KERNEL);
+			e->procname = kstrdup(buf, GFP_KERNEL);
-		entry->mode = 0555;
+		}
-		entry->child = sd_alloc_ctl_cpu_table(i);
+		e->mode = 0555;
-		entry++;
+		e->child = sd_alloc_ctl_cpu_table(i);
+
+		__cpumask_clear_cpu(i, sd_sysctl_cpus);
 	}
 
 	WARN_ON(sd_sysctl_header);
 	sd_sysctl_header = register_sysctl_table(sd_ctl_root);
 }
 
+void dirty_sched_domain_sysctl(int cpu)
+{
+	if (cpumask_available(sd_sysctl_cpus))
+		__cpumask_set_cpu(cpu, sd_sysctl_cpus);
+}
+
 /* may be called multiple times per register */
 void unregister_sched_domain_sysctl(void)
 {
 	unregister_sysctl_table(sd_sysctl_header);
 	sd_sysctl_header = NULL;
-	if (sd_ctl_dir[0].child)
-		sd_free_ctl_entry(&sd_ctl_dir[0].child);
 }
 #endif /* CONFIG_SYSCTL */
 #endif /* CONFIG_SMP */

kernel/sched/sched.h

@@ -1120,11 +1120,15 @@ extern int group_balance_cpu(struct sched_group *sg);
 
 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
 void register_sched_domain_sysctl(void);
+void dirty_sched_domain_sysctl(int cpu);
 void unregister_sched_domain_sysctl(void);
 #else
 static inline void register_sched_domain_sysctl(void)
 {
 }
+static inline void dirty_sched_domain_sysctl(int cpu)
+{
+}
 static inline void unregister_sched_domain_sysctl(void)
 {
 }

kernel/sched/topology.c

@@ -459,6 +459,7 @@ cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
 	rq_attach_root(rq, rd);
 	tmp = rq->sd;
 	rcu_assign_pointer(rq->sd, sd);
+	dirty_sched_domain_sysctl(cpu);
 	destroy_sched_domains(tmp);
 
 	update_top_cache_domain(cpu);