[PATCH] sched: integrate cpu hotplug and sched domains

Register a cpu hotplug notifier which reinitializes the scheduler domains
hierarchy.  The notifier temporarily attaches all running cpus to a "dummy"
domain (like we currently do during boot) to avoid balancing.  It then calls
arch_init_sched_domains which rebuilds the "real" domains and reattaches the
cpus to them.

Also change __init attributes to __devinit where necessary.
Signed-off-by: Nathan Lynch <nathanl@austin.ibm.com>

Alterations from Nick Piggin:

* Detach all domains in CPU_UP|DOWN_PREPARE notifiers. Reinitialise and
  reattach in CPU_ONLINE|DEAD|UP_CANCELED. This ensures the domains as
  seen from the scheduler won't become out of synch with the cpu_online_map.

* This allows us to remove runtime cpu_online verifications. Do that.

* Dummy domains are __devinitdata.

* Remove the hackery in arch_init_sched_domains to work around the fact that
  the domains used to work with cpu_possible maps, but node_to_cpumask returned
  a cpu_online map.
Signed-off-by: Nick Piggin <nickpiggin@yahoo.com.au>
Signed-off-by: Ingo Molnar <mingo@elte.hu>
Signed-off-by: Andrew Morton <akpm@osdl.org>
Signed-off-by: Linus Torvalds <torvalds@osdl.org>

kernel/sched.c

@@ -1087,8 +1087,7 @@ static int wake_idle(int cpu, task_t *p)
 	if (!(sd->flags & SD_WAKE_IDLE))
 		return cpu;
 
-	cpus_and(tmp, sd->span, cpu_online_map);
-	cpus_and(tmp, tmp, p->cpus_allowed);
+	cpus_and(tmp, sd->span, p->cpus_allowed);
 
 	for_each_cpu_mask(i, tmp) {
 		if (idle_cpu(i))
@@ -1640,8 +1639,7 @@ static int find_idlest_cpu(struct task_struct *p, int this_cpu,
 	min_cpu = UINT_MAX;
 	min_load = ULONG_MAX;
 
-	cpus_and(mask, sd->span, cpu_online_map);
-	cpus_and(mask, mask, p->cpus_allowed);
+	cpus_and(mask, sd->span, p->cpus_allowed);
 
 	for_each_cpu_mask(i, mask) {
 		load = target_load(i);
@@ -1893,7 +1891,6 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
 	max_load = this_load = total_load = total_pwr = 0;
 
 	do {
-		cpumask_t tmp;
 		unsigned long load;
 		int local_group;
 		int i, nr_cpus = 0;
@@ -1902,11 +1899,8 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
 
 		/* Tally up the load of all CPUs in the group */
 		avg_load = 0;
-		cpus_and(tmp, group->cpumask, cpu_online_map);
-		if (unlikely(cpus_empty(tmp)))
-			goto nextgroup;
 
-		for_each_cpu_mask(i, tmp) {
+		for_each_cpu_mask(i, group->cpumask) {
 			/* Bias balancing toward cpus of our domain */
 			if (local_group)
 				load = target_load(i);
@@ -2025,13 +2019,11 @@ find_busiest_group(struct sched_domain *sd, int this_cpu,
  */
 static runqueue_t *find_busiest_queue(struct sched_group *group)
 {
-	cpumask_t tmp;
 	unsigned long load, max_load = 0;
 	runqueue_t *busiest = NULL;
 	int i;
 
-	cpus_and(tmp, group->cpumask, cpu_online_map);
-	for_each_cpu_mask(i, tmp) {
+	for_each_cpu_mask(i, group->cpumask) {
 		load = source_load(i);
 
 		if (load > max_load) {
@@ -2232,18 +2224,13 @@ static void active_load_balance(runqueue_t *busiest, int busiest_cpu)
 
 	group = sd->groups;
 	do {
-		cpumask_t tmp;
 		runqueue_t *rq;
 		int push_cpu = 0;
 
 		if (group == busy_group)
 			goto next_group;
 
-		cpus_and(tmp, group->cpumask, cpu_online_map);
-		if (!cpus_weight(tmp))
-			goto next_group;
-
-		for_each_cpu_mask(i, tmp) {
+		for_each_cpu_mask(i, group->cpumask) {
 			if (!idle_cpu(i))
 				goto next_group;
 			push_cpu = i;
@@ -2512,7 +2499,7 @@ static inline void wake_sleeping_dependent(int this_cpu, runqueue_t *this_rq)
 	 */
 	spin_unlock(&this_rq->lock);
 
-	cpus_and(sibling_map, sd->span, cpu_online_map);
+	sibling_map = sd->span;
 
 	for_each_cpu_mask(i, sibling_map)
 		spin_lock(&cpu_rq(i)->lock);
@@ -2557,7 +2544,7 @@ static inline int dependent_sleeper(int this_cpu, runqueue_t *this_rq)
 	 * wake_sleeping_dependent():
 	 */
 	spin_unlock(&this_rq->lock);
-	cpus_and(sibling_map, sd->span, cpu_online_map);
+	sibling_map = sd->span;
 	for_each_cpu_mask(i, sibling_map)
 		spin_lock(&cpu_rq(i)->lock);
 	cpu_clear(this_cpu, sibling_map);
@@ -4209,7 +4196,10 @@ spinlock_t kernel_flag __cacheline_aligned_in_smp = SPIN_LOCK_UNLOCKED;
 EXPORT_SYMBOL(kernel_flag);
 
 #ifdef CONFIG_SMP
-/* Attach the domain 'sd' to 'cpu' as its base domain */
+/*
+ * Attach the domain 'sd' to 'cpu' as its base domain.  Callers must
+ * hold the hotplug lock.
+ */
 static void cpu_attach_domain(struct sched_domain *sd, int cpu)
 {
 	migration_req_t req;
@@ -4217,8 +4207,6 @@ static void cpu_attach_domain(struct sched_domain *sd, int cpu)
 	runqueue_t *rq = cpu_rq(cpu);
 	int local = 1;
 
-	lock_cpu_hotplug();
-
 	spin_lock_irqsave(&rq->lock, flags);
 
 	if (cpu == smp_processor_id() || !cpu_online(cpu)) {
@@ -4237,8 +4225,6 @@ static void cpu_attach_domain(struct sched_domain *sd, int cpu)
 		wake_up_process(rq->migration_thread);
 		wait_for_completion(&req.done);
 	}
-
-	unlock_cpu_hotplug();
 }
 
 /*
@@ -4258,7 +4244,7 @@ static void cpu_attach_domain(struct sched_domain *sd, int cpu)
  *
  * Should use nodemask_t.
  */
-static int __init find_next_best_node(int node, unsigned long *used_nodes)
+static int __devinit find_next_best_node(int node, unsigned long *used_nodes)
 {
 	int i, n, val, min_val, best_node = 0;
 
@@ -4294,7 +4280,7 @@ static int __init find_next_best_node(int node, unsigned long *used_nodes)
  * should be one that prevents unnecessary balancing, but also spreads tasks
  * out optimally.
  */
-static cpumask_t __init sched_domain_node_span(int node)
+static cpumask_t __devinit sched_domain_node_span(int node)
 {
 	int i;
 	cpumask_t span;
@@ -4314,7 +4300,7 @@ static cpumask_t __init sched_domain_node_span(int node)
 	return span;
 }
 #else /* SD_NODES_PER_DOMAIN */
-static cpumask_t __init sched_domain_node_span(int node)
+static cpumask_t __devinit sched_domain_node_span(int node)
 {
 	return cpu_possible_map;
 }
@@ -4324,7 +4310,7 @@ static cpumask_t __init sched_domain_node_span(int node)
 #ifdef CONFIG_SCHED_SMT
 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
 static struct sched_group sched_group_cpus[NR_CPUS];
-__init static int cpu_to_cpu_group(int cpu)
+static int __devinit cpu_to_cpu_group(int cpu)
 {
 	return cpu;
 }
@@ -4332,7 +4318,7 @@ __init static int cpu_to_cpu_group(int cpu)
 
 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
 static struct sched_group sched_group_phys[NR_CPUS];
-__init static int cpu_to_phys_group(int cpu)
+static int __devinit cpu_to_phys_group(int cpu)
 {
 #ifdef CONFIG_SCHED_SMT
 	return first_cpu(cpu_sibling_map[cpu]);
@@ -4345,7 +4331,7 @@ __init static int cpu_to_phys_group(int cpu)
 
 static DEFINE_PER_CPU(struct sched_domain, node_domains);
 static struct sched_group sched_group_nodes[MAX_NUMNODES];
-__init static int cpu_to_node_group(int cpu)
+static int __devinit cpu_to_node_group(int cpu)
 {
 	return cpu_to_node(cpu);
 }
@@ -4355,9 +4341,9 @@ __init static int cpu_to_node_group(int cpu)
 static struct sched_group sched_group_isolated[NR_CPUS];
 
 /* cpus with isolated domains */
-cpumask_t __initdata cpu_isolated_map = CPU_MASK_NONE;
+cpumask_t __devinitdata cpu_isolated_map = CPU_MASK_NONE;
 
-__init static int cpu_to_isolated_group(int cpu)
+static int __devinit cpu_to_isolated_group(int cpu)
 {
 	return cpu;
 }
@@ -4387,7 +4373,7 @@ __setup ("isolcpus=", isolated_cpu_setup);
  * covered by the given span, and will set each group's ->cpumask correctly,
  * and ->cpu_power to 0.
  */
-__init static void init_sched_build_groups(struct sched_group groups[],
+static void __devinit init_sched_build_groups(struct sched_group groups[],
 			cpumask_t span, int (*group_fn)(int cpu))
 {
 	struct sched_group *first = NULL, *last = NULL;
@@ -4421,10 +4407,16 @@ __init static void init_sched_build_groups(struct sched_group groups[],
 	last->next = first;
 }
 
-__init static void arch_init_sched_domains(void)
+/*
+ * Set up scheduler domains and groups.  Callers must hold the hotplug lock.
+ */
+static void __devinit arch_init_sched_domains(void)
 {
 	int i;
 	cpumask_t cpu_default_map;
+	cpumask_t cpu_isolated_online_map;
+
+	cpus_and(cpu_isolated_online_map, cpu_isolated_map, cpu_online_map);
 
 	/*
 	 * Setup mask for cpus without special case scheduling requirements.
@@ -4432,10 +4424,10 @@ __init static void arch_init_sched_domains(void)
 	 * exclude other special cases in the future.
 	 */
 	cpus_complement(cpu_default_map, cpu_isolated_map);
-	cpus_and(cpu_default_map, cpu_default_map, cpu_possible_map);
+	cpus_and(cpu_default_map, cpu_default_map, cpu_online_map);
 
 	/* Set up domains */
-	for_each_cpu(i) {
+	for_each_online_cpu(i) {
 		int group;
 		struct sched_domain *sd = NULL, *p;
 		cpumask_t nodemask = node_to_cpumask(cpu_to_node(i));
@@ -4447,7 +4439,7 @@ __init static void arch_init_sched_domains(void)
 		 * Unlike those of other cpus, the domains and groups are
 		 * single level, and span a single cpu.
 		 */
-		if (cpu_isset(i, cpu_isolated_map)) {
+		if (cpu_isset(i, cpu_isolated_online_map)) {
 #ifdef CONFIG_SCHED_SMT
 			sd = &per_cpu(cpu_domains, i);
 #else
@@ -4478,11 +4470,7 @@ __init static void arch_init_sched_domains(void)
 		sd = &per_cpu(phys_domains, i);
 		group = cpu_to_phys_group(i);
 		*sd = SD_CPU_INIT;
-#ifdef CONFIG_NUMA
 		sd->span = nodemask;
-#else
-		sd->span = cpu_possible_map;
-#endif
 		sd->parent = p;
 		sd->groups = &sched_group_phys[group];
 
@@ -4500,7 +4488,7 @@ __init static void arch_init_sched_domains(void)
 
 #ifdef CONFIG_SCHED_SMT
 	/* Set up CPU (sibling) groups */
-	for_each_cpu(i) {
+	for_each_online_cpu(i) {
 		cpumask_t this_sibling_map = cpu_sibling_map[i];
 		cpus_and(this_sibling_map, this_sibling_map, cpu_default_map);
 		if (i != first_cpu(this_sibling_map))
@@ -4512,15 +4500,12 @@ __init static void arch_init_sched_domains(void)
 #endif
 
 	/* Set up isolated groups */
-	for_each_cpu_mask(i, cpu_isolated_map) {
-		cpumask_t mask;
-		cpus_clear(mask);
-		cpu_set(i, mask);
+	for_each_cpu_mask(i, cpu_isolated_online_map) {
+		cpumask_t mask = cpumask_of_cpu(i);
 		init_sched_build_groups(sched_group_isolated, mask,
 						&cpu_to_isolated_group);
 	}
 
-#ifdef CONFIG_NUMA
 	/* Set up physical groups */
 	for (i = 0; i < MAX_NUMNODES; i++) {
 		cpumask_t nodemask = node_to_cpumask(i);
@@ -4532,10 +4517,6 @@ __init static void arch_init_sched_domains(void)
 		init_sched_build_groups(sched_group_phys, nodemask,
 						&cpu_to_phys_group);
 	}
-#else
-	init_sched_build_groups(sched_group_phys, cpu_possible_map,
-							&cpu_to_phys_group);
-#endif
 
 #ifdef CONFIG_NUMA
 	/* Set up node groups */
@@ -4568,7 +4549,7 @@ __init static void arch_init_sched_domains(void)
 	}
 
 	/* Attach the domains */
-	for_each_cpu(i) {
+	for_each_online_cpu(i) {
 		struct sched_domain *sd;
 #ifdef CONFIG_SCHED_SMT
 		sd = &per_cpu(cpu_domains, i);
@@ -4579,21 +4560,25 @@ __init static void arch_init_sched_domains(void)
 	}
 }
 
+static void __devinit arch_destroy_sched_domains(void)
+{
+	/* Do nothing: everything is statically allocated. */
+}
+
 #undef SCHED_DOMAIN_DEBUG
 #ifdef SCHED_DOMAIN_DEBUG
 void sched_domain_debug(void)
 {
 	int i;
 
-	for_each_cpu(i) {
+	for_each_online_cpu(i) {
 		runqueue_t *rq = cpu_rq(i);
 		struct sched_domain *sd;
 		int level = 0;
 
 		sd = rq->sd;
 
-		printk(KERN_DEBUG "CPU%d: %s\n",
-				i, (cpu_online(i) ? " online" : "offline"));
+		printk(KERN_DEBUG "CPU%d:\n", i);
 
 		do {
 			int j;
@@ -4659,10 +4644,60 @@ void sched_domain_debug(void)
 #define sched_domain_debug() {}
 #endif
 
+#ifdef CONFIG_SMP
+/* Initial dummy domain for early boot and for hotplug cpu */
+static __devinitdata struct sched_domain sched_domain_dummy;
+static __devinitdata struct sched_group sched_group_dummy;
+#endif
+
+#ifdef CONFIG_HOTPLUG_CPU
+/*
+ * Force a reinitialization of the sched domains hierarchy.  The domains
+ * and groups cannot be updated in place without racing with the balancing
+ * code, so we temporarily attach all running cpus to a "dummy" domain
+ * which will prevent rebalancing while the sched domains are recalculated.
+ */
+static int update_sched_domains(struct notifier_block *nfb,
+				unsigned long action, void *hcpu)
+{
+	int i;
+
+	switch (action) {
+	case CPU_UP_PREPARE:
+	case CPU_DOWN_PREPARE:
+		for_each_online_cpu(i)
+			cpu_attach_domain(&sched_domain_dummy, i);
+		arch_destroy_sched_domains();
+		return NOTIFY_OK;
+
+	case CPU_UP_CANCELED:
+	case CPU_ONLINE:
+	case CPU_DEAD:
+		/*
+		 * Fall through and re-initialise the domains.
+		 */
+		break;
+	default:
+		return NOTIFY_DONE;
+	}
+
+	/* The hotplug lock is already held by cpu_up/cpu_down */
+	arch_init_sched_domains();
+
+	sched_domain_debug();
+
+	return NOTIFY_OK;
+}
+#endif
+
 void __init sched_init_smp(void)
 {
+	lock_cpu_hotplug();
 	arch_init_sched_domains();
 	sched_domain_debug();
+	unlock_cpu_hotplug();
+	/* XXX: Theoretical race here - CPU may be hotplugged now */
+	hotcpu_notifier(update_sched_domains, 0);
 }
 #else
 void __init sched_init_smp(void)
@@ -4686,20 +4721,18 @@ void __init sched_init(void)
 
 #ifdef CONFIG_SMP
 	/* Set up an initial dummy domain for early boot */
-	static struct sched_domain sched_domain_init;
-	static struct sched_group sched_group_init;
-
-	memset(&sched_domain_init, 0, sizeof(struct sched_domain));
-	sched_domain_init.span = CPU_MASK_ALL;
-	sched_domain_init.groups = &sched_group_init;
-	sched_domain_init.last_balance = jiffies;
-	sched_domain_init.balance_interval = INT_MAX; /* Don't balance */
-	sched_domain_init.busy_factor = 1;
-
-	memset(&sched_group_init, 0, sizeof(struct sched_group));
-	sched_group_init.cpumask = CPU_MASK_ALL;
-	sched_group_init.next = &sched_group_init;
-	sched_group_init.cpu_power = SCHED_LOAD_SCALE;
+
+	memset(&sched_domain_dummy, 0, sizeof(struct sched_domain));
+	sched_domain_dummy.span = CPU_MASK_ALL;
+	sched_domain_dummy.groups = &sched_group_dummy;
+	sched_domain_dummy.last_balance = jiffies;
+	sched_domain_dummy.balance_interval = INT_MAX; /* Don't balance */
+	sched_domain_dummy.busy_factor = 1;
+
+	memset(&sched_group_dummy, 0, sizeof(struct sched_group));
+	sched_group_dummy.cpumask = CPU_MASK_ALL;
+	sched_group_dummy.next = &sched_group_dummy;
+	sched_group_dummy.cpu_power = SCHED_LOAD_SCALE;
 #endif
 
 	for (i = 0; i < NR_CPUS; i++) {
@@ -4712,7 +4745,7 @@ void __init sched_init(void)
 		rq->best_expired_prio = MAX_PRIO;
 
 #ifdef CONFIG_SMP
-		rq->sd = &sched_domain_init;
+		rq->sd = &sched_domain_dummy;
 		rq->cpu_load = 0;
 		rq->active_balance = 0;
 		rq->push_cpu = 0;