2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
4 * This code is licenced under the GPL.
6 #include <linux/proc_fs.h>
8 #include <linux/init.h>
9 #include <linux/notifier.h>
10 #include <linux/sched.h>
11 #include <linux/unistd.h>
12 #include <linux/cpu.h>
13 #include <linux/oom.h>
14 #include <linux/rcupdate.h>
15 #include <linux/export.h>
16 #include <linux/bug.h>
17 #include <linux/kthread.h>
18 #include <linux/stop_machine.h>
19 #include <linux/mutex.h>
20 #include <linux/gfp.h>
21 #include <linux/suspend.h>
22 #include <linux/lockdep.h>
23 #include <linux/tick.h>
24 #include <linux/irq.h>
25 #include <trace/events/power.h>
30 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
31 static DEFINE_MUTEX(cpu_add_remove_lock);
34 * The following two APIs (cpu_maps_update_begin/done) must be used when
35 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
36 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
37 * hotplug callback (un)registration performed using __register_cpu_notifier()
38 * or __unregister_cpu_notifier().
40 void cpu_maps_update_begin(void)
42 mutex_lock(&cpu_add_remove_lock);
44 EXPORT_SYMBOL(cpu_notifier_register_begin);
46 void cpu_maps_update_done(void)
48 mutex_unlock(&cpu_add_remove_lock);
50 EXPORT_SYMBOL(cpu_notifier_register_done);
52 static RAW_NOTIFIER_HEAD(cpu_chain);
54 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
55 * Should always be manipulated under cpu_add_remove_lock
57 static int cpu_hotplug_disabled;
59 #ifdef CONFIG_HOTPLUG_CPU
62 struct task_struct *active_writer;
63 /* wait queue to wake up the active_writer */
65 /* verifies that no writer will get active while readers are active */
68 * Also blocks the new readers during
69 * an ongoing cpu hotplug operation.
73 #ifdef CONFIG_DEBUG_LOCK_ALLOC
74 struct lockdep_map dep_map;
77 .active_writer = NULL,
78 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
79 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
80 #ifdef CONFIG_DEBUG_LOCK_ALLOC
81 .dep_map = {.name = "cpu_hotplug.lock" },
85 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
86 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
87 #define cpuhp_lock_acquire_tryread() \
88 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
89 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
90 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
93 void get_online_cpus(void)
96 if (cpu_hotplug.active_writer == current)
98 cpuhp_lock_acquire_read();
99 mutex_lock(&cpu_hotplug.lock);
100 atomic_inc(&cpu_hotplug.refcount);
101 mutex_unlock(&cpu_hotplug.lock);
103 EXPORT_SYMBOL_GPL(get_online_cpus);
105 void put_online_cpus(void)
109 if (cpu_hotplug.active_writer == current)
112 refcount = atomic_dec_return(&cpu_hotplug.refcount);
113 if (WARN_ON(refcount < 0)) /* try to fix things up */
114 atomic_inc(&cpu_hotplug.refcount);
116 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
117 wake_up(&cpu_hotplug.wq);
119 cpuhp_lock_release();
122 EXPORT_SYMBOL_GPL(put_online_cpus);
125 * This ensures that the hotplug operation can begin only when the
126 * refcount goes to zero.
128 * Note that during a cpu-hotplug operation, the new readers, if any,
129 * will be blocked by the cpu_hotplug.lock
131 * Since cpu_hotplug_begin() is always called after invoking
132 * cpu_maps_update_begin(), we can be sure that only one writer is active.
134 * Note that theoretically, there is a possibility of a livelock:
135 * - Refcount goes to zero, last reader wakes up the sleeping
137 * - Last reader unlocks the cpu_hotplug.lock.
138 * - A new reader arrives at this moment, bumps up the refcount.
139 * - The writer acquires the cpu_hotplug.lock finds the refcount
140 * non zero and goes to sleep again.
142 * However, this is very difficult to achieve in practice since
143 * get_online_cpus() not an api which is called all that often.
146 void cpu_hotplug_begin(void)
150 cpu_hotplug.active_writer = current;
151 cpuhp_lock_acquire();
154 mutex_lock(&cpu_hotplug.lock);
155 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
156 if (likely(!atomic_read(&cpu_hotplug.refcount)))
158 mutex_unlock(&cpu_hotplug.lock);
161 finish_wait(&cpu_hotplug.wq, &wait);
164 void cpu_hotplug_done(void)
166 cpu_hotplug.active_writer = NULL;
167 mutex_unlock(&cpu_hotplug.lock);
168 cpuhp_lock_release();
172 * Wait for currently running CPU hotplug operations to complete (if any) and
173 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
174 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
175 * hotplug path before performing hotplug operations. So acquiring that lock
176 * guarantees mutual exclusion from any currently running hotplug operations.
178 void cpu_hotplug_disable(void)
180 cpu_maps_update_begin();
181 cpu_hotplug_disabled++;
182 cpu_maps_update_done();
184 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
186 void cpu_hotplug_enable(void)
188 cpu_maps_update_begin();
189 WARN_ON(--cpu_hotplug_disabled < 0);
190 cpu_maps_update_done();
192 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
193 #endif /* CONFIG_HOTPLUG_CPU */
195 /* Need to know about CPUs going up/down? */
196 int register_cpu_notifier(struct notifier_block *nb)
199 cpu_maps_update_begin();
200 ret = raw_notifier_chain_register(&cpu_chain, nb);
201 cpu_maps_update_done();
205 int __register_cpu_notifier(struct notifier_block *nb)
207 return raw_notifier_chain_register(&cpu_chain, nb);
210 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
215 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
218 return notifier_to_errno(ret);
221 static int cpu_notify(unsigned long val, void *v)
223 return __cpu_notify(val, v, -1, NULL);
226 #ifdef CONFIG_HOTPLUG_CPU
228 static void cpu_notify_nofail(unsigned long val, void *v)
230 BUG_ON(cpu_notify(val, v));
232 EXPORT_SYMBOL(register_cpu_notifier);
233 EXPORT_SYMBOL(__register_cpu_notifier);
235 void unregister_cpu_notifier(struct notifier_block *nb)
237 cpu_maps_update_begin();
238 raw_notifier_chain_unregister(&cpu_chain, nb);
239 cpu_maps_update_done();
241 EXPORT_SYMBOL(unregister_cpu_notifier);
243 void __unregister_cpu_notifier(struct notifier_block *nb)
245 raw_notifier_chain_unregister(&cpu_chain, nb);
247 EXPORT_SYMBOL(__unregister_cpu_notifier);
250 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
253 * This function walks all processes, finds a valid mm struct for each one and
254 * then clears a corresponding bit in mm's cpumask. While this all sounds
255 * trivial, there are various non-obvious corner cases, which this function
256 * tries to solve in a safe manner.
258 * Also note that the function uses a somewhat relaxed locking scheme, so it may
259 * be called only for an already offlined CPU.
261 void clear_tasks_mm_cpumask(int cpu)
263 struct task_struct *p;
266 * This function is called after the cpu is taken down and marked
267 * offline, so its not like new tasks will ever get this cpu set in
268 * their mm mask. -- Peter Zijlstra
269 * Thus, we may use rcu_read_lock() here, instead of grabbing
270 * full-fledged tasklist_lock.
272 WARN_ON(cpu_online(cpu));
274 for_each_process(p) {
275 struct task_struct *t;
278 * Main thread might exit, but other threads may still have
279 * a valid mm. Find one.
281 t = find_lock_task_mm(p);
284 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
290 static inline void check_for_tasks(int dead_cpu)
292 struct task_struct *g, *p;
294 read_lock(&tasklist_lock);
295 for_each_process_thread(g, p) {
299 * We do the check with unlocked task_rq(p)->lock.
300 * Order the reading to do not warn about a task,
301 * which was running on this cpu in the past, and
302 * it's just been woken on another cpu.
305 if (task_cpu(p) != dead_cpu)
308 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
309 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
311 read_unlock(&tasklist_lock);
314 struct take_cpu_down_param {
319 /* Take this CPU down. */
320 static int take_cpu_down(void *_param)
322 struct take_cpu_down_param *param = _param;
325 /* Ensure this CPU doesn't handle any more interrupts. */
326 err = __cpu_disable();
330 cpu_notify(CPU_DYING | param->mod, param->hcpu);
331 /* Give up timekeeping duties */
332 tick_handover_do_timer();
333 /* Park the stopper thread */
334 stop_machine_park((long)param->hcpu);
338 /* Requires cpu_add_remove_lock to be held */
339 static int _cpu_down(unsigned int cpu, int tasks_frozen)
341 int err, nr_calls = 0;
342 void *hcpu = (void *)(long)cpu;
343 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
344 struct take_cpu_down_param tcd_param = {
349 if (num_online_cpus() == 1)
352 if (!cpu_online(cpu))
357 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
360 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
361 pr_warn("%s: attempt to take down CPU %u failed\n",
367 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
368 * and RCU users of this state to go away such that all new such users
371 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
372 * not imply sync_sched(), so wait for both.
374 * Do sync before park smpboot threads to take care the rcu boost case.
376 if (IS_ENABLED(CONFIG_PREEMPT))
377 synchronize_rcu_mult(call_rcu, call_rcu_sched);
381 smpboot_park_threads(cpu);
384 * Prevent irq alloc/free while the dying cpu reorganizes the
385 * interrupt affinities.
390 * So now all preempt/rcu users must observe !cpu_active().
392 err = stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
394 /* CPU didn't die: tell everyone. Can't complain. */
395 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
399 BUG_ON(cpu_online(cpu));
402 * The migration_call() CPU_DYING callback will have removed all
403 * runnable tasks from the cpu, there's only the idle task left now
404 * that the migration thread is done doing the stop_machine thing.
406 * Wait for the stop thread to go away.
408 while (!per_cpu(cpu_dead_idle, cpu))
410 smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
411 per_cpu(cpu_dead_idle, cpu) = false;
413 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
416 hotplug_cpu__broadcast_tick_pull(cpu);
417 /* This actually kills the CPU. */
420 /* CPU is completely dead: tell everyone. Too late to complain. */
421 tick_cleanup_dead_cpu(cpu);
422 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
424 check_for_tasks(cpu);
429 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
433 int cpu_down(unsigned int cpu)
437 cpu_maps_update_begin();
439 if (cpu_hotplug_disabled) {
444 err = _cpu_down(cpu, 0);
447 cpu_maps_update_done();
450 EXPORT_SYMBOL(cpu_down);
451 #endif /*CONFIG_HOTPLUG_CPU*/
454 * Unpark per-CPU smpboot kthreads at CPU-online time.
456 static int smpboot_thread_call(struct notifier_block *nfb,
457 unsigned long action, void *hcpu)
459 int cpu = (long)hcpu;
461 switch (action & ~CPU_TASKS_FROZEN) {
463 case CPU_DOWN_FAILED:
465 smpboot_unpark_threads(cpu);
475 static struct notifier_block smpboot_thread_notifier = {
476 .notifier_call = smpboot_thread_call,
477 .priority = CPU_PRI_SMPBOOT,
480 void smpboot_thread_init(void)
482 register_cpu_notifier(&smpboot_thread_notifier);
485 /* Requires cpu_add_remove_lock to be held */
486 static int _cpu_up(unsigned int cpu, int tasks_frozen)
488 int ret, nr_calls = 0;
489 void *hcpu = (void *)(long)cpu;
490 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
491 struct task_struct *idle;
495 if (cpu_online(cpu) || !cpu_present(cpu)) {
500 idle = idle_thread_get(cpu);
506 ret = smpboot_create_threads(cpu);
510 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
513 pr_warn("%s: attempt to bring up CPU %u failed\n",
518 /* Arch-specific enabling code. */
519 ret = __cpu_up(cpu, idle);
523 BUG_ON(!cpu_online(cpu));
525 /* Now call notifier in preparation. */
526 cpu_notify(CPU_ONLINE | mod, hcpu);
530 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
537 int cpu_up(unsigned int cpu)
541 if (!cpu_possible(cpu)) {
542 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
544 #if defined(CONFIG_IA64)
545 pr_err("please check additional_cpus= boot parameter\n");
550 err = try_online_node(cpu_to_node(cpu));
554 cpu_maps_update_begin();
556 if (cpu_hotplug_disabled) {
561 err = _cpu_up(cpu, 0);
564 cpu_maps_update_done();
567 EXPORT_SYMBOL_GPL(cpu_up);
569 #ifdef CONFIG_PM_SLEEP_SMP
570 static cpumask_var_t frozen_cpus;
572 int disable_nonboot_cpus(void)
574 int cpu, first_cpu, error = 0;
576 cpu_maps_update_begin();
577 first_cpu = cpumask_first(cpu_online_mask);
579 * We take down all of the non-boot CPUs in one shot to avoid races
580 * with the userspace trying to use the CPU hotplug at the same time
582 cpumask_clear(frozen_cpus);
584 pr_info("Disabling non-boot CPUs ...\n");
585 for_each_online_cpu(cpu) {
586 if (cpu == first_cpu)
588 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
589 error = _cpu_down(cpu, 1);
590 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
592 cpumask_set_cpu(cpu, frozen_cpus);
594 pr_err("Error taking CPU%d down: %d\n", cpu, error);
600 BUG_ON(num_online_cpus() > 1);
602 pr_err("Non-boot CPUs are not disabled\n");
605 * Make sure the CPUs won't be enabled by someone else. We need to do
606 * this even in case of failure as all disable_nonboot_cpus() users are
607 * supposed to do enable_nonboot_cpus() on the failure path.
609 cpu_hotplug_disabled++;
611 cpu_maps_update_done();
615 void __weak arch_enable_nonboot_cpus_begin(void)
619 void __weak arch_enable_nonboot_cpus_end(void)
623 void enable_nonboot_cpus(void)
627 /* Allow everyone to use the CPU hotplug again */
628 cpu_maps_update_begin();
629 WARN_ON(--cpu_hotplug_disabled < 0);
630 if (cpumask_empty(frozen_cpus))
633 pr_info("Enabling non-boot CPUs ...\n");
635 arch_enable_nonboot_cpus_begin();
637 for_each_cpu(cpu, frozen_cpus) {
638 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
639 error = _cpu_up(cpu, 1);
640 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
642 pr_info("CPU%d is up\n", cpu);
645 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
648 arch_enable_nonboot_cpus_end();
650 cpumask_clear(frozen_cpus);
652 cpu_maps_update_done();
655 static int __init alloc_frozen_cpus(void)
657 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
661 core_initcall(alloc_frozen_cpus);
664 * When callbacks for CPU hotplug notifications are being executed, we must
665 * ensure that the state of the system with respect to the tasks being frozen
666 * or not, as reported by the notification, remains unchanged *throughout the
667 * duration* of the execution of the callbacks.
668 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
670 * This synchronization is implemented by mutually excluding regular CPU
671 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
672 * Hibernate notifications.
675 cpu_hotplug_pm_callback(struct notifier_block *nb,
676 unsigned long action, void *ptr)
680 case PM_SUSPEND_PREPARE:
681 case PM_HIBERNATION_PREPARE:
682 cpu_hotplug_disable();
685 case PM_POST_SUSPEND:
686 case PM_POST_HIBERNATION:
687 cpu_hotplug_enable();
698 static int __init cpu_hotplug_pm_sync_init(void)
701 * cpu_hotplug_pm_callback has higher priority than x86
702 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
703 * to disable cpu hotplug to avoid cpu hotplug race.
705 pm_notifier(cpu_hotplug_pm_callback, 0);
708 core_initcall(cpu_hotplug_pm_sync_init);
710 #endif /* CONFIG_PM_SLEEP_SMP */
713 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
714 * @cpu: cpu that just started
716 * This function calls the cpu_chain notifiers with CPU_STARTING.
717 * It must be called by the arch code on the new cpu, before the new cpu
718 * enables interrupts and before the "boot" cpu returns from __cpu_up().
720 void notify_cpu_starting(unsigned int cpu)
722 unsigned long val = CPU_STARTING;
724 #ifdef CONFIG_PM_SLEEP_SMP
725 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
726 val = CPU_STARTING_FROZEN;
727 #endif /* CONFIG_PM_SLEEP_SMP */
728 cpu_notify(val, (void *)(long)cpu);
731 #endif /* CONFIG_SMP */
734 * cpu_bit_bitmap[] is a special, "compressed" data structure that
735 * represents all NR_CPUS bits binary values of 1<<nr.
737 * It is used by cpumask_of() to get a constant address to a CPU
738 * mask value that has a single bit set only.
741 /* cpu_bit_bitmap[0] is empty - so we can back into it */
742 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
743 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
744 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
745 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
747 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
749 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
750 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
751 #if BITS_PER_LONG > 32
752 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
753 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
756 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
758 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
759 EXPORT_SYMBOL(cpu_all_bits);
761 #ifdef CONFIG_INIT_ALL_POSSIBLE
762 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
765 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
767 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
768 EXPORT_SYMBOL(cpu_possible_mask);
770 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
771 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
772 EXPORT_SYMBOL(cpu_online_mask);
774 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
775 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
776 EXPORT_SYMBOL(cpu_present_mask);
778 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
779 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
780 EXPORT_SYMBOL(cpu_active_mask);
782 void set_cpu_possible(unsigned int cpu, bool possible)
785 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
787 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
790 void set_cpu_present(unsigned int cpu, bool present)
793 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
795 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
798 void set_cpu_online(unsigned int cpu, bool online)
801 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
802 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
804 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
808 void set_cpu_active(unsigned int cpu, bool active)
811 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
813 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
816 void init_cpu_present(const struct cpumask *src)
818 cpumask_copy(to_cpumask(cpu_present_bits), src);
821 void init_cpu_possible(const struct cpumask *src)
823 cpumask_copy(to_cpumask(cpu_possible_bits), src);
826 void init_cpu_online(const struct cpumask *src)
828 cpumask_copy(to_cpumask(cpu_online_bits), src);