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 static void cpu_notify_nofail(unsigned long val, void *v)
228 BUG_ON(cpu_notify(val, v));
230 EXPORT_SYMBOL(register_cpu_notifier);
231 EXPORT_SYMBOL(__register_cpu_notifier);
233 void unregister_cpu_notifier(struct notifier_block *nb)
235 cpu_maps_update_begin();
236 raw_notifier_chain_unregister(&cpu_chain, nb);
237 cpu_maps_update_done();
239 EXPORT_SYMBOL(unregister_cpu_notifier);
241 void __unregister_cpu_notifier(struct notifier_block *nb)
243 raw_notifier_chain_unregister(&cpu_chain, nb);
245 EXPORT_SYMBOL(__unregister_cpu_notifier);
247 #ifdef CONFIG_HOTPLUG_CPU
249 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
252 * This function walks all processes, finds a valid mm struct for each one and
253 * then clears a corresponding bit in mm's cpumask. While this all sounds
254 * trivial, there are various non-obvious corner cases, which this function
255 * tries to solve in a safe manner.
257 * Also note that the function uses a somewhat relaxed locking scheme, so it may
258 * be called only for an already offlined CPU.
260 void clear_tasks_mm_cpumask(int cpu)
262 struct task_struct *p;
265 * This function is called after the cpu is taken down and marked
266 * offline, so its not like new tasks will ever get this cpu set in
267 * their mm mask. -- Peter Zijlstra
268 * Thus, we may use rcu_read_lock() here, instead of grabbing
269 * full-fledged tasklist_lock.
271 WARN_ON(cpu_online(cpu));
273 for_each_process(p) {
274 struct task_struct *t;
277 * Main thread might exit, but other threads may still have
278 * a valid mm. Find one.
280 t = find_lock_task_mm(p);
283 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
289 static inline void check_for_tasks(int dead_cpu)
291 struct task_struct *g, *p;
293 read_lock(&tasklist_lock);
294 for_each_process_thread(g, p) {
298 * We do the check with unlocked task_rq(p)->lock.
299 * Order the reading to do not warn about a task,
300 * which was running on this cpu in the past, and
301 * it's just been woken on another cpu.
304 if (task_cpu(p) != dead_cpu)
307 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
308 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
310 read_unlock(&tasklist_lock);
313 struct take_cpu_down_param {
318 /* Take this CPU down. */
319 static int take_cpu_down(void *_param)
321 struct take_cpu_down_param *param = _param;
324 /* Ensure this CPU doesn't handle any more interrupts. */
325 err = __cpu_disable();
329 cpu_notify(CPU_DYING | param->mod, param->hcpu);
330 /* Give up timekeeping duties */
331 tick_handover_do_timer();
332 /* Park the stopper thread */
333 stop_machine_park((long)param->hcpu);
337 /* Requires cpu_add_remove_lock to be held */
338 static int _cpu_down(unsigned int cpu, int tasks_frozen)
340 int err, nr_calls = 0;
341 void *hcpu = (void *)(long)cpu;
342 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
343 struct take_cpu_down_param tcd_param = {
348 if (num_online_cpus() == 1)
351 if (!cpu_online(cpu))
356 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
359 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
360 pr_warn("%s: attempt to take down CPU %u failed\n",
366 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
367 * and RCU users of this state to go away such that all new such users
370 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
371 * not imply sync_sched(), so wait for both.
373 * Do sync before park smpboot threads to take care the rcu boost case.
375 if (IS_ENABLED(CONFIG_PREEMPT))
376 synchronize_rcu_mult(call_rcu, call_rcu_sched);
380 smpboot_park_threads(cpu);
383 * Prevent irq alloc/free while the dying cpu reorganizes the
384 * interrupt affinities.
389 * So now all preempt/rcu users must observe !cpu_active().
391 err = stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
393 /* CPU didn't die: tell everyone. Can't complain. */
394 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
398 BUG_ON(cpu_online(cpu));
401 * The migration_call() CPU_DYING callback will have removed all
402 * runnable tasks from the cpu, there's only the idle task left now
403 * that the migration thread is done doing the stop_machine thing.
405 * Wait for the stop thread to go away.
407 while (!per_cpu(cpu_dead_idle, cpu))
409 smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
410 per_cpu(cpu_dead_idle, cpu) = false;
412 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
415 hotplug_cpu__broadcast_tick_pull(cpu);
416 /* This actually kills the CPU. */
419 /* CPU is completely dead: tell everyone. Too late to complain. */
420 tick_cleanup_dead_cpu(cpu);
421 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
423 check_for_tasks(cpu);
428 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
432 int cpu_down(unsigned int cpu)
436 cpu_maps_update_begin();
438 if (cpu_hotplug_disabled) {
443 err = _cpu_down(cpu, 0);
446 cpu_maps_update_done();
449 EXPORT_SYMBOL(cpu_down);
450 #endif /*CONFIG_HOTPLUG_CPU*/
453 * Unpark per-CPU smpboot kthreads at CPU-online time.
455 static int smpboot_thread_call(struct notifier_block *nfb,
456 unsigned long action, void *hcpu)
458 int cpu = (long)hcpu;
460 switch (action & ~CPU_TASKS_FROZEN) {
462 case CPU_DOWN_FAILED:
464 smpboot_unpark_threads(cpu);
474 static struct notifier_block smpboot_thread_notifier = {
475 .notifier_call = smpboot_thread_call,
476 .priority = CPU_PRI_SMPBOOT,
479 void smpboot_thread_init(void)
481 register_cpu_notifier(&smpboot_thread_notifier);
484 /* Requires cpu_add_remove_lock to be held */
485 static int _cpu_up(unsigned int cpu, int tasks_frozen)
487 int ret, nr_calls = 0;
488 void *hcpu = (void *)(long)cpu;
489 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
490 struct task_struct *idle;
494 if (cpu_online(cpu) || !cpu_present(cpu)) {
499 idle = idle_thread_get(cpu);
505 ret = smpboot_create_threads(cpu);
509 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
512 pr_warn("%s: attempt to bring up CPU %u failed\n",
517 /* Arch-specific enabling code. */
518 ret = __cpu_up(cpu, idle);
522 BUG_ON(!cpu_online(cpu));
524 /* Now call notifier in preparation. */
525 cpu_notify(CPU_ONLINE | mod, hcpu);
529 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
536 int cpu_up(unsigned int cpu)
540 if (!cpu_possible(cpu)) {
541 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
543 #if defined(CONFIG_IA64)
544 pr_err("please check additional_cpus= boot parameter\n");
549 err = try_online_node(cpu_to_node(cpu));
553 cpu_maps_update_begin();
555 if (cpu_hotplug_disabled) {
560 err = _cpu_up(cpu, 0);
563 cpu_maps_update_done();
566 EXPORT_SYMBOL_GPL(cpu_up);
568 #ifdef CONFIG_PM_SLEEP_SMP
569 static cpumask_var_t frozen_cpus;
571 int disable_nonboot_cpus(void)
573 int cpu, first_cpu, error = 0;
575 cpu_maps_update_begin();
576 first_cpu = cpumask_first(cpu_online_mask);
578 * We take down all of the non-boot CPUs in one shot to avoid races
579 * with the userspace trying to use the CPU hotplug at the same time
581 cpumask_clear(frozen_cpus);
583 pr_info("Disabling non-boot CPUs ...\n");
584 for_each_online_cpu(cpu) {
585 if (cpu == first_cpu)
587 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
588 error = _cpu_down(cpu, 1);
589 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
591 cpumask_set_cpu(cpu, frozen_cpus);
593 pr_err("Error taking CPU%d down: %d\n", cpu, error);
599 BUG_ON(num_online_cpus() > 1);
601 pr_err("Non-boot CPUs are not disabled\n");
604 * Make sure the CPUs won't be enabled by someone else. We need to do
605 * this even in case of failure as all disable_nonboot_cpus() users are
606 * supposed to do enable_nonboot_cpus() on the failure path.
608 cpu_hotplug_disabled++;
610 cpu_maps_update_done();
614 void __weak arch_enable_nonboot_cpus_begin(void)
618 void __weak arch_enable_nonboot_cpus_end(void)
622 void enable_nonboot_cpus(void)
626 /* Allow everyone to use the CPU hotplug again */
627 cpu_maps_update_begin();
628 WARN_ON(--cpu_hotplug_disabled < 0);
629 if (cpumask_empty(frozen_cpus))
632 pr_info("Enabling non-boot CPUs ...\n");
634 arch_enable_nonboot_cpus_begin();
636 for_each_cpu(cpu, frozen_cpus) {
637 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
638 error = _cpu_up(cpu, 1);
639 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
641 pr_info("CPU%d is up\n", cpu);
644 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
647 arch_enable_nonboot_cpus_end();
649 cpumask_clear(frozen_cpus);
651 cpu_maps_update_done();
654 static int __init alloc_frozen_cpus(void)
656 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
660 core_initcall(alloc_frozen_cpus);
663 * When callbacks for CPU hotplug notifications are being executed, we must
664 * ensure that the state of the system with respect to the tasks being frozen
665 * or not, as reported by the notification, remains unchanged *throughout the
666 * duration* of the execution of the callbacks.
667 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
669 * This synchronization is implemented by mutually excluding regular CPU
670 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
671 * Hibernate notifications.
674 cpu_hotplug_pm_callback(struct notifier_block *nb,
675 unsigned long action, void *ptr)
679 case PM_SUSPEND_PREPARE:
680 case PM_HIBERNATION_PREPARE:
681 cpu_hotplug_disable();
684 case PM_POST_SUSPEND:
685 case PM_POST_HIBERNATION:
686 cpu_hotplug_enable();
697 static int __init cpu_hotplug_pm_sync_init(void)
700 * cpu_hotplug_pm_callback has higher priority than x86
701 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
702 * to disable cpu hotplug to avoid cpu hotplug race.
704 pm_notifier(cpu_hotplug_pm_callback, 0);
707 core_initcall(cpu_hotplug_pm_sync_init);
709 #endif /* CONFIG_PM_SLEEP_SMP */
712 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
713 * @cpu: cpu that just started
715 * This function calls the cpu_chain notifiers with CPU_STARTING.
716 * It must be called by the arch code on the new cpu, before the new cpu
717 * enables interrupts and before the "boot" cpu returns from __cpu_up().
719 void notify_cpu_starting(unsigned int cpu)
721 unsigned long val = CPU_STARTING;
723 #ifdef CONFIG_PM_SLEEP_SMP
724 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
725 val = CPU_STARTING_FROZEN;
726 #endif /* CONFIG_PM_SLEEP_SMP */
727 cpu_notify(val, (void *)(long)cpu);
730 #endif /* CONFIG_SMP */
733 * cpu_bit_bitmap[] is a special, "compressed" data structure that
734 * represents all NR_CPUS bits binary values of 1<<nr.
736 * It is used by cpumask_of() to get a constant address to a CPU
737 * mask value that has a single bit set only.
740 /* cpu_bit_bitmap[0] is empty - so we can back into it */
741 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
742 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
743 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
744 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
746 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
748 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
749 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
750 #if BITS_PER_LONG > 32
751 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
752 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
755 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
757 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
758 EXPORT_SYMBOL(cpu_all_bits);
760 #ifdef CONFIG_INIT_ALL_POSSIBLE
761 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
764 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
766 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
767 EXPORT_SYMBOL(cpu_possible_mask);
769 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
770 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
771 EXPORT_SYMBOL(cpu_online_mask);
773 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
774 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
775 EXPORT_SYMBOL(cpu_present_mask);
777 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
778 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
779 EXPORT_SYMBOL(cpu_active_mask);
781 void set_cpu_possible(unsigned int cpu, bool possible)
784 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
786 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
789 void set_cpu_present(unsigned int cpu, bool present)
792 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
794 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
797 void set_cpu_online(unsigned int cpu, bool online)
800 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
801 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
803 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
807 void set_cpu_active(unsigned int cpu, bool active)
810 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
812 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
815 void init_cpu_present(const struct cpumask *src)
817 cpumask_copy(to_cpumask(cpu_present_bits), src);
820 void init_cpu_possible(const struct cpumask *src)
822 cpumask_copy(to_cpumask(cpu_possible_bits), src);
825 void init_cpu_online(const struct cpumask *src)
827 cpumask_copy(to_cpumask(cpu_online_bits), src);