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 <linux/smpboot.h>
27 #include <trace/events/power.h>
28 #define CREATE_TRACE_POINTS
29 #include <trace/events/cpuhp.h>
34 * cpuhp_cpu_state - Per cpu hotplug state storage
35 * @state: The current cpu state
36 * @target: The target state
37 * @thread: Pointer to the hotplug thread
38 * @should_run: Thread should execute
39 * @cb_stat: The state for a single callback (install/uninstall)
40 * @cb: Single callback function (install/uninstall)
41 * @result: Result of the operation
42 * @done: Signal completion to the issuer of the task
44 struct cpuhp_cpu_state {
45 enum cpuhp_state state;
46 enum cpuhp_state target;
48 struct task_struct *thread;
50 enum cpuhp_state cb_state;
51 int (*cb)(unsigned int cpu);
53 struct completion done;
57 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
60 * cpuhp_step - Hotplug state machine step
61 * @name: Name of the step
62 * @startup: Startup function of the step
63 * @teardown: Teardown function of the step
64 * @skip_onerr: Do not invoke the functions on error rollback
65 * Will go away once the notifiers are gone
66 * @cant_stop: Bringup/teardown can't be stopped at this step
70 int (*startup)(unsigned int cpu);
71 int (*teardown)(unsigned int cpu);
76 static DEFINE_MUTEX(cpuhp_state_mutex);
77 static struct cpuhp_step cpuhp_bp_states[];
78 static struct cpuhp_step cpuhp_ap_states[];
81 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
82 * @cpu: The cpu for which the callback should be invoked
83 * @step: The step in the state machine
84 * @cb: The callback function to invoke
86 * Called from cpu hotplug and from the state register machinery
88 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state step,
89 int (*cb)(unsigned int))
91 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
95 trace_cpuhp_enter(cpu, st->target, step, cb);
97 trace_cpuhp_exit(cpu, st->state, step, ret);
103 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
104 static DEFINE_MUTEX(cpu_add_remove_lock);
105 bool cpuhp_tasks_frozen;
106 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
109 * The following two APIs (cpu_maps_update_begin/done) must be used when
110 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
111 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
112 * hotplug callback (un)registration performed using __register_cpu_notifier()
113 * or __unregister_cpu_notifier().
115 void cpu_maps_update_begin(void)
117 mutex_lock(&cpu_add_remove_lock);
119 EXPORT_SYMBOL(cpu_notifier_register_begin);
121 void cpu_maps_update_done(void)
123 mutex_unlock(&cpu_add_remove_lock);
125 EXPORT_SYMBOL(cpu_notifier_register_done);
127 static RAW_NOTIFIER_HEAD(cpu_chain);
129 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
130 * Should always be manipulated under cpu_add_remove_lock
132 static int cpu_hotplug_disabled;
134 #ifdef CONFIG_HOTPLUG_CPU
137 struct task_struct *active_writer;
138 /* wait queue to wake up the active_writer */
139 wait_queue_head_t wq;
140 /* verifies that no writer will get active while readers are active */
143 * Also blocks the new readers during
144 * an ongoing cpu hotplug operation.
148 #ifdef CONFIG_DEBUG_LOCK_ALLOC
149 struct lockdep_map dep_map;
152 .active_writer = NULL,
153 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
154 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
155 #ifdef CONFIG_DEBUG_LOCK_ALLOC
156 .dep_map = {.name = "cpu_hotplug.lock" },
160 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
161 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
162 #define cpuhp_lock_acquire_tryread() \
163 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
164 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
165 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
168 void get_online_cpus(void)
171 if (cpu_hotplug.active_writer == current)
173 cpuhp_lock_acquire_read();
174 mutex_lock(&cpu_hotplug.lock);
175 atomic_inc(&cpu_hotplug.refcount);
176 mutex_unlock(&cpu_hotplug.lock);
178 EXPORT_SYMBOL_GPL(get_online_cpus);
180 void put_online_cpus(void)
184 if (cpu_hotplug.active_writer == current)
187 refcount = atomic_dec_return(&cpu_hotplug.refcount);
188 if (WARN_ON(refcount < 0)) /* try to fix things up */
189 atomic_inc(&cpu_hotplug.refcount);
191 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
192 wake_up(&cpu_hotplug.wq);
194 cpuhp_lock_release();
197 EXPORT_SYMBOL_GPL(put_online_cpus);
200 * This ensures that the hotplug operation can begin only when the
201 * refcount goes to zero.
203 * Note that during a cpu-hotplug operation, the new readers, if any,
204 * will be blocked by the cpu_hotplug.lock
206 * Since cpu_hotplug_begin() is always called after invoking
207 * cpu_maps_update_begin(), we can be sure that only one writer is active.
209 * Note that theoretically, there is a possibility of a livelock:
210 * - Refcount goes to zero, last reader wakes up the sleeping
212 * - Last reader unlocks the cpu_hotplug.lock.
213 * - A new reader arrives at this moment, bumps up the refcount.
214 * - The writer acquires the cpu_hotplug.lock finds the refcount
215 * non zero and goes to sleep again.
217 * However, this is very difficult to achieve in practice since
218 * get_online_cpus() not an api which is called all that often.
221 void cpu_hotplug_begin(void)
225 cpu_hotplug.active_writer = current;
226 cpuhp_lock_acquire();
229 mutex_lock(&cpu_hotplug.lock);
230 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
231 if (likely(!atomic_read(&cpu_hotplug.refcount)))
233 mutex_unlock(&cpu_hotplug.lock);
236 finish_wait(&cpu_hotplug.wq, &wait);
239 void cpu_hotplug_done(void)
241 cpu_hotplug.active_writer = NULL;
242 mutex_unlock(&cpu_hotplug.lock);
243 cpuhp_lock_release();
247 * Wait for currently running CPU hotplug operations to complete (if any) and
248 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
249 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
250 * hotplug path before performing hotplug operations. So acquiring that lock
251 * guarantees mutual exclusion from any currently running hotplug operations.
253 void cpu_hotplug_disable(void)
255 cpu_maps_update_begin();
256 cpu_hotplug_disabled++;
257 cpu_maps_update_done();
259 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
261 void cpu_hotplug_enable(void)
263 cpu_maps_update_begin();
264 WARN_ON(--cpu_hotplug_disabled < 0);
265 cpu_maps_update_done();
267 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
268 #endif /* CONFIG_HOTPLUG_CPU */
270 /* Need to know about CPUs going up/down? */
271 int register_cpu_notifier(struct notifier_block *nb)
274 cpu_maps_update_begin();
275 ret = raw_notifier_chain_register(&cpu_chain, nb);
276 cpu_maps_update_done();
280 int __register_cpu_notifier(struct notifier_block *nb)
282 return raw_notifier_chain_register(&cpu_chain, nb);
285 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
288 unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
289 void *hcpu = (void *)(long)cpu;
293 ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
296 return notifier_to_errno(ret);
299 static int cpu_notify(unsigned long val, unsigned int cpu)
301 return __cpu_notify(val, cpu, -1, NULL);
304 /* Notifier wrappers for transitioning to state machine */
305 static int notify_prepare(unsigned int cpu)
310 ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
313 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
315 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
320 static int notify_online(unsigned int cpu)
322 cpu_notify(CPU_ONLINE, cpu);
326 static int notify_starting(unsigned int cpu)
328 cpu_notify(CPU_STARTING, cpu);
332 static int bringup_cpu(unsigned int cpu)
334 struct task_struct *idle = idle_thread_get(cpu);
337 /* Arch-specific enabling code. */
338 ret = __cpu_up(cpu, idle);
340 cpu_notify(CPU_UP_CANCELED, cpu);
343 BUG_ON(!cpu_online(cpu));
348 * Hotplug state machine related functions
350 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st,
351 struct cpuhp_step *steps)
353 for (st->state++; st->state < st->target; st->state++) {
354 struct cpuhp_step *step = steps + st->state;
356 if (!step->skip_onerr)
357 cpuhp_invoke_callback(cpu, st->state, step->startup);
361 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
362 struct cpuhp_step *steps, enum cpuhp_state target)
364 enum cpuhp_state prev_state = st->state;
367 for (; st->state > target; st->state--) {
368 struct cpuhp_step *step = steps + st->state;
370 ret = cpuhp_invoke_callback(cpu, st->state, step->teardown);
372 st->target = prev_state;
373 undo_cpu_down(cpu, st, steps);
380 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st,
381 struct cpuhp_step *steps)
383 for (st->state--; st->state > st->target; st->state--) {
384 struct cpuhp_step *step = steps + st->state;
386 if (!step->skip_onerr)
387 cpuhp_invoke_callback(cpu, st->state, step->teardown);
391 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
392 struct cpuhp_step *steps, enum cpuhp_state target)
394 enum cpuhp_state prev_state = st->state;
397 while (st->state < target) {
398 struct cpuhp_step *step;
401 step = steps + st->state;
402 ret = cpuhp_invoke_callback(cpu, st->state, step->startup);
404 st->target = prev_state;
405 undo_cpu_up(cpu, st, steps);
413 * The cpu hotplug threads manage the bringup and teardown of the cpus
415 static void cpuhp_create(unsigned int cpu)
417 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
419 init_completion(&st->done);
422 static int cpuhp_should_run(unsigned int cpu)
424 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
426 return st->should_run;
429 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
430 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
432 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
434 return cpuhp_down_callbacks(cpu, st, cpuhp_ap_states, target);
437 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
438 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
440 return cpuhp_up_callbacks(cpu, st, cpuhp_ap_states, st->target);
444 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
445 * callbacks when a state gets [un]installed at runtime.
447 static void cpuhp_thread_fun(unsigned int cpu)
449 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
453 * Paired with the mb() in cpuhp_kick_ap_work and
454 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
460 st->should_run = false;
462 /* Single callback invocation for [un]install ? */
464 if (st->cb_state < CPUHP_AP_ONLINE) {
466 ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
469 ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
472 /* Cannot happen .... */
473 BUG_ON(st->state < CPUHP_KICK_AP_THREAD);
475 /* Regular hotplug work */
476 if (st->state < st->target)
477 ret = cpuhp_ap_online(cpu, st);
478 else if (st->state > st->target)
479 ret = cpuhp_ap_offline(cpu, st);
485 /* Invoke a single callback on a remote cpu */
486 static int cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state,
487 int (*cb)(unsigned int))
489 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
491 if (!cpu_online(cpu))
494 st->cb_state = state;
497 * Make sure the above stores are visible before should_run becomes
498 * true. Paired with the mb() above in cpuhp_thread_fun()
501 st->should_run = true;
502 wake_up_process(st->thread);
503 wait_for_completion(&st->done);
507 /* Regular hotplug invocation of the AP hotplug thread */
508 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
513 * Make sure the above stores are visible before should_run becomes
514 * true. Paired with the mb() above in cpuhp_thread_fun()
517 st->should_run = true;
518 wake_up_process(st->thread);
521 static int cpuhp_kick_ap_work(unsigned int cpu)
523 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
524 enum cpuhp_state state = st->state;
526 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
527 __cpuhp_kick_ap_work(st);
528 wait_for_completion(&st->done);
529 trace_cpuhp_exit(cpu, st->state, state, st->result);
533 static struct smp_hotplug_thread cpuhp_threads = {
534 .store = &cpuhp_state.thread,
535 .create = &cpuhp_create,
536 .thread_should_run = cpuhp_should_run,
537 .thread_fn = cpuhp_thread_fun,
538 .thread_comm = "cpuhp/%u",
542 void __init cpuhp_threads_init(void)
544 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
545 kthread_unpark(this_cpu_read(cpuhp_state.thread));
548 #ifdef CONFIG_HOTPLUG_CPU
549 EXPORT_SYMBOL(register_cpu_notifier);
550 EXPORT_SYMBOL(__register_cpu_notifier);
551 void unregister_cpu_notifier(struct notifier_block *nb)
553 cpu_maps_update_begin();
554 raw_notifier_chain_unregister(&cpu_chain, nb);
555 cpu_maps_update_done();
557 EXPORT_SYMBOL(unregister_cpu_notifier);
559 void __unregister_cpu_notifier(struct notifier_block *nb)
561 raw_notifier_chain_unregister(&cpu_chain, nb);
563 EXPORT_SYMBOL(__unregister_cpu_notifier);
566 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
569 * This function walks all processes, finds a valid mm struct for each one and
570 * then clears a corresponding bit in mm's cpumask. While this all sounds
571 * trivial, there are various non-obvious corner cases, which this function
572 * tries to solve in a safe manner.
574 * Also note that the function uses a somewhat relaxed locking scheme, so it may
575 * be called only for an already offlined CPU.
577 void clear_tasks_mm_cpumask(int cpu)
579 struct task_struct *p;
582 * This function is called after the cpu is taken down and marked
583 * offline, so its not like new tasks will ever get this cpu set in
584 * their mm mask. -- Peter Zijlstra
585 * Thus, we may use rcu_read_lock() here, instead of grabbing
586 * full-fledged tasklist_lock.
588 WARN_ON(cpu_online(cpu));
590 for_each_process(p) {
591 struct task_struct *t;
594 * Main thread might exit, but other threads may still have
595 * a valid mm. Find one.
597 t = find_lock_task_mm(p);
600 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
606 static inline void check_for_tasks(int dead_cpu)
608 struct task_struct *g, *p;
610 read_lock(&tasklist_lock);
611 for_each_process_thread(g, p) {
615 * We do the check with unlocked task_rq(p)->lock.
616 * Order the reading to do not warn about a task,
617 * which was running on this cpu in the past, and
618 * it's just been woken on another cpu.
621 if (task_cpu(p) != dead_cpu)
624 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
625 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
627 read_unlock(&tasklist_lock);
630 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
632 BUG_ON(cpu_notify(val, cpu));
635 static int notify_down_prepare(unsigned int cpu)
637 int err, nr_calls = 0;
639 err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
642 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
643 pr_warn("%s: attempt to take down CPU %u failed\n",
649 static int notify_dying(unsigned int cpu)
651 cpu_notify(CPU_DYING, cpu);
655 /* Take this CPU down. */
656 static int take_cpu_down(void *_param)
658 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
659 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
660 int err, cpu = smp_processor_id();
662 /* Ensure this CPU doesn't handle any more interrupts. */
663 err = __cpu_disable();
667 /* Invoke the former CPU_DYING callbacks */
668 for (; st->state > target; st->state--) {
669 struct cpuhp_step *step = cpuhp_ap_states + st->state;
671 cpuhp_invoke_callback(cpu, st->state, step->teardown);
673 /* Give up timekeeping duties */
674 tick_handover_do_timer();
675 /* Park the stopper thread */
676 stop_machine_park(cpu);
680 static int takedown_cpu(unsigned int cpu)
685 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
686 * and RCU users of this state to go away such that all new such users
689 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
690 * not imply sync_sched(), so wait for both.
692 * Do sync before park smpboot threads to take care the rcu boost case.
694 if (IS_ENABLED(CONFIG_PREEMPT))
695 synchronize_rcu_mult(call_rcu, call_rcu_sched);
699 /* Park the hotplug thread */
700 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
703 * Prevent irq alloc/free while the dying cpu reorganizes the
704 * interrupt affinities.
709 * So now all preempt/rcu users must observe !cpu_active().
711 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
713 /* CPU didn't die: tell everyone. Can't complain. */
714 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
718 BUG_ON(cpu_online(cpu));
721 * The migration_call() CPU_DYING callback will have removed all
722 * runnable tasks from the cpu, there's only the idle task left now
723 * that the migration thread is done doing the stop_machine thing.
725 * Wait for the stop thread to go away.
727 while (!per_cpu(cpu_dead_idle, cpu))
729 smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
730 per_cpu(cpu_dead_idle, cpu) = false;
732 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
735 hotplug_cpu__broadcast_tick_pull(cpu);
736 /* This actually kills the CPU. */
739 tick_cleanup_dead_cpu(cpu);
743 static int notify_dead(unsigned int cpu)
745 cpu_notify_nofail(CPU_DEAD, cpu);
746 check_for_tasks(cpu);
751 #define notify_down_prepare NULL
752 #define takedown_cpu NULL
753 #define notify_dead NULL
754 #define notify_dying NULL
757 #ifdef CONFIG_HOTPLUG_CPU
759 /* Requires cpu_add_remove_lock to be held */
760 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
761 enum cpuhp_state target)
763 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
764 int prev_state, ret = 0;
765 bool hasdied = false;
767 if (num_online_cpus() == 1)
770 if (!cpu_present(cpu))
775 cpuhp_tasks_frozen = tasks_frozen;
777 prev_state = st->state;
780 * If the current CPU state is in the range of the AP hotplug thread,
781 * then we need to kick the thread.
783 if (st->state >= CPUHP_KICK_AP_THREAD) {
784 ret = cpuhp_kick_ap_work(cpu);
786 * The AP side has done the error rollback already. Just
787 * return the error code..
793 * We might have stopped still in the range of the AP hotplug
794 * thread. Nothing to do anymore.
796 if (st->state >= CPUHP_KICK_AP_THREAD)
800 * The AP brought itself down below CPUHP_KICK_AP_THREAD. So we need
801 * to do the further cleanups.
803 ret = cpuhp_down_callbacks(cpu, st, cpuhp_bp_states, target);
805 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
808 /* This post dead nonsense must die */
810 cpu_notify_nofail(CPU_POST_DEAD, cpu);
814 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
818 cpu_maps_update_begin();
820 if (cpu_hotplug_disabled) {
825 err = _cpu_down(cpu, 0, target);
828 cpu_maps_update_done();
831 int cpu_down(unsigned int cpu)
833 return do_cpu_down(cpu, CPUHP_OFFLINE);
835 EXPORT_SYMBOL(cpu_down);
836 #endif /*CONFIG_HOTPLUG_CPU*/
839 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
840 * @cpu: cpu that just started
842 * This function calls the cpu_chain notifiers with CPU_STARTING.
843 * It must be called by the arch code on the new cpu, before the new cpu
844 * enables interrupts and before the "boot" cpu returns from __cpu_up().
846 void notify_cpu_starting(unsigned int cpu)
848 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
849 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
851 while (st->state < target) {
852 struct cpuhp_step *step;
855 step = cpuhp_ap_states + st->state;
856 cpuhp_invoke_callback(cpu, st->state, step->startup);
861 * Called from the idle task. We need to set active here, so we can kick off
862 * the stopper thread.
864 static int cpuhp_set_cpu_active(unsigned int cpu)
866 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
868 /* The cpu is marked online, set it active now */
869 set_cpu_active(cpu, true);
870 /* Unpark the stopper thread and the hotplug thread */
871 stop_machine_unpark(cpu);
872 kthread_unpark(st->thread);
876 /* Requires cpu_add_remove_lock to be held */
877 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
879 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
880 struct task_struct *idle;
885 if (!cpu_present(cpu)) {
891 * The caller of do_cpu_up might have raced with another
892 * caller. Ignore it for now.
894 if (st->state >= target)
897 if (st->state == CPUHP_OFFLINE) {
898 /* Let it fail before we try to bring the cpu up */
899 idle = idle_thread_get(cpu);
906 cpuhp_tasks_frozen = tasks_frozen;
910 * If the current CPU state is in the range of the AP hotplug thread,
911 * then we need to kick the thread once more.
913 if (st->state >= CPUHP_KICK_AP_THREAD) {
914 ret = cpuhp_kick_ap_work(cpu);
916 * The AP side has done the error rollback already. Just
917 * return the error code..
924 * Try to reach the target state. We max out on the BP at
925 * CPUHP_KICK_AP_THREAD. After that the AP hotplug thread is
926 * responsible for bringing it up to the target state.
928 target = min((int)target, CPUHP_KICK_AP_THREAD);
929 ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
935 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
939 if (!cpu_possible(cpu)) {
940 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
942 #if defined(CONFIG_IA64)
943 pr_err("please check additional_cpus= boot parameter\n");
948 err = try_online_node(cpu_to_node(cpu));
952 cpu_maps_update_begin();
954 if (cpu_hotplug_disabled) {
959 err = _cpu_up(cpu, 0, target);
961 cpu_maps_update_done();
965 int cpu_up(unsigned int cpu)
967 return do_cpu_up(cpu, CPUHP_ONLINE);
969 EXPORT_SYMBOL_GPL(cpu_up);
971 #ifdef CONFIG_PM_SLEEP_SMP
972 static cpumask_var_t frozen_cpus;
974 int disable_nonboot_cpus(void)
976 int cpu, first_cpu, error = 0;
978 cpu_maps_update_begin();
979 first_cpu = cpumask_first(cpu_online_mask);
981 * We take down all of the non-boot CPUs in one shot to avoid races
982 * with the userspace trying to use the CPU hotplug at the same time
984 cpumask_clear(frozen_cpus);
986 pr_info("Disabling non-boot CPUs ...\n");
987 for_each_online_cpu(cpu) {
988 if (cpu == first_cpu)
990 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
991 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
992 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
994 cpumask_set_cpu(cpu, frozen_cpus);
996 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1002 BUG_ON(num_online_cpus() > 1);
1004 pr_err("Non-boot CPUs are not disabled\n");
1007 * Make sure the CPUs won't be enabled by someone else. We need to do
1008 * this even in case of failure as all disable_nonboot_cpus() users are
1009 * supposed to do enable_nonboot_cpus() on the failure path.
1011 cpu_hotplug_disabled++;
1013 cpu_maps_update_done();
1017 void __weak arch_enable_nonboot_cpus_begin(void)
1021 void __weak arch_enable_nonboot_cpus_end(void)
1025 void enable_nonboot_cpus(void)
1029 /* Allow everyone to use the CPU hotplug again */
1030 cpu_maps_update_begin();
1031 WARN_ON(--cpu_hotplug_disabled < 0);
1032 if (cpumask_empty(frozen_cpus))
1035 pr_info("Enabling non-boot CPUs ...\n");
1037 arch_enable_nonboot_cpus_begin();
1039 for_each_cpu(cpu, frozen_cpus) {
1040 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1041 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1042 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1044 pr_info("CPU%d is up\n", cpu);
1047 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1050 arch_enable_nonboot_cpus_end();
1052 cpumask_clear(frozen_cpus);
1054 cpu_maps_update_done();
1057 static int __init alloc_frozen_cpus(void)
1059 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1063 core_initcall(alloc_frozen_cpus);
1066 * When callbacks for CPU hotplug notifications are being executed, we must
1067 * ensure that the state of the system with respect to the tasks being frozen
1068 * or not, as reported by the notification, remains unchanged *throughout the
1069 * duration* of the execution of the callbacks.
1070 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1072 * This synchronization is implemented by mutually excluding regular CPU
1073 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1074 * Hibernate notifications.
1077 cpu_hotplug_pm_callback(struct notifier_block *nb,
1078 unsigned long action, void *ptr)
1082 case PM_SUSPEND_PREPARE:
1083 case PM_HIBERNATION_PREPARE:
1084 cpu_hotplug_disable();
1087 case PM_POST_SUSPEND:
1088 case PM_POST_HIBERNATION:
1089 cpu_hotplug_enable();
1100 static int __init cpu_hotplug_pm_sync_init(void)
1103 * cpu_hotplug_pm_callback has higher priority than x86
1104 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1105 * to disable cpu hotplug to avoid cpu hotplug race.
1107 pm_notifier(cpu_hotplug_pm_callback, 0);
1110 core_initcall(cpu_hotplug_pm_sync_init);
1112 #endif /* CONFIG_PM_SLEEP_SMP */
1114 #endif /* CONFIG_SMP */
1116 /* Boot processor state steps */
1117 static struct cpuhp_step cpuhp_bp_states[] = {
1124 [CPUHP_CREATE_THREADS]= {
1125 .name = "threads:create",
1126 .startup = smpboot_create_threads,
1130 [CPUHP_NOTIFY_PREPARE] = {
1131 .name = "notify:prepare",
1132 .startup = notify_prepare,
1133 .teardown = notify_dead,
1137 [CPUHP_BRINGUP_CPU] = {
1138 .name = "cpu:bringup",
1139 .startup = bringup_cpu,
1143 [CPUHP_TEARDOWN_CPU] = {
1144 .name = "cpu:teardown",
1146 .teardown = takedown_cpu,
1149 [CPUHP_CPU_SET_ACTIVE] = {
1150 .name = "cpu:active",
1151 .startup = cpuhp_set_cpu_active,
1154 [CPUHP_KICK_AP_THREAD] = {
1155 .name = "cpuhp:kickthread",
1156 .startup = cpuhp_kick_ap_work,
1157 .teardown = cpuhp_kick_ap_work,
1160 [CPUHP_BP_ONLINE] = {
1167 /* Application processor state steps */
1168 static struct cpuhp_step cpuhp_ap_states[] = {
1170 [CPUHP_AP_NOTIFY_STARTING] = {
1171 .name = "notify:starting",
1172 .startup = notify_starting,
1173 .teardown = notify_dying,
1177 [CPUHP_AP_SMPBOOT_THREADS] = {
1178 .name = "smpboot:threads",
1179 .startup = smpboot_unpark_threads,
1180 .teardown = smpboot_park_threads,
1182 [CPUHP_AP_NOTIFY_ONLINE] = {
1183 .name = "notify:online",
1184 .startup = notify_online,
1185 .teardown = notify_down_prepare,
1195 /* Sanity check for callbacks */
1196 static int cpuhp_cb_check(enum cpuhp_state state)
1198 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1203 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1205 if (state >= CPUHP_AP_OFFLINE && state <= CPUHP_AP_ONLINE)
1207 return state > CPUHP_BP_ONLINE;
1210 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1212 struct cpuhp_step *sp;
1214 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1218 static void cpuhp_store_callbacks(enum cpuhp_state state,
1220 int (*startup)(unsigned int cpu),
1221 int (*teardown)(unsigned int cpu))
1223 /* (Un)Install the callbacks for further cpu hotplug operations */
1224 struct cpuhp_step *sp;
1226 mutex_lock(&cpuhp_state_mutex);
1227 sp = cpuhp_get_step(state);
1228 sp->startup = startup;
1229 sp->teardown = teardown;
1231 mutex_unlock(&cpuhp_state_mutex);
1234 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1236 return cpuhp_get_step(state)->teardown;
1240 * Call the startup/teardown function for a step either on the AP or
1241 * on the current CPU.
1243 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1244 int (*cb)(unsigned int), bool bringup)
1251 * The non AP bound callbacks can fail on bringup. On teardown
1252 * e.g. module removal we crash for now.
1255 if (cpuhp_is_ap_state(state))
1256 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1258 ret = cpuhp_invoke_callback(cpu, state, cb);
1260 ret = cpuhp_invoke_callback(cpu, state, cb);
1262 BUG_ON(ret && !bringup);
1267 * Called from __cpuhp_setup_state on a recoverable failure.
1269 * Note: The teardown callbacks for rollback are not allowed to fail!
1271 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1272 int (*teardown)(unsigned int cpu))
1279 /* Roll back the already executed steps on the other cpus */
1280 for_each_present_cpu(cpu) {
1281 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1282 int cpustate = st->state;
1284 if (cpu >= failedcpu)
1287 /* Did we invoke the startup call on that cpu ? */
1288 if (cpustate >= state)
1289 cpuhp_issue_call(cpu, state, teardown, false);
1294 * Returns a free for dynamic slot assignment of the Online state. The states
1295 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1296 * by having no name assigned.
1298 static int cpuhp_reserve_state(enum cpuhp_state state)
1302 mutex_lock(&cpuhp_state_mutex);
1303 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1304 if (cpuhp_ap_states[i].name)
1307 cpuhp_ap_states[i].name = "Reserved";
1308 mutex_unlock(&cpuhp_state_mutex);
1311 mutex_unlock(&cpuhp_state_mutex);
1312 WARN(1, "No more dynamic states available for CPU hotplug\n");
1317 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1318 * @state: The state to setup
1319 * @invoke: If true, the startup function is invoked for cpus where
1320 * cpu state >= @state
1321 * @startup: startup callback function
1322 * @teardown: teardown callback function
1324 * Returns 0 if successful, otherwise a proper error code
1326 int __cpuhp_setup_state(enum cpuhp_state state,
1327 const char *name, bool invoke,
1328 int (*startup)(unsigned int cpu),
1329 int (*teardown)(unsigned int cpu))
1334 if (cpuhp_cb_check(state) || !name)
1339 /* currently assignments for the ONLINE state are possible */
1340 if (state == CPUHP_AP_ONLINE_DYN) {
1342 ret = cpuhp_reserve_state(state);
1348 cpuhp_store_callbacks(state, name, startup, teardown);
1350 if (!invoke || !startup)
1354 * Try to call the startup callback for each present cpu
1355 * depending on the hotplug state of the cpu.
1357 for_each_present_cpu(cpu) {
1358 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1359 int cpustate = st->state;
1361 if (cpustate < state)
1364 ret = cpuhp_issue_call(cpu, state, startup, true);
1366 cpuhp_rollback_install(cpu, state, teardown);
1367 cpuhp_store_callbacks(state, NULL, NULL, NULL);
1373 if (!ret && dyn_state)
1377 EXPORT_SYMBOL(__cpuhp_setup_state);
1380 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1381 * @state: The state to remove
1382 * @invoke: If true, the teardown function is invoked for cpus where
1383 * cpu state >= @state
1385 * The teardown callback is currently not allowed to fail. Think
1386 * about module removal!
1388 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1390 int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1393 BUG_ON(cpuhp_cb_check(state));
1397 if (!invoke || !teardown)
1401 * Call the teardown callback for each present cpu depending
1402 * on the hotplug state of the cpu. This function is not
1403 * allowed to fail currently!
1405 for_each_present_cpu(cpu) {
1406 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1407 int cpustate = st->state;
1409 if (cpustate >= state)
1410 cpuhp_issue_call(cpu, state, teardown, false);
1413 cpuhp_store_callbacks(state, NULL, NULL, NULL);
1416 EXPORT_SYMBOL(__cpuhp_remove_state);
1418 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1419 static ssize_t show_cpuhp_state(struct device *dev,
1420 struct device_attribute *attr, char *buf)
1422 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1424 return sprintf(buf, "%d\n", st->state);
1426 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1428 static ssize_t write_cpuhp_target(struct device *dev,
1429 struct device_attribute *attr,
1430 const char *buf, size_t count)
1432 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1433 struct cpuhp_step *sp;
1436 ret = kstrtoint(buf, 10, &target);
1440 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1441 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1444 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1448 ret = lock_device_hotplug_sysfs();
1452 mutex_lock(&cpuhp_state_mutex);
1453 sp = cpuhp_get_step(target);
1454 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1455 mutex_unlock(&cpuhp_state_mutex);
1459 if (st->state < target)
1460 ret = do_cpu_up(dev->id, target);
1462 ret = do_cpu_down(dev->id, target);
1464 unlock_device_hotplug();
1465 return ret ? ret : count;
1468 static ssize_t show_cpuhp_target(struct device *dev,
1469 struct device_attribute *attr, char *buf)
1471 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1473 return sprintf(buf, "%d\n", st->target);
1475 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1477 static struct attribute *cpuhp_cpu_attrs[] = {
1478 &dev_attr_state.attr,
1479 &dev_attr_target.attr,
1483 static struct attribute_group cpuhp_cpu_attr_group = {
1484 .attrs = cpuhp_cpu_attrs,
1489 static ssize_t show_cpuhp_states(struct device *dev,
1490 struct device_attribute *attr, char *buf)
1492 ssize_t cur, res = 0;
1495 mutex_lock(&cpuhp_state_mutex);
1496 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1497 struct cpuhp_step *sp = cpuhp_get_step(i);
1500 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1505 mutex_unlock(&cpuhp_state_mutex);
1508 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1510 static struct attribute *cpuhp_cpu_root_attrs[] = {
1511 &dev_attr_states.attr,
1515 static struct attribute_group cpuhp_cpu_root_attr_group = {
1516 .attrs = cpuhp_cpu_root_attrs,
1521 static int __init cpuhp_sysfs_init(void)
1525 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1526 &cpuhp_cpu_root_attr_group);
1530 for_each_possible_cpu(cpu) {
1531 struct device *dev = get_cpu_device(cpu);
1535 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1541 device_initcall(cpuhp_sysfs_init);
1545 * cpu_bit_bitmap[] is a special, "compressed" data structure that
1546 * represents all NR_CPUS bits binary values of 1<<nr.
1548 * It is used by cpumask_of() to get a constant address to a CPU
1549 * mask value that has a single bit set only.
1552 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1553 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
1554 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1555 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1556 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1558 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1560 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
1561 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
1562 #if BITS_PER_LONG > 32
1563 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
1564 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
1567 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1569 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1570 EXPORT_SYMBOL(cpu_all_bits);
1572 #ifdef CONFIG_INIT_ALL_POSSIBLE
1573 struct cpumask __cpu_possible_mask __read_mostly
1576 struct cpumask __cpu_possible_mask __read_mostly;
1578 EXPORT_SYMBOL(__cpu_possible_mask);
1580 struct cpumask __cpu_online_mask __read_mostly;
1581 EXPORT_SYMBOL(__cpu_online_mask);
1583 struct cpumask __cpu_present_mask __read_mostly;
1584 EXPORT_SYMBOL(__cpu_present_mask);
1586 struct cpumask __cpu_active_mask __read_mostly;
1587 EXPORT_SYMBOL(__cpu_active_mask);
1589 void init_cpu_present(const struct cpumask *src)
1591 cpumask_copy(&__cpu_present_mask, src);
1594 void init_cpu_possible(const struct cpumask *src)
1596 cpumask_copy(&__cpu_possible_mask, src);
1599 void init_cpu_online(const struct cpumask *src)
1601 cpumask_copy(&__cpu_online_mask, src);
1605 * Activate the first processor.
1607 void __init boot_cpu_init(void)
1609 int cpu = smp_processor_id();
1611 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1612 set_cpu_online(cpu, true);
1613 set_cpu_active(cpu, true);
1614 set_cpu_present(cpu, true);
1615 set_cpu_possible(cpu, true);
1619 * Must be called _AFTER_ setting up the per_cpu areas
1621 void __init boot_cpu_state_init(void)
1623 per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;