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 * @rollback: Perform a rollback
40 * @cb_stat: The state for a single callback (install/uninstall)
41 * @cb: Single callback function (install/uninstall)
42 * @result: Result of the operation
43 * @done: Signal completion to the issuer of the task
45 struct cpuhp_cpu_state {
46 enum cpuhp_state state;
47 enum cpuhp_state target;
49 struct task_struct *thread;
52 enum cpuhp_state cb_state;
53 int (*cb)(unsigned int cpu);
55 struct completion done;
59 static DEFINE_PER_CPU(struct cpuhp_cpu_state, cpuhp_state);
62 * cpuhp_step - Hotplug state machine step
63 * @name: Name of the step
64 * @startup: Startup function of the step
65 * @teardown: Teardown function of the step
66 * @skip_onerr: Do not invoke the functions on error rollback
67 * Will go away once the notifiers are gone
68 * @cant_stop: Bringup/teardown can't be stopped at this step
72 int (*startup)(unsigned int cpu);
73 int (*teardown)(unsigned int cpu);
78 static DEFINE_MUTEX(cpuhp_state_mutex);
79 static struct cpuhp_step cpuhp_bp_states[];
80 static struct cpuhp_step cpuhp_ap_states[];
83 * cpuhp_invoke_callback _ Invoke the callbacks for a given state
84 * @cpu: The cpu for which the callback should be invoked
85 * @step: The step in the state machine
86 * @cb: The callback function to invoke
88 * Called from cpu hotplug and from the state register machinery
90 static int cpuhp_invoke_callback(unsigned int cpu, enum cpuhp_state step,
91 int (*cb)(unsigned int))
93 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
97 trace_cpuhp_enter(cpu, st->target, step, cb);
99 trace_cpuhp_exit(cpu, st->state, step, ret);
105 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
106 static DEFINE_MUTEX(cpu_add_remove_lock);
107 bool cpuhp_tasks_frozen;
108 EXPORT_SYMBOL_GPL(cpuhp_tasks_frozen);
111 * The following two APIs (cpu_maps_update_begin/done) must be used when
112 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
113 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
114 * hotplug callback (un)registration performed using __register_cpu_notifier()
115 * or __unregister_cpu_notifier().
117 void cpu_maps_update_begin(void)
119 mutex_lock(&cpu_add_remove_lock);
121 EXPORT_SYMBOL(cpu_notifier_register_begin);
123 void cpu_maps_update_done(void)
125 mutex_unlock(&cpu_add_remove_lock);
127 EXPORT_SYMBOL(cpu_notifier_register_done);
129 static RAW_NOTIFIER_HEAD(cpu_chain);
131 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
132 * Should always be manipulated under cpu_add_remove_lock
134 static int cpu_hotplug_disabled;
136 #ifdef CONFIG_HOTPLUG_CPU
139 struct task_struct *active_writer;
140 /* wait queue to wake up the active_writer */
141 wait_queue_head_t wq;
142 /* verifies that no writer will get active while readers are active */
145 * Also blocks the new readers during
146 * an ongoing cpu hotplug operation.
150 #ifdef CONFIG_DEBUG_LOCK_ALLOC
151 struct lockdep_map dep_map;
154 .active_writer = NULL,
155 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
156 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
157 #ifdef CONFIG_DEBUG_LOCK_ALLOC
158 .dep_map = {.name = "cpu_hotplug.lock" },
162 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
163 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
164 #define cpuhp_lock_acquire_tryread() \
165 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
166 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
167 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
170 void get_online_cpus(void)
173 if (cpu_hotplug.active_writer == current)
175 cpuhp_lock_acquire_read();
176 mutex_lock(&cpu_hotplug.lock);
177 atomic_inc(&cpu_hotplug.refcount);
178 mutex_unlock(&cpu_hotplug.lock);
180 EXPORT_SYMBOL_GPL(get_online_cpus);
182 void put_online_cpus(void)
186 if (cpu_hotplug.active_writer == current)
189 refcount = atomic_dec_return(&cpu_hotplug.refcount);
190 if (WARN_ON(refcount < 0)) /* try to fix things up */
191 atomic_inc(&cpu_hotplug.refcount);
193 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
194 wake_up(&cpu_hotplug.wq);
196 cpuhp_lock_release();
199 EXPORT_SYMBOL_GPL(put_online_cpus);
202 * This ensures that the hotplug operation can begin only when the
203 * refcount goes to zero.
205 * Note that during a cpu-hotplug operation, the new readers, if any,
206 * will be blocked by the cpu_hotplug.lock
208 * Since cpu_hotplug_begin() is always called after invoking
209 * cpu_maps_update_begin(), we can be sure that only one writer is active.
211 * Note that theoretically, there is a possibility of a livelock:
212 * - Refcount goes to zero, last reader wakes up the sleeping
214 * - Last reader unlocks the cpu_hotplug.lock.
215 * - A new reader arrives at this moment, bumps up the refcount.
216 * - The writer acquires the cpu_hotplug.lock finds the refcount
217 * non zero and goes to sleep again.
219 * However, this is very difficult to achieve in practice since
220 * get_online_cpus() not an api which is called all that often.
223 void cpu_hotplug_begin(void)
227 cpu_hotplug.active_writer = current;
228 cpuhp_lock_acquire();
231 mutex_lock(&cpu_hotplug.lock);
232 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
233 if (likely(!atomic_read(&cpu_hotplug.refcount)))
235 mutex_unlock(&cpu_hotplug.lock);
238 finish_wait(&cpu_hotplug.wq, &wait);
241 void cpu_hotplug_done(void)
243 cpu_hotplug.active_writer = NULL;
244 mutex_unlock(&cpu_hotplug.lock);
245 cpuhp_lock_release();
249 * Wait for currently running CPU hotplug operations to complete (if any) and
250 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
251 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
252 * hotplug path before performing hotplug operations. So acquiring that lock
253 * guarantees mutual exclusion from any currently running hotplug operations.
255 void cpu_hotplug_disable(void)
257 cpu_maps_update_begin();
258 cpu_hotplug_disabled++;
259 cpu_maps_update_done();
261 EXPORT_SYMBOL_GPL(cpu_hotplug_disable);
263 void cpu_hotplug_enable(void)
265 cpu_maps_update_begin();
266 WARN_ON(--cpu_hotplug_disabled < 0);
267 cpu_maps_update_done();
269 EXPORT_SYMBOL_GPL(cpu_hotplug_enable);
270 #endif /* CONFIG_HOTPLUG_CPU */
272 /* Need to know about CPUs going up/down? */
273 int register_cpu_notifier(struct notifier_block *nb)
276 cpu_maps_update_begin();
277 ret = raw_notifier_chain_register(&cpu_chain, nb);
278 cpu_maps_update_done();
282 int __register_cpu_notifier(struct notifier_block *nb)
284 return raw_notifier_chain_register(&cpu_chain, nb);
287 static int __cpu_notify(unsigned long val, unsigned int cpu, int nr_to_call,
290 unsigned long mod = cpuhp_tasks_frozen ? CPU_TASKS_FROZEN : 0;
291 void *hcpu = (void *)(long)cpu;
295 ret = __raw_notifier_call_chain(&cpu_chain, val | mod, hcpu, nr_to_call,
298 return notifier_to_errno(ret);
301 static int cpu_notify(unsigned long val, unsigned int cpu)
303 return __cpu_notify(val, cpu, -1, NULL);
306 static void cpu_notify_nofail(unsigned long val, unsigned int cpu)
308 BUG_ON(cpu_notify(val, cpu));
311 /* Notifier wrappers for transitioning to state machine */
312 static int notify_prepare(unsigned int cpu)
317 ret = __cpu_notify(CPU_UP_PREPARE, cpu, -1, &nr_calls);
320 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
322 __cpu_notify(CPU_UP_CANCELED, cpu, nr_calls, NULL);
327 static int notify_online(unsigned int cpu)
329 cpu_notify(CPU_ONLINE, cpu);
333 static int notify_starting(unsigned int cpu)
335 cpu_notify(CPU_STARTING, cpu);
339 static int bringup_wait_for_ap(unsigned int cpu)
341 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
343 wait_for_completion(&st->done);
347 static int bringup_cpu(unsigned int cpu)
349 struct task_struct *idle = idle_thread_get(cpu);
352 /* Arch-specific enabling code. */
353 ret = __cpu_up(cpu, idle);
355 cpu_notify(CPU_UP_CANCELED, cpu);
358 ret = bringup_wait_for_ap(cpu);
359 BUG_ON(!cpu_online(cpu));
364 * Hotplug state machine related functions
366 static void undo_cpu_down(unsigned int cpu, struct cpuhp_cpu_state *st,
367 struct cpuhp_step *steps)
369 for (st->state++; st->state < st->target; st->state++) {
370 struct cpuhp_step *step = steps + st->state;
372 if (!step->skip_onerr)
373 cpuhp_invoke_callback(cpu, st->state, step->startup);
377 static int cpuhp_down_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
378 struct cpuhp_step *steps, enum cpuhp_state target)
380 enum cpuhp_state prev_state = st->state;
383 for (; st->state > target; st->state--) {
384 struct cpuhp_step *step = steps + st->state;
386 ret = cpuhp_invoke_callback(cpu, st->state, step->teardown);
388 st->target = prev_state;
389 undo_cpu_down(cpu, st, steps);
396 static void undo_cpu_up(unsigned int cpu, struct cpuhp_cpu_state *st,
397 struct cpuhp_step *steps)
399 for (st->state--; st->state > st->target; st->state--) {
400 struct cpuhp_step *step = steps + st->state;
402 if (!step->skip_onerr)
403 cpuhp_invoke_callback(cpu, st->state, step->teardown);
407 static int cpuhp_up_callbacks(unsigned int cpu, struct cpuhp_cpu_state *st,
408 struct cpuhp_step *steps, enum cpuhp_state target)
410 enum cpuhp_state prev_state = st->state;
413 while (st->state < target) {
414 struct cpuhp_step *step;
417 step = steps + st->state;
418 ret = cpuhp_invoke_callback(cpu, st->state, step->startup);
420 st->target = prev_state;
421 undo_cpu_up(cpu, st, steps);
429 * The cpu hotplug threads manage the bringup and teardown of the cpus
431 static void cpuhp_create(unsigned int cpu)
433 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
435 init_completion(&st->done);
438 static int cpuhp_should_run(unsigned int cpu)
440 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
442 return st->should_run;
445 /* Execute the teardown callbacks. Used to be CPU_DOWN_PREPARE */
446 static int cpuhp_ap_offline(unsigned int cpu, struct cpuhp_cpu_state *st)
448 enum cpuhp_state target = max((int)st->target, CPUHP_TEARDOWN_CPU);
450 return cpuhp_down_callbacks(cpu, st, cpuhp_ap_states, target);
453 /* Execute the online startup callbacks. Used to be CPU_ONLINE */
454 static int cpuhp_ap_online(unsigned int cpu, struct cpuhp_cpu_state *st)
456 return cpuhp_up_callbacks(cpu, st, cpuhp_ap_states, st->target);
460 * Execute teardown/startup callbacks on the plugged cpu. Also used to invoke
461 * callbacks when a state gets [un]installed at runtime.
463 static void cpuhp_thread_fun(unsigned int cpu)
465 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
469 * Paired with the mb() in cpuhp_kick_ap_work and
470 * cpuhp_invoke_ap_callback, so the work set is consistent visible.
476 st->should_run = false;
478 /* Single callback invocation for [un]install ? */
480 if (st->cb_state < CPUHP_AP_ONLINE) {
482 ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
485 ret = cpuhp_invoke_callback(cpu, st->cb_state, st->cb);
487 } else if (st->rollback) {
488 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
490 undo_cpu_down(cpu, st, cpuhp_ap_states);
492 * This is a momentary workaround to keep the notifier users
493 * happy. Will go away once we got rid of the notifiers.
495 cpu_notify_nofail(CPU_DOWN_FAILED, cpu);
496 st->rollback = false;
498 /* Cannot happen .... */
499 BUG_ON(st->state < CPUHP_AP_ONLINE_IDLE);
501 /* Regular hotplug work */
502 if (st->state < st->target)
503 ret = cpuhp_ap_online(cpu, st);
504 else if (st->state > st->target)
505 ret = cpuhp_ap_offline(cpu, st);
511 /* Invoke a single callback on a remote cpu */
512 static int cpuhp_invoke_ap_callback(int cpu, enum cpuhp_state state,
513 int (*cb)(unsigned int))
515 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
517 if (!cpu_online(cpu))
520 st->cb_state = state;
523 * Make sure the above stores are visible before should_run becomes
524 * true. Paired with the mb() above in cpuhp_thread_fun()
527 st->should_run = true;
528 wake_up_process(st->thread);
529 wait_for_completion(&st->done);
533 /* Regular hotplug invocation of the AP hotplug thread */
534 static void __cpuhp_kick_ap_work(struct cpuhp_cpu_state *st)
539 * Make sure the above stores are visible before should_run becomes
540 * true. Paired with the mb() above in cpuhp_thread_fun()
543 st->should_run = true;
544 wake_up_process(st->thread);
547 static int cpuhp_kick_ap_work(unsigned int cpu)
549 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
550 enum cpuhp_state state = st->state;
552 trace_cpuhp_enter(cpu, st->target, state, cpuhp_kick_ap_work);
553 __cpuhp_kick_ap_work(st);
554 wait_for_completion(&st->done);
555 trace_cpuhp_exit(cpu, st->state, state, st->result);
559 static struct smp_hotplug_thread cpuhp_threads = {
560 .store = &cpuhp_state.thread,
561 .create = &cpuhp_create,
562 .thread_should_run = cpuhp_should_run,
563 .thread_fn = cpuhp_thread_fun,
564 .thread_comm = "cpuhp/%u",
568 void __init cpuhp_threads_init(void)
570 BUG_ON(smpboot_register_percpu_thread(&cpuhp_threads));
571 kthread_unpark(this_cpu_read(cpuhp_state.thread));
574 #ifdef CONFIG_HOTPLUG_CPU
575 EXPORT_SYMBOL(register_cpu_notifier);
576 EXPORT_SYMBOL(__register_cpu_notifier);
577 void unregister_cpu_notifier(struct notifier_block *nb)
579 cpu_maps_update_begin();
580 raw_notifier_chain_unregister(&cpu_chain, nb);
581 cpu_maps_update_done();
583 EXPORT_SYMBOL(unregister_cpu_notifier);
585 void __unregister_cpu_notifier(struct notifier_block *nb)
587 raw_notifier_chain_unregister(&cpu_chain, nb);
589 EXPORT_SYMBOL(__unregister_cpu_notifier);
592 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
595 * This function walks all processes, finds a valid mm struct for each one and
596 * then clears a corresponding bit in mm's cpumask. While this all sounds
597 * trivial, there are various non-obvious corner cases, which this function
598 * tries to solve in a safe manner.
600 * Also note that the function uses a somewhat relaxed locking scheme, so it may
601 * be called only for an already offlined CPU.
603 void clear_tasks_mm_cpumask(int cpu)
605 struct task_struct *p;
608 * This function is called after the cpu is taken down and marked
609 * offline, so its not like new tasks will ever get this cpu set in
610 * their mm mask. -- Peter Zijlstra
611 * Thus, we may use rcu_read_lock() here, instead of grabbing
612 * full-fledged tasklist_lock.
614 WARN_ON(cpu_online(cpu));
616 for_each_process(p) {
617 struct task_struct *t;
620 * Main thread might exit, but other threads may still have
621 * a valid mm. Find one.
623 t = find_lock_task_mm(p);
626 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
632 static inline void check_for_tasks(int dead_cpu)
634 struct task_struct *g, *p;
636 read_lock(&tasklist_lock);
637 for_each_process_thread(g, p) {
641 * We do the check with unlocked task_rq(p)->lock.
642 * Order the reading to do not warn about a task,
643 * which was running on this cpu in the past, and
644 * it's just been woken on another cpu.
647 if (task_cpu(p) != dead_cpu)
650 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
651 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
653 read_unlock(&tasklist_lock);
656 static int notify_down_prepare(unsigned int cpu)
658 int err, nr_calls = 0;
660 err = __cpu_notify(CPU_DOWN_PREPARE, cpu, -1, &nr_calls);
663 __cpu_notify(CPU_DOWN_FAILED, cpu, nr_calls, NULL);
664 pr_warn("%s: attempt to take down CPU %u failed\n",
670 static int notify_dying(unsigned int cpu)
672 cpu_notify(CPU_DYING, cpu);
676 /* Take this CPU down. */
677 static int take_cpu_down(void *_param)
679 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
680 enum cpuhp_state target = max((int)st->target, CPUHP_AP_OFFLINE);
681 int err, cpu = smp_processor_id();
683 /* Ensure this CPU doesn't handle any more interrupts. */
684 err = __cpu_disable();
688 /* Invoke the former CPU_DYING callbacks */
689 for (; st->state > target; st->state--) {
690 struct cpuhp_step *step = cpuhp_ap_states + st->state;
692 cpuhp_invoke_callback(cpu, st->state, step->teardown);
694 /* Give up timekeeping duties */
695 tick_handover_do_timer();
696 /* Park the stopper thread */
697 stop_machine_park(cpu);
701 static int takedown_cpu(unsigned int cpu)
703 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
707 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
708 * and RCU users of this state to go away such that all new such users
711 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
712 * not imply sync_sched(), so wait for both.
714 * Do sync before park smpboot threads to take care the rcu boost case.
716 if (IS_ENABLED(CONFIG_PREEMPT))
717 synchronize_rcu_mult(call_rcu, call_rcu_sched);
721 /* Park the smpboot threads */
722 kthread_park(per_cpu_ptr(&cpuhp_state, cpu)->thread);
723 smpboot_park_threads(cpu);
726 * Prevent irq alloc/free while the dying cpu reorganizes the
727 * interrupt affinities.
732 * So now all preempt/rcu users must observe !cpu_active().
734 err = stop_machine(take_cpu_down, NULL, cpumask_of(cpu));
736 /* CPU refused to die */
738 /* Unpark the hotplug thread so we can rollback there */
739 kthread_unpark(per_cpu_ptr(&cpuhp_state, cpu)->thread);
742 BUG_ON(cpu_online(cpu));
745 * The migration_call() CPU_DYING callback will have removed all
746 * runnable tasks from the cpu, there's only the idle task left now
747 * that the migration thread is done doing the stop_machine thing.
749 * Wait for the stop thread to go away.
751 wait_for_completion(&st->done);
752 BUG_ON(st->state != CPUHP_AP_IDLE_DEAD);
754 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
757 hotplug_cpu__broadcast_tick_pull(cpu);
758 /* This actually kills the CPU. */
761 tick_cleanup_dead_cpu(cpu);
765 static int notify_dead(unsigned int cpu)
767 cpu_notify_nofail(CPU_DEAD, cpu);
768 check_for_tasks(cpu);
772 static void cpuhp_complete_idle_dead(void *arg)
774 struct cpuhp_cpu_state *st = arg;
779 void cpuhp_report_idle_dead(void)
781 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
783 BUG_ON(st->state != CPUHP_AP_OFFLINE);
784 rcu_report_dead(smp_processor_id());
785 st->state = CPUHP_AP_IDLE_DEAD;
787 * We cannot call complete after rcu_report_dead() so we delegate it
790 smp_call_function_single(cpumask_first(cpu_online_mask),
791 cpuhp_complete_idle_dead, st, 0);
795 #define notify_down_prepare NULL
796 #define takedown_cpu NULL
797 #define notify_dead NULL
798 #define notify_dying NULL
801 #ifdef CONFIG_HOTPLUG_CPU
803 /* Requires cpu_add_remove_lock to be held */
804 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen,
805 enum cpuhp_state target)
807 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
808 int prev_state, ret = 0;
809 bool hasdied = false;
811 if (num_online_cpus() == 1)
814 if (!cpu_present(cpu))
819 cpuhp_tasks_frozen = tasks_frozen;
821 prev_state = st->state;
824 * If the current CPU state is in the range of the AP hotplug thread,
825 * then we need to kick the thread.
827 if (st->state > CPUHP_TEARDOWN_CPU) {
828 ret = cpuhp_kick_ap_work(cpu);
830 * The AP side has done the error rollback already. Just
831 * return the error code..
837 * We might have stopped still in the range of the AP hotplug
838 * thread. Nothing to do anymore.
840 if (st->state > CPUHP_TEARDOWN_CPU)
844 * The AP brought itself down to CPUHP_TEARDOWN_CPU. So we need
845 * to do the further cleanups.
847 ret = cpuhp_down_callbacks(cpu, st, cpuhp_bp_states, target);
848 if (ret && st->state > CPUHP_TEARDOWN_CPU && st->state < prev_state) {
849 st->target = prev_state;
851 cpuhp_kick_ap_work(cpu);
854 hasdied = prev_state != st->state && st->state == CPUHP_OFFLINE;
857 /* This post dead nonsense must die */
859 cpu_notify_nofail(CPU_POST_DEAD, cpu);
863 static int do_cpu_down(unsigned int cpu, enum cpuhp_state target)
867 cpu_maps_update_begin();
869 if (cpu_hotplug_disabled) {
874 err = _cpu_down(cpu, 0, target);
877 cpu_maps_update_done();
880 int cpu_down(unsigned int cpu)
882 return do_cpu_down(cpu, CPUHP_OFFLINE);
884 EXPORT_SYMBOL(cpu_down);
885 #endif /*CONFIG_HOTPLUG_CPU*/
888 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
889 * @cpu: cpu that just started
891 * This function calls the cpu_chain notifiers with CPU_STARTING.
892 * It must be called by the arch code on the new cpu, before the new cpu
893 * enables interrupts and before the "boot" cpu returns from __cpu_up().
895 void notify_cpu_starting(unsigned int cpu)
897 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
898 enum cpuhp_state target = min((int)st->target, CPUHP_AP_ONLINE);
900 while (st->state < target) {
901 struct cpuhp_step *step;
904 step = cpuhp_ap_states + st->state;
905 cpuhp_invoke_callback(cpu, st->state, step->startup);
910 * Called from the idle task. We need to set active here, so we can kick off
911 * the stopper thread and unpark the smpboot threads. If the target state is
912 * beyond CPUHP_AP_ONLINE_IDLE we kick cpuhp thread and let it bring up the
915 void cpuhp_online_idle(enum cpuhp_state state)
917 struct cpuhp_cpu_state *st = this_cpu_ptr(&cpuhp_state);
918 unsigned int cpu = smp_processor_id();
920 /* Happens for the boot cpu */
921 if (state != CPUHP_AP_ONLINE_IDLE)
924 st->state = CPUHP_AP_ONLINE_IDLE;
926 /* The cpu is marked online, set it active now */
927 set_cpu_active(cpu, true);
928 /* Unpark the stopper thread and the hotplug thread of this cpu */
929 stop_machine_unpark(cpu);
930 kthread_unpark(st->thread);
932 /* Should we go further up ? */
933 if (st->target > CPUHP_AP_ONLINE_IDLE)
934 __cpuhp_kick_ap_work(st);
939 /* Requires cpu_add_remove_lock to be held */
940 static int _cpu_up(unsigned int cpu, int tasks_frozen, enum cpuhp_state target)
942 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
943 struct task_struct *idle;
948 if (!cpu_present(cpu)) {
954 * The caller of do_cpu_up might have raced with another
955 * caller. Ignore it for now.
957 if (st->state >= target)
960 if (st->state == CPUHP_OFFLINE) {
961 /* Let it fail before we try to bring the cpu up */
962 idle = idle_thread_get(cpu);
969 cpuhp_tasks_frozen = tasks_frozen;
973 * If the current CPU state is in the range of the AP hotplug thread,
974 * then we need to kick the thread once more.
976 if (st->state > CPUHP_BRINGUP_CPU) {
977 ret = cpuhp_kick_ap_work(cpu);
979 * The AP side has done the error rollback already. Just
980 * return the error code..
987 * Try to reach the target state. We max out on the BP at
988 * CPUHP_BRINGUP_CPU. After that the AP hotplug thread is
989 * responsible for bringing it up to the target state.
991 target = min((int)target, CPUHP_BRINGUP_CPU);
992 ret = cpuhp_up_callbacks(cpu, st, cpuhp_bp_states, target);
998 static int do_cpu_up(unsigned int cpu, enum cpuhp_state target)
1002 if (!cpu_possible(cpu)) {
1003 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
1005 #if defined(CONFIG_IA64)
1006 pr_err("please check additional_cpus= boot parameter\n");
1011 err = try_online_node(cpu_to_node(cpu));
1015 cpu_maps_update_begin();
1017 if (cpu_hotplug_disabled) {
1022 err = _cpu_up(cpu, 0, target);
1024 cpu_maps_update_done();
1028 int cpu_up(unsigned int cpu)
1030 return do_cpu_up(cpu, CPUHP_ONLINE);
1032 EXPORT_SYMBOL_GPL(cpu_up);
1034 #ifdef CONFIG_PM_SLEEP_SMP
1035 static cpumask_var_t frozen_cpus;
1037 int disable_nonboot_cpus(void)
1039 int cpu, first_cpu, error = 0;
1041 cpu_maps_update_begin();
1042 first_cpu = cpumask_first(cpu_online_mask);
1044 * We take down all of the non-boot CPUs in one shot to avoid races
1045 * with the userspace trying to use the CPU hotplug at the same time
1047 cpumask_clear(frozen_cpus);
1049 pr_info("Disabling non-boot CPUs ...\n");
1050 for_each_online_cpu(cpu) {
1051 if (cpu == first_cpu)
1053 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
1054 error = _cpu_down(cpu, 1, CPUHP_OFFLINE);
1055 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
1057 cpumask_set_cpu(cpu, frozen_cpus);
1059 pr_err("Error taking CPU%d down: %d\n", cpu, error);
1065 BUG_ON(num_online_cpus() > 1);
1067 pr_err("Non-boot CPUs are not disabled\n");
1070 * Make sure the CPUs won't be enabled by someone else. We need to do
1071 * this even in case of failure as all disable_nonboot_cpus() users are
1072 * supposed to do enable_nonboot_cpus() on the failure path.
1074 cpu_hotplug_disabled++;
1076 cpu_maps_update_done();
1080 void __weak arch_enable_nonboot_cpus_begin(void)
1084 void __weak arch_enable_nonboot_cpus_end(void)
1088 void enable_nonboot_cpus(void)
1092 /* Allow everyone to use the CPU hotplug again */
1093 cpu_maps_update_begin();
1094 WARN_ON(--cpu_hotplug_disabled < 0);
1095 if (cpumask_empty(frozen_cpus))
1098 pr_info("Enabling non-boot CPUs ...\n");
1100 arch_enable_nonboot_cpus_begin();
1102 for_each_cpu(cpu, frozen_cpus) {
1103 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
1104 error = _cpu_up(cpu, 1, CPUHP_ONLINE);
1105 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
1107 pr_info("CPU%d is up\n", cpu);
1110 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
1113 arch_enable_nonboot_cpus_end();
1115 cpumask_clear(frozen_cpus);
1117 cpu_maps_update_done();
1120 static int __init alloc_frozen_cpus(void)
1122 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
1126 core_initcall(alloc_frozen_cpus);
1129 * When callbacks for CPU hotplug notifications are being executed, we must
1130 * ensure that the state of the system with respect to the tasks being frozen
1131 * or not, as reported by the notification, remains unchanged *throughout the
1132 * duration* of the execution of the callbacks.
1133 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
1135 * This synchronization is implemented by mutually excluding regular CPU
1136 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
1137 * Hibernate notifications.
1140 cpu_hotplug_pm_callback(struct notifier_block *nb,
1141 unsigned long action, void *ptr)
1145 case PM_SUSPEND_PREPARE:
1146 case PM_HIBERNATION_PREPARE:
1147 cpu_hotplug_disable();
1150 case PM_POST_SUSPEND:
1151 case PM_POST_HIBERNATION:
1152 cpu_hotplug_enable();
1163 static int __init cpu_hotplug_pm_sync_init(void)
1166 * cpu_hotplug_pm_callback has higher priority than x86
1167 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
1168 * to disable cpu hotplug to avoid cpu hotplug race.
1170 pm_notifier(cpu_hotplug_pm_callback, 0);
1173 core_initcall(cpu_hotplug_pm_sync_init);
1175 #endif /* CONFIG_PM_SLEEP_SMP */
1177 #endif /* CONFIG_SMP */
1179 /* Boot processor state steps */
1180 static struct cpuhp_step cpuhp_bp_states[] = {
1187 [CPUHP_CREATE_THREADS]= {
1188 .name = "threads:create",
1189 .startup = smpboot_create_threads,
1194 * Preparatory and dead notifiers. Will be replaced once the notifiers
1195 * are converted to states.
1197 [CPUHP_NOTIFY_PREPARE] = {
1198 .name = "notify:prepare",
1199 .startup = notify_prepare,
1200 .teardown = notify_dead,
1204 /* Kicks the plugged cpu into life */
1205 [CPUHP_BRINGUP_CPU] = {
1206 .name = "cpu:bringup",
1207 .startup = bringup_cpu,
1212 * Handled on controll processor until the plugged processor manages
1215 [CPUHP_TEARDOWN_CPU] = {
1216 .name = "cpu:teardown",
1218 .teardown = takedown_cpu,
1224 /* Application processor state steps */
1225 static struct cpuhp_step cpuhp_ap_states[] = {
1227 /* Final state before CPU kills itself */
1228 [CPUHP_AP_IDLE_DEAD] = {
1229 .name = "idle:dead",
1232 * Last state before CPU enters the idle loop to die. Transient state
1233 * for synchronization.
1235 [CPUHP_AP_OFFLINE] = {
1236 .name = "ap:offline",
1239 /* First state is scheduler control. Interrupts are disabled */
1240 [CPUHP_AP_SCHED_STARTING] = {
1241 .name = "sched:starting",
1242 .startup = sched_cpu_starting,
1246 * Low level startup/teardown notifiers. Run with interrupts
1247 * disabled. Will be removed once the notifiers are converted to
1250 [CPUHP_AP_NOTIFY_STARTING] = {
1251 .name = "notify:starting",
1252 .startup = notify_starting,
1253 .teardown = notify_dying,
1257 /* Entry state on starting. Interrupts enabled from here on. Transient
1258 * state for synchronsization */
1259 [CPUHP_AP_ONLINE] = {
1260 .name = "ap:online",
1262 /* Handle smpboot threads park/unpark */
1263 [CPUHP_AP_SMPBOOT_THREADS] = {
1264 .name = "smpboot:threads",
1265 .startup = smpboot_unpark_threads,
1269 * Online/down_prepare notifiers. Will be removed once the notifiers
1270 * are converted to states.
1272 [CPUHP_AP_NOTIFY_ONLINE] = {
1273 .name = "notify:online",
1274 .startup = notify_online,
1275 .teardown = notify_down_prepare,
1280 * The dynamically registered state space is here
1283 /* CPU is fully up and running. */
1291 /* Sanity check for callbacks */
1292 static int cpuhp_cb_check(enum cpuhp_state state)
1294 if (state <= CPUHP_OFFLINE || state >= CPUHP_ONLINE)
1299 static bool cpuhp_is_ap_state(enum cpuhp_state state)
1302 * The extra check for CPUHP_TEARDOWN_CPU is only for documentation
1303 * purposes as that state is handled explicitely in cpu_down.
1305 return state > CPUHP_BRINGUP_CPU && state != CPUHP_TEARDOWN_CPU;
1308 static struct cpuhp_step *cpuhp_get_step(enum cpuhp_state state)
1310 struct cpuhp_step *sp;
1312 sp = cpuhp_is_ap_state(state) ? cpuhp_ap_states : cpuhp_bp_states;
1316 static void cpuhp_store_callbacks(enum cpuhp_state state,
1318 int (*startup)(unsigned int cpu),
1319 int (*teardown)(unsigned int cpu))
1321 /* (Un)Install the callbacks for further cpu hotplug operations */
1322 struct cpuhp_step *sp;
1324 mutex_lock(&cpuhp_state_mutex);
1325 sp = cpuhp_get_step(state);
1326 sp->startup = startup;
1327 sp->teardown = teardown;
1329 mutex_unlock(&cpuhp_state_mutex);
1332 static void *cpuhp_get_teardown_cb(enum cpuhp_state state)
1334 return cpuhp_get_step(state)->teardown;
1338 * Call the startup/teardown function for a step either on the AP or
1339 * on the current CPU.
1341 static int cpuhp_issue_call(int cpu, enum cpuhp_state state,
1342 int (*cb)(unsigned int), bool bringup)
1349 * The non AP bound callbacks can fail on bringup. On teardown
1350 * e.g. module removal we crash for now.
1353 if (cpuhp_is_ap_state(state))
1354 ret = cpuhp_invoke_ap_callback(cpu, state, cb);
1356 ret = cpuhp_invoke_callback(cpu, state, cb);
1358 ret = cpuhp_invoke_callback(cpu, state, cb);
1360 BUG_ON(ret && !bringup);
1365 * Called from __cpuhp_setup_state on a recoverable failure.
1367 * Note: The teardown callbacks for rollback are not allowed to fail!
1369 static void cpuhp_rollback_install(int failedcpu, enum cpuhp_state state,
1370 int (*teardown)(unsigned int cpu))
1377 /* Roll back the already executed steps on the other cpus */
1378 for_each_present_cpu(cpu) {
1379 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1380 int cpustate = st->state;
1382 if (cpu >= failedcpu)
1385 /* Did we invoke the startup call on that cpu ? */
1386 if (cpustate >= state)
1387 cpuhp_issue_call(cpu, state, teardown, false);
1392 * Returns a free for dynamic slot assignment of the Online state. The states
1393 * are protected by the cpuhp_slot_states mutex and an empty slot is identified
1394 * by having no name assigned.
1396 static int cpuhp_reserve_state(enum cpuhp_state state)
1400 mutex_lock(&cpuhp_state_mutex);
1401 for (i = CPUHP_AP_ONLINE_DYN; i <= CPUHP_AP_ONLINE_DYN_END; i++) {
1402 if (cpuhp_ap_states[i].name)
1405 cpuhp_ap_states[i].name = "Reserved";
1406 mutex_unlock(&cpuhp_state_mutex);
1409 mutex_unlock(&cpuhp_state_mutex);
1410 WARN(1, "No more dynamic states available for CPU hotplug\n");
1415 * __cpuhp_setup_state - Setup the callbacks for an hotplug machine state
1416 * @state: The state to setup
1417 * @invoke: If true, the startup function is invoked for cpus where
1418 * cpu state >= @state
1419 * @startup: startup callback function
1420 * @teardown: teardown callback function
1422 * Returns 0 if successful, otherwise a proper error code
1424 int __cpuhp_setup_state(enum cpuhp_state state,
1425 const char *name, bool invoke,
1426 int (*startup)(unsigned int cpu),
1427 int (*teardown)(unsigned int cpu))
1432 if (cpuhp_cb_check(state) || !name)
1437 /* currently assignments for the ONLINE state are possible */
1438 if (state == CPUHP_AP_ONLINE_DYN) {
1440 ret = cpuhp_reserve_state(state);
1446 cpuhp_store_callbacks(state, name, startup, teardown);
1448 if (!invoke || !startup)
1452 * Try to call the startup callback for each present cpu
1453 * depending on the hotplug state of the cpu.
1455 for_each_present_cpu(cpu) {
1456 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1457 int cpustate = st->state;
1459 if (cpustate < state)
1462 ret = cpuhp_issue_call(cpu, state, startup, true);
1464 cpuhp_rollback_install(cpu, state, teardown);
1465 cpuhp_store_callbacks(state, NULL, NULL, NULL);
1471 if (!ret && dyn_state)
1475 EXPORT_SYMBOL(__cpuhp_setup_state);
1478 * __cpuhp_remove_state - Remove the callbacks for an hotplug machine state
1479 * @state: The state to remove
1480 * @invoke: If true, the teardown function is invoked for cpus where
1481 * cpu state >= @state
1483 * The teardown callback is currently not allowed to fail. Think
1484 * about module removal!
1486 void __cpuhp_remove_state(enum cpuhp_state state, bool invoke)
1488 int (*teardown)(unsigned int cpu) = cpuhp_get_teardown_cb(state);
1491 BUG_ON(cpuhp_cb_check(state));
1495 if (!invoke || !teardown)
1499 * Call the teardown callback for each present cpu depending
1500 * on the hotplug state of the cpu. This function is not
1501 * allowed to fail currently!
1503 for_each_present_cpu(cpu) {
1504 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, cpu);
1505 int cpustate = st->state;
1507 if (cpustate >= state)
1508 cpuhp_issue_call(cpu, state, teardown, false);
1511 cpuhp_store_callbacks(state, NULL, NULL, NULL);
1514 EXPORT_SYMBOL(__cpuhp_remove_state);
1516 #if defined(CONFIG_SYSFS) && defined(CONFIG_HOTPLUG_CPU)
1517 static ssize_t show_cpuhp_state(struct device *dev,
1518 struct device_attribute *attr, char *buf)
1520 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1522 return sprintf(buf, "%d\n", st->state);
1524 static DEVICE_ATTR(state, 0444, show_cpuhp_state, NULL);
1526 static ssize_t write_cpuhp_target(struct device *dev,
1527 struct device_attribute *attr,
1528 const char *buf, size_t count)
1530 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1531 struct cpuhp_step *sp;
1534 ret = kstrtoint(buf, 10, &target);
1538 #ifdef CONFIG_CPU_HOTPLUG_STATE_CONTROL
1539 if (target < CPUHP_OFFLINE || target > CPUHP_ONLINE)
1542 if (target != CPUHP_OFFLINE && target != CPUHP_ONLINE)
1546 ret = lock_device_hotplug_sysfs();
1550 mutex_lock(&cpuhp_state_mutex);
1551 sp = cpuhp_get_step(target);
1552 ret = !sp->name || sp->cant_stop ? -EINVAL : 0;
1553 mutex_unlock(&cpuhp_state_mutex);
1557 if (st->state < target)
1558 ret = do_cpu_up(dev->id, target);
1560 ret = do_cpu_down(dev->id, target);
1562 unlock_device_hotplug();
1563 return ret ? ret : count;
1566 static ssize_t show_cpuhp_target(struct device *dev,
1567 struct device_attribute *attr, char *buf)
1569 struct cpuhp_cpu_state *st = per_cpu_ptr(&cpuhp_state, dev->id);
1571 return sprintf(buf, "%d\n", st->target);
1573 static DEVICE_ATTR(target, 0644, show_cpuhp_target, write_cpuhp_target);
1575 static struct attribute *cpuhp_cpu_attrs[] = {
1576 &dev_attr_state.attr,
1577 &dev_attr_target.attr,
1581 static struct attribute_group cpuhp_cpu_attr_group = {
1582 .attrs = cpuhp_cpu_attrs,
1587 static ssize_t show_cpuhp_states(struct device *dev,
1588 struct device_attribute *attr, char *buf)
1590 ssize_t cur, res = 0;
1593 mutex_lock(&cpuhp_state_mutex);
1594 for (i = CPUHP_OFFLINE; i <= CPUHP_ONLINE; i++) {
1595 struct cpuhp_step *sp = cpuhp_get_step(i);
1598 cur = sprintf(buf, "%3d: %s\n", i, sp->name);
1603 mutex_unlock(&cpuhp_state_mutex);
1606 static DEVICE_ATTR(states, 0444, show_cpuhp_states, NULL);
1608 static struct attribute *cpuhp_cpu_root_attrs[] = {
1609 &dev_attr_states.attr,
1613 static struct attribute_group cpuhp_cpu_root_attr_group = {
1614 .attrs = cpuhp_cpu_root_attrs,
1619 static int __init cpuhp_sysfs_init(void)
1623 ret = sysfs_create_group(&cpu_subsys.dev_root->kobj,
1624 &cpuhp_cpu_root_attr_group);
1628 for_each_possible_cpu(cpu) {
1629 struct device *dev = get_cpu_device(cpu);
1633 ret = sysfs_create_group(&dev->kobj, &cpuhp_cpu_attr_group);
1639 device_initcall(cpuhp_sysfs_init);
1643 * cpu_bit_bitmap[] is a special, "compressed" data structure that
1644 * represents all NR_CPUS bits binary values of 1<<nr.
1646 * It is used by cpumask_of() to get a constant address to a CPU
1647 * mask value that has a single bit set only.
1650 /* cpu_bit_bitmap[0] is empty - so we can back into it */
1651 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
1652 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
1653 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
1654 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
1656 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
1658 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
1659 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
1660 #if BITS_PER_LONG > 32
1661 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
1662 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
1665 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
1667 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
1668 EXPORT_SYMBOL(cpu_all_bits);
1670 #ifdef CONFIG_INIT_ALL_POSSIBLE
1671 struct cpumask __cpu_possible_mask __read_mostly
1674 struct cpumask __cpu_possible_mask __read_mostly;
1676 EXPORT_SYMBOL(__cpu_possible_mask);
1678 struct cpumask __cpu_online_mask __read_mostly;
1679 EXPORT_SYMBOL(__cpu_online_mask);
1681 struct cpumask __cpu_present_mask __read_mostly;
1682 EXPORT_SYMBOL(__cpu_present_mask);
1684 struct cpumask __cpu_active_mask __read_mostly;
1685 EXPORT_SYMBOL(__cpu_active_mask);
1687 void init_cpu_present(const struct cpumask *src)
1689 cpumask_copy(&__cpu_present_mask, src);
1692 void init_cpu_possible(const struct cpumask *src)
1694 cpumask_copy(&__cpu_possible_mask, src);
1697 void init_cpu_online(const struct cpumask *src)
1699 cpumask_copy(&__cpu_online_mask, src);
1703 * Activate the first processor.
1705 void __init boot_cpu_init(void)
1707 int cpu = smp_processor_id();
1709 /* Mark the boot cpu "present", "online" etc for SMP and UP case */
1710 set_cpu_online(cpu, true);
1711 set_cpu_active(cpu, true);
1712 set_cpu_present(cpu, true);
1713 set_cpu_possible(cpu, true);
1717 * Must be called _AFTER_ setting up the per_cpu areas
1719 void __init boot_cpu_state_init(void)
1721 per_cpu_ptr(&cpuhp_state, smp_processor_id())->state = CPUHP_ONLINE;