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 <trace/events/power.h>
29 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
30 static DEFINE_MUTEX(cpu_add_remove_lock);
33 * The following two APIs (cpu_maps_update_begin/done) must be used when
34 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
35 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
36 * hotplug callback (un)registration performed using __register_cpu_notifier()
37 * or __unregister_cpu_notifier().
39 void cpu_maps_update_begin(void)
41 mutex_lock(&cpu_add_remove_lock);
43 EXPORT_SYMBOL(cpu_notifier_register_begin);
45 void cpu_maps_update_done(void)
47 mutex_unlock(&cpu_add_remove_lock);
49 EXPORT_SYMBOL(cpu_notifier_register_done);
51 static RAW_NOTIFIER_HEAD(cpu_chain);
53 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
54 * Should always be manipulated under cpu_add_remove_lock
56 static int cpu_hotplug_disabled;
58 #ifdef CONFIG_HOTPLUG_CPU
61 struct task_struct *active_writer;
62 /* wait queue to wake up the active_writer */
64 /* verifies that no writer will get active while readers are active */
67 * Also blocks the new readers during
68 * an ongoing cpu hotplug operation.
72 #ifdef CONFIG_DEBUG_LOCK_ALLOC
73 struct lockdep_map dep_map;
76 .active_writer = NULL,
77 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
78 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
79 #ifdef CONFIG_DEBUG_LOCK_ALLOC
80 .dep_map = {.name = "cpu_hotplug.lock" },
84 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
85 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
86 #define cpuhp_lock_acquire_tryread() \
87 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
88 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
89 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
92 void get_online_cpus(void)
95 if (cpu_hotplug.active_writer == current)
97 cpuhp_lock_acquire_read();
98 mutex_lock(&cpu_hotplug.lock);
99 atomic_inc(&cpu_hotplug.refcount);
100 mutex_unlock(&cpu_hotplug.lock);
102 EXPORT_SYMBOL_GPL(get_online_cpus);
104 bool try_get_online_cpus(void)
106 if (cpu_hotplug.active_writer == current)
108 if (!mutex_trylock(&cpu_hotplug.lock))
110 cpuhp_lock_acquire_tryread();
111 atomic_inc(&cpu_hotplug.refcount);
112 mutex_unlock(&cpu_hotplug.lock);
115 EXPORT_SYMBOL_GPL(try_get_online_cpus);
117 void put_online_cpus(void)
121 if (cpu_hotplug.active_writer == current)
124 refcount = atomic_dec_return(&cpu_hotplug.refcount);
125 if (WARN_ON(refcount < 0)) /* try to fix things up */
126 atomic_inc(&cpu_hotplug.refcount);
128 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
129 wake_up(&cpu_hotplug.wq);
131 cpuhp_lock_release();
134 EXPORT_SYMBOL_GPL(put_online_cpus);
137 * This ensures that the hotplug operation can begin only when the
138 * refcount goes to zero.
140 * Note that during a cpu-hotplug operation, the new readers, if any,
141 * will be blocked by the cpu_hotplug.lock
143 * Since cpu_hotplug_begin() is always called after invoking
144 * cpu_maps_update_begin(), we can be sure that only one writer is active.
146 * Note that theoretically, there is a possibility of a livelock:
147 * - Refcount goes to zero, last reader wakes up the sleeping
149 * - Last reader unlocks the cpu_hotplug.lock.
150 * - A new reader arrives at this moment, bumps up the refcount.
151 * - The writer acquires the cpu_hotplug.lock finds the refcount
152 * non zero and goes to sleep again.
154 * However, this is very difficult to achieve in practice since
155 * get_online_cpus() not an api which is called all that often.
158 void cpu_hotplug_begin(void)
162 cpu_hotplug.active_writer = current;
163 cpuhp_lock_acquire();
166 mutex_lock(&cpu_hotplug.lock);
167 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
168 if (likely(!atomic_read(&cpu_hotplug.refcount)))
170 mutex_unlock(&cpu_hotplug.lock);
173 finish_wait(&cpu_hotplug.wq, &wait);
176 void cpu_hotplug_done(void)
178 cpu_hotplug.active_writer = NULL;
179 mutex_unlock(&cpu_hotplug.lock);
180 cpuhp_lock_release();
184 * Wait for currently running CPU hotplug operations to complete (if any) and
185 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
186 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
187 * hotplug path before performing hotplug operations. So acquiring that lock
188 * guarantees mutual exclusion from any currently running hotplug operations.
190 void cpu_hotplug_disable(void)
192 cpu_maps_update_begin();
193 cpu_hotplug_disabled = 1;
194 cpu_maps_update_done();
197 void cpu_hotplug_enable(void)
199 cpu_maps_update_begin();
200 cpu_hotplug_disabled = 0;
201 cpu_maps_update_done();
204 #endif /* CONFIG_HOTPLUG_CPU */
206 /* Need to know about CPUs going up/down? */
207 int __ref register_cpu_notifier(struct notifier_block *nb)
210 cpu_maps_update_begin();
211 ret = raw_notifier_chain_register(&cpu_chain, nb);
212 cpu_maps_update_done();
216 int __ref __register_cpu_notifier(struct notifier_block *nb)
218 return raw_notifier_chain_register(&cpu_chain, nb);
221 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
226 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
229 return notifier_to_errno(ret);
232 static int cpu_notify(unsigned long val, void *v)
234 return __cpu_notify(val, v, -1, NULL);
237 #ifdef CONFIG_HOTPLUG_CPU
239 static void cpu_notify_nofail(unsigned long val, void *v)
241 BUG_ON(cpu_notify(val, v));
243 EXPORT_SYMBOL(register_cpu_notifier);
244 EXPORT_SYMBOL(__register_cpu_notifier);
246 void __ref unregister_cpu_notifier(struct notifier_block *nb)
248 cpu_maps_update_begin();
249 raw_notifier_chain_unregister(&cpu_chain, nb);
250 cpu_maps_update_done();
252 EXPORT_SYMBOL(unregister_cpu_notifier);
254 void __ref __unregister_cpu_notifier(struct notifier_block *nb)
256 raw_notifier_chain_unregister(&cpu_chain, nb);
258 EXPORT_SYMBOL(__unregister_cpu_notifier);
261 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
264 * This function walks all processes, finds a valid mm struct for each one and
265 * then clears a corresponding bit in mm's cpumask. While this all sounds
266 * trivial, there are various non-obvious corner cases, which this function
267 * tries to solve in a safe manner.
269 * Also note that the function uses a somewhat relaxed locking scheme, so it may
270 * be called only for an already offlined CPU.
272 void clear_tasks_mm_cpumask(int cpu)
274 struct task_struct *p;
277 * This function is called after the cpu is taken down and marked
278 * offline, so its not like new tasks will ever get this cpu set in
279 * their mm mask. -- Peter Zijlstra
280 * Thus, we may use rcu_read_lock() here, instead of grabbing
281 * full-fledged tasklist_lock.
283 WARN_ON(cpu_online(cpu));
285 for_each_process(p) {
286 struct task_struct *t;
289 * Main thread might exit, but other threads may still have
290 * a valid mm. Find one.
292 t = find_lock_task_mm(p);
295 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
301 static inline void check_for_tasks(int dead_cpu)
303 struct task_struct *g, *p;
305 read_lock_irq(&tasklist_lock);
306 do_each_thread(g, p) {
310 * We do the check with unlocked task_rq(p)->lock.
311 * Order the reading to do not warn about a task,
312 * which was running on this cpu in the past, and
313 * it's just been woken on another cpu.
316 if (task_cpu(p) != dead_cpu)
319 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
320 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
321 } while_each_thread(g, p);
322 read_unlock_irq(&tasklist_lock);
325 struct take_cpu_down_param {
330 /* Take this CPU down. */
331 static int __ref take_cpu_down(void *_param)
333 struct take_cpu_down_param *param = _param;
336 /* Ensure this CPU doesn't handle any more interrupts. */
337 err = __cpu_disable();
341 cpu_notify(CPU_DYING | param->mod, param->hcpu);
342 /* Give up timekeeping duties */
343 tick_handover_do_timer();
344 /* Park the stopper thread */
345 kthread_park(current);
349 /* Requires cpu_add_remove_lock to be held */
350 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
352 int err, nr_calls = 0;
353 void *hcpu = (void *)(long)cpu;
354 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
355 struct take_cpu_down_param tcd_param = {
360 if (num_online_cpus() == 1)
363 if (!cpu_online(cpu))
368 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
371 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
372 pr_warn("%s: attempt to take down CPU %u failed\n",
378 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
379 * and RCU users of this state to go away such that all new such users
382 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
383 * not imply sync_sched(), so explicitly call both.
385 * Do sync before park smpboot threads to take care the rcu boost case.
387 #ifdef CONFIG_PREEMPT
392 smpboot_park_threads(cpu);
395 * So now all preempt/rcu users must observe !cpu_active().
398 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
400 /* CPU didn't die: tell everyone. Can't complain. */
401 smpboot_unpark_threads(cpu);
402 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
405 BUG_ON(cpu_online(cpu));
408 * The migration_call() CPU_DYING callback will have removed all
409 * runnable tasks from the cpu, there's only the idle task left now
410 * that the migration thread is done doing the stop_machine thing.
412 * Wait for the stop thread to go away.
414 while (!idle_cpu(cpu))
417 hotplug_cpu__broadcast_tick_pull(cpu);
418 /* This actually kills the CPU. */
421 /* CPU is completely dead: tell everyone. Too late to complain. */
422 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
424 check_for_tasks(cpu);
429 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
433 int __ref cpu_down(unsigned int cpu)
437 cpu_maps_update_begin();
439 if (cpu_hotplug_disabled) {
444 err = _cpu_down(cpu, 0);
447 cpu_maps_update_done();
450 EXPORT_SYMBOL(cpu_down);
451 #endif /*CONFIG_HOTPLUG_CPU*/
453 /* Requires cpu_add_remove_lock to be held */
454 static int _cpu_up(unsigned int cpu, int tasks_frozen)
456 int ret, nr_calls = 0;
457 void *hcpu = (void *)(long)cpu;
458 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
459 struct task_struct *idle;
463 if (cpu_online(cpu) || !cpu_present(cpu)) {
468 idle = idle_thread_get(cpu);
474 ret = smpboot_create_threads(cpu);
478 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
481 pr_warn("%s: attempt to bring up CPU %u failed\n",
486 /* Arch-specific enabling code. */
487 ret = __cpu_up(cpu, idle);
490 BUG_ON(!cpu_online(cpu));
492 /* Wake the per cpu threads */
493 smpboot_unpark_threads(cpu);
495 /* Now call notifier in preparation. */
496 cpu_notify(CPU_ONLINE | mod, hcpu);
500 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
507 int cpu_up(unsigned int cpu)
511 if (!cpu_possible(cpu)) {
512 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
514 #if defined(CONFIG_IA64)
515 pr_err("please check additional_cpus= boot parameter\n");
520 err = try_online_node(cpu_to_node(cpu));
524 cpu_maps_update_begin();
526 if (cpu_hotplug_disabled) {
531 err = _cpu_up(cpu, 0);
534 cpu_maps_update_done();
537 EXPORT_SYMBOL_GPL(cpu_up);
539 #ifdef CONFIG_PM_SLEEP_SMP
540 static cpumask_var_t frozen_cpus;
542 int disable_nonboot_cpus(void)
544 int cpu, first_cpu, error = 0;
546 cpu_maps_update_begin();
547 first_cpu = cpumask_first(cpu_online_mask);
549 * We take down all of the non-boot CPUs in one shot to avoid races
550 * with the userspace trying to use the CPU hotplug at the same time
552 cpumask_clear(frozen_cpus);
554 pr_info("Disabling non-boot CPUs ...\n");
555 for_each_online_cpu(cpu) {
556 if (cpu == first_cpu)
558 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
559 error = _cpu_down(cpu, 1);
560 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
562 cpumask_set_cpu(cpu, frozen_cpus);
564 pr_err("Error taking CPU%d down: %d\n", cpu, error);
570 BUG_ON(num_online_cpus() > 1);
571 /* Make sure the CPUs won't be enabled by someone else */
572 cpu_hotplug_disabled = 1;
574 pr_err("Non-boot CPUs are not disabled\n");
576 cpu_maps_update_done();
580 void __weak arch_enable_nonboot_cpus_begin(void)
584 void __weak arch_enable_nonboot_cpus_end(void)
588 void __ref enable_nonboot_cpus(void)
592 /* Allow everyone to use the CPU hotplug again */
593 cpu_maps_update_begin();
594 cpu_hotplug_disabled = 0;
595 if (cpumask_empty(frozen_cpus))
598 pr_info("Enabling non-boot CPUs ...\n");
600 arch_enable_nonboot_cpus_begin();
602 for_each_cpu(cpu, frozen_cpus) {
603 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
604 error = _cpu_up(cpu, 1);
605 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
607 pr_info("CPU%d is up\n", cpu);
610 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
613 arch_enable_nonboot_cpus_end();
615 cpumask_clear(frozen_cpus);
617 cpu_maps_update_done();
620 static int __init alloc_frozen_cpus(void)
622 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
626 core_initcall(alloc_frozen_cpus);
629 * When callbacks for CPU hotplug notifications are being executed, we must
630 * ensure that the state of the system with respect to the tasks being frozen
631 * or not, as reported by the notification, remains unchanged *throughout the
632 * duration* of the execution of the callbacks.
633 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
635 * This synchronization is implemented by mutually excluding regular CPU
636 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
637 * Hibernate notifications.
640 cpu_hotplug_pm_callback(struct notifier_block *nb,
641 unsigned long action, void *ptr)
645 case PM_SUSPEND_PREPARE:
646 case PM_HIBERNATION_PREPARE:
647 cpu_hotplug_disable();
650 case PM_POST_SUSPEND:
651 case PM_POST_HIBERNATION:
652 cpu_hotplug_enable();
663 static int __init cpu_hotplug_pm_sync_init(void)
666 * cpu_hotplug_pm_callback has higher priority than x86
667 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
668 * to disable cpu hotplug to avoid cpu hotplug race.
670 pm_notifier(cpu_hotplug_pm_callback, 0);
673 core_initcall(cpu_hotplug_pm_sync_init);
675 #endif /* CONFIG_PM_SLEEP_SMP */
678 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
679 * @cpu: cpu that just started
681 * This function calls the cpu_chain notifiers with CPU_STARTING.
682 * It must be called by the arch code on the new cpu, before the new cpu
683 * enables interrupts and before the "boot" cpu returns from __cpu_up().
685 void notify_cpu_starting(unsigned int cpu)
687 unsigned long val = CPU_STARTING;
689 #ifdef CONFIG_PM_SLEEP_SMP
690 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
691 val = CPU_STARTING_FROZEN;
692 #endif /* CONFIG_PM_SLEEP_SMP */
693 cpu_notify(val, (void *)(long)cpu);
696 #endif /* CONFIG_SMP */
699 * cpu_bit_bitmap[] is a special, "compressed" data structure that
700 * represents all NR_CPUS bits binary values of 1<<nr.
702 * It is used by cpumask_of() to get a constant address to a CPU
703 * mask value that has a single bit set only.
706 /* cpu_bit_bitmap[0] is empty - so we can back into it */
707 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
708 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
709 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
710 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
712 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
714 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
715 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
716 #if BITS_PER_LONG > 32
717 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
718 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
721 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
723 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
724 EXPORT_SYMBOL(cpu_all_bits);
726 #ifdef CONFIG_INIT_ALL_POSSIBLE
727 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
730 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
732 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
733 EXPORT_SYMBOL(cpu_possible_mask);
735 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
736 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
737 EXPORT_SYMBOL(cpu_online_mask);
739 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
740 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
741 EXPORT_SYMBOL(cpu_present_mask);
743 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
744 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
745 EXPORT_SYMBOL(cpu_active_mask);
747 void set_cpu_possible(unsigned int cpu, bool possible)
750 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
752 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
755 void set_cpu_present(unsigned int cpu, bool present)
758 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
760 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
763 void set_cpu_online(unsigned int cpu, bool online)
766 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
767 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
769 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
773 void set_cpu_active(unsigned int cpu, bool active)
776 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
778 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
781 void init_cpu_present(const struct cpumask *src)
783 cpumask_copy(to_cpumask(cpu_present_bits), src);
786 void init_cpu_possible(const struct cpumask *src)
788 cpumask_copy(to_cpumask(cpu_possible_bits), src);
791 void init_cpu_online(const struct cpumask *src)
793 cpumask_copy(to_cpumask(cpu_online_bits), src);