2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
4 * This code is licenced under the GPL.
6 #include <linux/proc_fs.h>
8 #include <linux/init.h>
9 #include <linux/notifier.h>
10 #include <linux/sched.h>
11 #include <linux/unistd.h>
12 #include <linux/cpu.h>
13 #include <linux/oom.h>
14 #include <linux/rcupdate.h>
15 #include <linux/export.h>
16 #include <linux/bug.h>
17 #include <linux/kthread.h>
18 #include <linux/stop_machine.h>
19 #include <linux/mutex.h>
20 #include <linux/gfp.h>
21 #include <linux/suspend.h>
22 #include <linux/lockdep.h>
23 #include <linux/tick.h>
24 #include <linux/irq.h>
25 #include <trace/events/power.h>
30 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
31 static DEFINE_MUTEX(cpu_add_remove_lock);
34 * The following two APIs (cpu_maps_update_begin/done) must be used when
35 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
36 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
37 * hotplug callback (un)registration performed using __register_cpu_notifier()
38 * or __unregister_cpu_notifier().
40 void cpu_maps_update_begin(void)
42 mutex_lock(&cpu_add_remove_lock);
44 EXPORT_SYMBOL(cpu_notifier_register_begin);
46 void cpu_maps_update_done(void)
48 mutex_unlock(&cpu_add_remove_lock);
50 EXPORT_SYMBOL(cpu_notifier_register_done);
52 static RAW_NOTIFIER_HEAD(cpu_chain);
54 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
55 * Should always be manipulated under cpu_add_remove_lock
57 static int cpu_hotplug_disabled;
59 #ifdef CONFIG_HOTPLUG_CPU
62 struct task_struct *active_writer;
63 /* wait queue to wake up the active_writer */
65 /* verifies that no writer will get active while readers are active */
68 * Also blocks the new readers during
69 * an ongoing cpu hotplug operation.
73 #ifdef CONFIG_DEBUG_LOCK_ALLOC
74 struct lockdep_map dep_map;
77 .active_writer = NULL,
78 .wq = __WAIT_QUEUE_HEAD_INITIALIZER(cpu_hotplug.wq),
79 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
80 #ifdef CONFIG_DEBUG_LOCK_ALLOC
81 .dep_map = {.name = "cpu_hotplug.lock" },
85 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
86 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
87 #define cpuhp_lock_acquire_tryread() \
88 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
89 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
90 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
93 void get_online_cpus(void)
96 if (cpu_hotplug.active_writer == current)
98 cpuhp_lock_acquire_read();
99 mutex_lock(&cpu_hotplug.lock);
100 atomic_inc(&cpu_hotplug.refcount);
101 mutex_unlock(&cpu_hotplug.lock);
103 EXPORT_SYMBOL_GPL(get_online_cpus);
105 bool try_get_online_cpus(void)
107 if (cpu_hotplug.active_writer == current)
109 if (!mutex_trylock(&cpu_hotplug.lock))
111 cpuhp_lock_acquire_tryread();
112 atomic_inc(&cpu_hotplug.refcount);
113 mutex_unlock(&cpu_hotplug.lock);
116 EXPORT_SYMBOL_GPL(try_get_online_cpus);
118 void put_online_cpus(void)
122 if (cpu_hotplug.active_writer == current)
125 refcount = atomic_dec_return(&cpu_hotplug.refcount);
126 if (WARN_ON(refcount < 0)) /* try to fix things up */
127 atomic_inc(&cpu_hotplug.refcount);
129 if (refcount <= 0 && waitqueue_active(&cpu_hotplug.wq))
130 wake_up(&cpu_hotplug.wq);
132 cpuhp_lock_release();
135 EXPORT_SYMBOL_GPL(put_online_cpus);
138 * This ensures that the hotplug operation can begin only when the
139 * refcount goes to zero.
141 * Note that during a cpu-hotplug operation, the new readers, if any,
142 * will be blocked by the cpu_hotplug.lock
144 * Since cpu_hotplug_begin() is always called after invoking
145 * cpu_maps_update_begin(), we can be sure that only one writer is active.
147 * Note that theoretically, there is a possibility of a livelock:
148 * - Refcount goes to zero, last reader wakes up the sleeping
150 * - Last reader unlocks the cpu_hotplug.lock.
151 * - A new reader arrives at this moment, bumps up the refcount.
152 * - The writer acquires the cpu_hotplug.lock finds the refcount
153 * non zero and goes to sleep again.
155 * However, this is very difficult to achieve in practice since
156 * get_online_cpus() not an api which is called all that often.
159 void cpu_hotplug_begin(void)
163 cpu_hotplug.active_writer = current;
164 cpuhp_lock_acquire();
167 mutex_lock(&cpu_hotplug.lock);
168 prepare_to_wait(&cpu_hotplug.wq, &wait, TASK_UNINTERRUPTIBLE);
169 if (likely(!atomic_read(&cpu_hotplug.refcount)))
171 mutex_unlock(&cpu_hotplug.lock);
174 finish_wait(&cpu_hotplug.wq, &wait);
177 void cpu_hotplug_done(void)
179 cpu_hotplug.active_writer = NULL;
180 mutex_unlock(&cpu_hotplug.lock);
181 cpuhp_lock_release();
185 * Wait for currently running CPU hotplug operations to complete (if any) and
186 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
187 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
188 * hotplug path before performing hotplug operations. So acquiring that lock
189 * guarantees mutual exclusion from any currently running hotplug operations.
191 void cpu_hotplug_disable(void)
193 cpu_maps_update_begin();
194 cpu_hotplug_disabled = 1;
195 cpu_maps_update_done();
198 void cpu_hotplug_enable(void)
200 cpu_maps_update_begin();
201 cpu_hotplug_disabled = 0;
202 cpu_maps_update_done();
205 #endif /* CONFIG_HOTPLUG_CPU */
207 /* Need to know about CPUs going up/down? */
208 int __ref register_cpu_notifier(struct notifier_block *nb)
211 cpu_maps_update_begin();
212 ret = raw_notifier_chain_register(&cpu_chain, nb);
213 cpu_maps_update_done();
217 int __ref __register_cpu_notifier(struct notifier_block *nb)
219 return raw_notifier_chain_register(&cpu_chain, nb);
222 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
227 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
230 return notifier_to_errno(ret);
233 static int cpu_notify(unsigned long val, void *v)
235 return __cpu_notify(val, v, -1, NULL);
238 #ifdef CONFIG_HOTPLUG_CPU
240 static void cpu_notify_nofail(unsigned long val, void *v)
242 BUG_ON(cpu_notify(val, v));
244 EXPORT_SYMBOL(register_cpu_notifier);
245 EXPORT_SYMBOL(__register_cpu_notifier);
247 void __ref unregister_cpu_notifier(struct notifier_block *nb)
249 cpu_maps_update_begin();
250 raw_notifier_chain_unregister(&cpu_chain, nb);
251 cpu_maps_update_done();
253 EXPORT_SYMBOL(unregister_cpu_notifier);
255 void __ref __unregister_cpu_notifier(struct notifier_block *nb)
257 raw_notifier_chain_unregister(&cpu_chain, nb);
259 EXPORT_SYMBOL(__unregister_cpu_notifier);
262 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
265 * This function walks all processes, finds a valid mm struct for each one and
266 * then clears a corresponding bit in mm's cpumask. While this all sounds
267 * trivial, there are various non-obvious corner cases, which this function
268 * tries to solve in a safe manner.
270 * Also note that the function uses a somewhat relaxed locking scheme, so it may
271 * be called only for an already offlined CPU.
273 void clear_tasks_mm_cpumask(int cpu)
275 struct task_struct *p;
278 * This function is called after the cpu is taken down and marked
279 * offline, so its not like new tasks will ever get this cpu set in
280 * their mm mask. -- Peter Zijlstra
281 * Thus, we may use rcu_read_lock() here, instead of grabbing
282 * full-fledged tasklist_lock.
284 WARN_ON(cpu_online(cpu));
286 for_each_process(p) {
287 struct task_struct *t;
290 * Main thread might exit, but other threads may still have
291 * a valid mm. Find one.
293 t = find_lock_task_mm(p);
296 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
302 static inline void check_for_tasks(int dead_cpu)
304 struct task_struct *g, *p;
306 read_lock_irq(&tasklist_lock);
307 do_each_thread(g, p) {
311 * We do the check with unlocked task_rq(p)->lock.
312 * Order the reading to do not warn about a task,
313 * which was running on this cpu in the past, and
314 * it's just been woken on another cpu.
317 if (task_cpu(p) != dead_cpu)
320 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
321 p->comm, task_pid_nr(p), dead_cpu, p->state, p->flags);
322 } while_each_thread(g, p);
323 read_unlock_irq(&tasklist_lock);
326 struct take_cpu_down_param {
331 /* Take this CPU down. */
332 static int __ref take_cpu_down(void *_param)
334 struct take_cpu_down_param *param = _param;
337 /* Ensure this CPU doesn't handle any more interrupts. */
338 err = __cpu_disable();
342 cpu_notify(CPU_DYING | param->mod, param->hcpu);
343 /* Give up timekeeping duties */
344 tick_handover_do_timer();
345 /* Park the stopper thread */
346 kthread_park(current);
350 /* Requires cpu_add_remove_lock to be held */
351 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
353 int err, nr_calls = 0;
354 void *hcpu = (void *)(long)cpu;
355 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
356 struct take_cpu_down_param tcd_param = {
361 if (num_online_cpus() == 1)
364 if (!cpu_online(cpu))
369 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
372 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
373 pr_warn("%s: attempt to take down CPU %u failed\n",
379 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
380 * and RCU users of this state to go away such that all new such users
383 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
384 * not imply sync_sched(), so explicitly call both.
386 * Do sync before park smpboot threads to take care the rcu boost case.
388 #ifdef CONFIG_PREEMPT
393 smpboot_park_threads(cpu);
396 * Prevent irq alloc/free while the dying cpu reorganizes the
397 * interrupt affinities.
402 * So now all preempt/rcu users must observe !cpu_active().
404 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
406 /* CPU didn't die: tell everyone. Can't complain. */
407 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
411 BUG_ON(cpu_online(cpu));
414 * The migration_call() CPU_DYING callback will have removed all
415 * runnable tasks from the cpu, there's only the idle task left now
416 * that the migration thread is done doing the stop_machine thing.
418 * Wait for the stop thread to go away.
420 while (!per_cpu(cpu_dead_idle, cpu))
422 smp_mb(); /* Read from cpu_dead_idle before __cpu_die(). */
423 per_cpu(cpu_dead_idle, cpu) = false;
425 /* Interrupts are moved away from the dying cpu, reenable alloc/free */
428 hotplug_cpu__broadcast_tick_pull(cpu);
429 /* This actually kills the CPU. */
432 /* CPU is completely dead: tell everyone. Too late to complain. */
433 tick_cleanup_dead_cpu(cpu);
434 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
436 check_for_tasks(cpu);
441 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
445 int __ref cpu_down(unsigned int cpu)
449 cpu_maps_update_begin();
451 if (cpu_hotplug_disabled) {
456 err = _cpu_down(cpu, 0);
459 cpu_maps_update_done();
462 EXPORT_SYMBOL(cpu_down);
463 #endif /*CONFIG_HOTPLUG_CPU*/
466 * Unpark per-CPU smpboot kthreads at CPU-online time.
468 static int smpboot_thread_call(struct notifier_block *nfb,
469 unsigned long action, void *hcpu)
471 int cpu = (long)hcpu;
473 switch (action & ~CPU_TASKS_FROZEN) {
475 case CPU_DOWN_FAILED:
477 smpboot_unpark_threads(cpu);
487 static struct notifier_block smpboot_thread_notifier = {
488 .notifier_call = smpboot_thread_call,
489 .priority = CPU_PRI_SMPBOOT,
492 void smpboot_thread_init(void)
494 register_cpu_notifier(&smpboot_thread_notifier);
497 /* Requires cpu_add_remove_lock to be held */
498 static int _cpu_up(unsigned int cpu, int tasks_frozen)
500 int ret, nr_calls = 0;
501 void *hcpu = (void *)(long)cpu;
502 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
503 struct task_struct *idle;
507 if (cpu_online(cpu) || !cpu_present(cpu)) {
512 idle = idle_thread_get(cpu);
518 ret = smpboot_create_threads(cpu);
522 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
525 pr_warn("%s: attempt to bring up CPU %u failed\n",
530 /* Arch-specific enabling code. */
531 ret = __cpu_up(cpu, idle);
535 BUG_ON(!cpu_online(cpu));
537 /* Now call notifier in preparation. */
538 cpu_notify(CPU_ONLINE | mod, hcpu);
542 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
549 int cpu_up(unsigned int cpu)
553 if (!cpu_possible(cpu)) {
554 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
556 #if defined(CONFIG_IA64)
557 pr_err("please check additional_cpus= boot parameter\n");
562 err = try_online_node(cpu_to_node(cpu));
566 cpu_maps_update_begin();
568 if (cpu_hotplug_disabled) {
573 err = _cpu_up(cpu, 0);
576 cpu_maps_update_done();
579 EXPORT_SYMBOL_GPL(cpu_up);
581 #ifdef CONFIG_PM_SLEEP_SMP
582 static cpumask_var_t frozen_cpus;
584 int disable_nonboot_cpus(void)
586 int cpu, first_cpu, error = 0;
588 cpu_maps_update_begin();
589 first_cpu = cpumask_first(cpu_online_mask);
591 * We take down all of the non-boot CPUs in one shot to avoid races
592 * with the userspace trying to use the CPU hotplug at the same time
594 cpumask_clear(frozen_cpus);
596 pr_info("Disabling non-boot CPUs ...\n");
597 for_each_online_cpu(cpu) {
598 if (cpu == first_cpu)
600 trace_suspend_resume(TPS("CPU_OFF"), cpu, true);
601 error = _cpu_down(cpu, 1);
602 trace_suspend_resume(TPS("CPU_OFF"), cpu, false);
604 cpumask_set_cpu(cpu, frozen_cpus);
606 pr_err("Error taking CPU%d down: %d\n", cpu, error);
612 BUG_ON(num_online_cpus() > 1);
613 /* Make sure the CPUs won't be enabled by someone else */
614 cpu_hotplug_disabled = 1;
616 pr_err("Non-boot CPUs are not disabled\n");
618 cpu_maps_update_done();
622 void __weak arch_enable_nonboot_cpus_begin(void)
626 void __weak arch_enable_nonboot_cpus_end(void)
630 void __ref enable_nonboot_cpus(void)
634 /* Allow everyone to use the CPU hotplug again */
635 cpu_maps_update_begin();
636 cpu_hotplug_disabled = 0;
637 if (cpumask_empty(frozen_cpus))
640 pr_info("Enabling non-boot CPUs ...\n");
642 arch_enable_nonboot_cpus_begin();
644 for_each_cpu(cpu, frozen_cpus) {
645 trace_suspend_resume(TPS("CPU_ON"), cpu, true);
646 error = _cpu_up(cpu, 1);
647 trace_suspend_resume(TPS("CPU_ON"), cpu, false);
649 pr_info("CPU%d is up\n", cpu);
652 pr_warn("Error taking CPU%d up: %d\n", cpu, error);
655 arch_enable_nonboot_cpus_end();
657 cpumask_clear(frozen_cpus);
659 cpu_maps_update_done();
662 static int __init alloc_frozen_cpus(void)
664 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
668 core_initcall(alloc_frozen_cpus);
671 * When callbacks for CPU hotplug notifications are being executed, we must
672 * ensure that the state of the system with respect to the tasks being frozen
673 * or not, as reported by the notification, remains unchanged *throughout the
674 * duration* of the execution of the callbacks.
675 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
677 * This synchronization is implemented by mutually excluding regular CPU
678 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
679 * Hibernate notifications.
682 cpu_hotplug_pm_callback(struct notifier_block *nb,
683 unsigned long action, void *ptr)
687 case PM_SUSPEND_PREPARE:
688 case PM_HIBERNATION_PREPARE:
689 cpu_hotplug_disable();
692 case PM_POST_SUSPEND:
693 case PM_POST_HIBERNATION:
694 cpu_hotplug_enable();
705 static int __init cpu_hotplug_pm_sync_init(void)
708 * cpu_hotplug_pm_callback has higher priority than x86
709 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
710 * to disable cpu hotplug to avoid cpu hotplug race.
712 pm_notifier(cpu_hotplug_pm_callback, 0);
715 core_initcall(cpu_hotplug_pm_sync_init);
717 #endif /* CONFIG_PM_SLEEP_SMP */
720 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
721 * @cpu: cpu that just started
723 * This function calls the cpu_chain notifiers with CPU_STARTING.
724 * It must be called by the arch code on the new cpu, before the new cpu
725 * enables interrupts and before the "boot" cpu returns from __cpu_up().
727 void notify_cpu_starting(unsigned int cpu)
729 unsigned long val = CPU_STARTING;
731 #ifdef CONFIG_PM_SLEEP_SMP
732 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
733 val = CPU_STARTING_FROZEN;
734 #endif /* CONFIG_PM_SLEEP_SMP */
735 cpu_notify(val, (void *)(long)cpu);
738 #endif /* CONFIG_SMP */
741 * cpu_bit_bitmap[] is a special, "compressed" data structure that
742 * represents all NR_CPUS bits binary values of 1<<nr.
744 * It is used by cpumask_of() to get a constant address to a CPU
745 * mask value that has a single bit set only.
748 /* cpu_bit_bitmap[0] is empty - so we can back into it */
749 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
750 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
751 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
752 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
754 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
756 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
757 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
758 #if BITS_PER_LONG > 32
759 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
760 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
763 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
765 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
766 EXPORT_SYMBOL(cpu_all_bits);
768 #ifdef CONFIG_INIT_ALL_POSSIBLE
769 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
772 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
774 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
775 EXPORT_SYMBOL(cpu_possible_mask);
777 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
778 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
779 EXPORT_SYMBOL(cpu_online_mask);
781 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
782 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
783 EXPORT_SYMBOL(cpu_present_mask);
785 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
786 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
787 EXPORT_SYMBOL(cpu_active_mask);
789 void set_cpu_possible(unsigned int cpu, bool possible)
792 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
794 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
797 void set_cpu_present(unsigned int cpu, bool present)
800 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
802 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
805 void set_cpu_online(unsigned int cpu, bool online)
808 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
809 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
811 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
815 void set_cpu_active(unsigned int cpu, bool active)
818 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
820 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
823 void init_cpu_present(const struct cpumask *src)
825 cpumask_copy(to_cpumask(cpu_present_bits), src);
828 void init_cpu_possible(const struct cpumask *src)
830 cpumask_copy(to_cpumask(cpu_possible_bits), src);
833 void init_cpu_online(const struct cpumask *src)
835 cpumask_copy(to_cpumask(cpu_online_bits), src);