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>
26 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
27 static DEFINE_MUTEX(cpu_add_remove_lock);
30 * The following two API's must be used when attempting
31 * to serialize the updates to cpu_online_mask, cpu_present_mask.
33 void cpu_maps_update_begin(void)
35 mutex_lock(&cpu_add_remove_lock);
38 void cpu_maps_update_done(void)
40 mutex_unlock(&cpu_add_remove_lock);
43 static RAW_NOTIFIER_HEAD(cpu_chain);
45 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
46 * Should always be manipulated under cpu_add_remove_lock
48 static int cpu_hotplug_disabled;
50 #ifdef CONFIG_HOTPLUG_CPU
53 struct task_struct *active_writer;
54 struct mutex lock; /* Synchronizes accesses to refcount, */
56 * Also blocks the new readers during
57 * an ongoing cpu hotplug operation.
61 .active_writer = NULL,
62 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
66 void get_online_cpus(void)
69 if (cpu_hotplug.active_writer == current)
71 mutex_lock(&cpu_hotplug.lock);
72 cpu_hotplug.refcount++;
73 mutex_unlock(&cpu_hotplug.lock);
76 EXPORT_SYMBOL_GPL(get_online_cpus);
78 void put_online_cpus(void)
80 if (cpu_hotplug.active_writer == current)
82 mutex_lock(&cpu_hotplug.lock);
83 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
84 wake_up_process(cpu_hotplug.active_writer);
85 mutex_unlock(&cpu_hotplug.lock);
88 EXPORT_SYMBOL_GPL(put_online_cpus);
91 * This ensures that the hotplug operation can begin only when the
92 * refcount goes to zero.
94 * Note that during a cpu-hotplug operation, the new readers, if any,
95 * will be blocked by the cpu_hotplug.lock
97 * Since cpu_hotplug_begin() is always called after invoking
98 * cpu_maps_update_begin(), we can be sure that only one writer is active.
100 * Note that theoretically, there is a possibility of a livelock:
101 * - Refcount goes to zero, last reader wakes up the sleeping
103 * - Last reader unlocks the cpu_hotplug.lock.
104 * - A new reader arrives at this moment, bumps up the refcount.
105 * - The writer acquires the cpu_hotplug.lock finds the refcount
106 * non zero and goes to sleep again.
108 * However, this is very difficult to achieve in practice since
109 * get_online_cpus() not an api which is called all that often.
112 static void cpu_hotplug_begin(void)
114 cpu_hotplug.active_writer = current;
117 mutex_lock(&cpu_hotplug.lock);
118 if (likely(!cpu_hotplug.refcount))
120 __set_current_state(TASK_UNINTERRUPTIBLE);
121 mutex_unlock(&cpu_hotplug.lock);
126 static void cpu_hotplug_done(void)
128 cpu_hotplug.active_writer = NULL;
129 mutex_unlock(&cpu_hotplug.lock);
132 #else /* #if CONFIG_HOTPLUG_CPU */
133 static void cpu_hotplug_begin(void) {}
134 static void cpu_hotplug_done(void) {}
135 #endif /* #else #if CONFIG_HOTPLUG_CPU */
137 /* Need to know about CPUs going up/down? */
138 int __ref register_cpu_notifier(struct notifier_block *nb)
141 cpu_maps_update_begin();
142 ret = raw_notifier_chain_register(&cpu_chain, nb);
143 cpu_maps_update_done();
147 static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
152 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
155 return notifier_to_errno(ret);
158 static int cpu_notify(unsigned long val, void *v)
160 return __cpu_notify(val, v, -1, NULL);
163 #ifdef CONFIG_HOTPLUG_CPU
165 static void cpu_notify_nofail(unsigned long val, void *v)
167 BUG_ON(cpu_notify(val, v));
169 EXPORT_SYMBOL(register_cpu_notifier);
171 void __ref unregister_cpu_notifier(struct notifier_block *nb)
173 cpu_maps_update_begin();
174 raw_notifier_chain_unregister(&cpu_chain, nb);
175 cpu_maps_update_done();
177 EXPORT_SYMBOL(unregister_cpu_notifier);
180 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
183 * This function walks all processes, finds a valid mm struct for each one and
184 * then clears a corresponding bit in mm's cpumask. While this all sounds
185 * trivial, there are various non-obvious corner cases, which this function
186 * tries to solve in a safe manner.
188 * Also note that the function uses a somewhat relaxed locking scheme, so it may
189 * be called only for an already offlined CPU.
191 void clear_tasks_mm_cpumask(int cpu)
193 struct task_struct *p;
196 * This function is called after the cpu is taken down and marked
197 * offline, so its not like new tasks will ever get this cpu set in
198 * their mm mask. -- Peter Zijlstra
199 * Thus, we may use rcu_read_lock() here, instead of grabbing
200 * full-fledged tasklist_lock.
202 WARN_ON(cpu_online(cpu));
204 for_each_process(p) {
205 struct task_struct *t;
208 * Main thread might exit, but other threads may still have
209 * a valid mm. Find one.
211 t = find_lock_task_mm(p);
214 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
220 static inline void check_for_tasks(int cpu)
222 struct task_struct *p;
224 write_lock_irq(&tasklist_lock);
225 for_each_process(p) {
226 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
227 (p->utime || p->stime))
228 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
229 "(state = %ld, flags = %x)\n",
230 p->comm, task_pid_nr(p), cpu,
233 write_unlock_irq(&tasklist_lock);
236 struct take_cpu_down_param {
241 /* Take this CPU down. */
242 static int __ref take_cpu_down(void *_param)
244 struct take_cpu_down_param *param = _param;
247 /* Ensure this CPU doesn't handle any more interrupts. */
248 err = __cpu_disable();
252 cpu_notify(CPU_DYING | param->mod, param->hcpu);
256 /* Requires cpu_add_remove_lock to be held */
257 static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
259 int err, nr_calls = 0;
260 void *hcpu = (void *)(long)cpu;
261 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
262 struct take_cpu_down_param tcd_param = {
267 if (num_online_cpus() == 1)
270 if (!cpu_online(cpu))
275 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
278 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
279 printk("%s: attempt to take down CPU %u failed\n",
283 smpboot_park_threads(cpu);
285 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
287 /* CPU didn't die: tell everyone. Can't complain. */
288 smpboot_unpark_threads(cpu);
289 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
292 BUG_ON(cpu_online(cpu));
295 * The migration_call() CPU_DYING callback will have removed all
296 * runnable tasks from the cpu, there's only the idle task left now
297 * that the migration thread is done doing the stop_machine thing.
299 * Wait for the stop thread to go away.
301 while (!idle_cpu(cpu))
304 /* This actually kills the CPU. */
307 /* CPU is completely dead: tell everyone. Too late to complain. */
308 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
310 check_for_tasks(cpu);
315 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
319 int __ref cpu_down(unsigned int cpu)
323 cpu_maps_update_begin();
325 if (cpu_hotplug_disabled) {
330 err = _cpu_down(cpu, 0);
333 cpu_maps_update_done();
336 EXPORT_SYMBOL(cpu_down);
337 #endif /*CONFIG_HOTPLUG_CPU*/
339 /* Requires cpu_add_remove_lock to be held */
340 static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
342 int ret, nr_calls = 0;
343 void *hcpu = (void *)(long)cpu;
344 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
345 struct task_struct *idle;
347 if (cpu_online(cpu) || !cpu_present(cpu))
352 idle = idle_thread_get(cpu);
358 ret = smpboot_create_threads(cpu);
362 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
365 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
370 /* Arch-specific enabling code. */
371 ret = __cpu_up(cpu, idle);
374 BUG_ON(!cpu_online(cpu));
376 /* Wake the per cpu threads */
377 smpboot_unpark_threads(cpu);
379 /* Now call notifier in preparation. */
380 cpu_notify(CPU_ONLINE | mod, hcpu);
384 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
391 int __cpuinit cpu_up(unsigned int cpu)
395 #ifdef CONFIG_MEMORY_HOTPLUG
400 if (!cpu_possible(cpu)) {
401 printk(KERN_ERR "can't online cpu %d because it is not "
402 "configured as may-hotadd at boot time\n", cpu);
403 #if defined(CONFIG_IA64)
404 printk(KERN_ERR "please check additional_cpus= boot "
410 #ifdef CONFIG_MEMORY_HOTPLUG
411 nid = cpu_to_node(cpu);
412 if (!node_online(nid)) {
413 err = mem_online_node(nid);
418 pgdat = NODE_DATA(nid);
421 "Can't online cpu %d due to NULL pgdat\n", cpu);
425 if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
426 mutex_lock(&zonelists_mutex);
427 build_all_zonelists(NULL, NULL);
428 mutex_unlock(&zonelists_mutex);
432 cpu_maps_update_begin();
434 if (cpu_hotplug_disabled) {
439 err = _cpu_up(cpu, 0);
442 cpu_maps_update_done();
445 EXPORT_SYMBOL_GPL(cpu_up);
447 #ifdef CONFIG_PM_SLEEP_SMP
448 static cpumask_var_t frozen_cpus;
450 int disable_nonboot_cpus(void)
452 int cpu, first_cpu, error = 0;
454 cpu_maps_update_begin();
455 first_cpu = cpumask_first(cpu_online_mask);
457 * We take down all of the non-boot CPUs in one shot to avoid races
458 * with the userspace trying to use the CPU hotplug at the same time
460 cpumask_clear(frozen_cpus);
462 printk("Disabling non-boot CPUs ...\n");
463 for_each_online_cpu(cpu) {
464 if (cpu == first_cpu)
466 error = _cpu_down(cpu, 1);
468 cpumask_set_cpu(cpu, frozen_cpus);
470 printk(KERN_ERR "Error taking CPU%d down: %d\n",
477 BUG_ON(num_online_cpus() > 1);
478 /* Make sure the CPUs won't be enabled by someone else */
479 cpu_hotplug_disabled = 1;
481 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
483 cpu_maps_update_done();
487 void __weak arch_enable_nonboot_cpus_begin(void)
491 void __weak arch_enable_nonboot_cpus_end(void)
495 void __ref enable_nonboot_cpus(void)
499 /* Allow everyone to use the CPU hotplug again */
500 cpu_maps_update_begin();
501 cpu_hotplug_disabled = 0;
502 if (cpumask_empty(frozen_cpus))
505 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
507 arch_enable_nonboot_cpus_begin();
509 for_each_cpu(cpu, frozen_cpus) {
510 error = _cpu_up(cpu, 1);
512 printk(KERN_INFO "CPU%d is up\n", cpu);
515 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
518 arch_enable_nonboot_cpus_end();
520 cpumask_clear(frozen_cpus);
522 cpu_maps_update_done();
525 static int __init alloc_frozen_cpus(void)
527 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
531 core_initcall(alloc_frozen_cpus);
534 * Prevent regular CPU hotplug from racing with the freezer, by disabling CPU
535 * hotplug when tasks are about to be frozen. Also, don't allow the freezer
536 * to continue until any currently running CPU hotplug operation gets
538 * To modify the 'cpu_hotplug_disabled' flag, we need to acquire the
539 * 'cpu_add_remove_lock'. And this same lock is also taken by the regular
540 * CPU hotplug path and released only after it is complete. Thus, we
541 * (and hence the freezer) will block here until any currently running CPU
542 * hotplug operation gets completed.
544 void cpu_hotplug_disable_before_freeze(void)
546 cpu_maps_update_begin();
547 cpu_hotplug_disabled = 1;
548 cpu_maps_update_done();
553 * When tasks have been thawed, re-enable regular CPU hotplug (which had been
554 * disabled while beginning to freeze tasks).
556 void cpu_hotplug_enable_after_thaw(void)
558 cpu_maps_update_begin();
559 cpu_hotplug_disabled = 0;
560 cpu_maps_update_done();
564 * When callbacks for CPU hotplug notifications are being executed, we must
565 * ensure that the state of the system with respect to the tasks being frozen
566 * or not, as reported by the notification, remains unchanged *throughout the
567 * duration* of the execution of the callbacks.
568 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
570 * This synchronization is implemented by mutually excluding regular CPU
571 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
572 * Hibernate notifications.
575 cpu_hotplug_pm_callback(struct notifier_block *nb,
576 unsigned long action, void *ptr)
580 case PM_SUSPEND_PREPARE:
581 case PM_HIBERNATION_PREPARE:
582 cpu_hotplug_disable_before_freeze();
585 case PM_POST_SUSPEND:
586 case PM_POST_HIBERNATION:
587 cpu_hotplug_enable_after_thaw();
598 static int __init cpu_hotplug_pm_sync_init(void)
600 pm_notifier(cpu_hotplug_pm_callback, 0);
603 core_initcall(cpu_hotplug_pm_sync_init);
605 #endif /* CONFIG_PM_SLEEP_SMP */
608 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
609 * @cpu: cpu that just started
611 * This function calls the cpu_chain notifiers with CPU_STARTING.
612 * It must be called by the arch code on the new cpu, before the new cpu
613 * enables interrupts and before the "boot" cpu returns from __cpu_up().
615 void __cpuinit notify_cpu_starting(unsigned int cpu)
617 unsigned long val = CPU_STARTING;
619 #ifdef CONFIG_PM_SLEEP_SMP
620 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
621 val = CPU_STARTING_FROZEN;
622 #endif /* CONFIG_PM_SLEEP_SMP */
623 cpu_notify(val, (void *)(long)cpu);
626 #endif /* CONFIG_SMP */
629 * cpu_bit_bitmap[] is a special, "compressed" data structure that
630 * represents all NR_CPUS bits binary values of 1<<nr.
632 * It is used by cpumask_of() to get a constant address to a CPU
633 * mask value that has a single bit set only.
636 /* cpu_bit_bitmap[0] is empty - so we can back into it */
637 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
638 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
639 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
640 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
642 const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
644 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
645 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
646 #if BITS_PER_LONG > 32
647 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
648 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
651 EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
653 const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
654 EXPORT_SYMBOL(cpu_all_bits);
656 #ifdef CONFIG_INIT_ALL_POSSIBLE
657 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
660 static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
662 const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
663 EXPORT_SYMBOL(cpu_possible_mask);
665 static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
666 const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
667 EXPORT_SYMBOL(cpu_online_mask);
669 static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
670 const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
671 EXPORT_SYMBOL(cpu_present_mask);
673 static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
674 const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
675 EXPORT_SYMBOL(cpu_active_mask);
677 void set_cpu_possible(unsigned int cpu, bool possible)
680 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
682 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
685 void set_cpu_present(unsigned int cpu, bool present)
688 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
690 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
693 void set_cpu_online(unsigned int cpu, bool online)
696 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
698 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
701 void set_cpu_active(unsigned int cpu, bool active)
704 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
706 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
709 void init_cpu_present(const struct cpumask *src)
711 cpumask_copy(to_cpumask(cpu_present_bits), src);
714 void init_cpu_possible(const struct cpumask *src)
716 cpumask_copy(to_cpumask(cpu_possible_bits), src);
719 void init_cpu_online(const struct cpumask *src)
721 cpumask_copy(to_cpumask(cpu_online_bits), src);