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 <trace/events/power.h>
28 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
29 static DEFINE_MUTEX(cpu_add_remove_lock
);
32 * The following two APIs (cpu_maps_update_begin/done) must be used when
33 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
34 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
35 * hotplug callback (un)registration performed using __register_cpu_notifier()
36 * or __unregister_cpu_notifier().
38 void cpu_maps_update_begin(void)
40 mutex_lock(&cpu_add_remove_lock
);
42 EXPORT_SYMBOL(cpu_notifier_register_begin
);
44 void cpu_maps_update_done(void)
46 mutex_unlock(&cpu_add_remove_lock
);
48 EXPORT_SYMBOL(cpu_notifier_register_done
);
50 static RAW_NOTIFIER_HEAD(cpu_chain
);
52 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
53 * Should always be manipulated under cpu_add_remove_lock
55 static int cpu_hotplug_disabled
;
57 #ifdef CONFIG_HOTPLUG_CPU
60 struct task_struct
*active_writer
;
61 struct mutex lock
; /* Synchronizes accesses to refcount, */
63 * Also blocks the new readers during
64 * an ongoing cpu hotplug operation.
67 /* And allows lockless put_online_cpus(). */
68 atomic_t puts_pending
;
70 #ifdef CONFIG_DEBUG_LOCK_ALLOC
71 struct lockdep_map dep_map
;
74 .active_writer
= NULL
,
75 .lock
= __MUTEX_INITIALIZER(cpu_hotplug
.lock
),
77 #ifdef CONFIG_DEBUG_LOCK_ALLOC
78 .dep_map
= {.name
= "cpu_hotplug.lock" },
82 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
83 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
84 #define cpuhp_lock_acquire_tryread() \
85 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
86 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
87 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
89 void get_online_cpus(void)
92 if (cpu_hotplug
.active_writer
== current
)
94 cpuhp_lock_acquire_read();
95 mutex_lock(&cpu_hotplug
.lock
);
96 cpu_hotplug
.refcount
++;
97 mutex_unlock(&cpu_hotplug
.lock
);
99 EXPORT_SYMBOL_GPL(get_online_cpus
);
101 bool try_get_online_cpus(void)
103 if (cpu_hotplug
.active_writer
== current
)
105 if (!mutex_trylock(&cpu_hotplug
.lock
))
107 cpuhp_lock_acquire_tryread();
108 cpu_hotplug
.refcount
++;
109 mutex_unlock(&cpu_hotplug
.lock
);
112 EXPORT_SYMBOL_GPL(try_get_online_cpus
);
114 void put_online_cpus(void)
116 if (cpu_hotplug
.active_writer
== current
)
118 if (!mutex_trylock(&cpu_hotplug
.lock
)) {
119 atomic_inc(&cpu_hotplug
.puts_pending
);
120 cpuhp_lock_release();
124 if (WARN_ON(!cpu_hotplug
.refcount
))
125 cpu_hotplug
.refcount
++; /* try to fix things up */
127 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
128 wake_up_process(cpu_hotplug
.active_writer
);
129 mutex_unlock(&cpu_hotplug
.lock
);
130 cpuhp_lock_release();
133 EXPORT_SYMBOL_GPL(put_online_cpus
);
136 * This ensures that the hotplug operation can begin only when the
137 * refcount goes to zero.
139 * Note that during a cpu-hotplug operation, the new readers, if any,
140 * will be blocked by the cpu_hotplug.lock
142 * Since cpu_hotplug_begin() is always called after invoking
143 * cpu_maps_update_begin(), we can be sure that only one writer is active.
145 * Note that theoretically, there is a possibility of a livelock:
146 * - Refcount goes to zero, last reader wakes up the sleeping
148 * - Last reader unlocks the cpu_hotplug.lock.
149 * - A new reader arrives at this moment, bumps up the refcount.
150 * - The writer acquires the cpu_hotplug.lock finds the refcount
151 * non zero and goes to sleep again.
153 * However, this is very difficult to achieve in practice since
154 * get_online_cpus() not an api which is called all that often.
157 void cpu_hotplug_begin(void)
159 cpu_hotplug
.active_writer
= current
;
161 cpuhp_lock_acquire();
163 mutex_lock(&cpu_hotplug
.lock
);
164 if (atomic_read(&cpu_hotplug
.puts_pending
)) {
167 delta
= atomic_xchg(&cpu_hotplug
.puts_pending
, 0);
168 cpu_hotplug
.refcount
-= delta
;
170 if (likely(!cpu_hotplug
.refcount
))
172 __set_current_state(TASK_UNINTERRUPTIBLE
);
173 mutex_unlock(&cpu_hotplug
.lock
);
178 void cpu_hotplug_done(void)
180 cpu_hotplug
.active_writer
= NULL
;
181 mutex_unlock(&cpu_hotplug
.lock
);
182 cpuhp_lock_release();
186 * Wait for currently running CPU hotplug operations to complete (if any) and
187 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
188 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
189 * hotplug path before performing hotplug operations. So acquiring that lock
190 * guarantees mutual exclusion from any currently running hotplug operations.
192 void cpu_hotplug_disable(void)
194 cpu_maps_update_begin();
195 cpu_hotplug_disabled
= 1;
196 cpu_maps_update_done();
199 void cpu_hotplug_enable(void)
201 cpu_maps_update_begin();
202 cpu_hotplug_disabled
= 0;
203 cpu_maps_update_done();
206 #endif /* CONFIG_HOTPLUG_CPU */
208 /* Need to know about CPUs going up/down? */
209 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
212 cpu_maps_update_begin();
213 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
214 cpu_maps_update_done();
218 int __ref
__register_cpu_notifier(struct notifier_block
*nb
)
220 return raw_notifier_chain_register(&cpu_chain
, nb
);
223 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
228 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
231 return notifier_to_errno(ret
);
234 static int cpu_notify(unsigned long val
, void *v
)
236 return __cpu_notify(val
, v
, -1, NULL
);
239 #ifdef CONFIG_HOTPLUG_CPU
241 static void cpu_notify_nofail(unsigned long val
, void *v
)
243 BUG_ON(cpu_notify(val
, v
));
245 EXPORT_SYMBOL(register_cpu_notifier
);
246 EXPORT_SYMBOL(__register_cpu_notifier
);
248 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
250 cpu_maps_update_begin();
251 raw_notifier_chain_unregister(&cpu_chain
, nb
);
252 cpu_maps_update_done();
254 EXPORT_SYMBOL(unregister_cpu_notifier
);
256 void __ref
__unregister_cpu_notifier(struct notifier_block
*nb
)
258 raw_notifier_chain_unregister(&cpu_chain
, nb
);
260 EXPORT_SYMBOL(__unregister_cpu_notifier
);
263 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
266 * This function walks all processes, finds a valid mm struct for each one and
267 * then clears a corresponding bit in mm's cpumask. While this all sounds
268 * trivial, there are various non-obvious corner cases, which this function
269 * tries to solve in a safe manner.
271 * Also note that the function uses a somewhat relaxed locking scheme, so it may
272 * be called only for an already offlined CPU.
274 void clear_tasks_mm_cpumask(int cpu
)
276 struct task_struct
*p
;
279 * This function is called after the cpu is taken down and marked
280 * offline, so its not like new tasks will ever get this cpu set in
281 * their mm mask. -- Peter Zijlstra
282 * Thus, we may use rcu_read_lock() here, instead of grabbing
283 * full-fledged tasklist_lock.
285 WARN_ON(cpu_online(cpu
));
287 for_each_process(p
) {
288 struct task_struct
*t
;
291 * Main thread might exit, but other threads may still have
292 * a valid mm. Find one.
294 t
= find_lock_task_mm(p
);
297 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
303 static inline void check_for_tasks(int dead_cpu
)
305 struct task_struct
*g
, *p
;
307 read_lock_irq(&tasklist_lock
);
308 do_each_thread(g
, p
) {
312 * We do the check with unlocked task_rq(p)->lock.
313 * Order the reading to do not warn about a task,
314 * which was running on this cpu in the past, and
315 * it's just been woken on another cpu.
318 if (task_cpu(p
) != dead_cpu
)
321 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
322 p
->comm
, task_pid_nr(p
), dead_cpu
, p
->state
, p
->flags
);
323 } while_each_thread(g
, p
);
324 read_unlock_irq(&tasklist_lock
);
327 struct take_cpu_down_param
{
332 /* Take this CPU down. */
333 static int __ref
take_cpu_down(void *_param
)
335 struct take_cpu_down_param
*param
= _param
;
338 /* Ensure this CPU doesn't handle any more interrupts. */
339 err
= __cpu_disable();
343 cpu_notify(CPU_DYING
| param
->mod
, param
->hcpu
);
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 /* This actually kills the CPU. */
420 /* CPU is completely dead: tell everyone. Too late to complain. */
421 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
423 check_for_tasks(cpu
);
428 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
432 int __ref
cpu_down(unsigned int cpu
)
436 cpu_maps_update_begin();
438 if (cpu_hotplug_disabled
) {
443 err
= _cpu_down(cpu
, 0);
446 cpu_maps_update_done();
449 EXPORT_SYMBOL(cpu_down
);
450 #endif /*CONFIG_HOTPLUG_CPU*/
452 /* Requires cpu_add_remove_lock to be held */
453 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
455 int ret
, nr_calls
= 0;
456 void *hcpu
= (void *)(long)cpu
;
457 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
458 struct task_struct
*idle
;
462 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
467 idle
= idle_thread_get(cpu
);
473 ret
= smpboot_create_threads(cpu
);
477 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
480 pr_warn("%s: attempt to bring up CPU %u failed\n",
485 /* Arch-specific enabling code. */
486 ret
= __cpu_up(cpu
, idle
);
489 BUG_ON(!cpu_online(cpu
));
491 /* Wake the per cpu threads */
492 smpboot_unpark_threads(cpu
);
494 /* Now call notifier in preparation. */
495 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
499 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
506 int cpu_up(unsigned int cpu
)
510 if (!cpu_possible(cpu
)) {
511 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
513 #if defined(CONFIG_IA64)
514 pr_err("please check additional_cpus= boot parameter\n");
519 err
= try_online_node(cpu_to_node(cpu
));
523 cpu_maps_update_begin();
525 if (cpu_hotplug_disabled
) {
530 err
= _cpu_up(cpu
, 0);
533 cpu_maps_update_done();
536 EXPORT_SYMBOL_GPL(cpu_up
);
538 #ifdef CONFIG_PM_SLEEP_SMP
539 static cpumask_var_t frozen_cpus
;
541 int disable_nonboot_cpus(void)
543 int cpu
, first_cpu
, error
= 0;
545 cpu_maps_update_begin();
546 first_cpu
= cpumask_first(cpu_online_mask
);
548 * We take down all of the non-boot CPUs in one shot to avoid races
549 * with the userspace trying to use the CPU hotplug at the same time
551 cpumask_clear(frozen_cpus
);
553 pr_info("Disabling non-boot CPUs ...\n");
554 for_each_online_cpu(cpu
) {
555 if (cpu
== first_cpu
)
557 trace_suspend_resume(TPS("CPU_OFF"), cpu
, true);
558 error
= _cpu_down(cpu
, 1);
559 trace_suspend_resume(TPS("CPU_OFF"), cpu
, false);
561 cpumask_set_cpu(cpu
, frozen_cpus
);
563 pr_err("Error taking CPU%d down: %d\n", cpu
, error
);
569 BUG_ON(num_online_cpus() > 1);
570 /* Make sure the CPUs won't be enabled by someone else */
571 cpu_hotplug_disabled
= 1;
573 pr_err("Non-boot CPUs are not disabled\n");
575 cpu_maps_update_done();
579 void __weak
arch_enable_nonboot_cpus_begin(void)
583 void __weak
arch_enable_nonboot_cpus_end(void)
587 void __ref
enable_nonboot_cpus(void)
591 /* Allow everyone to use the CPU hotplug again */
592 cpu_maps_update_begin();
593 cpu_hotplug_disabled
= 0;
594 if (cpumask_empty(frozen_cpus
))
597 pr_info("Enabling non-boot CPUs ...\n");
599 arch_enable_nonboot_cpus_begin();
601 for_each_cpu(cpu
, frozen_cpus
) {
602 trace_suspend_resume(TPS("CPU_ON"), cpu
, true);
603 error
= _cpu_up(cpu
, 1);
604 trace_suspend_resume(TPS("CPU_ON"), cpu
, false);
606 pr_info("CPU%d is up\n", cpu
);
609 pr_warn("Error taking CPU%d up: %d\n", cpu
, error
);
612 arch_enable_nonboot_cpus_end();
614 cpumask_clear(frozen_cpus
);
616 cpu_maps_update_done();
619 static int __init
alloc_frozen_cpus(void)
621 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
625 core_initcall(alloc_frozen_cpus
);
628 * When callbacks for CPU hotplug notifications are being executed, we must
629 * ensure that the state of the system with respect to the tasks being frozen
630 * or not, as reported by the notification, remains unchanged *throughout the
631 * duration* of the execution of the callbacks.
632 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
634 * This synchronization is implemented by mutually excluding regular CPU
635 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
636 * Hibernate notifications.
639 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
640 unsigned long action
, void *ptr
)
644 case PM_SUSPEND_PREPARE
:
645 case PM_HIBERNATION_PREPARE
:
646 cpu_hotplug_disable();
649 case PM_POST_SUSPEND
:
650 case PM_POST_HIBERNATION
:
651 cpu_hotplug_enable();
662 static int __init
cpu_hotplug_pm_sync_init(void)
665 * cpu_hotplug_pm_callback has higher priority than x86
666 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
667 * to disable cpu hotplug to avoid cpu hotplug race.
669 pm_notifier(cpu_hotplug_pm_callback
, 0);
672 core_initcall(cpu_hotplug_pm_sync_init
);
674 #endif /* CONFIG_PM_SLEEP_SMP */
677 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
678 * @cpu: cpu that just started
680 * This function calls the cpu_chain notifiers with CPU_STARTING.
681 * It must be called by the arch code on the new cpu, before the new cpu
682 * enables interrupts and before the "boot" cpu returns from __cpu_up().
684 void notify_cpu_starting(unsigned int cpu
)
686 unsigned long val
= CPU_STARTING
;
688 #ifdef CONFIG_PM_SLEEP_SMP
689 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
690 val
= CPU_STARTING_FROZEN
;
691 #endif /* CONFIG_PM_SLEEP_SMP */
692 cpu_notify(val
, (void *)(long)cpu
);
695 #endif /* CONFIG_SMP */
698 * cpu_bit_bitmap[] is a special, "compressed" data structure that
699 * represents all NR_CPUS bits binary values of 1<<nr.
701 * It is used by cpumask_of() to get a constant address to a CPU
702 * mask value that has a single bit set only.
705 /* cpu_bit_bitmap[0] is empty - so we can back into it */
706 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
707 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
708 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
709 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
711 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
713 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
714 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
715 #if BITS_PER_LONG > 32
716 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
717 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
720 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
722 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
723 EXPORT_SYMBOL(cpu_all_bits
);
725 #ifdef CONFIG_INIT_ALL_POSSIBLE
726 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
729 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
731 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
732 EXPORT_SYMBOL(cpu_possible_mask
);
734 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
735 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
736 EXPORT_SYMBOL(cpu_online_mask
);
738 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
739 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
740 EXPORT_SYMBOL(cpu_present_mask
);
742 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
743 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
744 EXPORT_SYMBOL(cpu_active_mask
);
746 void set_cpu_possible(unsigned int cpu
, bool possible
)
749 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
751 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
754 void set_cpu_present(unsigned int cpu
, bool present
)
757 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
759 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
762 void set_cpu_online(unsigned int cpu
, bool online
)
765 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
766 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
768 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
772 void set_cpu_active(unsigned int cpu
, bool active
)
775 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
777 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
780 void init_cpu_present(const struct cpumask
*src
)
782 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
785 void init_cpu_possible(const struct cpumask
*src
)
787 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
790 void init_cpu_online(const struct cpumask
*src
)
792 cpumask_copy(to_cpumask(cpu_online_bits
), src
);