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 static void apply_puts_pending(int max
)
93 if (atomic_read(&cpu_hotplug
.puts_pending
) >= max
) {
94 delta
= atomic_xchg(&cpu_hotplug
.puts_pending
, 0);
95 cpu_hotplug
.refcount
-= delta
;
99 void get_online_cpus(void)
102 if (cpu_hotplug
.active_writer
== current
)
104 cpuhp_lock_acquire_read();
105 mutex_lock(&cpu_hotplug
.lock
);
106 apply_puts_pending(65536);
107 cpu_hotplug
.refcount
++;
108 mutex_unlock(&cpu_hotplug
.lock
);
110 EXPORT_SYMBOL_GPL(get_online_cpus
);
112 bool try_get_online_cpus(void)
114 if (cpu_hotplug
.active_writer
== current
)
116 if (!mutex_trylock(&cpu_hotplug
.lock
))
118 cpuhp_lock_acquire_tryread();
119 apply_puts_pending(65536);
120 cpu_hotplug
.refcount
++;
121 mutex_unlock(&cpu_hotplug
.lock
);
124 EXPORT_SYMBOL_GPL(try_get_online_cpus
);
126 void put_online_cpus(void)
128 if (cpu_hotplug
.active_writer
== current
)
130 if (!mutex_trylock(&cpu_hotplug
.lock
)) {
131 atomic_inc(&cpu_hotplug
.puts_pending
);
132 cpuhp_lock_release();
136 if (WARN_ON(!cpu_hotplug
.refcount
))
137 cpu_hotplug
.refcount
++; /* try to fix things up */
139 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
140 wake_up_process(cpu_hotplug
.active_writer
);
141 mutex_unlock(&cpu_hotplug
.lock
);
142 cpuhp_lock_release();
145 EXPORT_SYMBOL_GPL(put_online_cpus
);
148 * This ensures that the hotplug operation can begin only when the
149 * refcount goes to zero.
151 * Note that during a cpu-hotplug operation, the new readers, if any,
152 * will be blocked by the cpu_hotplug.lock
154 * Since cpu_hotplug_begin() is always called after invoking
155 * cpu_maps_update_begin(), we can be sure that only one writer is active.
157 * Note that theoretically, there is a possibility of a livelock:
158 * - Refcount goes to zero, last reader wakes up the sleeping
160 * - Last reader unlocks the cpu_hotplug.lock.
161 * - A new reader arrives at this moment, bumps up the refcount.
162 * - The writer acquires the cpu_hotplug.lock finds the refcount
163 * non zero and goes to sleep again.
165 * However, this is very difficult to achieve in practice since
166 * get_online_cpus() not an api which is called all that often.
169 void cpu_hotplug_begin(void)
171 cpu_hotplug
.active_writer
= current
;
173 cpuhp_lock_acquire();
175 mutex_lock(&cpu_hotplug
.lock
);
176 apply_puts_pending(1);
177 if (likely(!cpu_hotplug
.refcount
))
179 __set_current_state(TASK_UNINTERRUPTIBLE
);
180 mutex_unlock(&cpu_hotplug
.lock
);
185 void cpu_hotplug_done(void)
187 cpu_hotplug
.active_writer
= NULL
;
188 mutex_unlock(&cpu_hotplug
.lock
);
189 cpuhp_lock_release();
193 * Wait for currently running CPU hotplug operations to complete (if any) and
194 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
195 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
196 * hotplug path before performing hotplug operations. So acquiring that lock
197 * guarantees mutual exclusion from any currently running hotplug operations.
199 void cpu_hotplug_disable(void)
201 cpu_maps_update_begin();
202 cpu_hotplug_disabled
= 1;
203 cpu_maps_update_done();
206 void cpu_hotplug_enable(void)
208 cpu_maps_update_begin();
209 cpu_hotplug_disabled
= 0;
210 cpu_maps_update_done();
213 #endif /* CONFIG_HOTPLUG_CPU */
215 /* Need to know about CPUs going up/down? */
216 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
219 cpu_maps_update_begin();
220 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
221 cpu_maps_update_done();
225 int __ref
__register_cpu_notifier(struct notifier_block
*nb
)
227 return raw_notifier_chain_register(&cpu_chain
, nb
);
230 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
235 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
238 return notifier_to_errno(ret
);
241 static int cpu_notify(unsigned long val
, void *v
)
243 return __cpu_notify(val
, v
, -1, NULL
);
246 #ifdef CONFIG_HOTPLUG_CPU
248 static void cpu_notify_nofail(unsigned long val
, void *v
)
250 BUG_ON(cpu_notify(val
, v
));
252 EXPORT_SYMBOL(register_cpu_notifier
);
253 EXPORT_SYMBOL(__register_cpu_notifier
);
255 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
257 cpu_maps_update_begin();
258 raw_notifier_chain_unregister(&cpu_chain
, nb
);
259 cpu_maps_update_done();
261 EXPORT_SYMBOL(unregister_cpu_notifier
);
263 void __ref
__unregister_cpu_notifier(struct notifier_block
*nb
)
265 raw_notifier_chain_unregister(&cpu_chain
, nb
);
267 EXPORT_SYMBOL(__unregister_cpu_notifier
);
270 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
273 * This function walks all processes, finds a valid mm struct for each one and
274 * then clears a corresponding bit in mm's cpumask. While this all sounds
275 * trivial, there are various non-obvious corner cases, which this function
276 * tries to solve in a safe manner.
278 * Also note that the function uses a somewhat relaxed locking scheme, so it may
279 * be called only for an already offlined CPU.
281 void clear_tasks_mm_cpumask(int cpu
)
283 struct task_struct
*p
;
286 * This function is called after the cpu is taken down and marked
287 * offline, so its not like new tasks will ever get this cpu set in
288 * their mm mask. -- Peter Zijlstra
289 * Thus, we may use rcu_read_lock() here, instead of grabbing
290 * full-fledged tasklist_lock.
292 WARN_ON(cpu_online(cpu
));
294 for_each_process(p
) {
295 struct task_struct
*t
;
298 * Main thread might exit, but other threads may still have
299 * a valid mm. Find one.
301 t
= find_lock_task_mm(p
);
304 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
310 static inline void check_for_tasks(int dead_cpu
)
312 struct task_struct
*g
, *p
;
314 read_lock_irq(&tasklist_lock
);
315 do_each_thread(g
, p
) {
319 * We do the check with unlocked task_rq(p)->lock.
320 * Order the reading to do not warn about a task,
321 * which was running on this cpu in the past, and
322 * it's just been woken on another cpu.
325 if (task_cpu(p
) != dead_cpu
)
328 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
329 p
->comm
, task_pid_nr(p
), dead_cpu
, p
->state
, p
->flags
);
330 } while_each_thread(g
, p
);
331 read_unlock_irq(&tasklist_lock
);
334 struct take_cpu_down_param
{
339 /* Take this CPU down. */
340 static int __ref
take_cpu_down(void *_param
)
342 struct take_cpu_down_param
*param
= _param
;
345 /* Ensure this CPU doesn't handle any more interrupts. */
346 err
= __cpu_disable();
350 cpu_notify(CPU_DYING
| param
->mod
, param
->hcpu
);
351 /* Park the stopper thread */
352 kthread_park(current
);
356 /* Requires cpu_add_remove_lock to be held */
357 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
359 int err
, nr_calls
= 0;
360 void *hcpu
= (void *)(long)cpu
;
361 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
362 struct take_cpu_down_param tcd_param
= {
367 if (num_online_cpus() == 1)
370 if (!cpu_online(cpu
))
375 err
= __cpu_notify(CPU_DOWN_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
378 __cpu_notify(CPU_DOWN_FAILED
| mod
, hcpu
, nr_calls
, NULL
);
379 pr_warn("%s: attempt to take down CPU %u failed\n",
385 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
386 * and RCU users of this state to go away such that all new such users
389 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
390 * not imply sync_sched(), so explicitly call both.
392 * Do sync before park smpboot threads to take care the rcu boost case.
394 #ifdef CONFIG_PREEMPT
399 smpboot_park_threads(cpu
);
402 * So now all preempt/rcu users must observe !cpu_active().
405 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
407 /* CPU didn't die: tell everyone. Can't complain. */
408 smpboot_unpark_threads(cpu
);
409 cpu_notify_nofail(CPU_DOWN_FAILED
| mod
, hcpu
);
412 BUG_ON(cpu_online(cpu
));
415 * The migration_call() CPU_DYING callback will have removed all
416 * runnable tasks from the cpu, there's only the idle task left now
417 * that the migration thread is done doing the stop_machine thing.
419 * Wait for the stop thread to go away.
421 while (!idle_cpu(cpu
))
424 /* This actually kills the CPU. */
427 /* CPU is completely dead: tell everyone. Too late to complain. */
428 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
430 check_for_tasks(cpu
);
435 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
439 int __ref
cpu_down(unsigned int cpu
)
443 cpu_maps_update_begin();
445 if (cpu_hotplug_disabled
) {
450 err
= _cpu_down(cpu
, 0);
453 cpu_maps_update_done();
456 EXPORT_SYMBOL(cpu_down
);
457 #endif /*CONFIG_HOTPLUG_CPU*/
459 /* Requires cpu_add_remove_lock to be held */
460 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
462 int ret
, nr_calls
= 0;
463 void *hcpu
= (void *)(long)cpu
;
464 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
465 struct task_struct
*idle
;
469 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
474 idle
= idle_thread_get(cpu
);
480 ret
= smpboot_create_threads(cpu
);
484 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
487 pr_warn("%s: attempt to bring up CPU %u failed\n",
492 /* Arch-specific enabling code. */
493 ret
= __cpu_up(cpu
, idle
);
496 BUG_ON(!cpu_online(cpu
));
498 /* Wake the per cpu threads */
499 smpboot_unpark_threads(cpu
);
501 /* Now call notifier in preparation. */
502 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
506 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
513 int cpu_up(unsigned int cpu
)
517 if (!cpu_possible(cpu
)) {
518 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
520 #if defined(CONFIG_IA64)
521 pr_err("please check additional_cpus= boot parameter\n");
526 err
= try_online_node(cpu_to_node(cpu
));
530 cpu_maps_update_begin();
532 if (cpu_hotplug_disabled
) {
537 err
= _cpu_up(cpu
, 0);
540 cpu_maps_update_done();
543 EXPORT_SYMBOL_GPL(cpu_up
);
545 #ifdef CONFIG_PM_SLEEP_SMP
546 static cpumask_var_t frozen_cpus
;
548 int disable_nonboot_cpus(void)
550 int cpu
, first_cpu
, error
= 0;
552 cpu_maps_update_begin();
553 first_cpu
= cpumask_first(cpu_online_mask
);
555 * We take down all of the non-boot CPUs in one shot to avoid races
556 * with the userspace trying to use the CPU hotplug at the same time
558 cpumask_clear(frozen_cpus
);
560 pr_info("Disabling non-boot CPUs ...\n");
561 for_each_online_cpu(cpu
) {
562 if (cpu
== first_cpu
)
564 trace_suspend_resume(TPS("CPU_OFF"), cpu
, true);
565 error
= _cpu_down(cpu
, 1);
566 trace_suspend_resume(TPS("CPU_OFF"), cpu
, false);
568 cpumask_set_cpu(cpu
, frozen_cpus
);
570 pr_err("Error taking CPU%d down: %d\n", cpu
, error
);
576 BUG_ON(num_online_cpus() > 1);
577 /* Make sure the CPUs won't be enabled by someone else */
578 cpu_hotplug_disabled
= 1;
580 pr_err("Non-boot CPUs are not disabled\n");
582 cpu_maps_update_done();
586 void __weak
arch_enable_nonboot_cpus_begin(void)
590 void __weak
arch_enable_nonboot_cpus_end(void)
594 void __ref
enable_nonboot_cpus(void)
598 /* Allow everyone to use the CPU hotplug again */
599 cpu_maps_update_begin();
600 cpu_hotplug_disabled
= 0;
601 if (cpumask_empty(frozen_cpus
))
604 pr_info("Enabling non-boot CPUs ...\n");
606 arch_enable_nonboot_cpus_begin();
608 for_each_cpu(cpu
, frozen_cpus
) {
609 trace_suspend_resume(TPS("CPU_ON"), cpu
, true);
610 error
= _cpu_up(cpu
, 1);
611 trace_suspend_resume(TPS("CPU_ON"), cpu
, false);
613 pr_info("CPU%d is up\n", cpu
);
616 pr_warn("Error taking CPU%d up: %d\n", cpu
, error
);
619 arch_enable_nonboot_cpus_end();
621 cpumask_clear(frozen_cpus
);
623 cpu_maps_update_done();
626 static int __init
alloc_frozen_cpus(void)
628 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
632 core_initcall(alloc_frozen_cpus
);
635 * When callbacks for CPU hotplug notifications are being executed, we must
636 * ensure that the state of the system with respect to the tasks being frozen
637 * or not, as reported by the notification, remains unchanged *throughout the
638 * duration* of the execution of the callbacks.
639 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
641 * This synchronization is implemented by mutually excluding regular CPU
642 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
643 * Hibernate notifications.
646 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
647 unsigned long action
, void *ptr
)
651 case PM_SUSPEND_PREPARE
:
652 case PM_HIBERNATION_PREPARE
:
653 cpu_hotplug_disable();
656 case PM_POST_SUSPEND
:
657 case PM_POST_HIBERNATION
:
658 cpu_hotplug_enable();
669 static int __init
cpu_hotplug_pm_sync_init(void)
672 * cpu_hotplug_pm_callback has higher priority than x86
673 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
674 * to disable cpu hotplug to avoid cpu hotplug race.
676 pm_notifier(cpu_hotplug_pm_callback
, 0);
679 core_initcall(cpu_hotplug_pm_sync_init
);
681 #endif /* CONFIG_PM_SLEEP_SMP */
684 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
685 * @cpu: cpu that just started
687 * This function calls the cpu_chain notifiers with CPU_STARTING.
688 * It must be called by the arch code on the new cpu, before the new cpu
689 * enables interrupts and before the "boot" cpu returns from __cpu_up().
691 void notify_cpu_starting(unsigned int cpu
)
693 unsigned long val
= CPU_STARTING
;
695 #ifdef CONFIG_PM_SLEEP_SMP
696 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
697 val
= CPU_STARTING_FROZEN
;
698 #endif /* CONFIG_PM_SLEEP_SMP */
699 cpu_notify(val
, (void *)(long)cpu
);
702 #endif /* CONFIG_SMP */
705 * cpu_bit_bitmap[] is a special, "compressed" data structure that
706 * represents all NR_CPUS bits binary values of 1<<nr.
708 * It is used by cpumask_of() to get a constant address to a CPU
709 * mask value that has a single bit set only.
712 /* cpu_bit_bitmap[0] is empty - so we can back into it */
713 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
714 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
715 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
716 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
718 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
720 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
721 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
722 #if BITS_PER_LONG > 32
723 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
724 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
727 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
729 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
730 EXPORT_SYMBOL(cpu_all_bits
);
732 #ifdef CONFIG_INIT_ALL_POSSIBLE
733 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
736 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
738 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
739 EXPORT_SYMBOL(cpu_possible_mask
);
741 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
742 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
743 EXPORT_SYMBOL(cpu_online_mask
);
745 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
746 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
747 EXPORT_SYMBOL(cpu_present_mask
);
749 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
750 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
751 EXPORT_SYMBOL(cpu_active_mask
);
753 void set_cpu_possible(unsigned int cpu
, bool possible
)
756 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
758 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
761 void set_cpu_present(unsigned int cpu
, bool present
)
764 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
766 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
769 void set_cpu_online(unsigned int cpu
, bool online
)
772 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
773 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
775 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
779 void set_cpu_active(unsigned int cpu
, bool active
)
782 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
784 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
787 void init_cpu_present(const struct cpumask
*src
)
789 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
792 void init_cpu_possible(const struct cpumask
*src
)
794 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
797 void init_cpu_online(const struct cpumask
*src
)
799 cpumask_copy(to_cpumask(cpu_online_bits
), src
);