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.
68 #ifdef CONFIG_DEBUG_LOCK_ALLOC
69 struct lockdep_map dep_map
;
72 .active_writer
= NULL
,
73 .lock
= __MUTEX_INITIALIZER(cpu_hotplug
.lock
),
75 #ifdef CONFIG_DEBUG_LOCK_ALLOC
76 .dep_map
= {.name
= "cpu_hotplug.lock" },
80 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
81 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
82 #define cpuhp_lock_acquire_tryread() \
83 lock_map_acquire_tryread(&cpu_hotplug.dep_map)
84 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
85 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
87 void get_online_cpus(void)
90 if (cpu_hotplug
.active_writer
== current
)
92 cpuhp_lock_acquire_read();
93 mutex_lock(&cpu_hotplug
.lock
);
94 cpu_hotplug
.refcount
++;
95 mutex_unlock(&cpu_hotplug
.lock
);
97 EXPORT_SYMBOL_GPL(get_online_cpus
);
99 bool try_get_online_cpus(void)
101 if (cpu_hotplug
.active_writer
== current
)
103 if (!mutex_trylock(&cpu_hotplug
.lock
))
105 cpuhp_lock_acquire_tryread();
106 cpu_hotplug
.refcount
++;
107 mutex_unlock(&cpu_hotplug
.lock
);
110 EXPORT_SYMBOL_GPL(try_get_online_cpus
);
112 void put_online_cpus(void)
114 if (cpu_hotplug
.active_writer
== current
)
116 mutex_lock(&cpu_hotplug
.lock
);
118 if (WARN_ON(!cpu_hotplug
.refcount
))
119 cpu_hotplug
.refcount
++; /* try to fix things up */
121 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
122 wake_up_process(cpu_hotplug
.active_writer
);
123 mutex_unlock(&cpu_hotplug
.lock
);
124 cpuhp_lock_release();
127 EXPORT_SYMBOL_GPL(put_online_cpus
);
130 * This ensures that the hotplug operation can begin only when the
131 * refcount goes to zero.
133 * Note that during a cpu-hotplug operation, the new readers, if any,
134 * will be blocked by the cpu_hotplug.lock
136 * Since cpu_hotplug_begin() is always called after invoking
137 * cpu_maps_update_begin(), we can be sure that only one writer is active.
139 * Note that theoretically, there is a possibility of a livelock:
140 * - Refcount goes to zero, last reader wakes up the sleeping
142 * - Last reader unlocks the cpu_hotplug.lock.
143 * - A new reader arrives at this moment, bumps up the refcount.
144 * - The writer acquires the cpu_hotplug.lock finds the refcount
145 * non zero and goes to sleep again.
147 * However, this is very difficult to achieve in practice since
148 * get_online_cpus() not an api which is called all that often.
151 void cpu_hotplug_begin(void)
153 cpu_hotplug
.active_writer
= current
;
155 cpuhp_lock_acquire();
157 mutex_lock(&cpu_hotplug
.lock
);
158 if (likely(!cpu_hotplug
.refcount
))
160 __set_current_state(TASK_UNINTERRUPTIBLE
);
161 mutex_unlock(&cpu_hotplug
.lock
);
166 void cpu_hotplug_done(void)
168 cpu_hotplug
.active_writer
= NULL
;
169 mutex_unlock(&cpu_hotplug
.lock
);
170 cpuhp_lock_release();
174 * Wait for currently running CPU hotplug operations to complete (if any) and
175 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
176 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
177 * hotplug path before performing hotplug operations. So acquiring that lock
178 * guarantees mutual exclusion from any currently running hotplug operations.
180 void cpu_hotplug_disable(void)
182 cpu_maps_update_begin();
183 cpu_hotplug_disabled
= 1;
184 cpu_maps_update_done();
187 void cpu_hotplug_enable(void)
189 cpu_maps_update_begin();
190 cpu_hotplug_disabled
= 0;
191 cpu_maps_update_done();
194 #endif /* CONFIG_HOTPLUG_CPU */
196 /* Need to know about CPUs going up/down? */
197 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
200 cpu_maps_update_begin();
201 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
202 cpu_maps_update_done();
206 int __ref
__register_cpu_notifier(struct notifier_block
*nb
)
208 return raw_notifier_chain_register(&cpu_chain
, nb
);
211 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
216 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
219 return notifier_to_errno(ret
);
222 static int cpu_notify(unsigned long val
, void *v
)
224 return __cpu_notify(val
, v
, -1, NULL
);
227 #ifdef CONFIG_HOTPLUG_CPU
229 static void cpu_notify_nofail(unsigned long val
, void *v
)
231 BUG_ON(cpu_notify(val
, v
));
233 EXPORT_SYMBOL(register_cpu_notifier
);
234 EXPORT_SYMBOL(__register_cpu_notifier
);
236 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
238 cpu_maps_update_begin();
239 raw_notifier_chain_unregister(&cpu_chain
, nb
);
240 cpu_maps_update_done();
242 EXPORT_SYMBOL(unregister_cpu_notifier
);
244 void __ref
__unregister_cpu_notifier(struct notifier_block
*nb
)
246 raw_notifier_chain_unregister(&cpu_chain
, nb
);
248 EXPORT_SYMBOL(__unregister_cpu_notifier
);
251 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
254 * This function walks all processes, finds a valid mm struct for each one and
255 * then clears a corresponding bit in mm's cpumask. While this all sounds
256 * trivial, there are various non-obvious corner cases, which this function
257 * tries to solve in a safe manner.
259 * Also note that the function uses a somewhat relaxed locking scheme, so it may
260 * be called only for an already offlined CPU.
262 void clear_tasks_mm_cpumask(int cpu
)
264 struct task_struct
*p
;
267 * This function is called after the cpu is taken down and marked
268 * offline, so its not like new tasks will ever get this cpu set in
269 * their mm mask. -- Peter Zijlstra
270 * Thus, we may use rcu_read_lock() here, instead of grabbing
271 * full-fledged tasklist_lock.
273 WARN_ON(cpu_online(cpu
));
275 for_each_process(p
) {
276 struct task_struct
*t
;
279 * Main thread might exit, but other threads may still have
280 * a valid mm. Find one.
282 t
= find_lock_task_mm(p
);
285 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
291 static inline void check_for_tasks(int dead_cpu
)
293 struct task_struct
*g
, *p
;
295 read_lock_irq(&tasklist_lock
);
296 do_each_thread(g
, p
) {
300 * We do the check with unlocked task_rq(p)->lock.
301 * Order the reading to do not warn about a task,
302 * which was running on this cpu in the past, and
303 * it's just been woken on another cpu.
306 if (task_cpu(p
) != dead_cpu
)
309 pr_warn("Task %s (pid=%d) is on cpu %d (state=%ld, flags=%x)\n",
310 p
->comm
, task_pid_nr(p
), dead_cpu
, p
->state
, p
->flags
);
311 } while_each_thread(g
, p
);
312 read_unlock_irq(&tasklist_lock
);
315 struct take_cpu_down_param
{
320 /* Take this CPU down. */
321 static int __ref
take_cpu_down(void *_param
)
323 struct take_cpu_down_param
*param
= _param
;
326 /* Ensure this CPU doesn't handle any more interrupts. */
327 err
= __cpu_disable();
331 cpu_notify(CPU_DYING
| param
->mod
, param
->hcpu
);
332 /* Park the stopper thread */
333 kthread_park(current
);
337 /* Requires cpu_add_remove_lock to be held */
338 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
340 int err
, nr_calls
= 0;
341 void *hcpu
= (void *)(long)cpu
;
342 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
343 struct take_cpu_down_param tcd_param
= {
348 if (num_online_cpus() == 1)
351 if (!cpu_online(cpu
))
356 err
= __cpu_notify(CPU_DOWN_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
359 __cpu_notify(CPU_DOWN_FAILED
| mod
, hcpu
, nr_calls
, NULL
);
360 pr_warn("%s: attempt to take down CPU %u failed\n",
366 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
367 * and RCU users of this state to go away such that all new such users
370 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
371 * not imply sync_sched(), so explicitly call both.
373 * Do sync before park smpboot threads to take care the rcu boost case.
375 #ifdef CONFIG_PREEMPT
380 smpboot_park_threads(cpu
);
383 * So now all preempt/rcu users must observe !cpu_active().
386 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
388 /* CPU didn't die: tell everyone. Can't complain. */
389 smpboot_unpark_threads(cpu
);
390 cpu_notify_nofail(CPU_DOWN_FAILED
| mod
, hcpu
);
393 BUG_ON(cpu_online(cpu
));
396 * The migration_call() CPU_DYING callback will have removed all
397 * runnable tasks from the cpu, there's only the idle task left now
398 * that the migration thread is done doing the stop_machine thing.
400 * Wait for the stop thread to go away.
402 while (!idle_cpu(cpu
))
405 /* This actually kills the CPU. */
408 /* CPU is completely dead: tell everyone. Too late to complain. */
409 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
411 check_for_tasks(cpu
);
416 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
420 int __ref
cpu_down(unsigned int cpu
)
424 cpu_maps_update_begin();
426 if (cpu_hotplug_disabled
) {
431 err
= _cpu_down(cpu
, 0);
434 cpu_maps_update_done();
437 EXPORT_SYMBOL(cpu_down
);
438 #endif /*CONFIG_HOTPLUG_CPU*/
440 /* Requires cpu_add_remove_lock to be held */
441 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
443 int ret
, nr_calls
= 0;
444 void *hcpu
= (void *)(long)cpu
;
445 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
446 struct task_struct
*idle
;
450 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
455 idle
= idle_thread_get(cpu
);
461 ret
= smpboot_create_threads(cpu
);
465 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
468 pr_warn("%s: attempt to bring up CPU %u failed\n",
473 /* Arch-specific enabling code. */
474 ret
= __cpu_up(cpu
, idle
);
477 BUG_ON(!cpu_online(cpu
));
479 /* Wake the per cpu threads */
480 smpboot_unpark_threads(cpu
);
482 /* Now call notifier in preparation. */
483 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
487 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
494 int cpu_up(unsigned int cpu
)
498 if (!cpu_possible(cpu
)) {
499 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
501 #if defined(CONFIG_IA64)
502 pr_err("please check additional_cpus= boot parameter\n");
507 err
= try_online_node(cpu_to_node(cpu
));
511 cpu_maps_update_begin();
513 if (cpu_hotplug_disabled
) {
518 err
= _cpu_up(cpu
, 0);
521 cpu_maps_update_done();
524 EXPORT_SYMBOL_GPL(cpu_up
);
526 #ifdef CONFIG_PM_SLEEP_SMP
527 static cpumask_var_t frozen_cpus
;
529 int disable_nonboot_cpus(void)
531 int cpu
, first_cpu
, error
= 0;
533 cpu_maps_update_begin();
534 first_cpu
= cpumask_first(cpu_online_mask
);
536 * We take down all of the non-boot CPUs in one shot to avoid races
537 * with the userspace trying to use the CPU hotplug at the same time
539 cpumask_clear(frozen_cpus
);
541 pr_info("Disabling non-boot CPUs ...\n");
542 for_each_online_cpu(cpu
) {
543 if (cpu
== first_cpu
)
545 trace_suspend_resume(TPS("CPU_OFF"), cpu
, true);
546 error
= _cpu_down(cpu
, 1);
547 trace_suspend_resume(TPS("CPU_OFF"), cpu
, false);
549 cpumask_set_cpu(cpu
, frozen_cpus
);
551 pr_err("Error taking CPU%d down: %d\n", cpu
, error
);
557 BUG_ON(num_online_cpus() > 1);
558 /* Make sure the CPUs won't be enabled by someone else */
559 cpu_hotplug_disabled
= 1;
561 pr_err("Non-boot CPUs are not disabled\n");
563 cpu_maps_update_done();
567 void __weak
arch_enable_nonboot_cpus_begin(void)
571 void __weak
arch_enable_nonboot_cpus_end(void)
575 void __ref
enable_nonboot_cpus(void)
579 /* Allow everyone to use the CPU hotplug again */
580 cpu_maps_update_begin();
581 cpu_hotplug_disabled
= 0;
582 if (cpumask_empty(frozen_cpus
))
585 pr_info("Enabling non-boot CPUs ...\n");
587 arch_enable_nonboot_cpus_begin();
589 for_each_cpu(cpu
, frozen_cpus
) {
590 trace_suspend_resume(TPS("CPU_ON"), cpu
, true);
591 error
= _cpu_up(cpu
, 1);
592 trace_suspend_resume(TPS("CPU_ON"), cpu
, false);
594 pr_info("CPU%d is up\n", cpu
);
597 pr_warn("Error taking CPU%d up: %d\n", cpu
, error
);
600 arch_enable_nonboot_cpus_end();
602 cpumask_clear(frozen_cpus
);
604 cpu_maps_update_done();
607 static int __init
alloc_frozen_cpus(void)
609 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
613 core_initcall(alloc_frozen_cpus
);
616 * When callbacks for CPU hotplug notifications are being executed, we must
617 * ensure that the state of the system with respect to the tasks being frozen
618 * or not, as reported by the notification, remains unchanged *throughout the
619 * duration* of the execution of the callbacks.
620 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
622 * This synchronization is implemented by mutually excluding regular CPU
623 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
624 * Hibernate notifications.
627 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
628 unsigned long action
, void *ptr
)
632 case PM_SUSPEND_PREPARE
:
633 case PM_HIBERNATION_PREPARE
:
634 cpu_hotplug_disable();
637 case PM_POST_SUSPEND
:
638 case PM_POST_HIBERNATION
:
639 cpu_hotplug_enable();
650 static int __init
cpu_hotplug_pm_sync_init(void)
653 * cpu_hotplug_pm_callback has higher priority than x86
654 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
655 * to disable cpu hotplug to avoid cpu hotplug race.
657 pm_notifier(cpu_hotplug_pm_callback
, 0);
660 core_initcall(cpu_hotplug_pm_sync_init
);
662 #endif /* CONFIG_PM_SLEEP_SMP */
665 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
666 * @cpu: cpu that just started
668 * This function calls the cpu_chain notifiers with CPU_STARTING.
669 * It must be called by the arch code on the new cpu, before the new cpu
670 * enables interrupts and before the "boot" cpu returns from __cpu_up().
672 void notify_cpu_starting(unsigned int cpu
)
674 unsigned long val
= CPU_STARTING
;
676 #ifdef CONFIG_PM_SLEEP_SMP
677 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
678 val
= CPU_STARTING_FROZEN
;
679 #endif /* CONFIG_PM_SLEEP_SMP */
680 cpu_notify(val
, (void *)(long)cpu
);
683 #endif /* CONFIG_SMP */
686 * cpu_bit_bitmap[] is a special, "compressed" data structure that
687 * represents all NR_CPUS bits binary values of 1<<nr.
689 * It is used by cpumask_of() to get a constant address to a CPU
690 * mask value that has a single bit set only.
693 /* cpu_bit_bitmap[0] is empty - so we can back into it */
694 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
695 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
696 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
697 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
699 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
701 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
702 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
703 #if BITS_PER_LONG > 32
704 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
705 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
708 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
710 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
711 EXPORT_SYMBOL(cpu_all_bits
);
713 #ifdef CONFIG_INIT_ALL_POSSIBLE
714 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
717 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
719 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
720 EXPORT_SYMBOL(cpu_possible_mask
);
722 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
723 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
724 EXPORT_SYMBOL(cpu_online_mask
);
726 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
727 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
728 EXPORT_SYMBOL(cpu_present_mask
);
730 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
731 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
732 EXPORT_SYMBOL(cpu_active_mask
);
734 void set_cpu_possible(unsigned int cpu
, bool possible
)
737 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
739 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
742 void set_cpu_present(unsigned int cpu
, bool present
)
745 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
747 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
750 void set_cpu_online(unsigned int cpu
, bool online
)
753 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
754 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
756 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
760 void set_cpu_active(unsigned int cpu
, bool active
)
763 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
765 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
768 void init_cpu_present(const struct cpumask
*src
)
770 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
773 void init_cpu_possible(const struct cpumask
*src
)
775 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
778 void init_cpu_online(const struct cpumask
*src
)
780 cpumask_copy(to_cpumask(cpu_online_bits
), src
);