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() lock_map_acquire(&cpu_hotplug.dep_map)
83 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
85 void get_online_cpus(void)
88 if (cpu_hotplug
.active_writer
== current
)
90 cpuhp_lock_acquire_read();
91 mutex_lock(&cpu_hotplug
.lock
);
92 cpu_hotplug
.refcount
++;
93 mutex_unlock(&cpu_hotplug
.lock
);
96 EXPORT_SYMBOL_GPL(get_online_cpus
);
98 void put_online_cpus(void)
100 if (cpu_hotplug
.active_writer
== current
)
102 mutex_lock(&cpu_hotplug
.lock
);
104 if (WARN_ON(!cpu_hotplug
.refcount
))
105 cpu_hotplug
.refcount
++; /* try to fix things up */
107 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
108 wake_up_process(cpu_hotplug
.active_writer
);
109 mutex_unlock(&cpu_hotplug
.lock
);
110 cpuhp_lock_release();
113 EXPORT_SYMBOL_GPL(put_online_cpus
);
116 * This ensures that the hotplug operation can begin only when the
117 * refcount goes to zero.
119 * Note that during a cpu-hotplug operation, the new readers, if any,
120 * will be blocked by the cpu_hotplug.lock
122 * Since cpu_hotplug_begin() is always called after invoking
123 * cpu_maps_update_begin(), we can be sure that only one writer is active.
125 * Note that theoretically, there is a possibility of a livelock:
126 * - Refcount goes to zero, last reader wakes up the sleeping
128 * - Last reader unlocks the cpu_hotplug.lock.
129 * - A new reader arrives at this moment, bumps up the refcount.
130 * - The writer acquires the cpu_hotplug.lock finds the refcount
131 * non zero and goes to sleep again.
133 * However, this is very difficult to achieve in practice since
134 * get_online_cpus() not an api which is called all that often.
137 void cpu_hotplug_begin(void)
139 cpu_hotplug
.active_writer
= current
;
141 cpuhp_lock_acquire();
143 mutex_lock(&cpu_hotplug
.lock
);
144 if (likely(!cpu_hotplug
.refcount
))
146 __set_current_state(TASK_UNINTERRUPTIBLE
);
147 mutex_unlock(&cpu_hotplug
.lock
);
152 void cpu_hotplug_done(void)
154 cpu_hotplug
.active_writer
= NULL
;
155 mutex_unlock(&cpu_hotplug
.lock
);
156 cpuhp_lock_release();
160 * Wait for currently running CPU hotplug operations to complete (if any) and
161 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
162 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
163 * hotplug path before performing hotplug operations. So acquiring that lock
164 * guarantees mutual exclusion from any currently running hotplug operations.
166 void cpu_hotplug_disable(void)
168 cpu_maps_update_begin();
169 cpu_hotplug_disabled
= 1;
170 cpu_maps_update_done();
173 void cpu_hotplug_enable(void)
175 cpu_maps_update_begin();
176 cpu_hotplug_disabled
= 0;
177 cpu_maps_update_done();
180 #endif /* CONFIG_HOTPLUG_CPU */
182 /* Need to know about CPUs going up/down? */
183 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
186 cpu_maps_update_begin();
187 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
188 cpu_maps_update_done();
192 int __ref
__register_cpu_notifier(struct notifier_block
*nb
)
194 return raw_notifier_chain_register(&cpu_chain
, nb
);
197 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
202 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
205 return notifier_to_errno(ret
);
208 static int cpu_notify(unsigned long val
, void *v
)
210 return __cpu_notify(val
, v
, -1, NULL
);
213 #ifdef CONFIG_HOTPLUG_CPU
215 static void cpu_notify_nofail(unsigned long val
, void *v
)
217 BUG_ON(cpu_notify(val
, v
));
219 EXPORT_SYMBOL(register_cpu_notifier
);
220 EXPORT_SYMBOL(__register_cpu_notifier
);
222 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
224 cpu_maps_update_begin();
225 raw_notifier_chain_unregister(&cpu_chain
, nb
);
226 cpu_maps_update_done();
228 EXPORT_SYMBOL(unregister_cpu_notifier
);
230 void __ref
__unregister_cpu_notifier(struct notifier_block
*nb
)
232 raw_notifier_chain_unregister(&cpu_chain
, nb
);
234 EXPORT_SYMBOL(__unregister_cpu_notifier
);
237 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
240 * This function walks all processes, finds a valid mm struct for each one and
241 * then clears a corresponding bit in mm's cpumask. While this all sounds
242 * trivial, there are various non-obvious corner cases, which this function
243 * tries to solve in a safe manner.
245 * Also note that the function uses a somewhat relaxed locking scheme, so it may
246 * be called only for an already offlined CPU.
248 void clear_tasks_mm_cpumask(int cpu
)
250 struct task_struct
*p
;
253 * This function is called after the cpu is taken down and marked
254 * offline, so its not like new tasks will ever get this cpu set in
255 * their mm mask. -- Peter Zijlstra
256 * Thus, we may use rcu_read_lock() here, instead of grabbing
257 * full-fledged tasklist_lock.
259 WARN_ON(cpu_online(cpu
));
261 for_each_process(p
) {
262 struct task_struct
*t
;
265 * Main thread might exit, but other threads may still have
266 * a valid mm. Find one.
268 t
= find_lock_task_mm(p
);
271 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
277 static inline void check_for_tasks(int cpu
)
279 struct task_struct
*p
;
280 cputime_t utime
, stime
;
282 write_lock_irq(&tasklist_lock
);
283 for_each_process(p
) {
284 task_cputime(p
, &utime
, &stime
);
285 if (task_cpu(p
) == cpu
&& p
->state
== TASK_RUNNING
&&
287 pr_warn("Task %s (pid = %d) is on cpu %d (state = %ld, flags = %x)\n",
288 p
->comm
, task_pid_nr(p
), cpu
,
291 write_unlock_irq(&tasklist_lock
);
294 struct take_cpu_down_param
{
299 /* Take this CPU down. */
300 static int __ref
take_cpu_down(void *_param
)
302 struct take_cpu_down_param
*param
= _param
;
305 /* Ensure this CPU doesn't handle any more interrupts. */
306 err
= __cpu_disable();
310 cpu_notify(CPU_DYING
| param
->mod
, param
->hcpu
);
311 /* Park the stopper thread */
312 kthread_park(current
);
316 /* Requires cpu_add_remove_lock to be held */
317 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
319 int err
, nr_calls
= 0;
320 void *hcpu
= (void *)(long)cpu
;
321 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
322 struct take_cpu_down_param tcd_param
= {
327 if (num_online_cpus() == 1)
330 if (!cpu_online(cpu
))
335 err
= __cpu_notify(CPU_DOWN_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
338 __cpu_notify(CPU_DOWN_FAILED
| mod
, hcpu
, nr_calls
, NULL
);
339 pr_warn("%s: attempt to take down CPU %u failed\n",
345 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
346 * and RCU users of this state to go away such that all new such users
349 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
350 * not imply sync_sched(), so explicitly call both.
352 * Do sync before park smpboot threads to take care the rcu boost case.
354 #ifdef CONFIG_PREEMPT
359 smpboot_park_threads(cpu
);
362 * So now all preempt/rcu users must observe !cpu_active().
365 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
367 /* CPU didn't die: tell everyone. Can't complain. */
368 smpboot_unpark_threads(cpu
);
369 cpu_notify_nofail(CPU_DOWN_FAILED
| mod
, hcpu
);
372 BUG_ON(cpu_online(cpu
));
375 * The migration_call() CPU_DYING callback will have removed all
376 * runnable tasks from the cpu, there's only the idle task left now
377 * that the migration thread is done doing the stop_machine thing.
379 * Wait for the stop thread to go away.
381 while (!idle_cpu(cpu
))
384 /* This actually kills the CPU. */
387 /* CPU is completely dead: tell everyone. Too late to complain. */
388 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
390 check_for_tasks(cpu
);
395 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
399 int __ref
cpu_down(unsigned int cpu
)
403 cpu_maps_update_begin();
405 if (cpu_hotplug_disabled
) {
410 err
= _cpu_down(cpu
, 0);
413 cpu_maps_update_done();
416 EXPORT_SYMBOL(cpu_down
);
417 #endif /*CONFIG_HOTPLUG_CPU*/
420 * Unpark per-CPU smpboot kthreads at CPU-online time.
422 static int smpboot_thread_call(struct notifier_block
*nfb
,
423 unsigned long action
, void *hcpu
)
425 int cpu
= (long)hcpu
;
427 switch (action
& ~CPU_TASKS_FROZEN
) {
430 smpboot_unpark_threads(cpu
);
440 static struct notifier_block smpboot_thread_notifier
= {
441 .notifier_call
= smpboot_thread_call
,
442 .priority
= CPU_PRI_SMPBOOT
,
445 void __cpuinit
smpboot_thread_init(void)
447 register_cpu_notifier(&smpboot_thread_notifier
);
450 /* Requires cpu_add_remove_lock to be held */
451 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
453 int ret
, nr_calls
= 0;
454 void *hcpu
= (void *)(long)cpu
;
455 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
456 struct task_struct
*idle
;
460 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
465 idle
= idle_thread_get(cpu
);
471 ret
= smpboot_create_threads(cpu
);
475 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
478 pr_warn("%s: attempt to bring up CPU %u failed\n",
483 /* Arch-specific enabling code. */
484 ret
= __cpu_up(cpu
, idle
);
487 BUG_ON(!cpu_online(cpu
));
489 /* Now call notifier in preparation. */
490 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
494 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
501 int cpu_up(unsigned int cpu
)
505 if (!cpu_possible(cpu
)) {
506 pr_err("can't online cpu %d because it is not configured as may-hotadd at boot time\n",
508 #if defined(CONFIG_IA64)
509 pr_err("please check additional_cpus= boot parameter\n");
514 err
= try_online_node(cpu_to_node(cpu
));
518 cpu_maps_update_begin();
520 if (cpu_hotplug_disabled
) {
525 err
= _cpu_up(cpu
, 0);
528 cpu_maps_update_done();
531 EXPORT_SYMBOL_GPL(cpu_up
);
533 #ifdef CONFIG_PM_SLEEP_SMP
534 static cpumask_var_t frozen_cpus
;
536 int disable_nonboot_cpus(void)
538 int cpu
, first_cpu
, error
= 0;
540 cpu_maps_update_begin();
541 first_cpu
= cpumask_first(cpu_online_mask
);
543 * We take down all of the non-boot CPUs in one shot to avoid races
544 * with the userspace trying to use the CPU hotplug at the same time
546 cpumask_clear(frozen_cpus
);
548 pr_info("Disabling non-boot CPUs ...\n");
549 for_each_online_cpu(cpu
) {
550 if (cpu
== first_cpu
)
552 trace_suspend_resume(TPS("CPU_OFF"), cpu
, true);
553 error
= _cpu_down(cpu
, 1);
554 trace_suspend_resume(TPS("CPU_OFF"), cpu
, false);
556 cpumask_set_cpu(cpu
, frozen_cpus
);
558 pr_err("Error taking CPU%d down: %d\n", cpu
, error
);
564 BUG_ON(num_online_cpus() > 1);
565 /* Make sure the CPUs won't be enabled by someone else */
566 cpu_hotplug_disabled
= 1;
568 pr_err("Non-boot CPUs are not disabled\n");
570 cpu_maps_update_done();
574 void __weak
arch_enable_nonboot_cpus_begin(void)
578 void __weak
arch_enable_nonboot_cpus_end(void)
582 void __ref
enable_nonboot_cpus(void)
586 /* Allow everyone to use the CPU hotplug again */
587 cpu_maps_update_begin();
588 cpu_hotplug_disabled
= 0;
589 if (cpumask_empty(frozen_cpus
))
592 pr_info("Enabling non-boot CPUs ...\n");
594 arch_enable_nonboot_cpus_begin();
596 for_each_cpu(cpu
, frozen_cpus
) {
597 trace_suspend_resume(TPS("CPU_ON"), cpu
, true);
598 error
= _cpu_up(cpu
, 1);
599 trace_suspend_resume(TPS("CPU_ON"), cpu
, false);
601 pr_info("CPU%d is up\n", cpu
);
604 pr_warn("Error taking CPU%d up: %d\n", cpu
, error
);
607 arch_enable_nonboot_cpus_end();
609 cpumask_clear(frozen_cpus
);
611 cpu_maps_update_done();
614 static int __init
alloc_frozen_cpus(void)
616 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
620 core_initcall(alloc_frozen_cpus
);
623 * When callbacks for CPU hotplug notifications are being executed, we must
624 * ensure that the state of the system with respect to the tasks being frozen
625 * or not, as reported by the notification, remains unchanged *throughout the
626 * duration* of the execution of the callbacks.
627 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
629 * This synchronization is implemented by mutually excluding regular CPU
630 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
631 * Hibernate notifications.
634 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
635 unsigned long action
, void *ptr
)
639 case PM_SUSPEND_PREPARE
:
640 case PM_HIBERNATION_PREPARE
:
641 cpu_hotplug_disable();
644 case PM_POST_SUSPEND
:
645 case PM_POST_HIBERNATION
:
646 cpu_hotplug_enable();
657 static int __init
cpu_hotplug_pm_sync_init(void)
660 * cpu_hotplug_pm_callback has higher priority than x86
661 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
662 * to disable cpu hotplug to avoid cpu hotplug race.
664 pm_notifier(cpu_hotplug_pm_callback
, 0);
667 core_initcall(cpu_hotplug_pm_sync_init
);
669 #endif /* CONFIG_PM_SLEEP_SMP */
672 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
673 * @cpu: cpu that just started
675 * This function calls the cpu_chain notifiers with CPU_STARTING.
676 * It must be called by the arch code on the new cpu, before the new cpu
677 * enables interrupts and before the "boot" cpu returns from __cpu_up().
679 void notify_cpu_starting(unsigned int cpu
)
681 unsigned long val
= CPU_STARTING
;
683 #ifdef CONFIG_PM_SLEEP_SMP
684 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
685 val
= CPU_STARTING_FROZEN
;
686 #endif /* CONFIG_PM_SLEEP_SMP */
687 cpu_notify(val
, (void *)(long)cpu
);
690 #endif /* CONFIG_SMP */
693 * cpu_bit_bitmap[] is a special, "compressed" data structure that
694 * represents all NR_CPUS bits binary values of 1<<nr.
696 * It is used by cpumask_of() to get a constant address to a CPU
697 * mask value that has a single bit set only.
700 /* cpu_bit_bitmap[0] is empty - so we can back into it */
701 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
702 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
703 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
704 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
706 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
708 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
709 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
710 #if BITS_PER_LONG > 32
711 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
712 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
715 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
717 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
718 EXPORT_SYMBOL(cpu_all_bits
);
720 #ifdef CONFIG_INIT_ALL_POSSIBLE
721 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
724 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
726 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
727 EXPORT_SYMBOL(cpu_possible_mask
);
729 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
730 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
731 EXPORT_SYMBOL(cpu_online_mask
);
733 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
734 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
735 EXPORT_SYMBOL(cpu_present_mask
);
737 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
738 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
739 EXPORT_SYMBOL(cpu_active_mask
);
741 void set_cpu_possible(unsigned int cpu
, bool possible
)
744 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
746 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
749 void set_cpu_present(unsigned int cpu
, bool present
)
752 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
754 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
757 void set_cpu_online(unsigned int cpu
, bool online
)
760 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
761 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
763 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
767 void set_cpu_active(unsigned int cpu
, bool active
)
770 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
772 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
775 void init_cpu_present(const struct cpumask
*src
)
777 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
780 void init_cpu_possible(const struct cpumask
*src
)
782 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
785 void init_cpu_online(const struct cpumask
*src
)
787 cpumask_copy(to_cpumask(cpu_online_bits
), src
);