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>
27 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
28 static DEFINE_MUTEX(cpu_add_remove_lock
);
31 * The following two APIs (cpu_maps_update_begin/done) must be used when
32 * attempting to serialize the updates to cpu_online_mask & cpu_present_mask.
33 * The APIs cpu_notifier_register_begin/done() must be used to protect CPU
34 * hotplug callback (un)registration performed using __register_cpu_notifier()
35 * or __unregister_cpu_notifier().
37 void cpu_maps_update_begin(void)
39 mutex_lock(&cpu_add_remove_lock
);
41 EXPORT_SYMBOL(cpu_notifier_register_begin
);
43 void cpu_maps_update_done(void)
45 mutex_unlock(&cpu_add_remove_lock
);
47 EXPORT_SYMBOL(cpu_notifier_register_done
);
49 static RAW_NOTIFIER_HEAD(cpu_chain
);
51 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
52 * Should always be manipulated under cpu_add_remove_lock
54 static int cpu_hotplug_disabled
;
56 #ifdef CONFIG_HOTPLUG_CPU
59 struct task_struct
*active_writer
;
60 struct mutex lock
; /* Synchronizes accesses to refcount, */
62 * Also blocks the new readers during
63 * an ongoing cpu hotplug operation.
67 #ifdef CONFIG_DEBUG_LOCK_ALLOC
68 struct lockdep_map dep_map
;
71 .active_writer
= NULL
,
72 .lock
= __MUTEX_INITIALIZER(cpu_hotplug
.lock
),
74 #ifdef CONFIG_DEBUG_LOCK_ALLOC
75 .dep_map
= {.name
= "cpu_hotplug.lock" },
79 /* Lockdep annotations for get/put_online_cpus() and cpu_hotplug_begin/end() */
80 #define cpuhp_lock_acquire_read() lock_map_acquire_read(&cpu_hotplug.dep_map)
81 #define cpuhp_lock_acquire() lock_map_acquire(&cpu_hotplug.dep_map)
82 #define cpuhp_lock_release() lock_map_release(&cpu_hotplug.dep_map)
84 void get_online_cpus(void)
87 if (cpu_hotplug
.active_writer
== current
)
89 cpuhp_lock_acquire_read();
90 mutex_lock(&cpu_hotplug
.lock
);
91 cpu_hotplug
.refcount
++;
92 mutex_unlock(&cpu_hotplug
.lock
);
95 EXPORT_SYMBOL_GPL(get_online_cpus
);
97 void put_online_cpus(void)
99 if (cpu_hotplug
.active_writer
== current
)
101 mutex_lock(&cpu_hotplug
.lock
);
103 if (WARN_ON(!cpu_hotplug
.refcount
))
104 cpu_hotplug
.refcount
++; /* try to fix things up */
106 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
107 wake_up_process(cpu_hotplug
.active_writer
);
108 mutex_unlock(&cpu_hotplug
.lock
);
109 cpuhp_lock_release();
112 EXPORT_SYMBOL_GPL(put_online_cpus
);
115 * This ensures that the hotplug operation can begin only when the
116 * refcount goes to zero.
118 * Note that during a cpu-hotplug operation, the new readers, if any,
119 * will be blocked by the cpu_hotplug.lock
121 * Since cpu_hotplug_begin() is always called after invoking
122 * cpu_maps_update_begin(), we can be sure that only one writer is active.
124 * Note that theoretically, there is a possibility of a livelock:
125 * - Refcount goes to zero, last reader wakes up the sleeping
127 * - Last reader unlocks the cpu_hotplug.lock.
128 * - A new reader arrives at this moment, bumps up the refcount.
129 * - The writer acquires the cpu_hotplug.lock finds the refcount
130 * non zero and goes to sleep again.
132 * However, this is very difficult to achieve in practice since
133 * get_online_cpus() not an api which is called all that often.
136 void cpu_hotplug_begin(void)
138 cpu_hotplug
.active_writer
= current
;
140 cpuhp_lock_acquire();
142 mutex_lock(&cpu_hotplug
.lock
);
143 if (likely(!cpu_hotplug
.refcount
))
145 __set_current_state(TASK_UNINTERRUPTIBLE
);
146 mutex_unlock(&cpu_hotplug
.lock
);
151 void cpu_hotplug_done(void)
153 cpu_hotplug
.active_writer
= NULL
;
154 mutex_unlock(&cpu_hotplug
.lock
);
155 cpuhp_lock_release();
159 * Wait for currently running CPU hotplug operations to complete (if any) and
160 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
161 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
162 * hotplug path before performing hotplug operations. So acquiring that lock
163 * guarantees mutual exclusion from any currently running hotplug operations.
165 void cpu_hotplug_disable(void)
167 cpu_maps_update_begin();
168 cpu_hotplug_disabled
= 1;
169 cpu_maps_update_done();
172 void cpu_hotplug_enable(void)
174 cpu_maps_update_begin();
175 cpu_hotplug_disabled
= 0;
176 cpu_maps_update_done();
179 #endif /* CONFIG_HOTPLUG_CPU */
181 /* Need to know about CPUs going up/down? */
182 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
185 cpu_maps_update_begin();
186 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
187 cpu_maps_update_done();
191 int __ref
__register_cpu_notifier(struct notifier_block
*nb
)
193 return raw_notifier_chain_register(&cpu_chain
, nb
);
196 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
201 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
204 return notifier_to_errno(ret
);
207 static int cpu_notify(unsigned long val
, void *v
)
209 return __cpu_notify(val
, v
, -1, NULL
);
212 #ifdef CONFIG_HOTPLUG_CPU
214 static void cpu_notify_nofail(unsigned long val
, void *v
)
216 BUG_ON(cpu_notify(val
, v
));
218 EXPORT_SYMBOL(register_cpu_notifier
);
219 EXPORT_SYMBOL(__register_cpu_notifier
);
221 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
223 cpu_maps_update_begin();
224 raw_notifier_chain_unregister(&cpu_chain
, nb
);
225 cpu_maps_update_done();
227 EXPORT_SYMBOL(unregister_cpu_notifier
);
229 void __ref
__unregister_cpu_notifier(struct notifier_block
*nb
)
231 raw_notifier_chain_unregister(&cpu_chain
, nb
);
233 EXPORT_SYMBOL(__unregister_cpu_notifier
);
236 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
239 * This function walks all processes, finds a valid mm struct for each one and
240 * then clears a corresponding bit in mm's cpumask. While this all sounds
241 * trivial, there are various non-obvious corner cases, which this function
242 * tries to solve in a safe manner.
244 * Also note that the function uses a somewhat relaxed locking scheme, so it may
245 * be called only for an already offlined CPU.
247 void clear_tasks_mm_cpumask(int cpu
)
249 struct task_struct
*p
;
252 * This function is called after the cpu is taken down and marked
253 * offline, so its not like new tasks will ever get this cpu set in
254 * their mm mask. -- Peter Zijlstra
255 * Thus, we may use rcu_read_lock() here, instead of grabbing
256 * full-fledged tasklist_lock.
258 WARN_ON(cpu_online(cpu
));
260 for_each_process(p
) {
261 struct task_struct
*t
;
264 * Main thread might exit, but other threads may still have
265 * a valid mm. Find one.
267 t
= find_lock_task_mm(p
);
270 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
276 static inline void check_for_tasks(int cpu
)
278 struct task_struct
*p
;
279 cputime_t utime
, stime
;
281 write_lock_irq(&tasklist_lock
);
282 for_each_process(p
) {
283 task_cputime(p
, &utime
, &stime
);
284 if (task_cpu(p
) == cpu
&& p
->state
== TASK_RUNNING
&&
286 printk(KERN_WARNING
"Task %s (pid = %d) is on cpu %d "
287 "(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 printk("%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*/
419 /* Requires cpu_add_remove_lock to be held */
420 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
422 int ret
, nr_calls
= 0;
423 void *hcpu
= (void *)(long)cpu
;
424 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
425 struct task_struct
*idle
;
429 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
434 idle
= idle_thread_get(cpu
);
440 ret
= smpboot_create_threads(cpu
);
444 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
447 printk(KERN_WARNING
"%s: attempt to bring up CPU %u failed\n",
452 /* Arch-specific enabling code. */
453 ret
= __cpu_up(cpu
, idle
);
456 BUG_ON(!cpu_online(cpu
));
458 /* Wake the per cpu threads */
459 smpboot_unpark_threads(cpu
);
461 /* Now call notifier in preparation. */
462 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
466 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
473 int cpu_up(unsigned int cpu
)
477 if (!cpu_possible(cpu
)) {
478 printk(KERN_ERR
"can't online cpu %d because it is not "
479 "configured as may-hotadd at boot time\n", cpu
);
480 #if defined(CONFIG_IA64)
481 printk(KERN_ERR
"please check additional_cpus= boot "
487 err
= try_online_node(cpu_to_node(cpu
));
491 cpu_maps_update_begin();
493 if (cpu_hotplug_disabled
) {
498 err
= _cpu_up(cpu
, 0);
501 cpu_maps_update_done();
504 EXPORT_SYMBOL_GPL(cpu_up
);
506 #ifdef CONFIG_PM_SLEEP_SMP
507 static cpumask_var_t frozen_cpus
;
509 int disable_nonboot_cpus(void)
511 int cpu
, first_cpu
, error
= 0;
513 cpu_maps_update_begin();
514 first_cpu
= cpumask_first(cpu_online_mask
);
516 * We take down all of the non-boot CPUs in one shot to avoid races
517 * with the userspace trying to use the CPU hotplug at the same time
519 cpumask_clear(frozen_cpus
);
521 printk("Disabling non-boot CPUs ...\n");
522 for_each_online_cpu(cpu
) {
523 if (cpu
== first_cpu
)
525 error
= _cpu_down(cpu
, 1);
527 cpumask_set_cpu(cpu
, frozen_cpus
);
529 printk(KERN_ERR
"Error taking CPU%d down: %d\n",
536 BUG_ON(num_online_cpus() > 1);
537 /* Make sure the CPUs won't be enabled by someone else */
538 cpu_hotplug_disabled
= 1;
540 printk(KERN_ERR
"Non-boot CPUs are not disabled\n");
542 cpu_maps_update_done();
546 void __weak
arch_enable_nonboot_cpus_begin(void)
550 void __weak
arch_enable_nonboot_cpus_end(void)
554 void __ref
enable_nonboot_cpus(void)
558 /* Allow everyone to use the CPU hotplug again */
559 cpu_maps_update_begin();
560 cpu_hotplug_disabled
= 0;
561 if (cpumask_empty(frozen_cpus
))
564 printk(KERN_INFO
"Enabling non-boot CPUs ...\n");
566 arch_enable_nonboot_cpus_begin();
568 for_each_cpu(cpu
, frozen_cpus
) {
569 error
= _cpu_up(cpu
, 1);
571 printk(KERN_INFO
"CPU%d is up\n", cpu
);
574 printk(KERN_WARNING
"Error taking CPU%d up: %d\n", cpu
, error
);
577 arch_enable_nonboot_cpus_end();
579 cpumask_clear(frozen_cpus
);
581 cpu_maps_update_done();
584 static int __init
alloc_frozen_cpus(void)
586 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
590 core_initcall(alloc_frozen_cpus
);
593 * When callbacks for CPU hotplug notifications are being executed, we must
594 * ensure that the state of the system with respect to the tasks being frozen
595 * or not, as reported by the notification, remains unchanged *throughout the
596 * duration* of the execution of the callbacks.
597 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
599 * This synchronization is implemented by mutually excluding regular CPU
600 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
601 * Hibernate notifications.
604 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
605 unsigned long action
, void *ptr
)
609 case PM_SUSPEND_PREPARE
:
610 case PM_HIBERNATION_PREPARE
:
611 cpu_hotplug_disable();
614 case PM_POST_SUSPEND
:
615 case PM_POST_HIBERNATION
:
616 cpu_hotplug_enable();
627 static int __init
cpu_hotplug_pm_sync_init(void)
630 * cpu_hotplug_pm_callback has higher priority than x86
631 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
632 * to disable cpu hotplug to avoid cpu hotplug race.
634 pm_notifier(cpu_hotplug_pm_callback
, 0);
637 core_initcall(cpu_hotplug_pm_sync_init
);
639 #endif /* CONFIG_PM_SLEEP_SMP */
642 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
643 * @cpu: cpu that just started
645 * This function calls the cpu_chain notifiers with CPU_STARTING.
646 * It must be called by the arch code on the new cpu, before the new cpu
647 * enables interrupts and before the "boot" cpu returns from __cpu_up().
649 void notify_cpu_starting(unsigned int cpu
)
651 unsigned long val
= CPU_STARTING
;
653 #ifdef CONFIG_PM_SLEEP_SMP
654 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
655 val
= CPU_STARTING_FROZEN
;
656 #endif /* CONFIG_PM_SLEEP_SMP */
657 cpu_notify(val
, (void *)(long)cpu
);
660 #endif /* CONFIG_SMP */
663 * cpu_bit_bitmap[] is a special, "compressed" data structure that
664 * represents all NR_CPUS bits binary values of 1<<nr.
666 * It is used by cpumask_of() to get a constant address to a CPU
667 * mask value that has a single bit set only.
670 /* cpu_bit_bitmap[0] is empty - so we can back into it */
671 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
672 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
673 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
674 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
676 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
678 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
679 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
680 #if BITS_PER_LONG > 32
681 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
682 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
685 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
687 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
688 EXPORT_SYMBOL(cpu_all_bits
);
690 #ifdef CONFIG_INIT_ALL_POSSIBLE
691 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
694 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
696 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
697 EXPORT_SYMBOL(cpu_possible_mask
);
699 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
700 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
701 EXPORT_SYMBOL(cpu_online_mask
);
703 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
704 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
705 EXPORT_SYMBOL(cpu_present_mask
);
707 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
708 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
709 EXPORT_SYMBOL(cpu_active_mask
);
711 void set_cpu_possible(unsigned int cpu
, bool possible
)
714 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
716 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
719 void set_cpu_present(unsigned int cpu
, bool present
)
722 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
724 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
727 void set_cpu_online(unsigned int cpu
, bool online
)
730 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
732 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
735 void set_cpu_active(unsigned int cpu
, bool active
)
738 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
740 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
743 void init_cpu_present(const struct cpumask
*src
)
745 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
748 void init_cpu_possible(const struct cpumask
*src
)
750 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
753 void init_cpu_online(const struct cpumask
*src
)
755 cpumask_copy(to_cpumask(cpu_online_bits
), src
);