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>
26 /* Serializes the updates to cpu_online_mask, cpu_present_mask */
27 static DEFINE_MUTEX(cpu_add_remove_lock
);
30 * The following two API's must be used when attempting
31 * to serialize the updates to cpu_online_mask, cpu_present_mask.
33 void cpu_maps_update_begin(void)
35 mutex_lock(&cpu_add_remove_lock
);
38 void cpu_maps_update_done(void)
40 mutex_unlock(&cpu_add_remove_lock
);
43 static RAW_NOTIFIER_HEAD(cpu_chain
);
45 /* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
46 * Should always be manipulated under cpu_add_remove_lock
48 static int cpu_hotplug_disabled
;
50 #ifdef CONFIG_HOTPLUG_CPU
53 struct task_struct
*active_writer
;
54 struct mutex lock
; /* Synchronizes accesses to refcount, */
56 * Also blocks the new readers during
57 * an ongoing cpu hotplug operation.
61 .active_writer
= NULL
,
62 .lock
= __MUTEX_INITIALIZER(cpu_hotplug
.lock
),
66 void get_online_cpus(void)
69 if (cpu_hotplug
.active_writer
== current
)
71 mutex_lock(&cpu_hotplug
.lock
);
72 cpu_hotplug
.refcount
++;
73 mutex_unlock(&cpu_hotplug
.lock
);
76 EXPORT_SYMBOL_GPL(get_online_cpus
);
78 void put_online_cpus(void)
80 if (cpu_hotplug
.active_writer
== current
)
82 mutex_lock(&cpu_hotplug
.lock
);
84 if (WARN_ON(!cpu_hotplug
.refcount
))
85 cpu_hotplug
.refcount
++; /* try to fix things up */
87 if (!--cpu_hotplug
.refcount
&& unlikely(cpu_hotplug
.active_writer
))
88 wake_up_process(cpu_hotplug
.active_writer
);
89 mutex_unlock(&cpu_hotplug
.lock
);
92 EXPORT_SYMBOL_GPL(put_online_cpus
);
95 * This ensures that the hotplug operation can begin only when the
96 * refcount goes to zero.
98 * Note that during a cpu-hotplug operation, the new readers, if any,
99 * will be blocked by the cpu_hotplug.lock
101 * Since cpu_hotplug_begin() is always called after invoking
102 * cpu_maps_update_begin(), we can be sure that only one writer is active.
104 * Note that theoretically, there is a possibility of a livelock:
105 * - Refcount goes to zero, last reader wakes up the sleeping
107 * - Last reader unlocks the cpu_hotplug.lock.
108 * - A new reader arrives at this moment, bumps up the refcount.
109 * - The writer acquires the cpu_hotplug.lock finds the refcount
110 * non zero and goes to sleep again.
112 * However, this is very difficult to achieve in practice since
113 * get_online_cpus() not an api which is called all that often.
116 void cpu_hotplug_begin(void)
118 cpu_hotplug
.active_writer
= current
;
121 mutex_lock(&cpu_hotplug
.lock
);
122 if (likely(!cpu_hotplug
.refcount
))
124 __set_current_state(TASK_UNINTERRUPTIBLE
);
125 mutex_unlock(&cpu_hotplug
.lock
);
130 void cpu_hotplug_done(void)
132 cpu_hotplug
.active_writer
= NULL
;
133 mutex_unlock(&cpu_hotplug
.lock
);
137 * Wait for currently running CPU hotplug operations to complete (if any) and
138 * disable future CPU hotplug (from sysfs). The 'cpu_add_remove_lock' protects
139 * the 'cpu_hotplug_disabled' flag. The same lock is also acquired by the
140 * hotplug path before performing hotplug operations. So acquiring that lock
141 * guarantees mutual exclusion from any currently running hotplug operations.
143 void cpu_hotplug_disable(void)
145 cpu_maps_update_begin();
146 cpu_hotplug_disabled
= 1;
147 cpu_maps_update_done();
150 void cpu_hotplug_enable(void)
152 cpu_maps_update_begin();
153 cpu_hotplug_disabled
= 0;
154 cpu_maps_update_done();
157 #endif /* CONFIG_HOTPLUG_CPU */
159 /* Need to know about CPUs going up/down? */
160 int __ref
register_cpu_notifier(struct notifier_block
*nb
)
163 cpu_maps_update_begin();
164 ret
= raw_notifier_chain_register(&cpu_chain
, nb
);
165 cpu_maps_update_done();
169 static int __cpu_notify(unsigned long val
, void *v
, int nr_to_call
,
174 ret
= __raw_notifier_call_chain(&cpu_chain
, val
, v
, nr_to_call
,
177 return notifier_to_errno(ret
);
180 static int cpu_notify(unsigned long val
, void *v
)
182 return __cpu_notify(val
, v
, -1, NULL
);
185 #ifdef CONFIG_HOTPLUG_CPU
187 static void cpu_notify_nofail(unsigned long val
, void *v
)
189 BUG_ON(cpu_notify(val
, v
));
191 EXPORT_SYMBOL(register_cpu_notifier
);
193 void __ref
unregister_cpu_notifier(struct notifier_block
*nb
)
195 cpu_maps_update_begin();
196 raw_notifier_chain_unregister(&cpu_chain
, nb
);
197 cpu_maps_update_done();
199 EXPORT_SYMBOL(unregister_cpu_notifier
);
202 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
205 * This function walks all processes, finds a valid mm struct for each one and
206 * then clears a corresponding bit in mm's cpumask. While this all sounds
207 * trivial, there are various non-obvious corner cases, which this function
208 * tries to solve in a safe manner.
210 * Also note that the function uses a somewhat relaxed locking scheme, so it may
211 * be called only for an already offlined CPU.
213 void clear_tasks_mm_cpumask(int cpu
)
215 struct task_struct
*p
;
218 * This function is called after the cpu is taken down and marked
219 * offline, so its not like new tasks will ever get this cpu set in
220 * their mm mask. -- Peter Zijlstra
221 * Thus, we may use rcu_read_lock() here, instead of grabbing
222 * full-fledged tasklist_lock.
224 WARN_ON(cpu_online(cpu
));
226 for_each_process(p
) {
227 struct task_struct
*t
;
230 * Main thread might exit, but other threads may still have
231 * a valid mm. Find one.
233 t
= find_lock_task_mm(p
);
236 cpumask_clear_cpu(cpu
, mm_cpumask(t
->mm
));
242 static inline void check_for_tasks(int cpu
)
244 struct task_struct
*p
;
245 cputime_t utime
, stime
;
247 write_lock_irq(&tasklist_lock
);
248 for_each_process(p
) {
249 task_cputime(p
, &utime
, &stime
);
250 if (task_cpu(p
) == cpu
&& p
->state
== TASK_RUNNING
&&
252 printk(KERN_WARNING
"Task %s (pid = %d) is on cpu %d "
253 "(state = %ld, flags = %x)\n",
254 p
->comm
, task_pid_nr(p
), cpu
,
257 write_unlock_irq(&tasklist_lock
);
260 struct take_cpu_down_param
{
265 /* Take this CPU down. */
266 static int __ref
take_cpu_down(void *_param
)
268 struct take_cpu_down_param
*param
= _param
;
271 /* Ensure this CPU doesn't handle any more interrupts. */
272 err
= __cpu_disable();
276 cpu_notify(CPU_DYING
| param
->mod
, param
->hcpu
);
277 /* Park the stopper thread */
278 kthread_park(current
);
282 /* Requires cpu_add_remove_lock to be held */
283 static int __ref
_cpu_down(unsigned int cpu
, int tasks_frozen
)
285 int err
, nr_calls
= 0;
286 void *hcpu
= (void *)(long)cpu
;
287 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
288 struct take_cpu_down_param tcd_param
= {
293 if (num_online_cpus() == 1)
296 if (!cpu_online(cpu
))
301 err
= __cpu_notify(CPU_DOWN_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
304 __cpu_notify(CPU_DOWN_FAILED
| mod
, hcpu
, nr_calls
, NULL
);
305 printk("%s: attempt to take down CPU %u failed\n",
309 smpboot_park_threads(cpu
);
312 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
313 * and RCU users of this state to go away such that all new such users
316 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
317 * not imply sync_sched(), so explicitly call both.
319 #ifdef CONFIG_PREEMPT
325 * So now all preempt/rcu users must observe !cpu_active().
328 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
330 /* CPU didn't die: tell everyone. Can't complain. */
331 smpboot_unpark_threads(cpu
);
332 cpu_notify_nofail(CPU_DOWN_FAILED
| mod
, hcpu
);
335 BUG_ON(cpu_online(cpu
));
338 * The migration_call() CPU_DYING callback will have removed all
339 * runnable tasks from the cpu, there's only the idle task left now
340 * that the migration thread is done doing the stop_machine thing.
342 * Wait for the stop thread to go away.
344 while (!idle_cpu(cpu
))
347 /* This actually kills the CPU. */
350 /* CPU is completely dead: tell everyone. Too late to complain. */
351 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
353 check_for_tasks(cpu
);
358 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
362 int __ref
cpu_down(unsigned int cpu
)
366 cpu_maps_update_begin();
368 if (cpu_hotplug_disabled
) {
373 err
= _cpu_down(cpu
, 0);
376 cpu_maps_update_done();
379 EXPORT_SYMBOL(cpu_down
);
380 #endif /*CONFIG_HOTPLUG_CPU*/
382 /* Requires cpu_add_remove_lock to be held */
383 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
385 int ret
, nr_calls
= 0;
386 void *hcpu
= (void *)(long)cpu
;
387 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
388 struct task_struct
*idle
;
392 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
397 idle
= idle_thread_get(cpu
);
403 ret
= smpboot_create_threads(cpu
);
407 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
410 printk(KERN_WARNING
"%s: attempt to bring up CPU %u failed\n",
415 /* Arch-specific enabling code. */
416 ret
= __cpu_up(cpu
, idle
);
419 BUG_ON(!cpu_online(cpu
));
421 /* Wake the per cpu threads */
422 smpboot_unpark_threads(cpu
);
424 /* Now call notifier in preparation. */
425 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
429 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
436 int cpu_up(unsigned int cpu
)
440 #ifdef CONFIG_MEMORY_HOTPLUG
445 if (!cpu_possible(cpu
)) {
446 printk(KERN_ERR
"can't online cpu %d because it is not "
447 "configured as may-hotadd at boot time\n", cpu
);
448 #if defined(CONFIG_IA64)
449 printk(KERN_ERR
"please check additional_cpus= boot "
455 #ifdef CONFIG_MEMORY_HOTPLUG
456 nid
= cpu_to_node(cpu
);
457 if (!node_online(nid
)) {
458 err
= mem_online_node(nid
);
463 pgdat
= NODE_DATA(nid
);
466 "Can't online cpu %d due to NULL pgdat\n", cpu
);
470 if (pgdat
->node_zonelists
->_zonerefs
->zone
== NULL
) {
471 mutex_lock(&zonelists_mutex
);
472 build_all_zonelists(NULL
, NULL
);
473 mutex_unlock(&zonelists_mutex
);
477 cpu_maps_update_begin();
479 if (cpu_hotplug_disabled
) {
484 err
= _cpu_up(cpu
, 0);
487 cpu_maps_update_done();
490 EXPORT_SYMBOL_GPL(cpu_up
);
492 #ifdef CONFIG_PM_SLEEP_SMP
493 static cpumask_var_t frozen_cpus
;
495 int disable_nonboot_cpus(void)
497 int cpu
, first_cpu
, error
= 0;
499 cpu_maps_update_begin();
500 first_cpu
= cpumask_first(cpu_online_mask
);
502 * We take down all of the non-boot CPUs in one shot to avoid races
503 * with the userspace trying to use the CPU hotplug at the same time
505 cpumask_clear(frozen_cpus
);
507 printk("Disabling non-boot CPUs ...\n");
508 for_each_online_cpu(cpu
) {
509 if (cpu
== first_cpu
)
511 error
= _cpu_down(cpu
, 1);
513 cpumask_set_cpu(cpu
, frozen_cpus
);
515 printk(KERN_ERR
"Error taking CPU%d down: %d\n",
522 BUG_ON(num_online_cpus() > 1);
523 /* Make sure the CPUs won't be enabled by someone else */
524 cpu_hotplug_disabled
= 1;
526 printk(KERN_ERR
"Non-boot CPUs are not disabled\n");
528 cpu_maps_update_done();
532 void __weak
arch_enable_nonboot_cpus_begin(void)
536 void __weak
arch_enable_nonboot_cpus_end(void)
540 void __ref
enable_nonboot_cpus(void)
544 /* Allow everyone to use the CPU hotplug again */
545 cpu_maps_update_begin();
546 cpu_hotplug_disabled
= 0;
547 if (cpumask_empty(frozen_cpus
))
550 printk(KERN_INFO
"Enabling non-boot CPUs ...\n");
552 arch_enable_nonboot_cpus_begin();
554 for_each_cpu(cpu
, frozen_cpus
) {
555 error
= _cpu_up(cpu
, 1);
557 printk(KERN_INFO
"CPU%d is up\n", cpu
);
560 printk(KERN_WARNING
"Error taking CPU%d up: %d\n", cpu
, error
);
563 arch_enable_nonboot_cpus_end();
565 cpumask_clear(frozen_cpus
);
567 cpu_maps_update_done();
570 static int __init
alloc_frozen_cpus(void)
572 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
576 core_initcall(alloc_frozen_cpus
);
579 * When callbacks for CPU hotplug notifications are being executed, we must
580 * ensure that the state of the system with respect to the tasks being frozen
581 * or not, as reported by the notification, remains unchanged *throughout the
582 * duration* of the execution of the callbacks.
583 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
585 * This synchronization is implemented by mutually excluding regular CPU
586 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
587 * Hibernate notifications.
590 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
591 unsigned long action
, void *ptr
)
595 case PM_SUSPEND_PREPARE
:
596 case PM_HIBERNATION_PREPARE
:
597 cpu_hotplug_disable();
600 case PM_POST_SUSPEND
:
601 case PM_POST_HIBERNATION
:
602 cpu_hotplug_enable();
613 static int __init
cpu_hotplug_pm_sync_init(void)
616 * cpu_hotplug_pm_callback has higher priority than x86
617 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
618 * to disable cpu hotplug to avoid cpu hotplug race.
620 pm_notifier(cpu_hotplug_pm_callback
, 0);
623 core_initcall(cpu_hotplug_pm_sync_init
);
625 #endif /* CONFIG_PM_SLEEP_SMP */
628 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
629 * @cpu: cpu that just started
631 * This function calls the cpu_chain notifiers with CPU_STARTING.
632 * It must be called by the arch code on the new cpu, before the new cpu
633 * enables interrupts and before the "boot" cpu returns from __cpu_up().
635 void notify_cpu_starting(unsigned int cpu
)
637 unsigned long val
= CPU_STARTING
;
639 #ifdef CONFIG_PM_SLEEP_SMP
640 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
641 val
= CPU_STARTING_FROZEN
;
642 #endif /* CONFIG_PM_SLEEP_SMP */
643 cpu_notify(val
, (void *)(long)cpu
);
646 #endif /* CONFIG_SMP */
649 * cpu_bit_bitmap[] is a special, "compressed" data structure that
650 * represents all NR_CPUS bits binary values of 1<<nr.
652 * It is used by cpumask_of() to get a constant address to a CPU
653 * mask value that has a single bit set only.
656 /* cpu_bit_bitmap[0] is empty - so we can back into it */
657 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
658 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
659 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
660 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
662 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
664 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
665 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
666 #if BITS_PER_LONG > 32
667 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
668 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
671 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
673 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
674 EXPORT_SYMBOL(cpu_all_bits
);
676 #ifdef CONFIG_INIT_ALL_POSSIBLE
677 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
680 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
682 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
683 EXPORT_SYMBOL(cpu_possible_mask
);
685 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
686 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
687 EXPORT_SYMBOL(cpu_online_mask
);
689 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
690 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
691 EXPORT_SYMBOL(cpu_present_mask
);
693 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
694 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
695 EXPORT_SYMBOL(cpu_active_mask
);
697 void set_cpu_possible(unsigned int cpu
, bool possible
)
700 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
702 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
705 void set_cpu_present(unsigned int cpu
, bool present
)
708 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
710 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
713 void set_cpu_online(unsigned int cpu
, bool online
)
716 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
718 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
721 void set_cpu_active(unsigned int cpu
, bool active
)
724 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
726 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
729 void init_cpu_present(const struct cpumask
*src
)
731 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
734 void init_cpu_possible(const struct cpumask
*src
)
736 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
739 void init_cpu_online(const struct cpumask
*src
)
741 cpumask_copy(to_cpumask(cpu_online_bits
), src
);