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
);
311 err
= __stop_machine(take_cpu_down
, &tcd_param
, cpumask_of(cpu
));
313 /* CPU didn't die: tell everyone. Can't complain. */
314 smpboot_unpark_threads(cpu
);
315 cpu_notify_nofail(CPU_DOWN_FAILED
| mod
, hcpu
);
318 BUG_ON(cpu_online(cpu
));
321 * The migration_call() CPU_DYING callback will have removed all
322 * runnable tasks from the cpu, there's only the idle task left now
323 * that the migration thread is done doing the stop_machine thing.
325 * Wait for the stop thread to go away.
327 while (!idle_cpu(cpu
))
330 /* This actually kills the CPU. */
333 /* CPU is completely dead: tell everyone. Too late to complain. */
334 cpu_notify_nofail(CPU_DEAD
| mod
, hcpu
);
336 check_for_tasks(cpu
);
341 cpu_notify_nofail(CPU_POST_DEAD
| mod
, hcpu
);
345 int __ref
cpu_down(unsigned int cpu
)
349 cpu_maps_update_begin();
351 if (cpu_hotplug_disabled
) {
356 err
= _cpu_down(cpu
, 0);
359 cpu_maps_update_done();
362 EXPORT_SYMBOL(cpu_down
);
363 #endif /*CONFIG_HOTPLUG_CPU*/
365 /* Requires cpu_add_remove_lock to be held */
366 static int _cpu_up(unsigned int cpu
, int tasks_frozen
)
368 int ret
, nr_calls
= 0;
369 void *hcpu
= (void *)(long)cpu
;
370 unsigned long mod
= tasks_frozen
? CPU_TASKS_FROZEN
: 0;
371 struct task_struct
*idle
;
375 if (cpu_online(cpu
) || !cpu_present(cpu
)) {
380 idle
= idle_thread_get(cpu
);
386 ret
= smpboot_create_threads(cpu
);
390 ret
= __cpu_notify(CPU_UP_PREPARE
| mod
, hcpu
, -1, &nr_calls
);
393 printk(KERN_WARNING
"%s: attempt to bring up CPU %u failed\n",
398 /* Arch-specific enabling code. */
399 ret
= __cpu_up(cpu
, idle
);
402 BUG_ON(!cpu_online(cpu
));
404 /* Wake the per cpu threads */
405 smpboot_unpark_threads(cpu
);
407 /* Now call notifier in preparation. */
408 cpu_notify(CPU_ONLINE
| mod
, hcpu
);
412 __cpu_notify(CPU_UP_CANCELED
| mod
, hcpu
, nr_calls
, NULL
);
419 int cpu_up(unsigned int cpu
)
423 #ifdef CONFIG_MEMORY_HOTPLUG
428 if (!cpu_possible(cpu
)) {
429 printk(KERN_ERR
"can't online cpu %d because it is not "
430 "configured as may-hotadd at boot time\n", cpu
);
431 #if defined(CONFIG_IA64)
432 printk(KERN_ERR
"please check additional_cpus= boot "
438 #ifdef CONFIG_MEMORY_HOTPLUG
439 nid
= cpu_to_node(cpu
);
440 if (!node_online(nid
)) {
441 err
= mem_online_node(nid
);
446 pgdat
= NODE_DATA(nid
);
449 "Can't online cpu %d due to NULL pgdat\n", cpu
);
453 if (pgdat
->node_zonelists
->_zonerefs
->zone
== NULL
) {
454 mutex_lock(&zonelists_mutex
);
455 build_all_zonelists(NULL
, NULL
);
456 mutex_unlock(&zonelists_mutex
);
460 cpu_maps_update_begin();
462 if (cpu_hotplug_disabled
) {
467 err
= _cpu_up(cpu
, 0);
470 cpu_maps_update_done();
473 EXPORT_SYMBOL_GPL(cpu_up
);
475 #ifdef CONFIG_PM_SLEEP_SMP
476 static cpumask_var_t frozen_cpus
;
478 int disable_nonboot_cpus(void)
480 int cpu
, first_cpu
, error
= 0;
482 cpu_maps_update_begin();
483 first_cpu
= cpumask_first(cpu_online_mask
);
485 * We take down all of the non-boot CPUs in one shot to avoid races
486 * with the userspace trying to use the CPU hotplug at the same time
488 cpumask_clear(frozen_cpus
);
490 printk("Disabling non-boot CPUs ...\n");
491 for_each_online_cpu(cpu
) {
492 if (cpu
== first_cpu
)
494 error
= _cpu_down(cpu
, 1);
496 cpumask_set_cpu(cpu
, frozen_cpus
);
498 printk(KERN_ERR
"Error taking CPU%d down: %d\n",
505 BUG_ON(num_online_cpus() > 1);
506 /* Make sure the CPUs won't be enabled by someone else */
507 cpu_hotplug_disabled
= 1;
509 printk(KERN_ERR
"Non-boot CPUs are not disabled\n");
511 cpu_maps_update_done();
515 void __weak
arch_enable_nonboot_cpus_begin(void)
519 void __weak
arch_enable_nonboot_cpus_end(void)
523 void __ref
enable_nonboot_cpus(void)
527 /* Allow everyone to use the CPU hotplug again */
528 cpu_maps_update_begin();
529 cpu_hotplug_disabled
= 0;
530 if (cpumask_empty(frozen_cpus
))
533 printk(KERN_INFO
"Enabling non-boot CPUs ...\n");
535 arch_enable_nonboot_cpus_begin();
537 for_each_cpu(cpu
, frozen_cpus
) {
538 error
= _cpu_up(cpu
, 1);
540 printk(KERN_INFO
"CPU%d is up\n", cpu
);
543 printk(KERN_WARNING
"Error taking CPU%d up: %d\n", cpu
, error
);
546 arch_enable_nonboot_cpus_end();
548 cpumask_clear(frozen_cpus
);
550 cpu_maps_update_done();
553 static int __init
alloc_frozen_cpus(void)
555 if (!alloc_cpumask_var(&frozen_cpus
, GFP_KERNEL
|__GFP_ZERO
))
559 core_initcall(alloc_frozen_cpus
);
562 * When callbacks for CPU hotplug notifications are being executed, we must
563 * ensure that the state of the system with respect to the tasks being frozen
564 * or not, as reported by the notification, remains unchanged *throughout the
565 * duration* of the execution of the callbacks.
566 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
568 * This synchronization is implemented by mutually excluding regular CPU
569 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
570 * Hibernate notifications.
573 cpu_hotplug_pm_callback(struct notifier_block
*nb
,
574 unsigned long action
, void *ptr
)
578 case PM_SUSPEND_PREPARE
:
579 case PM_HIBERNATION_PREPARE
:
580 cpu_hotplug_disable();
583 case PM_POST_SUSPEND
:
584 case PM_POST_HIBERNATION
:
585 cpu_hotplug_enable();
596 static int __init
cpu_hotplug_pm_sync_init(void)
599 * cpu_hotplug_pm_callback has higher priority than x86
600 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
601 * to disable cpu hotplug to avoid cpu hotplug race.
603 pm_notifier(cpu_hotplug_pm_callback
, 0);
606 core_initcall(cpu_hotplug_pm_sync_init
);
608 #endif /* CONFIG_PM_SLEEP_SMP */
611 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
612 * @cpu: cpu that just started
614 * This function calls the cpu_chain notifiers with CPU_STARTING.
615 * It must be called by the arch code on the new cpu, before the new cpu
616 * enables interrupts and before the "boot" cpu returns from __cpu_up().
618 void notify_cpu_starting(unsigned int cpu
)
620 unsigned long val
= CPU_STARTING
;
622 #ifdef CONFIG_PM_SLEEP_SMP
623 if (frozen_cpus
!= NULL
&& cpumask_test_cpu(cpu
, frozen_cpus
))
624 val
= CPU_STARTING_FROZEN
;
625 #endif /* CONFIG_PM_SLEEP_SMP */
626 cpu_notify(val
, (void *)(long)cpu
);
629 #endif /* CONFIG_SMP */
632 * cpu_bit_bitmap[] is a special, "compressed" data structure that
633 * represents all NR_CPUS bits binary values of 1<<nr.
635 * It is used by cpumask_of() to get a constant address to a CPU
636 * mask value that has a single bit set only.
639 /* cpu_bit_bitmap[0] is empty - so we can back into it */
640 #define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
641 #define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
642 #define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
643 #define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
645 const unsigned long cpu_bit_bitmap
[BITS_PER_LONG
+1][BITS_TO_LONGS(NR_CPUS
)] = {
647 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
648 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
649 #if BITS_PER_LONG > 32
650 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
651 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
654 EXPORT_SYMBOL_GPL(cpu_bit_bitmap
);
656 const DECLARE_BITMAP(cpu_all_bits
, NR_CPUS
) = CPU_BITS_ALL
;
657 EXPORT_SYMBOL(cpu_all_bits
);
659 #ifdef CONFIG_INIT_ALL_POSSIBLE
660 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
663 static DECLARE_BITMAP(cpu_possible_bits
, CONFIG_NR_CPUS
) __read_mostly
;
665 const struct cpumask
*const cpu_possible_mask
= to_cpumask(cpu_possible_bits
);
666 EXPORT_SYMBOL(cpu_possible_mask
);
668 static DECLARE_BITMAP(cpu_online_bits
, CONFIG_NR_CPUS
) __read_mostly
;
669 const struct cpumask
*const cpu_online_mask
= to_cpumask(cpu_online_bits
);
670 EXPORT_SYMBOL(cpu_online_mask
);
672 static DECLARE_BITMAP(cpu_present_bits
, CONFIG_NR_CPUS
) __read_mostly
;
673 const struct cpumask
*const cpu_present_mask
= to_cpumask(cpu_present_bits
);
674 EXPORT_SYMBOL(cpu_present_mask
);
676 static DECLARE_BITMAP(cpu_active_bits
, CONFIG_NR_CPUS
) __read_mostly
;
677 const struct cpumask
*const cpu_active_mask
= to_cpumask(cpu_active_bits
);
678 EXPORT_SYMBOL(cpu_active_mask
);
680 void set_cpu_possible(unsigned int cpu
, bool possible
)
683 cpumask_set_cpu(cpu
, to_cpumask(cpu_possible_bits
));
685 cpumask_clear_cpu(cpu
, to_cpumask(cpu_possible_bits
));
688 void set_cpu_present(unsigned int cpu
, bool present
)
691 cpumask_set_cpu(cpu
, to_cpumask(cpu_present_bits
));
693 cpumask_clear_cpu(cpu
, to_cpumask(cpu_present_bits
));
696 void set_cpu_online(unsigned int cpu
, bool online
)
699 cpumask_set_cpu(cpu
, to_cpumask(cpu_online_bits
));
700 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
702 cpumask_clear_cpu(cpu
, to_cpumask(cpu_online_bits
));
706 void set_cpu_active(unsigned int cpu
, bool active
)
709 cpumask_set_cpu(cpu
, to_cpumask(cpu_active_bits
));
711 cpumask_clear_cpu(cpu
, to_cpumask(cpu_active_bits
));
714 void init_cpu_present(const struct cpumask
*src
)
716 cpumask_copy(to_cpumask(cpu_present_bits
), src
);
719 void init_cpu_possible(const struct cpumask
*src
)
721 cpumask_copy(to_cpumask(cpu_possible_bits
), src
);
724 void init_cpu_online(const struct cpumask
*src
)
726 cpumask_copy(to_cpumask(cpu_online_bits
), src
);