Loading...
1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/proc_fs.h>
7#include <linux/smp.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
24#include "smpboot.h"
25
26#ifdef CONFIG_SMP
27/* Serializes the updates to cpu_online_mask, cpu_present_mask */
28static DEFINE_MUTEX(cpu_add_remove_lock);
29
30/*
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().
36 */
37void cpu_maps_update_begin(void)
38{
39 mutex_lock(&cpu_add_remove_lock);
40}
41EXPORT_SYMBOL(cpu_notifier_register_begin);
42
43void cpu_maps_update_done(void)
44{
45 mutex_unlock(&cpu_add_remove_lock);
46}
47EXPORT_SYMBOL(cpu_notifier_register_done);
48
49static RAW_NOTIFIER_HEAD(cpu_chain);
50
51/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
52 * Should always be manipulated under cpu_add_remove_lock
53 */
54static int cpu_hotplug_disabled;
55
56#ifdef CONFIG_HOTPLUG_CPU
57
58static struct {
59 struct task_struct *active_writer;
60 struct mutex lock; /* Synchronizes accesses to refcount, */
61 /*
62 * Also blocks the new readers during
63 * an ongoing cpu hotplug operation.
64 */
65 int refcount;
66
67#ifdef CONFIG_DEBUG_LOCK_ALLOC
68 struct lockdep_map dep_map;
69#endif
70} cpu_hotplug = {
71 .active_writer = NULL,
72 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
73 .refcount = 0,
74#ifdef CONFIG_DEBUG_LOCK_ALLOC
75 .dep_map = {.name = "cpu_hotplug.lock" },
76#endif
77};
78
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)
83
84void get_online_cpus(void)
85{
86 might_sleep();
87 if (cpu_hotplug.active_writer == current)
88 return;
89 cpuhp_lock_acquire_read();
90 mutex_lock(&cpu_hotplug.lock);
91 cpu_hotplug.refcount++;
92 mutex_unlock(&cpu_hotplug.lock);
93
94}
95EXPORT_SYMBOL_GPL(get_online_cpus);
96
97void put_online_cpus(void)
98{
99 if (cpu_hotplug.active_writer == current)
100 return;
101 mutex_lock(&cpu_hotplug.lock);
102
103 if (WARN_ON(!cpu_hotplug.refcount))
104 cpu_hotplug.refcount++; /* try to fix things up */
105
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();
110
111}
112EXPORT_SYMBOL_GPL(put_online_cpus);
113
114/*
115 * This ensures that the hotplug operation can begin only when the
116 * refcount goes to zero.
117 *
118 * Note that during a cpu-hotplug operation, the new readers, if any,
119 * will be blocked by the cpu_hotplug.lock
120 *
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.
123 *
124 * Note that theoretically, there is a possibility of a livelock:
125 * - Refcount goes to zero, last reader wakes up the sleeping
126 * writer.
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.
131 *
132 * However, this is very difficult to achieve in practice since
133 * get_online_cpus() not an api which is called all that often.
134 *
135 */
136void cpu_hotplug_begin(void)
137{
138 cpu_hotplug.active_writer = current;
139
140 cpuhp_lock_acquire();
141 for (;;) {
142 mutex_lock(&cpu_hotplug.lock);
143 if (likely(!cpu_hotplug.refcount))
144 break;
145 __set_current_state(TASK_UNINTERRUPTIBLE);
146 mutex_unlock(&cpu_hotplug.lock);
147 schedule();
148 }
149}
150
151void cpu_hotplug_done(void)
152{
153 cpu_hotplug.active_writer = NULL;
154 mutex_unlock(&cpu_hotplug.lock);
155 cpuhp_lock_release();
156}
157
158/*
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.
164 */
165void cpu_hotplug_disable(void)
166{
167 cpu_maps_update_begin();
168 cpu_hotplug_disabled = 1;
169 cpu_maps_update_done();
170}
171
172void cpu_hotplug_enable(void)
173{
174 cpu_maps_update_begin();
175 cpu_hotplug_disabled = 0;
176 cpu_maps_update_done();
177}
178
179#endif /* CONFIG_HOTPLUG_CPU */
180
181/* Need to know about CPUs going up/down? */
182int __ref register_cpu_notifier(struct notifier_block *nb)
183{
184 int ret;
185 cpu_maps_update_begin();
186 ret = raw_notifier_chain_register(&cpu_chain, nb);
187 cpu_maps_update_done();
188 return ret;
189}
190
191int __ref __register_cpu_notifier(struct notifier_block *nb)
192{
193 return raw_notifier_chain_register(&cpu_chain, nb);
194}
195
196static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
197 int *nr_calls)
198{
199 int ret;
200
201 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
202 nr_calls);
203
204 return notifier_to_errno(ret);
205}
206
207static int cpu_notify(unsigned long val, void *v)
208{
209 return __cpu_notify(val, v, -1, NULL);
210}
211
212#ifdef CONFIG_HOTPLUG_CPU
213
214static void cpu_notify_nofail(unsigned long val, void *v)
215{
216 BUG_ON(cpu_notify(val, v));
217}
218EXPORT_SYMBOL(register_cpu_notifier);
219EXPORT_SYMBOL(__register_cpu_notifier);
220
221void __ref unregister_cpu_notifier(struct notifier_block *nb)
222{
223 cpu_maps_update_begin();
224 raw_notifier_chain_unregister(&cpu_chain, nb);
225 cpu_maps_update_done();
226}
227EXPORT_SYMBOL(unregister_cpu_notifier);
228
229void __ref __unregister_cpu_notifier(struct notifier_block *nb)
230{
231 raw_notifier_chain_unregister(&cpu_chain, nb);
232}
233EXPORT_SYMBOL(__unregister_cpu_notifier);
234
235/**
236 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
237 * @cpu: a CPU id
238 *
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.
243 *
244 * Also note that the function uses a somewhat relaxed locking scheme, so it may
245 * be called only for an already offlined CPU.
246 */
247void clear_tasks_mm_cpumask(int cpu)
248{
249 struct task_struct *p;
250
251 /*
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.
257 */
258 WARN_ON(cpu_online(cpu));
259 rcu_read_lock();
260 for_each_process(p) {
261 struct task_struct *t;
262
263 /*
264 * Main thread might exit, but other threads may still have
265 * a valid mm. Find one.
266 */
267 t = find_lock_task_mm(p);
268 if (!t)
269 continue;
270 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
271 task_unlock(t);
272 }
273 rcu_read_unlock();
274}
275
276static inline void check_for_tasks(int cpu)
277{
278 struct task_struct *p;
279 cputime_t utime, stime;
280
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 &&
285 (utime || stime))
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,
289 p->state, p->flags);
290 }
291 write_unlock_irq(&tasklist_lock);
292}
293
294struct take_cpu_down_param {
295 unsigned long mod;
296 void *hcpu;
297};
298
299/* Take this CPU down. */
300static int __ref take_cpu_down(void *_param)
301{
302 struct take_cpu_down_param *param = _param;
303 int err;
304
305 /* Ensure this CPU doesn't handle any more interrupts. */
306 err = __cpu_disable();
307 if (err < 0)
308 return err;
309
310 cpu_notify(CPU_DYING | param->mod, param->hcpu);
311 /* Park the stopper thread */
312 kthread_park(current);
313 return 0;
314}
315
316/* Requires cpu_add_remove_lock to be held */
317static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
318{
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 = {
323 .mod = mod,
324 .hcpu = hcpu,
325 };
326
327 if (num_online_cpus() == 1)
328 return -EBUSY;
329
330 if (!cpu_online(cpu))
331 return -EINVAL;
332
333 cpu_hotplug_begin();
334
335 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
336 if (err) {
337 nr_calls--;
338 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
339 printk("%s: attempt to take down CPU %u failed\n",
340 __func__, cpu);
341 goto out_release;
342 }
343
344 /*
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
347 * will observe it.
348 *
349 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
350 * not imply sync_sched(), so explicitly call both.
351 *
352 * Do sync before park smpboot threads to take care the rcu boost case.
353 */
354#ifdef CONFIG_PREEMPT
355 synchronize_sched();
356#endif
357 synchronize_rcu();
358
359 smpboot_park_threads(cpu);
360
361 /*
362 * So now all preempt/rcu users must observe !cpu_active().
363 */
364
365 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
366 if (err) {
367 /* CPU didn't die: tell everyone. Can't complain. */
368 smpboot_unpark_threads(cpu);
369 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
370 goto out_release;
371 }
372 BUG_ON(cpu_online(cpu));
373
374 /*
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.
378 *
379 * Wait for the stop thread to go away.
380 */
381 while (!idle_cpu(cpu))
382 cpu_relax();
383
384 /* This actually kills the CPU. */
385 __cpu_die(cpu);
386
387 /* CPU is completely dead: tell everyone. Too late to complain. */
388 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
389
390 check_for_tasks(cpu);
391
392out_release:
393 cpu_hotplug_done();
394 if (!err)
395 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
396 return err;
397}
398
399int __ref cpu_down(unsigned int cpu)
400{
401 int err;
402
403 cpu_maps_update_begin();
404
405 if (cpu_hotplug_disabled) {
406 err = -EBUSY;
407 goto out;
408 }
409
410 err = _cpu_down(cpu, 0);
411
412out:
413 cpu_maps_update_done();
414 return err;
415}
416EXPORT_SYMBOL(cpu_down);
417#endif /*CONFIG_HOTPLUG_CPU*/
418
419/* Requires cpu_add_remove_lock to be held */
420static int _cpu_up(unsigned int cpu, int tasks_frozen)
421{
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;
426
427 cpu_hotplug_begin();
428
429 if (cpu_online(cpu) || !cpu_present(cpu)) {
430 ret = -EINVAL;
431 goto out;
432 }
433
434 idle = idle_thread_get(cpu);
435 if (IS_ERR(idle)) {
436 ret = PTR_ERR(idle);
437 goto out;
438 }
439
440 ret = smpboot_create_threads(cpu);
441 if (ret)
442 goto out;
443
444 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
445 if (ret) {
446 nr_calls--;
447 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
448 __func__, cpu);
449 goto out_notify;
450 }
451
452 /* Arch-specific enabling code. */
453 ret = __cpu_up(cpu, idle);
454 if (ret != 0)
455 goto out_notify;
456 BUG_ON(!cpu_online(cpu));
457
458 /* Wake the per cpu threads */
459 smpboot_unpark_threads(cpu);
460
461 /* Now call notifier in preparation. */
462 cpu_notify(CPU_ONLINE | mod, hcpu);
463
464out_notify:
465 if (ret != 0)
466 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
467out:
468 cpu_hotplug_done();
469
470 return ret;
471}
472
473int cpu_up(unsigned int cpu)
474{
475 int err = 0;
476
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 "
482 "parameter\n");
483#endif
484 return -EINVAL;
485 }
486
487 err = try_online_node(cpu_to_node(cpu));
488 if (err)
489 return err;
490
491 cpu_maps_update_begin();
492
493 if (cpu_hotplug_disabled) {
494 err = -EBUSY;
495 goto out;
496 }
497
498 err = _cpu_up(cpu, 0);
499
500out:
501 cpu_maps_update_done();
502 return err;
503}
504EXPORT_SYMBOL_GPL(cpu_up);
505
506#ifdef CONFIG_PM_SLEEP_SMP
507static cpumask_var_t frozen_cpus;
508
509int disable_nonboot_cpus(void)
510{
511 int cpu, first_cpu, error = 0;
512
513 cpu_maps_update_begin();
514 first_cpu = cpumask_first(cpu_online_mask);
515 /*
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
518 */
519 cpumask_clear(frozen_cpus);
520
521 printk("Disabling non-boot CPUs ...\n");
522 for_each_online_cpu(cpu) {
523 if (cpu == first_cpu)
524 continue;
525 error = _cpu_down(cpu, 1);
526 if (!error)
527 cpumask_set_cpu(cpu, frozen_cpus);
528 else {
529 printk(KERN_ERR "Error taking CPU%d down: %d\n",
530 cpu, error);
531 break;
532 }
533 }
534
535 if (!error) {
536 BUG_ON(num_online_cpus() > 1);
537 /* Make sure the CPUs won't be enabled by someone else */
538 cpu_hotplug_disabled = 1;
539 } else {
540 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
541 }
542 cpu_maps_update_done();
543 return error;
544}
545
546void __weak arch_enable_nonboot_cpus_begin(void)
547{
548}
549
550void __weak arch_enable_nonboot_cpus_end(void)
551{
552}
553
554void __ref enable_nonboot_cpus(void)
555{
556 int cpu, error;
557
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))
562 goto out;
563
564 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
565
566 arch_enable_nonboot_cpus_begin();
567
568 for_each_cpu(cpu, frozen_cpus) {
569 error = _cpu_up(cpu, 1);
570 if (!error) {
571 printk(KERN_INFO "CPU%d is up\n", cpu);
572 continue;
573 }
574 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
575 }
576
577 arch_enable_nonboot_cpus_end();
578
579 cpumask_clear(frozen_cpus);
580out:
581 cpu_maps_update_done();
582}
583
584static int __init alloc_frozen_cpus(void)
585{
586 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
587 return -ENOMEM;
588 return 0;
589}
590core_initcall(alloc_frozen_cpus);
591
592/*
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.
598 *
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.
602 */
603static int
604cpu_hotplug_pm_callback(struct notifier_block *nb,
605 unsigned long action, void *ptr)
606{
607 switch (action) {
608
609 case PM_SUSPEND_PREPARE:
610 case PM_HIBERNATION_PREPARE:
611 cpu_hotplug_disable();
612 break;
613
614 case PM_POST_SUSPEND:
615 case PM_POST_HIBERNATION:
616 cpu_hotplug_enable();
617 break;
618
619 default:
620 return NOTIFY_DONE;
621 }
622
623 return NOTIFY_OK;
624}
625
626
627static int __init cpu_hotplug_pm_sync_init(void)
628{
629 /*
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.
633 */
634 pm_notifier(cpu_hotplug_pm_callback, 0);
635 return 0;
636}
637core_initcall(cpu_hotplug_pm_sync_init);
638
639#endif /* CONFIG_PM_SLEEP_SMP */
640
641/**
642 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
643 * @cpu: cpu that just started
644 *
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().
648 */
649void notify_cpu_starting(unsigned int cpu)
650{
651 unsigned long val = CPU_STARTING;
652
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);
658}
659
660#endif /* CONFIG_SMP */
661
662/*
663 * cpu_bit_bitmap[] is a special, "compressed" data structure that
664 * represents all NR_CPUS bits binary values of 1<<nr.
665 *
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.
668 */
669
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)
675
676const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
677
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),
683#endif
684};
685EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
686
687const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
688EXPORT_SYMBOL(cpu_all_bits);
689
690#ifdef CONFIG_INIT_ALL_POSSIBLE
691static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
692 = CPU_BITS_ALL;
693#else
694static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
695#endif
696const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
697EXPORT_SYMBOL(cpu_possible_mask);
698
699static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
700const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
701EXPORT_SYMBOL(cpu_online_mask);
702
703static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
704const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
705EXPORT_SYMBOL(cpu_present_mask);
706
707static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
708const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
709EXPORT_SYMBOL(cpu_active_mask);
710
711void set_cpu_possible(unsigned int cpu, bool possible)
712{
713 if (possible)
714 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
715 else
716 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
717}
718
719void set_cpu_present(unsigned int cpu, bool present)
720{
721 if (present)
722 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
723 else
724 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
725}
726
727void set_cpu_online(unsigned int cpu, bool online)
728{
729 if (online) {
730 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
731 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
732 } else {
733 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
734 }
735}
736
737void set_cpu_active(unsigned int cpu, bool active)
738{
739 if (active)
740 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
741 else
742 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
743}
744
745void init_cpu_present(const struct cpumask *src)
746{
747 cpumask_copy(to_cpumask(cpu_present_bits), src);
748}
749
750void init_cpu_possible(const struct cpumask *src)
751{
752 cpumask_copy(to_cpumask(cpu_possible_bits), src);
753}
754
755void init_cpu_online(const struct cpumask *src)
756{
757 cpumask_copy(to_cpumask(cpu_online_bits), src);
758}
1/* CPU control.
2 * (C) 2001, 2002, 2003, 2004 Rusty Russell
3 *
4 * This code is licenced under the GPL.
5 */
6#include <linux/proc_fs.h>
7#include <linux/smp.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/module.h>
14#include <linux/kthread.h>
15#include <linux/stop_machine.h>
16#include <linux/mutex.h>
17#include <linux/gfp.h>
18
19#ifdef CONFIG_SMP
20/* Serializes the updates to cpu_online_mask, cpu_present_mask */
21static DEFINE_MUTEX(cpu_add_remove_lock);
22
23/*
24 * The following two API's must be used when attempting
25 * to serialize the updates to cpu_online_mask, cpu_present_mask.
26 */
27void cpu_maps_update_begin(void)
28{
29 mutex_lock(&cpu_add_remove_lock);
30}
31
32void cpu_maps_update_done(void)
33{
34 mutex_unlock(&cpu_add_remove_lock);
35}
36
37static RAW_NOTIFIER_HEAD(cpu_chain);
38
39/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
40 * Should always be manipulated under cpu_add_remove_lock
41 */
42static int cpu_hotplug_disabled;
43
44#ifdef CONFIG_HOTPLUG_CPU
45
46static struct {
47 struct task_struct *active_writer;
48 struct mutex lock; /* Synchronizes accesses to refcount, */
49 /*
50 * Also blocks the new readers during
51 * an ongoing cpu hotplug operation.
52 */
53 int refcount;
54} cpu_hotplug = {
55 .active_writer = NULL,
56 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
57 .refcount = 0,
58};
59
60void get_online_cpus(void)
61{
62 might_sleep();
63 if (cpu_hotplug.active_writer == current)
64 return;
65 mutex_lock(&cpu_hotplug.lock);
66 cpu_hotplug.refcount++;
67 mutex_unlock(&cpu_hotplug.lock);
68
69}
70EXPORT_SYMBOL_GPL(get_online_cpus);
71
72void put_online_cpus(void)
73{
74 if (cpu_hotplug.active_writer == current)
75 return;
76 mutex_lock(&cpu_hotplug.lock);
77 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
78 wake_up_process(cpu_hotplug.active_writer);
79 mutex_unlock(&cpu_hotplug.lock);
80
81}
82EXPORT_SYMBOL_GPL(put_online_cpus);
83
84/*
85 * This ensures that the hotplug operation can begin only when the
86 * refcount goes to zero.
87 *
88 * Note that during a cpu-hotplug operation, the new readers, if any,
89 * will be blocked by the cpu_hotplug.lock
90 *
91 * Since cpu_hotplug_begin() is always called after invoking
92 * cpu_maps_update_begin(), we can be sure that only one writer is active.
93 *
94 * Note that theoretically, there is a possibility of a livelock:
95 * - Refcount goes to zero, last reader wakes up the sleeping
96 * writer.
97 * - Last reader unlocks the cpu_hotplug.lock.
98 * - A new reader arrives at this moment, bumps up the refcount.
99 * - The writer acquires the cpu_hotplug.lock finds the refcount
100 * non zero and goes to sleep again.
101 *
102 * However, this is very difficult to achieve in practice since
103 * get_online_cpus() not an api which is called all that often.
104 *
105 */
106static void cpu_hotplug_begin(void)
107{
108 cpu_hotplug.active_writer = current;
109
110 for (;;) {
111 mutex_lock(&cpu_hotplug.lock);
112 if (likely(!cpu_hotplug.refcount))
113 break;
114 __set_current_state(TASK_UNINTERRUPTIBLE);
115 mutex_unlock(&cpu_hotplug.lock);
116 schedule();
117 }
118}
119
120static void cpu_hotplug_done(void)
121{
122 cpu_hotplug.active_writer = NULL;
123 mutex_unlock(&cpu_hotplug.lock);
124}
125
126#else /* #if CONFIG_HOTPLUG_CPU */
127static void cpu_hotplug_begin(void) {}
128static void cpu_hotplug_done(void) {}
129#endif /* #else #if CONFIG_HOTPLUG_CPU */
130
131/* Need to know about CPUs going up/down? */
132int __ref register_cpu_notifier(struct notifier_block *nb)
133{
134 int ret;
135 cpu_maps_update_begin();
136 ret = raw_notifier_chain_register(&cpu_chain, nb);
137 cpu_maps_update_done();
138 return ret;
139}
140
141static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
142 int *nr_calls)
143{
144 int ret;
145
146 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
147 nr_calls);
148
149 return notifier_to_errno(ret);
150}
151
152static int cpu_notify(unsigned long val, void *v)
153{
154 return __cpu_notify(val, v, -1, NULL);
155}
156
157#ifdef CONFIG_HOTPLUG_CPU
158
159static void cpu_notify_nofail(unsigned long val, void *v)
160{
161 BUG_ON(cpu_notify(val, v));
162}
163EXPORT_SYMBOL(register_cpu_notifier);
164
165void __ref unregister_cpu_notifier(struct notifier_block *nb)
166{
167 cpu_maps_update_begin();
168 raw_notifier_chain_unregister(&cpu_chain, nb);
169 cpu_maps_update_done();
170}
171EXPORT_SYMBOL(unregister_cpu_notifier);
172
173static inline void check_for_tasks(int cpu)
174{
175 struct task_struct *p;
176
177 write_lock_irq(&tasklist_lock);
178 for_each_process(p) {
179 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
180 (!cputime_eq(p->utime, cputime_zero) ||
181 !cputime_eq(p->stime, cputime_zero)))
182 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
183 "(state = %ld, flags = %x)\n",
184 p->comm, task_pid_nr(p), cpu,
185 p->state, p->flags);
186 }
187 write_unlock_irq(&tasklist_lock);
188}
189
190struct take_cpu_down_param {
191 unsigned long mod;
192 void *hcpu;
193};
194
195/* Take this CPU down. */
196static int __ref take_cpu_down(void *_param)
197{
198 struct take_cpu_down_param *param = _param;
199 int err;
200
201 /* Ensure this CPU doesn't handle any more interrupts. */
202 err = __cpu_disable();
203 if (err < 0)
204 return err;
205
206 cpu_notify(CPU_DYING | param->mod, param->hcpu);
207 return 0;
208}
209
210/* Requires cpu_add_remove_lock to be held */
211static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
212{
213 int err, nr_calls = 0;
214 void *hcpu = (void *)(long)cpu;
215 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
216 struct take_cpu_down_param tcd_param = {
217 .mod = mod,
218 .hcpu = hcpu,
219 };
220
221 if (num_online_cpus() == 1)
222 return -EBUSY;
223
224 if (!cpu_online(cpu))
225 return -EINVAL;
226
227 cpu_hotplug_begin();
228
229 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
230 if (err) {
231 nr_calls--;
232 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
233 printk("%s: attempt to take down CPU %u failed\n",
234 __func__, cpu);
235 goto out_release;
236 }
237
238 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
239 if (err) {
240 /* CPU didn't die: tell everyone. Can't complain. */
241 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
242
243 goto out_release;
244 }
245 BUG_ON(cpu_online(cpu));
246
247 /*
248 * The migration_call() CPU_DYING callback will have removed all
249 * runnable tasks from the cpu, there's only the idle task left now
250 * that the migration thread is done doing the stop_machine thing.
251 *
252 * Wait for the stop thread to go away.
253 */
254 while (!idle_cpu(cpu))
255 cpu_relax();
256
257 /* This actually kills the CPU. */
258 __cpu_die(cpu);
259
260 /* CPU is completely dead: tell everyone. Too late to complain. */
261 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
262
263 check_for_tasks(cpu);
264
265out_release:
266 cpu_hotplug_done();
267 if (!err)
268 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
269 return err;
270}
271
272int __ref cpu_down(unsigned int cpu)
273{
274 int err;
275
276 cpu_maps_update_begin();
277
278 if (cpu_hotplug_disabled) {
279 err = -EBUSY;
280 goto out;
281 }
282
283 err = _cpu_down(cpu, 0);
284
285out:
286 cpu_maps_update_done();
287 return err;
288}
289EXPORT_SYMBOL(cpu_down);
290#endif /*CONFIG_HOTPLUG_CPU*/
291
292/* Requires cpu_add_remove_lock to be held */
293static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
294{
295 int ret, nr_calls = 0;
296 void *hcpu = (void *)(long)cpu;
297 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
298
299 if (cpu_online(cpu) || !cpu_present(cpu))
300 return -EINVAL;
301
302 cpu_hotplug_begin();
303 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
304 if (ret) {
305 nr_calls--;
306 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
307 __func__, cpu);
308 goto out_notify;
309 }
310
311 /* Arch-specific enabling code. */
312 ret = __cpu_up(cpu);
313 if (ret != 0)
314 goto out_notify;
315 BUG_ON(!cpu_online(cpu));
316
317 /* Now call notifier in preparation. */
318 cpu_notify(CPU_ONLINE | mod, hcpu);
319
320out_notify:
321 if (ret != 0)
322 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
323 cpu_hotplug_done();
324
325 return ret;
326}
327
328int __cpuinit cpu_up(unsigned int cpu)
329{
330 int err = 0;
331
332#ifdef CONFIG_MEMORY_HOTPLUG
333 int nid;
334 pg_data_t *pgdat;
335#endif
336
337 if (!cpu_possible(cpu)) {
338 printk(KERN_ERR "can't online cpu %d because it is not "
339 "configured as may-hotadd at boot time\n", cpu);
340#if defined(CONFIG_IA64)
341 printk(KERN_ERR "please check additional_cpus= boot "
342 "parameter\n");
343#endif
344 return -EINVAL;
345 }
346
347#ifdef CONFIG_MEMORY_HOTPLUG
348 nid = cpu_to_node(cpu);
349 if (!node_online(nid)) {
350 err = mem_online_node(nid);
351 if (err)
352 return err;
353 }
354
355 pgdat = NODE_DATA(nid);
356 if (!pgdat) {
357 printk(KERN_ERR
358 "Can't online cpu %d due to NULL pgdat\n", cpu);
359 return -ENOMEM;
360 }
361
362 if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
363 mutex_lock(&zonelists_mutex);
364 build_all_zonelists(NULL);
365 mutex_unlock(&zonelists_mutex);
366 }
367#endif
368
369 cpu_maps_update_begin();
370
371 if (cpu_hotplug_disabled) {
372 err = -EBUSY;
373 goto out;
374 }
375
376 err = _cpu_up(cpu, 0);
377
378out:
379 cpu_maps_update_done();
380 return err;
381}
382
383#ifdef CONFIG_PM_SLEEP_SMP
384static cpumask_var_t frozen_cpus;
385
386void __weak arch_disable_nonboot_cpus_begin(void)
387{
388}
389
390void __weak arch_disable_nonboot_cpus_end(void)
391{
392}
393
394int disable_nonboot_cpus(void)
395{
396 int cpu, first_cpu, error = 0;
397
398 cpu_maps_update_begin();
399 first_cpu = cpumask_first(cpu_online_mask);
400 /*
401 * We take down all of the non-boot CPUs in one shot to avoid races
402 * with the userspace trying to use the CPU hotplug at the same time
403 */
404 cpumask_clear(frozen_cpus);
405 arch_disable_nonboot_cpus_begin();
406
407 printk("Disabling non-boot CPUs ...\n");
408 for_each_online_cpu(cpu) {
409 if (cpu == first_cpu)
410 continue;
411 error = _cpu_down(cpu, 1);
412 if (!error)
413 cpumask_set_cpu(cpu, frozen_cpus);
414 else {
415 printk(KERN_ERR "Error taking CPU%d down: %d\n",
416 cpu, error);
417 break;
418 }
419 }
420
421 arch_disable_nonboot_cpus_end();
422
423 if (!error) {
424 BUG_ON(num_online_cpus() > 1);
425 /* Make sure the CPUs won't be enabled by someone else */
426 cpu_hotplug_disabled = 1;
427 } else {
428 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
429 }
430 cpu_maps_update_done();
431 return error;
432}
433
434void __weak arch_enable_nonboot_cpus_begin(void)
435{
436}
437
438void __weak arch_enable_nonboot_cpus_end(void)
439{
440}
441
442void __ref enable_nonboot_cpus(void)
443{
444 int cpu, error;
445
446 /* Allow everyone to use the CPU hotplug again */
447 cpu_maps_update_begin();
448 cpu_hotplug_disabled = 0;
449 if (cpumask_empty(frozen_cpus))
450 goto out;
451
452 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
453
454 arch_enable_nonboot_cpus_begin();
455
456 for_each_cpu(cpu, frozen_cpus) {
457 error = _cpu_up(cpu, 1);
458 if (!error) {
459 printk(KERN_INFO "CPU%d is up\n", cpu);
460 continue;
461 }
462 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
463 }
464
465 arch_enable_nonboot_cpus_end();
466
467 cpumask_clear(frozen_cpus);
468out:
469 cpu_maps_update_done();
470}
471
472static int alloc_frozen_cpus(void)
473{
474 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
475 return -ENOMEM;
476 return 0;
477}
478core_initcall(alloc_frozen_cpus);
479#endif /* CONFIG_PM_SLEEP_SMP */
480
481/**
482 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
483 * @cpu: cpu that just started
484 *
485 * This function calls the cpu_chain notifiers with CPU_STARTING.
486 * It must be called by the arch code on the new cpu, before the new cpu
487 * enables interrupts and before the "boot" cpu returns from __cpu_up().
488 */
489void __cpuinit notify_cpu_starting(unsigned int cpu)
490{
491 unsigned long val = CPU_STARTING;
492
493#ifdef CONFIG_PM_SLEEP_SMP
494 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
495 val = CPU_STARTING_FROZEN;
496#endif /* CONFIG_PM_SLEEP_SMP */
497 cpu_notify(val, (void *)(long)cpu);
498}
499
500#endif /* CONFIG_SMP */
501
502/*
503 * cpu_bit_bitmap[] is a special, "compressed" data structure that
504 * represents all NR_CPUS bits binary values of 1<<nr.
505 *
506 * It is used by cpumask_of() to get a constant address to a CPU
507 * mask value that has a single bit set only.
508 */
509
510/* cpu_bit_bitmap[0] is empty - so we can back into it */
511#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
512#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
513#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
514#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
515
516const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
517
518 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
519 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
520#if BITS_PER_LONG > 32
521 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
522 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
523#endif
524};
525EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
526
527const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
528EXPORT_SYMBOL(cpu_all_bits);
529
530#ifdef CONFIG_INIT_ALL_POSSIBLE
531static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
532 = CPU_BITS_ALL;
533#else
534static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
535#endif
536const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
537EXPORT_SYMBOL(cpu_possible_mask);
538
539static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
540const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
541EXPORT_SYMBOL(cpu_online_mask);
542
543static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
544const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
545EXPORT_SYMBOL(cpu_present_mask);
546
547static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
548const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
549EXPORT_SYMBOL(cpu_active_mask);
550
551void set_cpu_possible(unsigned int cpu, bool possible)
552{
553 if (possible)
554 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
555 else
556 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
557}
558
559void set_cpu_present(unsigned int cpu, bool present)
560{
561 if (present)
562 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
563 else
564 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
565}
566
567void set_cpu_online(unsigned int cpu, bool online)
568{
569 if (online)
570 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
571 else
572 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
573}
574
575void set_cpu_active(unsigned int cpu, bool active)
576{
577 if (active)
578 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
579 else
580 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
581}
582
583void init_cpu_present(const struct cpumask *src)
584{
585 cpumask_copy(to_cpumask(cpu_present_bits), src);
586}
587
588void init_cpu_possible(const struct cpumask *src)
589{
590 cpumask_copy(to_cpumask(cpu_possible_bits), src);
591}
592
593void init_cpu_online(const struct cpumask *src)
594{
595 cpumask_copy(to_cpumask(cpu_online_bits), src);
596}