projects/linux·blob v3.13-rc5 · refs

722 lines · 17.1 KiB

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
23#include "smpboot.h"
24
25#ifdef CONFIG_SMP
26/* Serializes the updates to cpu_online_mask, cpu_present_mask */
27static DEFINE_MUTEX(cpu_add_remove_lock);
28
29/*
30 * The following two API's must be used when attempting
31 * to serialize the updates to cpu_online_mask, cpu_present_mask.
32 */
33void cpu_maps_update_begin(void)
34{
35 mutex_lock(&cpu_add_remove_lock);
36}
37
38void cpu_maps_update_done(void)
39{
40 mutex_unlock(&cpu_add_remove_lock);
41}
42
43static RAW_NOTIFIER_HEAD(cpu_chain);
44
45/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
46 * Should always be manipulated under cpu_add_remove_lock
47 */
48static int cpu_hotplug_disabled;
49
50#ifdef CONFIG_HOTPLUG_CPU
51
52static struct {
53 struct task_struct *active_writer;
54 struct mutex lock; /* Synchronizes accesses to refcount, */
55 /*
56 * Also blocks the new readers during
57 * an ongoing cpu hotplug operation.
58 */
59 int refcount;
60} cpu_hotplug = {
61 .active_writer = NULL,
62 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
63 .refcount = 0,
64};
65
66void get_online_cpus(void)
67{
68 might_sleep();
69 if (cpu_hotplug.active_writer == current)
70 return;
71 mutex_lock(&cpu_hotplug.lock);
72 cpu_hotplug.refcount++;
73 mutex_unlock(&cpu_hotplug.lock);
74
75}
76EXPORT_SYMBOL_GPL(get_online_cpus);
77
78void put_online_cpus(void)
79{
80 if (cpu_hotplug.active_writer == current)
81 return;
82 mutex_lock(&cpu_hotplug.lock);
83
84 if (WARN_ON(!cpu_hotplug.refcount))
85 cpu_hotplug.refcount++; /* try to fix things up */
86
87 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
88 wake_up_process(cpu_hotplug.active_writer);
89 mutex_unlock(&cpu_hotplug.lock);
90
91}
92EXPORT_SYMBOL_GPL(put_online_cpus);
93
94/*
95 * This ensures that the hotplug operation can begin only when the
96 * refcount goes to zero.
97 *
98 * Note that during a cpu-hotplug operation, the new readers, if any,
99 * will be blocked by the cpu_hotplug.lock
100 *
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.
103 *
104 * Note that theoretically, there is a possibility of a livelock:
105 * - Refcount goes to zero, last reader wakes up the sleeping
106 * writer.
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.
111 *
112 * However, this is very difficult to achieve in practice since
113 * get_online_cpus() not an api which is called all that often.
114 *
115 */
116void cpu_hotplug_begin(void)
117{
118 cpu_hotplug.active_writer = current;
119
120 for (;;) {
121 mutex_lock(&cpu_hotplug.lock);
122 if (likely(!cpu_hotplug.refcount))
123 break;
124 __set_current_state(TASK_UNINTERRUPTIBLE);
125 mutex_unlock(&cpu_hotplug.lock);
126 schedule();
127 }
128}
129
130void cpu_hotplug_done(void)
131{
132 cpu_hotplug.active_writer = NULL;
133 mutex_unlock(&cpu_hotplug.lock);
134}
135
136/*
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.
142 */
143void cpu_hotplug_disable(void)
144{
145 cpu_maps_update_begin();
146 cpu_hotplug_disabled = 1;
147 cpu_maps_update_done();
148}
149
150void cpu_hotplug_enable(void)
151{
152 cpu_maps_update_begin();
153 cpu_hotplug_disabled = 0;
154 cpu_maps_update_done();
155}
156
157#endif /* CONFIG_HOTPLUG_CPU */
158
159/* Need to know about CPUs going up/down? */
160int __ref register_cpu_notifier(struct notifier_block *nb)
161{
162 int ret;
163 cpu_maps_update_begin();
164 ret = raw_notifier_chain_register(&cpu_chain, nb);
165 cpu_maps_update_done();
166 return ret;
167}
168
169static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
170 int *nr_calls)
171{
172 int ret;
173
174 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
175 nr_calls);
176
177 return notifier_to_errno(ret);
178}
179
180static int cpu_notify(unsigned long val, void *v)
181{
182 return __cpu_notify(val, v, -1, NULL);
183}
184
185#ifdef CONFIG_HOTPLUG_CPU
186
187static void cpu_notify_nofail(unsigned long val, void *v)
188{
189 BUG_ON(cpu_notify(val, v));
190}
191EXPORT_SYMBOL(register_cpu_notifier);
192
193void __ref unregister_cpu_notifier(struct notifier_block *nb)
194{
195 cpu_maps_update_begin();
196 raw_notifier_chain_unregister(&cpu_chain, nb);
197 cpu_maps_update_done();
198}
199EXPORT_SYMBOL(unregister_cpu_notifier);
200
201/**
202 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
203 * @cpu: a CPU id
204 *
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.
209 *
210 * Also note that the function uses a somewhat relaxed locking scheme, so it may
211 * be called only for an already offlined CPU.
212 */
213void clear_tasks_mm_cpumask(int cpu)
214{
215 struct task_struct *p;
216
217 /*
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.
223 */
224 WARN_ON(cpu_online(cpu));
225 rcu_read_lock();
226 for_each_process(p) {
227 struct task_struct *t;
228
229 /*
230 * Main thread might exit, but other threads may still have
231 * a valid mm. Find one.
232 */
233 t = find_lock_task_mm(p);
234 if (!t)
235 continue;
236 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
237 task_unlock(t);
238 }
239 rcu_read_unlock();
240}
241
242static inline void check_for_tasks(int cpu)
243{
244 struct task_struct *p;
245 cputime_t utime, stime;
246
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 &&
251 (utime || stime))
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,
255 p->state, p->flags);
256 }
257 write_unlock_irq(&tasklist_lock);
258}
259
260struct take_cpu_down_param {
261 unsigned long mod;
262 void *hcpu;
263};
264
265/* Take this CPU down. */
266static int __ref take_cpu_down(void *_param)
267{
268 struct take_cpu_down_param *param = _param;
269 int err;
270
271 /* Ensure this CPU doesn't handle any more interrupts. */
272 err = __cpu_disable();
273 if (err < 0)
274 return err;
275
276 cpu_notify(CPU_DYING | param->mod, param->hcpu);
277 /* Park the stopper thread */
278 kthread_park(current);
279 return 0;
280}
281
282/* Requires cpu_add_remove_lock to be held */
283static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
284{
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 = {
289 .mod = mod,
290 .hcpu = hcpu,
291 };
292
293 if (num_online_cpus() == 1)
294 return -EBUSY;
295
296 if (!cpu_online(cpu))
297 return -EINVAL;
298
299 cpu_hotplug_begin();
300
301 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
302 if (err) {
303 nr_calls--;
304 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
305 printk("%s: attempt to take down CPU %u failed\n",
306 __func__, cpu);
307 goto out_release;
308 }
309
310 /*
311 * By now we've cleared cpu_active_mask, wait for all preempt-disabled
312 * and RCU users of this state to go away such that all new such users
313 * will observe it.
314 *
315 * For CONFIG_PREEMPT we have preemptible RCU and its sync_rcu() might
316 * not imply sync_sched(), so explicitly call both.
317 *
318 * Do sync before park smpboot threads to take care the rcu boost case.
319 */
320#ifdef CONFIG_PREEMPT
321 synchronize_sched();
322#endif
323 synchronize_rcu();
324
325 smpboot_park_threads(cpu);
326
327 /*
328 * So now all preempt/rcu users must observe !cpu_active().
329 */
330
331 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
332 if (err) {
333 /* CPU didn't die: tell everyone. Can't complain. */
334 smpboot_unpark_threads(cpu);
335 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
336 goto out_release;
337 }
338 BUG_ON(cpu_online(cpu));
339
340 /*
341 * The migration_call() CPU_DYING callback will have removed all
342 * runnable tasks from the cpu, there's only the idle task left now
343 * that the migration thread is done doing the stop_machine thing.
344 *
345 * Wait for the stop thread to go away.
346 */
347 while (!idle_cpu(cpu))
348 cpu_relax();
349
350 /* This actually kills the CPU. */
351 __cpu_die(cpu);
352
353 /* CPU is completely dead: tell everyone. Too late to complain. */
354 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
355
356 check_for_tasks(cpu);
357
358out_release:
359 cpu_hotplug_done();
360 if (!err)
361 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
362 return err;
363}
364
365int __ref cpu_down(unsigned int cpu)
366{
367 int err;
368
369 cpu_maps_update_begin();
370
371 if (cpu_hotplug_disabled) {
372 err = -EBUSY;
373 goto out;
374 }
375
376 err = _cpu_down(cpu, 0);
377
378out:
379 cpu_maps_update_done();
380 return err;
381}
382EXPORT_SYMBOL(cpu_down);
383#endif /*CONFIG_HOTPLUG_CPU*/
384
385/* Requires cpu_add_remove_lock to be held */
386static int _cpu_up(unsigned int cpu, int tasks_frozen)
387{
388 int ret, nr_calls = 0;
389 void *hcpu = (void *)(long)cpu;
390 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
391 struct task_struct *idle;
392
393 cpu_hotplug_begin();
394
395 if (cpu_online(cpu) || !cpu_present(cpu)) {
396 ret = -EINVAL;
397 goto out;
398 }
399
400 idle = idle_thread_get(cpu);
401 if (IS_ERR(idle)) {
402 ret = PTR_ERR(idle);
403 goto out;
404 }
405
406 ret = smpboot_create_threads(cpu);
407 if (ret)
408 goto out;
409
410 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
411 if (ret) {
412 nr_calls--;
413 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
414 __func__, cpu);
415 goto out_notify;
416 }
417
418 /* Arch-specific enabling code. */
419 ret = __cpu_up(cpu, idle);
420 if (ret != 0)
421 goto out_notify;
422 BUG_ON(!cpu_online(cpu));
423
424 /* Wake the per cpu threads */
425 smpboot_unpark_threads(cpu);
426
427 /* Now call notifier in preparation. */
428 cpu_notify(CPU_ONLINE | mod, hcpu);
429
430out_notify:
431 if (ret != 0)
432 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
433out:
434 cpu_hotplug_done();
435
436 return ret;
437}
438
439int cpu_up(unsigned int cpu)
440{
441 int err = 0;
442
443 if (!cpu_possible(cpu)) {
444 printk(KERN_ERR "can't online cpu %d because it is not "
445 "configured as may-hotadd at boot time\n", cpu);
446#if defined(CONFIG_IA64)
447 printk(KERN_ERR "please check additional_cpus= boot "
448 "parameter\n");
449#endif
450 return -EINVAL;
451 }
452
453 err = try_online_node(cpu_to_node(cpu));
454 if (err)
455 return err;
456
457 cpu_maps_update_begin();
458
459 if (cpu_hotplug_disabled) {
460 err = -EBUSY;
461 goto out;
462 }
463
464 err = _cpu_up(cpu, 0);
465
466out:
467 cpu_maps_update_done();
468 return err;
469}
470EXPORT_SYMBOL_GPL(cpu_up);
471
472#ifdef CONFIG_PM_SLEEP_SMP
473static cpumask_var_t frozen_cpus;
474
475int disable_nonboot_cpus(void)
476{
477 int cpu, first_cpu, error = 0;
478
479 cpu_maps_update_begin();
480 first_cpu = cpumask_first(cpu_online_mask);
481 /*
482 * We take down all of the non-boot CPUs in one shot to avoid races
483 * with the userspace trying to use the CPU hotplug at the same time
484 */
485 cpumask_clear(frozen_cpus);
486
487 printk("Disabling non-boot CPUs ...\n");
488 for_each_online_cpu(cpu) {
489 if (cpu == first_cpu)
490 continue;
491 error = _cpu_down(cpu, 1);
492 if (!error)
493 cpumask_set_cpu(cpu, frozen_cpus);
494 else {
495 printk(KERN_ERR "Error taking CPU%d down: %d\n",
496 cpu, error);
497 break;
498 }
499 }
500
501 if (!error) {
502 BUG_ON(num_online_cpus() > 1);
503 /* Make sure the CPUs won't be enabled by someone else */
504 cpu_hotplug_disabled = 1;
505 } else {
506 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
507 }
508 cpu_maps_update_done();
509 return error;
510}
511
512void __weak arch_enable_nonboot_cpus_begin(void)
513{
514}
515
516void __weak arch_enable_nonboot_cpus_end(void)
517{
518}
519
520void __ref enable_nonboot_cpus(void)
521{
522 int cpu, error;
523
524 /* Allow everyone to use the CPU hotplug again */
525 cpu_maps_update_begin();
526 cpu_hotplug_disabled = 0;
527 if (cpumask_empty(frozen_cpus))
528 goto out;
529
530 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
531
532 arch_enable_nonboot_cpus_begin();
533
534 for_each_cpu(cpu, frozen_cpus) {
535 error = _cpu_up(cpu, 1);
536 if (!error) {
537 printk(KERN_INFO "CPU%d is up\n", cpu);
538 continue;
539 }
540 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
541 }
542
543 arch_enable_nonboot_cpus_end();
544
545 cpumask_clear(frozen_cpus);
546out:
547 cpu_maps_update_done();
548}
549
550static int __init alloc_frozen_cpus(void)
551{
552 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
553 return -ENOMEM;
554 return 0;
555}
556core_initcall(alloc_frozen_cpus);
557
558/*
559 * When callbacks for CPU hotplug notifications are being executed, we must
560 * ensure that the state of the system with respect to the tasks being frozen
561 * or not, as reported by the notification, remains unchanged *throughout the
562 * duration* of the execution of the callbacks.
563 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
564 *
565 * This synchronization is implemented by mutually excluding regular CPU
566 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
567 * Hibernate notifications.
568 */
569static int
570cpu_hotplug_pm_callback(struct notifier_block *nb,
571 unsigned long action, void *ptr)
572{
573 switch (action) {
574
575 case PM_SUSPEND_PREPARE:
576 case PM_HIBERNATION_PREPARE:
577 cpu_hotplug_disable();
578 break;
579
580 case PM_POST_SUSPEND:
581 case PM_POST_HIBERNATION:
582 cpu_hotplug_enable();
583 break;
584
585 default:
586 return NOTIFY_DONE;
587 }
588
589 return NOTIFY_OK;
590}
591
592
593static int __init cpu_hotplug_pm_sync_init(void)
594{
595 /*
596 * cpu_hotplug_pm_callback has higher priority than x86
597 * bsp_pm_callback which depends on cpu_hotplug_pm_callback
598 * to disable cpu hotplug to avoid cpu hotplug race.
599 */
600 pm_notifier(cpu_hotplug_pm_callback, 0);
601 return 0;
602}
603core_initcall(cpu_hotplug_pm_sync_init);
604
605#endif /* CONFIG_PM_SLEEP_SMP */
606
607/**
608 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
609 * @cpu: cpu that just started
610 *
611 * This function calls the cpu_chain notifiers with CPU_STARTING.
612 * It must be called by the arch code on the new cpu, before the new cpu
613 * enables interrupts and before the "boot" cpu returns from __cpu_up().
614 */
615void notify_cpu_starting(unsigned int cpu)
616{
617 unsigned long val = CPU_STARTING;
618
619#ifdef CONFIG_PM_SLEEP_SMP
620 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
621 val = CPU_STARTING_FROZEN;
622#endif /* CONFIG_PM_SLEEP_SMP */
623 cpu_notify(val, (void *)(long)cpu);
624}
625
626#endif /* CONFIG_SMP */
627
628/*
629 * cpu_bit_bitmap[] is a special, "compressed" data structure that
630 * represents all NR_CPUS bits binary values of 1<<nr.
631 *
632 * It is used by cpumask_of() to get a constant address to a CPU
633 * mask value that has a single bit set only.
634 */
635
636/* cpu_bit_bitmap[0] is empty - so we can back into it */
637#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
638#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
639#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
640#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
641
642const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
643
644 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
645 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
646#if BITS_PER_LONG > 32
647 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
648 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
649#endif
650};
651EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
652
653const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
654EXPORT_SYMBOL(cpu_all_bits);
655
656#ifdef CONFIG_INIT_ALL_POSSIBLE
657static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
658 = CPU_BITS_ALL;
659#else
660static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
661#endif
662const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
663EXPORT_SYMBOL(cpu_possible_mask);
664
665static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
666const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
667EXPORT_SYMBOL(cpu_online_mask);
668
669static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
670const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
671EXPORT_SYMBOL(cpu_present_mask);
672
673static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
674const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
675EXPORT_SYMBOL(cpu_active_mask);
676
677void set_cpu_possible(unsigned int cpu, bool possible)
678{
679 if (possible)
680 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
681 else
682 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
683}
684
685void set_cpu_present(unsigned int cpu, bool present)
686{
687 if (present)
688 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
689 else
690 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
691}
692
693void set_cpu_online(unsigned int cpu, bool online)
694{
695 if (online)
696 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
697 else
698 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
699}
700
701void set_cpu_active(unsigned int cpu, bool active)
702{
703 if (active)
704 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
705 else
706 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
707}
708
709void init_cpu_present(const struct cpumask *src)
710{
711 cpumask_copy(to_cpumask(cpu_present_bits), src);
712}
713
714void init_cpu_possible(const struct cpumask *src)
715{
716 cpumask_copy(to_cpumask(cpu_possible_bits), src);
717}
718
719void init_cpu_online(const struct cpumask *src)
720{
721 cpumask_copy(to_cpumask(cpu_online_bits), src);
722}