projects/linux·blob v3.12-rc3 · refs

725 lines · 17.0 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 smpboot_park_threads(cpu);
310
311 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
312 if (err) {
313 /* CPU didn't die: tell everyone. Can't complain. */
314 smpboot_unpark_threads(cpu);
315 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
316 goto out_release;
317 }
318 BUG_ON(cpu_online(cpu));
319
320 /*
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.
324 *
325 * Wait for the stop thread to go away.
326 */
327 while (!idle_cpu(cpu))
328 cpu_relax();
329
330 /* This actually kills the CPU. */
331 __cpu_die(cpu);
332
333 /* CPU is completely dead: tell everyone. Too late to complain. */
334 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
335
336 check_for_tasks(cpu);
337
338out_release:
339 cpu_hotplug_done();
340 if (!err)
341 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
342 return err;
343}
344
345int __ref cpu_down(unsigned int cpu)
346{
347 int err;
348
349 cpu_maps_update_begin();
350
351 if (cpu_hotplug_disabled) {
352 err = -EBUSY;
353 goto out;
354 }
355
356 err = _cpu_down(cpu, 0);
357
358out:
359 cpu_maps_update_done();
360 return err;
361}
362EXPORT_SYMBOL(cpu_down);
363#endif /*CONFIG_HOTPLUG_CPU*/
364
365/* Requires cpu_add_remove_lock to be held */
366static int _cpu_up(unsigned int cpu, int tasks_frozen)
367{
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;
372
373 cpu_hotplug_begin();
374
375 if (cpu_online(cpu) || !cpu_present(cpu)) {
376 ret = -EINVAL;
377 goto out;
378 }
379
380 idle = idle_thread_get(cpu);
381 if (IS_ERR(idle)) {
382 ret = PTR_ERR(idle);
383 goto out;
384 }
385
386 ret = smpboot_create_threads(cpu);
387 if (ret)
388 goto out;
389
390 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
391 if (ret) {
392 nr_calls--;
393 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
394 __func__, cpu);
395 goto out_notify;
396 }
397
398 /* Arch-specific enabling code. */
399 ret = __cpu_up(cpu, idle);
400 if (ret != 0)
401 goto out_notify;
402 BUG_ON(!cpu_online(cpu));
403
404 /* Wake the per cpu threads */
405 smpboot_unpark_threads(cpu);
406
407 /* Now call notifier in preparation. */
408 cpu_notify(CPU_ONLINE | mod, hcpu);
409
410out_notify:
411 if (ret != 0)
412 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
413out:
414 cpu_hotplug_done();
415
416 return ret;
417}
418
419int cpu_up(unsigned int cpu)
420{
421 int err = 0;
422
423#ifdef CONFIG_MEMORY_HOTPLUG
424 int nid;
425 pg_data_t *pgdat;
426#endif
427
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 "
433 "parameter\n");
434#endif
435 return -EINVAL;
436 }
437
438#ifdef CONFIG_MEMORY_HOTPLUG
439 nid = cpu_to_node(cpu);
440 if (!node_online(nid)) {
441 err = mem_online_node(nid);
442 if (err)
443 return err;
444 }
445
446 pgdat = NODE_DATA(nid);
447 if (!pgdat) {
448 printk(KERN_ERR
449 "Can't online cpu %d due to NULL pgdat\n", cpu);
450 return -ENOMEM;
451 }
452
453 if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
454 mutex_lock(&zonelists_mutex);
455 build_all_zonelists(NULL, NULL);
456 mutex_unlock(&zonelists_mutex);
457 }
458#endif
459
460 cpu_maps_update_begin();
461
462 if (cpu_hotplug_disabled) {
463 err = -EBUSY;
464 goto out;
465 }
466
467 err = _cpu_up(cpu, 0);
468
469out:
470 cpu_maps_update_done();
471 return err;
472}
473EXPORT_SYMBOL_GPL(cpu_up);
474
475#ifdef CONFIG_PM_SLEEP_SMP
476static cpumask_var_t frozen_cpus;
477
478int disable_nonboot_cpus(void)
479{
480 int cpu, first_cpu, error = 0;
481
482 cpu_maps_update_begin();
483 first_cpu = cpumask_first(cpu_online_mask);
484 /*
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
487 */
488 cpumask_clear(frozen_cpus);
489
490 printk("Disabling non-boot CPUs ...\n");
491 for_each_online_cpu(cpu) {
492 if (cpu == first_cpu)
493 continue;
494 error = _cpu_down(cpu, 1);
495 if (!error)
496 cpumask_set_cpu(cpu, frozen_cpus);
497 else {
498 printk(KERN_ERR "Error taking CPU%d down: %d\n",
499 cpu, error);
500 break;
501 }
502 }
503
504 if (!error) {
505 BUG_ON(num_online_cpus() > 1);
506 /* Make sure the CPUs won't be enabled by someone else */
507 cpu_hotplug_disabled = 1;
508 } else {
509 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
510 }
511 cpu_maps_update_done();
512 return error;
513}
514
515void __weak arch_enable_nonboot_cpus_begin(void)
516{
517}
518
519void __weak arch_enable_nonboot_cpus_end(void)
520{
521}
522
523void __ref enable_nonboot_cpus(void)
524{
525 int cpu, error;
526
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))
531 goto out;
532
533 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
534
535 arch_enable_nonboot_cpus_begin();
536
537 for_each_cpu(cpu, frozen_cpus) {
538 error = _cpu_up(cpu, 1);
539 if (!error) {
540 printk(KERN_INFO "CPU%d is up\n", cpu);
541 continue;
542 }
543 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
544 }
545
546 arch_enable_nonboot_cpus_end();
547
548 cpumask_clear(frozen_cpus);
549out:
550 cpu_maps_update_done();
551}
552
553static int __init alloc_frozen_cpus(void)
554{
555 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
556 return -ENOMEM;
557 return 0;
558}
559core_initcall(alloc_frozen_cpus);
560
561/*
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.
567 *
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.
571 */
572static int
573cpu_hotplug_pm_callback(struct notifier_block *nb,
574 unsigned long action, void *ptr)
575{
576 switch (action) {
577
578 case PM_SUSPEND_PREPARE:
579 case PM_HIBERNATION_PREPARE:
580 cpu_hotplug_disable();
581 break;
582
583 case PM_POST_SUSPEND:
584 case PM_POST_HIBERNATION:
585 cpu_hotplug_enable();
586 break;
587
588 default:
589 return NOTIFY_DONE;
590 }
591
592 return NOTIFY_OK;
593}
594
595
596static int __init cpu_hotplug_pm_sync_init(void)
597{
598 /*
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.
602 */
603 pm_notifier(cpu_hotplug_pm_callback, 0);
604 return 0;
605}
606core_initcall(cpu_hotplug_pm_sync_init);
607
608#endif /* CONFIG_PM_SLEEP_SMP */
609
610/**
611 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
612 * @cpu: cpu that just started
613 *
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().
617 */
618void notify_cpu_starting(unsigned int cpu)
619{
620 unsigned long val = CPU_STARTING;
621
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);
627}
628
629#endif /* CONFIG_SMP */
630
631/*
632 * cpu_bit_bitmap[] is a special, "compressed" data structure that
633 * represents all NR_CPUS bits binary values of 1<<nr.
634 *
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.
637 */
638
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)
644
645const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
646
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),
652#endif
653};
654EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
655
656const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
657EXPORT_SYMBOL(cpu_all_bits);
658
659#ifdef CONFIG_INIT_ALL_POSSIBLE
660static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
661 = CPU_BITS_ALL;
662#else
663static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
664#endif
665const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
666EXPORT_SYMBOL(cpu_possible_mask);
667
668static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
669const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
670EXPORT_SYMBOL(cpu_online_mask);
671
672static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
673const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
674EXPORT_SYMBOL(cpu_present_mask);
675
676static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
677const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
678EXPORT_SYMBOL(cpu_active_mask);
679
680void set_cpu_possible(unsigned int cpu, bool possible)
681{
682 if (possible)
683 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
684 else
685 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
686}
687
688void set_cpu_present(unsigned int cpu, bool present)
689{
690 if (present)
691 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
692 else
693 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
694}
695
696void set_cpu_online(unsigned int cpu, bool online)
697{
698 if (online)
699 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
700 else
701 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
702}
703
704void set_cpu_active(unsigned int cpu, bool active)
705{
706 if (active)
707 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
708 else
709 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
710}
711
712void init_cpu_present(const struct cpumask *src)
713{
714 cpumask_copy(to_cpumask(cpu_present_bits), src);
715}
716
717void init_cpu_possible(const struct cpumask *src)
718{
719 cpumask_copy(to_cpumask(cpu_possible_bits), src);
720}
721
722void init_cpu_online(const struct cpumask *src)
723{
724 cpumask_copy(to_cpumask(cpu_online_bits), src);
725}