projects/linux·blob v3.7-rc2 · refs

726 lines · 17.2 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 */
116static void 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
130static void cpu_hotplug_done(void)
131{
132 cpu_hotplug.active_writer = NULL;
133 mutex_unlock(&cpu_hotplug.lock);
134}
135
136#else /* #if CONFIG_HOTPLUG_CPU */
137static void cpu_hotplug_begin(void) {}
138static void cpu_hotplug_done(void) {}
139#endif /* #else #if CONFIG_HOTPLUG_CPU */
140
141/* Need to know about CPUs going up/down? */
142int __ref register_cpu_notifier(struct notifier_block *nb)
143{
144 int ret;
145 cpu_maps_update_begin();
146 ret = raw_notifier_chain_register(&cpu_chain, nb);
147 cpu_maps_update_done();
148 return ret;
149}
150
151static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
152 int *nr_calls)
153{
154 int ret;
155
156 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
157 nr_calls);
158
159 return notifier_to_errno(ret);
160}
161
162static int cpu_notify(unsigned long val, void *v)
163{
164 return __cpu_notify(val, v, -1, NULL);
165}
166
167#ifdef CONFIG_HOTPLUG_CPU
168
169static void cpu_notify_nofail(unsigned long val, void *v)
170{
171 BUG_ON(cpu_notify(val, v));
172}
173EXPORT_SYMBOL(register_cpu_notifier);
174
175void __ref unregister_cpu_notifier(struct notifier_block *nb)
176{
177 cpu_maps_update_begin();
178 raw_notifier_chain_unregister(&cpu_chain, nb);
179 cpu_maps_update_done();
180}
181EXPORT_SYMBOL(unregister_cpu_notifier);
182
183/**
184 * clear_tasks_mm_cpumask - Safely clear tasks' mm_cpumask for a CPU
185 * @cpu: a CPU id
186 *
187 * This function walks all processes, finds a valid mm struct for each one and
188 * then clears a corresponding bit in mm's cpumask. While this all sounds
189 * trivial, there are various non-obvious corner cases, which this function
190 * tries to solve in a safe manner.
191 *
192 * Also note that the function uses a somewhat relaxed locking scheme, so it may
193 * be called only for an already offlined CPU.
194 */
195void clear_tasks_mm_cpumask(int cpu)
196{
197 struct task_struct *p;
198
199 /*
200 * This function is called after the cpu is taken down and marked
201 * offline, so its not like new tasks will ever get this cpu set in
202 * their mm mask. -- Peter Zijlstra
203 * Thus, we may use rcu_read_lock() here, instead of grabbing
204 * full-fledged tasklist_lock.
205 */
206 WARN_ON(cpu_online(cpu));
207 rcu_read_lock();
208 for_each_process(p) {
209 struct task_struct *t;
210
211 /*
212 * Main thread might exit, but other threads may still have
213 * a valid mm. Find one.
214 */
215 t = find_lock_task_mm(p);
216 if (!t)
217 continue;
218 cpumask_clear_cpu(cpu, mm_cpumask(t->mm));
219 task_unlock(t);
220 }
221 rcu_read_unlock();
222}
223
224static inline void check_for_tasks(int cpu)
225{
226 struct task_struct *p;
227
228 write_lock_irq(&tasklist_lock);
229 for_each_process(p) {
230 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
231 (p->utime || p->stime))
232 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
233 "(state = %ld, flags = %x)\n",
234 p->comm, task_pid_nr(p), cpu,
235 p->state, p->flags);
236 }
237 write_unlock_irq(&tasklist_lock);
238}
239
240struct take_cpu_down_param {
241 unsigned long mod;
242 void *hcpu;
243};
244
245/* Take this CPU down. */
246static int __ref take_cpu_down(void *_param)
247{
248 struct take_cpu_down_param *param = _param;
249 int err;
250
251 /* Ensure this CPU doesn't handle any more interrupts. */
252 err = __cpu_disable();
253 if (err < 0)
254 return err;
255
256 cpu_notify(CPU_DYING | param->mod, param->hcpu);
257 return 0;
258}
259
260/* Requires cpu_add_remove_lock to be held */
261static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
262{
263 int err, nr_calls = 0;
264 void *hcpu = (void *)(long)cpu;
265 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
266 struct take_cpu_down_param tcd_param = {
267 .mod = mod,
268 .hcpu = hcpu,
269 };
270
271 if (num_online_cpus() == 1)
272 return -EBUSY;
273
274 if (!cpu_online(cpu))
275 return -EINVAL;
276
277 cpu_hotplug_begin();
278
279 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
280 if (err) {
281 nr_calls--;
282 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
283 printk("%s: attempt to take down CPU %u failed\n",
284 __func__, cpu);
285 goto out_release;
286 }
287 smpboot_park_threads(cpu);
288
289 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
290 if (err) {
291 /* CPU didn't die: tell everyone. Can't complain. */
292 smpboot_unpark_threads(cpu);
293 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
294 goto out_release;
295 }
296 BUG_ON(cpu_online(cpu));
297
298 /*
299 * The migration_call() CPU_DYING callback will have removed all
300 * runnable tasks from the cpu, there's only the idle task left now
301 * that the migration thread is done doing the stop_machine thing.
302 *
303 * Wait for the stop thread to go away.
304 */
305 while (!idle_cpu(cpu))
306 cpu_relax();
307
308 /* This actually kills the CPU. */
309 __cpu_die(cpu);
310
311 /* CPU is completely dead: tell everyone. Too late to complain. */
312 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
313
314 check_for_tasks(cpu);
315
316out_release:
317 cpu_hotplug_done();
318 if (!err)
319 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
320 return err;
321}
322
323int __ref cpu_down(unsigned int cpu)
324{
325 int err;
326
327 cpu_maps_update_begin();
328
329 if (cpu_hotplug_disabled) {
330 err = -EBUSY;
331 goto out;
332 }
333
334 err = _cpu_down(cpu, 0);
335
336out:
337 cpu_maps_update_done();
338 return err;
339}
340EXPORT_SYMBOL(cpu_down);
341#endif /*CONFIG_HOTPLUG_CPU*/
342
343/* Requires cpu_add_remove_lock to be held */
344static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
345{
346 int ret, nr_calls = 0;
347 void *hcpu = (void *)(long)cpu;
348 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
349 struct task_struct *idle;
350
351 if (cpu_online(cpu) || !cpu_present(cpu))
352 return -EINVAL;
353
354 cpu_hotplug_begin();
355
356 idle = idle_thread_get(cpu);
357 if (IS_ERR(idle)) {
358 ret = PTR_ERR(idle);
359 goto out;
360 }
361
362 ret = smpboot_create_threads(cpu);
363 if (ret)
364 goto out;
365
366 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
367 if (ret) {
368 nr_calls--;
369 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
370 __func__, cpu);
371 goto out_notify;
372 }
373
374 /* Arch-specific enabling code. */
375 ret = __cpu_up(cpu, idle);
376 if (ret != 0)
377 goto out_notify;
378 BUG_ON(!cpu_online(cpu));
379
380 /* Wake the per cpu threads */
381 smpboot_unpark_threads(cpu);
382
383 /* Now call notifier in preparation. */
384 cpu_notify(CPU_ONLINE | mod, hcpu);
385
386out_notify:
387 if (ret != 0)
388 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
389out:
390 cpu_hotplug_done();
391
392 return ret;
393}
394
395int __cpuinit cpu_up(unsigned int cpu)
396{
397 int err = 0;
398
399#ifdef CONFIG_MEMORY_HOTPLUG
400 int nid;
401 pg_data_t *pgdat;
402#endif
403
404 if (!cpu_possible(cpu)) {
405 printk(KERN_ERR "can't online cpu %d because it is not "
406 "configured as may-hotadd at boot time\n", cpu);
407#if defined(CONFIG_IA64)
408 printk(KERN_ERR "please check additional_cpus= boot "
409 "parameter\n");
410#endif
411 return -EINVAL;
412 }
413
414#ifdef CONFIG_MEMORY_HOTPLUG
415 nid = cpu_to_node(cpu);
416 if (!node_online(nid)) {
417 err = mem_online_node(nid);
418 if (err)
419 return err;
420 }
421
422 pgdat = NODE_DATA(nid);
423 if (!pgdat) {
424 printk(KERN_ERR
425 "Can't online cpu %d due to NULL pgdat\n", cpu);
426 return -ENOMEM;
427 }
428
429 if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
430 mutex_lock(&zonelists_mutex);
431 build_all_zonelists(NULL, NULL);
432 mutex_unlock(&zonelists_mutex);
433 }
434#endif
435
436 cpu_maps_update_begin();
437
438 if (cpu_hotplug_disabled) {
439 err = -EBUSY;
440 goto out;
441 }
442
443 err = _cpu_up(cpu, 0);
444
445out:
446 cpu_maps_update_done();
447 return err;
448}
449EXPORT_SYMBOL_GPL(cpu_up);
450
451#ifdef CONFIG_PM_SLEEP_SMP
452static cpumask_var_t frozen_cpus;
453
454int disable_nonboot_cpus(void)
455{
456 int cpu, first_cpu, error = 0;
457
458 cpu_maps_update_begin();
459 first_cpu = cpumask_first(cpu_online_mask);
460 /*
461 * We take down all of the non-boot CPUs in one shot to avoid races
462 * with the userspace trying to use the CPU hotplug at the same time
463 */
464 cpumask_clear(frozen_cpus);
465
466 printk("Disabling non-boot CPUs ...\n");
467 for_each_online_cpu(cpu) {
468 if (cpu == first_cpu)
469 continue;
470 error = _cpu_down(cpu, 1);
471 if (!error)
472 cpumask_set_cpu(cpu, frozen_cpus);
473 else {
474 printk(KERN_ERR "Error taking CPU%d down: %d\n",
475 cpu, error);
476 break;
477 }
478 }
479
480 if (!error) {
481 BUG_ON(num_online_cpus() > 1);
482 /* Make sure the CPUs won't be enabled by someone else */
483 cpu_hotplug_disabled = 1;
484 } else {
485 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
486 }
487 cpu_maps_update_done();
488 return error;
489}
490
491void __weak arch_enable_nonboot_cpus_begin(void)
492{
493}
494
495void __weak arch_enable_nonboot_cpus_end(void)
496{
497}
498
499void __ref enable_nonboot_cpus(void)
500{
501 int cpu, error;
502
503 /* Allow everyone to use the CPU hotplug again */
504 cpu_maps_update_begin();
505 cpu_hotplug_disabled = 0;
506 if (cpumask_empty(frozen_cpus))
507 goto out;
508
509 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
510
511 arch_enable_nonboot_cpus_begin();
512
513 for_each_cpu(cpu, frozen_cpus) {
514 error = _cpu_up(cpu, 1);
515 if (!error) {
516 printk(KERN_INFO "CPU%d is up\n", cpu);
517 continue;
518 }
519 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
520 }
521
522 arch_enable_nonboot_cpus_end();
523
524 cpumask_clear(frozen_cpus);
525out:
526 cpu_maps_update_done();
527}
528
529static int __init alloc_frozen_cpus(void)
530{
531 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
532 return -ENOMEM;
533 return 0;
534}
535core_initcall(alloc_frozen_cpus);
536
537/*
538 * Prevent regular CPU hotplug from racing with the freezer, by disabling CPU
539 * hotplug when tasks are about to be frozen. Also, don't allow the freezer
540 * to continue until any currently running CPU hotplug operation gets
541 * completed.
542 * To modify the 'cpu_hotplug_disabled' flag, we need to acquire the
543 * 'cpu_add_remove_lock'. And this same lock is also taken by the regular
544 * CPU hotplug path and released only after it is complete. Thus, we
545 * (and hence the freezer) will block here until any currently running CPU
546 * hotplug operation gets completed.
547 */
548void cpu_hotplug_disable_before_freeze(void)
549{
550 cpu_maps_update_begin();
551 cpu_hotplug_disabled = 1;
552 cpu_maps_update_done();
553}
554
555
556/*
557 * When tasks have been thawed, re-enable regular CPU hotplug (which had been
558 * disabled while beginning to freeze tasks).
559 */
560void cpu_hotplug_enable_after_thaw(void)
561{
562 cpu_maps_update_begin();
563 cpu_hotplug_disabled = 0;
564 cpu_maps_update_done();
565}
566
567/*
568 * When callbacks for CPU hotplug notifications are being executed, we must
569 * ensure that the state of the system with respect to the tasks being frozen
570 * or not, as reported by the notification, remains unchanged *throughout the
571 * duration* of the execution of the callbacks.
572 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
573 *
574 * This synchronization is implemented by mutually excluding regular CPU
575 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
576 * Hibernate notifications.
577 */
578static int
579cpu_hotplug_pm_callback(struct notifier_block *nb,
580 unsigned long action, void *ptr)
581{
582 switch (action) {
583
584 case PM_SUSPEND_PREPARE:
585 case PM_HIBERNATION_PREPARE:
586 cpu_hotplug_disable_before_freeze();
587 break;
588
589 case PM_POST_SUSPEND:
590 case PM_POST_HIBERNATION:
591 cpu_hotplug_enable_after_thaw();
592 break;
593
594 default:
595 return NOTIFY_DONE;
596 }
597
598 return NOTIFY_OK;
599}
600
601
602static int __init cpu_hotplug_pm_sync_init(void)
603{
604 pm_notifier(cpu_hotplug_pm_callback, 0);
605 return 0;
606}
607core_initcall(cpu_hotplug_pm_sync_init);
608
609#endif /* CONFIG_PM_SLEEP_SMP */
610
611/**
612 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
613 * @cpu: cpu that just started
614 *
615 * This function calls the cpu_chain notifiers with CPU_STARTING.
616 * It must be called by the arch code on the new cpu, before the new cpu
617 * enables interrupts and before the "boot" cpu returns from __cpu_up().
618 */
619void __cpuinit notify_cpu_starting(unsigned int cpu)
620{
621 unsigned long val = CPU_STARTING;
622
623#ifdef CONFIG_PM_SLEEP_SMP
624 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
625 val = CPU_STARTING_FROZEN;
626#endif /* CONFIG_PM_SLEEP_SMP */
627 cpu_notify(val, (void *)(long)cpu);
628}
629
630#endif /* CONFIG_SMP */
631
632/*
633 * cpu_bit_bitmap[] is a special, "compressed" data structure that
634 * represents all NR_CPUS bits binary values of 1<<nr.
635 *
636 * It is used by cpumask_of() to get a constant address to a CPU
637 * mask value that has a single bit set only.
638 */
639
640/* cpu_bit_bitmap[0] is empty - so we can back into it */
641#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
642#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
643#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
644#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
645
646const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
647
648 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
649 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
650#if BITS_PER_LONG > 32
651 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
652 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
653#endif
654};
655EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
656
657const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
658EXPORT_SYMBOL(cpu_all_bits);
659
660#ifdef CONFIG_INIT_ALL_POSSIBLE
661static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
662 = CPU_BITS_ALL;
663#else
664static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
665#endif
666const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
667EXPORT_SYMBOL(cpu_possible_mask);
668
669static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
670const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
671EXPORT_SYMBOL(cpu_online_mask);
672
673static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
674const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
675EXPORT_SYMBOL(cpu_present_mask);
676
677static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
678const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
679EXPORT_SYMBOL(cpu_active_mask);
680
681void set_cpu_possible(unsigned int cpu, bool possible)
682{
683 if (possible)
684 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
685 else
686 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
687}
688
689void set_cpu_present(unsigned int cpu, bool present)
690{
691 if (present)
692 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
693 else
694 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
695}
696
697void set_cpu_online(unsigned int cpu, bool online)
698{
699 if (online)
700 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
701 else
702 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
703}
704
705void set_cpu_active(unsigned int cpu, bool active)
706{
707 if (active)
708 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
709 else
710 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
711}
712
713void init_cpu_present(const struct cpumask *src)
714{
715 cpumask_copy(to_cpumask(cpu_present_bits), src);
716}
717
718void init_cpu_possible(const struct cpumask *src)
719{
720 cpumask_copy(to_cpumask(cpu_possible_bits), src);
721}
722
723void init_cpu_online(const struct cpumask *src)
724{
725 cpumask_copy(to_cpumask(cpu_online_bits), src);
726}