projects/linux·blob da89fb16 · refs

681 lines · 15.9 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/export.h>
14#include <linux/kthread.h>
15#include <linux/stop_machine.h>
16#include <linux/mutex.h>
17#include <linux/gfp.h>
18#include <linux/suspend.h>
19
20#include "smpboot.h"
21
22#ifdef CONFIG_SMP
23/* Serializes the updates to cpu_online_mask, cpu_present_mask */
24static DEFINE_MUTEX(cpu_add_remove_lock);
25
26/*
27 * The following two API's must be used when attempting
28 * to serialize the updates to cpu_online_mask, cpu_present_mask.
29 */
30void cpu_maps_update_begin(void)
31{
32 mutex_lock(&cpu_add_remove_lock);
33}
34
35void cpu_maps_update_done(void)
36{
37 mutex_unlock(&cpu_add_remove_lock);
38}
39
40static RAW_NOTIFIER_HEAD(cpu_chain);
41
42/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
43 * Should always be manipulated under cpu_add_remove_lock
44 */
45static int cpu_hotplug_disabled;
46
47#ifdef CONFIG_HOTPLUG_CPU
48
49static struct {
50 struct task_struct *active_writer;
51 struct mutex lock; /* Synchronizes accesses to refcount, */
52 /*
53 * Also blocks the new readers during
54 * an ongoing cpu hotplug operation.
55 */
56 int refcount;
57} cpu_hotplug = {
58 .active_writer = NULL,
59 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
60 .refcount = 0,
61};
62
63void get_online_cpus(void)
64{
65 might_sleep();
66 if (cpu_hotplug.active_writer == current)
67 return;
68 mutex_lock(&cpu_hotplug.lock);
69 cpu_hotplug.refcount++;
70 mutex_unlock(&cpu_hotplug.lock);
71
72}
73EXPORT_SYMBOL_GPL(get_online_cpus);
74
75void put_online_cpus(void)
76{
77 if (cpu_hotplug.active_writer == current)
78 return;
79 mutex_lock(&cpu_hotplug.lock);
80 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
81 wake_up_process(cpu_hotplug.active_writer);
82 mutex_unlock(&cpu_hotplug.lock);
83
84}
85EXPORT_SYMBOL_GPL(put_online_cpus);
86
87/*
88 * This ensures that the hotplug operation can begin only when the
89 * refcount goes to zero.
90 *
91 * Note that during a cpu-hotplug operation, the new readers, if any,
92 * will be blocked by the cpu_hotplug.lock
93 *
94 * Since cpu_hotplug_begin() is always called after invoking
95 * cpu_maps_update_begin(), we can be sure that only one writer is active.
96 *
97 * Note that theoretically, there is a possibility of a livelock:
98 * - Refcount goes to zero, last reader wakes up the sleeping
99 * writer.
100 * - Last reader unlocks the cpu_hotplug.lock.
101 * - A new reader arrives at this moment, bumps up the refcount.
102 * - The writer acquires the cpu_hotplug.lock finds the refcount
103 * non zero and goes to sleep again.
104 *
105 * However, this is very difficult to achieve in practice since
106 * get_online_cpus() not an api which is called all that often.
107 *
108 */
109static void cpu_hotplug_begin(void)
110{
111 cpu_hotplug.active_writer = current;
112
113 for (;;) {
114 mutex_lock(&cpu_hotplug.lock);
115 if (likely(!cpu_hotplug.refcount))
116 break;
117 __set_current_state(TASK_UNINTERRUPTIBLE);
118 mutex_unlock(&cpu_hotplug.lock);
119 schedule();
120 }
121}
122
123static void cpu_hotplug_done(void)
124{
125 cpu_hotplug.active_writer = NULL;
126 mutex_unlock(&cpu_hotplug.lock);
127}
128
129#else /* #if CONFIG_HOTPLUG_CPU */
130static void cpu_hotplug_begin(void) {}
131static void cpu_hotplug_done(void) {}
132#endif /* #else #if CONFIG_HOTPLUG_CPU */
133
134/* Need to know about CPUs going up/down? */
135int __ref register_cpu_notifier(struct notifier_block *nb)
136{
137 int ret;
138 cpu_maps_update_begin();
139 ret = raw_notifier_chain_register(&cpu_chain, nb);
140 cpu_maps_update_done();
141 return ret;
142}
143
144static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
145 int *nr_calls)
146{
147 int ret;
148
149 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
150 nr_calls);
151
152 return notifier_to_errno(ret);
153}
154
155static int cpu_notify(unsigned long val, void *v)
156{
157 return __cpu_notify(val, v, -1, NULL);
158}
159
160#ifdef CONFIG_HOTPLUG_CPU
161
162static void cpu_notify_nofail(unsigned long val, void *v)
163{
164 BUG_ON(cpu_notify(val, v));
165}
166EXPORT_SYMBOL(register_cpu_notifier);
167
168void __ref unregister_cpu_notifier(struct notifier_block *nb)
169{
170 cpu_maps_update_begin();
171 raw_notifier_chain_unregister(&cpu_chain, nb);
172 cpu_maps_update_done();
173}
174EXPORT_SYMBOL(unregister_cpu_notifier);
175
176static inline void check_for_tasks(int cpu)
177{
178 struct task_struct *p;
179
180 write_lock_irq(&tasklist_lock);
181 for_each_process(p) {
182 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
183 (p->utime || p->stime))
184 printk(KERN_WARNING "Task %s (pid = %d) is on cpu %d "
185 "(state = %ld, flags = %x)\n",
186 p->comm, task_pid_nr(p), cpu,
187 p->state, p->flags);
188 }
189 write_unlock_irq(&tasklist_lock);
190}
191
192struct take_cpu_down_param {
193 unsigned long mod;
194 void *hcpu;
195};
196
197/* Take this CPU down. */
198static int __ref take_cpu_down(void *_param)
199{
200 struct take_cpu_down_param *param = _param;
201 int err;
202
203 /* Ensure this CPU doesn't handle any more interrupts. */
204 err = __cpu_disable();
205 if (err < 0)
206 return err;
207
208 cpu_notify(CPU_DYING | param->mod, param->hcpu);
209 return 0;
210}
211
212/* Requires cpu_add_remove_lock to be held */
213static int __ref _cpu_down(unsigned int cpu, int tasks_frozen)
214{
215 int err, nr_calls = 0;
216 void *hcpu = (void *)(long)cpu;
217 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
218 struct take_cpu_down_param tcd_param = {
219 .mod = mod,
220 .hcpu = hcpu,
221 };
222
223 if (num_online_cpus() == 1)
224 return -EBUSY;
225
226 if (!cpu_online(cpu))
227 return -EINVAL;
228
229 cpu_hotplug_begin();
230
231 err = __cpu_notify(CPU_DOWN_PREPARE | mod, hcpu, -1, &nr_calls);
232 if (err) {
233 nr_calls--;
234 __cpu_notify(CPU_DOWN_FAILED | mod, hcpu, nr_calls, NULL);
235 printk("%s: attempt to take down CPU %u failed\n",
236 __func__, cpu);
237 goto out_release;
238 }
239
240 err = __stop_machine(take_cpu_down, &tcd_param, cpumask_of(cpu));
241 if (err) {
242 /* CPU didn't die: tell everyone. Can't complain. */
243 cpu_notify_nofail(CPU_DOWN_FAILED | mod, hcpu);
244
245 goto out_release;
246 }
247 BUG_ON(cpu_online(cpu));
248
249 /*
250 * The migration_call() CPU_DYING callback will have removed all
251 * runnable tasks from the cpu, there's only the idle task left now
252 * that the migration thread is done doing the stop_machine thing.
253 *
254 * Wait for the stop thread to go away.
255 */
256 while (!idle_cpu(cpu))
257 cpu_relax();
258
259 /* This actually kills the CPU. */
260 __cpu_die(cpu);
261
262 /* CPU is completely dead: tell everyone. Too late to complain. */
263 cpu_notify_nofail(CPU_DEAD | mod, hcpu);
264
265 check_for_tasks(cpu);
266
267out_release:
268 cpu_hotplug_done();
269 if (!err)
270 cpu_notify_nofail(CPU_POST_DEAD | mod, hcpu);
271 return err;
272}
273
274int __ref cpu_down(unsigned int cpu)
275{
276 int err;
277
278 cpu_maps_update_begin();
279
280 if (cpu_hotplug_disabled) {
281 err = -EBUSY;
282 goto out;
283 }
284
285 err = _cpu_down(cpu, 0);
286
287out:
288 cpu_maps_update_done();
289 return err;
290}
291EXPORT_SYMBOL(cpu_down);
292#endif /*CONFIG_HOTPLUG_CPU*/
293
294/* Requires cpu_add_remove_lock to be held */
295static int __cpuinit _cpu_up(unsigned int cpu, int tasks_frozen)
296{
297 int ret, nr_calls = 0;
298 void *hcpu = (void *)(long)cpu;
299 unsigned long mod = tasks_frozen ? CPU_TASKS_FROZEN : 0;
300 struct task_struct *idle;
301
302 if (cpu_online(cpu) || !cpu_present(cpu))
303 return -EINVAL;
304
305 cpu_hotplug_begin();
306
307 idle = idle_thread_get(cpu);
308 if (IS_ERR(idle)) {
309 ret = PTR_ERR(idle);
310 goto out;
311 }
312
313 ret = __cpu_notify(CPU_UP_PREPARE | mod, hcpu, -1, &nr_calls);
314 if (ret) {
315 nr_calls--;
316 printk(KERN_WARNING "%s: attempt to bring up CPU %u failed\n",
317 __func__, cpu);
318 goto out_notify;
319 }
320
321 /* Arch-specific enabling code. */
322 ret = __cpu_up(cpu, idle);
323 if (ret != 0)
324 goto out_notify;
325 BUG_ON(!cpu_online(cpu));
326
327 /* Now call notifier in preparation. */
328 cpu_notify(CPU_ONLINE | mod, hcpu);
329
330out_notify:
331 if (ret != 0)
332 __cpu_notify(CPU_UP_CANCELED | mod, hcpu, nr_calls, NULL);
333out:
334 cpu_hotplug_done();
335
336 return ret;
337}
338
339int __cpuinit cpu_up(unsigned int cpu)
340{
341 int err = 0;
342
343#ifdef CONFIG_MEMORY_HOTPLUG
344 int nid;
345 pg_data_t *pgdat;
346#endif
347
348 if (!cpu_possible(cpu)) {
349 printk(KERN_ERR "can't online cpu %d because it is not "
350 "configured as may-hotadd at boot time\n", cpu);
351#if defined(CONFIG_IA64)
352 printk(KERN_ERR "please check additional_cpus= boot "
353 "parameter\n");
354#endif
355 return -EINVAL;
356 }
357
358#ifdef CONFIG_MEMORY_HOTPLUG
359 nid = cpu_to_node(cpu);
360 if (!node_online(nid)) {
361 err = mem_online_node(nid);
362 if (err)
363 return err;
364 }
365
366 pgdat = NODE_DATA(nid);
367 if (!pgdat) {
368 printk(KERN_ERR
369 "Can't online cpu %d due to NULL pgdat\n", cpu);
370 return -ENOMEM;
371 }
372
373 if (pgdat->node_zonelists->_zonerefs->zone == NULL) {
374 mutex_lock(&zonelists_mutex);
375 build_all_zonelists(NULL);
376 mutex_unlock(&zonelists_mutex);
377 }
378#endif
379
380 cpu_maps_update_begin();
381
382 if (cpu_hotplug_disabled) {
383 err = -EBUSY;
384 goto out;
385 }
386
387 err = _cpu_up(cpu, 0);
388
389out:
390 cpu_maps_update_done();
391 return err;
392}
393EXPORT_SYMBOL_GPL(cpu_up);
394
395#ifdef CONFIG_PM_SLEEP_SMP
396static cpumask_var_t frozen_cpus;
397
398void __weak arch_disable_nonboot_cpus_begin(void)
399{
400}
401
402void __weak arch_disable_nonboot_cpus_end(void)
403{
404}
405
406int disable_nonboot_cpus(void)
407{
408 int cpu, first_cpu, error = 0;
409
410 cpu_maps_update_begin();
411 first_cpu = cpumask_first(cpu_online_mask);
412 /*
413 * We take down all of the non-boot CPUs in one shot to avoid races
414 * with the userspace trying to use the CPU hotplug at the same time
415 */
416 cpumask_clear(frozen_cpus);
417 arch_disable_nonboot_cpus_begin();
418
419 printk("Disabling non-boot CPUs ...\n");
420 for_each_online_cpu(cpu) {
421 if (cpu == first_cpu)
422 continue;
423 error = _cpu_down(cpu, 1);
424 if (!error)
425 cpumask_set_cpu(cpu, frozen_cpus);
426 else {
427 printk(KERN_ERR "Error taking CPU%d down: %d\n",
428 cpu, error);
429 break;
430 }
431 }
432
433 arch_disable_nonboot_cpus_end();
434
435 if (!error) {
436 BUG_ON(num_online_cpus() > 1);
437 /* Make sure the CPUs won't be enabled by someone else */
438 cpu_hotplug_disabled = 1;
439 } else {
440 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
441 }
442 cpu_maps_update_done();
443 return error;
444}
445
446void __weak arch_enable_nonboot_cpus_begin(void)
447{
448}
449
450void __weak arch_enable_nonboot_cpus_end(void)
451{
452}
453
454void __ref enable_nonboot_cpus(void)
455{
456 int cpu, error;
457
458 /* Allow everyone to use the CPU hotplug again */
459 cpu_maps_update_begin();
460 cpu_hotplug_disabled = 0;
461 if (cpumask_empty(frozen_cpus))
462 goto out;
463
464 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
465
466 arch_enable_nonboot_cpus_begin();
467
468 for_each_cpu(cpu, frozen_cpus) {
469 error = _cpu_up(cpu, 1);
470 if (!error) {
471 printk(KERN_INFO "CPU%d is up\n", cpu);
472 continue;
473 }
474 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
475 }
476
477 arch_enable_nonboot_cpus_end();
478
479 cpumask_clear(frozen_cpus);
480out:
481 cpu_maps_update_done();
482}
483
484static int __init alloc_frozen_cpus(void)
485{
486 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
487 return -ENOMEM;
488 return 0;
489}
490core_initcall(alloc_frozen_cpus);
491
492/*
493 * Prevent regular CPU hotplug from racing with the freezer, by disabling CPU
494 * hotplug when tasks are about to be frozen. Also, don't allow the freezer
495 * to continue until any currently running CPU hotplug operation gets
496 * completed.
497 * To modify the 'cpu_hotplug_disabled' flag, we need to acquire the
498 * 'cpu_add_remove_lock'. And this same lock is also taken by the regular
499 * CPU hotplug path and released only after it is complete. Thus, we
500 * (and hence the freezer) will block here until any currently running CPU
501 * hotplug operation gets completed.
502 */
503void cpu_hotplug_disable_before_freeze(void)
504{
505 cpu_maps_update_begin();
506 cpu_hotplug_disabled = 1;
507 cpu_maps_update_done();
508}
509
510
511/*
512 * When tasks have been thawed, re-enable regular CPU hotplug (which had been
513 * disabled while beginning to freeze tasks).
514 */
515void cpu_hotplug_enable_after_thaw(void)
516{
517 cpu_maps_update_begin();
518 cpu_hotplug_disabled = 0;
519 cpu_maps_update_done();
520}
521
522/*
523 * When callbacks for CPU hotplug notifications are being executed, we must
524 * ensure that the state of the system with respect to the tasks being frozen
525 * or not, as reported by the notification, remains unchanged *throughout the
526 * duration* of the execution of the callbacks.
527 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
528 *
529 * This synchronization is implemented by mutually excluding regular CPU
530 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
531 * Hibernate notifications.
532 */
533static int
534cpu_hotplug_pm_callback(struct notifier_block *nb,
535 unsigned long action, void *ptr)
536{
537 switch (action) {
538
539 case PM_SUSPEND_PREPARE:
540 case PM_HIBERNATION_PREPARE:
541 cpu_hotplug_disable_before_freeze();
542 break;
543
544 case PM_POST_SUSPEND:
545 case PM_POST_HIBERNATION:
546 cpu_hotplug_enable_after_thaw();
547 break;
548
549 default:
550 return NOTIFY_DONE;
551 }
552
553 return NOTIFY_OK;
554}
555
556
557static int __init cpu_hotplug_pm_sync_init(void)
558{
559 pm_notifier(cpu_hotplug_pm_callback, 0);
560 return 0;
561}
562core_initcall(cpu_hotplug_pm_sync_init);
563
564#endif /* CONFIG_PM_SLEEP_SMP */
565
566/**
567 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
568 * @cpu: cpu that just started
569 *
570 * This function calls the cpu_chain notifiers with CPU_STARTING.
571 * It must be called by the arch code on the new cpu, before the new cpu
572 * enables interrupts and before the "boot" cpu returns from __cpu_up().
573 */
574void __cpuinit notify_cpu_starting(unsigned int cpu)
575{
576 unsigned long val = CPU_STARTING;
577
578#ifdef CONFIG_PM_SLEEP_SMP
579 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
580 val = CPU_STARTING_FROZEN;
581#endif /* CONFIG_PM_SLEEP_SMP */
582 cpu_notify(val, (void *)(long)cpu);
583}
584
585#endif /* CONFIG_SMP */
586
587/*
588 * cpu_bit_bitmap[] is a special, "compressed" data structure that
589 * represents all NR_CPUS bits binary values of 1<<nr.
590 *
591 * It is used by cpumask_of() to get a constant address to a CPU
592 * mask value that has a single bit set only.
593 */
594
595/* cpu_bit_bitmap[0] is empty - so we can back into it */
596#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
597#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
598#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
599#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
600
601const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
602
603 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
604 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
605#if BITS_PER_LONG > 32
606 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
607 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
608#endif
609};
610EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
611
612const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
613EXPORT_SYMBOL(cpu_all_bits);
614
615#ifdef CONFIG_INIT_ALL_POSSIBLE
616static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
617 = CPU_BITS_ALL;
618#else
619static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
620#endif
621const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
622EXPORT_SYMBOL(cpu_possible_mask);
623
624static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
625const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
626EXPORT_SYMBOL(cpu_online_mask);
627
628static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
629const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
630EXPORT_SYMBOL(cpu_present_mask);
631
632static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
633const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
634EXPORT_SYMBOL(cpu_active_mask);
635
636void set_cpu_possible(unsigned int cpu, bool possible)
637{
638 if (possible)
639 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
640 else
641 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
642}
643
644void set_cpu_present(unsigned int cpu, bool present)
645{
646 if (present)
647 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
648 else
649 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
650}
651
652void set_cpu_online(unsigned int cpu, bool online)
653{
654 if (online)
655 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
656 else
657 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
658}
659
660void set_cpu_active(unsigned int cpu, bool active)
661{
662 if (active)
663 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
664 else
665 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
666}
667
668void init_cpu_present(const struct cpumask *src)
669{
670 cpumask_copy(to_cpumask(cpu_present_bits), src);
671}
672
673void init_cpu_possible(const struct cpumask *src)
674{
675 cpumask_copy(to_cpumask(cpu_possible_bits), src);
676}
677
678void init_cpu_online(const struct cpumask *src)
679{
680 cpumask_copy(to_cpumask(cpu_online_bits), src);
681}