projects/linux·blob v3.4-rc3 · refs

670 lines · 15.7 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#ifdef CONFIG_SMP
21/* Serializes the updates to cpu_online_mask, cpu_present_mask */
22static DEFINE_MUTEX(cpu_add_remove_lock);
23
24/*
25 * The following two API's must be used when attempting
26 * to serialize the updates to cpu_online_mask, cpu_present_mask.
27 */
28void cpu_maps_update_begin(void)
29{
30 mutex_lock(&cpu_add_remove_lock);
31}
32
33void cpu_maps_update_done(void)
34{
35 mutex_unlock(&cpu_add_remove_lock);
36}
37
38static RAW_NOTIFIER_HEAD(cpu_chain);
39
40/* If set, cpu_up and cpu_down will return -EBUSY and do nothing.
41 * Should always be manipulated under cpu_add_remove_lock
42 */
43static int cpu_hotplug_disabled;
44
45#ifdef CONFIG_HOTPLUG_CPU
46
47static struct {
48 struct task_struct *active_writer;
49 struct mutex lock; /* Synchronizes accesses to refcount, */
50 /*
51 * Also blocks the new readers during
52 * an ongoing cpu hotplug operation.
53 */
54 int refcount;
55} cpu_hotplug = {
56 .active_writer = NULL,
57 .lock = __MUTEX_INITIALIZER(cpu_hotplug.lock),
58 .refcount = 0,
59};
60
61void get_online_cpus(void)
62{
63 might_sleep();
64 if (cpu_hotplug.active_writer == current)
65 return;
66 mutex_lock(&cpu_hotplug.lock);
67 cpu_hotplug.refcount++;
68 mutex_unlock(&cpu_hotplug.lock);
69
70}
71EXPORT_SYMBOL_GPL(get_online_cpus);
72
73void put_online_cpus(void)
74{
75 if (cpu_hotplug.active_writer == current)
76 return;
77 mutex_lock(&cpu_hotplug.lock);
78 if (!--cpu_hotplug.refcount && unlikely(cpu_hotplug.active_writer))
79 wake_up_process(cpu_hotplug.active_writer);
80 mutex_unlock(&cpu_hotplug.lock);
81
82}
83EXPORT_SYMBOL_GPL(put_online_cpus);
84
85/*
86 * This ensures that the hotplug operation can begin only when the
87 * refcount goes to zero.
88 *
89 * Note that during a cpu-hotplug operation, the new readers, if any,
90 * will be blocked by the cpu_hotplug.lock
91 *
92 * Since cpu_hotplug_begin() is always called after invoking
93 * cpu_maps_update_begin(), we can be sure that only one writer is active.
94 *
95 * Note that theoretically, there is a possibility of a livelock:
96 * - Refcount goes to zero, last reader wakes up the sleeping
97 * writer.
98 * - Last reader unlocks the cpu_hotplug.lock.
99 * - A new reader arrives at this moment, bumps up the refcount.
100 * - The writer acquires the cpu_hotplug.lock finds the refcount
101 * non zero and goes to sleep again.
102 *
103 * However, this is very difficult to achieve in practice since
104 * get_online_cpus() not an api which is called all that often.
105 *
106 */
107static void cpu_hotplug_begin(void)
108{
109 cpu_hotplug.active_writer = current;
110
111 for (;;) {
112 mutex_lock(&cpu_hotplug.lock);
113 if (likely(!cpu_hotplug.refcount))
114 break;
115 __set_current_state(TASK_UNINTERRUPTIBLE);
116 mutex_unlock(&cpu_hotplug.lock);
117 schedule();
118 }
119}
120
121static void cpu_hotplug_done(void)
122{
123 cpu_hotplug.active_writer = NULL;
124 mutex_unlock(&cpu_hotplug.lock);
125}
126
127#else /* #if CONFIG_HOTPLUG_CPU */
128static void cpu_hotplug_begin(void) {}
129static void cpu_hotplug_done(void) {}
130#endif /* #else #if CONFIG_HOTPLUG_CPU */
131
132/* Need to know about CPUs going up/down? */
133int __ref register_cpu_notifier(struct notifier_block *nb)
134{
135 int ret;
136 cpu_maps_update_begin();
137 ret = raw_notifier_chain_register(&cpu_chain, nb);
138 cpu_maps_update_done();
139 return ret;
140}
141
142static int __cpu_notify(unsigned long val, void *v, int nr_to_call,
143 int *nr_calls)
144{
145 int ret;
146
147 ret = __raw_notifier_call_chain(&cpu_chain, val, v, nr_to_call,
148 nr_calls);
149
150 return notifier_to_errno(ret);
151}
152
153static int cpu_notify(unsigned long val, void *v)
154{
155 return __cpu_notify(val, v, -1, NULL);
156}
157
158#ifdef CONFIG_HOTPLUG_CPU
159
160static void cpu_notify_nofail(unsigned long val, void *v)
161{
162 BUG_ON(cpu_notify(val, v));
163}
164EXPORT_SYMBOL(register_cpu_notifier);
165
166void __ref unregister_cpu_notifier(struct notifier_block *nb)
167{
168 cpu_maps_update_begin();
169 raw_notifier_chain_unregister(&cpu_chain, nb);
170 cpu_maps_update_done();
171}
172EXPORT_SYMBOL(unregister_cpu_notifier);
173
174static inline void check_for_tasks(int cpu)
175{
176 struct task_struct *p;
177
178 write_lock_irq(&tasklist_lock);
179 for_each_process(p) {
180 if (task_cpu(p) == cpu && p->state == TASK_RUNNING &&
181 (p->utime || p->stime))
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}
382EXPORT_SYMBOL_GPL(cpu_up);
383
384#ifdef CONFIG_PM_SLEEP_SMP
385static cpumask_var_t frozen_cpus;
386
387void __weak arch_disable_nonboot_cpus_begin(void)
388{
389}
390
391void __weak arch_disable_nonboot_cpus_end(void)
392{
393}
394
395int disable_nonboot_cpus(void)
396{
397 int cpu, first_cpu, error = 0;
398
399 cpu_maps_update_begin();
400 first_cpu = cpumask_first(cpu_online_mask);
401 /*
402 * We take down all of the non-boot CPUs in one shot to avoid races
403 * with the userspace trying to use the CPU hotplug at the same time
404 */
405 cpumask_clear(frozen_cpus);
406 arch_disable_nonboot_cpus_begin();
407
408 printk("Disabling non-boot CPUs ...\n");
409 for_each_online_cpu(cpu) {
410 if (cpu == first_cpu)
411 continue;
412 error = _cpu_down(cpu, 1);
413 if (!error)
414 cpumask_set_cpu(cpu, frozen_cpus);
415 else {
416 printk(KERN_ERR "Error taking CPU%d down: %d\n",
417 cpu, error);
418 break;
419 }
420 }
421
422 arch_disable_nonboot_cpus_end();
423
424 if (!error) {
425 BUG_ON(num_online_cpus() > 1);
426 /* Make sure the CPUs won't be enabled by someone else */
427 cpu_hotplug_disabled = 1;
428 } else {
429 printk(KERN_ERR "Non-boot CPUs are not disabled\n");
430 }
431 cpu_maps_update_done();
432 return error;
433}
434
435void __weak arch_enable_nonboot_cpus_begin(void)
436{
437}
438
439void __weak arch_enable_nonboot_cpus_end(void)
440{
441}
442
443void __ref enable_nonboot_cpus(void)
444{
445 int cpu, error;
446
447 /* Allow everyone to use the CPU hotplug again */
448 cpu_maps_update_begin();
449 cpu_hotplug_disabled = 0;
450 if (cpumask_empty(frozen_cpus))
451 goto out;
452
453 printk(KERN_INFO "Enabling non-boot CPUs ...\n");
454
455 arch_enable_nonboot_cpus_begin();
456
457 for_each_cpu(cpu, frozen_cpus) {
458 error = _cpu_up(cpu, 1);
459 if (!error) {
460 printk(KERN_INFO "CPU%d is up\n", cpu);
461 continue;
462 }
463 printk(KERN_WARNING "Error taking CPU%d up: %d\n", cpu, error);
464 }
465
466 arch_enable_nonboot_cpus_end();
467
468 cpumask_clear(frozen_cpus);
469out:
470 cpu_maps_update_done();
471}
472
473static int __init alloc_frozen_cpus(void)
474{
475 if (!alloc_cpumask_var(&frozen_cpus, GFP_KERNEL|__GFP_ZERO))
476 return -ENOMEM;
477 return 0;
478}
479core_initcall(alloc_frozen_cpus);
480
481/*
482 * Prevent regular CPU hotplug from racing with the freezer, by disabling CPU
483 * hotplug when tasks are about to be frozen. Also, don't allow the freezer
484 * to continue until any currently running CPU hotplug operation gets
485 * completed.
486 * To modify the 'cpu_hotplug_disabled' flag, we need to acquire the
487 * 'cpu_add_remove_lock'. And this same lock is also taken by the regular
488 * CPU hotplug path and released only after it is complete. Thus, we
489 * (and hence the freezer) will block here until any currently running CPU
490 * hotplug operation gets completed.
491 */
492void cpu_hotplug_disable_before_freeze(void)
493{
494 cpu_maps_update_begin();
495 cpu_hotplug_disabled = 1;
496 cpu_maps_update_done();
497}
498
499
500/*
501 * When tasks have been thawed, re-enable regular CPU hotplug (which had been
502 * disabled while beginning to freeze tasks).
503 */
504void cpu_hotplug_enable_after_thaw(void)
505{
506 cpu_maps_update_begin();
507 cpu_hotplug_disabled = 0;
508 cpu_maps_update_done();
509}
510
511/*
512 * When callbacks for CPU hotplug notifications are being executed, we must
513 * ensure that the state of the system with respect to the tasks being frozen
514 * or not, as reported by the notification, remains unchanged *throughout the
515 * duration* of the execution of the callbacks.
516 * Hence we need to prevent the freezer from racing with regular CPU hotplug.
517 *
518 * This synchronization is implemented by mutually excluding regular CPU
519 * hotplug and Suspend/Hibernate call paths by hooking onto the Suspend/
520 * Hibernate notifications.
521 */
522static int
523cpu_hotplug_pm_callback(struct notifier_block *nb,
524 unsigned long action, void *ptr)
525{
526 switch (action) {
527
528 case PM_SUSPEND_PREPARE:
529 case PM_HIBERNATION_PREPARE:
530 cpu_hotplug_disable_before_freeze();
531 break;
532
533 case PM_POST_SUSPEND:
534 case PM_POST_HIBERNATION:
535 cpu_hotplug_enable_after_thaw();
536 break;
537
538 default:
539 return NOTIFY_DONE;
540 }
541
542 return NOTIFY_OK;
543}
544
545
546static int __init cpu_hotplug_pm_sync_init(void)
547{
548 pm_notifier(cpu_hotplug_pm_callback, 0);
549 return 0;
550}
551core_initcall(cpu_hotplug_pm_sync_init);
552
553#endif /* CONFIG_PM_SLEEP_SMP */
554
555/**
556 * notify_cpu_starting(cpu) - call the CPU_STARTING notifiers
557 * @cpu: cpu that just started
558 *
559 * This function calls the cpu_chain notifiers with CPU_STARTING.
560 * It must be called by the arch code on the new cpu, before the new cpu
561 * enables interrupts and before the "boot" cpu returns from __cpu_up().
562 */
563void __cpuinit notify_cpu_starting(unsigned int cpu)
564{
565 unsigned long val = CPU_STARTING;
566
567#ifdef CONFIG_PM_SLEEP_SMP
568 if (frozen_cpus != NULL && cpumask_test_cpu(cpu, frozen_cpus))
569 val = CPU_STARTING_FROZEN;
570#endif /* CONFIG_PM_SLEEP_SMP */
571 cpu_notify(val, (void *)(long)cpu);
572}
573
574#endif /* CONFIG_SMP */
575
576/*
577 * cpu_bit_bitmap[] is a special, "compressed" data structure that
578 * represents all NR_CPUS bits binary values of 1<<nr.
579 *
580 * It is used by cpumask_of() to get a constant address to a CPU
581 * mask value that has a single bit set only.
582 */
583
584/* cpu_bit_bitmap[0] is empty - so we can back into it */
585#define MASK_DECLARE_1(x) [x+1][0] = (1UL << (x))
586#define MASK_DECLARE_2(x) MASK_DECLARE_1(x), MASK_DECLARE_1(x+1)
587#define MASK_DECLARE_4(x) MASK_DECLARE_2(x), MASK_DECLARE_2(x+2)
588#define MASK_DECLARE_8(x) MASK_DECLARE_4(x), MASK_DECLARE_4(x+4)
589
590const unsigned long cpu_bit_bitmap[BITS_PER_LONG+1][BITS_TO_LONGS(NR_CPUS)] = {
591
592 MASK_DECLARE_8(0), MASK_DECLARE_8(8),
593 MASK_DECLARE_8(16), MASK_DECLARE_8(24),
594#if BITS_PER_LONG > 32
595 MASK_DECLARE_8(32), MASK_DECLARE_8(40),
596 MASK_DECLARE_8(48), MASK_DECLARE_8(56),
597#endif
598};
599EXPORT_SYMBOL_GPL(cpu_bit_bitmap);
600
601const DECLARE_BITMAP(cpu_all_bits, NR_CPUS) = CPU_BITS_ALL;
602EXPORT_SYMBOL(cpu_all_bits);
603
604#ifdef CONFIG_INIT_ALL_POSSIBLE
605static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly
606 = CPU_BITS_ALL;
607#else
608static DECLARE_BITMAP(cpu_possible_bits, CONFIG_NR_CPUS) __read_mostly;
609#endif
610const struct cpumask *const cpu_possible_mask = to_cpumask(cpu_possible_bits);
611EXPORT_SYMBOL(cpu_possible_mask);
612
613static DECLARE_BITMAP(cpu_online_bits, CONFIG_NR_CPUS) __read_mostly;
614const struct cpumask *const cpu_online_mask = to_cpumask(cpu_online_bits);
615EXPORT_SYMBOL(cpu_online_mask);
616
617static DECLARE_BITMAP(cpu_present_bits, CONFIG_NR_CPUS) __read_mostly;
618const struct cpumask *const cpu_present_mask = to_cpumask(cpu_present_bits);
619EXPORT_SYMBOL(cpu_present_mask);
620
621static DECLARE_BITMAP(cpu_active_bits, CONFIG_NR_CPUS) __read_mostly;
622const struct cpumask *const cpu_active_mask = to_cpumask(cpu_active_bits);
623EXPORT_SYMBOL(cpu_active_mask);
624
625void set_cpu_possible(unsigned int cpu, bool possible)
626{
627 if (possible)
628 cpumask_set_cpu(cpu, to_cpumask(cpu_possible_bits));
629 else
630 cpumask_clear_cpu(cpu, to_cpumask(cpu_possible_bits));
631}
632
633void set_cpu_present(unsigned int cpu, bool present)
634{
635 if (present)
636 cpumask_set_cpu(cpu, to_cpumask(cpu_present_bits));
637 else
638 cpumask_clear_cpu(cpu, to_cpumask(cpu_present_bits));
639}
640
641void set_cpu_online(unsigned int cpu, bool online)
642{
643 if (online)
644 cpumask_set_cpu(cpu, to_cpumask(cpu_online_bits));
645 else
646 cpumask_clear_cpu(cpu, to_cpumask(cpu_online_bits));
647}
648
649void set_cpu_active(unsigned int cpu, bool active)
650{
651 if (active)
652 cpumask_set_cpu(cpu, to_cpumask(cpu_active_bits));
653 else
654 cpumask_clear_cpu(cpu, to_cpumask(cpu_active_bits));
655}
656
657void init_cpu_present(const struct cpumask *src)
658{
659 cpumask_copy(to_cpumask(cpu_present_bits), src);
660}
661
662void init_cpu_possible(const struct cpumask *src)
663{
664 cpumask_copy(to_cpumask(cpu_possible_bits), src);
665}
666
667void init_cpu_online(const struct cpumask *src)
668{
669 cpumask_copy(to_cpumask(cpu_online_bits), src);
670}